content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
Gaussian processes (GPs) are flexible Bayesian non-parametric models that have achieved state-of-the-art performance in a range of applications~\citep{diggle2007springer,Snoek2012}. A key advantage of GP models is that they have large representational capacity, while being robust to overfitting~\citep{Rasmussen2006}. %
This property is especially important for robotic applications, where there may be an abundance of data in some parts of the space but a scarcity in others~\citep{deisenroth2011pilco}. Unfortunately, exact inference in GPs scales cubically in computation and quadratically in memory with the size of the training set, and is only available in closed form for Gaussian likelihoods.
To learn from large datasets, variational inference provides a principled way to find tractable approximations to the true posterior. A common approach to approximate GP inference is to form a sparse variational posterior, which is designed by conditioning the prior process at a small set of \emph{inducing points}~\citep{titsias2009variational}. The sparse variational framework trades accuracy against computation, but its complexities still scale superlinearly in the number of inducing points. %
Consequently, the representation power of the approximate distribution is greatly limited.
Various attempts have been made to reduce the complexities in order to scale up GP models for better approximation.
Most of them, however, rely on certain assumptions on the kernel structure and input dimension. In the extreme, \citet{hartikainen2010kalman} show that, for 1D-input problems, exact GP inference can be solved in linear time for kernels with finitely many non-zero derivatives.
For low-dimensional inputs and stationary kernels, variational inference with
structured kernel approximation~\citep{wilson2015kernel} or Fourier features~\citep{hensman2016variational} has been proposed.
Both approaches, nevertheless, scale exponentially with input dimension,
except for the special case of sum-and-product kernels~\citep{gardner2018product}.
Approximate kernels have also been proposed as GP priors with low-rank structure~\citep{snelson2006sparse, quinonero2005unifying} or a sparse spectrum~\citep{quia2010sparse}.
Another family of methods partitions the input space into subsets and performs prediction aggregation~\citep{tresp2000bayesian, rasmussen2002infinite, nguyen2014fast,deisenroth2015distributed}, and
Bayesian aggregation of local experts with attractive theoretical properties is recently proposed by~\citet{rulliere2018nested}.
A recent decoupled framework~\citep{cheng2017variational} takes a different direction to address the complexity issue of GP inference.
In contrast to the above approaches, this decoupled framework is agnostic to problem setups (e.g. likelihoods, kernels, and input dimensions) and extends the original sparse variational formulation~\citep{titsias2009variational}.
The key idea is to represent the variational distribution in the reproducing kernel Hilbert space (RKHS) induced by the covariance function of the GP. The sparse variational posterior by~\citet{titsias2009variational} turns out to be equivalent to a particular parameterization in the RKHS, where the mean and covariance both share the same basis.
\citet{cheng2017variational} suggest to relax the requirement of basis sharing. Since the computation only scales linearly in the mean parameters, many more basis functions can be used for modeling the mean function to achieve higher accuracy in prediction.
However, the original decoupled basis~\citep{cheng2017variational} turns out to have optimization difficulties~\citep{havasi2018deep}.
In particular, the non-convexity of the optimization problem means that a suboptimal solution may be found, leading to performance that is potentially worse than the standard coupled case.
While \citet{havasi2018deep} suggest to use a pre-conditioner to amortize the problem, their algorithm incurs an additional cubic computational cost; therefore, its applicability is limited to small simple models.%
Inspired by the success of natural gradients in variational inference~\citep{hoffman2013stochastic, salimbeni2018natural}, we propose a novel RKHS parameterization of decoupled GPs that admits efficient natural gradient computation. We decompose the mean parametrization into a part that shares the basis with the covariance, and an orthogonal part that models the residues that the standard coupled approach fails to capture. We show that, with this particular choice, the natural gradient update rules further \emph{decouple} into the natural gradient descent of the coupled part and the functional gradient descent of the residual part. Based on these insights, we propose an efficient optimization algorithm that preserves the desired properties of decoupled GPs and converges faster than the original formulation~\citep{cheng2017variational}.
We demonstrate that our basis is more effective than the original decoupled formulation on a range of classification and regression tasks. We show that the natural gradient updates improve convergence considerably and can lead to much better performance in practice. Crucially, we show also that our basis is more effective than the standard coupled basis for a fixed computational budget.
\section{Background}
We consider the inference problem of GP models. Given a dataset $\mathcal{D}}\def\EE{\mathcal{E}}\def\FF{\mathcal{F} = \{(x_n, y_n) \}_{n=1}^N$ and a GP prior on a latent function $f$, the goal is to infer the (approximate) posterior of $f(x^*)$ for any query input $x^*$. In this work, we adopt the recent decoupled RKHS reformulation of variational inference~\citep{cheng2017variational}, and, without loss of generality, we will assume $f$ is a scalar function.
For notation, we use boldface to distinguish finite-dimensional vectors and matrices that are used in computation from scalar and abstract mathematical objects.
\subsection{Gaussian Processes and their RKHS Representation} \label{sec:GP and RKHS}
We first review the primal and dual representations of GPs, which form the foundation of the RKHS reformulation.
Let $\XX \subseteq \Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb^d$ be the domain of the latent function. A GP is a distribution of functions, which is described by a mean function $m: \XX \to \Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$ and a covariance function $k: \XX \times \XX \to \Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$.
We say a latent function $f$ is distributed according to $\GG\PP(m, k)$, if for any $x, x' \in \XX$,
$\Ebb[f(x)] = m(x)$, $\C[f(x), f(x')] = k(x,x')$, and for any finite subset $\{ f(x_n) : x_n \in \XX \}_{n=1}^N$ is Gaussian distributed. %
We call the above definition, in terms of the \emph{function values} $m(x)$ and $k(x,x')$, the \emph{primal} representation of GPs. Alternatively, one can adopt a \emph{dual} representation of GPs, by treating \emph{functions} $m$ and $k$ as RKHS objects~\citep{cheng2016incremental}.
This is based on observing that the covariance function $k$ satisfies the definition of positive semi-definite functions, %
so $k$ can also be viewed as a reproducing kernel \citep{aronszajn1950theory}.
Specifically, given $\GG\PP(m, k)$, without loss of generality, we can find an RKHS $\HH$ such that
\begin{align}\label{eq:RKHS form}
m(x) = \phi(x)^\top \mu, \qquad k(x,x') = \phi(x)^\top \Sigma \phi(x')
\end{align}
for some $\mu \in \HH$, bounded positive semidefinite self-adjoint operator $\Sigma: \HH \to \HH$, and feature map $\phi: \XX \to \HH$. Here we use $^\top$ to denote the inner product in $\HH$, even when $\dim{\HH}$ is infinite.
For notational clarity, we use symbols $m$ and $k$ (or $s$) to denote the mean and covariance functions, and symbols $\mu$ and $\Sigma$ to denote the RKHS objects; we use $s$ to distinguish the (approximate) posterior covariance function from the prior covariance function $k$.
If $ f \sim \GG\PP(m, k)$ satisfying~\eqref{eq:RKHS form}, we also write $f \sim \GG\PP_\HH( \mu, \Sigma)$.\footnote{
This notation only denotes that $m$ and $k$ can be represented as RKHS objects, not that the sampled functions of $\GG\PP(m, k)$ necessarily reside in $\HH$ (which only holds for the special when $\Sigma$ has finite trace).}
To concretely illustrate the primal-dual connection, we consider the GP regression problem. Suppose $f \sim \GG\PP(0,k)$ in prior and $y_n = f(x_n) + \epsilon_n$, where $\epsilon_n \sim \NN(\epsilon_n | 0, \sigma^2)$.
Let $X = \{x_n\}_{n=1}^N$ and $\mathbf{y}}\def\zb{\mathbf{z} = (y_n)_{n=1}^N \in \Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb^N$, where the notation $(\cdot)_{n = \cdot}^\cdot$ denotes stacking the elements.
Then, with $\mathcal{D}}\def\EE{\mathcal{E}}\def\FF{\mathcal{F}$ observed, it can be shown that $f \sim \GG\PP(m, s)$ where
\begin{align} \label{eq:GPR}
m(x) = \kb_{x,X} (\Kb_X + \sigma^2 \Ib)^{-1} \mathbf{y}}\def\zb{\mathbf{z}, \qquad
s(x,x') = k_{x,x'} - \kb_{x,X} (\Kb_X + \sigma^2 \Ib)^{-1} \kb_{X,x'},
\end{align}
where $k_{\cdot,\cdot}$, $\mbf{k_{\cdot,\cdot}}$ and $\mbf{K}_{\cdot,\cdot}$ denote the covariances between the subscripted sets,\footnote{If the two sets are the same, only one is listed.}
We can also equivalently write the posterior GP in~\eqref{eq:GPR} in its dual RKHS representation: suppose the feature map $\phi$ is selected such that $k(x,x') = \phi(x)^\top \phi(x')$, then \textit{a priori} $f \sim \GG\PP_\HH( 0, I)$ and \textit{a posteriori} %
$f \sim \GG\PP_\HH( \mu, \Sigma)$, %
\begin{align} \label{eq:GPR in RKHS}
\mu = \Phi_{X} (\Kb_X + \sigma^2 \Ib)^{-1} \mathbf{y}}\def\zb{\mathbf{z}, \qquad
\Sigma = I - \Phi_{X} (\Kb_X + \sigma^2 \Ib)^{-1} \Phi_{X}^\top,
\end{align}
where $\Phi_{X} = [\phi(x_1), \dots, \phi(x_N)]$.
\subsection{Variational Inference Problem}
Inference in GP models is challenging because the closed-form expressions in~\eqref{eq:GPR} have computational complexity that is cubic in the size of the training dataset, and are only applicable for Gaussian likelihoods.
For non-Gaussian likelihoods (e.g. classification) or for large datasets (i.e. more than 10,000 data points), we must adopt approximate inference.
Variational inference provides a principled approach to search for an approximate but tractable posterior.
It seeks a variational posterior $q$ that is close to the true posterior $p(f | \mathcal{D}}\def\EE{\mathcal{E}}\def\FF{\mathcal{F})$ in terms of KL divergence, i.e. it solves $\min_{q} \KL{q(f)}{p(f | \mathcal{D}}\def\EE{\mathcal{E}}\def\FF{\mathcal{F})}$.
For GP models, the variational posterior must be defined over the entire function, so a natural choice is to use another GP. This choice is also motivated by the fact that the exact posterior is a GP in the case of a Gaussian likelihood as shown in~\eqref{eq:GPR}.
Using the results from Section~\ref{sec:GP and RKHS}, we can represent this posterior process via a mean and a covariance function or, equivalently, through their associated RKHS objects.
We denote these RKHS objects as $\mu$ and $\Sigma$, which uniquely determine the GP posterior $ \GG\PP_\HH( \mu, \Sigma)$. In the following, without loss of generality, we shall assume that the prior GP is zero-mean and the RKHS is selected such that $f \sim \GG\PP_\HH( 0, I)$ \textit{a priori}.
The variational inference problem in GP models leads to the optimization problem
\begin{align} \label{eq:VI problem}
\min_{q = \GG\PP_\HH( \mu, \Sigma) } \LL (q)\,,
\qquad
\LL(q) = - \sum\nolimits_{n=1}^{N} \Ebb_{q(f(x_n))}[ \log p(y_n| f(x_n))] + \KL{q(f)}{p(f)}\,,
\end{align}
where $p(f) = \GG\PP_\HH( 0, I)$ and $\KL{q(f)}{p(f)}= \int \log \frac{q(f)}{p(f)} \mathrm{d} q(f)$ is the KL divergence between the approximate posterior GP $q(f)$ and the prior GP $p(f)$. It can be shown that $\LL(q) = \KL{q(f)}{p(f|\mathcal{D}}\def\EE{\mathcal{E}}\def\FF{\mathcal{F})}$ up to an additive constant~\citep{cheng2017variational}.
\subsection{Decoupled Gaussian Processes}
Directly optimizing the possibly infinite-dimensional RKHS objects $\mu$ and $\Sigma$ is computationally intractable except for the special case of a Gaussian likelihood and a small training set size $N$. Therefore, in practice, we need to impose a certain sparse structure on $\mu$ and $\Sigma$.
Inspired by the functional form of the exact solution in~\eqref{eq:GPR in RKHS}, \citet{cheng2017variational} propose to model the approximate posterior GP in the \emph{decoupled subspace parametrization} (which we will refer to as \emph{decoupled basis} for short) with
\begin{align} \label{eq:general subspace parametrization}
\mu = \Psi_\alpha \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}, \qquad
\Sigma = I + \Psi_\beta \mathbf{A}}\def\Bb{\mathbf{B}}\def\Cb{\mathbf{C} \Psi_\beta^\top
\end{align}
where $\alpha$ and $\beta$ are the sets of \emph{inducing points} to specify the bases $\Psi_\alpha$ and $\Psi_\beta$ in the RKHS, and $\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c} \in \Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb^{|\alpha|}$ and $\mathbf{A}}\def\Bb{\mathbf{B}}\def\Cb{\mathbf{C} \in \Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb^{|\beta| \times |\beta|}$ are the coefficients such that $\Sigma \succeq 0$. With only finite perturbations from the prior, the construction in~\eqref{eq:general subspace parametrization} ensures the KL divergence $\KL{q(f)}{p(f)}$ is finite~\citep{matthews2016sparse,cheng2017variational} (see Appendix~\ref{app:variational inference}).
Importantly, this parameterization decouples the variational parameters $(\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}, \alpha)$ for the mean $\mu$ and the variational parameters $(\mathbf{A}}\def\Bb{\mathbf{B}}\def\Cb{\mathbf{C}, \beta)$ for the covariance $\Sigma$. As a result, the computation complexities related to the two parts become independent, and a \emph{large} set of parameters can adopted for the mean to model complicated functions, as discussed below.
\paragraph{Coupled Basis}
The form in~\eqref{eq:general subspace parametrization} covers the sparse variational posterior~\citep{titsias2009variational}.
Let $Z=\{z_n \in\XX \}_{n=1}^{M}$ be some fictitious inputs and let
$\fb_{Z} = (f(z_n))_{n=1}^M$ be the vector of function values.
Based on the primal viewpoint of GPs, \citet{titsias2009variational} constructs the variational posterior as the posterior GP conditioned on $Z$ with marginal $q(\fb_{Z})=\mathcal N(\fb_{Z} | \mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}, \mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U})$,
where $\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o} \in \Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb^M$ and $\mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U} \succeq 0 \in \Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb^{M \times M}$.
The elements in $Z$ along with $\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}$ and $\mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U}$ are the variational parameters to optimize.
The mean and covariance functions of this process $\GG\PP(m,s)$ are %
\begin{align}
m(x) = \kb_{x, Z} \Kb^{-1}_{Z} \mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}, \qquad
s(x,x') = k_{x, x'} + \kb_{x, Z} \mbf{K}_{Z}^{-1} ( \mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U} - \mbf{K}_{Z}) \mbf{K}_{Z}^{-1} \kb_{Z, x} \,,
\end{align}
which is reminiscent of the exact result in~\eqref{eq:GPR}.
Equivalently, it has the dual representation
\begin{align} \label{eq:coupled RKHS}
\mu = \Psi_{Z} \Kb^{-1}_{Z}\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o},
\qquad
\Sigma = I + \Psi_{Z} \mbf{K}_{Z}^{-1} ( \mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U} - \mbf{K}_{Z}) \mbf{K}_{Z}^{-1} \Psi_{Z}^\top ,
\end{align}
which conforms with the form in~\eqref{eq:general subspace parametrization}. The computational complexity of using the coupled basis reduces from $O(N^3)$ to $O(M^3 + M^2 N )$. Therefore, when $M \ll N$ is selected, the GP can be applied to learning from large datasets~\citep{titsias2009variational}.
\paragraph{Inversely Parametrized Decoupled Basis}%
Directly parameterizing the dual representation in~\eqref{eq:general subspace parametrization} admits more flexibility than the primal function-valued perspective.
To ensure that the covariance of the posterior strictly decreases compared with the prior, \citet{cheng2017variational} propose a decoupled basis with an inversely parametrized covariance operator
\begin{align} \label{eq:decoupled RKHS (cheng 2017)}
\mu = \Psi_\alpha \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}, \qquad
\Sigma = (I + \Psi_\beta \Bb \Psi_\beta^\top)^{-1},
\end{align}
where $\Bb \succeq 0 \in \Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb^{|\beta| \times |\beta|}$ and is further parametrized by its Cholesky factors in implementation. It can be shown that the choice in~\eqref{eq:decoupled RKHS (cheng 2017)} is equivalent to setting $\mathbf{A}}\def\Bb{\mathbf{B}}\def\Cb{\mathbf{C} = - \Kb_\beta^{-1} + (\Kb_\beta \Bb \Kb_\beta + \Kb_\beta )^{-1}$
in~\eqref{eq:general subspace parametrization}.
In this parameterization, because the bases for the mean and the covariance are decoupled, the computational complexity of solving~\eqref{eq:VI problem} with the decoupled basis in~\eqref{eq:decoupled RKHS (cheng 2017)}
becomes $O(|\alpha| + |\beta|^3)$, as opposed to $O(M^3)$ of~\eqref{eq:coupled RKHS}.
Therefore, while it is usually assumed that $|\beta| $ is in the order of $M$, with a decoupled basis, we can freely choose $|\alpha| \gg |\beta|$ for modeling complex mean functions accurately.
\section{Orthogonally Decoupled Variational Gaussian Processes }
While the particular decoupled basis in~\eqref{eq:decoupled RKHS (cheng 2017)} is more expressive, its optimization problem is ill-conditioned and non-convex, and empirically slow convergence has been observed~\citep{havasi2018deep}.
To improve the speed of learning decoupled models, %
we consider the use of natural gradient descent~\citep{amari1998natural}. In particular, we are interested in the update rule for \emph{natural parameters}, which has empirically demonstrated impressive convergence performance over other choices of parametrizations~\citep{salimbeni2018natural}
However, it is unclear what the natural parameters~\eqref{eq:general subspace parametrization} for the general decoupled basis in~\eqref{eq:general subspace parametrization} are and whether finite-dimensional natural parameters even exist for such a model. In this paper, we show that when a decoupled basis is appropriately structured, then natural parameters do exist. Moreover, they admit a very efficient (approximate) natural gradient update rule as detailed in Section~\ref{sec:NGD}.
As a result, large-scale decoupled models can be quickly learned, joining the fast convergence property from the coupled approach~\citep{hensman2013gaussian} and the flexibility of the decoupled approach~\citep{cheng2017variational}.
\subsection{Alternative Decoupled Bases}
To motivate the proposed approach, let us first introduce some alternative decoupled bases for improving optimization properties~\eqref{eq:decoupled RKHS (cheng 2017)} and discuss their limitations. The inversely parameterized decoupled basis~\eqref{eq:decoupled RKHS (cheng 2017)} is likely to have different optimization properties from the standard coupled basis~\eqref{eq:coupled RKHS}, due to the inversion in its covariance parameterization. To avoid these potential difficulties, we reparameterize the covariance of~\eqref{eq:decoupled RKHS (cheng 2017)} as the one in~\eqref{eq:coupled RKHS} and consider instead the basis
\begin{align} \label{eq:decoupled_aS}
\mu = \Psi_\alpha \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c},
\qquad
\Sigma = (I - \Psi_\beta \Kb^{-1}_\beta \Psi_\beta^\top) +
\Psi_\beta \Kb^{-1}_{\beta} \mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U}\Kb^{-1}_{\beta}\Psi_\beta ^\top\,.
\end{align}
The basis \eqref{eq:decoupled_aS} can be viewed as a decoupled version of \eqref{eq:coupled RKHS}: it can be readily identified that setting $\alpha=\beta=Z$ and $\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}=\Kb_Z^{-1}\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}$ recovers \eqref{eq:coupled RKHS}. Note that we do not want to define a basis in terms of $\Kb_\alpha^{-1}$ as that incurs the cubic complexity that we intend to avoid. This basis gives a posterior process with
\begin{align} \label{eq:decoupled_1_posterior}
m(x) = \kb_{x,\alpha} \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}, \qquad
s(x, x') = k_{x,x'} - \kb_{x,\beta} \Kb^{-1}_\beta (\mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U} - \Kb_\beta)\Kb^{-1}_\beta\kb_{\beta, x'}.
\end{align}
The alternate choice~\eqref{eq:decoupled_aS} addresses the difficulty in optimizing the covariance operator, but it still suffers from one serious drawback: while using more inducing points, \eqref{eq:decoupled_aS} is not necessarily more expressive than the standard basis~\eqref{eq:coupled RKHS}, for example, when $\alpha$ is selected badly. To eliminate the worst-case setup, we can explicitly consider $\beta$ to be part of $\alpha$ and use
\begin{align}\label{eq:decoupled_gamma_beta_S}
\mu = \Psi_\gamma \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_{\gamma} + \Psi_\beta \Kb_\beta^{-1}\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}_{\beta},
\qquad
\Sigma = (I - \Psi_\beta \Kb^{-1}_\beta \Psi_\beta^\top) +
\Psi_\beta \Kb^{-1}_{\beta} \mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U}\Kb^{-1}_{\beta}\Psi_\beta ^\top.
\end{align}
where $\gamma = \alpha \setminus \beta$.
This is exactly the \emph{hybrid basis} suggested in the appendix of \citet{cheng2017variational}, which is strictly more expressive than~\eqref{eq:coupled RKHS} and yet has the complexity as~\eqref{eq:decoupled RKHS (cheng 2017)}. Also the explicit inclusion of $\beta$ inside $\alpha$ is pivotal to defining proper finite-dimensional natural parameters, which we will later discuss. This basis gives a posterior process with the same covariance as \eqref{eq:decoupled_1_posterior}, and mean $m(x) = \kb_{x,\gamma} \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\gamma + \kb_{x,\beta}\Kb_{\beta}^{-1}\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}_\beta $.
\subsection{Orthogonally Decoupled Representation} \label{sec:orthogonal}
But is~\eqref{eq:decoupled_gamma_beta_S} the best possible decoupled basis? Upon closer inspection, we find that there is redundancy in the parameterization of this basis: as $\Psi_\gamma$ is not orthogonal to $\Psi_\beta$ in general, optimizing $\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_{\gamma}$ and $\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}_{\beta}$ jointly would create coupling and make the optimization landscape more ill-conditioned.
To address this issue, under the partition that $\alpha = \{ \beta, \gamma\}$, we propose a new decoupled basis as
\begin{align} \label{eq:orthogonal RKHS}
\mu = (I - \Psi_\beta \Kb^{-1}_\beta \Psi_\beta^\top)\Psi_\gamma \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_{\gamma} + \Psi_\beta \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_{\beta},
\qquad
\Sigma = (I - \Psi_\beta \Kb^{-1}_\beta \Psi_\beta^\top) +
\Psi_\beta \Kb^{-1}_{\beta} \mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U}\Kb^{-1}_{\beta}\Psi_\beta ^\top,
\end{align}
where $\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\gamma \in \Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb^{|\gamma|}$, $\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_{\beta} \in \Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb^{|\beta|}$ and $\mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U} =\Lb\Lb^{\top}$ is parametrized by its Cholesky factor.
We call $(\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\gamma, \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\beta, \mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U})$ the \emph{model parameters}
and
refer to \eqref{eq:orthogonal RKHS} as the \emph{orthogonally} decoupled basis, because $(I - \Psi_\beta \Kb^{-1}_\beta \Psi_\beta^\top)$ is orthogonal to $\Psi_\beta$ (i.e. $(I - \Psi_\beta \Kb^{-1}_\beta \Psi_\beta^\top)^\top\Psi_\beta=0$).
By substituting $Z = \beta$ and $\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\beta = \Kb_Z^{-1} \mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}$, we can compare~\eqref{eq:orthogonal RKHS} to~\eqref{eq:coupled RKHS}: \eqref{eq:orthogonal RKHS} has an additional part parameterized by $\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\gamma$ to model the mean function residues that \emph{cannot} be captured by using the inducing points $\beta$ alone. In prediction, our basis has time complexity in $O(|\gamma| + |\beta|^3)$ because $\Kb_{\beta}^{-1} \Kb_{\beta, \gamma} \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_{\gamma}$ can be precomputed. The orthogonally decoupled basis results in a posterior process with
\begin{align*}%
m(x) = (\kb_{x,\gamma} - \kb_{x, \beta} \Kb_{\beta, \gamma} ) \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_{\gamma} + \kb_{x, \beta} \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_{\beta}, \quad
s(x, x') = k_{x,x'} - \kb_{x,\beta} \Kb^{-1}_{\beta} (\mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U}-\Kb^{-1}_{\beta})\Kb^{-1}_{\beta} \kb_{\beta, x'}.
\end{align*}
This decoupled basis can also be derived from the perspective of \citet{titsias2009variational} by conditioning the prior on a finite set of inducing points\footnote{We thank an anonymous reviewer for highlighting this connection.}. Details of this construction are in Appendix \ref{app:inducing_point_construction}.
Compared with the original decoupled basis in~\eqref{eq:decoupled RKHS (cheng 2017)}, our choice in~\eqref{eq:orthogonal RKHS} has attractive properties:
\begin{enumerate}%
\item The explicit inclusion of $\beta$ as a subset of $\alpha$ leads to the existence of natural parameters.
\item If the likelihood is strictly log-concave (e.g. Gaussian and Bernoulli likelihoods), then the variational inference problem in~\eqref{eq:VI problem} is strictly convex in $(\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\gamma, \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\beta, \Lb)$ (see Appendix~\ref{app:convexity}).
\end{enumerate}
Our setup in~\eqref{eq:orthogonal RKHS} introduces a projection operator before $\Psi_{\gamma}\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\gamma$ in the basis~\eqref{eq:decoupled_gamma_beta_S} and therefore it can be viewed as the \emph{unique} hybrid parametrization, which confines the function modeled by $\gamma$ to be orthogonal to the span the $\beta$ basis. Consequently, there is no correlation between optimizing $\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\gamma$ and $\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\beta$, making the problem more well-conditioned.
\subsection{Natural Parameters and Expectation Parameters}
\label{sec:parameters}
To identify the natural parameter of GPs structured as~\eqref{eq:orthogonal RKHS},
we revisit the definition of natural parameters in exponential families. A distribution $p(x)$ belongs to an exponential family if we can write $p(x) = h(x) \exp( t(x)^\top\eta - A(\eta))$, where $t(x)$ is the sufficient statistics, $\eta$ is the natural parameter, $A$ is the log-partition function, and $h(x)$ is the carrier measure.
Based on this definition, we can see that the choice of natural parameters is not unique. Suppose $\eta = H \tilde{\eta} + b$ for some constant matrix $H$ and vector $b$. Then $\tilde{\eta}$ is also an admissible natural parameter, because we can write $p(x) = \tilde{h}(x) \exp( \tilde{t}(x)^\top \tilde{\eta} - \tilde{A}(\tilde\eta))$,
where $\tilde{t}(x) = H^\top t(x)$, $\tilde{h}(x) = h(x) \exp(t(x)^\top b)$, and $\tilde{A}(\tilde{\eta}) = A(H\tilde{\eta} + b)$.
In other words, the natural parameter is only unique up to affine transformations.
If the natural parameter is transformed, the corresponding expectation parameter $\theta = \Ebb_{p}[t(x)]$ also transforms accordingly to $\tilde{\theta} = H^\top \theta$. It can be shown that the Legendre primal-dual relationship between $\eta$ and $\theta$ is also preserved: $\tilde{A}$ is also convex, and it satisfies $\tilde{\theta} = \nabla \tilde{A}(\tilde{\eta})$ and $\tilde{\eta} = \nabla \tilde{A}^*(\tilde{\theta})$, where $*$ denotes the Legendre dual function (see Appendix~\ref{app:nat param}).
We use this trick to identify the natural and expectation parameters of~\eqref{eq:orthogonal RKHS}.\footnote{While GPs do not admit a density function, the property of transforming natural parameters described above still applies. An alternate proof can be derived using KL divergence.
}
The relationships between natural, expectation, and model parameters are summarized in Figure~\ref{fig:relationship}.
\paragraph{Natural Parameters}
Recall that for Gaussian distributions the natural parameters are conventionally defined as $(\Sigma^{-1} \mu, \frac{1}{2}\Sigma^{-1})$. Therefore, to find the natural parameters of~\eqref{eq:orthogonal RKHS}, it suffices to show that $(\Sigma^{-1} \mu, \frac{1}{2}\Sigma^{-1})$ of~\eqref{eq:orthogonal RKHS} can be written as an affine transformation of some finite-dimensional parameters.
The matrix inversion lemma and
the orthogonality
of $(I - \Psi_\beta\Kb_\beta^{-1}\Psi_\beta^{\top})$ and $\Psi_\beta$
yield
\begin{align} \label{eq:natural parameters}
&\Sigma^{-1} \mu = ( I - \Psi_{\beta} \Kb_{\beta}^{-1} \Psi_{\beta}^\top ) \Psi_{\gamma} \mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\gamma
+ \Psi_{\beta} \mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\beta \,,\qquad \tfrac{1}{2}\Sigma^{-1} = \tfrac{1}{2}( I - \Psi_\beta\Kb_\beta^{-1} \Psi_\beta^{\top}) + \Psi_\beta \bm{\Theta} \Psi_\beta^{\top}, \nonumber \\
&\textstyle
\text{where} \qquad
\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\gamma = \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\gamma,
\qquad
\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\beta = \mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U}^{-1} \Kb_\beta \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\beta,
\qquad
\bm{\Theta} = \frac{1}{2} \mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U}^{-1}.
\end{align}
Therefore, we call $(\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\gamma, \mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\beta, \bm\Theta)$ the natural parameters of~\eqref{eq:orthogonal RKHS}. This choice is unique in the sense that $\Sigma^{-1} \mu $ is orthogonally parametrized.\footnote{The hybrid parameterization~\eqref{eq:decoupled_gamma_beta_S} in~\citep[Appendix]{cheng2017variational}, which also considers $\beta$ explicitly in $\mu$, admits natural parameters as well. However, their relationship and the natural gradient update rule turn out to be more convoluted; we provide a thorough discussion in Appendix~\ref{app:nat param}.}
The explicit inclusion of $\beta$ as part of $\alpha$ is important; otherwise there will be a constraint on $\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\alpha$ and $\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\beta$ because $\mu$ can only be parametrized by the $\alpha$-basis (see Appendix~\ref{app:nat param}).
\paragraph{Expectation Parameters} Once the new natural parameters are selected, we can also derive the corresponding expectation parameters. Recall for the natural parameters $(\Sigma^{-1} \mu, \frac{1}{2}\Sigma^{-1})$, the associated expectation parameters are $(\mu, -(\Sigma + \mu\mu^\top) )$.
Using the relationship between transformed natural and expectation parameters, we find the expected parameters of~\eqref{eq:orthogonal RKHS} using the adjoint operators:
$
[
( I - \Psi_{\beta} \Kb_{\beta}^{-1} \Psi_{\beta}^\top ) \Psi_{\gamma},
\Psi_{\beta}
]^\top \mu
=
[
\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}_{\gamma \perp \beta},
\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}_{\beta}
]^\top$
and
$
- \Psi_\beta^\top (\Sigma + \mu\mu^\top) \Psi_\beta = \bm\Lambda
$,
where we have
\begin{align} \label{eq:expectation parameters}
\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}_{\gamma \perp \beta} = (\Kb_{\gamma} - \Kb_{\gamma, \beta} \Kb_\beta^{-1} \Kb_{\beta, \gamma} ) \mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\gamma,
\qquad
\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}_{ \beta} = \mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U} \mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\beta,
\qquad
\bm\Lambda = - \mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U} - \mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}_{ \beta}\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}_{ \beta}^\top.
\end{align}
Note the equations for $\beta$ in~\eqref{eq:natural parameters} and~\eqref{eq:expectation parameters} have exactly the same relationship between the natural and expectation parameters in the standard Gaussian case, i.e. $(\Sigma^{-1} \mu, \frac{1}{2}\Sigma^{-1}) \leftrightarrow (\mu, -(\Sigma + \mu\mu^\top) )$.
\begin{figure}[t]
%
\includegraphics[width=\textwidth]{fig1.pdf}
\caption{The relationship between the three parameterizations of the orthogonally decoupled basis. The box highlights the parameters in common with the standard coupled basis, which are decoupled from the additional $\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\gamma$ parameter. This is a unique property of our orthogonal basis}
\label{fig:relationship}
\end{figure}
\subsection{Natural Gradient Descent} \label{sec:NGD}
Natural gradient descent updates parameters according to the information geometry induced by the KL divergence~\citep{amari1998natural}. It is invariant to reparametrization and can normalize the problem to be well conditioned~\citep{martens2014new}.
Let $F(\eta) = \nabla^2 \KL{q}{p_\eta}|_{q = p_\eta}$ be the Fisher information matrix, where $p_\eta$ denotes a distribution with natural parameter $\eta$. Natural gradient descent for natural parameters performs the update
$\eta \leftarrow \eta - \tau F(\eta)^{-1}\nabla_{\eta} \LL,$
where $\tau > 0$ is the step size. Because directly computing the inverse $ F(\eta)^{-1}$ is computationally expensive, we use the duality between natural and expectation parameters in exponential families
and adopt the equivalent update
$\eta \leftarrow \eta - \tau \nabla_{\theta} \LL$~\cite{hoffman2013stochastic,salimbeni2018natural}.
\paragraph{Exact Update Rules}
For our basis in~\eqref{eq:orthogonal RKHS}, the natural gradient descent step can be written as
\begin{align} \label{eq:ngd orthogonal GP}
\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l} \leftarrow \mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l} - \tau \nabla_{\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}} \LL,
\qquad
\bm\Theta \leftarrow \bm\Theta - \tau \nabla_{\bm\Lambda} \LL,
\end{align}
where we recall $\LL$ is the negative variational lower bound in~\eqref{eq:VI problem}, $\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l} = [\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\gamma, \mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\beta]$ in~\eqref{eq:natural parameters}, and $\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o} = [\mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}_{\gamma \perp \beta}, \mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}_\beta]$ in~\eqref{eq:expectation parameters}.
As $\LL$ is defined in terms of $(\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_{\gamma}, \mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_{\beta}, \mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U})$, to compute these derivatives we use chain rule (provided by the relationship in Figure~\ref{fig:relationship}) and obtain
\begin{subequations} \label{eq:ngd orthogonal GP (details)}
\begin{align}
\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\gamma &\leftarrow \mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\gamma - \tau (\Kb_{\gamma} - \Kb_{\gamma, \beta} \Kb_\beta^{-1} \Kb_{\gamma, \beta} )^{-1} \nabla_{\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\gamma} \LL, \label{eq:ngd for j_gamma}
\\
\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\beta &\leftarrow \mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\beta - \tau ( \Kb_{\beta}^{-1} \nabla_{\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\beta} \LL - 2 \nabla_{\mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U}} \LL \mathbf{m}}\def\nb{\mathbf{n}}\def\ob{\mathbf{o}_{ \beta} ), \label{eq:ngd for j_beta}
\\
\bm\Theta &\leftarrow \bm\Theta + \tau \nabla_{\mathbf{S}}\def\Tb{\mathbf{T}}\def\Ub{\mathbf{U}} \LL.
\label{eq:ngd for Theta}
\end{align}
\end{subequations}
Due to the orthogonal choice of natural parameter definition, the update for the $\gamma$ and the $\beta$ parts are independent. Furthermore, one can show that the update for $\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\beta$ and $\bm\Theta$ is exactly the same as the natural gradient descent rule for the standard coupled basis~\citep{hensman2013gaussian}, and that the update for the residue part $\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\gamma$ is equivalent to functional gradient descent~\citep{kivinen2004online} in the subspace orthogonal to the span of $\Psi_\beta$.%
\paragraph{Approximate Update Rule}
\label{sec:preconditioning}
We described the natural gradient descent update for the orthogonally decoupled GPs in~\eqref{eq:orthogonal RKHS}. However, in the regime where $\abs{\gamma} \gg \abs{\beta}$, computing~\eqref{eq:ngd for j_gamma} becomes infeasible. Here we propose an approximation of~\eqref{eq:ngd for j_gamma} by approximating $\Kb_\gamma$ with a diagonal-plus-low-rank structure. Because the inducing points $\beta$ are selected to globally approximate the function landscape, one sensible choice is to approximate $\Kb_\gamma$ with a Nystr\"om approximation based on $\beta$ and a diagonal correction term: $\Kb_\gamma \approx \mathbf{D}}\def\Eb{\mathbf{E}}\def\Fb{\mathbf{F}_{\gamma | \beta} + \Kb_{\gamma | \beta}$, where
$\mathbf{D}}\def\Eb{\mathbf{E}}\def\Fb{\mathbf{F}_{\gamma | \beta} = \mathrm{diag}(\Kb_\gamma - {\Kb}_{\gamma | \beta})$, ${\Kb}_{\gamma | \beta} = \Kb_{\gamma, \beta} \Kb_\beta^{-1} \Kb_{\beta, \gamma} $, and $\mathrm{diag}$ denotes extracting the diagonal part of a matrix. FITC~\citep{snelson2006sparse} uses a similar idea to approximate the prior distribution~\citep{quinonero2005unifying}, whereas here it is used to derive an approximate update rule without changing the problem.
This leads to a simple update rule
\begin{align} \label{eq:approximate ngd for j_beta}
\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\gamma \leftarrow \mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\gamma - \tau (\mathbf{D}}\def\Eb{\mathbf{E}}\def\Fb{\mathbf{F}_{\gamma | \beta} + \epsilon \Ib )^{-1} \nabla_{\mathbf{a}}\def\bb{\mathbf{b}}\def\cb{\mathbf{c}_\gamma} \LL,
\end{align}
where a jitter $\epsilon > 0$ is added to ensure stability.
This update rule uses a diagonal scaling $ (\mathbf{D}}\def\Eb{\mathbf{E}}\def\Fb{\mathbf{F}_{\gamma | \beta} + \epsilon \Ib )^{-1}$, which is independent of $\mathbf{j}}\def\kb{\mathbf{k}}\def\lb{\mathbf{l}_\beta$ and $\bm{\Theta}$. Therefore, while one could directly use the update~\eqref{eq:approximate ngd for j_beta}, in implementation, we propose to replace~\eqref{eq:approximate ngd for j_beta} with an adaptive coordinate-wise gradient descent algorithm (e.g. ADAM~\cite{Kingma2014Adam:Optimization}) to update the $\gamma$-part.
Due to the orthogonal structure, the overall computational complexity is in $O(|\gamma||\beta| + |\beta|^3)$. While this is more than the $O(|\gamma| + |\beta|^3)$ complexity of the original decoupled approach~\citep{cheng2017variational}; the experimental results suggest the additional computation is worth the large performance improvement.
\section{Results}
We empirically assess the performance of our algorithm in multiple regression and classification tasks. We show that
\begin{enumerate*}[label=\itshape\alph*\upshape)]
\item given fixed wall-clock time, the proposed orthogonally decoupled basis outperforms existing approaches;
\item given the same number of inducing points for the covariance, our method almost always improves on the coupled approach (which is in contrast to the previous decoupled bases);
\item using natural gradients can dramatically improve performance, especially in regression.
\end{enumerate*}
We compare updating our orthogonally decoupled basis with adaptive gradient descent using the Adam optimizer \citep{Kingma2014Adam:Optimization} (\textsc{Orth}\xspace), and using the approximate natural gradient descent rule described in Section \ref{sec:preconditioning} (\textsc{OrthNat}\xspace).
As baselines, we consider the original decoupled approach of \citet{cheng2017variational} (\textsc{Decoupled}\xspace) %
and the hybrid approach suggested in their Appendix (\textsc{Hybrid}\xspace). We compare also to the standard coupled basis with and without natural gradients (\textsc{CoupledNat}\xspace and \textsc{Coupled}\xspace, respectively). We make generic choices for hyperparameters, inducing point initializations, and data processing, which are detailed in \suppmat~\ref{sec:experimental details}.
Our code \footnote{\href{https://github.com/hughsalimbeni/orth\_decoupled\_var\_gps}{\color{blue}\texttt{https://github.com/hughsalimbeni/orth\_decoupled\_var\_gps}}} and datasets \footnote{\href{https://github.com/hughsalimbeni/bayesian_benchmarks}{\color{blue}\texttt{https://github.com/hughsalimbeni/bayesian\_benchmarks}}} are publicly available.
\begin{figure*}[t!]
\centering
\begin{subfigure}{0.33\textwidth}
\centering
\includegraphics[height=3.2cm]{TestRMSE_josh}
\caption{Test RMSE}
\label{fig:test rmse}
\end{subfigure}%
\begin{subfigure}{0.33\textwidth}
\centering
\includegraphics[height=3.2cm]{Testlogliklihood_josh}
\caption{Test log-likelihood}
\label{fig:test log-likelihood}
\end{subfigure}
\begin{subfigure}{0.33\textwidth}
\centering
\includegraphics[height=3.2cm]{ELBO_josh}
\caption{ELBO}
\label{fig:elbo}
\end{subfigure}
\caption{Training curves for our models in three different settings. %
Panel~(\subref{fig:elbo}) has $|\gamma|=3500, |\beta|=1500$ for the decoupled bases and $|\beta|=2000$ for the coupled bases. Panels~(\subref{fig:test rmse}) and~(\subref{fig:test log-likelihood}) have $|\gamma|=|\beta|=500$ and fixed hyperparameters and full batches to highlight the convergence properties of the approaches. Panels~(\subref{fig:test rmse}) and ~(\subref{fig:elbo}) use a Gaussian likelihood. Panel~(\subref{fig:test log-likelihood}) uses a Bernoulli likelihood.}
\end{figure*}
\paragraph{Illustration} Figures \ref{fig:test rmse} and \ref{fig:test log-likelihood} show a simplified setting to illustrate the difference between the methods. In this example, we fixed the inducing inputs and hyperparameters, and optimized the rest of the variational parameters. All the decoupled methods then have the same global optimum, so we can easily assess their convergence property. With a Gaussian likelihood we also computed the optimal solution analytically as an optimal baseline, %
although this requires inverting an $|\alpha|$-sized matrix, and therefore is not useful as a practical method. We include the expressions of the optimal solution in \suppmat~\ref{sec:analytic_expressions}. We set $|\gamma|=|\beta|=500$ for all bases and conducted experiments on \dataset{3droad} dataset ($N=434874$, $D=3$) for regression with a Gaussian likelihood and \dataset{ringnorm} data ($N=7400$, $D=21$) for classification with a Bernoulli likelihood.
Overall, the natural gradient methods are much faster to converge than their ordinary gradient counterparts. \textsc{Decoupled}\xspace fails to converge to the optimum after 20K iterations, even in the Gaussian case. We emphasize that, unlike our proposed approaches, \textsc{Decoupled}\xspace leads to a non-convex optimization problem.
\paragraph{Wall-clock comparison} To investigate large-scale performance, we used \dataset{3droad} with a large number of inducing points. We used a computer with a Tesla K40 GPU and found that, in wall-clock time, the orthogonally decoupled basis with $|\gamma|=3500,~|\beta|=1500$ was equivalent to a coupled model with $|\beta|=2000$ (about $0.7$ seconds per iteration) in our tensorflow~\citep{abadi2016tensorflow} implementation. %
Under this setting, we show the ELBO in Figure~\ref{fig:elbo} and the test log-likelihood and accuracy in Figure~\ref{fig:extra_plots} of the \suppmat~\ref{sec:further_results}. \textsc{OrthNat}\xspace performs the best, both in terms of log-likelihood and accuracy. The the highest test log-likelihood of \textsc{OrthNat}\xspace is $-3.25$, followed by \textsc{Orth}\xspace ($-3.26$). \textsc{Coupled}\xspace ($-3.37$) and \textsc{CoupledNat}\xspace ($-3.33$) both outperform \textsc{Hybrid}\xspace ($-3.39$) and \textsc{Decoupled}\xspace ($-3.66$).
\paragraph{Regression benchmarks} We applied our models on 12 regression datasets ranging from 15K to 2M points. To enable feasible computation on multiple datasets, we downscale (but keep the same ratio) the number of inducing points to $|\gamma|=700,|\beta|=300$ for the decoupled models, and $|\beta|=400$ for the coupled mode.
We compare also to the coupled model with $|\beta|=300$ to establish whether extra computation always improves the performance of the decoupled basis. The test mean absolute error (MAE) results are summarized in Table~\ref{table:regression_acc}, and the full results for both test log-likelihood and MAE are given in Appendix~\ref{sec:further_results}. \textsc{OrthNat} overall is the most competitive basis. And, by all measures, the orthogonal bases outperform their coupled counterparts with the same $\beta$, except for \textsc{Hybrid} and \textsc{Decoupled}.
\begin{table}[h]
\scalebox{0.72}{
\begin{tabular}{lccccccccc} %
\toprule
& \textsc{Coupled}$\dagger$ & \textsc{CoupledNat}$\dagger$ & \textsc{Coupled} & \textsc{CoupledNat} & \textsc{Orth} & \textsc{OrthNat} & \textsc{Hybrid} & \textsc{Decoupled} \\
\midrule
Mean & 0.298 & 0.295 & 0.291 & 0.290 & 0.284 & \textbf{0.282} & 0.298 & 0.361 \\
Median & 0.221 & 0.219 & 0.215 & 0.213 & 0.211 & \textbf{0.210} & 0.225 & 0.299 \\
Avg Rank & 6.083(0.19)& 5.00(0.33)& 3.750(0.26)& 2.417(0.31)& 2.500(0.47)& \textbf{1.833(0.35)}& 6.417(0.23)& 8(0.00)\\
\bottomrule
\end{tabular}
}
\caption{Regression summary for normalized test MAE on 12 regression datasets, with standard errors for the average ranks. The coupled bases had $|\beta|=400$ ($|\beta|=300$ for the $\dagger$ bases), and the decoupled all had $\gamma=700$, $\beta=300$. See Appendix~\ref{sec:further_results} for the full results.}
\label{table:regression_acc}
\end{table}
\paragraph{Classification benchmarks}
We compare our method with state-of-the-art fully connected neural networks with Selu activations~\citep{klambauer2017self}. We adopted the experimental setup from \citep{klambauer2017self}, using the largest 19 datasets (4898 to 130000 data points). For the binary datasets we used the Bernoulli likelihood, and for the multiclass datasets we used the robust-max likelihood~\citep{hernandez2011robust}. The same basis settings as for the regression benchmarks were used here.
\textsc{Orth}\xspace performs the best in terms of median, and \textsc{OrthNat}\xspace is best ranked. The neural network wins in terms of mean, because it substantially outperforms all the GP models in one particular dataset (\dataset{chess-krvk}), which skews the mean performance over the 19 datasets. We see that our orthogonal bases on average improve the coupled bases with equivalent wall-clock time, although for some datasets the coupled bases are superior. Unlike in the regression case, it is not always true that using natural gradients improve performance, although on average they do. This holds for both the coupled and decoupled bases.
\begin{table}[h]
\label{table:classification}
\centering
\scalebox{0.85}{
\begin{tabular}{lccccccc}%
\toprule
& Selu & \textsc{Coupled} & \textsc{CoupledNat} & \textsc{Orth} & \textsc{OrthNat} & \textsc{Hybrid} & \textsc{Decoupled} \\
\midrule Mean & \textbf{91.6} & 90.4 & 90.2 & 90.6 & 90.3 & 89.9 & 89.0 \\
Median & 93.1 & 94.8 & 93.6 & \textbf{95.6} & 93.6 & 93.4 & 92.0 \\
Average rank &4.16(0.67)& 3.89(0.42)& 3.53(0.45) &3.68(0.35)& \textbf{3.42(0.31)} &3.89(0.38)& 5.42(0.51) \\
\bottomrule
\end{tabular}
}
\caption{Classification test accuracy(\%) results for our models, showing also the results from \cite{klambauer2017self}, with standard errors for the average ranks. See Table~\ref{table:classification_acc_full} in the \suppmat for the complete results.}
\label{table:classification_acc}
\end{table}
Overall, the empirical results demonstrate that the orthogonally decoupled basis is superior to the coupled basis with the same wall-clock time, averaged over datasets. It is important to note that for the \emph{same} $\beta$, adding extra $\gamma$ increases performance for the orthogonally decoupled basis in almost all cases, but not for \textsc{Hybrid}\xspace of \textsc{Decoupled}\xspace. While this does add additional computation, the ratio between the extra computation for additional $\beta$ and that for additional $\gamma$ decreases to zero as $\beta$ increases. That is, eventually the cubic scaling in $|\beta|$ will dominate the linear scaling in $|\gamma|$.
\section{Conclusion}
We present a novel orthogonally decoupled basis for variational inference in GP models. Our basis is constructed by extending the standard coupled basis with an additional component to model the mean residues. Therefore, it extends the standard coupled basis~\citep{titsias2009variational,hensman2013gaussian} and achieves better performance.
We show how the natural parameters of our decoupled basis can be identified and propose an approximate natural gradient update rule, which %
significantly improves the optimization performance over original decoupled approach~\citep{cheng2017variational}. %
Empirically, our method demonstrates strong performance in multiple regression and classification tasks. %
\medskip
{\footnotesize
\bibliographystyle{abbrvnat}
|
\section{Introduction}
Person re-identification is the problem of identifying persons across images using different cameras or across time using a single camera. Automatic person re-identification has become essential in surveillance systems due to the rapid expansion of large-scale distributed multi-camera systems. However, many issues still prevent person re-id of achieving high accuracy as compared to other image recognition tasks; its performance is still far from optimal. These issues relate to the fact that person re-id usually needs to match the person images captured by surveillance cameras working in wide angle mode with a very low resolution and unstable lighting conditions. Despite the increasing attention given by researchers to solve the person re-id problem, it remains a challenging task in a practical environment. Some of these challenges are depicted in \ref{fig:challegences} and are: dramatic variations in visual appearance and ambient environment (a), human pose variation across time and space (b), background clutter and occlusions (c), and different individuals sharing similar appearances (d) among others.
Current approaches to solving person re-id generally follow a verification or identification framework or both \cite{Varior2016Gated, zheng2016discriminatively}. Such a framework takes as input a pair of images and outputs a similarity score or a classification result. Moreover, a Siamese convolution neural network architecture \cite{Li2014Pairing, Varior2016Gated,Zhao2017PartAligned, zheng2016discriminatively} which consists of two copies of the same network has recently emerged. The two networks are connected by a cost function used to evaluate the relationship between the pair. Other architecture models, driven by a triplet loss function, have resulted into part-based networks \cite{Cheng2016, Li2017DeepContext, Sun2017RPP} where the first convolution layers learn low-level features while fully connected layers concentrate on learning higher-level features. All the part contribute to the training process jointly.
In the last decade, with the advent of deep generative models, GAN-based models \cite{yu2017cross, Zhang2018Crossing, zheng2017unlabeled, zhong2018camera} have slightly increased the performance of person re-id task; however, results from these works show that there is still a room for improvement.
In this paper, inspired by biology and the recent success of attention mechanism on Recurrent Neural Network \cite{Chorowski2015,Hermann2015,Mnih2014HardAttention,Shan2017Speech, Zeyer2018Speech} and Convolution Neural Network \cite{Jaderberg2015STN, Li2018Harmonious, Liu2017End2End, Rahimpour2017Attention,Rodriguez2017AgeGender, Wang2017ResAttention, Wu2018CoAttention, Xu2016AskAttendAnswer, Yang2016StackedAttention, Zhang2016Picking, Zhao2017Diversified}, we propose Self Attention Grid (SAG) for person re-identification. Attention mechanism gives a network a capability to focus on specific parts of the visual input to compute the adequate response. In other words, It helps the network select the most pertinent piece of information, rather than using all available information. This is particularly important in person re-id where input images might contain information that is actually irrelevant for computing visual responses. In a classical person re-id model, the whole input image is used to predict the output regardless of the fact that not all pixels are equally important. As a result, we use a variant of self-attention mechanism to overcome such limitations. The contributions of this work are:
\begin{enumerate}
\item a simple feed-forward attention mechanism with multiple non-overlapping attention regions;
\item an attention module solely based on self-attention that can extract high discriminative feature representation from a high-resolution image and preserve low-level information from the initial representation;
\item a fully differentiable attention module that can be easily plugged into any existing network architecture without much effort.
\end{enumerate}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{images/challenges}
\caption{Person re-id challenges. Images with a same border color represent the same identities within a given group.}
\label{fig:challegences}
\end{figure}
\section{Related Works}
In this section, we present the works relevant to understand our approach. This include attention mechanism background together with their application to person re-identification task.
\subsection{Attention Mechanism}
Inspired by neuroscience and biology, attention mechanism allows deep neural network to focus on specific parts of the input vector. Not only it allows a network to learn the most discriminative feature, but also it effectively reduces the computational burden of processing the whole input vector.
Attention mechanism started in the field of Natural Language Processing (NPL) \cite{Bahdanau2015} and became central to many deep learning approaches, especially Recurrent Neural Networks (RNNs). It has been successfully applied to various interesting tasks such as text-based question answering \cite{Hermann2015}, image captioning \cite{Xu2015ShowAttendTell}, visual question answering \cite{Yang2016StackedAttention, Xu2016AskAttendAnswer}, speech recognition \cite{Chorowski2015, Shan2017Speech, Zeyer2018Speech} and fine-grained image classification \cite{Jaderberg2015STN, Wang2017ResAttention, Zhang2016Picking, Zhao2017Diversified}. Following its success in machine translation, many researchers started exploring its application into the computer vision fields by proposing various forms of attention mechanisms: hard-attention \cite{Mnih2014HardAttention}, soft-attention, global attention and local-attention to cite but a few.
In particular, self-attention mechanism, also referred to as \textit{intra-attention} in \cite{Cheng2016LSTMMR, Parikh2016Decomposable} attends to different parts of a single sequence by using the internal representations of the same sequence. \cite{Vaswani2017Attention} proposed scaled dot-product attention combining self-attention with a scaling factor and successfully achieved state-of-art in machine translation. \cite{Parmar2018Transformer} generalized an autoregressive model architecture based on self-attention for image generation and \cite{Xiaolong2018NonLocal} formalized self-attention for machine translation as a class of non-local filtering operation that can be applied to video sequences.
While many researchers \cite{Liu2017End2End, Li2018Harmonious, Rahimpour2017Attention, Wu2018CoAttention} have investigated the application of soft and hard attention mechanism to person re-identification; however, to the best of our knowledge, multi-depth regions solely based on self-attention mechanism has not yet been explored in this context. Our work, exploit the attention provided by the attention mechanism at different levels to capture information and efficiently learn to focus on specific part of the image by using only the internal representations of the same image at each time.
\subsection{Person Re-Identification}
Person re-id works can be roughly divided into two groups: distance metric learning and deep machine learning based approaches. The first group, also named discriminative distance metric focus on learning local and global feature similarities by leveraging inter-personal and intra-personal distances \cite{ Chen2016, Kostinger2012, Liao2015, Liao2015a, Xiong2014, Zhang2011, Zheng2015}. The second group is CNN-based with a goal to jointly learn the best feature representation and a distance metric. Some feature learning approaches \cite{Cheng2016, Li2017DeepContext, Sun2017RPP} decompose the images into part based. Other methods \cite{Li2014Pairing, Varior2016Gated, Zhao2017PartAligned, zheng2016discriminatively} used a siamese convolution neural network architecture for simultaneously learning a discriminative feature and a similarity metric. Given a pair of input images, they predict if it belongs to the same subject or not through a similarity score. To improve the similarity score, \cite{Paisitkriangkrai2015, Zhong2017reranking} proposed to optimize the evaluation metrics commonly used in person re-id. Recently \cite{yu2017cross, Zhang2018Crossing, zheng2017unlabeled, zhong2018camera} proposed to address the problem of lack of large datasets in person re-id by training a CNN based architecture and a GAN \cite{Goodfellow2014GAN} generated samples through a regularization method \cite{Szegedy2016}. It was particularly observed that generated images improve the re-id accuracy when combined with a training sample.
Similar to our work, \cite{Liu2017End2End} proposed an end-to-end Comparative Attention Network (CAN) to progressively compare the appearance of a pair of images and determine whether the pair belongs to the same person. During training, a triplet of raw images is fed into CAN for discriminative feature learning and local comparative visual attention generation. Their network architecture, made of two parts require much higher computation cost compared to our work. \cite{Li2018Harmonious} proposed a complex CNN model for jointly learning soft and hard attention. The two attention mechanisms with feature representation learning are simultaneously optimized. Finally, \cite{Rahimpour2017Attention} proposed gradient-based attention mechanism to solve the problem of pose and illumination found in re-id problem in a triplet architecture and \cite{Wu2018CoAttention} recommended Co-attention based comparator to learn a co-dependent feature of an image pair by attending to distinct regions relative to each pair.
In general, attention mechanism is used in re-identification task to discover the most discriminative information for further processing.
We depart from these works and propose our attention based-CNN model in next section.
\section{Our Approach}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{images/framework_module}
\caption{The baseline architecture model (left) coupled with our Self Attention Grid (SAG) Module (right). We show how SAG modules can be easily integrated into existing network without much efforts.}
\label{fig:framework_module}
\end{figure*}
In this section, we introduce our Self Attention Grid (SAG) mechanism coupled with ResNet50 baseline specifically designed for person re-identification tasks. We first describe the overall network architecture and then elaborate on the design of the SAG module.
\subsection{Overview}
As described in Figure \ref{fig:framework_module}, the overall network architecture of the proposed attention consists of two branches sharing the same weights. After each residual unit (\textit{layer$_i$}), we introduce a SAG module, we refer to this position as the attention depth ($D_i$). The network is fed with two copies of the same image $I_1 \in \mathbb{R}^{C\times H_1 \times W_1}$ and $I_2 \in \mathbb{R}^{C \times H_2 \times W_2}$ such that the spatial dimension of $I_1$ is twice the spatial dimension of $I_2$ i.e $(H_1,W_1)=(2H_2, 2W_2)$. To do this, we upsampled $I_1$ by a factor of two using bilinear interpolation as follows:
\begin{equation} \label{eq:bilinear}
\begin{split}
H_{1} & = \left \lfloor H_{2} \times \text{scale\_factor} \right\rfloor \\
W_{1} & = \left \lfloor W_{2} \times \text{scale\_factor} \right\rfloor
\end{split}
\end{equation}
where $\text{scale\_factor}=2$, $(H_1,W_1)$ the height and width of the high resolution image $I_1$, $(H_2, W_2)$ the spatial dimension of the original image $I_2$ and $C$ the channel.
At depth $D_i$, the first branch (upper-branch) processes the high resolution image $I_1$ and outputs a high dimensional features $f_{1}^{D_i}$, whereas the second branch (lower-branch) processes the low-resolution image $I_2$ and outputs a low dimensional features $f_{2}^{D_i}$ representing the filtered feature map, with our filtering attention grid putting focus on the interesting part of the original image. Given $D_i$, the network computes the attention response $f_{2}^{D_i}$ weighted by an importance score predicted by the SAG modules. In the next sections, for the sake of brevity we simply refers to $f_{1}^{D_i},f^{D_i}_{2}$ as $f_1, f_2$.
Given the high dimension image, we learn a discriminative feature using the upper CNN branch before element-wise multiplied by the output of the attention grid coming from the lower branch. The proposed method goes beyond the traditional CNN based attention models in re-identification and proposes an attention grid network that can learn multiple discriminative parts of person images as depicted in Figure \ref{fig:multiregions}.
In general, attention model output a summary vector $z$ of a class probability $y_i$; focusing on the information of an input vector $x_i$. $z$ is usually a weighted arithmetic mean of $y_i$; with weights chosen by relevance of each $y_i$, given $x_i$. In our case, the output of the attention module is a multi-grid region which relevance is produced by a softmax operation.
The overall network architecture is finally trained at once for identity classification task using supervised learning. We use the conventional cross-entropy loss function defined by:
\begin{equation} \label{eq:crossentropy}
\mathcal{L}(\theta) = - \sum_{i=1}^{N} \log p(\tilde{y}_i=y_i)
\end{equation}
where $N$ denotes the number of output classes, $p(\tilde{y}_i=y_i)$ the vector class probability produced by the neural network, $\tilde{y}_i$ the predicted label, $y_i$ the ground-truth label and $\theta$ the network parameters. The network is trained to minimize Equation \ref{eq:crossentropy}
\subsection{The Self Attention Grid} \label{selfattentiongrid}
Let $X=\{(x^{(1)}, y^{(1)}),(x^{(2)}, y^{(2)}), \cdots (x^{(n)}, y^{(n)})\}$, be $n$ training samples where $x^{(i)}$ represents an image $\in \mathbb{R}^{C\times H \times W}$ and $y^{(i)} \in [1, \ldots, n]$ its corresponding label; we aim at learning a feature representation model for person matching across multiple views. As a result, we propose Self Attention Grid as an attention mechanism for locating the most discriminative pixels and regions at different depths. We consider a multi-branch network architecture with weights shared between the two branches.
Our Self Attention Grid module consists of a $1\times1$ convolution, a batch normalization, an activation function (ReLU) and a softmax function. A convention neural network represents a function $h(x;\theta)$ defined by:
\begin{equation}
h(x) = \bm{w}^T_k\cdot x + b_k
\end{equation}
where $(\bm{w}_k, b_k)$ denote the network parameters that are automatically learned from data. Given $x_i$, the output $h(x_i)$ of a residual unit $L_i$ is fed into the $1\times1$ convolution layer to produce an attention heatmap $e_{ij}$. We use the softmax operation to normalize the full self attention so as to make the sum equal 1 using:
\begin{equation}
S_{ij} = softmax(e_{ij}) = \frac{\exp(e_{ij})}{\sum_{j=0}^{N} \exp(e_{ij})}
\end{equation}
The softmax layer computes the maximum of relevance of the variable $x^{(i)}$ according to $e{ij}$.
$S_{ij}$ is a two dimensional vector which we then convert into a grid attention $\bm{G}_{ij}$ $\in \mathbb{R}^{N\times 1\times h \times w}$ such that :
\begin{equation}
\sum_{i=1}^{N} \bm{G}_{ij} = 1
\end{equation}
The output of the each attention grid consists of one spatial channel dimension and defines a function $\phi : \mathbb{R}^{N \times C \times H \times W} \rightarrow \mathbb{R}^{N \times 1 \times h \times w}$. The output of the attention module $f_2$ is element-wise multiplied by each channel dimension of $f_1$ produced by the same layer sequence processing the high image resolution:
\begin{equation} \label{eq:sagproduct}
v = f_1 \odot f_2
\end{equation}
where $f_1 \in \mathbb{R}^{C \times H \times W}$ and $f_2 \in \mathbb{R}^{1 \times h \times w}$. To make this possible, we applied a sampled-based discretization process to downsample the high feature representation $f_1$, thus reducing its dimensionality and allowing the feature contained in the sub-regions of $f_1$ to correspond to the initial sub-region representations of $f_2$.
Finally, we performed $L_2$ normalization over the attended regions $v$ using:
\begin{equation} \label{eq:normalization}
f^I = \frac{v}{\max(\lVert v\rVert_2, \epsilon)}
\end{equation}
where $\lVert \cdot \rVert_2$ is the Euclidean norm and $\epsilon = 1e-12$ a small value to avoid division by zero. This Equation \ref{eq:normalization} helps keep the overall error small. We show in section \ref{discussion}, how this can be used to improve the classification accuracy.
In general, our approach can be considered as a kind of region-based-attention for it searches through the image multiples regions that match what the network is interested in for further processing.
\begin{table*}[t]
\centering
\caption{Impact depth of SAG modules on three datasets. We report results with \textbf{no} re-ranking \cite{Zhong2017reranking}. Basel.+ $D_i$ refers to our baseline architecture with SAG module at depth \textit{i}}
\label{tab:ablationresults}
\resizebox{\textwidth}{!}{%
\begin{tabular}{|l|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{SAG Components}& \multicolumn{4}{c|}{Market-1501 \cite{Zheng2015}} & \multicolumn{4}{|c|}{CUHK03 \cite{Li2014Pairing}} & \multicolumn{4}{c|}{DukeMTMC \cite{zheng2017unlabeled}}\\
\cline{2-13}
&R1&R5&R10&mAP&R1&R5&R10&mAP&R1&R5&R10&mAP\\
\hline
Baseline(ResNet50)&83.49&93.82&95.90&63.87&71.04&92.89&96.63&80.46&72.04&84.96&89.09&54.36\\
Basel.+ $D_1$ (h=40,w=16)&84.35&94.00&96.08&64.40&78.50&95.60&98.22&85.93&73.74&84.87&88.24&51.99\\
Basel.+ $D_2$ (h=20,w=8)&85.18&94.12&96.20&65.90&78.99&95.63&98.13&86.25&74.28&85.05&89.18&52.20\\
Basel.+ $D_3$ (h=10,w=4)&87.65&95.07&96.91&70.03&81.41&96.62&98.51&88.00&76.30&87.25&90.80&56.62\\
Basel.+ $D_4$ (h=5,w=2)&\textbf{90.17}&\textbf{96.38}&\textbf{97.48}&\textbf{73.87}&\textbf{82.46}&\textbf{96.44}&\textbf{98.42}&\textbf{88.64}&\textbf{79.94}&\textbf{89.68}&\textbf{92.15}&\textbf{60.88}\\
Basel.+ $D_{1,2}$&84.09&93.97&96.17&61.28&77.60&95.79&98.50&85.52&74.55&84.96&88.69&51.65\\
Basel.+ $D_{1,2,3}$&85.10&94.24&96.23&65.60&77.77&95.81&98.34&85.59&72.44&84.52&88.47&51.03\\
Basel.+ $D_{1,2,3,4}$&82.13&92.79&95.72&60.81&76.51&94.86&97.86&84.53&70.42&82.27&86.58&4876\\
\hline
\end{tabular}
}
\end{table*}
\section{Experiments} \label{experiments}
To validate the effectiveness of the proposed attention mechanism, we intensively conduct experiments and ablation study on three widely used datasets \footnote{The code of the experiments is available at \url{https://github.com/jpainam/self_attention_grid}}.
\subsection{Person Re-ID Datasets}
Three datasets were used to evaluate the proposed attention scheme. They include Market-1501, CUHK03 and DukeMTMC-ReID.
\textbf{Market-1501} \cite{Zheng2015} is one of the largest and most realistic dataset in person re-identification. It is collected using six overlapping cameras. The image bounding boxes were automatically detected using the Deformable Part Model (DPM) \cite{Felzenszwalb2010}. The dataset contains $32,668$ from $1,501$ identities divided into $12,936$ images for the training set and $ 19,732$ images for the testing. There are $751$ identities in the training set, $750$ identities in test set, $3,368$ query images and $2,793$ distractors. In this work, we use all the training set for training and all the test set for testing.
\textbf{CUHK03} \cite{Li2014Pairing} contains $13,164$ images from $1,467$ identities. The dataset is captured by six cameras, but each identity only appears in two disjoint camera views with an average of $4.8$ images in each view. The dataset is split into two subsets, one set contains manually cropped bounding boxes, and the other set is automatically detected using the Deformable Part Model \cite{Felzenszwalb2010}. In this work, we use the detected set.
\textbf{DukeMTMC-reID} \cite{zheng2017unlabeled} is a subset of a pedestrian tracking dataset DukeMTMC \cite{Ristani2016}. The original dataset is a collection of handcrafted bounding boxes and high resolution videos data set recorded by 8 synchronized cameras over $2,000$ identities. In this work, we use the subset defined in \cite{zheng2017unlabeled}. The subset follows the Market-1501 format and contains $36,411$ images from $1,404$ identities divided into $16,522$ images from $702$ identities for the training set and $17,661$ images from $702$ identities for the test set. There are $2,228$ query images and $17,661$ gallery images.
For a fair comparison, we follow the dataset split strategy of each dataset as described in their first released. Table \ref{tab:datasets} gives a summary of the split strategy.
\begin{table}
\centering
\caption{Dataset split details. The total number of images (\textit{QueryImgs, GalleryImgs, TrainImgs}), together with the total number of identities (\textit{TrainID, TestID}) are listed.}
\label{tab:datasets}
\begin{tabular}{|l|c|c|c|c|}
\hline
Dataset &Market&CUHK03&Duke\\
\hline\hline
Number of IDs&1501&1,467&1404\\
Number of Images&36,036&14,097&36,411\\
Cameras&6&2&8\\
\#Train IDs&751&1367&702\\
\#Train Images&12,936&13,113&16,522\\
\#Test IDs&750&100&702\\
\#Query Images&3,368&984&2,228\\
\#Gallery Images&19,732&984&17,661\\
\hline
\end{tabular}
\end{table}
\subsection{Evaluation Metrics}
We adopt the quantitative metrics cumulative matching curve (CMC) and Mean Average Precision (mAP) as they are commonly used in person re-id.
\textbf{Cumulative Matching Curve} is a precision curve that provides recognition for each rank. Rank-\textit{k} accuracy denotes the probabilities of one or more correctly matched images appearing in top-\textit{k}. Given an query set $Q = \{I_i\}^n_{i=0}$ from $n$-identity, we compute the $L_2$ distance between the query image and all gallery images and return a list of the top-$n$ images. If the returned list contains the query image at a position $k$-th, we consider this query as success at rank-\textit{k} and set it to $1$; if the top-\textit{k} ranked gallery samples do not contain the query identity, we set it to $0$. The final CMC curve is computed by averaging rank-\textit{k} over all the queries.
\textbf{Re-ranking} Recent works \cite{Bai2017,Wang2018, Zhong2017reranking} choose to perform an additional re-ranking to improve the re-identification accuracy. In this work, we use re-ranking with \textit{k}-reciprocal encoding \cite{Zhong2017reranking}, which combines the Euclidian distance and Jaccard distance.
Note that, all the CMC score for the CUHK03 and DukeMTMC datasets are computed with the single-shot setting. Only, experiments on Market-1501 dataset are under both the single-query and multi-query evaluation settings.
\subsection{Implementation details}
We use ResNet50 \cite{Kaiming2015}, pre-trained on imageNet as baseline and fine-tune the model according to the number of classes i.e. $751; 1,367$ and $702$ units for Market-1501, CUHK03 and DukeMTMC-ReID respectively. All the input images are resized to $160\times 64$ before random horizontal flipping. We scale the pixels in the range of $-1$ and $1$ and apply zero-center by mean pixel and random erasing \cite{Zhong2017}. Finally, we train the model for $200$ epochs using stochastic gradient descent (SGD) with a batch size of $32$, a momentum of $0.9$ and a weight decay of $5 \times 10 ^{-4}$. We use a base learning rate $lr$ of $0.01$ for upper layers and $0.1$ for fully connected layers. To further improve the training capability, we gradually decrease $lr$ by a factor of $0.1$ every $30$ epochs using an exponential policy: $lr=lr^{(0)} \times \gamma^{\frac{k}{step\_size}}$ where $lr^{(0)}$ is base learning rate, $\gamma=0.1$, $step size=30$ and $k$ the index of the current mini-batch iteration. We use a validation set to evaluate intermediate models and select the one with maximum performance for testing.
\subsection{Ablation study} \label{ablation}
In this section, we give detailed analysis and investigate the impact produced by SAG module when we vary the attention depth $D_i (i \in [1, \ldots 4])$. We systematically introduce SAG modules at different levels (i.e. after each residual unit) to capture information and observe accuracy variation. Figure \ref{fig:rank1bar} shows that SAG helps minimize the representation learning risk and improve the descriptive power of the baseline. We first show the effect of applying a single SAG module at different depths and then examine the impact of multiple (up to $4$) SAG modules. We conduct a series of experiments on three datasets and report results in Table \ref{tab:ablationresults}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{images/rank1bar}
\caption{The effect of the position of SAG within a CNN architecture. It can be observed that deeper position yields better results compared to shallow}
\label{fig:rank1bar}
\end{figure}
\textbf{Baseline model without attention module} In Table \ref{tab:ablationresults}, \textit{Baseline(ResNet50)} refers to the ResNet50 baseline trained for person re-id task. We changed the last fully connected layer to match the number of classes of the three dataset and trained the network using supervised learning. As shown in section \ref{comparison}, the average precision accuracy of our baseline model already outperform previous state-of-art methods. However, this performance is low compared to our attention based model.
\textbf{Baseline model with single SAG module} The $D_1-D_3$ architectures consist only of three or two convolution layers and are computationally efficient, whereas $D_4$ is deeper ($4$-conv layers), computationally more expensive, but has better performance. As it can be seen in Table \ref{tab:ablationresults}, deep $D_i$ results in better accuracy, reaching saturation at depth $D_4$. It results that, applying our attention module on deeper layers yields better results than on upper layers. These experiments show that SAG enforces the network to learn more discriminative representation. We therefore recommend the application of our module on last layers.
\textbf{Baseline model with an accumulation of SAG modules} We test the importance of multiple attention modules at different depths at the same time by increasingly stacking attention layers at $D_{1-4}$ and after each residual unit. Table \ref{tab:ablationresults} shows that stacking multiple SAG modules at different depth results in a accuracy drop of $1\%$ each time.
We achieved the best results with $D_4$ settings on all the datasets.
\subsection{Comparison with the state-of-arts} \label{comparison}
Tables \ref{tab:marketresults} \ref{tab:cuhk03results} \ref{tab:dukeresults} show comparison results with state-of-art methods. '-' means that no reported results is available and * means paper on ArXiv but not published. In the results, \textbf{SAG} represents our method with ResNet50 as baseline and \textbf{SAG+RR} represents our model with re-ranking.
\begin{table}
\centering
\caption{Comparison results on Market-1501.}
\label{tab:marketresults}
\begin{tabular}{|l|c|c|c|c|}
\hline
Query type& \multicolumn{2}{c|}{Single Query} & \multicolumn{2}{c|}{Multi Query}\\
\hline
Methods(\%)&R1&mAP&R1&mAP\\
\hline
CAN \cite{Liu2017End2End} &60.3&35.9&72.1&47.9\\
DNS \cite{Zhang2016}&61.02&35.68&71.56&46.03\\
Gated Reid \cite{Varior2016Gated}&65.88&39.55&76.04&48.45\\
MR B-CNN \cite{Ustinova2015}&66.36&85.01&90.17&41.17\\
Cross-GAN \cite{Zhang2018Crossing}*&72.15&-&94.3&48.24\\
SOMAnet ~\cite{Barros2018}&73.87&47.89&81.29&56.98\\
HydraPlus-Net \cite{liu2017hydraplus}&76.9&91.3&94.5&-\\
Verif.Identif \cite{zheng2016discriminatively}&79.51&59.87&85.47&70.33\\
MSCAN \cite{Li2017DeepContext} &80.31&57.53&86.79&66.70\\
SVDNet \cite{Sun2017SVDNet}&82.3&62.1&-&-\\
DeepTransfer \cite{Geng2016}*&83.7&65.5&89.6&73.80\\
LSRO \cite{zheng2017unlabeled}&83.97&66.07&88.42&76.10\\
JLML \cite{Li2017DeepJoint} &85.1&65.5&89.7&74.5\\
KFM-ReID \cite{Shen2018End2End} &90.1&75.3&-&-\\
\hline \hline
\textbf{SAG}&\textbf{90.17}&\textbf{73.87}&\textbf{92.76}&\textbf{80.15}\\
\textbf{SAG+RR}&\textbf{92.04}&\textbf{89.28}&\textbf{94.60}&\textbf{85.32}\\
\hline
\end{tabular}
\end{table}
\textbf{Evaluation on Market-1501} We compared our model with existing works in Market-1501 datasets and showed the superiority of our model. We achieved a $90.17\%$ rank-1 accuracy and $73.87\%$ mAP on single query setting. Our method outperforms JLML \cite{Li2017DeepJoint}(hard attention) by a factor of $5.07\%$ and slightly outperforms KFM-ReID \cite{Shen2018End2End} (Residual Self Attention) by a factor of $0.07\%$ on rank-1 accuracy achieving state-of-arts on attention-based CNN models for person re-id.
\begin{table}
\centering
\caption{Comparison result with state-of-arts on CUHK03.}
\label{tab:cuhk03results}
\begin{tabular}{|l|c|c|c|c|}
\hline
\multirow{2}{*}{Methods}&\multicolumn{4}{c|}{CUHK03}\\
\cline{2-5}
&R1&R5&R10&mAP \\
\hline \hline
SI-CI ~\cite{Wang2016}&52.20&84.30&94.8&-\\
DNS ~\cite{Zhang2016} &54.7&80.1&88.30&-\\
FisherNet ~\cite{Wu2016FisherNet}&63.23&89.95&92.73&44.11\\
MR B-CNN ~\cite{Ustinova2015}&63.67&89.15&94.66&-\\
Gated ReID ~\cite{Varior2016Gated}&68.1&88.1& 94.6&58.8\\
SOMAnet ~\cite{Barros2018}&72.40&92.10&95.80&-\\
SSM ~\cite{Bai2017}&72.7&92.4&96.1&-\\
SVDNet ~\cite{Sun2017SVDNet}&81.8&95.2&97.2&84.8\\
\hline \hline
\textbf{SAG}&\textbf{82.46}&\textbf{96.44}&\textbf{98.42}&\textbf{88.64}\\
\hline
\end{tabular}
\end{table}
\textbf{Evaluation on CUHK03} On this dataset, we achieved an $82.46\%$ rank-1 accuracy and $88.64\%$ mAP respectively. We improve the baseline by a factor of $11.42\%$ on rank-1 accuracy and $8.18\%$ on mAP respectively.
\begin{table}
\centering
\caption{Comparison results of the state-of-arts methods on DukeMTMCReID.}
\label{tab:dukeresults}
\begin{tabular}{|l|c|c|c|c|}
\hline
\multirow{2}{*}{Methods} & \multicolumn{4}{c|}{DukeMTMCReID}\\
\cline{2-5}
&R1&R5&R10&mAP \\
\hline \hline
PUL \cite{Fan2017PUL}&36.5&52.6&57.9&21.5\\
SPGAN \cite{Weijian2018Image2Image}&46.9&62.6&68.5&26.4\\
LSRO ~\cite{zheng2017unlabeled}&67.68&-&-&47.13\\
OIM ~\cite{xiaoli2017joint}&68.1&-&-&47.4\\
TriNet ~\cite{Hermans2017Triplet}*&72.44&-&-&53.50\\
SVDNet~\cite{Sun2017SVDNet}&76.7&86.4&89.9&56.8\\
DCC \cite{Wu2018CoAttention}&\textcolor{blue}{\textbf{80.3}}&\textcolor{blue}{\textbf{92.0}}&\textcolor{blue}{\textbf{97.1}}&\textcolor{blue}{\textbf{59.2}}\\
\hline \hline
\textbf{SAG}&\textbf{79.94}&\textbf{89.68}&\textbf{92.15}&\textbf{60.88}\\
\textbf{SAG+RR}&\textbf{85.28}&\textbf{91.07}&\textbf{93.76}&\textbf{81.05}\\
\hline
\end{tabular}
\end{table}
\textbf{Evaluation on DukeMTMC-ReID} On this dataset, we achieved competitive result with DCC \cite{Wu2018CoAttention} which Co-Attention model exceeds our Attention-Grid by a small factor of $0.36\%$ $(79.94-80.3)$.
\textbf{Visualization of the Self Attention Grid}
We visualize the attention grid at four different depths. Figure \ref{fig:multiregions} shows the visualization of our proposed attention mechanism. For example, with $(h=5,w=2)$, the results show how the SAG module can extract multi-parts and discriminative regions of the input images (e.g., backpack, legs, person's face, things in their hands, t-shirts). Also, it can be easily observed that our attention model successfully ignore the image's background.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{images/attended_regions}
\caption{Visualization of the attention grid predicted by $D_1, D_2, D_3, D_4$ architectures. The first column shows he original images. The second and last columns show the attention grid learned with $h=5,10,20,40$ and $w=2,4,8,16$. High attentions are shown in red and yellow. The attention grid also predicts low values for backgrounds. Best viewed in color}
\label{fig:multiregions}
\end{figure}
\subsection{Discussion} \label{discussion}
An end to end CNN-based model which incorporates a self-attention mechanism at different levels has been proposed. The attention module can be plugged into any existing system and is fully differentiable. The parameters of the SAG module can be learned at the same time during training. In other words, the attention layers and the network are trained simultaneously using back-propagation. The main contribution of this paper is a deep attention grid that can focus on multiple regions of an image with high resolution and at the time preserve the internal information in the low resolution.
Previous approaches \cite{Li2017DeepJoint, Li2018Harmonious, Shen2018End2End, Wu2018CoAttention} for applying attention to re-identification mainly tend on finding a single attention region on the image for further processing; while it is true that a single regions can describe a person, we argue that paying attention to more than one region simultaneously can improve the person re-id. In general, we describe our proposed attention mechanism as modular, architecture independent, fast and simple.
During training, the gradient of our network can be decomposed into two additives terms with the first term propagating information directly through the first branch, without any attention information and the second term propagating information back to the attention zone units
This approach can be considered as a soft attention as the gradient is directly computed during training and the update of network parameters achieved through the use of the conventional Gradient Descent algorithm.
We also observed classification accuracy improvement only when we add $L_2$ normalization to the $1^{st},2^{nd}$ and $3^{rd}$ residual units, but in $4^{th}$ layer, the classification accuracy drastically decrease. We therefore recommend using Equation \ref{eq:normalization} only on upper layers.
This is due to the non-sparseness propriety of the $L_2$ norm which positively affects the classification accuracy by leveraging high-level features but negatively disturb learned low-level features.
\section{Conclusion}
This paper proposed Self Attention Grid (SAG) based-CNN model for person re-identification tasks. We proposed SAG modules to find the most informative regions of input images at different depth levels and combine them with the output feature maps of the same layer. Our proposed attention only uses the internal representation of the input vector at each step to update the attention response. We further performed an ablation study to demonstrate that the model generalizes well when applied to deep layers. Design choices for implementing the attention model (attention mode, weight sharing, output normalization and attention depth) has been proposed and compared using three popular datasets Market-1501 \cite{Zheng2015}, CUHK03 \cite{Li2014Pairing} and DukeMTMC-ReID \cite{zheng2017unlabeled}.
We successfully improve the accuracy of the baseline CNN model and outperform a vast range of state-of-art methods. In general, our attention grid mechanism can be adopted for any re-id task such as vehicle re-id.
\section{Acknowledgement}
This work is supported by the Ministry of Science and Technology of Sichuan province (Grant No. 2017JY0073) and Fundamental Research Funds for the Central Universities
in China (Grant No. ZYGX2016J083).
{\small
\bibliographystyle{ieee}
|
\section{Background}
\label{sec:background}
Modern big-data applications have vast datasets that dwarf capacity-limited SRAM caches and reside in memory~\cite{bakhshalipour2018domino, ferdman:clearing}. Such applications frequently access the off-chip memory for data, putting significant pressure on the DRAM modules. Consequently, substantial performance is lost solely because of bandwidth limitations of the off-chip memory.
Recent research suggests using the die-stacked DRAM to break the bandwidth wall~\cite{Black_2006_DSM_1194816_1194860, Loh_3d}. As die-stacked DRAM cannot accommodate the whole datasets of big-data applications, prior proposals use it either as a large cache or as a part of main memory. In this section, we explore and discuss the design space of the die-stacked DRAM.
\subsection {Die-Stacked DRAM as a Memory-Side Cache}
\label{sec:background:cache}
One major challenge in architecting a giga-scale DRAM cache is microarchitecting the tags. Since the size of die-stacked DRAMs is in the range of several gigabytes, maintaining the tag information of such a large structure requires megabytes of storage. For example, the tag array of a 4~GB block-based\footnote{By a block-based design, we refer to a die-stacked DRAM cache whose caching granularity is equal to the block size of the processor's SRAM caches (e.g., 64~bytes). In contrast, by a page-based design, we refer to a DRAM cache whose caching granularity is more coarse-grained (e.g., 4~KB).} DRAM cache requires in excess of 64~MB of storage. Managing the DRAM as a page-based cache reduces the tag storage to nearly 14~MB, which is still significant. Obviously, affording such large structures in SRAM is impractical; hence, most practical approaches store the information within the die-stacked DRAM itself. However, storing the tag information in the die-stacked DRAM imposes latency and bandwidth overheads, because, before every data access, the corresponding tag information should be checked via accessing the die-stacked DRAM. Many pieces of prior work targeted this overhead and suggested techniques to improve latency and/or bandwidth efficiency of cache accesses.
Various design objectives (e.g., hit latency, bandwidth-efficiency and hit ratio) have been considered in architecting giga-scale DRAM cache designs. \methodname{Alloy Cache}~\cite{qureshi:alloy} is a block-based cache design that targets minimizing hit latency using a direct-mapped organization. \methodname{Alloy Cache} locates data and the corresponding tag, adjacently, and streams them out upon each access. This way, the latency of tag-checking is eliminated from the cache access if the access hits in the DRAM cache. Nevertheless, \methodname{Alloy Cache} suffers from two fundamental inefficiencies: (1) extra die-stacked DRAM bandwidth is consumed for accessing tag information, and (2) low hit ratio because of being a block-based cache and employing a direct-mapped organization.
\methodname{Unison Cache}~\cite{jevdjic:unison} improves the hit ratio of a DRAM cache by caching the data at the page granularity with a set-associative structure. As the size of the tag array, even for page-based designs, is significant, \methodname{Unison Cache} likewise co-locates the tag and data in the die-stacked DRAM. Caching the data at the page granularity improves the hit ratio but imposes substantial bandwidth overhead, because many blocks of a page are never used during the residency of the page in the DRAM cache~\cite{jang2016efficient, jevdjic:footprint}. Moreover, employing a set-associative organization necessitates searching all cache ways before touching the data, which adds latency and bandwidth overheads to the cache accesses.
Naively implementing a page-based design not only does not improve performance but may also harm it, mainly because of bandwidth inefficiency~\cite{jang2016efficient, jevdjic:footprint, jiang2010chop, yu2017banshee}. To reduce the traffic of page-based caches, \methodname{Unison Cache} takes advantage of a predictor for fetching the blocks within a page that will be used while the page is in the cache (i.e., referred to as the \emph{footprint} of the page~\cite{jevdjic:footprint}). Upon a cache miss, only those blocks within the requested page that are predicted to be used will be brought into the cache to reduce the traffic. While this technique (i.e., footprint caching) is useful at reducing the bandwidth usage, its efficiency is restricted by the accuracy of the predictor. For every misprediction, either the cache suffers from an extra miss due to not having the requested block in the cache, or non-useful blocks will be brought into the cache, wasting valuable bandwidth of off-chip and die-stacked DRAM. Moreover, the naive implementation of a set-associative structure requires two serialized accesses to the DRAM cache to get a cache block: (1) one access to read the tags and identify the location of the piece of data (i.e., way) and (2) another access to read the piece of data. To reduce the cache access latency, \methodname{Unison Cache} uses a \emph{way predictor} to access the requested block with two overlapped read operations for tag and data. Unfortunately, every time the predictor makes a mistake, the cache has to be accessed one more time, which increases the cache access latency and wastes valuable cache bandwidth. Moreover, even when the prediction is correct, extra bandwidth is consumed for accessing tag information, which puts further pressure on the die-stacked DRAM.
Another kind of DRAM caching approaches, referred to as \methodname{Tagless Dram Cache (TDC)}~\cite{jang2016efficient, Lee:2015:FAT:2749469.2750383}, align the granularity of caching with OS pages to track the tags of cache-resident pages within the Page Table and Translation Lookaside Buffers (TLBs). With such approaches, whenever a TLB miss occurs, the information of residency of the page in the DRAM cache (e.g., whether the page is cached or not and the location of the page in the DRAM cache if it is cached) are fetched together with the corresponding Page Table Entry (PTE). This way, the latency of tag-checking becomes virtually zero, but at the cost of significant complexities that are pushed into both software and hardware. In such approaches, the contents of the TLBs of \emph{all} the cores should (seem to) be coherent. Therefore, they use system-wide TLB shootdowns for updating the content of all TLBs whenever a piece of data is replaced in the DRAM cache. In addition to the complexities that are imposed because of run-time hardware-software co-operations, \emph{frequent} TLB shootdowns cause significant performance degradation due to costly software interventions and present a scalability challenge, as the latency of system-wide TLB shootdowns increases with increasing the core count~\cite{meswani:heterogeneous, romanescu2010unified}.
Another major drawback of these methods is the massive bandwidth overhead which is consumed because of fetching data at page granularity. Techniques like footprint caching~\cite{jevdjic:footprint} have limited applicability for such schemes, as these methods are restricted to use otherwise-unused bits in the PTEs to store footprint metadata, which is not adequately-spacious storage. For example, recent 64-bit Intel Xeon Phi processors~\cite{sodani2015knights} use 64-bit PTEs, in which, 18-bits are left unused. However, a 4~KB page includes 64 cache blocks, and hence, 64-bits are required to \emph{precisely} record the corresponding footprint information (i.e., one bit for each block, representing whether or not the block was touched during the last residency of the page in the cache). \methodname{Footprint-augmented TDC (F-TDC)}~\cite{jang2016efficient} attempts to solve this problem by storing $m$-line granularity footprint metadata, where a single bit is used to represent the existence of $m$ cache blocks in the footprint of each page, in order to fit within the capacity limitations of PTEs. Unfortunately, this strategy leads to the over-fetching of data and bandwidth pollution, as the precise knowledge of the residency of cache blocks is lost. On each misprediction, up to $m$ blocks are over-fetched, which leads to further bandwidth pollution. The problem is even more significant because modern big-data applications heavily use large (e.g., 4~MB) and huge (e.g., 1~GB) OS pages for increasing the TLB reach, as discussed in the recent work~\cite{papadopoulou2015prediction}. Employing $m$-line granularity footprint prediction for 1~GB pages requires saving 1 bit for every 450~K blocks. As a consequence, upon every misprediction, up to 450~K cache blocks are over-fetched, imposing an unbearable bandwidth overhead.
Lastly, since these methods use \emph{different} address spaces for caches (i.e., SRAM caches plus a DRAM cache) and off-chip memory, and because they do not flush the \emph{stale} entries of SRAM caches after each page remapping, they sacrifice consistency among physical addresses and are prone to result in wrong execution\footnote{We do not further discuss this problem, as it is entirely addressed in the recent work~\cite{yu2017banshee}, where it is called the \emph{address consistency problem}.}.
Recent work, namely the \methodname{Banshee Cache}~\cite{yu2017banshee}, addresses the first problem by \emph{coalescing} Page-Table updates. The \methodname{Banshee Cache} caches the details of recently-remapped pages (i.e., pages that were recently placed in or evicted from the DRAM cache) in an auxiliary SRAM structure, named the \componentname{Tag Buffer}. Whenever the occupancy of the \componentname{Tag Buffer} exceeds a certain threshold, a software interrupt is triggered to flush its entries into the Page Table and the TLBs. By provisioning enough capacity for the \componentname{Tag Buffer} (e.g., $\sim$5~KB), the frequency of Page Table updates decreases, which helps amortize the high cost of software interrupts. However, employing such a structure imposes SRAM storage overhead and may present scalability challenges as the core count increases. With more cores, the pressure on the \componentname{Tag Buffer} increases as the frequency of page re-mapping increases, resulting in more frequent flushes and costly software interventions. Therefore, the size of the \componentname{Tag Buffer} should commensurately be increased when increasing the number of cores, to preserve performance scalability. Furthermore, the complexity of run-time hardware-software co-operation is still a problem, as software is responsible for flushing the \componentname{Tag Buffer} into the Page Table and TLBs. To reduce the over-fetching problem, the \methodname{Banshee Cache} uses a bandwidth-aware replacement policy, in which, pages are replaced \emph{lazily} to lower the bandwidth pressure on die-stacked and off-chip DRAM. While the replacement policy is effective at reducing the movement of pages, it still consumes extra bandwidth to access page metadata (e.g., tag and replacement information) stored in the die-stacked DRAM. Finally, the \methodname{Banshee Cache} solves the address consistency problem of previously-proposed Page Table based approaches by using the \emph{same} address space for the die-stacked and off-chip DRAMs.
\subsection {Die-Stacked DRAM as a Part of Main Memory}
Using the die-stacked DRAM as a part of main memory eliminates the necessity of tags, and hence tag overheads, the major drawback of cache designs. The OS allocates some of the pages in the die-stacked DRAM physical address space and the rest in the off-chip DRAM. Such a scheme, however, cannot respond to dynamic changes in the data working sets of applications. Even if the OS applies intelligent algorithms for allocating hot pages in the high-bandwidth die-stacked memory, over time, there would be many pages in the die-stacked DRAM that are not currently being used, wasting its precious capacity. This happens because modern big-data applications have many \emph{dynamic data-dependent behaviors}, which cannot be detected/exploited statically before the execution of an application at page-allocation time~\cite{Huang_DIP, Sudan_MID}.
Several pieces of prior work~\cite{dong2010simple, meswani:heterogeneous, Sim:2014:THM:2742155.2742158} propose \emph{swapping} pages between the high- and low-bandwidth memory, periodically, to capture the dynamic data-dependent behavior of applications throughout their execution. In the state-of-the-art approach~\cite{meswani:heterogeneous}, regularly, the OS \emph{ranks} all pages based on their usage and moves hot pages into the die-stacked DRAM and cold pages out.
Unfortunately, such approaches suffer significantly from the massive cost of page swapping concerning both latency and bandwidth. In each interval and after the page swapping, the OS should update \emph{all} PTEs and shoot down \emph{all} TLBs for coherence, which takes a considerable amount of time. Moreover, exchanging many pages between die-stacked and off-chip DRAM consumes considerable bandwidth, as whole pages (and not their footprints, as in cache designs) should be transferred. For example, swapping two 4~KB pages, \componentname{P1} and \componentname{P2}, requires 16~KB of data transfer bandwidth: 4~KB for reading \componentname{P1}, 4~KB for writing \componentname{P1}, 4~KB for reading \componentname{P2}, and 4~KB for writing \componentname{P2}. If an application has many transient pages, the number of pages that should be swapped in each interval is also increased, exacerbating the bandwidth inefficiency of these approaches. Therefore, these methods perform the page swapping at \emph{very coarse granularity} (e.g., hundreds of milliseconds) to amortize the associated latency and bandwidth overheads. Consequently, the pages that are highly-utilized for a short duration of time (which we call \emph{transient pages}) cannot be captured by these approaches.
Another drawback of these methods is the limited ability to detect hot pages. As the OS has no precise information about the utilization of pages at \emph{run-time}, it cannot accurately rank them, and correctly identify hot and cold pages. Usually, there is a single bit in each PTE that indicates whether the page is used or not, and no knowledge of access frequency is available at run-time. Therefore, precisely ranking the pages at execution time requires non-trivial changes to both hardware and software, which makes such designs more complex.
\section{Conclusion}
Die-stacked DRAM has shown the potential to break the bandwidth wall. The research community has evaluated two extreme use cases of the die-stacked DRAM: (1) as a sizable memory-side cache, and (2) as a part of software-visible main memory. In this work, we showed that both designs are suboptimal to a scheme that uses the die-stacked DRAM partly as main memory and partly as a cache. By analyzing the access behavior of various big-data applications, we observed that there are many hot pages with a significant number of accesses. We proposed \methodname{MemCache}, an approach that uses a portion of the die-stacked DRAM as the main memory for hosting hot pages and the rest as a cache for capturing the dynamic behavior of applications.
\section{Evaluation}
\label{sec:eval}
\subsection{Performance}
\label{sec:eval:perf}
Figure~\ref{fig:performance} shows the performance of all designs, normalized to a system without die-stacked DRAM. Moreover, solid dots in the figure indicate Misses Per Kilo Instructions (MPKI) of the competing designs. On average, \methodname{MemCache-S/D} offer 107\%/114\% performance improvement, which is 21\%/28\% higher than \methodname{Banshee Cache}, the best-performing prior proposal.
\methodname{Unison Cache} improves the performance by 12\% on average and has the lowest performance enhancement among the evaluated designs. While it offers a high hit ratio (thanks to the page-based organization and ideal footprint predictor), it suffers from inefficiencies in utilizing the die-stacked DRAM bandwidth, as we show in Section~\ref{sec:eval:on-chip-traffic}. Due to frequent replacements of large pages and checking/updating the tags, a significant bandwidth overhead is imposed on the die-stacked DRAM, preventing it from offering considerable performance improvement.
As compared to \methodname{Unison Cache}, \methodname{Alloy Cache} has better performance improvement (38\% on average), mainly because of imposing less bandwidth pressure on the die-stacked DRAM. \methodname{Alloy Cache} caches data at block granularity, and hence, consumes less bandwidth upon replacements. Moreover, it uses a direct-mapped organization which frees it from the tag-updating bandwidth (e.g., LRU bits). However, \methodname{Alloy Cache} still consumes significant bandwidth for probing the tag and speculatively loading data, as we show in Section~\ref{sec:eval:on-chip-traffic}. For minimizing hit latency, \methodname{Alloy Cache} co-locates tag and data adjacently and streams them out in a single access. Whenever a request hits in the DRAM cache, the latency of accessing data is a single DRAM access, but extra bandwidth is consumed for loading the tag. In case of a cache miss, not only the tag but also the speculatively-loaded data impose bandwidth overhead. Another drawback of \methodname{Alloy Cache} is its relatively high MPKI. Caching data at block granularity and employing a direct-mapped organization synergistically decrease the cache hit ratio because of not exploiting the spatial locality and the increased conflict misses.
\methodname{HMA} has no tag overhead but suffers extensively from the low hit ratio of the die-stacked DRAM. As the periods of page swapping (OS intervals) are too long, the technique is unable to serve requests to transient datasets. Therefore, requests to data objects that are highly utilized for only a short period of time are served from the bandwidth-limited off-chip memory. We emphasize that reducing the period of OS intervals is not feasible because of the colossal associated latency and the bandwidth overhead of swapping numerous large pages.
\methodname{Banshee Cache} tracks tags of cache-resident pages through TLBs and hence eliminates the latency of tag-checking. Moreover, its bandwidth-aware replacement policy significantly reduces the frequency of replacements, resulting in less bandwidth pressure on the die-stacked and off-chip DRAM. However, its performance is far from that of \methodname{Infinite}. Lazily replacing pages performed by the bandwidth-aware replacement policy of \methodname{Banshee Cache} reduces the hit ratio, resulting in serving more requests from the bandwidth-limited off-chip memory. Moreover, in order to reach the metadata of pages (e.g., tag and counter) that are stored in the die-stacked DRAM, \methodname{Banshee Cache} consumes considerable bandwidth, which prevents it from reaching the peak performance. Finally, costly software interrupts that read and flush the Tag Buffer entries to the Page Table and TLBs take significant system cycles (each accounts for tens of kilo-cycles), resulting in performance degradation.
The performance improvement of \methodname{MemCache-S/D} ranges from 18\%/19\% to 294\%/310\% with an average of 107\%/114\%. Compared to \methodname{Banshee Cache}, the best-performing previous proposal, \methodname{MemCache-S/D} improves the performance by 21\%/28\% on average and up to 85\%/138\%. The performance improvement is more evident in bandwidth-hungry throughput-oriented applications (e.g.,~\appname{pr}). The performance improvement of \methodname{MemCache-S/D} over \methodname{Banshee Cache} comes from:
\begin{figure*}[t]
\centering
\makebox[\textwidth][c]{\includegraphics[width=\textwidth]{figures/stacked-bandwidth.pdf}}
\caption{The breakdown of die-stacked DRAM traffic.
\label{fig:die-stacked-bandwidth}}
\end{figure*}
\begin{figure*}[b]
\centering
\makebox[\textwidth][c]{\includegraphics[width=\textwidth]{figures/offchip-bandwidth.pdf}}
\caption{Off-chip DRAM traffic of evaluated designs.
\label{fig:off-chip-bandwidth}}
\end{figure*}
\begin{enumerate}
\item \methodname{MemCache-S/D} mitigates wrong replacement decisions made by \methodname{Banshee Cache} (i.e., not caching pages\footnote{\methodname{Banshee Cache} caches a page only if, intuitively, the page is identified hotter than other pages resident on the same set. This is problematic when more than four hot pages are mapped to the same set, preventing having all/many of them in the high-bandwidth DRAM. Increasing the associativity may mitigate this problem, but will exacerbate the bandwidth pressure on the die-stacked DRAM as more bytes need to be be transferred upon each metadata access. \methodname{MemCache} virtually eliminates this problem by allocating hot pages in the memory portion of the die-stacked DRAM.}). Therefore, it reduces both miss ratio and off-chip bandwidth of \methodname{Banshee Cache}. On average, \methodname{MemCache-S/D} reduces the MPKI of \methodname{Banshee Cache} by 8\%/15\% and up to 36\%/45\%. Moreover, it reduces the off-chip bandwidth by 22\%/27\%, on average, and up to 56\%/59\%.
\item \methodname{MemCache} reduces the bandwidth pressure of the die-stacked DRAM that is imposed because of accessing metadata in \methodname{Banshee Cache} for either replacement or tag-probing. It comes from the fact that, a significant fraction of requests is served from the memory portion of the die-stacked DRAM, for which, no metadata bandwidth is consumed. On average, \methodname{MemCache-S/D} consumes 13\%/15\% less stacked bandwidth (up to 31\%/30\%) as compared to \methodname{Banshee Cache}.
\item The number of costly software interrupts is reduced with \methodname{MemCache}, as compared to \methodname{Banshee Cache}. The main reason is that, with serving a significant fraction of requests from the memory portion of the die-stacked DRAM, there is a lighter load on the \componentname{Tag Buffer}, and hence, it is filled up more slowly. \methodname{MemCache-S/D} reduces the number of \componentname{Tag Buffer} flushes by 78\%/87\% on average and up to 94\%/99\%.
\end{enumerate}
\methodname{MemCache-D} offers higher performance as compared to \methodname{MemCache-S}. \methodname{MemCache-D} outperforms \methodname{MemCache-S} by 7\% on average and up to 84\%, mainly because of more intelligent partitioning of the die-stacked DRAM capacity. However, in few cases, the performance improvement of \methodname{MemCache-D} is slightly less (up to 2\%) than that of \methodname{MemCache-S}. It is because we partition the die-stacked DRAM capacity (i.e., Algorithm~\ref{alg:ahf}) towards memory, \emph{conservatively}. That is, we dedicate frames to the memory as much as the whole design (i.e., the composition of memory and cache) becomes susceptible to fall behind the full-cache design. However, as discussed in Section~\ref{motiv:cache-vs-mem}, since the memory has better latency/bandwidth characterizations as compared to the cache, the optimal point in capacity partitioning might be where the memory fraction is slightly higher than what Algorithm~\ref{alg:ahf} determines. One solution may be enhancing the offline partitioning algorithm by considering factors other than hit ratio, such as bandwidth and/or latency. However, doing so adds complications to the offline process since the compiler should run a \emph{timing simulation} rather than simply a trace-driven simulation. We conclude that the minor performance improvement does not justify this increased complexity.
Impractical \methodname{Infinite} has the highest average performance improvement. \methodname{Infinite} improves the performance by 149\% on average and up to 275\%. However, in several cases (e.g., \appname{cc}, \appname{pr}, \appname{sssp}, and \appname{mcf}), its performance is less than that of \methodname{MemCache-S/D} (in \appname{mcf} also less than \methodname{HMA}). This is because, with \methodname{Infinite}, applications benefit only from the bandwidth of the die-stacked DRAM and not the off-chip DRAM. Therefore, with \methodname{Infinite}, the total bandwidth available to applications is less than other approaches in which the bandwidth of both die-stacked and off-chip DRAM are available.
\begin{figure*}[b]
\centering
\makebox[\textwidth][c]{\includegraphics[width=\textwidth]{figures/sens-cache-size.pdf}}
\caption{The sensitivity of performance (bars) and MPKI (solid dots) to memory-cache capacity partitioning. Performance is normalized to a baseline without die-stacked DRAM. \textit{X/Y} represents a design that dedicates \textit{X}~GB to memory and \textit{Y}~GB to cache.
\label{fig:memcache-partition}}
\end{figure*}
\subsection{Die-Stacked DRAM Traffic}
\label{sec:eval:on-chip-traffic}
As throughput-oriented applications (like graph processing) are able to utilize the bandwidth of die-stacked DRAM, managing die-stacked DRAM in a bandwidth-efficient manner is crucial for reaching the peak performance.
Figure~\ref{fig:die-stacked-bandwidth} shows the breakdown of die-stacked DRAM traffic for the competing designs. We use \emph{Bytes Per Instruction}~\cite{yu2017banshee} as the metric for bandwidth intensity of applications. The breakdown consists of \componentname{Data}, \componentname{Metadata}, \componentname{Spec. Data}, and \componentname{Replacement}. \componentname{Data} refers to the traffic that is used for reading and updating data in the die-stacked DRAM. \componentname{Metadata} includes the bandwidth consumed for reading/updating tag and replacement metadata (e.g., LRU bits). \componentname{Spec. Data} corresponds to the traffic that is imposed to speculatively load data for requests that miss in the die-stacked DRAM. This type of traffic is specific to \methodname{Alloy Cache} and \methodname{Unison Cache}. Finally, \componentname{Replacement} shows the bandwidth consumed for replacing cache blocks/pages.
\methodname{HMA} has the lowest traffic as it touches the die-stacked DRAM only for data\footnote{Repeatedly, we do not model the bandwidth of page-swapping in \methodname{HMA}.}. \methodname{Alloy Cache} and \methodname{Unison Cache} impose an order of magnitude higher traffic to the die-stacked DRAM, mainly because of metadata accesses. The metadata traffic is higher in \methodname{Unison Cache} because it employs a set-associative structure and is required to update metadata information upon each access. \methodname{Banshee Cache} has lower bandwidth pressure because it has fewer metadata accesses and lazily replaces pages in the die-stacked DRAM. Meanwhile, it has higher bandwidth usage as compared to \methodname{MemCache}. By serving a significant fraction of requests from the memory portion of the die-stacked DRAM, \methodname{MemCache-S/D} reduces the die-stacked bandwidth usage of \methodname{Banshee Cache} by 13\%/15\%. The reduction mainly comes from the cut down on metadata and replacement traffic.
\subsection{Off-Chip DRAM Traffic}
Figure~\ref{fig:off-chip-bandwidth} shows the traffic of the off-chip DRAM for the evaluated designs. Off-chip DRAM traffic is more important than that of the die-stacked DRAM because more gap between the bandwidth of die-stacked and off-chip DRAM is expected with the future technologies~\cite{tran2016era}.
\methodname{HMA} has the highest traffic because of its lower hit ratio. As discussed previously, \methodname{HMA} is unable to capture the dynamic behavior of applications, due to its long OS intervals. Therefore, in this technique, virtually all requests to transient data objects are served from the off-chip DRAM, imposing a significant bandwidth overhead. \methodname{Unison Cache} and \methodname{Alloy Cache} impose traffic on the off-chip DRAM due to cache misses and dirty writebacks. \methodname{Banshee Cache} has a less off-chip bandwidth, as compared to \methodname{Unison Cache} and \methodname{Alloy Cache}, because it has fewer replacements. \methodname{MemCache} has the lowest off-chip bandwidth among the evaluated designs. The bandwidth reduction of \methodname{MemCache-S/D} over \methodname{Banshee Cache} is 22\%/27\% (up to 56\%/59\%) and comes from mitigating wrong replacement decisions made by \methodname{Banshee Cache}, resulting in higher hit ratio of the die-stacked DRAM.
\subsection{Sensitivity to Memory-Cache Capacity Partitioning}
\label{sec:eval:capacity_partition}
The optimal capacity partitioning between the cache and memory in the die-stacked DRAM depends on the applications. Figure~\ref{fig:memcache-partition} shows the performance and MPKI of \methodname{MemCache-S} as the fraction of memory varies. Applications that have many hot pages (e.g., \appname{pr}) favor dedicating smaller capacity to the cache and larger to the memory. On the other hand, for applications with significant transient pages (e.g., \appname{gems}), partitioning the capacity towards more cache is beneficial. Among the different partitionings, the design which devotes 3~GB to memory and 1~GB to cache offers the highest average performance; consequently, we choose this partitioning for the \methodname{MemCache-S}. We note that on average, cache design (i.e., \emph{0/4}) performs better than full-memory design (i.e., \emph{4/0}). This confirms that due to the mismatch between the dataset size and die-stacked DRAM capacity, full-memory is suboptimal to full-cache designs. However, in few cases and despite the fact that full-memory design has higher MPKI than the cache design, its performance is slightly better. This is due to the fact that a full-memory design completely eliminates the whole overhead associated with the evaluated cache scheme (i.e., \methodname{Banshee Cache}): no metadata overhead, no replacement, and no expensive software interrupt.
An interesting data point is that, for applications that largely benefit from the die-stacked DRAM technology (i.e., the performance gets more than doubled), it is beneficial to dedicate more capacity to memory and less to cache. This is the main reason why the design that uses three-fourths of the die-stacked DRAM as memory offers a significant performance improvement of the dynamically-partitioned design. Modern big-data throughput-oriented applications (e.g., most of the graph-processing applications in our suite) have a significant number of hot pages (cf.~Section~\ref{motiv:hot-pages}) which can be hosted in the memory portion of the die-stacked DRAM, enabling better usage of the die-stacked DRAM technology, and hence higher performance improvement.
\subsection{The Architecture of the Cache Portion}
\label{sec:eval:cache_design}
The results reported for \methodname{MemCache} are obtained for a design where the cache part is organized as a \methodname{Banshee Cache} as it offers the highest performance. In this section, we evaluate other architectures for the cache portion of the die-stacked DRAM. Figure~\ref{fig:memcache-cache-design} shows the performance of \methodname{MemCache} with various cache designs, as compared to the performance of the corresponding cache design. \methodname{MemCache-S} with \methodname{Alloy Cache}/\methodname{Unison Cache}/\methodname{Banshee Cache} improves performance by 13\%/9\%/21\% over the corresponding cache design. Furthermore, \methodname{MemCache-D} with \methodname{Alloy Cache}/\methodname{Unison Cache}/\methodname{Banshee Cache} enhances the performance of the corresponding cache design by 26\%/25\%/28\%. The performance improvement of \methodname{MemCache-S/D} with \methodname{Alloy Cache} as compared to the full-cache design is 13\%/26\% on average and up to 45\%/103\%. The performance improvement comes from: (1) 18\%/38\% lower MPKI, (2) 16\%/27\% less die-stacked DRAM bandwidth usage, and (3) 17\%/36\% lower traffic on the off-chip DRAM. \methodname{MemCache-S/D} with \methodname{Unison Cache} improves performance by 9\%/25\% on average and up to 66\%/174\% over the full-cache desgin. The performance enhancement comes from: (1) 8\%/49\% MPKI reduction, (2) 18\%/35\% less bandwidth pressure on the die-stacked DRAM, and (3) 5\%/33\% lower off-chip DRAM traffic.
The results indicate that \methodname{MemCache} is not limited to a specific cache design and can be used with other cache organizations, as well. Moreover, even with \methodname{MemCache}, the performance is significantly affected by the cache architecture. Since most of the bottlenecks on the road to achieve the peak performance are cache-induced (e.g., tag-checking bandwidth and limited associativity), and not memory-induced, the architecture of the cache portion has a decisive role in the overall performance.
\begin{figure}[t]
\centering
\includegraphics[width=.48\textwidth]{figures/sens-cache-design.pdf}
\caption{The sensitivity of performance to the cache organization of \methodname{MemCache}. Performance is normalized to a baseline without die-stacked DRAM.
\label{fig:memcache-cache-design}}
\end{figure}
\subsection{How Long are the Pages Hot?}
\label{eval:how_long}
\begin{figure}[t]
\centering
\includegraphics[width=.48\textwidth]{figures/mem-hit-ratio-time.pdf}
\caption{Fraction of requests that are served from the memory part of MemCache over time.
\label{fig:hit-rate-time}}
\end{figure}
We fill the memory part of \methodname{MemCache} with hot pages that are distinguished at compile time. Any static and profile-based approach has the risk that its results become stale over time. Here, we measure the \emph{fraction of the requests that are served from the memory part}, as time spent from the point where timing simulations have finished. Figure~\ref{fig:hit-rate-time} shows how this fraction changes over 100 seconds for 3~GB of the hottest pages (i.e., \methodname{MemCache-S}). As shown, except for one workload (i.e., \appname{tc}), the metric remains steady. The results indicate that:
\begin{enumerate}
\item Hot pages stem fundamentally from the data access patterns of applications and stay hot throughout the whole execution.
\item A static compile-time analysis with a representative input-set is able to easily detect hot pages in the context of a giga-scale die-stacked DRAM.
\end{enumerate}
The \appname{tc} application exhibits significant dynamic data-dependent behavior (cf.~Figure~\ref{fig:mem-vs-cache}) that results in a sharp drop in the reuse of pages that are classified as hot at the beginning of the execution. One solution for such workloads is to dynamically re-identify and swap pages between die-stacked and off-chip memories, just like prior work~\cite{dong2010simple, meswani:heterogeneous, Sim:2014:THM:2742155.2742158}. Note that, even for such applications, as a part of the die-stacked DRAM in \methodname{MemCache} is managed as a cache, it quickly responds to the dynamic changes of applications. Therefore, there is room for significantly increasing the period of OS interventions (e.g., once every 100 seconds), in order to amortize the high latency/bandwidth overhead. Evaluating such ideas is beyond the reach of architectural simulators, and hence, we leave them for future work.
\subsection{Sensitivity to Inputset}
\label{eval:inputset}
To evaluate the impact of inputset on the effectiveness of \methodname{MemCache}, we consider two other inputsets (i.e., \componentname{Web}~\cite{davis2011university} and \componentname{Urand}~\cite{erdds1959random}) for \componentname{GAPBS} applications\footnote{We do not perform a similar experiment for \componentname{SPEC} programs since other standard inputsets of \componentname{SPEC} applications are unable to fill the capacity of the die-stacked DRAM in a reasonable simulation time.} and compare the results with those of the so-far--evaluated \componentname{Twitter} inputset~\cite{kwak2010twitter}.
Table~\ref{table:inputset} summarizes the key characteristics of applications when they run these inputsets. Corroborating the characterization study performed by Beamer et al.~\cite{beamer2015gap}, \componentname{Web} graphs, despite the large size, exhibit substantial locality owing to their topology and high average degree~\cite{davis2011university}. Synthetically-generated \componentname{Urand} graphs, on the other hand, manifest the worst-case locality as every vertex in the graph has equal probability to be a neighbor of every other vertex~\cite{erdds1959random}. \componentname{Twitter} graphs, as they come from real-world data, have characteristics that lie in between \componentname{Web} and \componentname{Urand}.
\begin{table}[h]
\sffamily
\scriptsize
\begin{center}
\caption{LLSC MPKI and memory footprint of evaluated graph-processing applications when they use different inputsets.}
\label{table:inputset}
\resizebox{.48\textwidth}{!}{
\renewcommand{\arraystretch}{1}
\begin{tabular}{| c || c | c | c | c | c | c |}
\hline
\multirow{2}{*}{\bf Application} & \multicolumn{3}{c|}{\bf {LLSC MPKI}} & \multicolumn{3}{c|}{\bf {Memory Footprint (GB)}} \\ \cline{2-7}
& \componentname{Twitter} & \componentname{Web} & \componentname{Urand} & \componentname{Twitter} & \componentname{Web} & \componentname{Urand} \\
\hline
{bc} & 61.4 & 10.2 & 71.3 & 92.9 & 107.9 & 304.4 \\
\hline
{bfs} & 32.9 & 8.0 & 58.4 & 113.3 & 68.6 & 234.8 \\
\hline
{cc} & 85.6 & 8.2 & 98.7 & 9.6 & 9.2 & 25.0 \\
\hline
{pr} & 129.9 & 19.2 & 179.1 & 76.5 & 216.9 & 102.6 \\
\hline
{sssp} & 73.4 & 23.8 & 98.2 & 140.1 & 94.9 & 496.0 \\
\hline
{tc} & 12.5 & 2.9 & 16.2 & 53.4 & 21.5 & 266.2 \\
\hline
\end{tabular}
}
\end{center}
\end{table}
Figure~\ref{fig:inputset-sensitivity} shows the average performance improvement of die-stacked DRAM organizations when the applications use different inputsets. On average, with the \componentname{Twitter}/\componentname{Web}/\componentname{Urand} inputset, \methodname{MemCache-S} outperforms the best of cache and memory by 21\%/6\%/10\%. \methodname{MemCache-D} also outperforms \methodname{MemCache-S} by 7\%/9\%/6\%. While considerable, the performance improvement with \componentname{Web} and \componentname{Urand} inputsets is less than the improvement with the \componentname{Twitter} inputset. The improvement with \componentname{Web} is relatively low because most of the accesses hit in SRAM caches due to the high locality of graphs (cf.~Table~\ref{table:inputset}), reducing the bandwidth requirements of applications, and hence, downplaying the effect of bandwidth-improvement techniques. The lack of a considerable locality in \componentname{Urand} graphs, on the other face of the coin, defeat the caching and hot-page--detection policies used in the die-stacked DRAM organizations, giving rise to less performance improvement with these approaches. By and large, with all inputsets, \methodname{MemCache-S/D} consistently outperform both cache and memory designs, reinforcing our stance that using the die-stacked DRAM, partly as main memory and partly as cache, is the right design choice for modern throughput-oriented big-data applications.
\begin{figure}[t]
\centering
\includegraphics[width=.48\textwidth]{figures/inputset.pdf}
\caption{Sensitivity of performance improvement of various die-stacked DRAM organizations to inputset.
\label{fig:inputset-sensitivity}}
\end{figure}
\subsection{Sensitivity to Die-Stacked DRAM Size}
\label{eval:dram_size}
Figure~\ref{fig:dram_size} shows the performance improvement of various methods with different die-stacked DRAM sizes. The figure shows that regardless of the die-stacked DRAM capacity, the hybrid design (i.e., \methodname{MemCache}) outperforms both cache and memory, because of \emph{simultaneously} reaping the benefits of both cache and memory. More to the point, with increasing the size of the die-stacked DRAM, the fraction of capacity that should be devoted to the memory in order to gain higher performance also increases. With increasing die-stacked DRAM capacity, the number of identified/allocated hot pages (cf.~Section~\ref{sec:motiv:what_frac}) and the accuracy of offline page classification (cf.~Section~\ref{sec:proposal:memory_portion}) increase, giving rise to increasing the memory fraction of optimal design.
\begin{figure}[t]
\centering
\includegraphics[width=.48\textwidth]{figures/dram-size.pdf}
\caption{The sensitivity of performance improvement of various die-stacked DRAM organizations to the total capacity of die-stacked DRAM. \textit{MemCache-X\%} represents a hybrid design that dedicates \textit{X\%} of die-stacked DRAM capacity to memory and the rest to cache.
\label{fig:dram_size}}
\end{figure}
\section{Introduction}
The increase in core count of chip multiprocessors (CMPs) has been driving the designs into the memory bandwidth wall, mainly because of pin count limitations~\cite{bakhshalipour2018fast, Huh:2001:EDS:645988.674164, rogers:scaling}. Current CMPs with tens of cores already lose performance because of the limited bandwidth of DIMM-based DDR memories, and the problem is exacerbated as the number of cores increases~\cite{kgil:picoserver, lotfi:sop}. Therefore, continuing performance scaling through core count scaling requires a commensurate enhancement in the bandwidth of the memory system.
Emerging die-stacked DRAM technology is a promising solution to fulfill the ever-increasing bandwidth requirements of multi-core processors. The progress in manufacturing has enabled stacking several DRAM modules on the active die using high-density through-silicon vias (TSVs). Compared to traditional DIMM-based DDR memories, die-stacked DRAM provides several orders of magnitude higher bandwidth, but with approximately the same latency~\cite{hmc2014hybrid, chang2013reevaluating, chou2017batman, Lee_SMA, sodani2015intel, sodani2016knights}. In recent years, several models have been developed for the die-stacked DRAM~\cite{wideio, standard2013high, micronhmc, hmc2014hybrid, 3dics}, and many commercial vendors have planned to use such models in their products~\cite{nvidia-volta, nvidia-pascal, amd-fiji, amdradeon, lakka2012xilinx, sodani2015knights}.
One critical feature of the die-stacked DRAM is its limited capacity. Technological constraints, such as power-delivery and thermal limitations, restrict the size of the die-stacked DRAM to utmost a few gigabytes~\cite{healy2009study, lewis2011designing}. As such, it cannot accommodate the whole datasets of modern big-data applications with the datasets ranging from hundreds of gigabytes to a few terabytes. Consequently, prior proposals use the die-stacked DRAM either as a memory-side cache~\cite{el2013dual, hameed2013simultaneously, jang2016efficient, jevdjic:unison, jevdjic:footprint, jiang2010chop, Lee:2015:FAT:2749469.2750383, Loh:2009:EED:1669112.1669139, loh:efficiently, qureshi:alloy, sim2012mostly, sim2013resilient, volos2017fat, yu2017banshee, zhao2007exploring} or as a \emph{part} of software-visible main memory~\cite{agarwal2015unlocking, Agarwal:2015:PPS:2694344.2694381, dong2010simple, loh2012challenges, meswani:heterogeneous, Sim:2014:THM:2742155.2742158}, as shown in Figure~\ref{fig:scheme_cache_mem_memcache}.
\begin{figure}[t]
\centering
\includegraphics[width=.48\textwidth]{figures/cache-mem-memcache.pdf}
\caption{Comparison of proposals for the die-stacked DRAM.
\label{fig:scheme_cache_mem_memcache}}
\end{figure}
DRAM caches can \emph{quickly react to changes} in the data working sets of applications, and use the DRAM capacity more effectively by evicting cold data and keeping actively-accessed objects. However, the DRAM cache designs \emph{suffer from the tag overhead}, which is substantially more than the tag cost of traditional SRAM caches. As the size of the die-stacked DRAM is quite large, managing it as a cache requires megabytes of storage, somewhere in the system, for storing the tag information. As placing such a large structure in the active die (i.e., SRAM) is impractical, most proposals put the tags in the die-stacked DRAM itself~\cite{Gulur:2014:BDC:2742155.2742160, huang2014atcache, jevdjic:unison, loh:efficiently, qureshi:alloy, yu2017banshee}. Nevertheless, placing the tags in the die-stacked DRAM can add significant latency to the critical path of cache accesses, as the DRAM should be accessed twice; once for the tag and then for the data. Some approaches~\cite{jevdjic:unison, qureshi:alloy} store the tags and data next to each other and stream them out together in a single access to avoid this serialization latency overhead. While these approaches are able to mitigate the latency overhead of tag-checking, they still incur significant extra bandwidth overhead.
On the other hand, using die-stacked DRAM as a part of main memory eliminates the main drawback of caches: \emph{there is no need for tags}. Consequently, no tag storage is required, and the tag latency/bandwidth overhead is eliminated. Memory, however, \emph{cannot respond to dynamic variations} in the data working sets of applications, due to its static nature. In memory designs, over time, there would be many pieces of cold data in the die-stacked DRAM that are not currently being used, wasting its capacity. Prior work~\cite{meswani:heterogeneous} proposed to swap pages between die-stacked and off-chip DRAM periodically, using run-time Operating System (OS) support. Upon each period, the OS identifies hot and cold pages in off-chip and die-stacked DRAM, respectively, and swaps them in order to have the currently-used pages in the high-bandwidth DRAM. Such approaches, nonetheless, are not cheaply-realizable and present significant challenges to both software and hardware. First, as OS intervention is very time-consuming, the periodic intervals should be large enough (e.g., hundreds of milliseconds) to amortize the enormous costs of interrupts, page migrations, and TLB shootdowns. As such, the pages that are highly-utilized for a short duration of time cannot be captured by these approaches. Second, as full OS pages should be transferred between die-stacked and off-chip DRAM, there is a significant bandwidth overhead associated with these approaches. Third, as the OS has no precise information about the utilization of pages at run-time, it cannot robustly rank them, and correctly identify hot and cold pages.
In this work, we make a case for using the die-stacked DRAM partly as main memory and partly as a cache (Figure~\ref{fig:scheme_cache_mem_memcache}-c). We corroborate prior work~\cite{Chou:2014:CTM:2742155.2742157, dong2010simple, meswani2014toward} that using the whole capacity of the die-stacked DRAM as a part of the main memory is suboptimal to using it as a cache. However, we observe that, in modern big-data applications, there are numerous hot pages with a large number of accesses. Based on this observation, \emph{we classify application datasets into two distinct categories: hot datasets (hot pages), and transient datasets (transient pages)}. Hot datasets refer to data objects that are used \emph{steadily} by an application and \emph{serve a significant fraction of memory accesses} throughout the whole execution of an application. Transient datasets point to data structures that \emph{are utilized only for a short period of time}. In this work, we suggest \methodname{MemCache} for exploiting such heterogeneity in the access behavior of applications in the context of multi-gigabyte die-stacked DRAM. We suggest to use a portion of the die-stacked DRAM as a part of main memory and allocate hot pages in this part and use the rest of the capacity of the die-stacked DRAM as a hardware-managed cache to capture the transient datasets of applications. By allocating hot pages in the memory portion of the die-stacked DRAM, a considerable fraction of accesses are served from the high bandwidth memory without the overhead of tag-checking. The cache portion of the die-stacked DRAM remains intact and caches the transient datasets, providing quick responses to the dynamic variations in the datasets of applications.
To identify hot pages, we use a static profile-based approach before the execution of an application. A software procedure, incorporated into the compiler, pre-processes the application and sorts the pages based on their access frequency. Then it picks the top pages and asks the OS to map them to the memory portion of the die-stacked DRAM. We show that by using a representative input-set, such an offline analysis can classify pages robustly.
We evaluate \methodname{MemCache} against the state-of-the-art cache and memory proposals for the die-stacked DRAM and show that it significantly outperforms them on various big-data applications. Compared to a baseline without die-stacked DRAM, \methodname{MemCache} offers 114\% performance improvement on average and up to 309\%. Meanwhile, \methodname{MemCache} outperforms the best-performing prior design (\methodname{Banshee Cache}~\cite{yu2017banshee}) by 28\% on average and up to 139\%.
\section{Methodology}
\label{sec:methodology}
We use \componentname{ZSim}~\cite{sanchez-zsim} to simulate a system whose configuration is shown in Table~\ref{table:config}. We model the system based on one \componentname{Sub-NUMA Cluster (SNC)} of Intel's \componentname{Knights Landing} processor~\cite{sodani2016knightsieeemicro}. The chip has 16 OoO cores with an 8~MB LLSC and a 4~GB die-stacked DRAM. One channel is used for accessing off-chip DRAM, providing a maximum bandwidth of 21~GB/s, and four channels are responsible for establishing communications with the die-stacked DRAM, providing up to 84~GB/s bandwidth. In comparison, Intel's \componentname{Knights Landing} processor has four \componentname{SNCs}, offering 72 cores, 36~MB LLSC (512~KB per core; 8~MB per 16 cores), 16~GB die-stacked DRAM (4~GB per \componentname{SNCs}), 90~GB/s off-chip peak bandwidth (1.25~GB/s per core; 20~GB/s per 16 cores), and up to 400~GB/s stacked bandwidth (5.5~GB/s per core; 88~GB/s per 16 cores). DRAM modules are modeled based on \componentname{DDR3-1333} technology, parametrized with data borrowed from commercial device specifications~\cite{ddr3}. Physical addresses are mapped to memory controllers at 4~KB granularity. The link width among memory controllers and die-stacked DRAM is 16~B, but the minimum data transfer size is 32~B~\cite{yu2015imp, yu2017banshee}.
\begin{table}[h]
\sffamily
\begin{center}
\caption{Evaluation parameters.}
\label{table:config}
{
\resizebox{.48\textwidth}{!}{
\begin{tabular}{| c || c |}
\hline
{\bf {Parameter}} & {\bf {Value}} \\
\hline
\hline
{Cores} & Sixteen 4-wide OoO cores, 2.8~GHz \\
\hline
{L1-D/I} & 32~KB, 2-way, 1-cycle load-to-use \\
\hline
{L2 Cache} & 8~MB, 16-way, 15-cycle hit latency \\
\hline
{Die-Stacked DRAM} & 4~GB, 4 channels, 4~KB interleaved \\
\hline
{Off-Chip DRAM} & 1 channel, up to 21~GB/s bandwidth \\
\hline
\multirow{3}{*}{DRAM Modules} & DDR 1333~MHz, 8~KB row buffer \\
& 4 ranks/channel, 8 banks/rank, 16-byte bus width \\
& tCAS-tRCD-tRP-tRAS = 10-10-10-24 \\
\hline
\end{tabular}
}
} \small
\end{center}
\end{table}
\subsection{Workloads}
Table~\ref{table:workloads} summarizes the key characteristics of our simulated workloads. We use all graph processing workloads from \componentname{GAPBS}~\cite{beamer2015gap} and run them with the \componentname{Twitter} inputset~\cite{kwak2010twitter}. While the primary target of this work (and the systems that employ die-stacked DRAM) is throughput-oriented workloads like graph processing, yet for reference, we also include workloads from the \componentname{SPEC}~\cite{henning2006spec} benchmark suite. We choose four \componentname{SPEC} benchmarks whose memory footprints exceed 4~GB and run them with the \componentname{reference} inputset in \componentname{RATE} mode (i.e., all sixteen cores run the same single-thread program). We also consider a \appname{mix} of the four \componentname{SPEC} programs in which the processor executes four copies of each program.
\begin{table}[t]
\sffamily
\scriptsize
\begin{center}
\caption{Application parameters.}
\label{table:workloads}
\resizebox{.42\textwidth}{!}{
\renewcommand{\arraystretch}{1}
\begin{tabular}{| c || c | c |}
\hline
{\bf {Application}} & {\bf {LLSC MPKI}} & {\bf {Memory Footprint (GB)}} \\
\hline
\hline
\multicolumn{3}{|c|}{\bf {GAPBS}} \\
\hline
\hline
{\appname{bc}} & 61.4 & 92.9 \\
\hline
{\appname{bfs}} & 32.9 & 113.3 \\
\hline
{\appname{cc}} & 85.6 & 9.6 \\
\hline
{\appname{pr}} & 129.9 & 76.5 \\
\hline
{\appname{sssp}} & 73.4 & 140.1 \\
\hline
{\appname{tc}} & 12.5 & 53.4 \\
\hline
\hline
\multicolumn{3}{|c|}{\bf {SPEC}} \\
\hline
\hline
{\appname{bwaves}} & 17.6 & 13.1 \\
\hline
{\appname{gems}} & 26.7 & 12.6 \\
\hline
{\appname{mcf}} & 66.9 & 26.1 \\
\hline
{\appname{milc}} & 17.0 & 10.2 \\
\hline
{\appname{mix}} & 17.8 & 15.4 \\
\hline
\end{tabular}
}
\end{center}
\end{table}
\subsection{Die-Stacked DRAM Organizations}
We compare the following die-stacked DRAM organizations:\\
\begin{figure*}[t]
\centering
\makebox[\textwidth][c]{\includegraphics[width=\textwidth]{figures/performance.pdf}}
\caption{Performance improvement (bars) and MPKI (solid dots) of evaluated designs.
\label{fig:performance}}
\end{figure*}
\noindent
\textbf{\methodname{Alloy Cache}:} A state-of-the-art block-based cache design that uses a direct-mapped organization to minimize hit latency. \methodname{Alloy Cache} has a simple structure and was shown to be quite effective~\cite{qureshi:alloy}. \\
\noindent
\textbf{\methodname{Unison Cache}:} A state-of-the-art page-based four-way set-associative cache design with LRU replacement policy that uses a footprint predictor~\cite{jevdjic:footprint} to reduce the off-chip bandwidth pressure. Moreover, it uses a way predictor to avoid the serialization latency of tag and data accesses. In this paper, in order to decouple the results of cache design from the accuracy of predictors, we use \emph{perfect} predictors for fetching expected-to-use blocks and the cache way where the data is located at. For modeling a perfect footprint predictor, we follow a similar methodology as prior work~\cite{yu2017banshee}. We first profile applications offline and measure the average number of blocks that are used from cache pages (i.e., average footprint size). In timing simulations, whenever we want to bring a page into the die-stacked DRAM, we transfer as many cache blocks as the average footprint size of the application. However, unlike prior work~\cite{yu2017banshee}, we consider an accurate 1-line granularity footprint predictor (instead of 4-line granularity).\\
\noindent
\textbf{\methodname{Banshee Cache}:} A state-of-the-art page-based four-way set-associative cache design that tracks the tags of cache-resident pages in the Page Table and \componentname{Tag Buffer}. We consider a 1K-entry 8-way associative \componentname{Tag Buffer} (5.5~KB of SRAM storage) and set the baseline sampling coefficient to 10\% based on the original proposal~\cite{yu2017banshee}. Whenever the occupancy of the \componentname{Tag Buffer} exceeds 70\%, it should be flushed to the Page Table (and TLBs) to make all information coherent. We consider 25{$\mu{}s$} latency for the software procedure that updates the Page Table and shoots down TLBs based on the original proposal.\\
\noindent
\textbf{\methodname{HMA}:} The whole die-stacked DRAM is used as a part of main memory. Upon every $10^8$ LLSC misses, a software procedure swaps the pages between the die-stacked and off-chip DRAM to have currently-used pages in the high-bandwidth memory. For every page in the system, we consider a counter that counts the number of accesses to that page. Upon each OS interval, we \emph{precisely} sort the pages and allocate highly-used ones in the high bandwidth memory.
By accurately modeling the bandwidth of page swapping, we observed that the performance (IPC) of this technique consistently drops to \textit{zero} in all workloads because of the massive bandwidth overhead. The simulated workloads already utilize the bandwidth of both die-stacked and off-chip DRAM, and hence, there is not much unused bandwidth left for swapping numerous large pages on each interval. Increasing the page swapping period (e.g., once every 10 seconds) may amortize the associated bandwidth overhead, but doing so will further reduce the ability of this technique to respond to the dynamic changes in the memory access behavior of applications (i.e., transient datasets). In this paper, we model \emph{an idealized case}, in which, page swapping consumes \emph{no bandwidth} and happens instantaneously with \emph{no latency}.\\
\noindent
\textbf{\methodname{MemCache-S}:} Based on the performance sensitivity analysis (see Section~\ref{sec:eval:capacity_partition}), 3~GB of the die-stacked DRAM is used as a part of main memory, and the remaining 1~GB operates as a cache. Software allocates the identified hot pages in the memory portion of the die-stacked DRAM based on the \emph{warm-up} instructions (i.e., the instructions and the part of inputset which are used for identifying hot pages and filling the die-stacked memory are not used/counted in the actual experiments). Without loss of generality, we consider a \methodname{Banshee Cache} architecture for the cache portion of the die-stacked DRAM\footnote{In Section~\ref{sec:eval:cache_design}, we evaluate other cache architectures for the cache portion of \methodname{MemCache}, as well.}.\\
\noindent
\textbf{\methodname{MemCache-D}:} For every application, the memory-cache capacity partitioning is determined using Algorithm~\ref{alg:ahf}. The other specifications are the same as \methodname{MemCache-S}.\\
\noindent
\textbf{\methodname{Infinite}:} Die-stacked DRAM has infinite size and is used as a memory to host the \emph{whole} datasets of applications.
\subsection{Simulation Parameters}
For trace-driven experiments, we gather 16~billion LLSC misses, using the first 2~billion for warm-up and the rest for measurements. For timing analysis, we run the simulations for 200~billion instructions and use the first 20~billion for warm-up and the next 180~billion for measurements. We ensure that after the warm-up period, the die-stacked DRAM has become well-filled (i.e., no empty frame has remained), and the statistics are in the steady state.
\section{Motivation}
We first show that, in modern big-data applications, there is a discrepancy in page usage: some of the pages are used more frequently than the others. Then we compare two extreme use cases of the die-stacked DRAM: (1) whole die-stacked DRAM as a cache, and (2) whole die-stacked DRAM as a part of main memory. Finally, we motivate to use a part of the die-stacked DRAM as main memory and the rest as a cache.
\subsection{Hot Pages}
\label{motiv:hot-pages}
Corroborating many pieces of prior work (e.g.,~\cite{agarwal2015unlocking, Agarwal:2015:PPS:2694344.2694381, jiang2010chop, meswani2014toward, Sudan_MID, volos2016effective, volos2017fat, yu2017banshee}), we observe that, in modern big-data applications, there are many hot pages with a large number of accesses. Hot pages refer to the phenomenon that some pages are accessed with higher frequency than the others. This causes a small fraction of an application's memory working set to serve a significant fraction of accesses (e.g., 10\% of data working set serve 80\% of requests).
\begin{figure}[b]
\centering
\includegraphics[width=.48\textwidth]{figures/hot-pages.pdf}
\caption{Cumulative distribution of data accesses with pages sorted from hottest to coldest.
\label{fig:hot_pages}}
\end{figure}
Figure~\ref{fig:hot_pages} shows the cumulative distribution of accesses to the main memory for several big-data applications. The figure shows that most of the applications contain many hot pages that serve a significant fraction of total memory accesses. For example, in \appname{pr}, only 10\% of pages ($\sim$7~GB of memory footprint) serve more than 91\% of requests. The figure further indicates that applications consist of many transient pages. As an example, in \appname{pr}, the remaining 9\% of accesses are distributed among 90\% of pages with nearly the same access frequencies. \emph{We find that hot pages that are responsible for a significant fraction of accesses are used {steadily} throughout the whole execution of applications}\footnote{We elaborate more on this observation later in Section~\ref{eval:how_long}.}. In contrast, transient pages are used only for a short period of time.
The existence of hot pages motivates identifying and allocating them into the die-stacked memory to efficiently exploit the higher bandwidth that it provides. However, if we use the whole capacity of the die-stacked DRAM as a part of main memory and allocate hot pages into it, transient pages will be served from the off-chip DRAM. In contrast, if we use the die-stacked DRAM as a cache, we can serve both hot and transient datasets from the die-stacked DRAM, but have to pay the tag overhead (cf.~Section~\ref{sec:background:cache}).
\subsection{Full-Cache Versus Full-Memory}
\label{motiv:cache-vs-mem}
If the goal is to use the die-stacked DRAM as a part of main memory, the best one can do is to place pages with the highest number of references (i.e., hottest pages) in the die-stacked memory. Figure~\ref{fig:mem-vs-cache} shows the number of requests that go off-chip in a 4~GB die-stacked memory normalized to a 4~GB page-based cache for several big-data applications. In this experiment, we go over the sequence of accesses and find pages with the highest number of accesses and place them in the die-stacked memory.
\begin{figure}[t]
\centering
\includegraphics[width=.48\textwidth]{figures/mem-vs-cache.pdf}
\caption{The number of off-chip accesses of a 4~GB die-stacked DRAM memory as compared to that of a cache.
\label{fig:mem-vs-cache}}
\end{figure}
As Figure~\ref{fig:mem-vs-cache} clearly shows, using the whole capacity of the die-stacked DRAM as a memory considerably increases the number of accesses to the bandwidth-limited off-chip DRAM, as compared to using it as a cache. It comes from the fact that a die-stacked DRAM as memory cannot adapt itself to the dynamic run-time changes in the data working set of applications and yet is not large enough to capture the entire datasets of big-data applications. The results indicate that even an ideal memory, which is oracularly filled with the hottest pages, increases the number of off-chip accesses from $1.4\times$ in \appname{sssp} to $60.5\times$ in \appname{tc}, as compared to a cache design. The results are consistent with those of prior work~\cite{ Chou:2014:CTM:2742155.2742157, dong2010simple, meswani2014toward} concluded that, for big-data applications, memory designs are inferior to caches at reducing the number of off-chip accesses.
\subsection{What Fraction of Die-Stacked DRAM Can Be Turned into Memory?}
\label{sec:motiv:what_frac}
Even though using the whole capacity of the die-stacked DRAM as a part of main memory increases the number of off-chip accesses as compared to caches, the existence of hot pages with a considerable number of accesses may suggest that we can use a part of the die-stacked DRAM as a memory for hosting a subset of such hot pages and use the rest of the capacity as a cache for capturing the dynamic data-dependent behavior of applications.
The single major drawback of using the die-stacked DRAM as memory is its inability to respond to dynamic changes of applications. Over time, many transient pages will emerge in the data working set of an application, and since such pages were not detected and allocated in the die-stacked DRAM, a large number of requests will be sent to the bandwidth-limited off-chip memory. This causes memory designs to offer an order of magnitude lower hit\footnote{In this paper, by \emph{hit}, we refer to a state where the requested data is in the die-stacked DRAM, regardless of the fact that it is managed as a cache or as a part of memory. Analogously, by \emph{miss}, we refer to the state, in which, the requested data is not in the die-stacked DRAM.} ratio (higher off-chip accesses) as compared to cache designs (cf.~Section~\ref{motiv:cache-vs-mem}). Therefore, the mixture design (i.e., the design which uses the die-stacked DRAM partly as main memory and partly as cache) \emph{will lose nothing compared to the full-cache design, if it is able to offer a hit ratio as high as the cache}.
In order to determine what fraction of a die-stacked DRAM can be turned into memory without losing the benefits of the cache design, we first consider the die-stacked DRAM as a set of \emph{frames}. Each frame in the die-stacked DRAM is a \emph{physical location} where a piece of data at the granularity of a page reside, regardless of the fact that the die-stacked DRAM is managed as a cache or as a memory. We first consider the whole die-stacked DRAM as a sizeable page-based cache and calculate the \componentname{Average number of Hits per Frame (AHF)}. Then \emph{we turn each frame into memory and allocate the hottest unallocated page in it until the \emph{\componentname{AHF}} becomes smaller than that of the cache}. Algorithm~\ref{alg:ahf} summarizes the steps that we take in order to determine the fraction of die-stacked DRAM that can be turned into a memory.
\begin{algorithm}[t]
\caption{Calculate the Memory Fraction}\label{alg:ahf}
\begin{algorithmic}[1]
\footnotesize
\State $cacheAHF=totalCacheHits/dramFrames;$
\State $pages=sort(pages);$\Comment{{\scriptsize Sort pages based on their access count}}
\State $memFrames=0;$
\State $totalAccessesToMemFrames=0;$\newline
\While{$(totalAccessesToMemFrames\geq memFrames\times cacheAHF)$}
\State $hotPage = pages.top();$\Comment{{\scriptsize The hottest unallocated page}}
\State $memFrames\mathrel{++};$\Comment{{\scriptsize Allocate the hot page in the die-stacked memory}}
\State $pages.removeTop();$
\State $totalAccessesToMemFrames\mathrel{+}= hotPage.accessCount;$
\EndWhile\label{euclidendwhile}
\State \textbf{done}\newline
\State $memFraction=memFrames/dramFrames;$
\end{algorithmic}
\end{algorithm}
\begin{figure}[b]
\centering
\includegraphics[width=.48\textwidth]{figures/mem-fract.pdf}
\caption{The fraction of the die-stacked DRAM that can be managed like memory and still offers the same hit ratio as a page-based cache.
\label{fig:max_memory}}
\end{figure}
Figure~\ref{fig:max_memory} shows the fraction of a 4~GB die-stacked DRAM that can be managed as memory and still offer the same hit ratio as a 4~GB page-based cache. As the figure shows, a significant fraction of the die-stacked DRAM can be managed as a memory without negatively increasing the number of off-chip accesses. The fraction ranges from 11\% in \appname{gems}\footnote{\appname{gems} is an exceptional case, in which, requests are distributed among all pages with nearly the same frequency (i.e., it does not have a considerable number of hot pages. cf.~Figure~\ref{fig:hot_pages}).} to 96\% in \appname{pr}, with an average of 65\%. This means that, on average, 65\% of a 4~GB DRAM can be turned into memory without losing the benefits of a full-cache design. Note that, this is the fraction, at which, the \emph{hit ratio} of the design that uses the die-stacked DRAM partly as main memory and partly as cache remains intact as compared to the full-cache design. While the hit ratio is important for die-stacked DRAM organizations, prior work showed that minimizing \emph{hit latency}~\cite{qureshi:alloy} or efficiently utilizing the \emph{die-stacked DRAM bandwidth}~\cite{yu2017banshee} are of equal importance if not more. As memory organizations generally offer lower access latency than cache organizations and do not waste DRAM bandwidth for replacements/tag manipulations, we expect that even a larger fraction of the die-stacked DRAM ($> 65\%$ on average) can be managed as a part of main memory. We determine the fraction of the die-stacked DRAM's capacity that can be managed as a memory using a performance sensitivity analysis in Section~\ref{sec:eval:capacity_partition}.
Taking advantage of this observation, we propose to use a portion of the die-stacked DRAM as a part of the main memory and allocate hot pages to it and use the rest of the capacity of the die-stacked DRAM as a hardware-managed cache to capture the transient data-dependent behavior of applications. By allocating hot pages in the memory portion of the die-stacked DRAM, a significant fraction of accesses is served from the high-bandwidth memory without the overhead of tag-checking (bandwidth and latency). The cache portion of the die-stacked DRAM also remains intact and caches the transient pages, providing quick responses to dynamic variations in the datasets of applications.
\section{The Proposal}
In order to exploit the heterogeneity in the access behavior of applications to hot and transient pieces of data, we propose \methodname{MemCache}, a design that uses a part of the die-stacked DRAM as main memory and the rest as a cache. The software is responsible for filling the die-stacked memory with the identified hot pages. The transient pages are also served from the hardware-managed cache portion of the die-stacked DRAM. For every \mbox{Last-Level SRAM Cache (LLSC)} miss, if the miss address is mapped to the memory portion of the die-stacked DRAM, it is served by the die-stacked memory (i.e., without the overhead of tag-checking). Otherwise, it checks the DRAM cache tag array. In case of a cache hit, the request is served by the cache portion of the die-stacked DRAM, and in case of a cache miss, it proceeds to the off-chip memory.
\subsection{Memory Portion}
\label{sec:proposal:memory_portion}
The memory frames of the die-stacked DRAM should be filled with the hot pages of applications. First off, the hot pages of an application should be identified. We use a \emph{profiling} approach in order to distinguish hot pages from cold ones. A software procedure, incorporated into the compiler, using profiling, determines which \emph{virtual pages} of the application are hot. To do so, the software procedure sorts virtual pages based on their access count, in descending order, and then, conveys this information as clues to the page allocation unit of the OS.
There are quite a few ways to implement the profiling step needed in our mechanism, two of which are more common and are elaborated upon here. One approach is that the compiler profiles the program by simulating the behavior of the on-chip cache hierarchy of the target machine~\cite{ebrahimi2009techniques}. The simulation is used to gather the access counts of pages, which are then used to identify hot pages. Note that, such a profiling approach does not necessitate an accurate timing simulation of the processor and cache hierarchy; rather, it requires simply a trace-driven simulation of memory operations in order to gather information about accesses that are not captured in on-chip SRAM caches (i.e., LLSC misses and evictions). Another approach for implementing the profiling step is relying on hardware support. In this strategy, the target machine provides support for profiling operations (e.g., \emph{Informing Load Operations}~\cite{Horowitz_IMO}). With this support, throughout the profiling run, the compiler gathers the access counts of application's pages, and at the end, sorts them accordingly. In this paper, we use the first approach as it requires no change in the hardware.
Identifying hot pages by the compiler demands for a \emph{sample inputset} on which the profiling step should be run. The sample inputset must be large enough to enable the compiler to observe various pages, gather adequate access counts on them, and distinguish hot pages from cold ones by sorting them. To shed light on how large the inputset should be, we perform the following experiment: First, we use quite large inputsets\footnote{The details of applications and their inputsets are explained in Section~\ref{sec:methodology}.}, such that simulating the on-chip cache hierarchy on them, produces 10~B LLSC misses/evictions for each application. Based on these inputsets, we find the hottest pages of every application that fill a 4~GB memory. We refer to the set of these hot pages (identified by evaluating 10~B LLSC misses/evictions) as \emph{ideal pages}. Then, we redo the same experiment based on just a part of the same inputsets, that results in 1~B LLSC misses/evictions for every application, and again identify the hottest pages that fill a 4~GB memory. Finally, we compare these hot pages with ideal pages and report the ratio of similar pages to all pages as the \emph{accuracy of page classification}. We also repeat the same experiment based on 2~B, 3~B, \ldots{}, and 9~B LLSC misses/evictions. Figure~\ref{fig:classification-accuracy} shows how the accuracy of page classification varies with increasing the sample inputset size.
\begin{figure}[t]
\centering
\includegraphics[width=.48\textwidth]{figures/classification-accuracy.pdf}
\caption{The effect of sample inputset size on the accuracy of page classification.
\label{fig:classification-accuracy}}
\end{figure}
A critical observation we make here is that \emph{by observing a small fraction of the inputset, a static profile-based approach is able to classify pages accurately}. This is especially true for modern throughput-oriented big-data graph applications. For example, in \appname{cc}, classifying pages into hot and cold based on observing the first 1~B accesses yields 98\% of the accuracy of an ideal classifier which is aware of the future and classifies pages based on the whole 10~B accesses\footnote{Later in Section~\ref{eval:how_long}, we show that not only is such a page classification accurate, but it also produces robust results. That is, the pages that are classified as hot remain hot throughout the whole execution of an application. }. On average across all applications, by observing only 1~B accesses, a profile-based approach is able to classify pages accurately, realizing roughly three-fourths of the accuracy of an ideal mechanism. Note that this fairly simple mechanism (i.e., profile-based classification of pages) works well mainly because of the fact that the size of the die-stacked DRAM is quite large. That is, the die-stacked DRAM is capacious enough to accommodate thousands and even millions of pages, and a few misclassifications do not have a significant effect on the overall accuracy\footnote{Generally, the accuracy of probabilistic methods (such as the employed profile-based method that estimates a set of hot pages for the whole execution of an application based on observing a small fraction of accesses) for \emph{Top-N Recommendation} increases with increasing \emph{N}~\cite{Barbieri_APM}.}.
After detection of hot pages, their details are coded into the program binary. Whenever the program gets executed, the \componentname{Loader} passes the required information of hot pages to the OS. Then, the OS tries\footnote{Sometimes, there might be conflicting preferences/constraints that get in the way of the OS allocating a page to a specific physical address space~\cite{bovet2005understanding}. Resolving such issues is beyond the scope of this paper. } to map such hot pages to physical locations that belong to the memory portion of the die-stacked DRAM. For the OS, allocating pages in the die-stacked and off-chip memory is similar to the same operations in Non-Uniform Memory Architecture (NUMA)~\cite{Larowe_ECM} systems\footnote{Here, NUMA regions are further broken into a die-stacked memory region and an off-chip memory region.}. Therefore, the OS, as well as system libraries, is able to use the same memory allocation algorithms for allocating pages in the die-stacked and off-chip memory. Furthermore, other operations, like virtual address translation and virtual memory management, for the die-stacked and off-chip memory are quite the same as in a NUMA system.
\subsection{Cache Portion}
The cache portion of the die-stacked DRAM filters requests to the pages that are allocated in the off-chip memory (i.e., pages that have not been identified as hot). The cache is transparent to the software and can be managed using any DRAM cache management technique. In this paper, we evaluate three cache architectures (\methodname{Alloy Cache}~\cite{qureshi:alloy}, \methodname{Unison Cache}~\cite{jevdjic:unison}, and \methodname{Banshee Cache}~\cite{yu2017banshee}) for the cache portion of the \methodname{MemCache}.
\subsection{Memory-Cache Capacity Partitioning}
The optimal capacity of the die-stacked DRAM which can be turned into a part of the main memory directly depends on the applications. For applications with significant hot pages (e.g., \appname{pr}), it is better to dedicate less space for the cache and more to memory. On the other hand, for applications with few hot pages (e.g., \appname{gems}), partitioning the capacity towards more cache would be beneficial. However, dynamically changing the partition based on the running applications makes the system more complex, especially when the system is supposed to run multiple applications and \emph{frequently} switch from one application to another.
In this paper, we consider two variants for \methodname{MemCache}; one design which its partitioning is tuned based on the application, and another which has fixed partitioning. In the first design, named \methodname{MemCache-D}, the compiler determines what fraction of the die-stacked DRAM capacity should be devoted to the memory and what fraction to the cache. To do so, the compiler runs Algorithm~\ref{alg:ahf}, thereby finding a fairly suitable partitioning\footnote{As discussed in Section~\ref{sec:motiv:what_frac}, in Algorithm~\ref{alg:ahf}, first a full-cache design is simulated. Then, the frames of the die-stacked DRAM are gradually turned into memory until the whole design becomes susceptible to fall behind a full-cache design.}. Then, the OS, using a specific instruction, announces the partitioning to the memory controllers. This way, the memory controllers become able to redirect requests to their correct locations\footnote{By checking the physical address of a request and comparing it with the physical address of the last frame which has been devoted to memory, the memory controller decides whether to send the request to the memory portion of the die-stacked DRAM or serve it from the cache portion.}. \methodname{MemCache-D} adapts itself based on the application behavior and is supposed to offer higher performance than a design that its partition is determined statically. However, it needs to change the partitioning for every application, which obviously is a costly operation, and may not be effective when the system runs multiple applications, frequently switching from one to another\footnote{When the memory-cache partitioning of the die-stacked DRAM changes, not only is a system-wide TLB shootdown required, but also the valid data in the memory frames and modified data in the cache frames should be written back to the off-chip memory if the functionality of the frames changes.}. Therefore, we believe that \methodname{MemCache-D} is a suitable design for systems where a specific application (or a set of specific applications\footnote{In this paper, we evaluate this technique when multiple \emph{dissimilar} applications co-run on the CMP.}) runs for a long time (e.g., datacenter applications).
The other variant of \methodname{MemCache} we consider in this work is \methodname{MemCache-S}. In \methodname{MemCache-S}, a \emph{fixed} fraction of the die-stacked DRAM is devoted to the memory and the rest to the cache. The partitioning is determined by the user, \emph{at boot time}, and is then reported to the OS by the BIOS, along with other information. While a fixed partitioning might not give the best performance for all of the applications, it realizes a significant fraction of the performance benefit of dynamic application-based partitioning (i.e., \methodname{MemCache-D}). In this paper, we choose the capacity partitioning of \methodname{MemCache-S} based on the performance sensitivity analysis of all applications (Section~\ref{sec:eval:capacity_partition}) and show that such a static design offers a level of performance close to that of a design that picks the best memory-cache partitioning for each application.
The case where multiple applications share the processor, and the system frequently switches from one application to another, presents complications to \methodname{MemCache} and other hybrid memory systems where multiple dissimilar memories are used (e.g., \componentname{DRAM+NVM} systems where \componentname{DRAM} has characteristics entirely different from \componentname{NVM}, or even \componentname{NUMA} systems where accessing \componentname{Local Memory} has lower latency than \componentname{Remote Memories}). The problem arises from the fact that when the hot pages of an application (or set of applications) are being allocated in the die-stacked memory, information about the hot pages of other applications that will possibly be running in the future is unknown; therefore, filling the die-stacked DRAM merely based on the current applications may result in suboptimal performance if future applications will be more bandwidth-hungry and require more capacity from the die-stacked DRAM. Various proposals (e.g., \cite{Agarwal_TAP, knyaginin2018profess, li2017utility, Ogasawara_NMM, Verghese_OSS}) have suggested to optimize memory management in such situations typically by gradually or periodically migrating application pages between different types of memories based on factors like programming model, application's criticality, sharing degree, and so on. Such approaches are orthogonal to \methodname{MemCache} and can augment it to provide performance/fairness benefits; however, we leave the analysis of enhancing our method with these techniques for future work, promoting a clearer understanding of our contribution in the context of prior die-stacked DRAM literature.
\section{Related Work}
\label{sec:related}
To the best of our knowledge, this is the first research that proposes to use the die-stacked DRAM partly as main memory and partly as a cache. Nevertheless, \methodname{MemCache}, in the sprite, is similar to \componentname{Hybrid} configuration of \componentname{Multi-Channel DRAM (MCDRAM)} modules in Intel's \componentname{Knights Landing} processor~\cite{sodani2016knightsieeemicro}. In \componentname{MCDRAM} modules of \componentname{Knights Landing} processor, similar to \methodname{MemCache-S}, a fixed portion of the die-stacked DRAM is devoted to memory, and the rest acts as a cache. The \emph{programmer} is responsible for allocating data structures in the memory portion of \componentname{MCDRAM} modules, using high-level instructions like {\ttfamily {hbw$\_$malloc}} and {\ttfamily FASTMEM}. In this paper, we showed that such a heavy burden on the programmer could be lifted with compile-time page classification. We showed that the large capacity of the die-stacked DRAM paves the way for offloading the hot-pages--identification task to the compiler, providing programming ease (cf.~Section~\ref{sec:proposal:memory_portion}). Moreover, we signified the potentials for dynamism in partitioning the capacity of the die-stacked DRAM, showing that the dynamic partitioning is able to outperform the static one by as much as 84\% (cf.~Section~\ref{sec:eval:perf}).
Some pieces of prior work~\cite{Agarwal:2015:PPS:2694344.2694381, loh2012challenges, meswani2014toward} also proposed to profile applications and manage the die-stacked DRAM based on the outcome. However, all of these approaches use the die-stacked DRAM entirely as a part of main memory and neither proposed nor discussed a design that uses the die-stacked DRAM partly as main memory and partly as a cache.
Chou et al.~\cite{Chou:2014:CTM:2742155.2742157} proposed \methodname{Cameo} to minimize the number of accesses to the backend storage (e.g., SSD or Disk). \methodname{Cameo} manages die-stacked and off-chip DRAM in different address spaces to increase the physical address space of the system while attempting to preserve the benefits of caching via keeping recently-accessed data in the high-bandwidth DRAM. \methodname{MemCache}, however, uses a part of the die-stacked DRAM as the main memory to enable efficient access to hot data structures without the overhead of tag-checking. \methodname{MemCache} is distinct from \methodname{Cameo} but brings most of its benefits for free. Yet, \methodname{Cameo} is orthogonal to \methodname{MemCache} since the cache part of the die-stacked DRAM in \methodname{MemCache} can be managed like \methodname{Cameo}.
\methodname{Tertiary Caching} was proposed in the context of multi-socket systems. In \methodname{Tertiary Caching}~\cite{thekkath1997evaluation, zhang1997reducing}, a portion of each node's local memory is managed as a cache for caching only data objects that are allocated in remote nodes. While there are similarities between \methodname{MemCache} and \methodname{Tertiary Caching}, they are conceptually different. In \methodname{MemCache}, the memory part is used for keeping only hot pages, while in \methodname{Tertiary Caching}, memory hosts all local pages. Besides, \methodname{MemCache} uses the cache for the transient datasets of applications, while \methodname{Tertiary Caching} dedicates a part of memory to the cache in order to cache only remotely-allocated data objects.
In addition to schemes thoroughly discussed in Section~\ref{sec:background}, there are other proposals for managing the die-stacked DRAM as a cache. Loh and Hill~\cite{loh:efficiently} suggested a set-associative DRAM cache in which the tags and data of each set fit in a single DRAM row, allowing it to serve cache hits without the need to open more than one DRAM row. Jiang et al.~\cite{jiang2010chop} proposed \methodname{Chop}, a design that caches only pages with high expected reuse to avoid the significant bandwidth overhead of page-based caches. Loh~\cite{Loh:2009:EED:1669112.1669139} proposed organizing each DRAM cache set as a multi-level queue to evict dead-on-arrival cache lines immediately after insertion. Gulur et al.~\cite{Gulur:2014:BDC:2742155.2742160} proposed \methodname{Bi-Modal Dram Cache} to obtain the advantages of both block-based and page-based caches via dynamically choosing the caching granularity. Sim et al.~\cite{sim2012mostly} proposed techniques that enable accessing off-chip DRAM when the die-stacked DRAM cache bandwidth is highly utilized. In this paper, we showed that a large DRAM cache is suboptimal to a design that uses both cache and memory at the same time.
Numerous strategies were proposed to improve the efficiency of die-stacked DRAM. \methodname{Dap}~\cite{gaur2017near} and \methodname{Batman}~\cite{chou2017batman} attempt to maximize the total bandwidth utilization of systems with both die-stacked and DIMM-based DRAM modules. These approaches forbid data movement from the off-chip DRAM to the die-stacked DRAM when the bandwidth usage of the die-stacked DRAM exceeds a certain threshold to efficiently exploit the available bandwidth of both DRAM components. Franey and Lipasti~\cite{franey2015tag} proposed \methodname{Tag Tables}, a technique for compressing the tag array of DRAM caches to enable fabricating it in SRAM. \methodname{Accord}~\cite{youngaccord} provides associativity for direct-mapped \methodname{Alloy Cache} using way-prediction and way-install. Chou et al.~\cite{baer-isca} suggested \methodname{Bear} for reducing the bandwidth pressure of die-stacked DRAM caches. \methodname{Bear} decreases the tag-checking bandwidth by maintaining certain status details of the DRAM cache in the active die. Moreover, it samples replacement decisions, effectively trading hit ratio for bandwidth efficiency. Young et al.~\cite{young2017dice} introduced \methodname{DICE}, a technique for compressing DRAM caches mainly for reducing bandwidth usage and then for having a higher effective capacity. \methodname{DICE} attempts to get multiple useful blocks in a single DRAM access by compressing data and adaptively adjusting the set-indexing scheme of the cache. Huang and Nagarajan~\cite{huang2014atcache} proposed \methodname{ATCache} to provide faster access to the tag array of a DRAM cache. \methodname{ATCache} exploits the empirically-observed spatio-temporal locality in tag accesses and caches/prefetches some tags of the die-stacked DRAM in an SRAM structure. \methodname{Candy}~\cite{chou2016candy} and \methodname{C$^3$D}~\cite{huang2016c} facilitate the use of DRAM caches in the context of multi-node systems by making them coherent. Most of these approaches are orthogonal to our work and can be used together.
|
\section{Introduction}
Since the pioneer work of Seiberg and Witten \cite{Seiberg1994,*Seiberg1994a}, 4D ${\mathcal{N}}=2$ super Yang-Mills (SYM) theories have manifested deep connections with integrability. Thus, in the Euclidean background $\mathbb{R}^4$, their BPS sector is described by a classical integrable system \cite{Gorsky1995,*Donagi1995}. This system becomes quantized after the (partial) omega-deformation of the background $\mathbb{R}^2_{\varepsilon_1}\times\mathbb{R}^2$, the parameter $\varepsilon_1$ playing the role of a Planck constant \cite{Nekrasov2009}. Furthermore, the partition function of the (fully) omega-deformed theories on $\mathbb{R}^2_{\varepsilon_1}\times\mathbb{R}_{\varepsilon_2}^2$ has been computed exactly by Nekrasov using a localization technique \cite{Nekrasov2003}. These quantities receive non-perturbative (instanton) corrections. They were further shown to be equal to the W-algebras' conformal blocks describing correlation functions of Liouville/Toda integrable conformal field theories under the celebrated AGT-correspondence \cite{Alday2010,*Alba2010}. Some of these integrable aspects follow from an action of Vasserot-Schiffmann's Spherical Hecke central (SHc) algebra \cite{Schiffmann2012,*Kanno2012a,*Kanno2013} (formally equivalent to the affine Yangian of $\mathfrak{gl}_1$ \cite{Prochazka2015,Tsymbaliuk2014}), for which representations of level $m$ contains the action of the $W_m$-algebra involved in the AGT correspondence. However, the complete description of algebraic structures sustaining all the observed integrable properties is still lacking.
As it turns out, the algebraic description is better understood when the gauge theory is lifted to five dimensions by an extra $S^1$. The corresponding gauge theories have ${\mathcal{N}}=1$ supersymmetry, and describe the low-energy dynamics of $(p,q)$-branes in type IIB string theory \cite{Aharony1997,Aharony1997a}. Alternatively, they can be obtained as the topological string amplitude on a toric Calabi-Yau threefold \cite{Aganagic2005,*Iqbal2007,*Awata2005}, and the toric diagram coincides with the $(p,q)$-brane web \cite{Leung1998}. This string theory realization is the basis of an algebraic construction, first proposed in \cite{Mironov2016,*Awata2016a}, based on the Ding-Iohara-Miki (DIM) algebra \cite{Ding1997,Miki2007}, which we call here \textit{algebraic engineering}. Representations of the DIM algebra are labeled by two integer levels, and are essentially made from two basic ones called \textit{vertical} $(0,m)$ \cite{feigin2011quantum} and \textit{horizontal} $(1,n)$ \cite{Feigin2009a}. In engineering ${\mathcal{N}}=1$ gauge theories, a representation of level $(q,p)$ is associated to each brane of charge $(p,q)$. Thus, D5 branes are associated to vertical representations $(0,1)$, and NS5-branes (possibly dressed by $n$ D5s) to horizontal ones. A refinement of this technique has been developed in which stacks of $m$ D5-branes (with different positions) correspond to vertical representations of higher levels $(0,m)$ \cite{Bourgine2017b}. The junctions between branes are described in string theory using the refined topological vertex, which has also been written in the form of an intertwiner for the DIM algebra \cite{Awata2011}. In this way, it has been possible to associate to the brane web of linear (and also D-type, see \cite{Bourgine2017c}) quivers a certain ${\mathcal{T}}$-operator acting on the tensor product of horizontal representations associated to external (dressed) NS5-branes \cite{Bourgine2017b}. The vacuum expectation value (vev) of this operator reproduces Nekrasov's instanton partition functions, while further insertions of DIM currents define the qq-characters \cite{NPS,*Nekrasov2015,*Kim2016,Bourgine2015c,Bourgine2016}.
In this note, we extend this algebraic construction to the case of 4D ${\mathcal{N}}=2$ theories. For this purpose, we introduce a formal current algebra seen as a degenerate limit of the DIM algebra. This algebra has a Hopf algebra structure with a Drinfeld-like coproduct, an important feature to define intertwiners. We proceed by constructing two different types of representations, which we call again vertical and horizontal, by analogy with the DIM algebra. The vertical representation is roughly equivalent to the action of SHc in which generators act on states labeled by a set of Young diagrams. On the other hand, the horizontal representation appears to be new. It is defined in terms of twisted vertex operators built upon the modes of a free boson. Then, we solve the intertwining relation coupling these two representations, thus defining a 4D equivalent of the topological vertex. From this point, the 5D method can be readily adapted to ${\mathcal{N}}=2$ theories, and, as an example, we construct the ${\mathcal{T}}$-operator for the pure $U(m)$ gauge theory, and re-derive both the instanton partition function and the fundamental qq-character. This ${\mathcal{T}}$-operator commutes with the action of the degenerate DIM algebra in a certain tensored horizontal representation. In the 5D case, the corresponding action of DIM can be decomposed into q-Heisenberg and q-Virasoro actions \cite{Mironov2016}, the latter being identify with Kimura-Pestun (KP) quiver $W$-algebra for the $A_1$-quiver \cite{Kimura2015}. A similar decomposition is observed in the degenerate case, leading to define the equivalent of KP's algebra for 4D theories. In this simple case, it is identified with a certain limit of the q-Virasoro algebra \cite{Shiraishi1995} that, surprisingly, also coincides with the Zamolodchikov-Faddeev (ZF) algebra for the sine-Gordon model \cite{Lukyanov1995}.
\section{Degenerate DIM algebra and representations}
\subsection{Definition of the algebra}
Following the analysis of the degenerate limit of the DIM algebra presented in \ref{AppA}, we introduce the following algebra obeyed by four currents denoted $x^\pm(z)$, $\psi^\pm(z)$, and a central element $c$,
\begin{align}\label{algebra}
\begin{split}
&[\psi^\pm(z),\psi^\pm(w)]=0,\quad \psi^\pm(z)\psi^\mp(w)=\dfrac{g(z-w-c\varepsilon_+/2)}{g(z-w+c\varepsilon_+/2)}\psi^\mp(w)\psi^\pm(z),\\
&\psi^+(z)x^\pm(w)=g(z-w\mp c\varepsilon_+/4)^{\pm1}x^\pm(w)\psi^+(z),\quad \psi^-(z)x^\pm(w)=g(z-w\pm c\varepsilon_+/4)^{\pm1}x^\pm(w)\psi^-(z),\\
&x^\pm(z)x^\pm(w)=g(z-w)^{\pm1}x^\pm(w)x^\pm(z),\\
&[x^+(z),x^-(w)]=-\dfrac{\varepsilon_1\varepsilon_2}{\varepsilon_+}\left[\delta(z-w+c\varepsilon_+/2)\psi^+(z+c\varepsilon_+/4)-\delta(z-w-c\varepsilon_+/2)\psi^-(z-c\varepsilon_+/4)\right].
\end{split}
\end{align}
This algebra depends on two parameters $\varepsilon_1$ and $\varepsilon_2$, identified with the omega-background equivariant parameters. It has been convenient to introduce the notation $\varepsilon_+=-\varepsilon_3=\varepsilon_1+\varepsilon_2$, and the function
\begin{equation}\label{def_g}
g(z)=\prod_{\alpha=1,2,3}\dfrac{z+\varepsilon_\alpha}{z-\varepsilon_\alpha},
\end{equation}
plays the role of a scattering factor. It obeys the unitarity property $g(z)g(-z)=1$. In contrast with the case of DIM, the delta function $\delta(z)$ here is not the multiplicative one, but it is simply defined as $2i\pi$ times the Dirac delta function.
We will not discuss in details the mathematical definition of the algebra \ref{algebra} that should correspond to a Drinfeld double of the affine Yangian of $\mathfrak{gl}_1$ \cite{Tsymbaliuk2014}.\footnote{We would like to thank Anton Nedelin, Sara Pasquetti and Yegor Zenkevich for communicating this observation to us.} Instead, we will present two types of representations: an horizontal representation of weight $u$, denoted $\rho_u^{(H)}$ and having level $\rho_u^{(H)}(c)=1$, and a set of vertical representations of rank $m$ and weight $\vec a$, denoted $\rho_{\vec a}^{(m)}$ and such that $\rho_{\vec a}^{(m)}(c)=0$. The 4D ${\mathcal{N}}=2$ SYM theories can also be obtained from a brane construction, this time using type IIA string theory \cite{Witten1997}. In this construction, D5 and NS5 branes are replaced by D4 and solitonic 5-branes respectively. The horizontal representation will be associated to the solitonic 5-branes, while stacks of $m$ D4-branes, on which act a $U(m)$ gauge group, will be associated to the rank $m$ vertical representation. The weights of representations will be identified with the branes positions. In addition, automorphisms of the DIM algebra encode the invariance of the brane-web under geometric transformations \cite{Bourgine2018a}. These automorphisms can also be defined on the algebra \ref{algebra}, except for Miki's automorphism $\mathcal{S}$ \cite{Miki2007} corresponding to a $90^\circ$ rotation of the brane-web that maps D5 on NS5 and vice versa, thus realizing the S-duality of type IIB string theory. On the other hand, the 180$^\circ$ rotation $\mathcal{S}^2$ is preserved:
\begin{align}\label{def_autom}
\begin{split}
&\mathcal{S}^2\cdot c=-c,\quad \mathcal{S}^2\cdot x^\pm(z)=-x^\mp(-z),\quad \mathcal{S}^2\cdot\psi^\pm(z)=\psi^\mp(-z),\\
&\tau_\omega \cdot c=c,\quad \tau_\omega \cdot x^\pm(z)=x^\pm(z+\omega ),\quad\tau_\omega \cdot\psi^\pm(z)=\psi^\pm(z+\omega ),\\
&{\bar\tau}_{\bar\omega}\cdot c=c,\quad {\bar\tau}_{\bar\omega}\cdot(x^\pm(z))={\bar\omega}^{\pm1}x^\pm(z),\quad {\bar\tau}_{\bar\omega}\cdot\psi^\pm(z)=\psi^\pm(z).
\end{split}
\end{align}
Here $\tau_\omega $ and ${\bar\tau}_{\bar\omega}$ define translations in the directions corresponding to the solitonic 5-branes and the D4-branes respectively. In addition, two reflection symmetries $\sigma_V$ and $\sigma_H$ (resp. along the directions of D5/D4-branes and NS5/solitonic 5-branes) were introduced in \cite{Bourgine2017b,Bourgine2017c} as involutive isomorphisms of algebra. Their definition extend to the algebra \ref{algebra}, mapping it to the same algebra but with the opposite parameters $\varepsilon_\alpha\to-\varepsilon_\alpha$, and acting on the generators as
\begin{align}
\begin{split}
&\sigma_V(c)=c,\quad \sigma_V\cdot x^\pm(z)=-x^\mp(z),\quad \sigma_V\cdot \psi^\pm(z)=\psi^\pm(z),\\
&\sigma_H(c)=-c,\quad \sigma_H\cdot x^\pm(z)=x^\pm(-z),\quad \sigma_H\cdot \psi^\pm(z)=\psi^\mp(-z).
\end{split}
\end{align}
Note that $\mathcal{S}^2$ is also involutive, and $\sigma_V\sigma_H=\mathcal{S}^2$.
The main difference between \ref{algebra} and the DIM algebra is the absence of a mode expansion for the currents $x^\pm(z)$ and $\psi^\pm(z)$. This is due to the presence of zero modes in the horizontal representation, but also to the fact that we use the Dirac delta function instead of the multiplicative one. In practice, the algebra \ref{algebra} is simply a convenient way to combine the algebraic relations satisfied by the currents in the two types of representations. Yet, its main interest lies in the possibility to define a co-algebraic structure and, later on, the corresponding intertwiners. The coproduct can be seen as a degenerate limit of the Drinfeld coproduct for the DIM algebra,
\begin{align}
\begin{split}\label{def_D}
&\Delta(x^+(z))=x^+(z)\otimes1+\psi^-(z-\varepsilon_+ c_{(1)}/4)\otimes x^+(z-\varepsilon_+ c_{(1)}/2),\\
&\Delta(x^-(z))=1\otimes x^-(z)+x^-(z-\varepsilon_+ c_{(2)}/2)\otimes\psi^+(z-\varepsilon_+ c_{(2)}/4),\\
&\Delta(\psi^\pm(z))=\psi^\pm(z\mp\varepsilon_+ c_{(2)}/4)\otimes\psi^\pm(z\pm\varepsilon_+ c_{(1)}/4),
\end{split}
\end{align}
$\Delta(c)=c_{(1)}+c_{(2)}$ with $c_{(1)}=c\otimes 1$ and $c_{(2)}=1\otimes c$. In fact, the whole Hopf algebra structure survives the limit, the co-unit taking the usual form $\epsilon(x^\pm(z))=\epsilon(c)=0$, $\epsilon(\psi^\pm(z))=\epsilon(1)=1$, and the antipode $\theta(c)=-c$, $\theta(\psi^\pm(z))=\psi^\pm(z)^{-1}$ and
\begin{equation}
\theta(x^+(z))=-\psi^-(z+\varepsilon_+c/4)^{-1}x^+(z+\varepsilon_+c/2),\quad \theta(x^-(z))=-x^-(z+\varepsilon_+ c/2)\psi^+(z+\varepsilon_+ c/4)^{-1}.
\end{equation}
As in \cite{Bourgine2018a}, the opposite (or permuted) coproduct $\Delta'$ can also be obtained as a twist of $\Delta$ by the automorphism $\mathcal{S}^2$.
\subsection{Horizontal representation}
The horizontal representation is built using vertex operators acting on the Fock space of a free boson. The modes $\alpha_n$ and $P$, $Q$ satisfy the standard Heisenberg commutation relations $[\alpha_n,\alpha_m]=n\delta_{n+m}$ and $[P,Q]=1$. The vacuum $\ket{\emptyset}$ is annihilated by the positive modes $\alpha_{n>0}$ and $P$, while negative modes $\alpha_{n<0}$ and $Q$ create excitations. As usual the normal ordering $:\cdots:$ is defined by moving the positive modes to the right. We also introduce the dual state $\bra{\emptyset}$ annihilated by negative modes, so that $\bra{\emptyset}:\cdots:\ket{\emptyset}=1$. In our construction, and just like in DIM's horizontal representations \cite{Feigin2009a}, twists between positive and negative modes have to be introduced, and it is convenient to define
\begin{equation}\label{def_vphi}
\varphi_+(z)=P\log z-\sum_{n>0}\dfrac{z^{-n}}{n}\alpha_n,\quad \varphi_-(z)=Q+\sum_{n>0}\dfrac{z^n}{n}\alpha_{-n}\quad\Rightarrow\quad e^{\varphi_+(z)}e^{\varphi_-(w)}=(z-w):e^{\varphi_+(z)}e^{\varphi_-(w)}:.
\end{equation}
In fact, the bosonic modes enter in the representation only through the combinations $V_-(z)=e^{\varphi_-(z-\varepsilon_2/2)-\varphi_-(z+\varepsilon_2/2)}$ and $V_+(z)=e^{\varphi_+(z+\varepsilon_1/2)-\varphi_+(z-\varepsilon_1/2)}$ that obey the normal-ordering property
\begin{equation}\label{def_S}
V_+(z)V_-(w)=S(z-w-\varepsilon_+/2)^{-1}:V_+(z)V_-(w):,\quad\text{with}\quad S(z)=\dfrac{(z+\varepsilon_1)(z+\varepsilon_2)}{z(z+\varepsilon_+)}.
\end{equation}
The function $S(z)$ is related to the scattering function $g(z)$ appearing in \ref{algebra} by $g(z)=S(z)/S(-z)$, it satisfies the crossing symmetry $S(-\varepsilon_+-z)=S(z)$. Then, the horizontal representation of weight $u$ can be defined in terms of the vertex operators
\begin{equation}\label{def_eta}
\eta^\pm(z)=V_-(z\pm\varepsilon_+/4)^{\pm1}V_+(z\mp\varepsilon_+/4)^{\pm1},\quad \xi^\pm(z)=V_\pm(z\mp\varepsilon_+/2)V_\pm(z\pm\varepsilon_+/2)^{-1},
\end{equation}
it reads
\begin{equation}
\rho_u^{(H)}(x^\pm(z))=u^{\pm1}\eta^\pm(z),\quad \rho_u^{(H)}(\psi^\pm(z))=\xi^\pm(z),\quad \rho_u^{(H)}(c)=1.
\end{equation}
Note that $\xi^+(z)$ (resp. $\xi^-(z)$) is expressed only in terms of positive (negative) modes. The algebraic relations \ref{algebra} follow from the normal-ordering properties of the vertex operators given in \ref{NO_eta_xi}. Here, product of operators are ordered along the imaginary axis: ${\mathcal{O}}_1(z){\mathcal{O}}_2(w)$ implies ${\rm Im ~}(z)<{\rm Im ~}(w)$, and the parameters $\varepsilon_1$, $\varepsilon_2$ are assumed to be real. The only non-trivial relation to be checked is the commutator $[x^+,x^-]$. It follows from the identities $\xi^\pm(z)=:\eta^+(z\mp\varepsilon_+/4)\eta^-(z\pm\varepsilon_+/4):$, together with pole decomposition of the function
\begin{equation}\label{dec_S}
S(z)=1+\dfrac{\varepsilon_1\varepsilon_2}{\varepsilon_+}\left(\dfrac1z-\dfrac1{z+\varepsilon_+}\right)\quad\Rightarrow\quad [S(z)]_+-[S(z)]_-=\dfrac{\varepsilon_1\varepsilon_2}{\varepsilon_+}\left(\delta(z)-\delta(z+\varepsilon_+)\right),
\end{equation}
where $[S(z)]_\pm=S(z\mp i0)$ defines the poles prescription (see \ref{AppA}).
\subsection{Vertical representations and SHc algebra}
Vertical representations of rank $m$ and weight $\vec a=a_1,\cdots,a_m$ act on a basis of states $\dket{\vec a,\vec \lambda}$ parameterized by a set of $m$ Young diagrams $\vec\lambda=\lambda^{(1)},\cdots,\lambda^{(m)}$,
\begin{align}
\begin{split}\label{def_vert}
&\rho^{(m)}_{\vec a}(x^+(z))\dket{\vec a,\vec \lambda}=\sum_{x\in A(\vec\lambda)}\delta(z-\phi_x)\mathop{\text{Res}}_{z=\phi_x}{\mathcal{Y}}_{\vec\lambda}(z)^{-1}\dket{\vec a,\vec \lambda+x},\\
&\rho^{(m)}_{\vec a}(x^-(z))\dket{\vec a,\vec \lambda}=\sum_{x\in R(\vec\lambda)}\delta(z-\phi_x)\mathop{\text{Res}}_{z=\phi_x}{\mathcal{Y}}_{\vec\lambda}(z+\varepsilon_+)\dket{\vec a,\vec \lambda-x},\\
&\rho^{(m)}_{\vec a}(\psi^\pm(z))\dket{\vec a,\vec \lambda}=\left[\Psi_{\vec\lambda}(z)\right]_\pm\dket{\vec a,\vec \lambda}.
\end{split}
\end{align}
We have denoted $A(\vec\lambda)$ (resp. $R(\vec\lambda)$) the set of boxes that can be added to (removed from) the Young diagrams composing $\vec\lambda$. In addition, to each box $x=(l,i,j)\in\vec\lambda$ with $(i,j)\in\lambda^{(l)}$, we have associated the complex number $\phi_x=a_l+(i-1)\varepsilon_1+(j-1)\varepsilon_2$ that is sometimes called \textit{instanton position} (in the moduli space). Finally, the functions ${\mathcal{Y}}_{\vec\lambda}(z)$ and $\Psi_{\vec\lambda}(z)$ appearing in \ref{def_vert} are defined by
\begin{equation}\label{def_CY}
\Psi_{\vec\lambda}(z)=\dfrac{{\mathcal{Y}}_{\vec\lambda}(z+\varepsilon_+)}{{\mathcal{Y}}_{\vec\lambda}(z)},\quad {\mathcal{Y}}_{\vec\lambda}(z)=\dfrac{\prod_{x\in A(\vec\lambda)}z-\phi_x}{\prod_{x\in R(\vec\lambda)}z-\varepsilon_+-\phi_x}.
\end{equation}
Parameters $a_l$ and $\varepsilon_1$, $\varepsilon_2$ are chosen real, so that the poles lie on the real axis. The commutator $[x^+,x^-]$ in \ref{algebra} follows from the pole decomposition of $\Psi_{\vec\lambda}(z)$. The other relations are derived using the shell formulas
\begin{equation}
{\mathcal{Y}}_{\vec\lambda}(z)=\prod_{l=1}^m(z-a_l)\prod_{x\in\vec\lambda}S(\phi_x-z),\quad \Psi_{\vec\lambda}(z)=\prod_{l=1}^m\dfrac{z+\varepsilon_+-a_l}{z-a_l}\prod_{x\in\vec\lambda}g(z-\phi_x).
\end{equation}
In order to define the dual intertwiner, we need to introduce a dual basis $\dbra{\vec a,\vec\lambda}$ on which acts the contragredient representation $\hat\rho^{(m)}_{\vec a}$ such that
\begin{equation}
\dbra{\vec a,\vec\lambda}\left(\rho^{(m)}_{\vec a}(e)\dket{\vec a,\vec\mu}\right)=\left(\dbra{\vec a,\vec\lambda}\hat\rho^{(m)}_{\vec a}(e)\right)\dket{\vec a,\vec\mu}.
\end{equation}
The form of the contragredient representation depends on the definition of the scalar product between vertical states. Like in the case of DIM's vertical representations \cite{Bourgine2017c,Bourgine2017b}, and in contrast with our previous choice \cite{Bourgine2015c}, it is convenient to introduce a non-trivial norm for these states, namely
\begin{equation}\label{def_al}
\dbra{\vec a,\vec\lambda}\vec a,\vec\mu\rangle\!\rangle=\mathcal{Z}_\text{vect.}(\vec a,\vec\lambda)^{-1}\delta_{\vec\lambda,\vec\mu},
\end{equation}
where $\mathcal{Z}_\text{vect.}(\vec a,\vec\lambda)$ denotes the vector contribution to the ${\mathcal{N}}=2$ instanton partition function.\footnote{We remind here the formula for the contributions to the instanton partition functions of the multiplets vector $\mathcal{Z}_\text{vect.}(\vec a,\vec\lambda)$ and bifundamental (with mass $\mathfrak{m}$) $\mathcal{Z}_\text{bfd.}(\vec a,\vec\lambda;\vec b,\vec\mu|\mathfrak{m})$: $\mathcal{Z}_\text{vect.}(\vec a,\vec\lambda)=\mathcal{Z}_\text{bfd.}(\vec a,\vec\lambda;\vec a,\vec\lambda|0)^{-1}$ and
\begin{equation}
\mathcal{Z}_\text{bfd.}(\vec a,\vec\lambda;\vec b,\vec\mu|\mathfrak{m})=\prod_{x\in\vec\lambda}\prod_{l}(\phi_x-b_l+\varepsilon_+-\mathfrak{m})\times\prod_{x\in\vec\mu}\prod_{l}(\phi_x-a_l+\mathfrak{m})\times\prod_{\superp{x\in\vec\lambda}{y\in\vec\mu}}S(\phi_x-\phi_y-\mathfrak{m}).
\end{equation}} With this choice, the contragredient representation reads exactly as in \ref{def_vert}, but with the action of $x^\pm(z)$ and $-x^\mp(z)$ exchanged. At first encounter, this choice might look like an artificial way to introduce the vector contribution. However, the underlying algebraic structure is in fact very rigid, and had we chosen a trivial norm, we would have recovered the vector contribution $\mathcal{Z}_\text{vect.}(\vec a,\vec\lambda)$ in the normalization factor $t_{\vec\lambda}^\ast$ of the dual intertwiner $\Phi^\ast$ defined below. Indirectly here, we have exploited the reflection symmetry $\sigma_V$ in order to simply the expression of $\Phi^\ast$.
The connection between the vertical representation \ref{def_vert} and the action of SHc algebra on the basis $\dket{\vec a,\vec\lambda}$ follows from the decomposition $\rho^{(m)}_{\vec a}(x^\pm(z))=[X^\pm(z)]_+-[X^\pm(z)]_-$ and $\psi^\pm(z)=[\Psi(z)]_\pm$ with
\begin{align}\label{act_Xpm}
\begin{split}
&X^+(z)\dket{\vec a,\vec \lambda}=\sum_{x\in A(\vec\lambda)}\dfrac1{z-\phi_x}\mathop{\text{Res}}_{z=\phi_x}{\mathcal{Y}}_{\vec\lambda}(z)^{-1}\dket{\vec a,\vec \lambda+x},\\
&X^-(z)\dket{\vec a,\vec \lambda}=\sum_{x\in R(\vec\lambda)}\dfrac1{z-\phi_x}\mathop{\text{Res}}_{z=\phi_x}{\mathcal{Y}}_{\vec\lambda}(z+\varepsilon_+)\dket{\vec a,\vec \lambda-x},
\end{split}
\end{align}
and $\Psi(z)\dket{\vec a,\vec \lambda}=\Psi_{\vec\lambda}(z)\dket{\vec a,\vec \lambda}$. This operator can be expressed using the \textit{half-boson} (i.e. containing only commuting positive modes) $\Phi(z)$,
\begin{equation}\label{def_Phi}
\Psi(z)=\prod_{l=1}^m\dfrac{z+\varepsilon_+-a_l}{z-a_l}\prod_{\alpha=1,2,3}e^{\Phi(z+\varepsilon_\alpha)-\Phi(z-\varepsilon_\alpha)},\quad \Phi(z)=\log(z)\Phi_0-\sum_{n=1}^\infty\dfrac1{nz^n}\Phi_{n},
\end{equation}
which acts diagonally on the states $\dket{\vec v,\vec\lambda}$,
\begin{equation}\label{act_Phi}
\Phi_n\dket{\vec v,\vec\lambda}=\left(\sum_{x\in\vec\lambda}\phi_x^n\right)\dket{\vec v,\vec\lambda},\quad \partial_z\Phi(z)\dket{\vec v,\vec\lambda}=\left(\sum_{x\in\vec\lambda}\dfrac1{z-\phi_x}\right)\dket{\vec v,\vec\lambda}.
\end{equation}
The new operators $X^\pm(z)$, $\Psi(z)$ and $\partial_z\Phi(z)$ obey the algebraic relations
\begin{equation}
[X^+(z),X^-(w)]=\dfrac{\varepsilon_1\varepsilon_2}{\varepsilon_+}\dfrac{\Psi(z)-\Psi(w)}{z-w},\quad [\partial_z\Phi(z),X^\pm(w)]=\pm\dfrac1{z-w}X^\pm(w),
\end{equation}
leading to identify them with the SHc currents $\sqrt{-\varepsilon_1\varepsilon_2}D_{\pm1}(z)$, $E(z)$ and $D_0(z)$ (respectively) in the holomorphic presentation of reference \cite{Bourgine2015c}. Furthermore, the actions \ref{act_Xpm} and \ref{act_Phi} on the states $\dket{\vec v,\vec\lambda}$ coincide with the one presented in \cite{Bourgine2015c}, up to a change of states normalization.\footnote{Alternatively, it is possible to expand the currents at infinity, thus defining the modes
\begin{equation}
X^\pm(z)=\sum_{n=0}^\infty z^{-n-1}X_n^\pm,\quad \Psi(z)=1+\varepsilon_+\sum_{n=0}^\infty z^{-n-1}\Psi_n,
\end{equation}
and identify $X_n^\pm$, $\Psi_n$ and $\Phi_n$ with, respectively, the representation of the SHc generators $\sqrt{-\varepsilon_1\varepsilon_2}D_{\pm1,n}$, $E_n$ and $D_{0,n+1}$. The latter satisfy the algebraic relations $[D_{0,n},D_{0,m}]=0$, $[D_{0,n+1},D_{\pm1,m}]=\pm D_{\pm1,n+m}$, $[D_{-1,n},D_{1,m}]=E_{n+m}$ and a certain exponential relation between $E_n$ and $D_{0,n}$ deriving from \ref{def_Phi}.}
\section{Reconstructing the gauge theories' BPS quantities}
\subsection{Intertwiners}
In the algebraic engineering of 5D ${\mathcal{N}}=1$ gauge theories, the role of the topological vertex is devoted to the intertwiners of the DIM algebra, first introduced by Awata, Feigin and Shiraishi (AFS) in \cite{Awata2011}, and further generalized in \cite{Bourgine2017b,Bourgine2018a}. In the 4D case, the intertwiners, denoted $\Phi^{(m)}[u,\vec a]$ and $\Phi^{\ast(m)}[u,\vec a]$, are obtained as solution to the following equations,
\begin{align}
\begin{split}\label{AFS_lemmas}
&\rho_{u'}^{(H)}(e)\Phi^{(m)}[u,\vec a]=\Phi^{(m)}[u,\vec a]\ \left(\rho_{\vec a}^{(m)}\otimes\rho_{u}^{(H)}\ \Delta(e)\right),\\
&\left(\rho_{\vec a}^{(m)}\otimes\rho_{u}^{(H)}\ \Delta'(e)\right)\ \Phi^{\ast(m)}[u,\vec a]=\Phi^{\ast(m)}[u,\vec a]\rho_{u'}^{(H)}(e),
\end{split}
\end{align}
where $e$ is any of the four currents defining the algebra \ref{algebra}, $\Delta$ is the coproduct given in \ref{def_D}, and $\Delta'$ the opposite coproduct obtained by permutation. Note that the charge $c$ is conserved in these relations. For each equation, the solution can be expanded on the vertical basis, their components are operators acting on the horizontal Fock space:
\begin{align}
\begin{split}\label{sol_AFS}
&\Phi^{(m)}[u,\vec a]=\sum_{\vec\lambda}\mathcal{Z}_\text{vect.}(\vec a,\vec\lambda)\ \Phi^{(m)}_{\vec\lambda}[u,\vec a]\ \dbra{\vec a,\vec\lambda},\quad \Phi^{(m)}_{\vec\lambda}[u,\vec a]=t^{(m)}_{\vec\lambda}[u,\vec a]\ :\prod_{l=1}^m\Phi_{\emptyset}(a_l)\prod_{x\in\vec\lambda}\eta^+(\phi_x):,\\
&\Phi^{\ast(m)}[u,\vec a]=\sum_{\vec\lambda}\mathcal{Z}_\text{vect.}(\vec a,\vec\lambda)\ \Phi^{\ast(m)}_{\vec\lambda}[u,\vec a]\ \dket{\vec a,\vec\lambda},\quad \Phi^{\ast(m)}_{\vec\lambda}[u,\vec a]=t^{\ast(m)}_{\vec\lambda}[u,\vec a]\ :\prod_{l=1}^m\Phi_{\emptyset}^{\ast}(a_l)\prod_{x\in\vec\lambda}\eta^-(\phi_x):,
\end{split}
\end{align}
where $t^{(m)}_{\vec\lambda}[u,\vec a]=(-1)^{m|\vec\lambda|}u^{|\vec\lambda|}$ and $t^{\ast(m)}_{\vec\lambda}[u,\vec a]=u^{-|\vec\lambda|}$ are normalization coefficients, and $|\vec\lambda|$ denotes the total number of boxes contained in $\vec\lambda$. In addition, the horizontal weights must also obey the constraint $u'=(-1)^mu$. The intertwiners \ref{sol_AFS} taken at $m=1$ define a degenerate version of the refined topological vertex, in the AFS free field presentation \cite{Awata2011}, that can be employed to construct directly 4D ${\mathcal{N}}=2$ gauge theories.
The definition of the vacuum components $\Phi_\emptyset(a)$ and $\Phi_\emptyset^\ast(a)$ in \ref{sol_AFS} requires some explanation. Physically, these operators define a sort of Fermi sea for which $\eta^\pm$ creates excitations. Thus, and by analogy with AFS intertwiners, we would like to define them as a (normal-ordered) infinite product of the operators $\eta^\pm(\phi_x)^{-1}$ over the boxes of a fully filled Young diagram, i.e. $\phi_x=a+(i-1)\varepsilon_1+(j-1)\varepsilon_2$ with $i,j=1\cdots\infty$. However, in contrast with the DIM scenario, it is not possible to express the resulting operators as vertex operators built upon the modes $\alpha_n$. We propose two different approaches to deal with this issue. The first one is to cut-off the size of the infinite Young diagram, effectively regularizing the double product to $i,j=1\cdots N$, and sending $N$ to infinity at the end of the calculation. This approach is a little cumbersome as it introduces extra factors that are eventually canceled. Alternatively, it is possible to introduce an auxiliary free boson $\tilde{\varphi}(z)$ such that $\varphi_-(z)=\tilde{\varphi}_-(z-\varepsilon_1/2)-\tilde{\varphi}_-(z+\varepsilon_1/2)$ and $\varphi_+(z)=\tilde{\varphi}_+(z+\varepsilon_2/2)-\tilde{\varphi}_+(z-\varepsilon_2/2)$. Assuming that the propagator of the boson $\tilde{\varphi}(z)$ is given by (minus) the log of a function $F(z)$ (conveniently shifted here by $+\varepsilon_+/2$), the relations between the positive/negative modes of the bosons $\varphi$ and $\tilde{\varphi}$ imply a certain functional relation for the propagator:\footnote{Note however, that the asymptotic expansion at $z=\infty$ of the function $\Gamma_2(z)$ (see, for instance, reference \cite{Spreafico}),
\begin{equation}\label{exp_G2}
\log\Gamma_2(z)\sim-\dfrac{z^2}{2\varepsilon_1\varepsilon_2}\left(\log(z)-\dfrac32\right)+\dfrac{\varepsilon_+}{2\varepsilon_1\varepsilon_2}z\left(\log(z)-1\right)+O(\log(z))
\end{equation}
is not compatible with the usual mode expansion \ref{def_vphi} of a free boson. To resolve this, it is possible to modify the expression of the propagator, defining $F(z)=\theta_2(z)\Gamma_2(z)$ where $\theta_2(z)$ obeys the same functional equation as $\Gamma_2(z)$ but with $1$ instead of $z$ in the RHS, and possesses an asymptotic expansion that cancels the unwanted terms in \ref{exp_G2}. However, it turns out that the one-loop determinant contributing to the partition function \cite{Losev2003} is usually regularized using the function $\Gamma_2(z)$ \cite{Bourgine2017}. Another possibility is to add extra zero-modes to the boson $\tilde{\varphi}(z)$ to generate the terms $z^2\log z$, $z\log z$,...}
\begin{equation}
e^{\tilde{\varphi}(z)}e^{\tilde{\varphi}(w)}=F(z-w+\varepsilon_+/2)^{-1}:e^{\tilde{\varphi}(z)}e^{\tilde{\varphi}(w)}:,\quad \dfrac{F(z+\varepsilon_1)F(z+\varepsilon_2)}{F(z)F(z+\varepsilon_+)}=z.
\end{equation}
This relation can be solved using the double-gamma function, $F(z)=\Gamma_2(z)$ with pseudo-periods $\varepsilon_1$ and $\varepsilon_2$ (see \cite{Spreafico,Bourgine2017}). Defining
\begin{equation}\label{def_Phi_vac}
\Phi_\emptyset(a)=e^{-\tilde{\varphi}_-(a-\varepsilon_+/4)}e^{-\tilde{\varphi}_+(a-3\varepsilon_+/4)},\quad \Phi_\emptyset^{\ast}(a)=e^{\tilde{\varphi}_-(a-3\varepsilon_+/4)}e^{\tilde{\varphi}_+(a-\varepsilon_+/4)},
\end{equation}
all the necessary normal-ordering relation can now be derived, they have been gathered in appendix \ref{NO_Phi_vac}. Adapting the method described in the appendix C of \cite{Bourgine2017b}, they can be used to show that the expressions given in \ref{sol_AFS} indeed solve the equations \ref{AFS_lemmas}.
\subsection{Instanton partition functions and qq-characters}
From this point, we can repeat the construction presented in \cite{Bourgine2017b} for the instanton partition functions and qq-characters of linear quivers gauge theory. For the purpose of illustration, we present here the main results concerning ($A_1$ quiver) pure $U(m)$ gauge theories. The corresponding ${\mathcal{T}}$-operators are obtained by coupling an intertwiner $\Phi^{(m)}[u,\vec a]$ to its dual $\Phi^{\ast(m)}[u^\ast,\vec a]$ through a scalar product $\cdot$ in their common vertical representation,
\begin{equation}\label{def_T_A1}
{\mathcal{T}}_{U(m)}=\Phi^{(m)}[u,\vec a]\cdot\Phi^{\ast(m)}[u^\ast,\vec a]=\sum_{\vec\lambda}\mathcal{Z}_\text{vect.}(\vec a,\vec\lambda)\ \Phi^{(m)}_{\vec\lambda}[u,\vec a]\otimes \Phi^{\ast(m)}_{\vec\lambda}[u^\ast,\vec a].
\end{equation}
This operator acts on the tensor product of two bosonic Fock spaces, it is easy to check that its vev reproduces the instanton partition function of the gauge theory,
\begin{equation}
\mathcal{Z}_\text{inst.}[U(m)]=\left(\bra{\emptyset}\otimes\bra{\emptyset}\right){\mathcal{T}}_{U(m)}\left(\ket{\emptyset}\otimes\ket{\emptyset}\right)=\sum_{\vec\lambda}\mathfrak{q}^{|\vec\lambda|}\mathcal{Z}_\text{vect.}(\vec a,\vec\lambda),
\end{equation}
under the identification of $\vec a$ with the Coulomb branch vevs, and $\mathfrak{q}=(-1)^m u/u^\ast$ with the exponentiated gauge coupling. The partition function can also be recovered as a scalar product of Gaiotto states, corresponding here to the vevs of the intertwiners:
\begin{align}
\begin{split}
&\dket{G,\vec a}=\left(1\otimes\bra{\emptyset}\right)\Phi^{\ast(m)}[u^\ast,\vec a]\ket{\emptyset}=\sum_{\vec\lambda}\mathcal{Z}_\text{vect.}(\vec a,\vec\lambda)\ \bra{\emptyset}\Phi^{\ast(m)}_{\vec\lambda}[u^\ast,\vec a]\ket{\emptyset}\ \dket{\vec a,\vec\lambda},\\
&\dbra{G,\vec a}=\bra{\emptyset}\Phi^{(m)}[u,\vec a]\left(1\otimes\ket{\emptyset}\right)=\sum_{\vec\lambda}\mathcal{Z}_\text{vect.}(\vec a,\vec\lambda)\ \bra{\emptyset}\Phi^{(m)}_{\vec\lambda}[u,\vec a]\ket{\emptyset}\ \dbra{\vec a,\vec\lambda},
\end{split}
\end{align}
so that $\left\langle{\mathcal{T}}_{U(m)}\right\rangle=\dbra{G,\vec a}\!G,\vec a\rangle\!\rangle$. In fact, these Gaiotto states coincide with those defined in \cite{Bourgine2015c} if we take into account the different normalization of the basis. The action of SHc currents $X^\pm(z)$ on these states has been derived in \cite{Bourgine2015c}. It can be recovered here using the intertwining property \ref{AFS_lemmas}. For instance,
\begin{align}
\begin{split}
&\rho_{\vec a}^{(m)}(x^+(z))\dket{G,\vec a}=-u^\ast\left([{\mathcal{Y}}(z+\varepsilon_+)_+-[{\mathcal{Y}}(z+\varepsilon_+)]_-\right)\dket{G,\vec a},\\
&\rho_{\vec a}^{(m)}(x^-(z))\dket{G,\vec a}=-(u^\ast)^{-1}\left([{\mathcal{Y}}(z)^{-1}]_+-[{\mathcal{Y}}(z)^{-1}]_-\right)\dket{G,\vec a},
\end{split}
\end{align}
where ${\mathcal{Y}}(z)$ is a diagonal operator in the basis $\dket{\vec a,\vec\lambda}$ with eigenvalues ${\mathcal{Y}}_{\vec\lambda}(z)$.
By construction, the operator ${\mathcal{T}}_{U(m)}$ commutes with the action of any element $e$ of the algebra \ref{algebra} in the appropriate representation,
\begin{equation}\label{covar_T}
\left(\rho_{u'}^{(H)}\otimes\rho_{u^\ast}^{(H)}\ \Delta'(e)\right){\mathcal{T}}_{U(m)}={\mathcal{T}}_{U(m)}\left(\rho_{u}^{(H)}\otimes\rho_{u^{\ast\prime}}^{(H)}\ \Delta'(e)\right).
\end{equation}
The fundamental qq-character can be obtained by insertion of the current $x^+(z)$ before taking the vev,\footnote{Note that, as consequence of the invariance of the brane web under the reflection encoded by $\sigma_V$, we find the same quantity if we replace $x^+(z)$ by $x^-(z)$ (and $(u^\ast)^{-1}$ by $u'$) in this formula.}
\begin{equation}\label{def_chi}
\chi(z-\varepsilon_+/2)=(u^\ast)^{-1}\dfrac{\left(\bra{\emptyset}\otimes\bra{\emptyset}\right)\left(\rho_{u'}^{(H)}\otimes\rho_{u^\ast}^{(H)}\ \Delta'(x^+(z))\right){\mathcal{T}}_{U(m)}\left(\ket{\emptyset}\otimes\ket{\emptyset}\right)}{\left(\bra{\emptyset}\otimes\bra{\emptyset}\right){\mathcal{T}}_{U(m)}\left(\ket{\emptyset}\otimes\ket{\emptyset}\right)}.
\end{equation}
After evaluation of the vev using the normal-ordering relations \ref{NO_Intw} of \ref{AppB}, we recover the well-known expression
\begin{equation}\label{expr_chi}
\chi(z)=\mathcal{Z}_\text{inst.}[U(m)]^{-1}\sum_{\vec\lambda}\mathfrak{q}^{|\vec\lambda|}\mathcal{Z}_\text{vect.}(\vec a,\vec\lambda)\left({\mathcal{Y}}_{\vec\lambda}(z+\varepsilon_+)+\dfrac{\mathfrak{q}}{{\mathcal{Y}}_{\vec\lambda}(z)}\right).
\end{equation}
The covariance property \ref{covar_T} of the operator ${\mathcal{T}}_{U(m)}$ implies the cancellation of poles between the two terms in \ref{expr_chi}, and so $\chi(z)$ is a polynomial of degree $m$. Higher qq-characters are constructed by insertion of multiple operators $x^+(z_i)$ in the vevs.
These results can be extended to linear quivers of higher rank. The corresponding ${\mathcal{T}}$-operator is obtained by coupling several of the operators ${\mathcal{T}}_{U(m)}$ defined in \ref{def_T_A1} in the horizontal channel. Bifundamental contributions $\mathcal{Z}_\text{bfd.}$ to the partition function appear as the product of intertwiners are normal-ordered following the two relations
\begin{align}\small
\begin{split}
&\Phi^{(m)}_{\vec\lambda}[u,\vec a]\Phi^{\ast(m^\ast)}_{\vec\lambda^\ast}[u^\ast,\vec a^\ast]=(-1)^{m|\vec\lambda^\ast|}\mathcal{Z}_\text{bfd.}(\vec a,\vec\lambda;\vec a^\ast,\vec\lambda^\ast|\varepsilon_+/2)\prod_{l,l^\ast=1}^{m,m^\ast}F(a_l-a_{l^\ast}^\ast+\varepsilon_+/2)\ :\Phi^{(m)}_{\vec\lambda}[u,\vec a]\Phi^{\ast(m^\ast)}_{\vec\lambda^\ast}[u^\ast,\vec a^\ast]:,\\
&\Phi^{\ast(m^\ast)}_{\vec\lambda^\ast}[u^\ast,\vec a^\ast]\Phi^{(m)}_{\vec\lambda}[u,\vec a]=(-1)^{m^\ast|\vec\lambda|}\mathcal{Z}_\text{bfd.}(\vec a,\vec\lambda;\vec a^\ast,\vec\lambda^\ast|\varepsilon_+/2)\prod_{l,l^\ast=1}^{m,m^\ast}F(a_{l^\ast}^\ast-a_l+\varepsilon_+/2)\ :\Phi^{(m)}_{\vec\lambda}[u,\vec a]\Phi^{\ast(m^\ast)}_{\vec\lambda^\ast}[u^\ast,\vec a^\ast]:.
\end{split}
\end{align}
In these formulas, the $F$-dependent factor can be interpreted as one-loop bifundamental contributions with the $\Gamma_2$-regularization of infinite products. The fundamental qq-characters attached to the two nodes at the extremity of the quiver are obtained by insertion of $x^\pm(z)$ (with repeated coproduct). On the other hand, qq-characters corresponding to middle nodes require the introduction of more involved operators \cite{Bourgine2017b}.
\subsection{Degenerate limit of Kimura-Pestun's quiver W-algebra}
Instanton partition functions of 5D ${\mathcal{N}}=1$ quiver gauge theories are covariant under the action of two distinct q-deformed W-algebras. The first one is involved in the q-deformed AGT correspondence \cite{Awata2009}, it appears at each individual node of the quiver. For instance, the partition function of a linear quiver theory bearing a $U(m)$ gauge group at each node coincides with a conformal block of the q-W$_m$-algebra. A second form of q-W-algebra covariance has appeared in the work of Kimura and Pestun (KP) \cite{Kimura2015}. In their construction, the Dynkin diagram of the algebra is identified with the quiver of the gauge theory. Thus, the partition function of a linear quiver with $k$ nodes is covariant under the action of a q-W$_{k+1}$-algebra. The DIM algebra offers a third perspective which is somehow more elementary as it allows us to understand the relation between these two q-W-algebras (at least in the case of linear quivers). For instance, each node of a quiver bearing a $U(m)$ gauge group corresponds to a stack of $m$ D5-branes in the $(p,q)$-brane construction. Accordingly, the node carries a vertical representation $(0,m)$ that contains the q-$W_m$ action involved in the q-AGT correspondence. On the other hand, the DIM algebra acts on the ${\mathcal{T}}$-operators in a different representation, obtained as a tensor product of horizontal representations $(1,n)$ attached to the external NS5-branes (possibly dressed by D5s). This tensored horizontal representation factorizes into a q-Heisenberg contribution, and the action of KP's q-W-algebra \cite{Mironov2016,Bourgine2017b}. Furthermore, the automorphism $\mathcal{S}$ discovered by Miki \cite{Miki2007} maps vertical to horizontal representations (and vice-versa), effectively implementing a rotation of the $(p,q)$-brane web, and thereby exchanging the role of the two different q-W-algebras. We refer to the upcoming paper \cite{Bourgine2018a} for a deeper discussion of this automorphism in the formalism of algebraic engineering.
In this section, we would like to define the degenerate version of KP's quiver W-algebra that acts on partition functions of 4D ${\mathcal{N}}=2$ gauge theories. We will restrict ourselves to the case of a single node quiver with $U(m)$ gauge group. For the 5D version of the gauge theory, the relevant q-W-algebra is the q-Virasoro algebra, constructed in \cite{Shiraishi1995} as a q-deformation of the usual Virasoro algebra. In the DIM description, the representation acting on the ${\mathcal{T}}$-operator has levels $(2,0)$, it is obtained by coproduct of the two representations $(1,n)\otimes(1,-n)$ associated to the two (dressed) NS5-branes. It can be decomposed as a tensor product of KP's q-Virasoro algebra and a q-Heisenberg algebra \cite{Mironov2016,Bourgine2017b}. The latter encodes the action of the currents $\bar\psi^\pm(Z)$, it can be factorized out of the action of $\bar x^\pm(Z)$, thereby providing the action of q-Virasoro on the ${\mathcal{T}}$-operator. Here, we would like to repeat this construction for the degenerate DIM algebra \ref{algebra}.
The first step is to isolate the q-Heisenberg action. For this purpose we need to introduce an auxiliary bosonic field $\varphi_W(z)$, with propagator $\log h(z)$, such that the (half) vertex operators $V_\pm(z)$ are decomposed into products of $W_\pm(z)=e^{\varphi_{W\pm}(z)}$:\footnote{In terms of bosonic fields, we have the relations $\varphi_{W+}(z+\varepsilon_+/2)+\varphi_{W+}(z-\varepsilon_+/2)=\varphi_+(z+\varepsilon_1/2)-\varphi_+(z-\varepsilon_1/2)$ and $\varphi_{W-}(z+\varepsilon_+/2)+\varphi_{W-}(z-\varepsilon_+/2)=\varphi_-(z-\varepsilon_2/2)-\varphi_-(z+\varepsilon_2/2)$ for positive/negative modes.}
\begin{equation}\label{dec_V_W}
V_\pm(z)=W_\pm(z+\varepsilon_+/2)W_\pm(z-\varepsilon_+/2),\quad W_+(z)W_-(w)=h(z-w):W_+(z)W_-(w):.
\end{equation}
It turns out convenient to introduce also the function $f(z)=h(z-\varepsilon_+/2)^{-1}h(z+\varepsilon_+/2)^{-1}$. Since the decomposition \ref{dec_V_W} has to reproduce the normal-ordering property \ref{def_S} for the vertex operators $V_\pm(z)$, we find the functional relation $f(z)f(z+\varepsilon_+)=S(z)$ constraining $f(z)$ in terms of the function $S(z)$. The exact formula for the function $f(z)$ will be specified later. Using the decomposition \ref{dec_V_W}, it is easy to observe that the representation $\rho^{(H)}\otimes\rho^{(H)}$ of the currents $x^\pm(z)$, $\psi^\pm(z)$ can be written as\footnote{It is possible to work with either coproduct $\Delta$ or $\Delta'$ by simply permuting the two Fock spaces. Here we chose $\Delta'$ in agreement with the previous section.}
\begin{align}\label{dec_UT}
\begin{split}
&\rho_{u_1}^{(H)}\otimes\rho_{u_2}^{(H)}\ \Delta'(x^\pm(z))=\sqrt{u_1u_2}^{\pm1}U_-(z\pm\varepsilon_+/2)^{\pm1}T(z)U_+(z\mp\varepsilon_+/2)^{\pm1},\\ &\rho_{u_1}^{(H)}\otimes\rho_{u_2}^{(H)}\ \Delta'(\psi^\pm(z))=U_\pm(z\mp\varepsilon_+)U_\pm(z\pm\varepsilon_+)^{-1},
\end{split}
\end{align}
where we have introduced the following quantities:
\begin{align}
\begin{split}
&U_\pm(z)=W_\pm(z\pm\varepsilon_+/4)\otimes W_\pm(z\mp\varepsilon_+/4),\quad T(z)=R(z)+:R(z-\varepsilon_+)^{-1}:,\\
&R(z)=\sqrt{\dfrac{u_2}{u_1}}W_-(z+\varepsilon_+/4)^{-1}W_+(z-\varepsilon_+/4)^{-1}\otimes W_-(z-\varepsilon_+/4)W_+(z+\varepsilon_+/4).
\end{split}
\end{align}
It is possible to refine the decomposition \ref{dec_UT}, observing that the vertex operators $U_\pm(z)$ and $R(w)$ commute as a result of the cancellation between the factors obtained in each channel when normal-ordering. Commuting the factors of $U_\pm(z)$ with $T(z)$, we can write the representation of the currents $x^\pm(z)$ and $\psi^\pm(z)$ in the form
\begin{equation}
\rho_{u_1}^{(H)}\otimes\rho_{u_2}^{(H)}\ \Delta'(x^\pm(z))=V_B^\pm(z)T(z),\quad \rho_{u_1}^{(H)}\otimes\rho_{u_2}^{(H)}\ \Delta'(\psi^\pm(z))=:V_B^+(z\mp\varepsilon_+/2)V_B^-(z\pm\varepsilon_+/2):,
\end{equation}
where $V_B^\pm(z)=\sqrt{u_1u_2}^{\pm1}U_-(z\pm\varepsilon_+/2)^{\pm1}U_+(z\mp\varepsilon_+/2)^{\pm1}$ commutes with $T(w)$. As in the case of the DIM algebra, we observe that the action of the currents $\psi^\pm(z)$ is given only in terms of the vertex operators $V_B^\pm(z)$. These operators can be factorized out of the action of $x^\pm(z)$, leaving only the operator $T(z)$. The latter obeys the following weighted
commutation relation,\footnote{This relation is obtained by first computing the normal-ordering
\begin{align}
\begin{split}
T(z)T(w)&=f(z-w)^{-1}\left[:R(z)R(w):+:R(z-\varepsilon_+)^{-1}R(w-\varepsilon_+)^{-1}:\right]+f(z-w+\varepsilon_+):R(z)R(w-\varepsilon_+)^{-1}:\\
&+f(z-w-\varepsilon_+):R(z-\varepsilon_+)^{-1}R(w):,
\end{split}
\end{align}
using \ref{dec_V_W}. Multiplying this relation with $f(z-w)$, the function $S(z-w)$ appears in the RHS. In order to recover \ref{rel_T}, we take the difference between the two normal-ordering relations with $z$ and $w$ exchanged. The first terms cancel, while the remaining ones produce the $\delta$-function through the pole decomposition \ref{dec_S} of the function $S(z)$.}
\begin{equation}\label{rel_T}
f(z-w)T(z)T(w)-f(w-z)T(w)T(z)=-\dfrac{\varepsilon_1\varepsilon_2}{\varepsilon_+}\left(\delta(z-w+\varepsilon_+)-\delta(z-w-\varepsilon_+)\right),
\end{equation}
that will be identified with a certain degenerate limit of the q-Virasoro algebra. We claim that this algebra is the degenerate version of KP's quiver W-algebra in the case of a single node. Repeating this procedure for an arbitrary quiver, it should be possible to formulate the general degenerate KP algebras.
The q-Virasoro algebra is defined in \cite{Shiraishi1995} in terms of its \textit{stress-energy tensor} ${\bar T}(Z)$, a current that obeys the algebraic relation
\begin{equation}\label{rel_bT}
{\bar f}(W/Z){\bar T}(Z){\bar T}(W)-{\bar f}(Z/W){\bar T}(W){\bar T}(Z)=\dfrac{(1-q_1)(1-q_2)}{(1-q_1q_2)}\left({\bar\delta}(Z/q_3W)-{\bar\delta}(q_3Z/W)\right),
\end{equation}
where $q_1=q$ and $q_2=t^{-1}$ are two complex parameters, $q_3=q_1^{-1}q_2^{-1}$, ${\bar\delta}(z)$ is the multiplicative delta function \ref{def_bd}, and the function ${\bar f}(Z)$ writes (the second equality is valid when $|q_3|<1$):
\begin{equation}
{\bar f}(Z)=\exp\left(\sum_{n=1}^\infty\dfrac1n\dfrac{(1-q_1^n)(1-q_2^n)}{(1+q_3^{-n})}Z^n\right)=\dfrac1{1-Z}\prod_{n=0}^\infty\dfrac{(1-q_1^{-1}q_3^{2n}Z)(1-q_2^{-1}q_3^{2n}Z)}{(1-q_1^{-1}q_3^{2n+1}Z)(1-q_2^{-1}q_3^{2n+1}Z)}.
\end{equation}
This function ${\bar f}(Z)$ satisfies the functional relation
\begin{equation}\label{fr_bf}
{\bar f}(Z){\bar f}(q_3^{-1}Z)=\bar S(Z),\quad \text{with}\quad \bar S(Z)=\dfrac{(1-q_1Z)(1-q_2Z)}{(1-Z)(1-q_1q_2Z)}.
\end{equation}
In order to recover the relation \ref{rel_T} for \ref{rel_bT}, we need to take a degenerate limit. Just like in the case of DIM (see \ref{AppA}), we set $q_\alpha=e^{R\varepsilon_\alpha}$, $Z=e^{Rz}$, $W=e^{Rw}$, and send $R\to0$. Then, $\bar S(Z)$ tends to $S(z)$ and the functional relation \ref{fr_bf} reproduces the one satisfied by the function $f(z)$ (which is in fact also satisfied by $f(-z)$ due to the crossing symmetry $S(-\varepsilon_+-z)=S(z)$). Since this functional equation is the only constraint for the (entire) function $f(z)$, we choose to identify $f(z)$ with the limit of ${\bar f}(Z^{-1})$:
\begin{equation}
f(z)=\lim_{R\to0}{\bar f}(e^{-Rz})=\dfrac{z-\varepsilon_2}{z}\prod_{n=0}^\infty\dfrac{(z+\varepsilon_2+2n\varepsilon_+)(z-\varepsilon_2+(2n+1)\varepsilon_+)}{(z-\varepsilon_2+2n\varepsilon_+)(z+\varepsilon_2+(2n+1)\varepsilon_+)}.
\end{equation}
Applying the limiting procedure of \ref{AppA} to treat the delta functions in the RHS of \ref{rel_bT}, it is easy to show that the limit $T(z)$ of ${\bar T}(Z)$ obeys \ref{rel_T}. Note that this degenerate limit is different from the usual one that sends q-Virasoro to Virasoro, and Macdonald to Jack polynomials.\footnote{Presumably, in our limit the variables of the polynomial $X_i=e^{R x_i}$ are also sent to one, so that the Macdonald operator does not reduce to the Calogero-Sutherland Hamiltonian, but instead produces a difference operator
\begin{equation}
D_{q,t}=\sum_i\prod_{j\neq i}\dfrac{q_2^{-1}X_i-X_j}{X_i-X_j}q_1^{X_i\partial_{X_i}}\to \sum_i\prod_{j\neq i}\dfrac{x_i-x_j-\varepsilon_2}{x_i-x_j}e^{\varepsilon_1\partial_{x_i}}.
\end{equation}}
This particular limit of the q-Virasoro algebra has already been encountered in the study of dualities for 3D $T[U(N)]$ theories, it provides an additive analogue of the q-Virasoro algebra called \textit{d-Virasoro} \cite{Nedelin2017}. Screening currents, vertex operators and Dotsenko-Fateev representations of conformal blocks were also defined in this paper. Furthermore, a correspondence with conformal blocks of ordinary Virasoro algebra was also observed, it suggests the presence of an analogue of Miki's automorphism in the degenerate case. Indeed, in the self-dual pure $U(2)$ gauge theory, the exchange of vertical/horizontal representations induced by the automorphism should reduce to an exchange of the role played by the degenerate KP algebra (d-Virasoro) and the ordinary Virasoro. We hope to come back to this problem in a near future.
Interestingly, the d-Virasoro algebra is also related to the 2D (massive) sine-Gordon model describing the scaling limit of the XYZ model.\footnote{This connection is a priori unrelated with the one observed in \cite{Mironov2009} between pure $SU(2)$ gauge theory and the 1D sine-Gordon model.} Indeed, in the short note \cite{Lukyanov1995}, Lukyanov noticed that the ratio\footnote{We use here the infinite product formula GR.1.431.1 and GR.1431.3 of \cite{Gradshteyn1996} for the trigonometric functions,
\begin{equation}
\sin z=z\prod_{k=1}^\infty\left(1-\dfrac{z^2}{k^2\pi^2}\right),\quad \cos z=\prod_{k=0}^\infty\left(1-\dfrac{4z^2}{(2k+1)^2\pi^2}\right).
\end{equation}}
\begin{equation}
\dfrac{f(z)}{f(-z)}=\dfrac{\tan\left(\pi(z+\varepsilon_2)/2\varepsilon_+\right)}{\tan\left(\pi(z-\varepsilon_2)/2\varepsilon_+\right)}
\end{equation}
reproduces the scattering factor of the sine-Gordon model
\begin{equation}
\dfrac{\sinh\beta+i\sin\pi\xi}{\sinh\beta-i\sin\pi\xi}
\end{equation}
setting $\beta=i\pi z/\varepsilon_+$, and $\xi=-\varepsilon_2/\varepsilon_+$ (note that the ordering ${\rm Im ~} z_1<{\rm Im ~} z_2$ becomes ${\rm Re ~}\beta_1>{\rm Re ~}\beta_2$). This observation, together with the existence of simple poles at $\beta=\pm i\pi$ (corresponding to the delta functions in \ref{rel_T}), led him to identify $T(z)$ with the generator of the corresponding Zamolodchikov-Faddeev (ZF) algebra \cite{Zamolodchikov1978,Faddeev1980}. In this context, the operator $T(\beta)$ creates a breather with rapidity $\beta$, the basic particle of the theory.\footnote{See also \cite{Fioravanti2005,*Fioravanti2006} for the derivation of the q-Virasoro algebra as the ZF algebra of the first breather in the scaling limit of the spin 1/2 XYZ model using Bethe Ansatz techniques.}
\section{Perspectives}
We have presented here an algebraic construction for the BPS quantities of 4D ${\mathcal{N}}=2$ SYM theories based on a degenerate version of the DIM algebra. The two essential ingredients are the action of the SHc algebra (vertical representation), and an algebra of twisted vertex operators (horizontal representation). It would be interesting to understand how the latter fits into the formalism of Maulik-Okounkov \cite{Maulik2012}, and to relate the intertwiners to Smirnov's instanton R-matrix \cite{Smirnov2013}. For linear quivers, the ${\mathcal{T}}$-operator constructed here is expected to be the Baxter T-operator of a 1D quantum integrable system, with two spectral parameters associated to vertical and horizontal weights, that produces the set of commuting Hamiltonians by expansion. Given the rich structure of this algebra, there is little doubt that the underlying statistical model will reveal itself of great interest.
Following the program initiated in \cite{Bourgine2014}, it should be possible to take the Nekrasov-Shatashvili limit $\varepsilon_2\to0$ of the algebra \ref{algebra}. In this way, one hopes to recover the quantum algebra behind the integrable system of the Bethe/gauge correspondence \cite{Nekrasov2009}. This construction should provide us with a proper interpretation for the $\varepsilon_2$-deformation of quantum integrable systems, and reproduce the subleading order terms in $\varepsilon_2$ derived in \cite{Bourgine2015,Bourgine2015b} by expansion of the partition function.
Finally, the second main result of this note is the identification of the degenerate version of KP's quiver W-algebra. In the $A_1$ case, we observed an unsuspected connection with the ZF algebra of the sine-Gordon model. This is particularly intriguing given the similarity between the algebraic engineering and the method developed by Lukyanov to compute form factors of massive integrable models \cite{Lukyanov1993}. Indeed, the vertical representation is reminiscent of the ZF algebra's representation $\pi_A$ associated to asymptotic states, while both horizontal representation and Lukyanov's $\pi_Z$ are built upon free fields. It appears essential to understand better the connection between the two methods. In this perspective, one could try to extend our algebraic construction to more general scattering functions $g(z)$. We hope to come back to this issue in a future publication.
\section*{Acknowledgments}
JEB would like to thank Davide Fioravanti and Vincent Pasquier for discussions, and the IPhT CEA-Saclay, DESY, and the University of Bologna for hospitality at various stages of this project. KZ(Hong Zhang) thanks Wei Li for guidance, and Shuichi Yokoyama for discussions. The work of KZ is partially supported by the General Financial Grant from the China Postdoctoral Science Foundation, with Grant No.
2017M611009.
|
\section{Introduction}
Jumping numbers measure the complexity of the singularities of a closed subscheme
of a variety. They are defined in terms of multiplier ideals of the subscheme. Multiplier
ideals form a nested sequence of ideals parametrized by rational numbers. The values
of the parameter where the multiplier ideal changes are called jumping numbers. For a
simple complete ideal in a local ring of a closed point on a smooth surface an explicit
formula has been provided by J\"arvilehto in~\cite{J}, which is based on the dissertation ~\cite{Jj}.
This result applies also to jumping numbers of an analytically irreducible plane curve.
The purpose of this article is to generalize this formula to any complete ideal.
Besides~\cite{J}, jumping numbers of simple complete ideals or analytically irreducible
plane curves have been independently investigated by several people (see~\cite{N}, \cite{T0} and~\cite{GM}).
In a local ring at a rational singularity of a complex surface, Tucker presented in~\cite{T}
an algorithm to compute the set of jumping numbers of any ideal. Recently, Alberich-Carrami\~nana,
Montaner and Dachs-Cadefau gave in~\cite{AM} another algoritm for this purpose.
But even in dimension two finding a closed formula for the general case has turned out
to be difficult. Kuwata calculated in~\cite{K} the smallest jumping number, the so called
log-canonical threshold, for a reduced plane curve with two branches. Galindo, Hernando
and Monserrat succeeded in~\cite{GHM} to generalize this to any number of branches.
Jumping numbers are defined by using an embedded resolution of the subscheme.
They depend on the exceptional divisors appearing in the resolution. We therefore look at
the dual graph of the resolution. Recall that the vertices of the dual graph correspond to
exceptional divisors and two vertices are connected by an edge if the exceptional divisors
in question intersect. To each vertex, we will attach a certain semigroup. We will then
describe jumping numbers in terms of these semigroups. In defining the semigroups we
use Zariski exponents of the valuations associated to the exceptional divisors.
To explain this in more detail, let $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ be a complete ideal of finite colength in a two-dimensional
regular local ring $R$ having an algebraically closed residue field. Let $X\longrightarrow \Spec R$
be a log resolution of the pair $(R,\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d})$. Let $E_1,\ldots,E_N$ be the exceptional divisors.
Let $\Gamma$ be the dual graph of $X$. Two vertices $\gamma$ and $\eta$ are called adjacent, denoted by
$\gamma\sim\eta$, if the corresponding exceptional divisors $E_\gamma$ and $E_\eta$ intersect. The valence $v_\Gamma(\mu)$
of a vertex $\mu$ means the number of vertices adjacent to it. A vertex with valence at most
one is called an end whereas a vertex of valence at least three is a star. Let $v_1,\ldots,v_N$
be the discrete valuations and $\fp_1,\ldots,\fp_N$ the simple ideals corresponding to $E_1,\ldots,E_N$,
respectively. Set $V_{\mu,\nu}=v_{\mu}(\fp_\nu)$ for all $\mu,\nu=1,\ldots,N$. The Zariski exponents are the
numbers $V_{\mu,\tau}$, where $\tau$ is end. Let $S^\mu$ denote the submonoid of $\mathbb N$ generated by $V_{\mu,\mu}$ and
the numbers
$$
s^\mu_\nu:=\gcd\left\{V_{\mu,\tau}\mid v_\Gamma(\tau)=1
\text{ and }
\tau\in\Gamma_\nu^\mu\right\},
$$
where $\Gamma^\mu_\nu$ is the branch emanating from $\mu$ towards $\nu$, i.~e., the maximal connected subgraph
of $\Gamma$ containing $\nu$ but not $\mu$. We will show that $S^\mu$ is a numerical semigroup
generated by at most two elements (see Remark \ref{Sgen2}).
Recall that a divisor $F=f_1E_1+\ldots +f_NE_N$ on $X$ is called antinef if $F \cdot E_\gamma \le 0$
for all $\gamma=1,\ldots,N$, where $F \cdot E_\gamma$ is the intersection product. Let $\{\widehat E_1,\dots,\widehat E_N\}$, where
$E_\mu\cdot\widehat E_\nu=-\delta_{\mu,\nu}$, denote the dual basis of $\{E_1,\dots,E_N\}$. Then $F=\widehat f_1\widehat E_1+\ldots+\widehat f_N\widehat E_N$
is antinef if and only if $\widehat f_i\ge 0$ for all $i=1,\ldots,N$. We call the numbers $\widehat f_1,\ldots, \widehat f_N$ as the
factors of $F$.
We make use of the observation made in~\cite{J} that jumping numbers of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ can be
parametrized by the antinef divisors. More precisely, the jumping number corresponding
to an antinef divisor $F$ is
$$
\xi_F:=\min_\gamma \frac{f_\gamma+k_\gamma+1}{d_\gamma},
$$
where $D=d_1E_1+\ldots+ d_NE_N$ is the divisor on $X$ such that $\cO_X(-D)=\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d} \cO_X$, and
$K=k_1E_1+\ldots+ k_NE_N$ denotes the canonical divisor. We say that $\xi$ is a jumping
number supported at a vertex $\mu$ if $\xi=\xi_F$ for some antinef divisor $F$ with
$$
\xi_F =\frac{f_\mu+k_\mu+1}{d_\mu}.
$$
We fix a vertex $\mu$ and concentrate on the set $\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ of jumping numbers supported at $\mu$.
Our main result, Theorem~\ref{2}, yields a formula for the set of the jumping numbers of
$\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ supported at $\mu$:
$$
\cH_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}
=\left\{\frac{t}{d_\mu}\middle | t+(v_{\Gamma}(\mu)-2)V_{\mu,\mu}
-\sum_{\nu\sim\mu}s^\mu_\nu\left\lceil t\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_i V_{\mu,i}}{s^\mu_\nu d_\mu}\right\rceil^{+}\in S^\mu\right\},
$$
where $\left\lceil\:\:\right\rceil^{+}$ means rounding up to the nearest positive integer. Note that a jumping
number is always supported at a vertex which is either a star or corresponds to a simple
factor of the ideal (see Lemma~\ref{3d}).
In the proof we look at the factors of divisors. Given a vertex $\mu$ we introduce two
transforms of divisors by means of which it is possible 'bring' factors from each branch
emanating from $\mu$ to the closest vertex adjacent to $\mu$ and 'distribute' a part of a factor
from $\mu$ to the adjacent vertices. Suppose that $\xi=\xi_F$ is supported at a $\mu$. Using these
transformations we can modify either $F$ or $D$ or both in such a way that we still have
$\xi=\xi_F$. In particular, we can assume that the divisor $D$ has factors only at the vertices
adjacent to $\mu$. In this process the properties of the mappings $\rho_{[\mu,\gamma]}\colon \Gamma\to\mathbb Q$, where $\mu$ and
$\gamma$ are fixed vertices, and
$$
\nu\mapsto\frac{V_{\gamma,\nu}}{V_{\mu,\nu}},
$$
play a crucial role. In particular, we prove in Lemma~\ref{V<V} that $\rho_{[\mu,\gamma]}$ is strictly increasing
along the path going from $\mu$ to $\gamma$, and stays constant on any path going away from this
path. Finally, we show in Example~\ref{generalexample} how our formula works in practice.
\section{Preliminaries}
In this paper, we make use of the Zariski-Lipman theory of complete ideals. The
general setting here is similar to that discussed in our paper \cite{HJ}. For the reader's convenience,
we collect here some basic concepts and notation. More details can be found
in~\cite{L}, \cite{C}, \cite{LW} and~\cite{J}.
\subsubsection*{About Zariski-Lipman theory}
Let $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ be a complete ideal of finite colength in a two-dimensional regular local ring $R$
having an algebraically closed residue field. Let $\pi \colon X\rightarrow\Spec(R)$ be a principalization
of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$. Then $X$ is a regular scheme and $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}\cO_X=\cO_X(-D)$ for an effective Cartier divisor
$D$. The morphism $\pi$ is a composition of point blowups of regular schemes
\begin{equation*}
\pi:X=X_{N+1}\xrightarrow{\pi_N}\cdots\xrightarrow{\pi_2} X_2\xrightarrow{\pi_1}X_1=\Spec R,
\end{equation*}
where $\pi_\mu$ is the blowup of $X_\mu$ at a closed point $x_\mu\in X_\mu$. Let $E_\mu$ be the strict and $E^*_\mu$ the
total transform of the exceptional divisor $\pi^{-1}_\mu\{x_\mu\}$ on $X$. We write $v_\mu$ for the discrete
valuation associated to the discrete valuation ring $\cO_{X,E_\mu}$, so that $v_\mu$ is the $\fm_{X_\mu,x_\mu}$-adic
order valuation.
A point $x_\mu$ is \textit{infinitely near} to a point $x_\nu$, if the projection $X_\mu \rightarrow X_\nu$ maps $x_\mu$ to
$x_\nu$. Further, $x_\mu$ is \textit{proximate} to $x_\nu$, denoted by $\mu\succ\nu$, if $x_\mu$ lies on the strict transform
of $\pi_\nu^{-1}\{x_\nu\}$ on $X_\mu.$ Note that a point can be proximate to at most two points. The
\textit{proximity matrix} is
\begin{equation*}\label{prox}
P:=(p_{\mu,\nu})_{N\times N},\text{ where }p_{\mu,\nu}=
\left\lbrace
\begin{array}{rl}
1,&\text{if }\mu=\nu;\\
-1,&\text{if }\mu\succ \nu;\\
0,&\text{otherwise.}
\end{array}
\right.
\end{equation*}
We write $Q=(q_{\mu,\nu})_{N\times N}:=P^{-1}$, so that $PQ=1$.
Besides the obvious one, the lattice $\Lambda:= \mathbb Z E_1 \oplus \ldots \oplus \mathbb Z E_N$ of exceptional divisors
on $X$
has two other convenient bases, namely $\{E_1^*,\dots,E_N^*\}$ and $\{\widehat E_1,\dots,\widehat E_N\}$, where
$E_\mu\cdot\widehat E_\nu=E_\mu^*\cdot E_\nu^*=-\delta_{\mu,\nu}$. Throughout this paper we use the practice that if an
upper case letter, say $G$, denotes a divisor $G\in\Lambda$, then the corresponding lower case
letter possibly with an accent mark denotes the coefficient vector with respect to the
appropriate base. In particular, writing
$$
G=g_1E_1+\ldots+g_NE_N=g_1^*E^*_1+\ldots+g_N^*E^*_N=\widehat g_1\widehat E_1+\ldots+\widehat g_N\widehat E_N
$$
with $g=(g_\nu), g^*=(g^*_\nu)$ and $\widehat g=(\widehat g_\nu)$, we get the following base change formulas:
\begin{equation}\label{BC}
g^*=gP^\textsc{t}\text{ and }\widehat g=gP^\textsc{t}P=g^*P.
\end{equation}
In many cases, we regard $\Lambda$ as a subset of $\Lambda_\mathbb Q := \mathbb Q\otimes\Lambda$. We call the vector $\widehat g$ the
\emph{factorization vector} and $g$ the \emph{valuation vector} of the divisor. Note that $g=\widehat gV$, where
$V:=(P^\textsc{t}P)^{-1}$ is called the \emph{valuation matrix}. Set
$$
w_\Gamma(\mu):=-E_\mu^2=1+\#\{\nu\mid \nu \succ \mu\}.
$$
We then get the formulas
\begin{equation}\label{P}
\widehat g_\mu=g^*_\mu-\sum_{\nu\succ \mu}g^*_\nu=w_\Gamma(\mu)g_\mu-\sum_{\nu\sim\mu}g_\nu\quad(\mu=1,\ldots,N).
\end{equation}
Especially, this yields
\begin{equation}\label{PW}
w_\Gamma(\eta)V_{\mu,\eta}=\sum_{i\sim\eta}V_{\mu,i}+\delta_{\mu,\eta}.
\end{equation}
Recall that a divisor $F\in \Lambda$ is \textit{antinef} if $\widehat f_\nu=-F \cdot E_\nu \ge 0$ for all $\nu=1,\ldots,N$.
Equivalently, the \textit{proximity inequalities}
\begin{equation*}
f^*_\mu\ge\sum_{\nu\succ\mu}f^*_\nu\quad(\mu=1,\ldots,N)
\end{equation*}
hold. Note that they can also be expressed in the form
\begin{equation*}
w_\Gamma(\mu)f_\mu\ge\sum_{\nu\sim\mu}f_\nu\quad(\mu=1,\ldots,N).
\end{equation*}
In fact, if $F\not=0$ is antinef, then also $f_\nu>0$ for all $\nu=1,\ldots,N$. There is a one to
one correspondence between the antinef divisors in $\Lambda$ and the complete ideals of finite
colength in $R$ generating invertible $\cO_X$-sheaves, given by $F\leftrightarrow\Gamma(X,\cO_X(-F))$.
An ideal is called \textit{simple} if it cannot be expressed as a product of two proper ideals.
By the famous result of Zariski, every complete ideal factorizes uniquely into a product
of simple complete ideals. More precisely, we can present a complete ideal $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ as a product
$$
\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=\fp_1^{\widehat d_1}\cdots\fp_N^{\widehat d_N},
$$
where $\fp_\mu\subset R$ denotes the simple complete ideal of finite colength corresponding to the
exceptional divisor $\widehat E_\mu$ and $\widehat d_i\in\mathbb N$ for every $i$. By~\eqref{BC}
\begin{equation*}
\widehat E_\mu=\sum_\nu q_{\mu,\nu}E^*_\nu=\sum_{\nu,\rho}q_{\nu,\rho}q_{\mu,\rho}E_\nu.
\end{equation*}
In particular, we observe the \textit{reciprocity formula}
\begin{equation*}
v_\nu(\fp_\mu)=\sum_{\rho}q_{\nu,\rho}q_{\mu,\rho}=v_\mu(\fp_\nu)\quad(\mu,\nu=1,\ldots,N),
\end{equation*}
in short, $V=V^\textsc t$.
Recall that the \textit{canonical divisor} is $K=\sum_\nu E^*_\nu$. If $k =(k_\nu)$ and
$\widehat k =(\widehat k_\nu)$ are the appropriate coefficient vectors, we have
$$
kE=\widehat k\widehat E=K.
$$
The formulas \eqref{BC} yield
\begin{equation}\label{K}
k_\nu=\sum_\mu q_{\nu,\mu}
\quad\hbox{and}\quad
\widehat k_\nu=E_\nu^2+2\quad(\nu=1,\ldots,N).
\end{equation}
\subsubsection*{Dual graph}
The \textit{dual graph} $\Gamma$ associated to our principalization is a tree, where the vertices
correspond one to one to the exceptional divisors and an edge between two adjacent
vertices, $\gamma\sim\eta$, means that the corresponding exceptional divisors $E_\gamma$ and $E_\eta$ intersect.
A vertex $\gamma$ corresponding to the exceptional divisor $E_\gamma$ is weighted by the number $w_\Gamma(\gamma)$.
We say that a vertex $\gamma$ is proximate to another vertex $\eta$ if $p_{\gamma,\eta}=-1$. It is \textit{free} if it is
proximate to at most one vertex. We may also say that $\gamma$ is infinitely near to $\eta$, and
write $\eta\subset\gamma$, if this is the case with the corresponding points. The \textit{root} of $\Gamma$ is the vertex
$\tau_0$ for which $\tau_0\subset\gamma$ for every $\gamma\in\Gamma$.
Blowing up a point on $E_\gamma$ expands the dual graph by adding a vertex $\nu$ corresponding
to the exceptional divisor of the blowup. The weight of the new vertex is one and
the weights of the adjacent vertices are increased by one. In \cite[Definition 5.1]{S1} such
expansions are called elementary modifications. There are two kinds of elementary modifications.
If $E_\gamma$ is the only exceptional divisor containing the center of blowup so that
$\gamma\sim\nu$ forms the only new edge, then the elementary modification is of the first kind:
\newline
\begin{picture}(400,75)(-40,-25)
\put(35,0){\circle*{5}}
\put(35,15){\makebox(0,0)[cc]{\small$w_\gamma$}}
\put(35,-14){\makebox(0,0)[cc]{\small$\gamma$}}
\put(35,0){\line(-2,1){10}}
\put(35,0){\line(-2,-1){10}}
\put(30,0){\line(-1,0){5}}
\put(35,0){\line(2,1){10}}
\put(35,0){\line(2,-1){10}}
\put(40,0){\line(1,0){5}}
\put(235,0){\line(-2,1){10}}
\put(235,0){\line(-2,-1){10}}
\put(230,0){\line(-1,0){5}}
\put(235,0){\line(2,1){10}}
\put(235,0){\line(2,-1){10}}
\put(240,0){\line(1,0){5}}
\put(130,-5){\large$\leadsto$}
\put(235,0){\circle*{5}}
\put(225,15){\makebox(0,0)[cc]{\small$w_\gamma+1$}}
\put(236,-14){\makebox(0,0)[cc]{\small$\gamma$}}
\put(235,0){\line(1,1){27}}
\put(262,27){\circle*{5}}
\put(265,16){\makebox(0,0)[cc]{\small$\nu$}}
\put(260,38){\makebox(0,0)[cc]{\small$1$}}
\end{picture}
If the center of blowup is the intersection point of $E_\gamma$ and another exceptional divisor,
say $E_\eta$, then the edge $\gamma\sim\eta$ is replaced by the edges $\gamma\sim\nu$ and $\nu\sim\eta$, and the elementary
modification is of the second kind:
\newline
\begin{picture}(400,75)(-40,-25)
\put(0,0){\circle*{5}}
\put(0,15){\makebox(0,0)[cc]{\small$w_\gamma$}}
\put(0,-14){\makebox(0,0)[cc]{\small$\gamma$}}
\put(0,0){\line(1,0){70}}
\put(0,0){\line(-2,1){10}}
\put(0,0){\line(-2,-1){10}}
\put(-5,0){\line(-1,0){5}}
\put(70,0){\line(2,1){10}}
\put(70,0){\line(2,-1){10}}
\put(75,0){\line(1,0){5}}
\put(210,0){\line(-2,1){10}}
\put(210,0){\line(-2,-1){10}}
\put(205,0){\line(-1,0){5}}
\put(310,0){\line(2,1){10}}
\put(310,0){\line(2,-1){10}}
\put(315,0){\line(1,0){5}}
\put(70,0){\circle*{5}}
\put(70,15){\makebox(0,0)[cc]{\small$w_\eta$}}
\put(70,-14){\makebox(0,0)[cc]{\small$\eta$}}
\put(130,-5){\large$\leadsto$}
\put(210,0){\circle*{5}}
\put(205,15){\makebox(0,0)[cc]{\small$w_\gamma+1$}}
\put(210,-14){\makebox(0,0)[cc]{\small$\gamma$}}
\put(210,0){\line(2,1){50}}
\put(310,0){\line(-2,1){50}}
\put(260,24.5){\circle*{5}}
\put(260,13){\makebox(0,0)[cc]{\small$\nu$}}
\put(260,36){\makebox(0,0)[cc]{\small$1$}}
\put(310,0){\circle*{5}}
\put(315,15){\makebox(0,0)[cc]{\small$w_\eta+1$}}
\put(310,-14){\makebox(0,0)[cc]{\small$\eta$}}
\end{picture}
Let us write
$$
\Gamma(\nu,U)\text{ where }U=\{\gamma\in\Gamma\mid\gamma\prec\nu\}
$$
for an elementary modification of the graph $\Gamma$ by adding a vertex $\nu$ adjacent to vertices
$\gamma\in U$. Note that $U$ consists of at most two vertices. Note also that if the graph is empty
then the elementary modification is defined to be of the first kind containing only the
root vertex. Following \cite[Definition 5.2]{S1}, a dual graph dominates a dual graph $\Gamma$, if it
can be obtained from $\Gamma$ by a sequence of elementary modifications. Obviously, a sequence
of point blowups correspond to a sequence of elementary modifications. Especially, the
dual graph of our principalization can be obtained from the graph containing only the
root vertex through successive elementary modifications (c.f. \cite[Remark 5.5]{S1}).
In a way, the matrix $P^\textsc {t}P$ represents the dual graph because the diagonal elements
$(P^\textsc{t}P)_{\nu,\nu}=-E_\nu^2$ correspond with the weights of the vertices while outside the diagonal
the element $(P^\textsc{t}P)_{\mu,\nu}=-E_\mu\cdot E_\nu$ is $-1$ if $E_\mu$ and $E_\nu$ intersect and otherwise zero.
The \textit{valence} $v_\Gamma(\mu)$ of a vertex $\mu$ means the number of vertices adjacent to it. If
$v_\Gamma(\mu)\ge 3$, then $\mu$ is called a \textit{star}. If $v_\Gamma(\mu)\le1$, then we call it an \textit{end}. The vertices
adjacent to $\mu$ correspond one to one to the branches emanating from $\mu$, which can be
defined as follows:
\begin{defn}
For any two vertices $\mu$ and $\nu$ in $\Gamma$, let $\Gamma^\mu_\nu$ denote the maximal connected
subgraph of $\Gamma$ containing $\nu$ but not $\mu$ $(\Gamma_\mu^\mu=\emptyset)$. We say $\Gamma^\mu_\nu$ is a branch emanating from
$\mu$ towards $\nu$. A branch $\Gamma^\mu_\nu$ is anterior to $\mu$, if $\mu$ is infinitely near to some of its vertices.
Otherwise we say it is posterior to $\mu$.
\end{defn}
Observe that every branch emanating from $\mu$ is either anterior or posterior to $\mu$, and
for those we immediately get the following result:
\begin{prop}
The unempty posterior branches of $\mu$ correspond one to one to the free
vertices, which are proximate to $\mu$, whereas the anterior branches are in one to one
correspondence with the vertices to which $\mu$ is proximate to.
\end{prop}
\begin{proof}
The claim is trivial if $\mu$ is the only vertex, meaning that there are no unempty
branches. We shall proceed by induction on the number of vertices. Suppose $\Gamma=\Gamma'(\eta,U)$
and the claim holds for $\Gamma'$. Observe that for any $\mu\in\Gamma$, $\mu$ is not proximate to $\eta$.
If $\mu=\eta$, then $\gamma\prec\mu$ exactly when $\gamma\sim\mu$, so that $\mu$ is infinitely near to any adjacent
vertex. Obviously, the branches $\Gamma^\mu_\gamma$ correspond one to one to the vertices $\gamma\sim\mu$. Thus
the claim is clear in this case.
Suppose that $\mu\neq\eta$. If $\eta$ is not a free vertex proximate to $\mu$, then the blowup just
augments an existing branch of $\Gamma'$, i.~e., the branches of $\Gamma$ emanating from $\mu$ correspond
one to one to those of $\Gamma'$. Because the proximity relations are preserved under blowup,
the claim follows. If $\eta$ is a free blowup of $\mu$, then $\mu\prec\eta$ and $\Gamma_\eta^\mu=\{\eta\}$ forms a new
branch, which corresponds to the vertex $\eta$. For the rest of the branches emanating from
$\mu$ the correspondence is inherited from $\Gamma'$.
\end{proof}
Recall that a vertex is proximate to at most two vertices. Subsequently, there are at
most two branches anterior to $\mu \in \Gamma$ depending on whether $\mu$ is free or not.
The \textit{distance} between two vertices $\mu, \nu\in \Gamma$ is defined as the length of the \textit{path} $[\nu,\mu]$,
i.~e.,
$$
d(\nu,\mu):=\min\{r\mid\nu=\nu_0\sim\cdots\sim\nu_r=\mu,
\text{ where }
\nu_0,\dots,\nu_r\in\Gamma\},
$$
Furthermore, if $T\subset \Gamma$, we set
$$
d(\nu,T):=\min\{d(\nu,\mu)\mid\mu\in T\}.
$$
If $d(\nu,T)=1$, then we write $\nu\sim T$.
\begin{defn}\label{pair}
A pair $(\gamma,\tau)$ is associated to $\mu$, if $\gamma$ and $\tau$ satisfy the following three
conditions:
\begin{itemize}
\item[i)] $\gamma\subset\tau\subset\mu$, i.~e., $\mu$ is infinitely near to $\tau$ which is infinitely near to $\gamma$;
\item[ii)] $\tau$ is free and infinitely near to every free vertex $\nu\subset\mu$;
\item[iii)] $\gamma$ is not free and infinitely near to every non free vertex $\nu\subset\tau$, unless every $\nu\subset\tau$
is free in which case $\gamma=\tau_0$ is the root.
\end{itemize}
The sequence of pairs $((\gamma_i,\tau_{i+1}))_{i=0}^{g}$ is associated to $\mu:=\gamma_{g+1}$, if it holds for $i=0,\dots,g$
that $(\gamma_i,\tau_{i+1})$ is the pair associated to $\gamma_{i+1}$.
\end{defn}
\begin{rem}\label{rem2}
Let $\Gamma$ be the dual graph of $\mu$, i.~e., the simple dual graph which consists
of all the vertices to which $\mu$ is infinitely near to. Observe that we may always reach
this situation by repeatedly blowing down any vertex different from $\mu$ having a weight one. If
$((\gamma_i,\tau_{i+1}))_{i=0}^{g}$ is now the sequence associated to $\mu$, then $\gamma_0=\tau_0$ is the root, $\tau_0,\dots,\tau_{g+1}$
are exactly the end vertices of $\Gamma$ while $\gamma_1,\dots,\gamma_g$ are its stars (cf. \cite[Proposition 4.3]{J}).
Note that the integer $g$, i.~e., the number of star vertices of the dual graph, is denoted
by $g^*$ in \cite[Notation 3.3]{J}.
\end{rem}
\begin{rem}\label{rem2kuva}
As the relation $\nu\subset\mu$ induces a partial order on $\Gamma$, we might give the
definition as follows: a pair $(\gamma,\tau)$ is associated to $\mu$, if $\tau$ is maximal among the free
points to which $\mu$ is infinitely near to, and $\gamma$ is maximal among the non free points to
which $\tau$ is infinitely near to. The graph below illustrates an example of a sequence of
pairs associated to a vertex.
\begin{picture}(350,154)(-37,-104)
\put(0,0){\circle{5}}
\put(0,0){\circle*{2}}
\put(50,0){\circle*{5}}
\put(50,-40){\circle{5}}
\put(50,-40){\circle*{2}}
\put(100,0){\circle*{5}}
\put(150,-40){\circle*{5}}
\put(150,-80){\circle{5}}
\put(150,-80){\circle*{2}}
\put(150,0){\circle*{5}}
\put(200,0){\circle{5}}
\put(250,0){\circle{5}}
\put(200,0){\circle*{2}}
\put(250,0){\circle*{2}}
\put(275,-20){\circle{5}}
\put(275,-20){\circle*{2}}
\put(75,-60){\circle{5}}
\put(75,-60){\circle*{2}}
\put(175,-60){\circle*{5}}
\put(75,20){\circle*{5}}
\put(225,20){\circle*{5}}
\thicklines
\put(2.5,0){\line(1,0){47.5}}
\multiput(54,0)(5,0){9}{\line(1,0){2}}
\put(100,0){\line(1,0){50}}
\put(150,0){\line(1,0){47.5}}
\multiput(204,0)(5,0){9}{\line(1,0){2}}
\put(50,0){\line(0,-1){37.5}}
\put(150,0){\line(0,-1){37.5}}
\multiput(150,-44.5)(0,-4){8}{\line(0,-1){2}}
\thinlines
\put(51.9,-41.7){\line(5,-4){21.1}}
\put(51.9,1.7){\line(5,4){21.1}}
\put(98.1,1.7){\line(-5,4){21.1}}
\put(201.9,1.7){\line(5,4){21.1}}
\put(248.1,1.7){\line(-5,4){21.1}}
\put(151.9,-41.7){\line(5,-4){21.1}}
\put(151.9,-78,3){\line(5,4){21.1}}
\put(251.9,-1.7){\line(5,-4){21.1}}
\put(-15,11){\makebox(0,0)[cc]{\small$\eta_1=\gamma_0=\tau_0$}}
\put(29,-45){\makebox(0,0)[cc]{\small$\eta_2=\tau_1$}}
\put(40,11){\makebox(0,0)[cc]{\small$\eta_3=\gamma_1$}}
\put(150,-90){\makebox(0,0)[cc]{\small$\eta_4=\tau_2$}}
\put(140,-45){\makebox(0,0)[cc]{\small$\eta_5$}}
\put(104,11){\makebox(0,0)[cc]{\small$\eta_6$}}
\put(150,11){\makebox(0,0)[cc]{\small$\eta_7=\gamma_2$}}
\put(197,11){\makebox(0,0)[cc]{\small$\eta_8$}}
\put(282,11){\makebox(0,0)[cc]{\small$\mu=\eta_9=\gamma_3=\tau_3$}}
\put(75,-70){\makebox(0,0)[cc]{\small$\eta_{11}$}}
\put(75,33){\makebox(0,0)[cc]{\small$\eta_{13}$}}
\put(187,-60){\makebox(0,0)[cc]{\small$\eta_{12}$}}
\put(225,33){\makebox(0,0)[cc]{\small$\eta_{14}$}}
\put(285,-11){\makebox(0,0)[cc]{\small$\eta_{10}$}}
{\color{lightgray}
\thinlines\thinlines
\put(33,-13){\vector(4,3){10}}
\qbezier(8,-3)(35,-4)(45,-35)
\begin{rotate}{-40}
\put(32,4.5){\makebox(0,0)[cc]{\tiny$(\gamma_0,\tau_1)$}}
\end{rotate}
\put(113,-27){\vector(4,3){15}}
\thinlines
\qbezier(58,-3)(120,-10)(145,-75)
\begin{rotate}{-42}
\put(101,50){\makebox(0,0)[cc]{\tiny$(\gamma_1,\tau_2)$}}
\end{rotate}
\put(208,-14){\vector(3,1){15}}
\thinlines
\qbezier(158,-5)(230,-25)(246,-5)
\begin{rotate}{-10}
\put(196,16){\makebox(0,0)[cc]{\tiny$(\gamma_2,\tau_3)$}}
\end{rotate}
}
\end{picture}
\noindent
Here the open circles represent free points. We now have $\eta_1 \subset \cdots \subset \eta_{10}$. Moreover,
$\eta_2\subset \eta_{11}$, $\eta_5\subset \eta_{12}$, $\eta_6\subset \eta_{13}$ and $\eta_9\subset \eta_{14}$. Since we are interested in the vertices
to which $\mu$ is infinitely near to, we may concentrate on the chain $\eta_1 \subset \cdots \subset \eta_9$ or, in
the dual graph, blow down the vertices $\eta_i$ with $i > 9$. The dashed lines in the graph
represent the edges emerging when blowing down. Obviously, the maximal free point to
which $\mu=\eta_9$ is infinitely near to is $\mu$ itself, and further, the maximal non free point to
which $\mu$ is infinitely near to is $\eta_7$. Thus the pair $(\eta_7,\mu)$ is associated to $\mu$. Similarly, the
pair $(\eta_3,\eta_4)$ is associated to $\eta_7$ and $(\eta_1,\eta_2)$ is associated to $\eta_3$.
\end{rem}
\subsubsection*{Jumping numbers}
We will next recall the definition of jumping numbers. A general reference for jumping
numbers is the fundamental article~\cite{ELSV}. For a nonnegative rational number $\xi$, the
\textit{multiplier ideal} $\mathcal{J}} \newcommand{\cK}{\mathcal{K}} \newcommand{\cL}{\mathcal{L}} \newcommand{\cM}{\mathcal{M}(\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}^\xi)$ is defined to be the ideal
$$
\mathcal{J}} \newcommand{\cK}{\mathcal{K}} \newcommand{\cL}{\mathcal{L}} \newcommand{\cM}{\mathcal{M}(\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}^\xi):=\Gamma\left(X,\cO_X\left(K-\left\lfloor\xi D\right\rfloor\right)\right)\subset R,
$$
where $D = d_1E_1 +\cdots+ d_NE_N$ is the divisor corresponding to $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ and $\left\lfloor \xi D\right\rfloor$ denotes the
integer part of $\xi D$. It is now known that there is an increasing discrete sequence
$$
0=\xi_0<\xi_1<\xi_2<\cdots
$$
of rational numbers $\xi_i$ characterized by the properties that $\mathcal{J}} \newcommand{\cK}{\mathcal{K}} \newcommand{\cL}{\mathcal{L}} \newcommand{\cM}{\mathcal{M}(\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}^\xi)=\mathcal{J}} \newcommand{\cK}{\mathcal{K}} \newcommand{\cL}{\mathcal{L}} \newcommand{\cM}{\mathcal{M}(\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}^{\xi_i})$ for $\xi\in[\xi_i,\xi_{i+1})$,
while $\mathcal{J}} \newcommand{\cK}{\mathcal{K}} \newcommand{\cL}{\mathcal{L}} \newcommand{\cM}{\mathcal{M}(\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}^{\xi_{i+1}})\subsetneq\mathcal{J}} \newcommand{\cK}{\mathcal{K}} \newcommand{\cL}{\mathcal{L}} \newcommand{\cM}{\mathcal{M}(\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}^{\xi_i})$ for every $i$. The numbers $\xi_1,\xi_2,\dots$, are called the
\textit{jumping numbers} of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$. The following Proposition~\ref{parametrisointi}, which is fundamental for the rest of
this article, results from~\cite[Proposition 6.7 and Proposition 7.2]{J}.
\begin{prop}\label{parametrisointi}
Let $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d} \subset R$ be a complete ideal of finite colength. Then $\xi$ is a jumping
number of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ if and only if there exists an antinef divisor $F=fE\in \Lambda$ such that
$$
\xi=\xi_F:=\min_\nu \frac{f_\nu+k_\nu+1}{d_\nu}.
$$
Moreover, if $\fb$ is the complete ideal corresponding to $F$, then
$$
\xi=\inf\{c\in\mathbb Q_{>0}\mid\mathcal{J}} \newcommand{\cK}{\mathcal{K}} \newcommand{\cL}{\mathcal{L}} \newcommand{\cM}{\mathcal{M}(\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}^{c})\nsupseteq\fb\}.
$$
\end{prop}
\begin{notation}
We write for any two divisors $F=fE,G=gE\in\Lambda_\mathbb Q$ and for any vertex $\nu$
$$
\lambda(F,G;\nu):=\frac{f_\nu+k_\nu+1}{g_\nu}.
$$
For any integer $a$ we set
$$
\lambda(a,\nu)=\lambda(a,D;\nu):=\lambda(aE,D;\nu).
$$
Furthermore, we call the set
$$
\{\nu\in\Gamma\mid\lambda(f_\nu,\nu)=\xi\}
$$
the \textit{support} of the jumping number $\xi$ with respect to the divisor $F$. The set of jumping
numbers of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ supported at a vertex $\mu\in\Gamma$ is denoted by
$$
\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}:=\{\xi_F\mid F\in\Lambda\textit{ is antinef and }\xi_F=\lambda(F,D;\mu)\}.
$$
\end{notation}
Recall that the function $\lambda_F:|\Gamma|\to\mathbb Q$, where $F=\sum_{\nu\in\Gamma}f_\nu E_\nu$ is a divisor and
$\lambda_F(\nu)=\lambda(f_\nu,\nu)$, makes the dual graph as an ordered tree. In \cite{HJ} we investigated this
kind of ordered tree structures, and further, we proved that a number being a jumping
number is equivalent to the existence of certain kind of ordered tree structures. In the
sequel, we make use of these results.
\begin{rem}\label{DualG}
Note that in \cite{HJ} and \cite{J} $\Gamma$ is the dual graph of the minimal principalization
of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$. We may loosen this restriction and consider the dual graph of a principalization
of the ideal. In the sequel, we may think $\Gamma$ as a dual graph of any ideal corresponding
to some antinef divisor in $\Lambda$. This is convenient, and it is possible because if $\fb$ is such
an ideal, then the principalization corresponding to $\Gamma$ is a principalization of $\fb$, and the
minimal principalization is obtained by blowing down. Observe that the ordered tree
structures behave accordingly. Suppose that the divisor corresponding to $\fb$ is $gE$ and
that the dual graph $\Gamma_\fb$ of its minimal principalization is obtained by blowing down a
vertex $\nu\in\Gamma$, then the valuation matrix of $\fb$ is just a restriction of that of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$. For a divisor
$fE\in\Lambda_\mathbb Q$ and for a vertex $\gamma\in\Gamma_\fb$ we get $\lambda(fE|_{\Gamma_\fb},gE|_{\Gamma_\fb};\gamma)=\lambda(fE,gE;\gamma)$. Thus the
ordered tree structures provided by $\lambda$ (see \cite{HJ}) can be obtained as restrictions, as well.
\end{rem}
Recall our main result in \cite[Theorem 1]{HJ}:
\begin{thm}[Theorem 1 in \cite{HJ}]
We have $\xi\in\mathcal H^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}_\mu$ if and only if there is a (connected) set
$U\subset\Gamma$ containing $\mu$ and a set of nonnegative integers $$\{a_{\eta}\in\mathbb N\mid d(\eta,U)\le1\}$$ satisfying
\begin{itemize}
\item [i)] $\lambda(a_\eta,\eta)>\xi=\lambda(a_\gamma,\gamma)$ for every $\gamma\in U$ and $\eta\sim U$;
\item [ii)] $w_{\Gamma}(\gamma)a_\gamma\ge\sum_{\nu\sim\gamma}a_\nu$ for every $\gamma\in U$.
\end{itemize}
\end{thm}
\noindent For further use, we also give here a refined versions of \cite[Lemma 5]{HJ} and \cite[Lemma 7]{HJ}.
\begin{lem}
\label{hj5}
Given any vertex $\gamma\in\Gamma$ and any nonnegative integer $a_\gamma$, we may choose for
every vertex $\eta\sim\gamma$ a nonnegative integer $a_\eta$ so that
$$
w_{\Gamma}(\gamma)a_{\gamma}\ge\sum_{\eta\sim\gamma}a_\eta
\text{ and }
\lambda(a_\eta,\eta)\ge\lambda(a_\gamma,\gamma),
$$
where the latter inequality holds for each $\eta\sim\gamma$ except at most one. More precisely, if
$$
\{\eta\mid\eta\sim\gamma\}=\{\eta_1,\dots\eta_m\},
$$
where $m>1$, then the following is true:
\begin{itemize}
\item[1)]
If it is possible to find a nonnegative integer $a_{\eta_1}$ with $\lambda(a_{\eta_1},\eta_1) = \lambda(a_\gamma,\gamma)$, then
one may choose the other integers $a_{\eta_j}$ so that
$$
\lambda(a_{\eta_2},\eta_2)\ge\lambda(a_\gamma,\gamma)
\text{ and }
\lambda(a_{\eta_j},\eta_j)>\lambda(a_\gamma,\gamma)
$$
for all $2< j\le m$.
\item[2)]
If it is possible to find a nonnegative integer $a_{\eta_1}$ satisfying $\lambda(a_{\eta_1},\eta_1)<\lambda(a_\gamma,\gamma)$ or,
in the case $\widehat d_\gamma>0$, $\lambda(a_{\eta_1},\eta_1)=\lambda(a_\gamma,\gamma)$, then one can choose the other integers $a_\eta$
in such a way that
$$
\lambda(a_{\eta_j},\eta_j)>\lambda(a_\gamma,\gamma)
$$
holds for every $1<j\le m$.
\end{itemize}
\end{lem}
\begin{proof}
The proof is conducted in \cite[Lemma 5]{HJ} except for the amendment in 1), which
claims that if we have nonnegative integers $a_{\eta_1}$ and $a_{\eta_2}$ satisfying
\begin{equation}\label{>=>}
\lambda(a_{\eta_2},\eta_2)
>\lambda(a_\gamma,\gamma)
=\lambda(a_{\eta_1},\eta_1)
>\lambda(a_{\eta_2}-1,\eta_2),
\end{equation}
then we may find nonnegative integers $a_{\eta_j}$ for $2<j\le m$ so that
$$
w_{\Gamma}(\gamma)a_{\gamma}=\sum_{j=1}^m a_{\eta_j}
\text{ and }
\lambda(a_{\eta_j},\eta_j)\ge\lambda(a_\gamma,\gamma),
$$
where the inequality is strict for $1<j\le m$.
To prove that, suppose that Equation \eqref{>=>} holds. For $\mu,\nu\in\Gamma$, write
$$
\alpha} \newcommand{\gb}{\beta} \newcommand{\gc}{\gamma_{\mu,\nu}:=k_\nu+1-\frac{d_\nu}{d_\mu}(k_\mu+1).
$$
Note that if $a_{\eta_2}=0$, then by \cite[Lemma 3 a)]{HJ}
$$
a_\gamma-1\le a_\gamma+\alpha} \newcommand{\gb}{\beta} \newcommand{\gc}{\gamma_{\eta_2,\gamma}=d_\gamma(\lambda(a_{\gamma},\gamma)-\lambda(a_{\eta_2},\eta_2)),
$$
which must be negative. Therefore $a_\gamma=0$, and then similarly, by \cite[Lemma 3 a)]{HJ},
$$
a_{\eta_1}-1
< a_{\eta_1}+\alpha} \newcommand{\gb}{\beta} \newcommand{\gc}{\gamma_{\gamma,\eta_1}
=d_{\eta_1}(\lambda(a_{\eta_1},\eta_1)-\lambda(a_{\gamma},\gamma))
=0,
$$
so that $a_{\eta_1}=0$, but then the claim follows from \cite[Lemma 3 b)]{HJ}.
Assume then that $a_{\eta_2}>0$ and set $a'_{\eta_2}:=a_{\eta_2}-1$. Now $\lambda(a'_{\eta_2},\eta_2)<\lambda(a_\gamma,\gamma)$, and by
\cite[Lemma 5]{HJ} we may find nonnegative integers $a'_{\eta_j}$ for $1\le j\le m$, $j\neq2$, so that
$$
w_{\Gamma}(\gamma)a_{\gamma}=\sum_{j=1}^m a'_{\eta_j}\text{ and }\lambda(a'_{\eta_j},\eta_j)>\lambda(a_\gamma,\gamma)\text{ for }j\neq2.
$$
But then $\lambda(a'_{\eta_1},\eta_1)>\lambda(a_{\eta_1},\eta_1)$. Clearly, we may choose the integers $a'_{\eta_j}$ so that $a'_{\eta_1}=a_{\eta_1}+1$.
It follows that
$$
w_{\Gamma}(\gamma)a_{\gamma}=a_{\eta_1}+a_{\eta_2}+\sum_{j=3}^m a'_{\eta_j}
\text{ and }
\lambda(a'_{\eta_j},\eta_j)>\lambda(a_\gamma,\gamma)
\text{ for }
2<j\le m.
$$
Choosing now $a_{\eta_j}=a'_{\eta_j}$ for $2<j\le m$ yields the claim.
\end{proof}
Practically, the lemma shows that if $\lambda_F(\mu)$ is a local minimum for a function $\lambda_F$
where $F$ is an effective divisor, then we may find an antinef divisor $A=\sum_\nu a_\nu E_\nu$ for
which $\lambda(a_\mu,\mu)=\lambda(f_\mu,\mu)$ is the global minimum of the function $\lambda_A$. The only problem
that may arise in finding such integers $a_\nu$ is the situation where we already have integers
$a_\gamma$ and $a_\tau$ with $\lambda(a_\gamma,\gamma)=\lambda(a_\tau,\tau)$, and $\tau$ is an end vertex. We cannot go on choosing
integers $a_\nu$ with $\tau\sim\nu\neq\gamma$ and $\lambda(a_\nu,\nu)=\lambda(a_\tau,\tau)$, because there are no such vertices,
and it may happen that $\widehat a_\tau:=w_\gamma(\tau)a_\tau-a_\gamma<0$. This is the reason why we cannot apply
1) of Lemma \ref{hj5} to an end. Nevertheless, rephrasing \cite[Lemma 7]{HJ}, the next Lemma shows
that 2) of Lemma \ref{hj5} is applicable even if the vertex in question is an end.
\begin{lem}\label{hj7}
Suppose $\tau$ is an end and $\gamma$ is adjacent to it. If $a_\tau$ and $a_\gamma$ are such integers
that $\lambda(a_\gamma,\gamma)\le\lambda(a_\tau,\tau)$, where the equality holds only if $\widehat d_\tau>0$, then
$$
w_\Gamma(\tau)a_\tau\ge a_\gamma.
$$
\end{lem}
\begin{proof}
By Equation \eqref{K} we know that $\widehat k_\tau=2-w_\Gamma(\tau)$. On the other hand, by Equation~\eqref{P}
we have $\widehat k_\tau=w_\Gamma(\tau)k_\tau-k_\gamma$. Thus $w_\Gamma(\tau)(k_\tau+1)=k_\gamma+2$. Moreover, by Equation~\eqref{P}
we get $w_\Gamma(\tau)d_\tau=d_\gamma+\widehat d_\tau$. This shows that
$$
\lambda(a_\tau,\tau)
=\frac{w_\Gamma(\tau)(a_\tau+k_\tau+1)}{w_\Gamma(\tau)d_\tau}
=\frac{w_\Gamma(\tau)a_\tau+k_\gamma+2}{d_\gamma+\widehat d_\tau}
$$
Therefore we see that
$$
\lambda(a_\gamma,\gamma)
=\frac{a_\gamma+k_\gamma+1}{d_\gamma}
<\frac{(w_\Gamma(\tau)a_\tau+1)+k_\gamma+1}{d_\gamma},
$$
which implies that $a_\gamma<w_\Gamma(\tau)a_\tau+1$, as wanted.
\end{proof}
By using these results we may construct suitable ordered tree structures, which in
turn can prove that certain rationals are jumping numbers for our ideal supported at the
desired vertex or vertices. The next lemma shows that in order to determine the jumping
numbers of an ideal, we just need to know the jumping numbers supported at a vertex
which is either a star or corresponds to a simple factor of the ideal.
\begin{lem}\label{3d}
A support of a jumping number contains a vertex which is either a star or
corresponds to a simple factor of the ideal.
\end{lem}
\begin{proof}
Let $\Gamma$ be a dual graph of an ideal $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=\prod_{\nu\in\Gamma}\fp_\nu^{\widehat d_\nu}$. Suppose $\xi$ is a jumping number
of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ supported at a vertex $\gamma\in\Gamma$, for which $\widehat d_\gamma=0$ and $v_\Gamma(\gamma)<3$. By Proposition \ref{parametrisointi} we
have an antinef divisor $F$ for which $\xi=\xi_F$. Further, we have $\xi=\lambda(f_\gamma,\gamma)\le\lambda(f_\nu,\nu)$ for
any $\nu\in\Gamma$. As $\widehat k_\gamma=2-w_\Gamma(\gamma)$ by \eqref{K}, and on the other hand, $\widehat k_\gamma=w_\Gamma(\gamma)k_\gamma-\sum_{\eta\sim\gamma}k_\eta$,
we see that $w_\Gamma(\gamma)(k_\gamma+1)=2+\sum_{\eta\sim\gamma}k_\eta$. By this and by \eqref{P} we obtain
$$
\lambda(f_\gamma,\gamma)
=\frac{w_\Gamma(\gamma)f_\gamma+w_\Gamma(\gamma)(k_\gamma+1)}{w_\Gamma(\gamma)d_\gamma}
=\frac{\widehat f_\gamma+2-v_\Gamma(\gamma)+\sum_{\eta\sim\gamma}(f_\eta+k_\eta+1)}{\widehat d_\gamma+\sum_{\eta\sim\gamma}d_\eta}.
$$
Since $\widehat d_\gamma=0$, $v_{\Gamma}(\gamma)<3$ and $\lambda(f_\eta,\eta)\ge\lambda(f_\gamma,\gamma)$, the above yields $v_\Gamma(\gamma)=2$, $\widehat f_\gamma=0$ and
$\lambda(f_\eta,\eta)=\xi$ for any $\eta\sim\gamma$. In other words, $\gamma$ has exactly two adjacent vertices, which
both are in the support of $\xi$. If neither of them is a star nor corresponds to a factor, we
may apply the above to them. Because the dual graph contains finitely many vertices,
we must eventually come up with a vertex in the support of $\xi$, which is either a star or
corresponds to a factor.
\end{proof}
\section{Modifications of the factorization}
Let $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$, $D$ and $\Gamma$ be as above and let $V$ be the valuation matrix, $\widehat d\in\mathbb Q^\Gamma$ the factorization
vector and $d = \widehat d V$ the valuation vector of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$. According to \cite[Theorem 1 and Lemma~6]{HJ},
we may find for any antinef divisor $F=fE\in\Lambda$ an antinef divisor $G=gE$ such that
$\widehat g_\nu>0$ only if $\nu$ is an end of $\Gamma$ and $\xi_F=\xi_{G}$. In this paper we further investigate divisors
corresponding to a jumping number and develop a method to modify them in order to
find ideals sharing the jumping numbers supported at a given vertex. To begin with, let
us consider the mapping $\rho_{[\mu,\gamma]}:\Gamma\to\mathbb Q$, where
$$
\rho_{[\mu,\gamma]}:\nu\mapsto\frac{V_{\gamma,\nu}}{V_{\mu,\nu}}.
$$
\begin{lem}\label{V<V}
The mapping $\rho_{[\mu,\gamma]}$ is strictly increasing on the path going from $\mu$ to $\gamma$ and it
stays constant on any path going away from $[\mu,\gamma]$, in other words, $\rho_{[\mu,\gamma]}(\nu_1)<\rho_{[\mu,\gamma]}(\nu_2)$
if and only if $[\mu,\nu_1]\cap[\mu,\gamma]\subsetneq[\mu,\nu_2]\cap[\mu,\gamma]$. Moreover, if $\mu\sim\gamma$, then
\begin{equation}\label{Oletus}
\rho_{[\mu,\gamma]}(\gamma)=\frac{V_{\gamma,\mu}+1}{V_{\mu,\mu}}.
\end{equation}
\end{lem}
\begin{proof}
Note first that since $\rho_{[\gamma,\mu]}(\nu)\rho_{[\mu,\gamma]}(\nu)=1$ for every $\nu$, the claim holds for $\rho_{[\gamma,\mu]}$
exactly when it holds for $\rho_{[\mu,\gamma]}$. If the claim holds whenever $\mu$ and $\gamma$ are adjacent vertices,
then we get the desired result by induction on the distance of $\mu$ and $\gamma$, as
$$
\rho_{[\mu,\gamma]}(\nu)=\rho_{[\mu,\eta]}(\nu)\rho_{[\eta,\gamma]}(\nu).
$$
Hence, it is enough to consider the cases where $\mu\sim\gamma$. We proceed by using induction
on the number of vertices of the dual graph.
If $\Gamma$ consists of only two adjacent vertices, then
$$
V=\hspace{-3pt}\left[\begin{array}{rr}1&1\\1&2\end{array}\right]
$$
and the case is clear.
Suppose that $\Gamma=\Gamma'(\eta, U)$ and that the claim holds on the graph $\Gamma'$. Note that the
valuation matrix of $\Gamma'$ is just a restriction of that of $\Gamma$. Moreover, since the valuation
matrix of $\Gamma$ is $V^{\textsc t}=V=(P^\textsc tP)^{-1}$, we see that $P_\eta V_\gamma=q_{\gamma,\eta}$, i.~e.,
\begin{equation}\label{Vi}
V_{\gamma,\eta}=\sum_{i\prec\eta}V_{\gamma,i}+\delta_{\eta,\gamma}.
\end{equation}
If now $\eta\notin[\mu,\gamma]$, then $\rho_{[\mu,\gamma]}(\nu)$ remains unaltered when $\nu\neq\eta$, and if $\nu=\eta$ then $j\in\Gamma^{\mu}_\gamma$
for any $j\prec\eta$ exactly when $\eta\in\Gamma_\gamma^\mu$. Therefore for any $j\prec\eta$,
$$
\rho_{[\mu,\gamma]}(\eta)
=\frac{\sum_{i\prec\eta}V_{\gamma,i}}{\sum_{i\prec\eta}V_{\mu,i}}
=\rho_{[\mu,\gamma]}(j).
$$
It remains to show that if $\mu\sim\eta$, then the claim holds for $\rho_{[\mu,\eta]}$, too. If $U=\{\mu\}$ and
$\nu\neq\eta$, then by Equation \eqref{Vi} we see that $V_{\eta,\nu}=V_{\mu,\nu}$, and further,
$$
\rho_{[\mu,\eta]}(\nu)
=\frac{V_{\eta,\mu}}{V_{\mu,\mu}}
<\frac{V_{\eta,\mu}+1}{V_{\mu,\mu}}
=\frac{V_{\eta,\eta}}{V_{\mu,\eta}}
=\rho_{[\mu,\eta]}(\eta),
$$
as wanted. Especially, Equation \eqref{Oletus} holds in this case.
Suppose then that $U=\{\mu,\gamma\}$. Together with \eqref{Oletus} Equation \eqref{Vi} yields
\begin{align*}
\rho_{[\mu,\eta]}(\gamma)
&=\frac{V_{\mu,\gamma}+V_{\gamma,\gamma}}{V_{\mu,\gamma}}=1+\frac{V_{\gamma,\mu}+1}{V_{\mu,\mu}}\\
&=1+\frac{V_{\gamma,\gamma}+V_{\gamma,\mu}+1}{V_{\mu,\gamma}+V_{\mu,\mu}}
=\frac{V_{\mu,\eta}+V_{\gamma,\eta}+1}{V_{\mu,\eta}}
=\rho_{[\mu,\eta]}(\eta).
\end{align*}
Moreover,
$$
\rho_{[\mu,\eta]}(\eta)
=\rho_{[\mu,\eta]}(\gamma)
=\frac{V_{\eta,\gamma}}{V_{\mu,\gamma}}=\frac{V_{\eta,\mu}+V_{\eta,\gamma}+1}{V_{\mu,\mu}+V_{\mu,\gamma}}
=\frac{V_{\eta,\mu}+1}{V_{\mu,\mu}}
>\rho_{[\mu,\eta]}(\mu).
$$
This shows that Equation \eqref{Oletus} holds for $\eta$. Since
$$
\rho_{[\mu,\eta]}(\nu)
=\frac{V_{\mu,\nu}+V_{\gamma,\nu}}{V_{\mu,\nu}}
=1+\rho_{[\mu,\gamma]}(\nu)
$$
for every $\nu\neq\eta$, we see that $\rho_{[\mu,\eta]}$ stays constant on any path going away from $[\mu,\eta]$.
Hence the claim holds for $\rho_{[\mu,\eta]}$, too.
\end{proof}
\begin{prop}\label{D}
Write $\mathbf 1_i=(\delta_{i,j})_{j\in\Gamma}$. For any vertices $\gamma$, $\mu$ and $\eta$, set
$$
\widehat r_{[\mu,\gamma]}
:=\mathbf 1_\gamma-\rho_{[\mu,\gamma]}(\mu)\mathbf 1_\mu
\:\:\:\text{ and }\:\:\:
\varphi_\eta(\nu)=\varphi_\eta^{[\mu,\gamma]}(\nu)
:=\frac{\left(\widehat r_{[\mu,\gamma]}V\right)_\nu}{V_{\eta,\nu}}.
$$
Then $\varphi_\eta(\nu)\ge0$, where the inequality is strict if and only if $\nu\in\Gamma^\mu_\gamma$. If $\nu,\nu'\in[\mu,\gamma]$ and
$d(\mu,\nu)<d(\mu,\nu')$, then
$$
\varphi_\eta(\nu)<\varphi_\eta(\nu').
$$
Further, if $\eta\in[\mu,\gamma]$ then $\varphi_\eta(\nu)$ is constant on any path intersecting $[\mu,\gamma]$ at most on
one point.
\end{prop}
\begin{proof}
We have
\begin{equation*}
\varphi_\eta(\nu)
=\rho_{[\eta,\gamma]}(\nu)-\rho_{[\mu,\gamma]}(\mu)\rho_{[\eta,\mu]}(\nu)
=\left(\rho_{[\mu,\gamma]}(\nu)-\rho_{[\mu,\gamma]}(\mu)\right)\rho_{[\eta,\mu]}(\nu)
\end{equation*}
By Lemma \ref{V<V} $\rho_{[\mu,\gamma]}(\nu)\ge\rho_{[\mu,\gamma]}(\mu)$ and thereby also $\varphi_\eta(\nu)\ge0$, where the equality holds
exactly when $\nu\in\Gamma\smallsetminus\Gamma^\mu_\gamma$.
Suppose $\nu,\nu'\in[\mu,\gamma]$ and $d(\mu,\nu)<d(\mu,\nu')$. If $]\mu,\eta]\cap[\mu,\gamma]=\emptyset$, then $\rho_{[\eta,\mu]}(\nu)$ does
not depend on $\nu\in[\mu,\gamma]$, and again by Lemma \ref{V<V} we know that $\rho_{[\mu,\gamma]}(\nu)$ is strictly
increasing on the path going from $\mu$ to $\gamma$, which proves the case. If $]\mu,\eta]\cap[\mu,\gamma]\neq\emptyset$,
then $\rho_{[\eta,\gamma]}(\nu)$ is strictly increasing on $[\eta,\gamma]$ and $\rho_{[\eta,\mu]}(\nu)$ is strictly decreasing on $[\mu,\eta]$,
while $\rho_{[\mu,\gamma]}(\mu)$ is a constant. Therefore
$$
\varphi_\eta(\nu)
=\rho_{[\eta,\gamma]}(\nu)-\rho_{[\mu,\gamma]}(\mu)\rho_{[\eta,\mu]}(\nu)
<\rho_{[\eta,\gamma]}(\nu')-\rho_{[\mu,\gamma]}(\mu)\rho_{[\eta,\mu]}(\nu')
=\varphi_\eta(\nu').
$$
The rest is now clear.
\end{proof}
In the sequel, we shall make use of the above especially in situations where we have
a divisor $F$ with $\xi=\xi_F=\lambda(F,D;\mu)$ for a vertex $\mu$ and we want to modify either $F$ or $D$
or both in such a way that we still have $\xi=\lambda(F',D';\mu)\le\lambda(F',D';\nu)$ for every $\nu\in\Gamma$.
For that we introduce a modified factorization vector: Let $\widehat f,\widehat g,\widehat h\in\mathbb Q^\Gamma$ and let $\mu\in\Gamma$.
We concentrate on $\mu$ and the vertices adjacent to it and modify $\widehat f$ with $\widehat g$ and $\widehat h$ so that
we 'bring' factors $\widehat g_i$ from each branch emanating from $\mu$ to the closest vertex adjacent
to $\mu$ and 'distribute' the factor $\sum_{\nu\sim\mu}\rho_{[\mu,\nu]}(\mu)\widehat h_\nu$ from $\mu$ to the adjacent vertices.
\begin{notation}
Let us write $\widehat f^{\,\mu}_{\langle\, \widehat g\,\rangle[\,\widehat h\,]}$ or, if the vertex $\mu$ is clear from the context, just
$$
\widehat f_{\langle\,\widehat g\,\rangle[\,\widehat h\,]}
:=\widehat f-\sum_{i\sim\mu}\sum_{j\in\Gamma_i^\mu}\widehat g_j\widehat r_{[i,j]}+\sum_{i\sim\mu}\widehat h_i\widehat r_{[\mu,i]}.
$$
If either $\widehat g$ or $\widehat h$ is zero, we may omit it in the notation. Let us also set
$$
\widehat f^\mathcal N:=\widehat f-\sum_{i\in\Gamma}\widehat f_i\widehat r_{[\mu,i]}.
$$
\end{notation}
\begin{rem}\label{Fuf}
Suppose $\widehat F=\widehat f_{\langle\,\widehat g\,\rangle[\,\widehat h\,]}$. Then obviously $\widehat f=\widehat F_{\langle\,-\widehat g\,\rangle[\,-\widehat h\,]}$. Moreover, we have
$$
\widehat f^\mathcal N=\widehat f_{\langle\,\widehat f\,\rangle [\,-\widehat f_{\langle\,\widehat f\,\rangle}\,]}
\text{\,,\, i.~e.,\, }
\widehat f^\mathcal N_{[\,\widehat f_{\langle\,\widehat f\,\rangle}\,]}=\widehat f_{\langle\,\widehat f\,\rangle}.
$$
\end{rem}
\begin{lem}\label{dM}
Let $fE=\widehat f\widehat E$ be a divisor. Write $U_i:=\Gamma^\mu_i\smallsetminus\{i\}$. Then
$$
\left(\widehat f_{\langle\, \widehat g\,\rangle[\,\widehat h\,]}\right)_i=
\begin{cases}\displaystyle
\widehat f_\mu-\sum_{\nu\sim\mu}\rho_{[\mu,\nu]}(\mu)\widehat h_\nu&\text{ if }i=\mu\\
\displaystyle\widehat f_i+\widehat h_i+\sum_{j\in U_i}\rho_{[i,j]}(\mu)\widehat g_j&\text{ if }i\sim\mu\\
\widehat f_i-\widehat g_i&\text{ otherwise. }
\end{cases}
$$
Furthermore,
$$
(\widehat f_{\langle\,\widehat g\,\rangle}V)_i=f_i-\sum_{\nu\sim\mu}\sum_{j\in\Gamma_\nu^\mu}\widehat g_j\varphi_\mu^{[\nu,j]}(i)V_{\mu,i}.
$$
It follows that if $\widehat g\ge0$, then $(\widehat f_{\langle\,\widehat g\,\rangle}V)_i\le f_i$, where the equality holds when $d(\mu,i)\le1$,
or more precisely, the inequality is strict exactly when $i\in\Gamma_j^\nu$ for some $\nu\sim\mu$ and some
$j\in\Gamma_\nu^\mu$ with $\widehat g_j>0$.
Similarly,
$$
(\widehat f_{[\,\widehat h\,]}V)_i=f_i+\sum_{\nu\sim\mu}\widehat h_\nu\varphi_\mu^{[\mu,\nu]}(i)V_{\mu,i},
$$
and if $\widehat h\ge0$, then $(\widehat f_{[\,\widehat h\,]}V)_i\ge f_i$, where the inequality is strict exactly when $i\in\Gamma_\nu^\mu$ for
such $\nu\sim\mu$ that $\widehat h_\nu>0$.
\end{lem}
\begin{proof}
Recall that $\widehat r_{[i,i]}=0$ for any $i$. Straightforward calculation shows that
\begin{align*}
\widehat f_{\langle\,\widehat g\,\rangle[\,\widehat h\,]}
&=\widehat f-\sum_{i\sim\mu}\sum_{j\in U_i}\widehat g_j(\mathbf 1_j-\rho_{[i,j]}(i)\mathbf 1_i)
+\sum_{i\sim\mu}\widehat h_i(\mathbf 1_i-\rho_{[\mu,i]}(\mu)\mathbf 1_\mu)\\
&=\widehat f-\sum_{i\sim\mu}\rho_{[\mu,i]}(\mu)\widehat h_i\mathbf 1_\mu
+\sum_{i\sim\mu}\left(\widehat h_i+\sum_{j\in U_i}\rho_{[i,j]}(\mu)\widehat g_j\right)\mathbf 1_i
-\sum_{i\sim\mu}\sum_{j\in U_i}\widehat g_j\mathbf 1_j
\end{align*}
as $\rho_{[i,j]}(i)=\rho_{[i,j]}(\mu)$ by Lemma \ref{V<V}. This proves the first assertion. Furthermore, by
Proposition \ref{D} we observe that
$$
(\widehat f_{\langle\,\widehat g\,\rangle}V)_i
=f_i-\sum_{\nu\sim\mu}\sum_{j\in\Gamma_\nu^\mu}\widehat g_j(\widehat r_{[\nu,j]}V)_i
=f_i-\sum_{\nu\sim\mu}\sum_{j\in\Gamma_\nu^\mu}\widehat g_j\varphi_\mu^{[\nu,j]}(i)V_{\mu,i},
$$
where $\varphi_\mu^{[\nu,j]}(i)\ge0$ is positive if and only if $i\in\Gamma_j^\nu$. Assuming $\widehat g\ge0$, this shows that
$$
\sum_{\nu\sim\mu}\sum_{j\in\Gamma_\nu^\mu}\widehat g_j\varphi_\mu^{[\nu,j]}(i)V_{\mu,i}>0
$$
if and only if $i\in\Gamma_j^\nu$ for some $\nu\sim\mu$ and some $j\in\Gamma_\nu^\mu$ with $\widehat g_j>0$.
Similarly, by Proposition \ref{D} we get
$$
(\widehat f_{[\,\widehat h\,]}V)_i
=f_i+\sum_{\nu\sim\mu}\widehat h_\nu(\widehat r_{[\mu,\nu]}V)_i
=f_i+\sum_{\nu\sim\mu}\widehat h_\nu\varphi_\mu^{[\mu,\nu]}(i)V_{\mu,i},
$$
where $\varphi_\mu^{[\mu,\nu]}(i)>0$ exactly when $i\in\Gamma^\mu_\nu$. Thus, if $\widehat h\ge0$,
$$
\sum_{\nu\sim\mu}\widehat h_\nu\varphi_\mu^{[\mu,\nu]}(i)V_{\mu,i}>0
$$
exactly when $i\in\Gamma^\mu_\nu$ for such $\nu\sim\mu$ that $\widehat h_\nu>0$.
\end{proof}
\begin{lem}\label{dM0}
For any divisor $\widehat f\widehat E$ we have
$$
\widehat f^\mathcal N=\sum_{i\in\Gamma}\rho_{[\mu,i]}(\mu)\widehat f_i\mathbf 1_\mu
\text{\: and \:}
\left(\widehat f^\mathcal NV\right)_\mu=\left(\widehat fV\right)_\mu.
$$
Furthermore, if for some divisor $\widehat g\widehat E$ holds $\lambda(\widehat f\widehat E,D;\mu)=\lambda(\widehat g\widehat E,D;\mu)$, then
$$
\widehat f^\mathcal N=\widehat g^\mathcal N.
$$
\end{lem}
\begin{proof}
The first equality comes straightforwardly from the definition. A direct calculation
shows that
$$
\left(\widehat f^\mathcal NV\right)_\mu
=\sum_{i\in\Gamma}\rho_{[\mu,i]}(\mu)\widehat f_i V_{\mu,\mu}
=\sum_{i\in\Gamma}\frac{V_{i,\mu}}{V_{\mu,\mu}}\widehat f_i V_{\mu,\mu}
=\sum_{i\in\Gamma}\widehat f_i V_{i,\mu}
=\left(\widehat fV\right)_\mu,
$$
as wanted.
Suppose next that $\lambda(\widehat f\widehat E,D;\mu)=\lambda(\widehat g\widehat E,D;\mu)$. Then $(\widehat f V)_\mu=(\widehat g V)_\mu$. By the above
we have $(\widehat f^\mathcal NV)_\mu=(\widehat g^\mathcal NV)_\mu$, and further, $\widehat f^\mathcal N_\mu V_{\mu,\mu}=\widehat g^\mathcal N_\mu V_{\mu,\mu}$. This is to say that $\widehat f^\mathcal N_\mu=\widehat g^\mathcal N_\mu$,
but then $\widehat f^\mathcal N=\widehat g^\mathcal N.$
\end{proof}
\begin{lem}\label{dM1}
For antinef divisors $fE\neq0$ and $G$ and for any vertex $\nu\in\Gamma$
$$
\lambda(G, \widehat f_{\langle \,\widehat f\,\rangle}\widehat E;\nu)\ge\lambda(G,fE;\nu),
$$
where the equality holds exactly when either $d(\mu,\nu)\le1$ or when $\widehat f_j=0$ for every $j\in\Gamma_\nu^i$,
where $i$ is the vertex in $[\mu,\nu[$ adjacent to $\mu$.
\end{lem}
\begin{proof}
By Lemma \ref{dM} we know that $\widehat f_{\langle \,\widehat f\,\rangle}V_\nu\le f_\nu$, where the equality holds exactly when
either $d(\mu,\nu)\le1$ or when $\widehat f_j=0$ for every $j\in\Gamma_\nu^i$, where $i$ is the vertex in $[\mu,\nu[$ adjacent
to $\mu$. Thus the claim is clear.
\end{proof}
\begin{lem}\label{dM2}
Let $F=fE$ be an antinef divisor and suppose $0\neq\widehat d\in\mathbb Q^\Gamma_{\ge0}$. If
$$
\min_{\nu}\lambda(F,\widehat d_{\langle\widehat d\,\rangle}\widehat E;\nu)=\lambda(F,\widehat d_{\langle\widehat d\,\rangle}\widehat E;\mu),
$$
then we can find an antinef divisor $G$ satisfying
$$
\min_{\nu}\lambda(G,dE;\nu)=\lambda(G,dE;\mu)=\lambda(F,dE;\mu).
$$
\end{lem}
\begin{proof}
Observe that for any divisors $U,V\in\Lambda_\mathbb Q$ and for any nonzero $n\in\mathbb Q$,
\begin{equation}\label{mUm}
\lambda(U,V;\nu)=n\lambda(U,nV;\nu).
\end{equation}
Thus it is not a restriction to assume that both $\widehat d$ and $\widehat d_{\langle\widehat d\,\rangle}$ are in
$\mathbb N^\Gamma$.
We need to show that there is a suitable vector $a\in\mathbb N^\Gamma$, for which the divisor $G=aE$
is as wanted. For $\nu$ with $d(\mu,\nu)\le1$ we set $a_\nu=f_\nu$, so that by Lemma \ref{dM1} we get
$\lambda(F,\widehat d_{\langle\widehat d\,\rangle}\widehat E;\nu)=\lambda(a_\nu,dE;\nu)$. It follows that $\lambda(a_\nu,dE;\nu)\ge\lambda(a_\mu,dE;\mu)$ and
$$
\widehat a_\mu:=w_\Gamma(\mu) a_\mu-\sum_{\nu\sim\mu}a_\nu=\widehat f_\mu\ge0.
$$
For any branch $\Gamma_\gamma^\mu$ with $\gamma\sim\mu$ we have two possible cases: either $\widehat d_\eta=0$ for every $\eta\in\Gamma_\gamma^\mu$, or
$\widehat d_\eta>0$ for some $\eta\in\Gamma_\gamma^\mu$. In the first case, we see by Lemma \ref{dM1} that
$\lambda(F,\widehat d_{\langle\widehat d\,\rangle}\widehat E;\nu)=\lambda(F,dE;\nu)$ for $\nu\in\Gamma_\gamma^\mu$. Hence we may choose $a_\nu=f_\nu$ for every $\nu\in\Gamma_\gamma^\mu$,
so that $\lambda(a_\nu,dE;\nu)\ge\lambda(a_\mu,dE;\mu)$ and $\widehat a_\nu:=w_\Gamma(\nu) a_\nu-\sum_{i\sim\nu}a_i\ge0$. In the latter case,
it may happen that $\lambda(F,dE;\nu)<\lambda(F,dE;\mu)$ for some $\nu\in\Gamma_\gamma^\mu\smallsetminus\{\gamma\}$, so that the integers
$f_\nu$ for $\nu\in\Gamma_\gamma^\mu\smallsetminus\{\gamma\}$ won't do. Therefore we must apply Lemma \ref{hj5} in selecting suitable set
of integers. If $\lambda(F,dE;\gamma)>\lambda(F,dE;\mu)$, then this would be straightforward, since then
we could by Lemma \ref{hj5} choose integers $a_\nu$ for $\nu\in\Gamma_\gamma^\mu\smallsetminus\{\gamma\}$ so that $\lambda(a_\nu,dE,\nu)$ strictly
increases on every path in $\Gamma_\gamma^\mu$ going away from $\gamma$, and $\widehat a_\nu\ge0$. Recall that by Lemma~\ref{hj7}
we can apply Lemma \ref{hj5}, 2) to end vertices, too. In general, we may by using lemma~\ref{hj5}
find such integers $a_\nu$ for $\nu\in\Gamma_\gamma^\mu\smallsetminus\{\gamma\}$, that $\lambda(a_\nu,dE,\nu)$ is increasing on the path $[\mu,\eta]$,
and strictly increases on every path in $\Gamma_\gamma^\mu$ going away from $[\mu,\eta]$, and $\widehat a_\nu\ge0$. This can
be shown as follows.
Let $\eta\in\Gamma_\gamma^\mu$ be such that $\widehat d_\eta>0$, and write $\mu=\eta_0$ and $\gamma=\eta_1$. We have a path of
adjacent vertices $\eta_0\sim\cdots\sim\eta_k=\eta$ for some positive integer $k$. Since $\lambda(a_{\eta_1},dE;\eta_1)\ge\lambda(a_{\eta_0},dE;\eta_0)$,
we may by Lemma \ref{hj5} choose integers $a_\nu$ for $\eta_0\neq\nu\sim\eta_1$ so that $\widehat a_{\eta_1}\ge0$
and $\lambda(a_\nu,dE;\nu)\ge\lambda(a_{\eta_1},dE;\eta_1)$, where the equality takes place only if $\nu=\eta_2$. Similarly,
if $0<i\le k$ and $\lambda(a_{\eta_i},dE;\eta_i)\ge\lambda(a_{\eta_{i-1}},dE;\eta_{i-1})$, we may by Lemma \ref{hj5} choose integers
$a_\nu$ for $\eta_{i-1}\neq\nu\sim\eta_i$ so that $\widehat a_{\eta_i}\ge0$ and $\lambda(a_\nu,dE;\nu)\ge\lambda(a_{\eta_i},dE;\eta_i)$, where the equality
takes place only if $i<k$ and $\nu=\eta_{i+1}$.
If $\theta\sim\eta_i$ for some $i\in\{1,\dots,k\}$ and $\theta\notin[\mu,\eta]$, then we have
$$
\lambda(a_{\theta},dE;\theta)>\lambda(a_{\eta_i},dE;\eta_i)\ge\lambda(a_{\mu},dE;\mu).
$$
Again by using Lemma \ref{hj5} we may choose integers $a_\nu$ for $\nu\in\Gamma_\theta^{\eta_i}$ so that $\widehat a_\nu\ge 0$ and
$\lambda(a_\nu,dE;\nu)>\lambda(a_{\theta},dE;\theta)$. Subsequently, by applying Lemma \ref{hj5} (and Lemma \ref{hj7}), we may
find a collection of non-negative integers which meets the requirements of \cite[Theorem~1]{HJ}.
Thereby we obtain the desired vector.
\end{proof}
\begin{rem}\label{nUn}
By Equation \eqref{mUm} at the beginning of the proof of Lemma \ref{dM2} we see that
$\xi\in\mathcal H_\mu^{\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}^n}$ if and only if $n\xi\in\mathcal H_\mu^{\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}}$. Thus we may always consider powers $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}^n$ with $n\in\mathbb N$ big
enough to achieve the situation where both $\widehat d$ and $\widehat d_{\langle\widehat d\,\rangle}$ are in $\mathbb N^\Gamma$.
\end{rem}
\begin{lem}\label{b-a}
Let $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ be an ideal with a factorization vector $\widehat d$. Suppose $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ is such that
$\widehat d_{\langle\widehat d\,\rangle}\in\mathbb N^\Gamma$ and let $\fb$ be the ideal corresponding to it. Then $\xi\in\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ if and only if $\xi\in\mathcal H_\mu^\fb$.
\end{lem}
\begin{proof}
If $\xi\in\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$, then there is an antinef divisor $F$ with $\xi_F=\lambda(F,D;\mu)$. It follows from
Lemma \ref{dM} that $\xi=\lambda(F,D;\mu)=\lambda(F,\widehat d_{\langle\widehat d\,\rangle}\widehat E;\mu)\le\lambda(F,\widehat d_{\langle\widehat d\,\rangle}\widehat E;\nu)$, as $(\widehat d_{\langle\widehat d\,\rangle} V)_i\le d_i$ where
the equality holds for $i$ with $d(\mu,i)\le1$. This means that $\xi\in\mathcal H_\mu^\fb$.
If $\xi\in\mathcal H_\mu^\fb$, then $\xi=\lambda(F,\widehat d_{\langle\widehat d\,\rangle};\mu)\le\lambda(F,\widehat d_{\langle\widehat d\,\rangle};\nu)$, then by Lemma \ref{dM2} we have such an
antinef divisor $G$ that $\xi=\lambda(G,D;\mu)\le\lambda(G,D;\nu)$, which shows that $\xi\in\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$.
\end{proof}
\section{Semigroup of values}
Let $\mathcal{SV}_\mu=(\mathcal{SV}_\mu,+)$ be the submonoid of $\mathbb N$ generated by values $V_{\mu,i}$, $i\in\Gamma$. This is
called the value semigroup of $v_\mu$. Recall that if $\Gamma$ is the dual graph of the simple ideal
$\fp_\mu$, then the Zariski exponents are the values of the form $V_{\mu,\tau}$ where $\tau$ is an end (see, e.~g.,
\cite[Remark 6.6]{J}). In general, with any dual graph, we may consider values $V_{\mu,\tau}$ where $\tau$
is an end of the graph. We then get the following:
\begin{prop}
Let $\Gamma$ be a dual graph containing $\mu$. As a submonoid of $\mathbb N$, the semigroup
$\mathcal{SV}_\mu$ is always generated by the set of Zariski exponents of $\mu$, i.~e., the values
$$
\{V_{\mu,\tau_i}\mid i=0\text{, or }i=1,\dots,g+1 \text{ and } \tau_i\neq\mu\},
$$
where $\tau_0$ is the root and the indices $\tau_1,\dots,\tau_{g+1}$ are as in Definition \ref{pair}. In general, we
may write
$$
\mathcal{SV}_\mu = \left\langle V_{\mu,\tau}\mid v_{\Gamma}(\tau)\le1\right\rangle.
$$
\end{prop}
\begin{proof}
If $\mu$ is the only vertex, then $\mathcal{SV}_\mu=\left\langle V_{\mu,\mu}\right\rangle$, and the claim is clear. Suppose that
$\Gamma=\Gamma'(\eta,U)$. If there is $\eta'\in\Gamma$ different from $\eta$, for which $w_\Gamma(\eta')=1$, then we may find
a graph $\Gamma''$ containing $\eta$, for which $\Gamma=\Gamma''(\eta',U')$. Thus we may in this situation choose
$\eta\neq\mu$. But if $\eta\neq\mu$, then
$$
V_{\mu,\eta}=\sum_{\nu\sim\eta}V_{\mu,\nu},
$$
and so the value semigroup remains unchanged under the blowup. Hence we may assume
that $\eta=\mu$ is the only vertex with weight one. This means that $\Gamma$ is the dual graph of
the minimal principalization of the simple ideal corresponding to the vertex $\mu$, and the
claim follows from Lemma \ref{Zar} below.
\end{proof}
Value semigroups are closely related to the jumping numbers. It follows from \cite[Theorem~6.2]{J}
(see also \cite[Lemma 6.1 and Remark 6.6]{J}) that if $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=\fp_\mu$ is simple, then $\xi$
is a jumping number of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ supported at $\mu$ if and only if
$$
\xi V_{\mu,\mu}-q_{\mu,\gamma}-V_{\mu,\tau}\in \left\langle q_{\mu,\gamma}, V_{\mu,\tau}\right\rangle,
$$
where $(\gamma,\tau)$ is the pair associated to $\mu$ (Definition \ref{pair}). Our aim is to generalize this
formula. Observe here that the first one of the two generators is not necessarily in $\mathcal{SV}_\mu$,
but as we shall see in Proposition \ref{Sv} below, $q_{\mu,\gamma}$ is the greatest common divisor of the
values $V_{\mu,i}$ with $i\in\Gamma_\gamma^\mu$, while $V_{\mu,\tau}$ is that of the values coming from the branch $\Gamma_\tau^\mu$.
Let $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ be a complete ideal in $R$ with a dual graph $\Gamma$. For vertices $\mu,\nu\in\Gamma$, let us
define a submonoid $(\mathcal{SV}^\mu_\nu,+)$ of $\mathcal{SV}_\mu$ corresponding to the branch $\Gamma_\nu^\mu$ by setting
$$
\mathcal{SV}^\mu_\nu:=\left\langle V_{\mu,i}\mid i\in\Gamma_\nu^\mu\cup\{\mu\}\right\rangle.
$$
Clearly, $\mathcal{SV}^\mu_\nu=\mathcal{SV}^\mu_\eta$, if $\nu$ and $\eta$ define the same branch. Set $s^\mu_\nu:=\gcd\mathcal{SV}^\mu_\nu$ and write
$(S^\mu,+)$ for the submonoid generated by these numbers, so that
\begin{equation}\label{SSS}
S^\mu:=\left\langle s^\mu_\nu\mid\nu\in\Gamma\right\rangle
\end{equation}
Note that if $\mu$ is the only vertex in $\Gamma$, then $S^\mu=\left\langle s^\mu_\mu\right\rangle=\mathbb N$. Otherwise $S^\mu=\left\langle s^\mu_\nu\mid \nu\sim\mu\right\rangle$.
\begin{lem}\label{Zar}
Let $\Gamma$ be the dual graph of the minimal principalization of a simple ideal
$\fp_\mu$ corresponding to the vertex $\mu$. Let $(\gamma_0,\tau_1),\dots,(\gamma_g,\tau_{g+1})=(\gamma,\tau)$ be the sequence of
pairs associated to $\mu$ and write $\tau_0=\gamma_0$ for the root. Then
$$
\mathcal{SV}^\mu_\gamma =\left\langle V_{\mu,\tau_0},\dots,V_{\mu,\tau_g}\right\rangle\text{ and }
\mathcal{SV}^\mu_\tau=\left\langle V_{\mu,\tau}\right\rangle.
$$
Furthermore, we have $s^\mu_\gamma=q_{\mu,\gamma}$ and $s^\mu_\tau=V_{\mu,\tau}$, and $s^\mu_\gamma s^\mu_\tau=V_{\mu,\mu}$. The greatest integral
multiple of $s^\mu_\gamma$ not in $\mathcal{SV}^\mu_\gamma$ is
$$
\sum_{i=1}^gV_{\mu,\gamma_i}-\sum_{i=0}^gV_{\mu,\tau_i}.
$$
\end{lem}
\begin{proof}
By Equation \eqref{PW} we have for $\nu\in\Gamma\smallsetminus\{\mu\}$
$$
w_{\Gamma}(\nu)V_{\mu,\nu}=\sum_{i\sim\nu}V_{\mu,i}.
$$
Subsequently, we observe that if $\eta$ is such that $d(\nu,\eta)\le1$ and a nonnegative integer $k$
divides $V_{\mu,i}$ for every $i\neq\eta$ with $d(\nu,i)\le1$, then $k$ divides $V_{\mu,\eta}$ also. It follows that
$$
\left\langle V_{\mu,i}\mid i\in\Gamma_{\tau_{n}}^{\gamma_n}\cup\{\gamma_n\} \right\rangle=\left\langle V_{\mu,\tau_{n}} \right\rangle
$$
as semigroups for any $n=0,\dots,g+1$, ($\gamma_{g+1}:=\mu$), so that $V_{\mu,\tau_{n}}$ is the greatest common
divisor of the set. Especially, $\mathcal{SV}_\tau^\mu=\left\langle V_{\mu,\tau}\right\rangle$ so that $s^\mu_\tau= V_{\mu,\tau}$. Let us define a submonoid
$\mathcal W_n$ as
$$
\mathcal W_n:=\left\langle V_{\mu,i}\mid i\in\Gamma_{\tau_0}^{\gamma_{n}}\cup\{\gamma_{n}\}\right\rangle
$$
for $n=0,\dots,g+1$, so that $\mathcal{SV}_\gamma^\mu=\mathcal W_{g+1}$. Let $\Sigma_n$ stand for the greatest integral multiple
of $\gcd\mathcal W_n$ not in $\mathcal W_n$. Clearly, $\mathcal W_0=\left\langle V_{\mu,\tau_0}\right\rangle$ and $\gcd\mathcal W_0=V_{\mu,\tau_0}$ and $\Sigma_0=-V_{\mu,\tau_0}$. To
complete our proof it is enough to show that for $n=0,\dots,g$,
\begin{equation}\label{WW}
\mathcal W_{n+1}=\left\langle \mathcal W_n, V_{\mu,\tau_n}\right\rangle, \gcd\mathcal W_{n+1}=q_{\mu,\gamma_n}
\text{ and }
\Sigma_{n+1}=\Sigma_n+V_{\mu,\gamma_{n}}-V_{\mu,\tau_n}<V_{\mu,\gamma_n}.
\end{equation}
At first, note that the vertices of the set $\left(\Gamma_{\tau_0}^{\gamma_{n+1}}\cup\{\gamma_{n+1}\}\right)\smallsetminus\left(\Gamma_{\tau_0}^{\gamma_{n}}\cup\Gamma_{\tau_n}^{\gamma_n}\right)$ yield a
path $\gamma_{n}=\eta_0^n\sim\cdots\sim\eta^n_{k_n}=\gamma_{n+1}$ for every $n=0,\dots,g$. Thus if $n=0$, it follows
from the observation we made at the beginning of our proof that $V_{\mu,\tau_0}$ divides $V_{\mu,\eta^0_1}$ and
subsequently $V_{\mu,\eta^0_i}$ for every $i=0,\dots,k_0$. Hence $\mathcal W_1=\left\langle V_{\mu,\tau_0}\right\rangle=\left\langle\mathcal W_0, V_{\mu,\tau_0} \right\rangle$ and the
greatest common divisor of the set is $q_{\mu,\gamma_0}= V_{\mu,\tau_0}$. Moreover, $\Sigma_1=\Sigma_0+V_{\mu,\gamma_0}-V_{\mu,\tau_0}=-V_{\mu,\gamma_0}$
so that $\Sigma_1<V_{\mu,\gamma_0}$.
Assume that Equation \eqref{WW} holds if $n<n_0$ for some $n_0\in\{1,\dots,g\}$. Suppose then
that $n=n_0$. Recall that by \cite[Lemma 6.1]{J} (see also \cite[Remark 6.6]{J}) we have
\begin{equation}\label{gcd1}
\gcd\{q_{\mu,\gamma_{n-1}},V_{\mu,\tau_{n}}\}=q_{\mu,\gamma_{n}}.
\end{equation}
Again, by the observation at the beginning of the proof, we see that $q_{\mu,\gamma_{n}}$ divides $V_{\mu,\nu}$
for every $\nu\neq\eta_1^n$ with $d(\gamma_n,\nu)\le1$, and therefore it divides also $V_{\mu,\eta_1^n}$. Subsequently, it
divides every $V_{\mu,\eta_i^n}$ with $i=0,\dots,k_n$. This shows that $\gcd\mathcal W_{n+1}=q_{\mu,\gamma_n}.$
Let us next verify that $\Sigma_{n+1}=\Sigma_{n}+V_{\mu,\gamma_n}-V_{\mu,\tau_n}$ is the greatest integral multiple
of $q_{\mu,\gamma_n}$ not in $\left\langle \mathcal W_n, V_{\mu,\tau_n}\right\rangle$. By \cite[Corollary 3.16]{J} we know that $q_{\mu,\gamma_{n-1}}=q_{\gamma_{n},\gamma_{n-1}}q_{\mu,\gamma_{n}}$.
Therefore by \cite[Lemma 6.1]{J} we get $V_{\mu,\gamma_{n}}=q_{\gamma_{n},\gamma_{n-1}}V_{\mu,\tau_{n}}$, and so
$$
\Sigma_{n+1}=\Sigma_{n}+(q_{\gamma_n,\gamma_{n-1}}-1)V_{\mu,\tau_{n}},
$$
which is clearly an integral multiple of $q_{\mu,\gamma_n}$.
Let $m\in\mathbb N$ be such that $\Sigma_{n+1}+mq_{\mu,\gamma_n}$ is in $\left\langle \mathcal W_n, V_{\mu,\tau_n}\right\rangle$. Equivalently, we may
write $\Sigma_{n+1}+mq_{\mu,\gamma_n}=s+tV_{\mu,\tau_{n}}$ for some $s\in\mathcal W_n$ and $t\in\{0,\dots,q_{\gamma_n,\gamma_{n-1}}-1\}$. Now
$u:=q_{\gamma_n,\gamma_{n-1}}-1-t$ belongs to the same set as $t$ and we may reformulate
$$
s=\Sigma_{n}+uV_{\mu,\tau_{n}}+mq_{\mu,\gamma_n}\in\mathcal W_{n}.
$$
This holds if and only if $uV_{\mu,\tau_{n}}+mq_{\mu,\gamma_n}$ is a positive multiple of $q_{\mu,\gamma_{n-1}}$. Observe that
the map
$$
\varphi:u\mapsto\frac{uV_{\mu,\tau_{n}}}{q_{\mu,\gamma_n}} \mod q_{\gamma_n,\gamma_{n-1}}
$$
is a bijection between the sets $\{0,\dots,q_{\gamma_n,\gamma_{n-1}}-1\}$ and $\mathbb Z_{q_{\gamma_n,\gamma_{n-1}}}$ following from the
fact that $\gcd\{q_{\gamma_n,\gamma_{n-1}},V_{\mu,\tau_{n}}/q_{\mu,\gamma_n}\}=1$. Therefore we may always find an integer $u\in\{0,\dots,q_{\gamma_n,\gamma_{n-1}}-1\}$
so that $\varphi(u)=-m\mod q_{\gamma_n,\gamma_{n-1}}$, but then $uV_{\mu,\tau_n}+mq_{\mu,\gamma_n}$ is
divisible by $q_{\mu,\gamma_{n-1}}$, and it is positive if and only if $m>0$. Thus $\Sigma_{n+1}$ is the greatest
integral multiple of $q_{\mu,\gamma_{n}}$ which does not belong to $\left\langle \mathcal W_{n},V_{\mu,\tau_n}\right\rangle$.
Finally, let us verify that $\Sigma_{n+1}<V_{\mu,\gamma_n}$ and that $\mathcal W_{n+1}=\left\langle\mathcal W_n, V_{\mu,\tau_n} \right\rangle$. By \eqref{BC} we get
$P_{\nu}V_\mu=q_{\mu,\nu}$, and since $PQ=1$ we see that $q_{\mu,\nu}=\sum_{i\succ\nu}q_{\mu,i}+\delta_{\mu,\nu}$ and subsequently
$q_{\mu,\nu}>0$ if $\nu\subset\mu$. Hence $V_{\mu,\nu'}<V_{\mu,\nu}$ if $\nu'\prec\nu\subset\mu$. It follows that $V_{\mu,\nu'}<V_{\mu,\nu}$ if
$\nu'\subset\nu\subset\mu$. Especially, $V_{\mu,\gamma_{n-1}}<V_{\mu,\tau_n}$, but then
$$
\Sigma_{n+1}<V_{\mu,\gamma_{n-1}}+V_{\mu,\gamma_n}-V_{\mu,\tau_n}<V_{\mu,\gamma_{n}}\le V_{\mu,\eta^n_i}
$$
for every $i=0,\dots,k_n$. As $\Sigma_{n+1}$ is the greatest integral multiple of $q_{\mu,\gamma_n}$ not in
$\left\langle\mathcal W_n, V_{\mu,\tau_n} \right\rangle$, we see that $V_{\mu,\eta^n_i}\in \left\langle\mathcal W_n, V_{\mu,\tau_n} \right\rangle$ for every $i=0,\dots,k_n$, but then
$$
\left\langle\mathcal W_n, V_{\mu,\tau_n} \right\rangle\subset\mathcal W_{n+1}\subset\left\langle\mathcal W_n, V_{\mu,\tau_n} \right\rangle,
$$
as wanted. Thus Equation \eqref{WW} holds for any $n=0,\dots,g$, and the claim follows.
\end{proof}
\begin{prop}\label{Sv}
The submonoid $\mathcal{SV}^\mu_\nu$ is generated by those values $V_{\mu,i}$ where $i\in\Gamma_\nu^\mu$ is
an end. More precisely, if $(\gamma,\tau)$ is the pair associated to $\mu$, then the following holds:
$$
\mathcal{SV}^\mu_\nu=
\begin{cases}
\left\langle V_{\mu,\mu}\right\rangle &\text{ if }\Gamma_\nu^\mu \text{ is posterior to } \mu;\\
\left\langle V_{\mu,\tau}\right\rangle &\text{ if }\nu \in \Gamma_\tau^\mu;\\
\left\langle V_{\mu,i}\mid v_{\Gamma}(i)=1, i\in\Gamma_\gamma^\mu\right\rangle &\text{ if }\nu \in
\Gamma_\gamma^\mu.
\end{cases}
$$
We have $s^\mu_\nu=q_{\mu,\gamma}$ if $\nu\in\Gamma_\gamma^\mu$, otherwise $s^\mu_\nu$ is the generator of the submonoid $\mathcal{SV}^\mu_\nu$. For
any $\nu\neq\mu$, the greatest integral multiple of $s^\mu_\nu$ not in $\mathcal{SV}^\mu_\nu$ is
$$
M_\nu=M^\mu_\nu:=\sum_{j\in\Gamma_\nu^\mu}(v_\Gamma(j)-2)V_{\mu,j},
$$
whereas $M_\mu:=-V_{\mu,\mu}$.
\end{prop}
\begin{proof}
Suppose that $\eta\neq\mu$ is a vertex in $\Gamma$ with $w_\Gamma(\eta)=1$. Let $\Gamma'$ be such that
$\Gamma=\Gamma'(\eta,U)$ and suppose that the claim holds for $\Gamma'$. In the case $\nu\in\Gamma'$, let $\mathcal{SV'}^\mu_\nu$ denote
the submonoid generated by the values $V_{\mu,i}$ where $i\in(\Gamma')^\mu_\nu\cup\{\mu\}$. Write $M'_\nu$ for the greatest
integral multiple of $\gcd \mathcal{SV'}^\mu_\nu$ not in $\mathcal{SV'}^\mu_\nu$.
As $U$ consists of vertices adjacent to $\eta$ we observe that if $\eta\in\Gamma_\nu^\mu$, then $U\subset\Gamma_\nu^\mu\cup\{\mu\}$.
By Equation \ref{PW} we then see that if $V_{\mu,i}\in \mathcal{SV'}^\mu_\nu$ for every $i\in U$, then $V_{\mu,\eta}\in \mathcal{SV'}^\mu_\nu$, as
$w_\Gamma(\eta)=1$ and $\delta_{\mu,\eta}=0$. Thus the submonoid $\mathcal{SV}^\mu_\nu$ is the same as the submonoid $\mathcal{SV'}^\mu_\nu$
when $\nu\neq\eta$. Further, $\mathcal{SV}^\mu_\eta=\mathcal{SV}^\mu_\nu$ for any $\nu\in\Gamma_\eta^\mu$, and in the case $\Gamma_\eta^\mu=\{\eta\}$ we see
that $\mathcal{SV}^\mu_\eta=\mathcal{SV'}^\mu_\mu$, as $\eta$ must be a free blowup of $\mu$, in which case Equation \eqref{PW} yields
$V_{\mu,\eta}=V_{\mu,\mu}$. Subsequently, the greatest common divisor of the values from a branch
remains unchanged under the blowup. Obviously, this is also the case with the greatest
of its integral multiples not in the submonoid.
To see that the formula for $M_\nu$ holds, note first that, $M_\nu=M'_\nu$ if $\eta\notin\Gamma_\nu^\mu$. If $\Gamma_{\nu}^\mu=\{\eta\}$,
then $v_{\Gamma}(\eta)=1$ and $M_\eta=(v_{\Gamma}(\eta)-2)V_{\mu,\eta}=-V_{\mu,\eta}$, as wanted, since $\mathcal{SV}^\mu_{\eta}=\left\langle V_{\mu,\eta}\right\rangle$ and
$V_{\mu,\eta}=V_{\mu,\mu}$. Suppose then that $\eta$ is not the only vertex on $\Gamma_\nu^\mu$. As $M_\nu=M_{\nu'}$ for every
$\nu'\in\Gamma_\nu^\mu$, we may assume that $\nu\neq\eta$. Then
$$
M_\nu=M'_\nu+\Delta_\eta,
$$
where
$$
\Delta_\eta:=(v_\Gamma(\eta)-2)V_{\mu,\eta}-\sum_{\mu\neq i\sim\eta}(v_{\Gamma'}(i)-v_{\Gamma}(i))V_{\mu,i}.
$$
If $\eta$ is free, then $v_{\Gamma}(\eta)=1$, but since there is only one vertex $i$ adjacent to it and $\eta$ is
not the only vertex on the branch, we see that $i\neq\mu$, $v_{\Gamma}(i)=v_{\Gamma'}(i)+1$ and $V_{\mu,\eta}=V_{\mu,i}$
by Equation \eqref{PW}. Therefore $\Delta_\eta=0$. If $\eta$ is not free, then $v_{\Gamma}(\eta)=2$ and $v_{\Gamma}(i)=v_{\Gamma'}(i)$
for any $i\sim\eta$. Hence $\Delta_\eta=0$ in any case. This shows that also the numbers $M_\nu$ remain
unchanged under the blowup.
By the above we may blow down $\eta$, and continuing this way, we may eventually
assume that $w_{\Gamma}(i)>1$ for every vertex $i\neq\mu$. This is to say that $\Gamma$ is the dual graph of
the minimal principalization of a simple ideal $\fp_\mu$ corresponding to the vertex $\mu$, but this
case is clear by Lemma \ref{Zar} above.
\end{proof}
\begin{rem}\label{Sgen2}
Because $V_{\mu,\mu}=q_{\mu,\gamma}V_{\mu,\tau}$, we see that $S^\mu=\left\langle q_{\mu,\gamma},V_{\mu,\tau}\right\rangle.$ Furthermore,
$\gcd\{q_{\mu,\gamma},V_{\mu,\tau}\}=q_{\mu,\mu}$ by \eqref{gcd1}. On the other hand, $q_{\mu,\mu}=1$ as easily seen since $Q=P^{-1}$.
It follows that $S^\mu$ is always a numerical semigroup, i.~e., a submonoid of $\bbN$ with finite
complement.
\end{rem}
\section{Main result}
\begin{thm}\label{2}
A positive number $\xi$ is a jumping number in $\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ exactly when
$$
h^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}_\mu(\xi):=d_\mu\xi
+(v_\Gamma(\mu)-2)V_{\mu,\mu}-\sum_{\nu\sim\mu}s^\mu_\nu\left\lceil\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_i V_{\mu,i}}{s^\mu_\nu}\,\xi\right\rceil^{+}\in S^\mu,
$$
where $S^\mu$ is the submonoid of $\mathbb N$ defined by Equation \eqref{SSS} and $\left\lceil\:\: \right\rceil^{+}$ means rounding up to
the nearest positive integer.
\end{thm}
\begin{rem}
This result yields a formula for the set of the jumping numbers of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ supported at $\mu$:
$$
\cH_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d} =\left\{\frac{t}{d_\mu}\middle | t+(v_{\Gamma}(\mu)-2)V_{\mu,\mu}
-\sum_{\nu\sim\mu}s^\mu_\nu\left\lceil t\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_i V_{\mu,i}}{ s^\mu_\nu d_\mu}\right\rceil^{+}\in S^\mu\right\}.
$$
\end{rem}
\begin{rem}
It easily follows from Proposition \ref{Sv} that the numbers $s^\mu_\nu$ and $V_{\mu,\mu}$ present
in the above formula generate $S^\mu$. Moreover,$$s^\mu_\nu=\gcd\left\{V_{\mu,\tau}\mid v_\Gamma(\tau)=1\text{ and }\tau\in\Gamma_\nu^\mu\right\}$$ for
every $\nu\sim\mu$.
\end{rem}
\begin{rem}
By Proposition \ref{Sv} $V_{\mu,\mu}\in S^\mu$ and $s^\mu_\nu=V_{\mu,\mu}$ for any $\nu$ such that $\Gamma^\mu_\nu$ is
posterior to $\mu$. It follows that, for any positive integers $t_\nu$,
$$
\sum_{\nu\sim\mu}s^\mu_\nu t_\nu-(v_\Gamma(\mu)-2)V_{\mu,\mu}\in S^\mu.
$$
Subsequently, if $\xi\in\cH_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$, then $d_\mu\xi\in S^\mu$. Note that the converse is not true: if for
example the ideal in question is the maximal ideal and $\xi=1$, then $\mu=\tau_0$ is the only
vertex of the dual graph of our ideal and $d_\mu=1$. Clearly, $S^\mu=\mathbb N$ so that $d_\mu\xi\in S^\mu$, but
as well known, $1$ is not a jumping number of the maximal ideal, see e.~g. Example \ref{Ex0}
below.
\end{rem}
\begin{proof}[Proof of Theorem \ref{2}]
Note first that the case $\mu$ is the only vertex of the dual graph is
trivial and is dealt with in Example \ref{Ex0} below. Thus we may assume $\mu$ has adjacent
vertices. Obviously, by Remark \ref{DualG} we could also choose any dual graph containing $\mu$ and
having other vertices, too. To begin with, observe that
$$
h^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}_\mu(\xi)=\frac{nd_\mu\xi}{n}+(v_{\Gamma}(\mu)-2)V_{\mu,\mu}
-\sum_{\nu\sim\mu}s^\mu_\nu\left\lceil\sum_{i\in\Gamma_\nu^\mu}\frac{n\widehat d_i V_{\mu,i}}{s^\mu_\nu}\cdot\frac{\xi}{n}\right\rceil^{+}\hspace{-3pt}
=h^{\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}^n}_\mu\hspace{-3pt}\left(\textstyle\frac{\xi}{n}\right)
$$
Hence $h^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}_\mu(\xi)\in S^\mu$ exactly, when $h^{\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}^n}_\mu\hspace{-3pt}\left(\textstyle\frac{\xi}{n}\right)\in S^\mu$. Together with Remark \ref{nUn} this shows
that, by considering $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}^n$ with $n$ big enough, we may assume $\widehat d_{\langle\widehat d\,\rangle}\in\mathbb N^\Gamma$.
Let $\fb$ be the ideal having the factorization vector $\widehat d_{\langle\widehat d\,\rangle}$. Note that $\fb$ is a product of
ideals $\fp_\nu$ where $d(\nu,\mu)\le1$, and we may regard $\Gamma$ as a dual graph and $V$ as a valuation
matrix of $\fb$ (see Remark \ref{DualG}). According to Lemma \ref{dM}
$$
(\widehat d_{\langle\widehat d\,\rangle} V)_\nu=d_\nu
\text{ and }
\sum_{i\in\Gamma_\nu^\mu}\left(\widehat d_{\langle\widehat d\,\rangle}\right)_i V_{\mu,i}=\sum_{i\in\Gamma_\nu^\mu}\widehat d_iV_{\mu,i}
$$
for $\nu$ with $d(\mu,\nu)\le1$. This shows, together with Lemma \ref{b-a}, that we may assume $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=\fb$,
i.~e., $\widehat d_\nu=0$ unless $d(\mu,\nu)\le1$. Thereby
$$
\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_i V_{\mu,i}}{s^\mu_\nu}\,\xi=\frac{\widehat d_\nu V_{\mu,\nu}}{s^\mu_\nu}\,\xi
$$
for every $\nu\sim\mu$.
Set $\widehat g:=\xi\widehat d+\widehat\kappa$, where $\widehat\kappa:=(v_\Gamma(\nu)-2)_{\nu\in\Gamma}$. Observe that because
$$
(1,\dots,1)P^\textsc t P=(w_{\Gamma}(\nu)-v_{\Gamma}(\nu))_{\nu\in\Gamma}
$$
and since $\widehat k=(2-w_{\Gamma}(\nu))_{\nu\in\Gamma}$ by \eqref{K}, we get $k_\nu+1=-(\widehat\kappa V)_\nu$. Subsequently, we obtain
$$
\lambda(\widehat g\widehat E, dE;\nu)=\frac{\xi d_\nu+(\widehat\kappa V)_\nu+k_\nu+1}{d_{\nu}}=\xi
$$
for any $\nu\in\Gamma$. Consider a vector$\widehat \phi:= \widehat g_{\langle\widehat\kappa\rangle[\zeta]}$where $\zeta_\nu:=0$ unless $\nu\sim\mu$ in which case
\begin{equation*}
\label{KD}
\zeta_\nu
:=\frac{s^\mu_\nu}{V_{\mu,\nu}}\left(\left\lceil\frac{\widehat d_\nu V_{\mu,\nu}}{s^\mu_\nu}\,\xi\right\rceil^+
-\frac{\widehat d_\nu V_{\mu,\nu}}{s^\mu_\nu}\,\xi\right).
\end{equation*}
Let $M_\nu$ be as in Proposition \ref{Sv}. By using Lemma \ref{dM} a direct calculation shows that
\begin{equation}\label{phihmu}
\widehat\phi_\mu=\xi\widehat d_\mu+(v_\Gamma(\mu)-2)-\sum_{\nu\sim\mu}\rho_{[\mu,\nu]}(\mu)\zeta_\nu=\frac{h^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}_\mu(\xi)}{V_{\mu,\mu}}
\end{equation}
and
\begin{equation}\label{phnuD}
\widehat\phi_\nu
=\xi\widehat d_\nu+\zeta_\nu+\sum_{i\in\Gamma_\nu^\mu}\rho_{[\nu,i]}(\mu)(v_\Gamma(i)-2)
=\frac{M_\nu}{V_{\mu,\nu}}+\frac{s^\mu_\nu}{V_{\mu,\nu}}\left\lceil\frac{\widehat d_\nu V_{\mu,\nu}}{s^\mu_\nu}\xi\right\rceil^+.
\end{equation}
Observe that since $\widehat\phi=(\xi\widehat d+\widehat\kappa_{\langle\widehat \kappa\,\rangle})_{[\zeta]}$, we have $\widehat\phi_\nu=0$ when $d(\mu,\nu)>1$. Furthermore,
as $\zeta_\nu\ge0$ where the equality may take place only if $\widehat d_\nu>0$, we know by Lemma \ref{dM} that
$(\widehat\phi V)_\mu=(\widehat g V)_\mu$ and
$$
(\widehat\phi V)_\nu\ge(\widehat g V)_\nu
$$
for every $\nu$ with $\nu\sim\mu$, where the equality may take place only if $\widehat d_\nu>0$.
Suppose now that $\xi\in\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$. By \cite[Theorem 1 and Lemma 6]{HJ} we have $\widehat f\in\mathbb N^\Gamma$ satisfying
$\widehat f_\nu>0$ only if $\nu$ is an end different from $\mu$, and further, $\lambda((\widehat f V)_\mu,\mu)=\xi$ and $\lambda((\widehat f V)_\nu,\nu)$
is increasing on every path going away from $\mu$. Because also $\lambda((\widehat\phi V)_\mu,\mu)=\lambda((\widehat g V)_\mu,\mu)=\,\xi$,
we get by applying Lemma \ref{dM0} $\widehat\phi^\mathcal N=\widehat f^\mathcal N$, and further
$$
\widehat\phi^\mathcal N_{\mu}V_{\mu,\mu}
=\left(\widehat\phi^\mathcal N V\right)_{\mu}
=\left(\widehat f^\mathcal N V\right)_{\mu}
=\sum_{\nu\sim\mu}\sum_{i\in\Gamma_\nu^\mu}\widehat f_i V_{\mu,i}.
$$
Thus Equation \eqref{phihmu} gives
$$
h^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}_\mu(\xi)=\left(\widehat\phi^\mathcal N_{\mu}-\sum_{\nu\sim\mu}\rho_{[\mu,\nu]}(\mu)\widehat\phi_\nu\right) V_{\mu,\mu}
=\sum_{\nu\sim\mu}\left(\sum_{i\in\Gamma_\nu^\mu}\widehat f_i V_{\mu,i}-\widehat\phi_\nu V_{\mu,\nu}\right).
$$
On the other hand, according to Equation \eqref{phnuD} and Proposition \ref{Sv}, $\widehat\phi_\nu V_{\mu,\nu}$ is the least
integral multiple of $s^\mu_\nu$, for which
$$
\widehat\phi_\nu V_{\mu,\nu}\ge M_\nu+\widehat d_\nu V_{\mu,\nu}\xi,
$$
where the inequality is strict if $\widehat d_\nu=0$. Observe that Proposition \ref{Sv} yields
$$
M_\nu+\widehat d_\nu V_{\mu,\nu}\xi
=\sum_{j\in\Gamma_\nu^\mu}(v_\Gamma(j)-2)V_{\mu,j}+\widehat d_\nu V_{\mu,\nu}\xi
=(\widehat g_{\langle\widehat\kappa\rangle})_\nu V_{\mu,\nu},
$$
where the last equality follows from Lemma \ref{dM}. Since every $V_{\mu,i}$ is divisible by $s^\mu_\nu$ for
$i\in\Gamma_\nu^\mu$, we see that $h^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}_\mu(\xi)\in S^\mu$, if for every $\nu\sim\mu$ holds
\begin{equation}
\label{fph}
\sum_{i\in\Gamma_\nu^\mu}\widehat f_i V_{\mu,i}\ge(\widehat g_{\langle\widehat\kappa\rangle})_\nu V_{\mu,\nu}.
\end{equation}
Observe that in the case $\widehat d_\nu=0$ the right hand side is equal to $M_\nu$ which is not in $\mathcal{SV}^\mu_\nu$
by Proposition \ref{Sv}, while the left hand side clearly is in $\mathcal{SV}^\mu_\nu$. Thus the inequality must
be strict in this case.
As we saw above, $\widehat g^\mathcal N = \widehat f^\mathcal N$ and $(\widehat g_{\langle\,\widehat\kappa\,\rangle})_i=0$ unless $d(\mu,i)\le1$. Therefore
$\widehat g_{\langle\,\widehat\kappa\,\rangle}=\widehat f^\mathcal N_{[\,\widehat g_{\langle\,\widehat\kappa\,\rangle}\,]}$, and $\widehat f_{\langle\,\widehat f\,\rangle}=\widehat f^\mathcal N_{[\,\widehat f_{\langle\,\widehat f\,\rangle}\,]}$ follows from Remark \ref{Fuf}. Recall that for any $\nu\sim\mu$
$$
\lambda((\widehat fV)_\nu,\nu)\ge\xi=\lambda((\widehat gV)_\nu,\nu).
$$
Hence, by applying Lemma \ref{dM}, we obtain for any $\nu\sim\mu$
$$
\lambda((\widehat f^\mathcal N_{[\,\widehat f_{\langle\,\widehat f\,\rangle}\,]}V)_\nu,\nu)
=\lambda((\widehat f_{\langle\,\widehat f\,\rangle}V)_\nu,\nu)
\ge\lambda((\widehat g_{\langle\,\widehat\kappa\,\rangle}V)_\nu,\nu)
=\lambda((\widehat f^\mathcal N_{[\,\widehat g_{\langle\,\widehat\kappa\,\rangle}\,]}V)_\nu,\nu),
$$
so that
\begin{align*}
\left(\widehat f^\mathcal N_{[\,\widehat f_{\langle\,\widehat f\,\rangle}\,]}V\right)_\nu
&=\left(\widehat f^\mathcal NV\right)_{\nu}+\sum_{i\sim\mu}\left(\widehat f_{\langle\,\widehat f\,\rangle}\right)_i\varphi_\mu^{[\mu,i]}(\nu)V_{\mu,\nu}\\
\ge
\left(\widehat f^\mathcal N_{[\,\widehat g_{\langle\,\widehat\kappa\,\rangle}\,]}V\right)_\nu
&=\left(\widehat f^\mathcal NV\right)_{\nu}+\sum_{i\sim\mu}\left(\widehat g_{\langle\,\widehat\kappa\,\rangle}\right)_i\varphi_\mu^{[\mu,i]}(\nu)V_{\mu,\nu}
\end{align*}
According to Proposition \ref{D} $\varphi_\mu^{[\mu,i]}(\nu)\ge0$ for $\nu\sim\mu$, where the equality takes place always
if $\mu\sim i\neq\nu$. Subsequently,
$$
\left(\widehat f_{\langle\,\widehat f\,\rangle}\right)_\nu\varphi_\mu^{[\mu,\nu]}(\nu)V_{\mu,\nu}\\
\ge
(\widehat g_{\langle\,\widehat\kappa\,\rangle})_\nu\varphi_\mu^{[\mu,\nu]}(\nu)V_{\mu,\nu}
$$
which yields Inequality \eqref{fph}. Thus $h^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}_\mu(\xi)\in S^\mu$.
Suppose next that $h_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}(\xi)\in S^\mu$. Then we may take such non-negative integers $m_\nu$ for
$\nu\sim\mu$, that
$$
h_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}(\xi)=\sum_{\nu\sim\mu}m_\nu s^\mu_\nu.
$$
Let $w:=\sum_{\nu\sim\mu}(m_\nu s^\mu_\nu/V_{\mu,\nu})\mathbf 1_\nu$ and define $\widehat\psi:=\widehat\phi_{\left[\,w\,\right]}$. Subsequently, we obtain by using
Equation \eqref{phihmu} and Lemma \ref{dM}
$$
\widehat \psi_\mu=\frac{h_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}(\xi)}{V_{\mu,\mu}}-\sum_{\nu\sim\mu}\frac{m_\nu s^\mu_\nu}{V_{\mu,\nu}}\frac{V_{\mu,\nu}}{V_{\mu,\mu}}=0,
$$
while for $\nu\sim\mu$ Equation \eqref{phnuD} together with Lemma \ref{dM} yields
$$
\widehat \psi_\nu
=\frac{s^\mu_\nu}{V_{\mu,\nu}}\left\lceil\frac{\widehat d_\nu V_{\mu,\nu}}{s^\mu_\nu}\,\xi\right\rceil^+
+\frac{M_\nu}{V_{\mu,\nu}}+\frac{m_\nu s^\mu_\nu}{V_{\mu,\nu}}
$$
Clearly, $\widehat\psi_\nu=0$ unless $\nu\sim\mu$, and further, since $s^\mu_\nu$ divides $\widehat\psi_\nu V_{\mu,\nu}$ and $\widehat\psi_\nu V_{\mu,\nu}>M_\nu$
we observe by Lemma \ref{Sv} that $\widehat\psi_\nu V_{\mu,\nu}\in \mathcal{SV}^\mu_\nu$ for every $\nu\sim\mu$. Therefore we may find
$\widehat f\in\mathbb N^\Gamma$ with $\widehat f_i>0$ only if $v_\Gamma(i)=1$ and $i\neq\mu$ satisfying
$$
\left(\widehat f_{\langle\,\widehat f\,\rangle}\right)_\nu=\widehat\psi_\nu
$$
for every $\nu\sim\mu$. It follows from Lemma \ref{dM} that for any $\nu$ with $d(\mu,\nu)\le1$
$$
\lambda((\widehat fV)_\nu,\nu)=\lambda((\widehat\psi V)_\nu,\nu)\ge\lambda((\widehat\phi V)_\nu,\nu)\ge \xi,
$$
where the equality holds for $\nu=\mu$ and otherwise it may take place only if $\widehat d_\nu>0$.
Subsequently, by choosing $a_i=\widehat fV_i$ for every $i$ with $d(i,\mu)\le1$, we may by using
\cite[Lemma~5 and Lemma~6]{HJ} achieve a connected set $U\subset\{\nu\in\Gamma\mid d(\nu,\mu)\le1\}$ and non-negative
integers $a_i$ for every $i\in\Gamma$ with $d(i,U)\le1$ satisfying the conditions of
\cite[Theorem~1]{HJ}. Thereby $\xi\in\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$.
\end{proof}
\section{Examples}
\subsubsection*{Cases of low valence}
\begin{exmp}\label{Ex0}
In the case $v_{\Gamma}(\mu)=0$ we have only one vertex $\mu$. Moreover, $\widehat d_\mu>0$ while
$V_{\mu,\mu}=1$ and thereby $d_{\mu}=\widehat d_\mu$. As the set of vertices adjacent to $\mu$ is empty, the set $S^\mu$
is $\mathbb N$. The claim of Theorem \ref{2} now says that $\xi$ is a jumping number if and only if
$$
\widehat d_\mu\xi-2\in\mathbb N.
$$
But this already follows from \cite[Theorem 6.2]{J} and Remark \ref{nUn}.
\end{exmp}
\begin{exmp}\label{Ex1}
Suppose $v_{\Gamma}(\mu)=1$. Let $\nu$ be the vertex adjacent to $\mu$. Then
$$
d_\mu\xi=\widehat d_\mu V_{\mu,\mu}\xi+\sum_{i\in\Gamma_\nu^\mu}s^\mu_\nu\frac{\widehat d_iV_{\mu,i}}{s^\mu_\nu}\xi.
$$
Since $(v_\Gamma(\mu)-2)V_{\mu,\mu}=-V_{\mu,\mu}$, we get by Theorem \ref{2}
$$
h_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}(\xi)=d_\mu\xi-V_{\mu,\mu}-s^\mu_\nu\left\lceil\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_iV_{\mu,i}}{s^\mu_\nu}\xi\right\rceil^+
$$
Putting these together shows that
$$
h_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}(\xi)
=(\widehat d_\mu\xi-1)V_{\mu,\mu}
+s^\mu_\nu\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_iV_{\mu,i}}{s^\mu_\nu}\xi
-s^\mu_\nu\left\lceil\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_iV_{\mu,i}}{s^\mu_\nu}\xi\right\rceil^+ <0
$$
always unless $\widehat d_\mu\xi\ge1$. Especially, $\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ is empty if $\widehat d_\mu=0$.
\end{exmp}
\begin{exmp}\label{Ex2}
Suppose $v_{\Gamma}(\mu)=2$. Since $(v_\Gamma(\mu)-2)V_{\mu,\mu}=0$, we get
\begin{align*}
h_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}(\xi)
&=d_\mu\xi-\sum_{\nu\sim\mu}s^\mu_\nu\left\lceil\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_iV_{\mu,i}}{s^\mu_\nu}\xi\right\rceil^+\\
&=\sum_{i\in\Gamma}\widehat d_i V_{\mu,i}\xi
-\sum_{\nu\sim\mu}s^\mu_\nu\left\lceil\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_iV_{\mu,i}}{s^\mu_\nu}\xi\right\rceil^+\\
&=\widehat d_\mu V_{\mu,\mu}\xi
+\sum_{\nu\sim\mu}\sum_{i\in\Gamma_\nu^\mu}\widehat d_i V_{\mu,i}\xi
-\sum_{\nu\sim\mu}s^\mu_\nu\left\lceil\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_iV_{\mu,i}}{s^\mu_\nu}\xi\right\rceil^+\\
&=\widehat d_\mu V_{\mu,\mu}\xi
+\sum_{\nu\sim\mu}s^\mu_\nu\left(\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_i V_{\mu,i}}{s^\mu_\nu}\xi
-\left\lceil\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_iV_{\mu,i}}{s^\mu_\nu}\xi\right\rceil^+\right).
\end{align*}
Suppose $\xi\in\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$. Then $h_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}(\xi)\ge0$, which implies that either $\widehat d_\mu>0$ or $\widehat d_\mu=0$ and
$$
\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_i V_{\mu,i}}{s^\mu_\nu}\xi=\left\lceil\sum_{i\in\Gamma_\nu^\mu}\frac{\widehat d_iV_{\mu,i}}{s^\mu_\nu}\xi\right\rceil^+
$$
for each $\nu\sim\mu$. Note that the latter is possible only if for each $\nu\sim\mu$ there is such $i\in\Gamma_\nu^\mu$
that $\widehat d_i>0$.
Let us then assume that $\widehat d_\mu=0$, and let $\widehat f$ be such that
\begin{equation}
\label{llx}
\lambda((\widehat f V)_\eta,\eta)\ge\lambda((\widehat f V)_\mu,\mu)=\xi.
\end{equation}
By Equation \eqref{PW} we have $w_\Gamma(\mu)V_{\mu,i}=\sum_{\nu\sim\mu}V_{\nu,i}+\delta_{\mu,i}$ for every $i\in\Gamma$. Moreover, by
Equations \eqref{P} and \eqref{K}
$$
w_\Gamma(\mu)(k_\mu+1)-\sum_{\nu\sim\mu}(k_\nu+1)=\widehat k_\mu=2-v_\Gamma(\mu)=0.
$$
Thereby
\begin{align*}
\lambda((\widehat f V)_\mu,\mu)
&=\frac{(\widehat f V)_\mu+k_\mu+1}{(\widehat d V)_\mu}\\
&=\frac{w_\Gamma(\mu)\left((\widehat f V)_\mu+k_\mu+1)\right)}{w_\Gamma(\mu)(\widehat d V)_\mu}\\
&=\frac{\sum_{\nu\sim\mu}(\widehat f V)_\nu+\widehat f_\mu+\sum_{\nu\sim\mu}(k_\nu+1)}{\sum_{\nu\sim\mu}(\widehat d V)_\nu}\\
&=\frac{(\widehat f V)_{\nu_1}+k_{\nu_1}+1+\widehat f_\mu+(\widehat f V)_{\nu_2}+k_{\nu_2}+1}
{\ \ (\widehat d V)_{\nu_1}\ \ \ \ \ \ \ \ +\ \ \ \ \ \ \ \ \ \ (\widehat d V)_{\nu_2}},
\end{align*}
where $\nu_1\sim\mu\sim\nu_2$. Furthermore, since we may assume that
$$
\lambda((\widehat f V)_{\nu_1},\nu_1)\le\lambda((\widehat f V)_{\nu_2},\nu_2),
$$
this shows that
$$
\lambda((\widehat f V)_{\nu_1},\nu_1)\le\lambda((\widehat f V)_{\mu},\mu)\le\lambda(\widehat f_\mu+(\widehat f V)_{\nu_2},\nu_2),
$$
where the equality holds on the left if and only if it holds on the right. Putting these
together with \eqref{llx} we observe that $\widehat f_\mu=0$ and
$$
\lambda((\widehat f V)_{\nu_1},\nu_1)=\xi=\lambda((\widehat f V)_{\nu_2},\nu_2).
$$
This is to say that both the vertices adjacent to $\mu$ support $\xi$. This means, informally
speaking, that $\mu$ doesn't support jumping numbers independently. Especially, $\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ is
empty if there is $\nu$ such that $\widehat d_i=0$ whenever $i\in\Gamma_\mu^\nu$.
\end{exmp}
\subsubsection*{Simple ideals}
\begin{exmp}
Suppose $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ is a simple ideal. If $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ is the maximal ideal, then the dual graph
consists of one vertex, but this case is already discussed in Example \ref{Ex0}. Thus we may
assume $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ is different from the maximal ideal. Since $\widehat d_i>0$ for only one vertex $i$, the
examples \ref{Ex1} and \ref{Ex2} show that if $v_\Gamma(\mu)<3$ and $\widehat d_\mu=0$ then $\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ is empty. Thus we may
suppose that $\mu$ is a vertex with $\widehat d_\mu=1$ or $v_\Gamma(\mu)=3$. Note that if $\widehat d_\mu=1$, then $v_\Gamma(\mu)<3$
and if $v_\Gamma(\mu)=3$ then $\widehat d_\mu=0$.Let $(\gamma,\tau)$ be the pair associated to $\mu$.
Consider first the case $\widehat d_\mu=1$ and $v_\Gamma(\mu)=1$. Then $\Gamma_\gamma^\mu$ is the only branch emanating
from $\mu$. By using Proposition \ref{Sv} we obtain $s^\mu_\gamma=q_{\mu,\gamma}$, and by applying equation $PQ=1$,
we see that $q_{\mu,\gamma}=q_{\mu,\mu}=1$. Thus $S^\mu=\mathbb N$, and by Theorem \ref{2}, $\xi\in\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ if and only if
$$
(\xi-1)d_\mu-1\in\mathbb N,\text{ i.~e., }\xi\in\left\{1+\frac{t+1}{d_\mu}\middle| t\in\mathbb N\right\}.
$$
Consider next the case $\widehat d_\mu=1$ and $v_\Gamma(\mu)=2$. It follows from Proposition \ref{Sv} that
$s^\mu_\gamma=a:=q_{\mu,\gamma}$. Furthermore, $s^\mu_{\tau}=b:=V_{\mu,\tau}$ and $d_\mu=ab$. Now $S^\mu=\left\langle a, b\right\rangle$, and
Theorem \ref{2} says that $\xi\in\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ if and only if
$$
ab\xi-a-b\in \left\langle a, b\right\rangle,\text{ i.~e., }\xi\in\left\{\frac{s+1}{a}+\frac{t+1}{b}\middle| s,t\in\mathbb N\right\}.
$$
In the case $v_\Gamma(\mu)=3$ we see, again by Proposition \ref{Sv}, that $s^\mu_\gamma=a:=q_{\mu,\gamma}$ and
$s^\mu_{\tau}=b:=V_{\mu,\tau}$ and $V_{\mu,\mu}=ab$, while $s^\mu_{\eta}=V_{\mu,\mu}$, where $\eta$ is the vertex corresponding to $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$.
According to Theorem \ref{2} $\xi\in\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ if and only if
$$
V_{\mu,\eta}\xi+V_{\mu,\mu}-a-b-V_{\mu,\mu}\left\lceil\frac{V_{\mu,\eta}\xi}{V_{\mu,\mu}}\right\rceil\in\left\langle a, b\right\rangle.
$$
Now $V_{\mu,\eta}/V_{\mu,\mu}=c:=q_{\eta,\mu}$, so the above is equivalent to
$$
\xi-\frac{\left\lceil c\xi\right\rceil-1}{c}=\frac{s+1}{ac}+\frac{t+1}{bc}\text{ for some }s,t\in\mathbb N.
$$
Obviously, the equation holds for $\xi+\frac{1}{c}$ if it holds for $\xi$. Subsequently, $\xi\in\mathcal H_\mu^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ if and
only if
$$
\xi\in\left\{\frac{s+1}{ac}+\frac{t+1}{bc}+\frac{m}{c}\middle|s,t,m\in\mathbb N,\frac{s+1}{ac}+\frac{t+1}{bc}\le\frac{1}{c}\right\}.
$$
Observe that $ac=q_{\eta,\gamma}$ and $bc=V_{\mu,\tau}q_{\eta,\mu}=V_{\eta,\tau}$ by, e.~g., \cite[Proposition 3.13]{J}.
The above shows that Theorem \ref{2} gives an alternative proof of the formula for jumping
numbers of a simple ideal (see \cite[Theorem 6.2]{J}).
\end{exmp}
\subsubsection*{General case}
\begin{exmp}\label{generalexample}
Let $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=\fp_1^2\fp_2^{\phantom{2}}\fp_3^2\fp_4^{\phantom{2}}\fp_5^3$ be an ideal, where
\begin{align*}
\fp_1&=\langle\overline{\phantom{^|}x^3y^3(x^3-y^2)\,,\,(x^3-y^2)^3+x^{11}\phantom{^|}}\rangle,\phantom{\sum^a}\\
\fp_2&=\langle\overline{\phantom{^|}x^2y^3 \,,\,(x^3-y^2)^2 \phantom{^|}}\rangle,\phantom{\sum^a}\\
\fp_3&=\langle\overline{\phantom{^|}xy^5 \,,\, x^3-y^7 \phantom{^|}}\rangle,\phantom{\sum^a}\\
\fp_4&=\langle\overline{\phantom{^|}x^{10} \,,\,(x^3-(x-y)^2)^3 \phantom{^|}}\rangle,\phantom{\sum^a}\\
\fp_5&=\langle\overline{\phantom{^|}y^2 \,,\, x-y \phantom{^|}}\rangle.\phantom{\sum^a}
\end{align*}
The dual graph of principalization of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ is as follows:
\unitlength=0.92mm
\begin{picture}(70,120)(-6,-27)
\put(-6,86){\makebox(0,0){\textbf{$\Gamma$}:}}
\put(3,24.5){\circle{6}}
\put(2,45){\circle{6}}
\put(2,75){\circle{6}}
\put(20,-15){\circle{6}}
\put(20,0){\circle{6}}
\put(20,15){\circle{6}}
\put(20,30){\circle{6}}
\put(20,45){\circle{6}}
\put(20,60){\circle{6}}
\put(20,75){\circle{6}}
\put(35,0){\circle{6}}
\put(40,15){\circle{6}}
\put(40,30){\circle{6}}
\put(50,0){\circle{6}}
\put(55,45){\circle{6}}
\put(70,30){\circle{6}}
\put(90,0){\circle{6}}
\put(90,15){\circle{6}}
\put(90,30){\circle{6}}
\put(110,30){\circle{6}}
\put(57.13,47.13){\vector(1,1){4}}
\put(22.13,77.13){\vector(1,1){4}}
\put(112.13,32.13){\vector(1,1){4}}
\put(112.13,27.87){\vector(1,-1){4}}
\put(0.87,26.63){\vector(-1,1){4}}
\put(0,24.5){\vector(-1,0){4}}
\put(0.87,22.37){\vector(-1,-1){4}}
\put(22.13,-17.13){\vector(1,-1){4}}
\put(17.87,-17.13){\vector(-1,-1){4}}
\put(23,0){\line(1,0){9}}
\put(20,-3){\line(0,-1){9}}
\put(20,3){\line(0,1){9}}
\put(38,0){\line(1,0){9}}
\put(20,27){\line(0,-1){9}}
\put(20,33){\line(0,1){9}}
\put(20,57){\line(0,-1){9}}
\put(20,63){\line(0,1){9}}
\put(23,30){\line(1,0){14}}
\put(73,30){\line(1,0){14}}
\put(93,30){\line(1,0){14}}
\put(40,27){\line(0,-1){9}}
\put(90,3){\line(0,1){9}}
\put(90,27){\line(0,-1){9}}
\put(17,75){\line(-1,0){12}}
\put(17,45){\line(-1,0){12}}
\put(42.13,32.13){\line(1,1){10.5}}
\put(67.87,32.13){\line(-1,1){10.5}}
\put(17.3,29.1){\line(-3,-1){11.5}}
\begin{tiny}
\put(3,24.5){\makebox(0,0){$34$}}
\put(2,45){\makebox(0,0){$34$}}
\put(2,75){\makebox(0,0){$70$}}
\put(20,-15){\makebox(0,0){$119$}}
\put(20,0){\makebox(0,0){$117$}}
\put(20,15){\makebox(0,0){$37$}}
\put(20,30){\makebox(0,0){$31$}}
\put(20,45){\makebox(0,0){$68$}}
\put(20,60){\makebox(0,0){$139$}}
\put(20,75){\makebox(0,0){$210$}}
\put(40,15){\makebox(0,0){$39$}}
\put(40,30){\makebox(0,0){$78$}}
\put(55,45){\makebox(0,0){$164$}}
\put(70,30){\makebox(0,0){$85$}}
\put(90,0){\makebox(0,0){$87$}}
\put(90,15){\makebox(0,0){$174$}}
\put(90,30){\makebox(0,0){$261$}}
\put(110,30){\makebox(0,0){$263$}}
\put(35,0){\makebox(0,0){$78$}}
\put(50,0){\makebox(0,0){$39$}}
\put(24,34){\makebox(0,0){$\gamma_1$}}
\put(44,19){\makebox(0,0){$\gamma_2$}}
\put(46,30){\makebox(0,0){$\gamma_3$}}
\put(74,34){\makebox(0,0){$\gamma_4$}}
\put(94,4){\makebox(0,0){$\gamma_5$}}
\put(94,19){\makebox(0,0){$\gamma_6$}}
\put(94,34){\makebox(0,0){$\gamma_7$}}
\put(116,30){\makebox(0,0){$\gamma_8$}}
\put(51,49){\makebox(0,0){$\gamma_9$}}
\put(24,19){\makebox(0,0){$\gamma_{10}$}}
\put(54,4){\makebox(0,0){$\gamma_{11}$}}
\put(39,4){\makebox(0,0){$\gamma_{12}$}}
\put(24,4){\makebox(0,0){$\gamma_{13}$}}
\put(24,-11){\makebox(0,0){$\gamma_{14}$}}
\put(6,49){\makebox(0,0){$\gamma_{15}$}}
\put(24,49){\makebox(0,0){$\gamma_{16}$}}
\put(6,79){\makebox(0,0){$\gamma_{17}$}}
\put(24,64){\makebox(0,0){$\gamma_{18}$}}
\put(26,75){\makebox(0,0){$\gamma_{19}$}}
\put(5,30){\makebox(0,0){$\gamma_{20}$}}
\end{tiny}
\end{picture}
\noindent
The factorization vector of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ is
$$
(0,0,0,0,0,0,0,\mathbf{2},\mathbf{1},0,0,0,0,\mathbf{2},0,0,0,0,\mathbf{1},\mathbf{3}),
$$
and
\vspace{3pt}
$$\left[
\begin{smallmatrix}
\vspace{3pt}
\hspace{1.5pt}\phantom{\widehat t}1\hspace{4pt} &\hspace{4pt}1\hspace{4pt} &\hspace{4pt}2\hspace{4pt} &\hspace{4pt}2
\hspace{4pt} &\hspace{4pt}2\hspace{4pt} &\hspace{4pt}4\hspace{4pt} &\hspace{4pt}6\hspace{4pt} &
\hspace{4pt}6\hspace{4pt} &\hspace{4pt}4\hspace{4pt} &\hspace{4pt}1\hspace{4pt} &\hspace{4pt}1
\hspace{4pt} &\hspace{4pt}2\hspace{4pt} &\hspace{4pt}3\hspace{4pt} &\hspace{4pt}3\hspace{4pt} &
\hspace{4pt}1\hspace{4pt} &\hspace{4pt}2\hspace{4pt} &\hspace{4pt}2\hspace{4pt} &\hspace{4pt}4
\hspace{4pt} &\hspace{4pt}6\hspace{4pt} &\hspace{4pt}1\hspace{4pt}\\
\vspace{3pt}1 &2 &3 &3 &3 &6 &9 &9 &6 &1 &1 &2 &3 &3 &1 &2 &2 &4 &6 &1\\
\vspace{3pt}2 &3 &6 &6 &6 &12 &18 &18 &12 &2 &2 &4 &6 &6 &2 &4 &4 &8 &12 &2\\
\vspace{3pt}2 &3 &6 &7 &7 &14 &21 &21 &13 &2 &2 &4 &6 &6 &2 &4 &4 &8 &12 &2\\
\vspace{3pt}2 &3 &6 &7 &8 &15 &22 &22 &13 &2 &2 &4 &6 &6 &2 &4 &4 &8 &12 &2\\
\vspace{3pt}4 &6 &12 &14 &15 &30 &44 &44 &26 &4 &4 &8 &12 &12 &4 &8 &8 &16 &24 &4\\
\vspace{3pt}6 &9 &18 &21 &22 &44 &66 &66 &39 &6 &6 &12 &18 &18 &6 &12 &12 &24 &36 &6\\
\vspace{3pt}6 &9 &18 &21 &22 &44 &66 &67 &39 &6 &6 &12 &18 &18 &6 &12 &12 &24 &36 &6\\
\vspace{3pt}4 &6 &12 &13 &13 &26 &39 &39 &26 &4 &4 &8 &12 &12 &4 &8 &8 &16 &24 &4\\
\vspace{3pt}1 &1 &2 &2 &2 &4 &6 &6 &4 &2 &2 &4 &6 &6 &1 &2 &2 &4 &6 &1\\
\vspace{3pt}1 &1 &2 &2 &2 &4 &6 &6 &4 &2 &3 &5 &7 &7 &1 &2 &2 &4 &6 &1\\
\vspace{3pt}2 &2 &4 &4 &4 &8 &12 &12 &8 &4 &5 &10 &14 &14 &2 &4 &4 &8 &12 &2\\
\vspace{3pt}3 &3 &6 &6 &6 &12 &18 &18 &12 &6 &7 &14 &21 &21 &3 &6 &6 &12 &18 &3\\
\vspace{3pt}3 &3 &6 &6 &6 &12 &18 &18 &12 &6 &7 &14 &21 &22 &3 &6 &6 &12 &18 &3\\
\vspace{3pt}1 &1 &2 &2 &2 &4 &6 &6 &4 &1 &1 &2 &3 &3 &2 &3 &3 &6 &9 &1\\
\vspace{3pt}2 &2 &4 &4 &4 &8 &12 &12 &8 &2 &2 &4 &6 &6 &3 &6 &6 &12 &18 &2\\
\vspace{3pt}2 &2 &4 &4 &4 &8 &12 &12 &8 &2 &2 &4 &6 &6 &3 &6 &7 &13 &20 &2\\
\vspace{3pt}4 &4 &8 &8 &8 &16 &24 &24 &16 &4 &4 &8 &12 &12 &6 &12 &13 &26 &39 &4\\
\vspace{3pt}6 &6 &12 &12 &12 &24 &36 &36 &24 &6 &6 &12 &18 &18 &9 &18 &20 &39 &60 &6\\
\vspace{3pt}1 &1 &2 &2 &2 &4 &6 &6 &4 &1 &1 &2 &3 &3 &1 &2 &2 &4 &6 &2
\end{smallmatrix}
\right]
$$
is the valuation matrix of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$. Recall that by Proposition \ref{3d} any jumping number is supported
at some vertex $\gamma$ with $\widehat d_\gamma>0$ or $v_{\Gamma}(\gamma)>2$. Therefore it is enough to consider
the sets $\cH_{\gamma_j}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ for $j=1, 3, 7, 8, 9, 13, 14, 16, 19$ and $20$. By Theorem \ref{2} we know that
$$
\cH_{\gamma_1}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=\left\{\frac{t}{31}\middle | t+(v_\Gamma(\gamma_1)-2)V_{\gamma_1,\gamma_1}-\sum_{\nu\sim\gamma_1}s^{\gamma_1}_\nu
\left\lceil t\sum_{i\in\Gamma_\nu^{\gamma_1}}\frac{\widehat d_i V_{\gamma_1,i}}{s^{\gamma_1}_\nu d_\mu}\right\rceil^{+}\in S^{\gamma_1}\right\}.
$$
Clearly, $v_\Gamma(\gamma_1)=4$ and $V_{\gamma_1,\gamma_1}=1$. Furthermore, we have $s^{\gamma_1}_\nu=1$ for every $\nu\sim\gamma_1$ so
that $S^{\gamma_1}=\mathbb N$, and if we write $\Psi_{\gamma,\nu}:=\sum_{i\in\Gamma_\nu^{\gamma}}\widehat d_iV_{\gamma,i}$, we see that
$$
(\Psi_{\gamma_1,\nu})_{\nu=\gamma_3,\gamma_{10},\gamma_{16},\gamma_{20}}=\: \left(16,6,6,3\right).
$$
Subsequently,
$$
\cH_{\gamma_1}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=\left\{\frac{t}{31}\middle | t+2-\left\lceil\frac{16\cdot t}{31}\right\rceil^{+}
-2\cdot\left\lceil\frac{6\cdot t}{31}\right\rceil^{+}-\left\lceil\frac{3\cdot t}{31}\right\rceil^{+}\ge0\right\}.
$$
In the case $j=3$ we see that $v_\Gamma(\gamma_3)=3$ and $V_{\gamma_3,\gamma_3}=6$. Moreover,
$(s^{\gamma_3}_\nu)_{\nu=\gamma_1,\gamma_2,\gamma_9}=(2,3,6)$ so that $S^{\gamma_3}=\left\langle 2, 3\right\rangle$, and
$$
(\Psi_{\gamma_3,\nu})_{\nu=\gamma_1,\gamma_2,\gamma_9}=(30,0,48).
$$
Since $\left\lceil 0\right\rceil^{+}=1$ we see that
\begin{align*}
\cH_{\gamma_3}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}
=&\left\{\frac{t}{78}\middle | t+6-2\cdot\left\lceil\frac{30\cdot t}{2\cdot78}\right\rceil^{+}-3\cdot\left\lceil\frac{0\cdot t}{3\cdot78}\right\rceil^{+}
-6\cdot\left\lceil\frac{48\cdot t}{6\cdot78}\right\rceil^{+}\in\mathbb N\smallsetminus\{1\}\right\}\\
=&\left\{\frac{t}{78}\middle | t+3-2\cdot\left\lceil\frac{30\cdot t}{2\cdot78}\right\rceil^{+}
-6\cdot\left\lceil\frac{48\cdot t}{6\cdot78}\right\rceil^{+}\in\mathbb N\smallsetminus\{1\}\right\}.
\end{align*}
Similarly,
\begin{align*}
\cH_{\gamma_7}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=&\left\{\frac{t}{261}\middle | t+44-3\cdot\left\lceil\frac{129\cdot t}{3\cdot261}\right\rceil^{+}
-66\cdot\left\lceil\frac{132\cdot t}{66\cdot261}\right\rceil^{+}\in \left\langle 3, 22\right\rangle\right\},\\
\cH_{\gamma_8}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=&\left\{\frac{t}{263}\middle | t-67-\left\lceil\frac{129\cdot t}{263}\right\rceil^{+}\ge0\right\},\\
\cH_{\gamma_9}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=&\left\{\frac{t}{164}\middle | t-2\cdot\left\lceil\frac{60\cdot t}{2\cdot164}\right\rceil^{+}
-13\cdot\left\lceil\frac{78\cdot t}{13\cdot164}\right\rceil^{+}\in \left\langle 2,13\right\rangle\right\},\\
\cH_{\gamma_{13}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=&\left\{\frac{t}{117}\middle | t+14-3\cdot\left\lceil\frac{75\cdot t}{3\cdot117}\right\rceil^{+}
-21\cdot\left\lceil\frac{42\cdot t}{21\cdot117}\right\rceil^{+}\in \left\langle 3, 7\right\rangle\right\},\\
\cH_{\gamma_{14}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=&\left\{\frac{t}{119}\middle | t-22-\left\lceil\frac{75\cdot t}{119}\right\rceil^{+}\ge0\right\},\\
\cH_{\gamma_{16}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=&\left\{\frac{t}{68}\middle | t+3-2\cdot\left\lceil\frac{50\cdot t}{2\cdot68}\right\rceil^{+}
-6\cdot\left\lceil\frac{18\cdot t}{6\cdot68}\right\rceil^{+}\in \mathbb N\smallsetminus\{1\}\right\},\\
\cH_{\gamma_{19}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=&\left\{\frac{t}{210}\middle | t-20-3\cdot\left\lceil\frac{150\cdot t}{3\cdot210}\right\rceil^{+}
\in\left\langle 3, 20\right\rangle\right\},
\end{align*}
and finally,
$$
\cH_{\gamma_{20}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=\left\{\frac{t}{34}\middle | t-2-\left\lceil\frac{28\cdot t}{34}\right\rceil^{+}\ge0\right\}.
$$
Thus we get
\begin{align*}
\cH_{\gamma_1}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d} =&\left\{\frac{t+10m}{31}+n\middle | t =3,4,5,7,8,9,10;\: m=0,1,2\right\}\cup\mathbb Z_{+},\\
\cH_{\gamma_3}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d} =&\left\{\frac{5+10t+2m}{78}+n\middle | t,m,n\in\mathbb N;\:t<8;\: m<3-\frac{t}{4}\right\}\cup\mathbb Z_{+},\\
\cH_{\gamma_7}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d} =&\left\{\frac{t+3m+129p}{261}+n\middle | t=46, 89; m,n\in\mathbb N; p=0, 1;\frac{t+3m+129p}{261}\le\frac{1+p}{2}\right\}\cup\mathbb Z_{+},\\
\cH_{\gamma_8}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d} =&\left\{\frac{t+132}{263}\middle | t \in\mathbb N\right\},\\
\cH_{\gamma_9}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d} =&\left\{\frac{19+21t+2m}{164}\middle | t,m\in\mathbb N\text{ and }\frac{3-t}{3}\le m\le4+\frac{16t}{5}\right\},\\
\cH_{\gamma_{13}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d} =&\left\{\frac{t+3m+57p}{117}+n\middle | t =22, 41;\,m,n\in\mathbb N;\,p=0, 1;\frac{t+3m+57p}{117}\le\frac{1+p}{2}\right\}\cup\mathbb Z_{+},\\
\cH_{\gamma_{14}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d} =&\left\{\frac{t+60}{119}\middle | t\in\mathbb N\right\},\\
\cH_{\gamma_{16}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d} =&\left\{\frac{t+2m}{68}+n\middle | t=11, 33, 55, 66; m=1, 2, 3, 4, 5, 6; n\in\mathbb N\text{ and }23\neq t+2m\le68\right\},\\
\cH_{\gamma_{19}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d} =&\left\{\frac{t+3m}{210}\middle | t = 71, 142, 210;\: m\in\mathbb N\right\},\\
\cH_{\gamma_{20}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=&\left\{\frac{t+12}{34}\middle | t\in\mathbb N\right\},
\end{align*}
and the set of jumping numbers of $\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}$ is
$$
\cH^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}=\cH_{\gamma_{1}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}\cup\cH_{\gamma_{3}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}\cup\cH_{\gamma_{7}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}\cup\cH_{\gamma_{8}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}\cup\cH_{\gamma_{9}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}
\cup\cH_{\gamma_{13}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}\cup\cH_{\gamma_{14}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}\cup\cH_{\gamma_{16}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}\cup\cH_{\gamma_{19}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}\cup\cH_{\gamma_{20}}^\mathfrak{a}} \newcommand{\fb}{\mathfrak{b}} \newcommand{\fd}{\mathfrak{d}.
$$
\end{exmp}
|
\section{Introduction}
\label{sec:introduction}
In complex statistical problems it is often of interest to share information across multiple model parameters and components.
For studies with multiple responses, the same unknown set of features may be associated with the responses.
In our motivating example of medical expenditure data from the Medical Expenditure Panel Survey (MEPS), many individuals record no medical expenditures (zero response) over the course of a year. As a consequence, the distribution of an individual's semi-continuous response of total yearly medical expenditures is a mixture of a point-mass at zero and a continuous distribution on the positive reals. Intuitively, the set of factors which predict whether an individual incurs \emph{no} medical expenditure may also be predictive of the \emph{magnitude} that individual's medical expenditure if one occurs.
An increasingly popular method for modeling nonparametric functions is the Bayesian additive regression trees (BART) framework introduced by \citet{chipman2010bart}. The BART framework has been successfully applied to a diverse set of problems including survival analysis \citep{sparapani2016nonparametric}, causal inference \citep{hahn2017bayesian,hill2011bayesian}, analysis of loglinear models \citep{murray2017log}, imputation of missing predictors \citep{xu2016sequential}, and high dimensional prediction and variable selection \citep{linero2016bayesian}.
In this paper, we introduce \emph{shared forests}, which nonparametrically model multiple model components using a single set of trees. By viewing BART as a method for learning data-adaptive basis expansions, shared forests restrict the basis functions across model components to be the same while allowing the corresponding coefficients to be different. A simulation study shows that sharing information across model components in this fashion can be very beneficial, particularly in sparse high-dimensional problems in which variable selection a is necessary step.
In addition to our shared forests model, we make several additional contributions which are of practical interest in their own right. Semi-continuous responses are routinely modeled via two-part mixture models, often called hurdle models in econometrics, with a binary component modeling the probability of a zero response, and a continuous distribution modeling the response given it is non-zero. We present two novel semiparametric hurdle models for analyzing semi-continuous responses. The first is a type of gamma hurdle model, which are popular for modeling rainfall data \citep{feuerverger1979some}, in which the mean of the gamma distribution and the probability of a zero response are both modeled nonparametrically.
The second model is a log-normal hurdle model \citep{aitchison1955distribution, xiao1999comparison} in which the log-mean, log-variance, and the probability of a zero response are all modeled nonparametrically. See \citet{tu2006zero} for a review of zero-inflated and hurdle models. To the best of our knowledge, we are first to adapt BART to the mean of a gamma distribution; this requires developing an analog of the usual Bayesian backfitting approach for fitting BART models of \citet{chipman2010bart}. Additionally, while nonparametric models for the variance have been considered in other Bayesian sum-of-trees approaches \citep{murray2017log, pratola2017heteroscedastic}, we are required to develop a different nonparametric approach for the log-variance of our log-normal hurdle model in order to allow the tree structures to be shared across the mean, variance, and probability of a zero response while preserving computational tractability. In order to prevent overfitting on the variance component of the model, our variance modeling framework is designed to be centered at, and to allow shrinking heavily towards, a parsimonious model with constant variance. This allows us to model heteroskedasticity in the data while preserving estimation efficiency when the variance of the response is actually constant.
Our approach has natural connections with several proposals in the machine learning literature on \emph{multi-task learning} and \emph{multi-output} learning; see \citet{borchani2015survey} for a review. Related methods include multi-objective decision trees, multi-task boosting, and multi-task kernel methods.
Additionally, there are several methods which share information across models in the BART framework.
The Bayesian causal forest (BCF) model of \citet{hahn2017bayesian} uses two separate forests to model the distribution of potential outcomes: the first forest accounts for the direct effect of confounders on the potential outcomes, and is identical for both the treatment and controls, while the second forest is unique to the distribution of the treated samples.
Another related work by \citet{starling2018functional} proposes a model in which a temporally indexed response is modeled using BART, with the forest being shared across time. We discuss these connections in Section~\ref{sec:related-methods}.
We apply our methodology to data from the Medical Expenditure Panel Survey (MEPS).
The outcome $Y$ is an individual's total health care expenditure during the course of the year 2015. We show that the heteroskedastic log-normal hurdle model fits this data very well, and we use a shared forest to jointly model (i) the probability of $Y = 0$, (ii) the mean of $\log Y$ given
$Y$ is nonzero, and (iii) the variance of $\log Y$ given $Y$ is nonzero. By examining the fit of the mean and variance components, we are able to validate the earlier observation of \citet{blough2000using} that the variance of $Y$ is roughly proportional to $E(Y)^{1.5}$ for MEPS data. However, we are also able to identify sources of heterogeneity which are not explained by this relationship between the variance and the mean.
In Section~\ref{sec:shared}, we introduce our shared forests framework. In Section~\ref{sec:models}, we develop and give default prior specifications for the gamma hurdle and log-normal hurdle models we use later to analyze the MEPS data. In Section~\ref{sec:simulation}, we conduct a simulation study which illustrates the potential benefits of sharing information across model components. In Section~\ref{sec:analysis}, we apply the methodology developed here to the MEPS dataset. We conclude in Section~\ref{sec:discussion} with a discussion. Additional computational details, and details about the analysis of the MEPS dataset, are given in a web appendix.
\section{Shared Forests}
\label{sec:shared}
\subsection{Review of Bayesian Additive Regression Trees}
\label{sec:review}
We briefly review the Bayesian Additive Regression Trees (BART) framework \citep{chipman2010bart}, an extremely useful tool for constructing highly flexible Bayesian semiparametric models. BART models typically outperform comparable linear models and often outperform machine learning techniques such as boosted decision trees and random forests.
We assume that the unknown function of interest $h(\bm x)$ can be expressed as a sum of $T$ regression trees depending on tree structures $\mathcal T_t$ and leaf-node parameters $\mathcal M_t$,
\begin{align}
\label{eq:basis}
h(\bm x) = \sum_{t=1}^T g(\bm x; \mathcal T_t, \mathcal M_t),
\end{align}
where $g(\bm x; \mathcal T_t, \mathcal M_t) = \mu_{t\ell}$ if the predictor value $\bm x$ is associated to leaf node $\ell$ of tree $t$. As illustrated in Figure~\ref{fig:GraphFigure}, the decision tree $\mathcal T_t$ encodes a recursive partition of the predictor space $\mathcal X = [0,1]^P$, with $g(\bm x; \mathcal T_t, \mathcal M_t)$ being piecewise-constant. Let $\mathcal L_t$ denote the leaf nodes of the tree, so that $\mathcal M_t = \{\mu_{t\ell} : \ell \in \mathcal L_t\}$.
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{./figure/GraphFigure}
\caption{Schematic illustration of the construction of a decision tree (bottom) with the induced recursive partitioning of the predictor space $\mathcal X = [0,1]^2$. After the decision tree is constructed, parameters associated to leaf node $\ell$ are given a mean parameter $\mu_{\ell}$.}
\label{fig:GraphFigure}
\end{figure}
The prior for $h(\cdot)$ in (\ref{eq:basis}) consists of a prior mass function $\pi_\mathcal T(\cdot)$ for the tree structures $\mathcal T_t$ and a prior on the leaf node parameters $\mathcal M_t$. \citet{chipman2010bart} propose a branching process prior for $\mathcal T_t$. A draw from this prior is obtained by generating, for each node at depth $d$, two child nodes with probability $q(d) = \gamma(1 + d)^{-\zeta}$; otherwise, the node becomes a leaf node (which defines a new equivalence class). This process iterates for $d = 0,1,2,\ldots$ until we reach a depth $d$ at which all of the nodes are leaves.
Note that $q(d)$ is not a mass function over $d$, but instead is the prior probability of a given leaf node being converted to a branch node. The case $\zeta = 0$ corresponds to the Galton-Watson process \citep{athreya2004branching}. A sufficient condition for termination of this branching process is $\zeta > 0$.
After the tree topology is generated, each branch node $b$ is associated to a decision rule of the form $[x_j \le C_b]$ where the coordinate $j\in \{1,\ldots,P\}$ is selected independently for each branch with probability $s_j$.
Throughout, we will use the sparsity inducing Dirichlet prior $(s_1, \ldots, s_P) \sim \mathcal D(\xi/P, \ldots, \xi/P)$ proposed by \citet{linero2016bayesian}. This prior concentrates on neighborhoods of sparse probability vectors, a fact which has been leveraged to perform variable selection in linear models \citep{bhattacharya2015dirichlet}, and adapt to irrelevant predictors in Gaussian process models \citep{bhattacharya2014anisotropic}. Intuitively, if $s_j$ is very small (e.g., $s_j < 10^{-10}$), then predictor $x_j$ is highly unlikely to appear within any splitting rule, effectively eliminating $x_j$ from the model. The Dirichlet prior encourages these extreme values of $s_j$, allowing the model to perform fully-Bayesian variable selection.
In this paper, we set $C_b \sim \operatorname{Uniform}(L_j,U_j)$ for the cut-points $C_b$ conditional on the tree topology, selected coordinate $j$, and the parameters of $b$'s ``ancestor nodes''.
Here $(L_1, U_1) \times \cdots \times (L_P, U_P)$ is the hyperrectangle defined by all values of $\bm x$ which lead to branch $b$.
Let $\bm \omega$ be a vector of non-tree-specific parameters, such as the variance $\sigma^2 = \operatorname{Var}(Y_i \mid \bm X_i)$ for a regression model with constant variance. Our model for the response $Y_i$ in this setting is expressed as $(Y_i \mid \bm X_i = \bm x, h, \bm \omega) \sim f\{y \mid h(\bm x), \bm \omega\}$ where $\{f(\cdot \mid \mu, \bm \omega)\}$ is a parametric family. Conditional on the trees $\mathcal T_1,\ldots,\mathcal T_T$, the leaf node parameters $\{\mu_{t\ell} : \ell \in \mathcal L_t, 1 \le t \le T\}$ are given iid priors $\mu_{t\ell} \sim \pi_\mu$. Usually $\pi_\mu$ is chosen to ensure that the integrated likelihood
\begin{align}
\label{eq:integrated-like}
\Lambda(\mathcal T_t)
=
\pi_\mathcal T(\mathcal T_t)
\int
\prod_{i = 1}^n f\{Y_i \mid h(\bm X_i), \bm \omega\} \,
\prod_{\ell \in \mathcal L_t} \pi_\mu(\mu_{t\ell}) \ d\mu_{t\ell}
\end{align}
has a closed form expression. For example, in the regression setting,
a popular choice for $\pi_{\mu}$ is the $ \mathcal N(0, \sigma^2_\mu)$ density. When using Markov Chain Monte Carlo (MCMC) to conduct Bayesian inference, $\mathcal T_t$ can be updated using Metropolis-Hastings, with $\Lambda(\mathcal T_t)$ used to compute the acceptance probability; see \citet{pratola2016efficient} for further details. In the original paper of \citet{chipman2010bart}, both $\Lambda(\mathcal T_t)$ and the full conditionals for the $\mu_{t\ell}$'s are calculated using Bayesian backfitting.
A large reason for the success of BART is the existence of highly effective ``default'' priors which can be expected to provide a reasonable baseline level of performance without requiring tuning by the user. As a default, we set $\gamma = 0.95$, $\zeta = 2$, and $\xi/(\xi + P) \sim \operatorname{Beta}(0.5,1)$; other prior specifications are model specific. Additionally, BART models have highly desirable theoretical properties \citep{linero2017abayesian, rockova2017posterior}; in particular, in regression problems, certain BART models attain near-minimax optimal rates of estimation for functions $h(\bm x)$ with low-order interactions. For an in-depth review of Bayesian regression tree methods, see \citet{chipman2013bayesian} and \citet{linero2017review}.
A fact which will be useful for specifying priors later, and for making
connections with other approaches, is that, under the conditions $E(\mu_{t\ell})
= 0$ and $\operatorname{Var}(\mu_{t\ell}) = \sigma^2_{\mu} / T$, the prior converges to a
Gaussian process as $T \to \infty$. To see this heuristically, note that we can
write $h(\bm x) = T^{-1/2} \sum_{t = 1}^T \mu^\star_{t\ell}$ where the
$\mu^{\star}_{t\ell}$'s are iid with mean $0$ and variance $\sigma^2_\mu$. In
general, for fixed $\bm x, \bm x'$, we have $\operatorname{Cov}\{h(\bm x), h(\bm x')\} = \sigma^2_\mu
\Pr(\bm x \sim \bm x')$ where the event $[\bm x \sim \bm x']$ denotes that $\bm x$ and
$\bm x'$ share the same terminal node in $\mathcal T_1$. An application of the
multivariate central limit theorem then establishes convergence of the finite
dimensional marginals to a multivariate normal
distribution.
A natural alternative method is to simply apply Gaussian processes in practice.
BART has both practical and theoretical benefits over Gaussian process models.
First, the computational complexity of Gaussian process methods typically is
$O(n^3)$ where $n$ is the number of samples; conversely, in practice, BART has
computations which scale slightly faster than $O(nT)$ \citep{chipman2010bart}.
Second, the recent works of \citet{rockova2017posterior} and
\citet{linero2017abayesian} show that BART models are capable of adapting to
low-order interactions in the covariates. Third, as we will see in
Section~\ref{sec:simulation}, empirically, BART with a finite $T$ tends to
perform better than the limiting model as $T \to \infty$.
\subsection{The shared forests model}
\label{sec:multi}
We consider a generalization of (\ref{eq:basis}) and set
\begin{math}
(Y_i \mid \bm X = \bm x, \bm h, \bm \omega)
\sim f\{y \mid \bm h(\bm x), \bm \omega\},
\end{math}
where $\bm h= (h_1, \ldots, h_M)$ is a collection of $m$ functions and each $h_m(\bm x)$
is modeled non-parametrically as a sum of regression trees. Note that, as illustrated in Figure~\ref{fig:GraphFigure}, we are effectively modeling $h_m(\bm x)$ in terms of a basis function expansion from an overcomplete family of basis functions
\begin{align*}
h_m(\bm x) = \sum_{t = 1}^T \sum_{\ell \in \mathcal L_t^{(m)}} \mu^{(m)}_{t\ell} \psi^{(m)}_{t\ell}(\bm x),
\qquad
\qquad
\psi^{(m)}_{t\ell}(\bm x) = I[\bm x \leadsto (t,\ell)],
\end{align*}
where $[\bm x \leadsto (t,\ell)]$ occurs if $\bm x$ is associated to leaf $\ell$ of tree $\mathcal T_t^{(m)}$ and $I(A)$ is the indicator that the event $A$ occurs. We can then view $\{\psi_{t\ell}^{(m)} : 1 \le t \le T, 1 \le m \le M, \ell \in \mathcal L_t^{(m)}\}$ as a collection of features which are adaptively learned from the data to approximate $\{h_1(\bm x), \ldots, h_M(\bm x)\}$.
Our proposed shared forest framework assumes that these basis functions are shared across $M$ model components; that is, we assume $\psi^{(m)}_{t\ell}(\bm x) \equiv \psi_{t\ell}(\bm x)$ for $m=1,\cdots,M$. Equivalently, we assume that the features which are useful for approximating $h_m(\bm x)$ are the same features that are useful for approximating $h_{m'}(\bm x)$.
Note, however, that a given feature $\psi_{t\ell}(\bm x)$ has
different unknown coefficients (effects) $\mu_{t\ell}^{(1)}$ and
$\mu_{t\ell}^{(2)}$ respectively for $h_1(\bm x)$ and $h_2(\bm x)$.
This shared basis function framework is imposed by assuming that the $h_m(\bm x)$'s are modeled using $T$ regression trees with the same collection of trees for all $M$ model components that are potentially affected by the covariate vector $\bm x$. That is, we assume
\begin{align}
\label{eq:newtree}
h_m(\bm x) = \sum_{t=1}^T g(\bm x ; \mathcal T_t, \mathcal M_t^{(m)}) \ ,
\end{align}
where $\bm \mu_{t\ell} = (\mu_{t\ell}^{(m)} : 1 \le m \le M)$. We will assume the multivariate prior density $\bm \mu_{t\ell} \sim \pi_{\bm \mu}$, potentially allowing dependence across parameters $\mu_{t\ell}^{(m)}$ for different values of $m$.
\begin{example}
\label{ex:mixed}
Consider a mixed-scale response $\bm Y_i = (Y_{i1}, Y_{i2}, Z_i)$ in which $Y_{ij} = h_j(\bm x) + \epsilon_j$ where $\bm \epsilon \sim \mathcal N(0, \bm \Sigma)$, $Z_i \sim \operatorname{Bernoulli}[\Phi\{h_3(\bm x)\}]$, and $\bm h$ is modeled with a shared forest with $\bm \mu_{t\ell} \sim \mathcal N(\bm 0, \Sigma_\mu / T)$. We consider a variant of this model in Section~\ref{sec:simulation}.
\end{example}
\begin{example}
\label{ex:semi}
Consider a semicontinuous response $(Y_i \mid \bm X_i = \bm x, \bm h, \bm \omega)$ where $Y_i > 0$ occurs with probability $\Phi\{h_1(\bm x)\}$ and $(Y_i \mid Y_i > 0, \bm X_i = \bm x, \bm h, \bm \omega) \sim \operatorname{Gam}[\alpha, \alpha \exp\{h_2(\bm x)\}]$. We refer to this model as the gamma hurdle model; see Section~\ref{sec:gamma}.
\end{example}
Bayesian inference for the shared forest model of (\ref{eq:newtree}) can be conducted by extending \eqref{eq:integrated-like} to incorporate priors on the parameters for the leaf nodes across the multiple model components giving the integrated likelihood
\begin{align}
\label{eq:integrated-multi}
\begin{split}
\Lambda(\mathcal T_t)
&=
\pi_\mathcal T(\mathcal T_t)
\int
\prod_{i = 1}^n f\{Y_i \mid \bm h(\bm X_i), \bm \omega\} \,
\left[
\prod_{\ell \in \mathcal L_t} \pi_{\bm \mu}(\bm \mu_{t\ell}) \ d\bm\mu_{t\ell}
\right]
\\&=
\pi_\mathcal T(\mathcal T_t)
\prod_{\ell \in \mathcal L_t}
\int
\prod_{i: \bm X_i \leadsto (t,\ell)}
f\{Y_i \mid \bm h(\bm X_i), \bm \omega\} \,
\pi_{\bm \mu}(\bm \mu_{t\ell}) \ d\bm \mu_{t\ell}\ .
\end{split}
\end{align}
As before, if \eqref{eq:integrated-multi} has a closed form then one can update $\mathcal T_t$ within an MCMC algorithm using standard Metropolis-Hastings proposals.
\subsection{Related methods}
\label{sec:related-methods}
There are several proposals for BART models which are related to our shared forests model. \citet{hahn2017bayesian} consider a related structure in the context of causal inference; given a binary treatment $z$, they model potential outcomes $Y_i(z)$ as $Y_i(z) = h(\bm X_i) + z\alpha(\bm X_i) + \epsilon_i$, with both $h(\bm x)$ and $\alpha(\bm x)$ modeled using BART priors. This is referred to as a \emph{Bayesian causal forest} (BCF). The function $h(\bm x)$ represents the prognostic effects of the covariates $\bm X_i$, which is shared across both potential outcomes, while $\alpha(\bm x)$ represents a treatment-covariate interaction, which is unique to the treated group. This differs from the shared forests framework we present in that we only require sharing the tree topologies across model components, while the BCF model shares the entire function $h(\bm x)$. Alternatively, one may view the BCF model for $h(\bm x)$ as a shared forest in which the values in leaf $\ell$, given by $(\mu_{\ell,0}, \mu_{\ell,1})$, are perfectly correlated. In the context of causal inference, this separation of the effect into a perfectly-shared forest $h(\bm x)$ and a completely separate treatment effect $\alpha(\bm x)$ is desirable because it gives the analyst a great deal of control over the prior information expressed about individual-level treatment effects.
In the context of functional regression, \citet{starling2018functional} model a temporally-observed response using a BART model as $Y_i(t) = h_t(\bm X_i) + \epsilon_{i}(t)$ where here $t \in \mathscr T$ indexes the observation time. The parameters of the leaf nodes of the trees in the ensemble are then modeled as random functions $\mu_\ell(t)$, with Gaussian process priors. The distinction between how $t$ and $\bm x$ are incorporated into their model is referred to as \emph{targeted smoothing}, as the model induces a higher degree of smoothing over $t$ than $\bm x$. This approach can also be cast as a type of shared forest model in which the collection of regression functions $\{h_t(\bm x) : t \in \mathscr T\}$ share the same tree topology. The dependence between $h_{t}(\bm x)$ and $h_{t'}(\bm x)$ induced using Gaussian processes is analogous to using the multivariate normal prior described in Example~\ref{ex:mixed}.
Shared forests have natural connections to many proposals for \emph{multi-task} or \emph{multi-output} methods in machine learning. The most immediate connections are with multi-objective decision trees (MODTs) initially proposed by \citet{de2002multivariate}. MODTs are grown in a CART-like fashion, but use a multivariate purity function for evaluating the quality of splits. In this way, splits are useful for predicting all outputs simultaneously. Our shared forests model with $T = 1$ is essentially a Bayesian version of a MODT, as the marginal likelihood \eqref{eq:integrated-multi} will be large when $\mathcal T_t$ gives good predictions across all tasks. MODTs can be ensembled in the usual ways via the bagging and random forests algorithms \citep{kocev2007ensembles}.
Our characterization of the shared forests model in terms of a shared basis function expansion is similar to the assumption of the multi-task feature learning approach of \citet{argyriou2007multi}. Our approach can also be related to the FIRE algorithm for fitting rule ensembles proposed by \citet{aho2012multi}. Each $\psi_{t\ell}(\bm x)$ in the ensemble is a ``rule'' and the $\mu_{t\ell}$'s are task-specific weights assigned to each rule. Additionally, in the same way that BART is analogous to gradient boosted decision trees \citep{chipman2010bart, freund1999short}, the shared forests model is analogous to the boosted multi-task learning approach of \citet{chapelle2011boosted}.
Using the connection between BART and Gaussian processes, we can also interpret the shared forests model in terms of multi-task kernel methods. Recall from Section~\ref{sec:review} that BART can be thought of as approximately implementing Gaussian process regression when the number of trees $T$ is large, with kernel $K(\bm x,\bm x') = \sigma^2_\mu\Pr(\bm x \sim \bm x')$. If the leaf nodes of the ensemble across the tasks are endowed with the prior $\bm \mu_{t\ell} \sim \mathcal N(0, \Sigma_\mu / T)$ then the cross-task kernel function is given by $\operatorname{Cov}\{h_j(\bm x), h_k(\bm x)\} = \Sigma_{ij} \Pr(\bm x \sim \bm x')$. This matches the proposal of \citet{bonilla2008multi} for multi-task Gaussian processes. Sharing of information across tasks for the shared forest occurs at a deeper level still, however, as in the finite-$T$ case the rule-sharing interpretation of our approach still applies even if $\Sigma_{ij} = 0$; as we show in the simulation study of Section~\ref{sec:simulation}, substantial gains are possible even with $\Sigma_{ij} = 0$.
\section{Models for semicontinuous data}
\label{sec:models}
\subsection{Probit-based hurdle models}
Motivated by the MEPS dataset, we present two models for analyzing zero-inflated responses. Throughout, let $\pi(\bm x) = \Pr(Y_i > 0 \mid \bm X_i = \bm x, \bm h, \bm \omega)$ denote the probability of a non-zero response. The gamma hurdle and log-normal hurdle models below are special cases of the probit-based hurdle model, where
\begin{math}
\pi(\bm x) = \Phi\{\theta_0 + h_\theta(\bm x)\},
\end{math}
and
\begin{math}
(Y_i \mid Y_i > 0, \bm X_i = \bm x, \bm h, \bm \omega)
\sim f\{y \mid \bm h_u(\bm x), \bm \omega\}.
\end{math}
Here, $\{f(\cdot \mid \mu, \bm \omega)\}$ is a parametric family of densities for the positive part of $Y_i$. We model $\bm h = (h_\theta, \bm h_u)$ with a shared forest. Let $\theta_{t\ell}$ denote the parameter associated to leaf $\ell$ of $\mathcal T_t$ for $h_\theta$ and $u_{t\ell}$ the parameter associated to leaf $\ell$ of $\mathcal T_t$ for $\bm h_u$. We use independent priors for the $\theta_{t\ell}$'s and $u_{t\ell}$'s and, following \citet{chipman2010bart}, set $\theta_{t\ell} \stackrel{\textnormal{iid}}{\sim} \mathcal N(0, \sigma^2_\theta)$.
For the sake of computational convenience, we do not use \eqref{eq:integrated-multi} directly, but instead augment the data with latent variables $Z_i \stackrel{\textnormal{indep}}{\sim} \mathcal N\{\theta_0 + h_\theta(\bm X_i), 1\}$ such that $Y_i > 0$ if-and-only-if $Z_i > 0$ \citep{albert1993bayesian}. Before computing \eqref{eq:integrated-multi} we first sample the $Z_i$'s from a $\mathcal N\{\theta_0 + h_\theta(\bm X_i), 1\}$ distribution, truncated to $(-\infty,0)$ or $(0,\infty)$ according as $Y_i = 0$ or $Y_i > 0$. We then compute the integrated likelihood
\begin{align}
\label{eq:probit-marginal}
\begin{split}
\Lambda(\mathcal T_t)
&=
\pi_\mathcal T(\mathcal T_t) \,
\prod_{\ell \in \mathcal L_t}
\left[ \int \prod_{i : \bm X_i \leadsto (t,\ell)} \mathcal N\{Z_i \mid \theta_0 + h_\theta(\bm X_i), 1\} \ \mathcal N(\theta_{t\ell} \mid 0, \sigma^2_\theta) \ d\theta_{t\ell}\right. \\
&\qquad \times \left. \int \prod_{i: Z_i > 0, \bm X_i \leadsto (t,\ell)} f\{Y_i \mid \bm h_u(\bm X_i), \bm \omega\} \ \pi_u(u_{t\ell}) \ du_{t\ell} \right] \\
&= \pi_\mathcal T(\mathcal T_t) \prod_{\ell \in \mathcal L_t} L_\theta(t,\ell) \cdot L_u(t,\ell)\ .
\end{split}
\end{align}
Notice that $L_\theta(t,\ell)$ does not depend on our choice for the distribution of the non-zero $Y_i$'s and can be computed in closed form; an expression for $L_\theta(t,\ell)$ is given in the web appendix. Hence, all that must be done to apply the probit-based hurdle model is to be able to compute $L_u(t,\ell)$ in closed form.
\subsection{Gamma hurdle models}
\label{sec:gamma}
Our semiparametric gamma hurdle model sets $Y_i \sim \operatorname{Gam}[\alpha, \alpha \exp\{\lambda_0 + h_\lambda(\bm x)\}]$ conditional on $Y_i > 0$ and $\bm X_i = \bm x$,
where $\operatorname{Gam}(\alpha,\beta)$ is parameterized to have mean $\alpha / \beta$ and variance $\alpha/\beta^2$. We model $h_\theta(\bm x)$ and $h_\lambda(\bm x)$ with a shared forest,
\begin{math}
h_\theta(\bm x) = \sum_{t = 1}^T g(\bm x ; \mathcal T_t, \mathcal M_{\theta,t}),
\end{math}
and
\begin{math}
h_\lambda(\bm x) = \sum_{t = 1}^T g(\bm x ; \mathcal T_t, \mathcal M_{\lambda,t}).
\end{math}
Note that, under this model, we have
\begin{align}
\label{eq:gamma-moments}
\begin{split}
E(Y_i \mid Y_i > 0, \bm X_i = \bm x, \bm h, \alpha)
&=
\exp\left\{ -\lambda_0 - h_\lambda(\bm x) \right\},
\\
\operatorname{Var}(Y_i \mid Y_i > 0, \bm X_i = \bm x, \bm h, \alpha)
&=
\frac{\exp [-2 \{\lambda_0 + h_\lambda (\bm x)\}]}{\alpha}
,
\end{split}
\end{align}
so that the conditional standard deviation of $Y_i$ is proportional to its mean.
The leaf-specific parameters for $h_\lambda(\bm x)$ are given log-gamma priors $\lambda_{t\ell} \sim \log \operatorname{Gam}(\alpha_\lambda, \beta_\lambda)$. The log-gamma prior is chosen because it is conjugate to the gamma likelihood and makes computation of \eqref{eq:probit-marginal} tractable. Under this prior for the leaf parameters, the gamma hurdle model is immediately applicable provided that we can compute the likelihood factor
\begin{align*}
L_\lambda(t,\ell)
=
\int \prod_{i \in \ell: Y_i > 0}
\operatorname{Gam}[Y_i \mid \alpha, \alpha \exp\left\{ \lambda_0 + h_\lambda(\bm x) \right\}]
\,
\log \operatorname{Gam}\{\lambda_{t\ell} \mid \alpha_\lambda, \beta_\lambda\}.
\end{align*}
To do this, similar to \citet{murray2017log} for loglinear models, we define
\begin{math}
\eta_i
=
\alpha
\exp \{\lambda_0 + h_\lambda(\bm X_i) - g(\bm X_i ; \mathcal T_t, \mathcal M_{\lambda,t})\}.
\end{math}
By analogy with the usual Bayesian backfitting algorithm of \citet{chipman2010bart}, the $\eta_i$'s play the role of the backfitted response. Let $A_\ell = \{i : i \in \ell, Y_i > 0\}$ and $N_\ell = |A_\ell|$. Then
\begin{align*}
\prod_{i \in A_\ell}
\operatorname{Gam}[Y_i \mid \alpha,
\alpha \exp\left\{ \lambda_0 + h_\lambda(\bm X_i) \right\}]
=
\left(\prod_{i \in A_\ell} \frac{(Y_i \eta_i)^\alpha}{Y_i \Gamma(\alpha)}\right)
\exp\left( \alpha N_{\ell} \lambda_{t\ell} - \sum_{i \in A_\ell} Y_i \eta_i e^{\lambda_{t\ell}} \right).
\end{align*}
Integrating against the $\log\operatorname{Gam}(\lambda_{t\ell} \mid \alpha_\lambda, \beta_\lambda)$ density gives
\begin{align*}
L_\lambda(t,\ell)
=
\frac{\prod_{i \in A_\ell} \eta_i^\alpha Y_i^{\alpha-1}}{\Gamma(\alpha)^{N_\ell}}
\cdot
\frac{\beta_\lambda^{\alpha_\lambda}}{\Gamma(\alpha_\lambda)}
\cdot
\frac{\Gamma(\alpha_\lambda + N_\ell \alpha)}
{(\beta_\lambda + \sum_{i \in A_\ell} Y_i \eta_i)^{\alpha_\lambda + N_\ell \alpha}}.
\end{align*}
Hence \eqref{eq:probit-marginal} can be computed in closed form. Additionally, by conjugacy of the log-gamma distribution, we have the full conditionals
\begin{align}
\label{eq:lambda-full-conditional}
\lambda_{t\ell} \sim \log \operatorname{Gam}\left(\alpha_\lambda + \alpha N_\ell, \beta_\lambda + \sum_{i \in A_\ell} Y_i \eta_i\right).
\end{align}
A detailed Markov chain Monte Carlo algorithm is given in the web appendix.
\subsection{Log-normal hurdle model}
\label{sec:lognormal}
A shortcoming of the gamma hurdle model is that the relationship between $\bm x$ and the variance is captured entirely through the mean. As an alternative, we propose the heteroskedastic log-normal hurdle model with $\pi(\bm x) = \Phi\{\theta_0 + h(\bm x)\}$ and $(Y_i \mid Y_i > 0, \bm X_i = \bm x, h, mu, \sigma^2) \sim \log \Normal\{\mu(\bm x), \sigma^2(\bm x)\}$.
We again use a shared forest to model the three functions
\begin{alignat*}{3}
\mu(\bm x) &= \sum_{t = 1}^T g(\bm x ; \mathcal T_t, \mathcal M_{\mu,t}), &\qquad&&
\sigma^{-2}(\bm x) &= \exp\left\{ \lambda_0 + \sum_{t=1}^T g(\bm x ; \mathcal T_t, \mathcal M_{\lambda,t}) \right\}, \\
h(\bm x) &= \sum_{t=1}^T g(\bm x ; \mathcal T_t, \mathcal M_{\theta,t}).
\end{alignat*}
The resulting model for the mean and variance of $(Y_i \mid Y_i > 0, \bm X_i = \bm x)$ is given by
$m(\bm x) = \exp\{\mu(\bm x) + \sigma^2(\bm x) / 2\}$ and $s^2(\bm x) = m(\bm x)^2 [\exp\{\sigma^2(\bm x)\} - 1]$.
Like the gamma hurdle model, when a \emph{homoskedastic} model for $\log Y_i$ is used, we find that the mean $m(\bm x)$ is proportional to the standard deviation $s(\bm x)$. By modeling $\sigma^2(\bm x)$ nonparametrically, however, we allow for more complex relationships between $m(\bm x)$ and $s(\bm x)$. Our heteroskedastic model for the $\log Y_i$'s is similar to the heteroskedastic BART models proposed by \citet{murray2017log} and \citet{pratola2017heteroscedastic}, but our model differs in two respects. First, the trees used to model the mean and variance functions are shared, which is helpful because the variance function $\sigma^2(\bm x)$ is generally much more weakly identified than the mean $\mu(\bm x)$. Second, our choice of prior for $\sigma^2(\bm x)$ will explicitly shrink the posterior model towards a homoskedastic model; see Section~\ref{sec:prior-lognormal}.
Let $\mu_{t\ell}$ and $\lambda_{t\ell}$ be the leaf parameters associated to leaf $\ell$ of $\mathcal T_t$ for $\mu(\cdot)$ and $\sigma(\cdot)$ respectively and let $\tau_{t\ell} = \exp(\lambda_{t\ell})$. We use a normal-gamma prior for $(\mu_{t\ell}, \tau_{t\ell})$, i.e.,
\begin{math}
\tau_{t\ell} \sim \operatorname{Gam}(\alpha_\lambda, \beta_\lambda)
\end{math}
and
\begin{math}
\mu_{t\ell} \sim \mathcal N\{0, 1/(\kappa \tau_{t\ell})\}.
\end{math}
This normal-gamma prior allows for computation of the likelihood factor
\begin{align*}
L_{\mu,\lambda}(t,\ell)
&=
\int \prod_{i: \bm X_i \leadsto (t,\ell), Y_i > 0}
\log \mathcal N\{Y_i \mid \mu(\bm X_i), \sigma^2(\bm X_i)\}
\\
&\qquad\times
\mathcal N\{\mu_{t\ell} \mid 0, 1/(\kappa \tau_{t\ell})\}
\
\operatorname{Gam}(\tau_{t\ell} \mid \alpha_\lambda, \beta_\lambda)
\
d\mu_{t\ell} \ d\tau_{t\ell}.
\end{align*}
Let $W_i = \log Y_i$ and suppose $\bm X_i \leadsto (t,\ell)$. Then, conditional on $Y_i > 0$, we have
\begin{math}
W_i \sim \mathcal N(\eta_i + \mu_{t\ell} , \frac{1}{\nu_i \tau_{t\ell}}),
\end{math}
where
\begin{math}
\eta_i = \sum_{j \ne t} g(\bm X_i; \mathcal T_j, \mathcal M_{\mu,j}),
\end{math}
and
\begin{math}
\nu_i = \exp\left\{ \lambda_0 + \sum_{j \ne t} g(\bm X_i; \mathcal T_j, \mathcal M_{\lambda,j})\right\}.
\end{math}
Let $Q_i = W_i - \eta_i$ and $A(\ell) = \{i : \bm X_i \leadsto (t,\ell), Y_i > 0\}$; $Q_i$ and $\nu_i$ are analogous to the backfitted response in the usual Bayesian backfitting algorithm. We have
\begin{align*}
\prod_{i \in A(\ell)} \mathcal N\left(W_i \mid \eta_i + \mu_{t\ell}, \frac 1 {\nu_i \tau_{t\ell}}\right)
=
\prod_{i \in A(\ell)}
\left( \frac{\nu_i \tau_{t\ell}}{2\pi} \right)^{1/2}
\exp\left\{ -\frac{\nu_i \tau_{t\ell}}{2} (Q_i - \mu_{t\ell})^2 \right\} .
\end{align*}
This likelihood is conjugate to the normal-gamma prior for $(\mu_{t\ell}, \tau_{t\ell})$, and routine computations give the expression
\begin{align*}
\left( \prod_i \sqrt{\frac{\nu_i}{2\pi}} \right)
\sqrt{\frac{\kappa}{\kappa + w_\ell}}
\frac{\beta_\lambda^{\alpha_\lambda} \Gamma(\alpha_\lambda + N_\ell / 2)}{\Gamma(\alpha_\lambda)}
\left( \beta_\lambda + \frac{S^2_{\ell}}{2} + \frac{\kappa w_\ell \bar Q_\ell^2}{2(\kappa + w_\ell)} \right)^{-(\alpha_\lambda + N_{\ell}/2)},
\end{align*}
for $L_{\mu,\lambda}(t,\ell)$ where
\begin{align*}
\bar Q_\ell = \frac{\sum_{i \in A(\ell)} \nu_i Q_i}{\sum_{i \in A(\ell)} \nu_i},
\qquad
w_\ell = \sum_{i \in A(\ell)} \nu_i,
\qquad \text{and} \qquad
S^2_\ell = \sum_{i \in A(\ell)} \nu_i (Q_i - \bar Q_\ell)^2.
\end{align*}
We again have a closed form for \eqref{eq:probit-marginal}. Moreover, we also have the following full conditionals for the leaf parameters:
\begin{align}
\label{eq:ng-full-conditional}
\tau_{t\ell} \sim \operatorname{Gam}(\widehat \alpha_\ell, \widehat \beta_\ell),
\qquad
\text{and}
\qquad
\mu_{t\ell} \sim \mathcal N\{\widehat \mu_\ell, 1/(\widehat \kappa_\ell \tau_{t\ell})\},
\end{align}
where
\begin{alignat*}{3}
\widehat \alpha_\ell &= \alpha_\lambda + N_\ell / 2,
&&\qquad&
\widehat \beta_\ell &= \beta_\lambda + \frac{S^2_\ell}{2} + \frac{\bar Q_\ell^2 \kappa w_\ell}{2(\kappa + w_\ell)},
\\
\widehat \kappa_\ell &= \kappa + w_\ell,
&&\qquad&
\widehat \mu_\ell &= \frac{\sum_{i \in A(\ell)} v_i Q_i}{\widehat \kappa}.
\end{alignat*}
Additional details for the various steps of the MCMC algorithm are deferred to the web appendix.
\subsection{Prior specification}
\label{sec:prior-lognormal}
An advantage of the BART framework is that there exist standard ``default'' priors which have proven to work remarkably well in practice. In particular, very little tuning is required to obtain an acceptable baseline level of performance. We develop default priors for the gamma hurdle and log-normal hurdle models we consider here. For both models, we will use the default prior recommended by \citet{chipman2010bart} for the $\theta_{t\ell}$'s. Additionally, we apply a quantile normalization separately to each column of the design matrix $\bm X$ so that the predictors are distributed approximately uniformly on $[0,1]$.
We first give a prior specification for the log-normal hurdle model. As a preprocessing step, we work with $W_i = \log Y_i$; further, we standardize the finite $W_i$'s to have mean $0$ and standard error $1$. In order for the prior to be stable as the number of trees is increased, we choose the hyperparameters so that $E(\lambda_{t\ell}) = 0$ and $\operatorname{Var}(\lambda_{t\ell}) = a_\lambda^2 / T$, and similarly for $\mu_{t\ell}$ and $\theta_{t\ell}$. This ensures that the stochastic process $\sum_{t=1}^T g(\bm x; \mathcal T_t, \mathcal M_t)$ converges to a Gaussian process as $T \to \infty$ so that the prior is stable under adding additional trees.
Appropriate values for $(\alpha_\lambda, \beta_\lambda)$ can be obtained by solving the equations
\begin{align}
\label{eq:mean}
E(\lambda_{t\ell}) &= \psi(\alpha_\lambda) - \log \beta_\lambda = 0, \\
\label{eq:var}
\operatorname{Var}(\lambda_{t\ell}) &= \psi'(\alpha_\lambda) = a^2_\lambda / T.
\end{align}
Noting that $\psi'(\alpha) \approx \alpha^{-1}$, \eqref{eq:var} implies that for moderate values of $T$ we will have $\alpha \approx T/a_{\lambda}^2$. Additionally, noting that $\psi(\alpha) \approx \log(\alpha)$, \eqref{eq:mean} implies that $\alpha_\lambda \approx \beta_{\lambda}$; in particular, both $\alpha_\lambda$ and $\beta_{\lambda}$ are roughly proportional to $T$.
As there is typically less information in the data about the second order effect $\sigma^2(\bm x)$ than the first order effect $\mu(\bm x)$, it is sensible to shrink our model towards a homoskedastic model. Note that if all the $\lambda_{t\ell}$'s are equal to $0$ then the variance function reduces to $\sigma^{2}(\bm x) = \exp(-\lambda_0)$ so that the model is homoskedastic. Accordingly, we place a $\operatorname{half-Cauchy}(0,1)$ on the baseline standard deviation $\sigma_0 = \exp(-\lambda_0 / 2)$ and shrink the $\lambda_{t\ell}$'s heavily to zero.
As a default, we have found $a_\lambda = 0.5$ to work well in practice; alternatively, one might set $a_\lambda \sim \operatorname{half-Cauchy}(0,1)$ to allow the model to adaptively determine the amount of heteroskedasticity in the data.
Next, by analogy with the prior specification of \citet{chipman2010bart}, we ensure that the $\mu_{t\ell}$'s marginally have mean 0 and standard deviation $3/(k_\mu \sqrt{T})$ by noting that
\begin{math}
\operatorname{Var}(\mu_{t\ell}) = \beta / \{(\alpha - 1)\kappa\}.
\end{math}
As noted above, for moderate $T$ we will have $\alpha_{\lambda} \approx \beta_{\lambda} \propto T$, so that $\operatorname{Var}(\mu_{t\ell}) \approx \kappa^{-1}$. This suggests setting $\kappa^{-1/2} = k_{\mu} / \sqrt{T}$ (or giving $\kappa^{-1/2}$ a prior centered at this value). Here $k_\mu$ is a tuning parameter which controls the signal-to-noise ratio and by default we set $k_{\mu} = 1.5$.
We recommend a similar default prior for the gamma model. We first scale the non-zero $Y_i$'s to have mean $1$. As before, we impose the restrictions $E(\lambda_{t\ell}) = 0$ and $\operatorname{Var}(\lambda_{t\ell}) = a_\lambda^2 / T$ so that $h_{\lambda}(\bm x) \stackrel{\text{\scalebox{2}{\(.\)}}}{\sim} \mathcal N(0, a_\lambda^2)$. This can be accomplished by solving the system of equations~(\ref{eq:mean}, \ref{eq:var}). As a default, we set $a_\lambda = k_{\lambda}\sqrt{\operatorname{Var}(\log Y_i \mid Y_i > 0)}$ where $k_\lambda$ is a user-specified tuning parameter which we set to $1.5$. Additionally, we require a prior for the shape parameter $\alpha$. From \eqref{eq:gamma-moments} we see that $1/\alpha$ is a dispersion parameter. We use a weakly informative half-Cauchy prior $\alpha^{-1/2} \sim \operatorname{half-Cauchy}(0,A)$ for some $A > 0$. For the MEPS data in particular we set $A = 1$ to encourage small values of $\alpha$, as medical expenditures are highly right-skewed.
\subsection{Identifiability of the model components}
Given that the hurdle models we have proposed are mixture models, there is a question of whether the models we have defined here are identifiable. Let $\mathcal X$ denote a predictor space and $(\mathcal Y, \mathscr B)$ be a measurable space. Given a class $\{F_{\theta} : \theta \in \Theta\}$ where $F_{\theta} : \mathscr B \times \mathcal X \to [0,1]$ is a probability distribution on $(\mathcal Y, \mathscr B)$ for every $\bm x \in \mathcal X$, the parameter $\theta$ is called \emph{identifiable} if the mapping $\theta \mapsto F_{\theta}$ is one-to-one (\citealp{lehmann2006theory}, Definition 1.5.2). General forms of the hurdle model may not be identifiable, particularly when we are mixing a point mass at $0$ with a distribution that also is supported at $0$, or if the positive part has probability $0$. The following lemma shows that this is essentially the only case in which we might run into identifiability issues. A proof this result is given in the web appendix.
\begin{lemma}
\label{prop:ident}
Let $\mathcal X$ denote an arbitrary set, $(\mathbb R, \mathscr B)$ the Borel measurable reals, and $\delta_0$ the point-mass distribution at $0$.
Let $\mathscr G$ denote the set of conditional distributions with no atoms at $0$
\begin{align*}
\mathscr{G}= & \begin{Bmatrix}G:\mathscr{B}\times\mathcal{X}\to[0,1] & \text{such that } & G_{\bm x}(\cdot)\text{ is a probability distribution}\\
& \text{and} & G_{\bm x}(\{0\})=0\text{ for all }\bm{x}\in\mathcal{X}
\end{Bmatrix}
\end{align*}
and let $\mathscr P$ be the collection of conditional probabilities which are bounded by $1$,
\begin{math}
\mathscr P
=
\left\{
\pi: \mathcal X \to [0,1)
\right\}.
\end{math}
Let $\mathscr M$ denote the collection of conditional distributions on $(\mathbb R, \mathscr B)$ which are not identically $0$,
\begin{align*}
\mathscr{M}= & \begin{Bmatrix}M:\mathscr{B}\times\mathcal{X}\to[0,1] & \text{such that } & F_{\bm x}(\cdot)\text{ is a probability distribution}\\
& \text{and} & F_{\bm x}(\{0\}) \ne 1\text{ for all }\bm{x}\in\mathcal{X}
\end{Bmatrix}.
\end{align*}
Then the mapping $\mathscr G \times \mathscr P \to \mathscr M$ given by
\begin{math}
(G, \pi) \mapsto \pi(\bm x) \delta_0 + \{1-\pi(\bm x)\} G_{\bm x},
\end{math}
is a bijection.
\end{lemma}
A consequence of this result is that the semiparametric hurdle models developed in Section~\ref{sec:gamma} and Section~\ref{sec:lognormal} are also identifiable when the model parameters are understood to be the nonparametrically-specified functions $\bm h$ and the parametric component $\bm \omega$ (noting that $(\bm h, \bm \omega)$ maps in a one-to-one fashion to $(G,\pi)$).
The individual trees in the ensemble are, however, not identifiable, as the collection of possible regression trees is an overcomplete basis. In practice, we are usually only interested in recovering $\bm h$ rather than the individual trees, so that this lack of identifiability is not a concern.
\section{Simulation study}
\label{sec:simulation}
In this section, we examine the benefits of sharing information across related tasks using a simple simulation study. We consider a mixed response
\begin{align}
\label{eq:sim-model}
\Pr(Z_i = 1 \mid \bm X_i = \bm x)
= \Phi\{\sigma_\theta \, h(\bm x)\}, \qquad (Y_i \mid \bm X_i = \bm x) \sim \mathcal N\{h(\bm x), \sigma^2\},
\end{align}
with $(Z_i \perp Y_i \mid \bm X_i = \bm x)$. This is similar to the zero-inflated response setting, but with the continuous portion of the distribution always observed (see also Example~\ref{ex:mixed}). Note that the information in $\bm X_i$ is captured by the one-dimensional summary $h(\bm X_i)$ which is shared across both models. We emphasize that the structure \eqref{eq:sim-model} is not assumed by the shared forest model - only the basis functions are shared - and must effectively be learned from the data. We consider the benchmark function given by \citet{friedman1991multivariate}
\begin{align*}
h(\bm x) = 10 \sin(\pi x_1 x_2) + 20 (x_3 - 0.5)^2 + 10 x_4 + 5 x_5.
\end{align*}
We sample $\bm X_i$ uniformly distributed on $[0,1]^P$; if $P > 5$ then the predictors $X_{i6}, \ldots, X_{iP}$ have no influence on the response. We compare the shared forest to an approach which fits separate BART models to $(Z_i \mid \bm X_i = \bm x)$ and $(Y_i \mid \bm X_i = \bm x)$ so that information is not shared across tasks. Our focus is on how well these models estimate $\Pr(Z_i = 1 \mid \bm X_i) = \pi(\bm X_i)$ as measured by the cross-entropy between the true and estimated $\pi$'s,
\begin{align*}
\text{Loss}
= \int \left[\pi(\bm x) \log \left( \frac{\pi(\bm x)}{\widehat \pi(\bm x)} \right)
+ \{1 - \pi(\bm x)\} \log \left( \frac{1 - \pi(\bm x)}{1 - \widehat \pi(\bm x)} \right)
\right]
\ d\bm x,
\end{align*}
which is computed by Monte Carlo integration. We focus on the setting in which the continuous response $Y_i$ is relatively informative while the information contained in $Z_i$ is relatively weak by fixing $\sigma^2 = 1$. We consider a training set of size $n = 250$ for both the $Y_i$'s and the $Z_i$'s.
\begin{figure}
\centering
\includegraphics[width = 1\textwidth]{figure/simulation-share-noshare-1}
\caption{Left: average value of Loss for $\log P \in (\log 5, \log 250)$ averaged over 20 replications. Right: average value of Loss for $\sigma_\theta$ ranging from $1$ to $20$. ``No'' indicates the single BART model while ``Yes'' indicates use of the shared forest. This figure appears in color in the electronic version.}
\label{fig:simulation-share-noshare}
\end{figure}
Results are given in Figure~\ref{fig:simulation-share-noshare}, with $20$ replications per simulation setting. In the left panel, we fix $\sigma_\theta= 4$ (roughly corresponding to $\pi(\bm X_i) \sim \operatorname{Uniform}(0,1)$) and examine how sharing impacts the loss as $P$ varies from $P = 5$ to $P = 250$. We see that, as the variable selection task becomes more difficult, the model which does not share information is far more sensitive to irrelevant predictors than the model which does share. This is because the $Y_i$'s are much more informative about the relevant predictors than the $Z_i$'s, so that the shared model can do a much better job of selecting the relevant predictors. In the right panel, we fix $P = 20$ and vary the signal level $\sigma_\theta$ from $1$ to $20$. In this case, the gain from sharing is essentially constant, with higher losses for higher signal levels.
A potentially important tuning parameter in BART models is the number of trees $T$ used in the ensemble. Other works have supported the following overall trend: predictive performance of BART models is typically insensitive to the number of trees included, provided that we include sufficiently many. We find that this behavior holds for the shared forests model as well, with Figure~\ref{fig:tree-sim} summarizing the results of the simulation experiment with $P = 20$ and $\sigma_\theta = 4$ fixed as a function of $T$. As before, results are based on $20$ replications of the experiment for each setting. As expected we see a slight decrease in performance as $T$ increases from the optimal choice.
\begin{figure}
\centering
\includegraphics[width=.7\textwidth]{./figure/tree-simulation-1}
\caption{Average values of Loss for $T$ evenly space in $[10, 200]$ averaged over 20 replications. This figure appears in color in the electronic version.}
\label{fig:tree-sim}
\end{figure}
\section{Analysis of MEPS data}
\label{sec:analysis}
Our motivating example is from the 2015 Medical Expenditure Panel Survey (MEPS). The MEPS study \citep{natarajan2008variance} was developed to estimate national and regional health care use and expenditures in the United States. We first illustrate the capability of the proposed model to effectively capture heteroskedasticity in the MEPS data. We analyze data from 10,729 adult females who participated in the survey, focusing on the total medical expenditure during 2015, denoted $Y_i$. Previous analyses of this dataset have suggested taking $\operatorname{Var}(Y_i) = \phi E(Y_i)^{1.5}$ \citep{blough2000using, natarajan2008variance}. We consider a list of predictors including, among other things, age, race, family income, whether the individual smokes, perceived health, body mass index, and number of visits to the dentist over the survey period; a full list of predictors is given in the web appendix.
We fit the log-normal hurdle and gamma hurdle models to the data. We examine the fit of these models to the positive part of the data $(Y_i \mid Y_i > 0, \bm X = \bm x)$ by considering the generalized residuals \citep{cox1968general} $r_i = \Phi^{-1}\{\widehat F_{\bm X_i}(Y_i)\}$ where $\widehat F_{\bm x}$ is an estimate of the cdf of $(Y_i \mid Y_i > 0, \bm X_i = \bm x)$ obtained from the model. In the case of the log-normal hurdle model, $r_i$ is equivalent to the usual standardized residual $\{\log Y_i - \widehat \mu_i(\bm x)\} / \widehat \sigma(\bm x)$; for comparison, we also consider the raw residuals $\log Y_i - \widehat \mu_i(\bm x)$ to examine the effect of heteroskedasticity on the model fit.
\begin{figure}
\centering
\includegraphics[width=.9\textwidth]{figure/qq-1}
\caption{Quantile-quantile plots comparing the residuals $r_i$ for each model to a reference normal distribution. The top panels give the raw residuals $\log Y_i - \widehat \mu(\bm X_i)$ (left) and standardized residual $r_i$ for the log-normal hurdle model (right). The bottom panels give the residuals $r_i$ for the gamma hurdle (left) and the generalized gamma hurdle (right) models. This figure appears in color in the electronic version.}
\label{fig:qq}
\end{figure}
Quantile-quantile plots of the residuals compared to a reference Gaussian distribution are given in Figure~\ref{fig:qq}. We see that the log-normal hurdle model fits the data very well. Additionally, we see that ignoring heteroskedasticity causes a poor fit in the left tail of the data, corresponding to individuals with lower healthcare costs. By comparison, the gamma model fits poorly. We also consider a generalized gamma distribution \citep{stacy1962generalization} which models $Y_i^\phi$ with a gamma distribution, where $\phi$ is learned from the data. The generalized gamma model fits roughly as well as a homoskedastic log-normal model, but is inferior to the heteroskedastic log-normal model due to the stringent relationship between the mean and the variance implied by the generalized gamma model.
In addition to fitting the data well, the heteroskedastic log-normal model provides several interesting insights into the nature of the heteroskedasticity in the data. Let $\widehat m(\bm x)$ and $\widehat s(\bm x)$ denote the posterior mean of $m(\bm x)$ and $s(\bm x)$ given in Section~\ref{sec:lognormal}. The top panel of Figure~\ref{fig:hetero} gives a plot of $\widehat m(\bm X_i)$ against $\widehat s(\bm X_i)$ on the log-log scale. To aide visualization, points with similar values of $(\widehat m(\bm X_i), \widehat s(\bm X_i))$ are grouped into hexagonal tiles and are shaded according to the \emph{average number of dentist visits per subject} within each tile.
There are several interesting features of the top panel of Figure~\ref{fig:hetero}. First, there is near-linear relationship between $\log \widehat m(\bm X_i)$ and $\log \widehat s(\bm X_i)$. An ordinary least squares (OLS) fit of $\log \widehat m(\bm X_i)$ to $\log \widehat s(\bm X_i)$ has slope $0.7556$ and an $R^2$ of 82\%. Hence, the OLS fit suggests the approximation $\widehat s^2(\bm X_i) \propto \widehat m(\bm X_i)^{1.511}$, which agrees nearly exactly with \citet{blough2000using}. Second, by shading the hexagonal tiles by the number of dentist visits, we see clearly that the mean \emph{does not} account for all of the heteroskedasticity due to the predictors. We see, for example, that individuals with lower numbers of visits to the dentist tend to have a standard deviation which is higher than what would be predicted by the mean alone. To understand this relationship better, we let
\begin{math}
\delta = \log \widehat s(\bm X_i) - 0.7556 \log \widehat m(\bm X_i) - 2.672
\end{math}
denote the residual in predicting $\log \widehat s(\bm X_i)$ with $\log \widehat m(\bm X_i)$ by OLS. The bottom panel of Figure~\ref{fig:hetero} shows how the distribution of $\delta$ varies across the number of dentist trips and the individual's perceived health status. We see first that individuals with fewer dentist trips have standard deviations which are larger than what would be predicted using only the mean; similarly, individuals with higher perceived health status scores (corresponding to \emph{lower} perceived health) also tend to have higher variability than would be predicted by the mean alone.
\begin{figure}[!ht]
\centering
\includegraphics[width=.9\textwidth]{figure/hetero-1}
\includegraphics[width=.9\textwidth]{figure/hetero-2}
\caption{Top: Plot of $\widehat s^2(\bm X_i)$ against $\widehat m(\bm X_i)$ on the log-log scale; individual points are binned into hexagons, which are shaded according to the number of dentist visits the subject has. Bottom: boxplots of $\delta$ for the number of dentist trips and perceived health status. This figure appears in color in the electronic version.}
\label{fig:hetero}
\end{figure}
To assess whether there is a benefit of using the shared forests methodology for the MEPS data, we compute the log pseudo-marginal likelihood for the shared forest model and an equivalent model which does not share the trees across model components. Specifically, we fit the heteroskedastic BART (HBART) model of \citet{pratola2017heteroscedastic}, which sets $Y_i = g(\bm X_i) + s(X_i) \epsilon_i$ with $g(\bm X_i)$ and $\log s(\bm X_i)$ given BART priors. HBART was fit to the non-zero observed $Y_i$'s. The probability of a zero response was modeled using a binary BART model with a probit link, i.e., $\Pr(Y_i = 0 \mid \bm X_i = \bm x) = \Phi\{h(\bm x)\}$ where $h(\cdot)$ was given a BART prior. The log pseudo-marginal likelihood (LPML) is given by $\operatorname{LPML} = \sum_{i = 1}^n \log \operatorname{CPO}_i$ where $\operatorname{CPO}_i = f(Y_i \mid \bm Y_{-i}, \bm X)$ is the predictive density of the $i^{\text{th}}$ observation given $\bm Y_{-i} = (Y_1, \ldots, Y_{i-1}, Y_{i+1}, \ldots, Y_n)$ and $\bm X = (\bm X_1, \ldots, \bm X_n)$ \citep{geisser1979predictive}. Results are given in Table~\ref{tab:loo}. The $\operatorname{LPML}$ was computed using the Markov chain output using the \texttt{loo} package in \texttt{R} \citep{vehtari2017practical}. Multiple fits of the model using different seeds for the MCMC algorithm give qualitatively similar results.
\begin{table}
\centering
\begin{tabular}{lrr}
\toprule
& Shared & Not Shared \\
\midrule
Regression & -15166.9 & -15267.2 \\
Binary & -1552.8 & -2069.7 \\
Total & -16719.7 & -17336.9 \\
\bottomrule
\end{tabular}
\vspace{2em}
\caption{LPML of the model when the forests are shared across the mean, variance,
and hurdle components, compared with the LPML when the forests are
not shared. The row ``Regression'' gives the LPML contribution obtained
from $[Y_{i}\mid Y_{i}>0]$, while the row ``Binary'' give the LPML
contribution obtained from $I(Y_{i}>0)$; ``Total'' gives the final
LPML. }
\label{tab:loo}
\end{table}
We see that the shared forest gives a substantial boost in $\operatorname{LPML}$ for the binary component of the model. This suggests that the features learned from the continuous part of the model are helpful in determining whether an individual incurs any medical expense. We also observe a less substantial, but still large, improvement in performance for the regression model.
\section{Discussion}
\label{sec:discussion}
In this paper we introduced shared forests and demonstrated their usefulness on both simulated data and data from the MEPS dataset. Additionally, we introduced two novel models for semicontinuous data: a gamma hurdle model and a heteroskedastic log-normal hurdle model.
There are several promising areas for future work. First, there are other possibilities for sharing information across nonparametric components. Here we have restricted the components to share the same basis function expansion. To make the models more tightly coupled, one might consider shrinking together the coefficients of these expansions; an example where this might be useful is in meta-regression, where one would expect both that features across different studies will exert similar (but not necessarily identical) effects on the outcome. In the other direction, one might allow the models to share a \emph{subset} of the basis functions; for example, each model component might consist of a shared forest combined with an \emph{innovation forest} which is specific to each task. This structure is likely to be useful if only a subset of relevant features are shared across nonparametric components. A special case of such a construction is given by \citet{hahn2017bayesian} to estimate heterogeneous causal effects; in our terminology, their model consists of a shared forest which captures the prognostic features of covariates which are shared across treatment levels $z = 1$ and $z = 0$ and an innovation forest which is specific to the treatment $z = 1$.
Additionally, \citet{linero2017abayesian} recently demonstrated that the discrete nature of decision trees can lead to suboptimal performance on both a theoretical and practical level, and that this can be corrected by replacing the usual decision trees with \emph{smooth} decision trees. The shared forests framework can easily be extended to allow for smooth decision trees for the homoskedastic log-normal hurdle model, but non-trivial modifications are required to apply this strategy to the heteroskedastic log-normal and gamma hurdle models.
\section{Supporting Information}
A web appendix referenced in Sections 1, 3, and 5, and code implementing the methodology, are available with this paper at the Biometrics website on Wiley Online Library; code is also available at \url{www.github.com/theodds/SharedForestPaper}.
\section*{Acknowledgements}
This work was partially supported by NSF grant DMS-1712870 and by the Office of
the Secretary of Defense, Directorate of Operational Test and Evaluation and the
Test Resource Management Center under the Science of Test research consortium.
\bibliographystyle{apalike}
|
\section{Introduction}
It is not easy to describe a new approach to an old problem when the meaning of the new terminology is similar
but not exactly the same as in the established
publications. The longer way is to develop the required terminology
and definitions in
detail, but this takes time and the reader may lost patience before the essence of the new approach and its interest are
obtained.
Aiming to explain our approach in a relatively short exposition, we shall work here with a simple example with which the differences
between the two terminologies and approaches can be highlighted.
We shall present a very simple algorithm for two concurrent processes,
which we call ``Kishon's Poker Algorithm,''
and describe how two approaches handle its correctness proof:
the standard\footnote{When we refer here to the standard approach
it is just for convenience--we do not claim that the existing diversity can be reduced to a single one.} and
our proposed one. While this doesn't do justice to neither approaches, it
gives a leisurely discussion and a fairly good idea on their different
merits.
It should be said already at the beginning that our aim is not
to propose a better approach, but rather to enlarge the existing range of tools by an addition which may be valuable in certain
circumstances.
\section{Kishon's algorithm}
\label{KA} The writer and humorist
Ephraim Kishon is no longer popular as he was in my childhood and probably most of you have not read his sketch
``poker yehudi'' (Jewish Poker). Kishon meets his friend Arbinka who invented a game in which each of the two players has to think about a number
and the one who comes up with the greater number wins. At first, Kishon looses, but when he understands that he must bluff in order
to win he gets his revenge. Senseless as it is, the following algorithm is motivated by this game. The two players are the processes
$p_0$ and $p_1$, and they run concurrently their protocols (Figure \ref{JP}) just once.
\begin{figure}
\begin{tabular}{|l|l|}\hline
\begin{minipage}[t]{72mm}
$p_0$'s protocol\\
\begin{enumerate}
\item[$1_0$] $n_0 :=$ pick-a-number$(\,)$;
\item[$2_0$] $R_0 := n_0$;
\item[$3_0$] $v_0:= R_1$;
\item[$4_0$] \IF\ $(v_0=0\vee v_0=n_0)$ \then\
$\> \Val_0 :=0$ \\
\elseif\ $v_0<n_0$ \then\ $val_0 :=1$ \\
\ELSE\ \hspace{1.5mm} $val_0 := -1$;
\item[$5_0$] \END.
\end{enumerate}
\end{minipage}
&
\begin{minipage}[t]{72mm}
$p_1$'s protocol\\
\begin{enumerate}
\item[$1_1$] $n_1 :=$ pick-a-number$(\,)$;
\item[$2_1$] $R_1 := n_1$;
\item[$3_1$] $v_1:= R_0$;
\item[$4_1$] \IF\ $(v_1=0\vee v_1=n_1)$ \then\ $val_1 :=0$\\
\elseif\ $v_1<n_1$ \then\ $val_1 :=1$ \\
\ELSE\ $val_1 := -1$;
\item[$5_1$] \END.
\end{enumerate}
\end{minipage}\\
\hline
\end{tabular}
\caption{The Kishon's Poker algorithm. Registers $R_0$ and $R_1$ are initially $0$. Procedure pick-a-number returns a natural
number $>0$.}
\label{JP}
\end{figure}
The processes communicate with two registers, $R_0$ and $R_1$, written by $p_0$ and $p_1$ respectively (and read by the other process).
These registers carry natural numbers and their initial value is $0$. We assume first, for simplicity, that the registers
are serial, and later on we shall deal with regular registers (which are even more interesting from our point of view,
see section \ref{S3.1}).
The local variables of process $p_i$ ($i=0,1$) are $n_i$ and $v_i$ which carry natural numbers and have initial value $0$, and the variable
$val_i$ which carries a value in $\{-1,0,1\}$ and is initially $0$.
We assume a procedure ``pick-a-number'' which returns a (randomly chosen) natural number greater than $0$.
If $E$ is an execution of the protocol by $p_i$ ($i=0,1$) then $E$ consists of four event, $E_1,\ldots,E_4$ which are executions of
the instructions at lines $1_i,\ldots,4_i$ of $p_i$'s protocol. Thus, for example, if $E$ is a protocol execution by $p_0$, then
$E_1$ is the event of invoking pick-a-number and assigning the returned non-zero natural number to variable $n_0$.
$E_2$ is the write event on register $R_0$ that corresponds to line 2, $E_3$ is the read
event of register $R_1$ and the assignment of the value obtained to variable $v_0$.
Finally, $E_4$ is the event of assigning a numeric value to variable $val_0$.
If $x$ is any of the program variables or the register $R_i$ of $p_i$, then there is exactly one instruction that can change the initial value of $x$,
and we denote with $x(E)$ the value of $x$ after it has been finally determined in $E$. For example, if $E$ is a protocol execution by $p_0$, then
$v_0(E)$ is the value of variable $v_0$ after event $E_3$. That is, $v_0(E)$ is the value (assigned to $v_0$) obtained by reading register $R_1$.
Suppose for example that $E$ is a protocol execution by $p_0$. Concerning $val_0(E)$ there are four distinct possibilities:
\begin{enumerate}
\item If $v_0(E)=0$, then $val_0(E)=0$. Otherwise, $v_0(E)$ is compared to $n_0(E)$ and we have the following possibilities:
\item If $v_0(E)=n_0(E)$ then $val_0(E)=0$.
\item If $v_0(E)<n_0(E)$
then $val_0(E)=1$ (indicating that $p_0$ believes to win the Kishon's game),
\item but if $v_0(E)> n_0(E)$ then $p_0$ sets $val_0(E)=-1$ (admitting it losts the play).
\end{enumerate}
The idea in the way that $p_0$ determines $val_0$ is simple. In case $v_0(E)=0$, $p_0$ ``knows'' that it has read the
initial value of register $R_1$ and not the value of $p_1$'s pick-a-number execution (which must be $>0$). In this case, $p_0$
has no way of knowing whether it is a winner of the game or not, and it sets $val_0$ to the neutral value $0$. In the three remaining
cases, $p_0$ has in its hands the two pick-a-number values, of $p_1$ and its own, and can decide whether $val_0$ should be $+1$ or $-1$
(or $0$ in case of equality).
We assume that the Kishon's poker game is played just once, and our aim is to prove that
the process that obtains the higher number in its pick-a-number wins the play (i.e. gets a higher $\Val_i$ value than the other player). We will prove the following theorem twice, and thus exemplify in a very simple setting two approaches to the problem of
proving the correctness of distributed algorithms.
\begin{theorem}
\label{T1}
Suppose $E$ and $F$ are concurrent protocol executions by $p_0$ and $p_1$ respectively of their algorithms. If
$n_0(E)<n_1(F)$, then $val_0(E)<val_1(F)$. And symmetrically if $n_1(F)<n_0(E)$ then $val_1(F)<val_0(E)$. If $n_0(E)=n_1(F)$ then $val_0(E)=val_1(F)=0$.
\end{theorem}
In this section we argue informally for this theorem, and to simplify the proof we assume for the moment that the registers are serial.
This assumption will be relaxed in sections \ref{SRR} and \ref{S4} which deal with regular registers.
Register $R$ is serial if:
\begin{enumerate}
\item the read/write events on that register are linearly ordered,
and
\item the value of any read $r$ of $R$ is equal to the value of the last
write event $w$ on $R$ that precedes $r$ or to the initial value of that register in case there is no write event
that precedes $r$.
\end{enumerate}
In section \ref{S3} we describe the invariant based
approach to the proof of this theorem (this is the standard approach), and then in section \ref{Sec6}
we reprove this theorem in our model theoretic formulation and approach.
\subsubsection*{Informal argument for Theorem \ref{T1}.}
\label{SSInf}
Let $E$ and $F$ be protocol execution by $p_0$ and $p_1$ (respectively) which are assumed to be executed concurrently.
Recal our convention to denote with $x(E)$ (and $x(F)$) the values of any variable $x$ of $p_0$ (respectively $p_1$) at the
end of $E$ ($F$).
Assume for example
that \begin{equation}
\label{E1}
n_0(E)<n_1(F).
\end{equation} That is, assume that the number obtained by $p_1$ in its pick-a-number
is greater than the one obtained by $p_0$. We have to show that
$val_0(E)<val_1(F)$ in this case. The other items of theorem \ref{T1} will follow by a similar argument.
We denote with $E_1, E_2, E_3, E_4$ the events in $E$ (executions of the four instructions at lines $1_0,\ldots,4_0$ of the protocol of $p_0$) and with
$F_1, \cdots, F_4$ the corresponding events in $F$. For any two events $a$ and $b$ we write $a<b$ to say that $a$ precedes $b$ (i.e. $a$
terminates before $b$ begins)\footnote{So $<$ is used for both the number ordering and the temporal precedence ordering. In simple situations such as ours this should create no confusion.}. Thus we have $E_1< E_2 < E_3 < E_4$
and $F_1<F_2<F_3<F_4$, but any interleaving of the $E$ events and the $F$ events is possible.
The key fact from which our theorem follows is that each process first writes and then reads. So $E_2<E_3$, i.e. the execution of
$R_0:= n_0$ precedes that of $v_0:= R_1$, and likewise $F_2<F_3$.
There is just one write event on register $R_0$, namely the write $E_2$ which
is of value $n_0(E)$.
Likewise there is just one write event on register $R_1$: the write $F_2$ of value $n_1(F)$.
Comparing the write $E_2$ on register $R_0$ with the read $F_3$ of that register by $p_1$ there are two possibilities.
\begin{enumerate}
\item $F_3<E_2$. In this case $F_3$ obtains the initial value $0$, and then $v_1(F)=0$ which implies that $val_1(F)=0$.
But as we have that $F_2<F_3<E_2<E_3$, we get that $F_2<E_3$, and hence the read $E_3$ of register $R_1$ obtains the value
of the write $F_2$ on that register which is of value $n_1(F)$. Since we assume that $n_0(E)<n_1(F)$, it follows that $val_0(E)=-1$. Thus we have in this
case that $val_0(E)<val_1(F)$.
\item $E_2<F_3$. In this case $F_3$ obtains the value $n_0(E)$ that was written by $E_2$ on $R_0$, and since we assume that
$n_0(E)<n_1(F)$, $val_1(F)=1$. Depending on the temporal relation between the write $F_2$ and the read $E_3$,
the read event $E_3$ can either return the initial value $0$ of register $R_1$ (in which case $val_0(E)=0$)
or else the value of the write $F_2$ (in which case $val_0(E)=-1$). In both cases we get that
$val_0(E)<val_1(F)$.
\end{enumerate}
I hope that my readers find this proof of Theorem \ref{T1} satisfactory, and so they may ask ``In what sense that proof is informal, and what is meant here by the term formal proof''? This question is discussed in section \ref{S3.1}, and at this stage let me only make the following remark. The problem with the proof outlined here is not so much
in that some details are missing or some assumptions are hidden, but rather in that the connection
between the algorithm text and the resulting executions is not established. In other words,
the proof provides no answer to the question of defining the structures that represent possible executions of the algorithm,
and hence the very base of the proof is missing.
The standard state-and-history approach gives a satisfactory answer to this question by defining what are executions (also called histories, i.e. sequences of states) and we describe in the following section this standard approach and find
an invariant that proves theorem \ref{T1}.
Then, in the later sections we outline a formal, mathematical proof of theorem \ref{T1} that follows the steps of the informal
description that was given above. That proof is necessarily longer because it requires some background preparation, namely the
definition of system-executions as Tarskian structures\footnote{The term ``structure'' is so overloaded that we have to be more
specific, and with {\em Tarskian structures} we refer here to those structures that are used in model theory
as interpretations of first-order languages (see also Section \ref{SReg}).}. As the reader will notice, that proof with Tarskian system executions has the same structure as the informal argument
that was given above, and this similarity
between the informal argument and its formal development is in my opinion an important advantage of the event-based model theoretic approach that we sketch in this paper.
\section{State and history proof of Kishon's algorithm}
\label{S3}
In this section the term {\em state} refers to the notion of {\em global state}, but in the
event-based approach, as we shall later see, a different notion of state is employed in which
only local states of individual serial processes are used.
A (global) state of a distributed system is a description of the system at some instant, as if taken by some global snapshot.
A state is a function that assigns values to all the state variables. The state variables of the Kishon's algorithm
are the program variables ($R_i, n_i,v_i$ and $val_i$ for $i=0,1$) and the two program counters $PC_0$ and $PC_1$ of $p_0$ and $p_1$.
If $S$ is a state and $x$ is a state variable, then $S(x)$ is the value
of $x$ at state $S$. $S(PC_0)=3_0$ for example means (intuitively) that $p_0$ is about to execute the instruction at line $3_0$ of its
protocol. A {\em step} is a pair of states
$(S,T)$ that describes an execution of an instruction of the protocol, either the instruction at line $S(PC_0)$ if this is a step by $p_0$
or the instruction at line $S(PC_1)$ if this is a step by $p_1$. And $T$ is the resulting state of that execution.
For example, a pair of states $(S,T)$ is a read-of-register-$R_1$ by $p_0$ (also said to be a $(3_0,4_0)$ step) if the following hold.
\begin{enumerate}
\item $S(PC_0)=3_0$ and $T(PC_0)=4_0$,
\item $T(v_0)=S(R_1)$,
\item for any state variable $x$ other than $PC_0$ and $v_0$, $T(x)=S(x)$.
\end{enumerate}
We leave it to the reader to define all eight classes of steps: $(1_0,2_0),\ldots,(4_0,5_0)$ and $(1_1,2_1),\ldots,(4_1,5_1)$.
An initial state, is a state $S$ for which $S(PC_0)=1_0$, $S(PC_1)=1_1$, and $S(x)=0$ for any other state variable.
(In particular $S(R_0)=S(R_1)=0$).
A final state is a state $S$ for which $S(PC_0)=5_0$ and $S(PC_1)=5_1$. Note that if $S$ is not a final state then there exists
a state $T$ such that $(S,T)$ is a step.
Each state variable has a type. For example, the type of $v_i$ is the set of natural numbers. We denote with $\tau$ the conjunction
which says that each variable is in its type. Clearly the value of each variable
in the initial state is in its type. A simple observation is that
if $(S,T)$ is any step and for every state variable $x$, $S(x)$ is in the type of $x$, then $T(x)$ is also in that type.
The basic sentential formulas are equality and comparison of state variables. For example $R_0=v_1$, $PC_0\geq 2$, and
$n_0>0$ are basic formulas. Statements about the types of variables are also basic. For example, ``$n_0$ is in $\mathbb N$''.
If $S$ is any state and $\varphi$ a sentential formula, then $S\models \varphi$ means that $\varphi$ holds in $S$. For example
$S\models v_0<n_0$ iff $S(v_0)<S(n_0)$. Similarly $S\models PC_0\leq 2_0$ iff $S(PC_0)=1_0,2_0$.
A sentential formula is obtained from basic (primitive) sentential formulas with logical connectives (conjunction, disjunction, negation and
implication).
A sentential formula $\varphi$ is said to be an invariant (also, an inductive invariant) if the following holds.
\begin{enumerate}
\item If $S$ is the initial state then $S\models\varphi$.
\item For every step $(S,T)$, if $S\models \varphi$ then also $T\models\varphi$.
\end{enumerate}
To prove that $\varphi$ is an invariant one has to prove that it holds in the initial state, and then to prove for every step $(S,T)$
that either $S\not\models\varphi$ or $T\models \varphi$.
A {\em history sequence} is a sequence of states $S_0,\ldots,S_k,\ldots$ such that (1) $S_0$ is the initial state, (2) for every state $S_j$ in the sequence that is not a final state, $S_{j+1}$ exists in the sequence and
$(S_j,S_{j+1})$ is a step by $p_0$ or by $p_1$. In our simple algorithm, every terminating history has
exactly four steps by $p_0$ and four by $p_1$. So a terminating history sequence
for the Kishon's algorithm has the form $S_0,\ldots,S_8$. There are as many histories as there are possible ways
of interleavings of $p_0$ and $p_1$ steps.
The following easy theorem is of prime importance even though its proof is immediate:
\begin{quote}
If $\varphi$ is an invariant formula, and $S_0,\ldots,S_k$ is any history sequence, then
$S_m\models \varphi$ for every index $0\leq m\leq k$.
\end{quote}
The counterpart of Theorem \ref{T1} in this history context is the following.
\begin{theorem}
\label{T2}
If $S_0,\ldots,S_8$ is a terminating history sequence of the Kishon's algorithm, then for each $i\in\{0,1\}$ \[S_8 \models (n_i<n_{1-i}\to val_i <val_{1-i}).\]
\end{theorem}
To prove this theorem we have to find an invariant $\varphi$ such that for each $i\in\{0,1\}$
\begin{equation}
\label{Eq}(\varphi\wedge PC_0=5_0\wedge PC_1=5_1)\to (n_i<n_{1-i}\to val_i <val_{1-i}).\end{equation}
\begin{equation}
\label{Eqq}(\varphi\wedge PC_0=5_0\wedge PC_1=5_1)\to (n_0=n_{1}\to val_0 = val_{1}).\end{equation}
The difficult part of the proof (which is a typical difficulty for this type of proofs) is to find the required invariant. Once it is found, the proof that $\varphi$ is indeed an invariant is a routine checking.
The invariant that I have found for this theorem is described now.
Let $\alpha$ be the conjunction of the following five sentences.
\begin{enumerate}
\item[$\alpha1.$] $PC_0\geq 2_0\to n_0>0$.
\item[$\alpha2.$] $PC_0\leq 2_0\to R_0=0$.
\item[$\alpha3.$] $PC_0\geq 3_0\to R_0=n_0$.
\item[$\alpha4.$] $v_0\not= 0\to v_0=R_1$.
\item[$\alpha5.$] $PC_0=5_0\to $
\hspace*{7pt}$(v_0=0\to val_0=0 )\wedge (v_0=n_0\to val_0=0)\wedge$\\
$\hspace*{7pt} (0<v_0<n_0\to val_0=1 )\wedge(v_0>n_0\to val_0=-1)$.
\end{enumerate}
Let $\beta$ the conjunction of the corresponding five sentences.
\begin{enumerate}
\item[$\beta1$.] $PC_1\geq 2_1\to n_1>0$.
\item[$\beta2$.] $PC_1\leq 2_1\to R_1=0$.
\item[$\beta3$.] $PC_1\geq 3_1\to R_1=n_1$.
\item[$\beta4$.] $v_1\not= 0\to v_1= R_0$.
\item[$\beta5$.] $PC_1=5_1\to $
$\hspace{2mm}(v_1=0\to val_1=0 )\wedge (v_1=n_1\to val_1=0)\wedge$\\
$\hspace*{2mm} (0<v_1<n_1\to val_1=1 ) \wedge (v_1>n_1\to val_1=-1)$.
\end{enumerate}
Let $\gamma$ be the following sentence.
\[PC_0\geq 4_0 \wedge PC_1\geq 4_1\, \to\, v_0=R_1 \vee v_1=R_0.\]
Let $\tau$ be the sentence saying that each state variable is in its type.
Our invariant $\varphi$ is the conjunction \[\alpha\wedge\beta\wedge \gamma\wedge \tau.\]
To prove that $\varphi$ is an invariant (of the Kishon's algorithm) we have first to prove that $\varphi$ holds in the initial state,
and then prove for every step $(S,T)$ that if $S\models \varphi$, then $T\models\varphi$ as well. In proving this implication, one can rely only on the definition of steps:
the program itself cannot be consulted at this stage.
\begin{theorem} $\varphi$ is an invariant.\end{theorem}
We will not go over this lengthy (and rather routine)
checking, since what interests us here are the concepts involved with the invariant method rather than the details of the proof.
Assuming that $\varphi$ is indeed an invariant, we shall conclude the proof of theorem \ref{T2} by proving that equations (\ref{Eq})
and (\ref{Eqq}) hold for $i=0,1$.
This is obtained immediately from the following lemma.
\begin{lemma}
\label{L3.2}
Assume that $S$ is any state such that
$S\models \varphi\wedge PC_0=5_0 \wedge PC_1=5_1$. Then
\begin{equation}
\label{E2}
S\models (n_0=n_1\to val_0=val_1 = 0), \text{ and }
\end{equation}
\begin{equation}
\label{E2a}
S\models n_i<n_{1-i} \to (val_i<val_{1-i})
\end{equation}
for $i=0,1$.
\end{lemma}
\begin{proof}
We prove only (\ref{E2a}) (since the proof of (\ref{E2}) uses similar arguments) and only for $i=0$.
Assume that $S$ is any state such that
$S\models \varphi\wedge PC_0=5_0 \wedge PC_1=5_1$.
Assume also that \[S\models n_0<n_1.\]
Since $S\models \gamma$, there are two cases in the proof that $S\models val_0<val_1$: $S\models v_0 = R_1$, and $S\models v_1 = R_0$.
{\bf Case} $S\models v_0=R_1$. Since $S\models PC_1\geq 3_1$, $\beta3$ implies that $R_1=n_1$. We assume that $n_0<n_1$, and
hence $n_0<v_0$ in $S$. This implies that $S\models \Val_0=-1$ (by $\alpha5$).
The possible values for $\Val_1$ are $-1,0,1$, and so it suffices to exclude the possibility of $\Val_1=-1$ in $S$ in order to
conclude that $S\models \Val_0<\Val_1$. Suppose, for a contradiction, that $\Val_1=-1$ in $S$. This implies (by $\beta5$)
that $v_1>n_1$ in $S$. Hence in particular $v_1\not=0$, and $\beta4$ implies that $v_1=R_0$ in $S$. But $R_0=n_0$ (by $\alpha3$)
and so $n_0=v_1>n_1$ follows in $S$ in contradiction to our assumption above.
{\bf Case} $S\models v_1=R_0$. By $\alpha3$, $R_0=n_0$, and since $n_0>0$ (by $\alpha1$) we get $0<n_0=R_0=v_1 $. Since $n_0< n_1$ in $S$,
$0<v_1<n_1$ in $S$.
So $\Val_1=1$ in $S$. Again, the possible values of $\Val_0$ are $-1,0,1$, and so it suffices to exclude the possibility of
$\Val_0=1$ in order to conclude that $S\models \Val_0<\Val_1$. Suppose for a contradiction that $\Val_0=1$, and conclude that
$0<v_0<n_0$ in $S$. But $v_0=R_1$ (by $\alpha4$) and $R_1=n_1$ in $S$ (by $\beta3$) and hence $n_1<n_0$ in contradiction
to our assumption.\end{proof}
\subsection{An intermediate discussion}
\label{S3.1}
The term ``formal proof'' has a rather definite meaning in mathematical logic; namely it refers to some set of deduction
rules and axioms in some formal language that determine which sequences of formulas in that language form a proof.
But, in this article, we intend a less stringent usage of this term. A formal-mathematical proof (as opposed to an intuitive informal
argument)
is a rigorous mathematical argument that establishes the truth of some sentence written in a formal language.
We require three prerequisites that an argument has to have in order to be considered a rigorous
mathematical proof in this sense.
(1) There has to be a formal language in which the theorem is stated and in which the main steps of the proof are formulated. (2) there
is a mathematical definition for the class of structures that
interpret the language and a definition for the satisfaction relation that determines when a sentence
$\varphi$ holds in a structure $M$. (3) A proof of a sentence $\varphi$ has to establish that $\varphi$ holds (is satisfied) in every structure
in the intended manifold of structures. In this sense, the proof of theorem \ref{T2} given here in the state-history approach is a satisfying formal-mathematical proof. Of course, some details were omitted and the definitions were not sufficiently
general (they were tailored to the specific Kishon's algorithm), but the desired prerequisites are there: the language is the sentential
language, and the relation $S\models\varphi$ for a state $S$ and sentence $\varphi$ is (as the reader knows)
well-defined. With the notions of steps and history
sequences, the framework for proving theorem \ref{T2} is mathematically satisfying, even if my presentation has left something to be desired.
This is not the case with the informal argument for theorem \ref{T1}: the language in which this theorem is enunciated is not defined, and more
importantly, the structures to which this theorem refers to are not defined. That is, we did not define those structures that
represent executions of the algorithm. As a consequence of these deficiencies the informal argument presented in \ref{SSInf}
does not establish any formal connection between the algorithm text and its executions. We will establish such a connection
in section \ref{SReg}, essentially by transforming a certain type of history
sequences into first-order structures. My aim in doing so is
to show how the informal argument for Kishon's algorithm can be transformed into a rigorous mathematical proof, and thus to exemplify the event-based proof framework for distributed systems. To say it in two words, the
formal language for this framework is first-order predicate language, and the structures are Tarskian structures, i.e. interpretations of that predicate language.
The invariant proof method is not without its own typical problems which should be exposed in order to promote the possible
relevance of alternative or complimentary methods.
\noindent
{\bf The difficulty in finding an invariant.} An advantage of the standard state and history based proofs is not only that they yield satisfying mathematical proofs but also that the required notions
on which such proofs are based
(states, steps, history-sequences etc.) are easily defined. Yet, and this is a major problem, these proofs are based on the notion of {\em invariant}, and these
invariants are notoriously difficult (and sometimes very difficult) to obtain. It took me almost a day to find the invariant
$\alpha\wedge\beta\wedge\gamma$ (checking that it is indeed an invariant can also be long, but with software tools this would be
an instantaneous work). The reader who is perhaps more experienced in finding invariants may need less
than a day, but surely more than the five or ten minutes that it takes to find a convincing intuitive argument for the event-based
theorem \ref{T1} which relies on the checking of different temporal event-scenarios.
{\bf The state and history approach does not show in a clear way where the assumptions on the communication devices are used.}
The role of an assumption in a mathematical proof is often clarified by either finding a proof with a weaker assumption or else
by showing the necessity of that assumption.
In contrast, Theorem \ref{T1} has an assumption that the registers are serial, but in the assertion-based proof of that theorem this assumption somehow disappears. It is hidden, in a sense, in the definition of steps.
History sequences, as defined here, can only have serial registers, and thus the status of register seriality as an assumption is unclear.
{\bf Seriality of the registers is a consequence, a theorem, not an assumption.}
From a different angle, the preceding point can be explained by considering a history sequence $H$ of some algorithm in which the registers
are assumed to be serial. Seriality of a register means that for every reading step $r$ of the register in $H$, the value of $r$ is the value of the last
previous write step in $H$ that precedes $r$ (or the initial value of $R$ in case no such write step exists).
Now the seriality of $R$ in history $H$ is a {\em theorem} not an assumption on the history. To prove this theorem, we can survey all steps and realize that
only write steps may change the value of the register in a state, and from this we can deduce that the value of any read step is the value of the last write step
in $H$ that precedes that read. Being a theorem, a consequence, how can we say that seriality is an assumption?
{\bf The problem with proving correctness for regular registers.}
This point can be clarified even further by considering regular and safe registers. We may ask: does Kishon's algorithm retain its properties
with only regular (or safe) registers? We will see in the following subsection that almost the same informal
argument for theorem \ref{T1} works for regular registers, but it would be a difficult challenge to prove the correctness
of Kishon's algorithm for regular registers in the state-histories approach. In order to represent a single-writer regular register
$R$, one has to represent each extended event
of process $p$ (such as a read or write event of some register) by a pair of {\em invoke} and {\em respond} actions by $p$ that may
appear in a history sequence with several actions of other processes in between. Moreover, the writing process has to record two values,
the current value and the previous value, and the reading process has to keep an active bag of possible values so that its return value is one of these possible values. It is not impossible to represents regular registers in such a way but it is quite complicated and I am not aware of any published invariant-based correctness proof for an algorithm that uses these registers.
\subsubsection{Regular registers}
\label{SRR}
The notions of safe, regular, and atomic registers were introduced by Lamport \cite{SE} in order to investigate
the situation where read and write operation executions can be concurrent. Thus, seriality is no longer assumed for these events.
A (single-writer) register $R$ is regular if there is a specific process that can write on $R$, and any read
of $R$ (by any process) returns a value $v$ that satisfies the following requirement. If there is no write event that
precedes that read, then $v$ is equal to the initial value of $R$ or to the value of some write event that is concurrent with
the read. If there is a write that precedes the read, then the value of the read is the value of the last write on $R$ that precedes the read, or the value of some
write on $R$ that is concurrent with the read.
In the context of regular registers, the precedence ordering $<$ on the events is a partial ordering which is not necessarily linear.
An important property of that partial ordering is the Russell-Wiener property.
\begin{equation}
\label{RW}
\text{For all events }a,b,c,d, \text{ if }a<b,c<d \text{ and } \neg( c<b), \text{ then } a<d.
\end{equation}
The justification for this property is evident when
we think about interval orderings (further details are in Section \ref{SReg}).
Let's repeat the intuitive proof of Theorem \ref{T1} for the Kishon's algorithm, but now for regular registers.
\noindent{\em Correctness of Kishon's Poker algorithm with regular registers.}
Let E and F be concurrent protocol execution by $p_0$ and $p_1$ (respectively). Recall (from section \ref{KA}) that $E$ consists of four
events $E_1,\ldots,E_4$ that correspond to executions of lines $1_0,\ldots,4_0$, and $F$ contains the corresponding four events of $p_1$:
$F_1,\ldots,F_4$. Assume for example that
\begin{equation}
\label{E7}
n_0 (E) < n_ 1 (F);
\end{equation}
The initial values of the registers is $0$, and there is just one write event on register $R_0$, namely the write $E_2$
which executes $R_0 := n_0$ and is of value $n_0(E)>0$. Hence (by the regularity of $R_0$) there are only two values that a
read of register $R_0$ can return: the initial value $0$ and the value $n_0(E)>0$ of the write event $E_2$. Applying this
observation to $F_3$ (the read event of $R_0$ by $p_1$), the possible values of $F_3$ are $0$ and $n_0(E)>0$. We observe that
if $E_2<F_3$, then the value of $F_3$ is necessarily $n_0(E)$.
Likewise there is only one write event on register $R_1$, namely $F_2$ which is of value $n_1(F)>0$, and hence the regularity
of $R_1$ implies that there are only two values that a read of register $R_1$ can return: the initial value $0$ and the
value $n_1(F)>0$ of the write event $F_2$. Applying this observation to the read $E_3$,
\begin{equation}
\label{Eq8}
\text{the possible values of } E_3 \ \text{are $0$ and $n_1(F)$.}
\end{equation} We observe that if $F_2<E_3$ then the value of $E_3$ is necessarily $n_1(F)$.
Taking into account our assumption (\ref{E7}), there are only two possible values for $\Val_0(E)$:
$0$ and $-1$ (use (\ref{Eq8})). And if $F_2<E_3$ then $\Val_0(E)$ is $-1$. Likewise, there are only two possible values
for $\Val_1(F)$: $0$ and $1$, and if $E_2<F_3$ then $\Val_1(F)$ is $1$.
So, in order to prove that $\Val_0(E)<\Val_1(F)$, we have to exclude the possibility that $\Val_0(E)=0 \wedge \Val_1(F)=0$.
We observed that this possibility is indeed excluded if $F_2<E_3$ or if $E_2<F_3$. Therefore the following lemma establishes
that $\Val_0(E)<\Val_1(F)$ and proves theorem \ref{T1} for regular registers.
\begin{lemma}
For every concurrent protocol executions $E$ and $F$ by $p_0$ and $p_1$, $F_2<E_3 \vee E_2<F_3$.
\end{lemma}
\begin{proof} Assume that $\neg (F_2< E_3)$. Then we have the following temporal relations.
\[ E_2<E_3,\ \neg (F_2< E_3),\ F_2<F_3.\]
Hence the Russell--Wiener property implies that $E_2<F_3$ as required. \end{proof}
The second part of our paper, sections \ref{SReg}, \ref{Sec6}, and \ref{S4}, is its main part in which we prove the correctness of
Kishon's Algorithm with regular registers in the event and model theoretic framework. Actually, the main part of the
proof takes just a couple of pages, but the description of its framework necessitates a redefinition of states as (finite) Tarskian
system executions which takes some place.
\section{Tarskian system executions and regular registers}
\label{SReg}
In this section we define the notion of Tarskian system execution that are used to explicate regular registers. This notion
relies on the notions of first-order language and interpretation which are defined and explained in any logic textbook.
Here we shall describe these notions in a concise way mainly by following an example.
A Tarskian structure is an interpretation of a first-order language. A language, $L$, is specified
by listing its symbols, that is its sorts (for we shall use many-sorted languages), its predicates, function symbols, and individual constants.
For example, the
language $L_{R}$ that we design now is used for specifying that $R$ is a single-writer regular register (owned by the writer process $p$). The following symbols are in $L_R$.
\begin{enumerate}
\item There are two sorts \Event\ and \Data.
\item The unary predicates are
$\Readit_R$ and $\Writeit_R$ and the predicate $p$.
Formula $\Read_R(e)$ for example says that event $e$ is a read event of register $R$, and formula $p(e)$ says that event $e$
is by process $p$ (the writer process).
\item A binary predicate $<$ (called the temporal precedence relation) is defined over the events.
\item There is one function symbol $\Val$, and one individual constants $d_R$ (called the initial value of register $R$).
\end{enumerate}
The language $L$ has an infinite set of variables, and it is also convenient to have separate sets of variables for each type.
We use here lower-case letters from the beginning of the alphabet, possibly with indexes, as \Event\ variables,
and the letters
$x$ and $y$ with or without indexes are used as general variables. It is convenient when designing $L_R$ to be more specific and
to determine that predicates $\Readit_R$, $\Writeit_R$, $p$, and $<$ apply to sort \Event, that $\Val$ is a function from sort \Event\ to sort
\Data, and that the constant $d_R$ is in sort \Data.
A Tarskian structure $M$ that interprets this language $L$ consists of a set $U = U^M$, the ``universe of $M$'' which is the disjoint
union of two subsets: $\Event^M$ and $\Data^M$ which are the interpretation of the \Event\ and \Data\ sorts by $M$. Then the unary predicates
$\Writeit_R$, $\Readit_R$, and $p$
are interpreted as subsets of $\Event^M$. The binary predicate $<$ is interpreted as a set of pairs $<^M\subset \Event^M\times \Event^M$
(called the temporal precedence relation). $\Val$ is
interpreted as a function $\Val^M:\Event^M\to \Data^M$. And the (initial value) constant $d_R$ is interpreted as a member
$d_R^M$ of $\Data^M$.
An interpretation
of $L_R$ is a rather arbitrary structure $M$ which does not necessarily correspond to the idea that we
have of register behavior. It is by means of $L_R$ formulas that we can impose some discipline on the interpretations, and thus
define the notion of register regularity as the class of those interpretations of $L_R$ that satisfy the required regularity formulas.
Formulas of a logical language are
formed by means of its predicates, function symbols and constants, together with the logical symbols: $\forall$ and $\exists$ (these are the quantifiers), and
$\wedge$, $\vee$, $\neg$, $\to$ (these are the connectives).
Here are some examples of formulas. $\forall x (\Event(x)\to \Writeit(x)\vee \Readit(x))$ says that every event is a \Writeit\ event or a \Readit\ event. Recalling that variable $a$ is restricted to events, we can write this as $\forall a(\Writeit(a)\vee \Readit(a))$. (This sentence doesn't exclude the possibility that an event is both a read and
a write, and we can add $\neg\exists a (\Writeit(a)\wedge\Readit(a))$ to do just that). A formula with no free variables is said to be a sentence, and any sentence $\varphi$ may be true in $M$, $M\models \varphi$, or false $M\not\models \varphi$.
We can say that
$<$ is a partial ordering on sort \Event\ (i.e. an irreflexive and transitive relation) as follows.
\[ \forall a (a\not < a)\wedge \forall a,b,c(a<b\wedge b<c\to a<c).\]
Here $a\not < a$ is another way to write $\neg(a<a)$.
If $a$ and $b$ are incomparable events in the $<$ temporal precedence relation then we say that $a$ and $b$ are concurrent.
That is, the formula ``$a$ is concurrent with $b$'' is $\neg(a<b\vee b<a)$. (Thus any event is concurrent with itself.)
\begin{definition}
\label{TD}
Let $M$ be an interpretation of our language $L_R$ (or a similarly defined language).
We say that $M$ is a ``Tarskian system-execution
interpretation of $L_R$''
if \end{definition}
\begin{enumerate}
\item $<^M$ is a partial ordering that satisfies the Russell-Wiener property (see equation \ref{RW}). Equivalently,
\begin{equation}
\forall a,b,c,d (a<b\wedge (b\text{ is concurrent with }c) \wedge c < d\to a<d).\end{equation}
\item
For every event $e\in E^M$, the set $\{x\in E^M\mid x <^M e\}$ is finite, and likewise
the set of events that are concurrent with $e$ is also finite. (Finiteness is not a first-order property.)
\end{enumerate}
The Russell-Wiener property makes sense if we think about events as entities that lie in time; they have (or can be represented by)
a nonempty temporal extension. Every event may be
thought of as being represented with a nonempty interval of instants (say of real numbers). Then $a<b$ means (intuitively) that the temporal extension of event $a$ lies completely to the left of $b$'s extension. So events
$a$ and $b$ are concurrent if and only if their temporal extensions have a common instant. Now the Russell-Wiener formula can be justified as follows.
Let $t$ be any instant that is both in the temporal interval of $b$ and of $c$. Since $a<b$, every instant of $a$ is before $t$, and similarly every instant of $d$ is after $t$, and hence every instant of $a$ is before every instant of $d$, that is $a<d$.
The finiteness property (item 2) can also be justified for the systems that we have in mind.
\begin{definition}
\label{D4.1}
Let $M$ be some Tarskian system execution interpretation of $L_R$. We say that $M$
models the regularity of the single-writer register $R$ owned by process $p$ iff it satisfies the following sentences.
\end{definition}
\begin{enumerate}
\item Process $p$ is serial and all write events are by $p$.
\begin{equation}
\begin{split}
\forall a,b: & \\
&p(a)\wedge p(b)\Rightarrow (a\leq b\vee b\leq a)\ \wedge\\
& \Writeit_R(a)\Rightarrow p(a).
\end{split}
\end{equation}
\item No event is both a $\Writeit_R$ and a $\Readit_R$ event.
\item For every $\Readit_R$ event $r$, $\Val(r)$ satisfies the following (not necessarily exclusive) disjunction.
\begin{enumerate}
\item For some $\Writeit_R$ event $w$ such that $\Val(r)=\Val(w)$, $w$ is concurrent with $r$, or
\item for some $\Writeit_R$ event $w$ such that $\Val(r)=\Val(w)$, $w<r$ and there is no $\Writeit_R$ event $w'$
with $w<w'<r$, or
\item there is no $\Writeit_R$ event $w$ such that $w<r$, and $\Val(r)$ is the initial value $d_R$ of the register.
\end{enumerate}
\end{enumerate}
When this definition of regularity is expressed as a sentence $\rho$ in $L_{R}$, we can meaningfully say
that a system execution $M$ that interprets $L_R$ is a model of register regularity iff $M\models \rho$.
It is
often the case that mathematical English is easier for us to read and understand, and hence most
statements are given here in mathematical English rather than in first-order formulas.
\section{Non-restricted system executions}
\label{S4}
We explain in this section the notion of non-restricted semantics of distributed systemד, a simple but essential
notion for our approach.
Suppose a distributed system composed of
serial processes that use diverse means of communication devices (such as shared memory registers, message passing channels etc.) where the operations of
each process are directed by some code that the process executes (serially). We ask: what can be said about the semantics of this code
when nothing can be said about its communication devices. On the positive side, we can say for example that instruction at line $k+1$
is executed after the one at line $k$, unless that one is transferring control to another line, we know how to execute
$\IF --- \THEN ---$ instructions etc. And on the negative side we cannot determine the value of a read instruction or the
effect of a write instruction. By {\em non-restricted} we mean the semantics of such a system
when there is
absolutely no restriction on the behavior of its communication devices.
So a read of a memory location
(a register) is not required to return the value of the last write event, and it may return any value in the type of that register--even
a value that has never been written. Likewise the contents of messages received is arbitrary, and a message received may have possibly never been sent. Surely not much can
be said about this non-restricted semantics when the different processes are not aware of the presence or absence of the other processes. However, as we shall see, by separating the semantics of the program from
the specification of the communication devices we obtains a greater flexibility and application range in our correctness proofs.
With reference to the code of Figure \ref{JP}, we ask about the semantics of the code for $p_i$ when only the minimal necessary assumptions are made on the read and write instruction executions.
These minimal assumptions suppose that
instructions have names and their actions have values. However, and here is the the idea of non-restriction, nothing
relates the value of a communication action to the value of another communication action.
For example, instruction $R_0:= n_0$ is called a ``write instruction on register $R_0$'' and the value of its execution is the
value of variable $n_0$,
and likewise, instruction $v_0:= R_1$ is called ``read of register $R_1$'' and its execution has a value that is assigned to variable $v_0$.
The only thing that can be said about the value of such a read event is that its value is in the type of $v_0$, i.e.
natural numbers in our case. The minimal assumptions, however,
do not relate this value to
the value of register $R_1$. In fact we even do not assume that there exists an object called $R_1$, and the expression
``the value of register
$R_1$'' is meaningless here. The semantics of the program under such minimal assumptions is the
non-restricted semantics of the algorithm, and a detailed definition is given next.
\subsection{An approximation to the non-restricted semantics}
\label{Sub5.1}
In order to give a better explanation of the non-restricted semantics of Kishon's Poker algorithm, we define here that
semantics with local states, local steps, and local history sequences (as opposed to
the global states and steps that were discussed in section \ref{S3}). This explanation is only an approximation, a presentation
of the idea, and a fuller presentation will be described in Section \ref{PNP} only after the benefit of non-restricted semantics
is made evident with the proof of Theorem \ref{T3.3}. So we begin this section with the notion of local variables and local states and steps.
{\em The local-state variables of process $p_i$ (for $i=0,1$) and their types are the following:} $n_i,v_0,\Val_i$ are of type
$\mathbb N$ (natural numbers),
and $PC_i$ is of type $\{1,\ldots,5\}$. So registers $R_0$ and $R_1$ are not among
the local state variables. A local state is a function that gives values to the local state variables in their types.
In the initial local state $S_0$ of $p_i$ we have that $S_0(PC_i)=1$ and $S_0(x)=\bot$ is the undefined value for any other local variable.
A non-restricted step by $p_i$ is a pair of $p_i$ local states, and as before we have $(1_0,2_0),\ldots, (4_0,5_0)$ local steps by $p_0$, and
$(1_1,2_1),\ldots,(4_1,5_1)$ local steps by $p_1$. The definition of the read and write local steps however is different from those
of Section \ref{S3}, and as an example we look at $p_0$ local read and write steps.
\begin{enumerate}
\item
A pair of local $p_0$ states $(S,T)$ is denoted ``write-on-register-$R_0$'' (also said to be a local $(2_0,3_0)$ step by $p_0$) when
\begin{enumerate}
\item $S(PC_0)=2$, $T(PC_0)=3$, and
\item all local variables other than $PC_0$ have the same value in $T$ as in $S$.
\end{enumerate}
We define the value of this step to be $S(n_0)$. This is something new in relation to Section \ref{S3}, that steps have values.
Registers do not exist and have no value, but steps do have values in the non-restricted framework. Registers appear in the {\em name} of the step (and this is important, they will serve as predicates)
but they do not record values.
\item
A pair of $p_0$ local states $(S,T)$ is denoted ``read-of-register-$R_1$'' (also said to be a local $(3_0,4_0)$ step by $p_0$)
when
\begin{enumerate}
\item $S(PC_0)=3$, $T(PC_0)=4$, $T(v_0 )\in \mathbb{N}$,
\item all local variables other than $PC_0$ and $v_0$ have the same value in $T$ as in $S$.
\end{enumerate}
We define the value of this step to be $T(v_0)$. There is no restriction on the value obtained in a read step (except that it has to be
in the right type).
\end{enumerate}
Finally, a non-restricted local
history of $p_i$ is a sequence of local $p_i$ states $S_0,\ldots$, beginning with the initial state $S_0$,
such that every pair $(S_m,S_{m+1})$ in the sequence is a local step by $p_i$.
A non-restricted history sequence is a good approximation to what we mean here by non-restricted semantics (of the Kishon's Poker algorithm
of $p_i$),
but the ``real'' definition is given in \ref{Def5.2} in the form of a class of
system executions whose language $L^i_{NR}$ (for $i=0,1$) is defined first.
\begin{definition} \label{DefLK}
The language $L^i_{NR}$ is a two-sorted language that contains the following features.
\end{definition}
\begin{enumerate}
\item There are two sorts: \Event\ and $\Data$. ($\Data$ has a fixed interpretation as the set ${\mathbb N}\cup\{-1\}$).
\item Unary predicates on \Event\ are: $p_i$, $\Assignmentto n_i$, $\Writeon R_i$, $\Readof R_{1-i}$, and
$\Return_i$.
\item
There are two binary
predicates denoted both
with the symbol $<$, one is the temporal precedence relation on the \Event\ sort, and the other is the ordering relation on the
natural numbers.
\item There is a function symbol $\Val:\Event\to \Data$.
\end{enumerate}
The language contains logical variables with which formulas and sentences can be formed. We reserve lower-case letters such as $a,b,c$
as variables over the \Event\ sort.
We are ready for the definition of non-restricted semantics of Kishon's Poker algorithm.
\begin{definition}
\label{Def5.2}
A system execution $M$ that interprets
$L^i_{NR}$ is said to be a non-restricted
execution of Kishon's Poker algorithm if it satisfies the properties enumerated in Figure \ref{P1}.
\end{definition}
\begin{figure
\fbox{
\begin{minipage}[t]{\columnwidth}
\begin{enumerate}
\item[0.] For every event $e$, $p_i(e)$.
There are exactly four events (of $p_i$) and they are linearly ordered by the temporal precedence relation $<$: $a_1<a_2<a_3<a_4$.
\item[1.]
The first $p_i$ event, $a_1$, is an \Assignmentto$n_i$ event. $\Val(a_1)$ is a non-zero natural number.
\item[2.]
The second $p_i$ event, $a_2$, is a \Writeon$R_i$ and $\Val(a_2)=\Val(a_1)$. No other
event is a $\Writeon R_i$ event.
\item[3.]
The third $p_i$ event, $a_3$, is a \Readof$R_{1-i}$ event. $\Val(a_3)\in {\mathbb{N}}$
is a natural number. No other event is predicated \Readof$R_{1-i}$.
\item[4.]
The fourth $p_i$ event, $a_4$, is a \Return\ event. As for
the value that $p_i$ returns we have the following:
\begin{enumerate}
\item If $\Val(a_3)=0$ or $\Val(a_3)=\Val(a_1)$ then $\Val(a_4)=0$,
\item If $0<\Val(a_3)<\Val(a_1)$, then $\Val(a_4)=1$,
\item If $\Val(a_3)>\Val(a_1)$, then $\Val(a_4)=-1$.
\end{enumerate}
\end{enumerate}
\end{minipage}
}
\caption{Properties of non-restricted executions of Kishon's protocol of $p_i$ expressed in the $L^i_{NR}$ language.
We say that $\Val(a_4)$ is the value returned by $p_i$.}
\label{P1}
\end{figure}
In figure \ref{P1} we collected the little that can be said about non-restricted executions by $p_i$.
The $p_i$ properties (for $i=0,1$) describe the ways $p_i$ plays the Kishon's game: first an arbitrary nonzero number is picked
and assigned to $n_i$. This action $a_1$ is predicated by $p_i$ and the $\Assignmentto n_i$ predicates. Then comes event $a_2$ predicated under $p_i$ and
\Writeon$R_i$ and whose value is
equal to the value of the first event. The third event $a_3$ is predicated by $p_i$ and \Readof$R_{1-i}$ and its value (a natural number) is
unspecified. Finally
the result of the game is calculated, this is $a_4$ the fourth event of $p_i$ whose value depends only on the values of the previous
events by $p_i$.
Although these properties of Figure \ref{P1} are written in English, they should formally be expressed in the first-order language
$L^i_{NR}$ defined above. Only with a formal language the satisfaction relation that a
given structure satisfies statements in its language is well-defined.
Note in particular item 3 in that list. It says for $i=0$ that
``The third $p_0$ event, $a_3$, is a \Readof$R_1$ event'' and that ``$\Val(a_3)\in {\mathbb{N}}$
is a natural number''. That is, under the minimality condition, $\Val(a_3)$ is unrelated to any write event,
and the only requirement is that it has the correct type.
Note also that none of the statements of Figure \ref{P1} relates events of $p_0$ to events of $p_1$.
No concurrency
is involved in the non-restricted specification. For that reason, it is expected that the proof that every non-restricted execution
of Kishon's Poker algorithm satisfies these statements of Figure \ref{P1} would be quite simple. Well, the proof is indeed simple,
but it requires some preliminary definitions which take time, and in order not to delay too much the correctness proof
we postpone these preliminary definitions to Section \ref{PNP} and proceed directly to that proof.
So, in reading the following section, the reader is asked to rely on intuition and to accept that even if the registers are
regular,
the minimal properties of Figure \ref{P1} hold in every execution by $p_i$.
With this assumption we shall prove in the following section the correctness of the Poker algorithm.
\section{Correctness of Kishon's Poker algorithm for regular registers}
\label{Sec6}
A {\em restricted} semantics is obtained as the conjunction of non-restricted semantics and a specification of the communication
devices. So ``restricted'' means imposing restrictions on the communication devices. In this section the restrictions on the registers
are that they are regular. We shall define in details the restricted semantics and prove Theorem \ref{T3.3} which says essentially that even
with regular registers the correctness condition holds.
Define the language $L_{NR}$ as the union of the languages $L^0_{NR}$ and $L^1_{NR}$ defined in the previous section and employed in
Figure \ref{P1}.
That is, the symbols of $L_{NR}$ are the symbols of $L^0_{NR}$ and of $L^1_{NR}$. So $L_{NR}$ is a two sorted language, with sorts
\Event\ and \Data. The predicates of $L_{NR}$ are $p_0$, $p_1$ etc. (and the symbols such as $<$ that appear in both languages).
If $M$ is a structure that interprets $L_{NR}$ then the {\em reduct} of $M$ to $p_0$ is the structure $M^0$ defined as follows.
The universe of $M^0$ consists of the set of events $p_0^M$ (that is all the events of $M$ that fall under predicate $p_0$)
and $\Data^{M_0}=\Data^M={\mathbb N}\cup \{-1\}$. All other predicates of $L^0_{NR}$ have the same
interpretation in $M_0$ as in $M$, and the function $\Val^{M_0}$ is the restriction of $\Val^M$ to the set of $p_0$ events of $M$. In a similar way the restriction of $M$ to $p_1$ is defined.
The following is the definition of the class of structures that model executions of the Kishon's Poker algorithm under the assumption
that the registers are regular.
\begin{definition}
\label{D5.6}
A restricted system-execution of Kishon's Poker algorithm for regular registers is a system execution interpretation of $L_{NR}$, $M$,
that satisfies the
following two conditions.
\begin{enumerate}
\item Every event in $M$ is either in $p_0$ or in $p_1$. That is, $M\models \forall e(p_0(e)\vee p_1(e))\wedge \neg\exists e(p_0(e)\wedge p_1(e))$.
\item For every $i=0,1$, $M^i$ satisfies the non-restricted properties of Figure \ref{P1}. ($M^i$ is the reduct of $M$ to $p_i$.)
\item The two (distinct) registers $R_0$ and $R_1$ are single writer regular registers of $p_0$ and $p_1$ (Definition \ref{D4.1}).
\end{enumerate}
\end{definition}
Recall (Section \ref{SReg}) that being a system-execution is already a certain restriction on the structure $M$. For example,
it is required that the
$<^M$ relation satisfies the Russell-Wiener property (equation \ref{RW}).
It follows from this definition that any restricted system-execution of Kishon's algorithm for regular registers
contains eight events: the four $p_0$ event $a_1,\ldots,a_4$ and the four $p_1$ events $b_1,\ldots,b_4$.
We now prove the correctness of the Kishon's Poker algorithm as an absract statement about a class of Tarskian system executions.
The
following theorem corresponds to theorem \ref{T1}.
\begin{theorem}
\label{T3.3}
Assume that $M$ is a system execution of the Kishon's Poker algorithm with regular registers as in Definition \ref{D5.6}. Then
the following hold in $M$.
Suppose that $a_1$ and $b_1$ are the Assignment events of $p_0$ and $p_1$ respectively, and $a_4$, $b_4$ are their Return events.
\begin{enumerate}
\item If $\Val(a_1)<\Val(b_1)$, then $\Val(a_4)<\Val(b_4)$.
\item If $\Val(a_1) > \Val(b_1)$, then $\Val(a_4) > \Val(b_4)$.
\item If $\Val(a_1) = \Val(b_1)$, then $\Val(a_4) = \Val(b_4)=0$.
\end{enumerate}
\end{theorem}
\begin{proof} We only prove the first item, since the second has a symmetric proof and the third a similar proof.
The following two lemmas are used for the proof of this theorem. These lemmas rely on the abstract
properties of Figure \ref{P1} and on
the assumed regularity of the registers.
\begin{lemma}
\label{ML}
It is not the case that $\Val(a_3)=\Val(b_3)=0$.
\end{lemma}
There are two cases in the proof of the lemma.
\begin{enumerate}
\item[Case 1:] $a_2< b_3$. Since $R_0$ is regular and $a_2$ is the only write on $R_0$, and as $b_3$ is a read
of that register, $\Val(a_2)=\Val(b_3)$.
But
$\Val(a_2)=\Val(a_1)>0$ (by items 1 and 2 of Figure \ref{P1}) and hence $\Val(b_3)>0$ in this case as required.
\item[Case 2:] not Case 1. Then $\neg(a_2<b_3)$. Together with $b_2 < b_3$ and $a_2 < a_3$,
the Russell-Wiener property implies that $b_2 < a_3$.
Hence $\Val(a_3)=\Val(b_2)$ by the symmetric argument using now the regularity of register $R_1$. Since $\Val(b_2)>0$, we get
that $\Val(a_3)>0$ as required.
\end{enumerate}
\begin{lemma}
\label{LM1}
Suppose that $\Val(a_1)<\Val(b_1)$.
\begin{enumerate}
\item If $\Val(b_3) \not=0$ then
\[ \Val(b_4)= 1.\]
\item
Symmetrically, if $\Val(a_3)\not=0$ then
\[ \Val(a_4)= -1.\]
\end{enumerate}
\end{lemma}
For the proof of this lemma suppose that $\Val(b_3)\not= 0$.
Event $b_3$ is a read of register $R_0$, and there is just one write event on that regular register, namely the write event $a_2$ whose value is the value of the \Assignmentto$n_0$ event $a_1$. The initial value of $R_0$ is $0$. Since
$R_0$ is a regular register, this implies immediately that $\Val(b_3)=0$ or else $\Val(b_3)=\Val(a_2)=\Val(a_1)$. (Indeed,
the value of the read $b_3$ is either the value of some write event, and only $a_2$ can be that event, or the initial
value of the register.)
Since $\Val(b_3) \not=0$ is assumed, then \[\Val(b_3)=\Val(a_1)\] follows. The lemma assumes that $\Val(a_1)<\Val(b_1)$,
that is $0<\Val(b_3)<\Val(b_1)$, and hence $\Val(b_4)= 1$ by property $4(b)$ of $p_1$.
The second item of the lemma follows symmetrically, and this ends the proofs of the two lemmas.
We now conclude the proof of the first item of theorem \ref{T3.3}. Assume that $\Val(a_1)<\Val(b_1)$. We shall prove that
$\Val(a_4)< \Val(b_4)$. There are two cases.
\begin{enumerate}
\item[Case 1] $\Val(a_3)=0$. Then $\Val(a_4)=0$ (by 4(a)). By lemma \ref{ML}, $\Val(b_3)\not= 0$, and hence by lemma \ref{LM1}
$\Val(b_4)=1$. So $\Val(a_4)<\Val(b_4)$ in this case.
\item[Case 2] $\Val(a_3)\not=0$. Then $\Val(a_4)=-1$ by lemma \ref{LM1}. Now $\Val(b_4)$ can be $0$, and in this case $\Val(a_4)<\Val(b_4)$ holds. But if not, if $\Val(b_4) \not = 0$,
then
$\Val(b_3)\not=0$ by $4_1(a)$. Hence $\Val(b_4)=1$ (again by lemma \ref{LM1}). Thus, in both cases $\Val(a_4)<\Val(b_4)$.
\end{enumerate} \end{proof}
\section{Proof of the non-restricted properties}
\label{PNP}
In order to complete the proof of Theorem \ref{T3.3} it remains to prove that the properties of Figure \ref{P1} hold for $p_i$
even when we make no assumptions whatsoever on the registers used. For such a simple protocol with its four instructions that
are executed consecutively, it is evident that every execution generates four actions as described in that figure. It is also evident
that no concurrency is involved in establishing these properties of a single process. However, if we seek a proof
that does not rely on our intuitive understanding of the protocol, we need some mathematical framework that relates the code of Figure
\ref{JP} with the properties of Figure \ref{P1}. That is, we need an explication of the non-restricted semantics with which it is possible to prove that every execution of the code of process $p_i$ results in a structure that satisfies the properties of Figure \ref{P1}.
The development of section \ref{Sub5.1} is not good enough simply because a local history sequence is not a Tarskian structure for which
it is meaningful to say that it satisfies (or not) these properties. In this section we describe a way to present the set of non-restricted executions by process $p_i$ as a set of Tarskian structures.
We first redefine the notion of {\em state} (i.e. a non-restricted state of $p_i$) not as a function that gives values to state variables, but rather as a finite Tarskian
structure. Moreover, that state records not only the attributes of the moment, but actually all the events that led up to that state. That is, a state (in this section) is an extended state structure that includes its own history as well\footnote{In fact, it suffices that the extended state
includes a {\em property} of its history.}. (For earlier work that
tries to elucidate the notion of state in a similar fashion we refer to \cite{WhatIs} and \cite{OnsystemEx}.)
Let \StateVar\ be some set of {\em state variables} and suppose that every $v\in \StateVar$ has a type $\Type(v)$. Then a state, in
the functional meaning of the word (as is usually defined), is a function $s$ defined over $\StateVar$ such that $s(v)\in \Type(v)$ for every state variable $v$.
Now suppose that $L$ is some first-order language such that every $v\in \StateVar$ is an individual constant in $L$, and its type is
a sort of $L$ that is
required to be interpreted as $\Type(v)$ in every interpretation of $L$. (Note that $v$ is not a variable of $L$, it is not quantifiable and cannot be a free variable in a formula. It is just a name
of a member of any interpreting structure of $L$.)
Besides these state variables and sorts, $L$ contains other symbols (predicates, function symbols, and constants).
Let $\calM$ be a class of system execution structures that interpret $L$ and are such that for every
$M\in \calM$ and any
$v\in \StateVar$, $v^M\in \Type(v)$. The {\em functional state} of $M$, $S=S(M)$, is defined by $S(v)=v^M$ for every
$v\in \StateVar$.
The structures in $\calM$ are said to be {\em extended states.}
The following example will clarify this definition of extended states.
we define first a logical language $L^0_{K}$ which will be used to define the non-restricted semantics of the Kishon's Poker protocol for
$p_0$. The $L^1_{K}$ language is defined analogously for $p_1$.
\begin{definition} \label{DefLKK}
The language $L^0_{K}$ is a two-sorted language that contains the following features.
\end{definition}
\begin{enumerate}
\item There are two sorts: \Event\ and $\Data$. ($\Data$ is in this context interpreted as the set ${\mathbb N}\cup\{-1\}$).
\item Individual constants are names of \Data\ values. The following local variables of $p_0$ are individual constants of $L^0_K$:
$n_0$, $v_0$, $\Val_0$ and
$PC_0$. We set $\StateVar= \{n_0,v_0,\Val_0, PC_0\}$. Also, $-1,0,1$ are individual constants with fixed interpretations.
\item Unary predicates on \Event\ are: $p_0$, $\Assignmentto n_0$, $\Writeon R_0$, $\Readof R_1$, and $\Return_0$.
\item
There are two binary
predicates denoted both
with the symbol $<$, one is the temporal precedence relation on the \Event\ sort, and the other is the ordering relation on the
natural numbers.
\item There is a function symbol $\Val:\Event\to \Data$.
\end{enumerate}
Note that $L^0_{K}$ is richer than the language $L^0_{NR}$ with which the properties of non-restricted executions of $p_0$ were formulated
(in Figure \ref{P1}). In fact, $L^0_K$ is obtained from $L^0_{NR}$ by the additions of the constants in $\StateVar$ and their
sorts. So any interpreting structure of $L^0_{K}$ is also an interpretation of $L_{NR}$.
Let $\calM$ be the set of all interpretations $M$ of $L_{K_0}$ such that
\begin{enumerate}
\item $\Event^M$ is finite, and $\Data^M={\mathbb N}\cup \{ -1\}$.
\item $n_0^M,v_0^M\in {\mathbb N}$ and $PC_0^M\in \{1,2,3,4,5\}$.
\end{enumerate}
Members of $\calM$ are said to be extended (non-restricted) states (of $p_0$), and if $M\in \calM$ then the functional
state $S=S(M)$ is defined by $S(x)=x^M$.
The initial extended state is the structure $M_0\in \calM$ such that $\Event^{M_0}=\emptyset$ (there are no events),
$PC_0^{M_0} = 1$, the predicates over the events have
empty interpretation (of course) and $x^{M_0}=0$ for any individual constant $x$ (other than $PC_0$).
If $M$ and $N$ are structures for $L^0_K$, then $N$ is said to be an {\em end-extension} of $M$ when $\Event^M$ is an initial
section of $\Event^N$ (in the $<^N$ ordering) and the reduct of $N$ to $\Event^M$ is the structure $M$.
The $p_0$ extended steps are defined as follows.
\begin{enumerate}
\item[$(1_0,2_0)$ steps] are pairs $(S,T)$ of extended $p_0$ states such that
\begin{enumerate}
\item $PC_0^S=1_0$, $PC_0^T=2_0$, $n_0^T$ is an arbitrary positive natural number.
\item For some $a_1\not\in \Event^S$ (we say that $a_1$ is a new member) $\Event^T=\Event^S\cup \{a_1\}$, and $S$ is the restriction
of $T$ to $\Event^S\cup {\mathbb N}\cup \{-1\}$. For every $x\in \Event^S$, $x<^T a_1$. That is, $a_1$ is added after all events
of $S$, and $T$ is an end-extension of $S$.
\item The following hold in $T$. $p_0(a_1)$. $\Assignmentto n_0(a_1)$. $\neg \Writeon R_0(a_1)$. $\neg \Readof R_1(a_1)$. $\neg \Return(a_1)$.
$\Val(a_1)=n_0$.
\end{enumerate}
\item[$(2_0,3_0)$ steps] are pairs $(S,T)$ of extended $p_0$ states such that
\begin{enumerate}
\item $PC_0^S =2_0$, $PC_0^T=3_0$, $n_0^T=n_0^S$.
\item For some new member $a_2$, $\Event^T=\Event^S\cup \{ a_2\}$. $T$ is an end-extension of $S$.
\item The following holds in $T$. $p_0(a_2)$. $\Writeon R_0(a_2)$. $\neg \Assignmentto n_0(a_2)$.
$\neg \Readof R_1(a_2)$. $\neg \Return(a_2)$. $\Val(a_2)= n_0$.
\end{enumerate}
\item[$(3_0,4_0)$ steps] are pairs $(S,T)$ of extended $p_0$ states such that
\begin{enumerate}
\item $PC_0^S =3_0$, $PC_0^T=4_0$, $n_0^T=n_0^S$, $v_0^T\in {\mathbb N}$ is arbitrary.
\item For some new member $a_3$, $\Event^T=\Event^S\cup \{ a_3\}$. $T$ is an end-extension of $S$.
\item The following holds in $T$. $p_0(a_3)$. $\Readof R_1(a_3)$. $\neg \Writeon R_0(a_3)$. $\neg \Assignmentto n_0(a_3)$.
$\neg \Return(a_3)$. $\Val(a_3)= v_0$.
\end{enumerate}
\item[$(4_0,5_0)$ steps] are pairs $(S,T)$ of extended $p_0$ states such that
\begin{enumerate}
\item $PC_0^S =4_0$, $PC_0^T=5_0$, $n_0^T=n_0^S$, $v_0^T=v_0^S$.
\item For some new member $a_4$, $\Event^T=\Event^S\cup \{ a_4\}$. $T$ is an end-extension of $S$.
\item The following holds in $T$. $p_0(a_4)$. $\Return(a_4)$. $\neg \Readof R_1(a_4)$. $\neg \Writeon R_0(a_4)$.
$\neg \Assignmentto n_0(a_4)$.
$\Val_0^T\in \{-1,0,1\}$ is such that $\Val^T(a_4)=\Val^T_0$ and is determined as follows.
If $(v^S_0=0\vee v^S_0=n^S_0)$ then
$\> \Val^T_0 :=0$ elseif $v^S_0<n^S_0$ then $\Val^T_0 :=1$ else $\Val_0 := -1$.
Hence the following hold in $T$. If $\Val(a_3)=0$ or $\Val(a_3) =\Val(a_4)$ then $\Val(a_4)=0$.
If $0< \Val(a_3)<\Val(a_1)$ then $\Val(a_4)= 1$. If $\Val(a_3)>\Val(a_1)$ then $\Val(a_4)=-1$.
\end{enumerate}
\end{enumerate}
The $p_1$ steps:
$(1_1,2_1),\ldots,(4_1,5_1)$ are similarly defined.
A $p_0$ invariant is a (first-order) sentence $\alpha$ in $L_{K_0}$ that holds in the initial extended state $M_0$ and is such that for every
extended step $(S,T)$ of $p_0$, if $S\models \alpha$ then $T\models \alpha$.
Note that if $M$ is any non-restricted extended state such that $PC_i^M = k_i\in\{1,\ldots,4\}$ then there is a state $N$ such that $(M,N)$ is a
$(k,(k+1))$ step.
\begin{definition}
A non-restricted extended history sequence of the Kishon's Poker algorithm for $p_i$ ($i=0,1$) is a sequence of Tarskian non-restricted extended states
$(M_0,\ldots)$ of $p_i$ such that $M_0$ is the initial state and every pair $(M_i,M_{i+1})$
in the sequence is a step by $p_i$.
\end{definition}
We note that in a maximal non-restricted extended history sequence there are five states.
\begin{definition}
\label{D5.4}
We say that $M$ is a non-restricted execution of the Kishon's protocol for $p_i$ if $M$ is the last state
in a maximal non-restricted extended history sequence for $p_i$.
\end{definition}
So if $M$ is a non-restricted execution of the Kishon protocol for $p_i$ then $M\models PC_i=5$.
\begin{theorem}
\label{ThmNR}
If $M$ is a non-restricted system execution of the Kishon's Poker algorithm, then $M$ satisfies the properties enumerated in
Figure \ref{P1}.
\end{theorem}
\begin{proof}
Let $M$ be a non-restricted system-execution for $p_0$. So there is
a maximal non-restricted extended history sequence for $p_0$, $M_1,\ldots,M_5$ such that $M=M_5$ is its last state.
We have to find an invariant $\alpha$ such that
\[ \alpha\wedge PC_0=5 \Rightarrow \text{properties 0--4 of Figure \ref{P1}}.\]
Let $\alpha$ be the conjunction of the properties listed in Figure \ref{P4}.
\begin{figure}[h!]
\fbox{
\begin{minipage}[t]{\columnwidth}
All events fall under predicate $p_0$.
\begin{enumerate}
\item[1.]
If $PC=1$ then sort \Event\ is empty.
\item[2.] If $PC=2$ then there is just one event, $a_1$, which is such that
\Assignmentto$n_0(a_1)$
and $\Val(a_1)>0$ is a natural number.
\item[3.] If $PC=3$, then there are exactly two events. The first is as in item 2 and the second event, $a_2$, is
such that $\Writeon R_0(a_2)$ and $\Val(a_2)=\Val(a_1)$.
\item[4.] If $PC=4$, then there are exactly three events. The first and second are as in items 2 and 3, and the third event, $a_3$,
is such that $\Readof R_1(a_3)$ and $\Val(a_3)\in {\mathbb{N}}$ is a natural number.
\item[5.] If $PC=5$ then there are exactly four events. The first three are as in items 1,2,3, the fourth event, $a_4$, is
such that $\Return_0(a_4)$ and $\Val(a_4)$ satisfies the following:
\begin{enumerate}
\item If $\Val(a_3)=0$ or $\Val(a_3)= \Val(a_1)$ then $\Val(a_4)=0$,
\item If $0<\Val(a_3)<\Val(a_1)$, then $\Val(a_4)=1$,
\item If $\Val(a_3)>\Val(a_1)$, then $\Val(a_4)=-1$.
\end{enumerate}
\end{enumerate}
\end{minipage}
}
\caption{Inductive properties expressed in the $L^0_{K}$ language which serve in the proof of Theorem \ref{ThmNR}.}
\label{P4}
\end{figure}
It is not difficult to prove that
$\alpha$ is an invariant. This shows that
$M = M_5\models \alpha$ and since $M_5\models PC_0=5$ it follows immediately that $M_5$ satisfies the properties enumerated in
Figure \ref{P1}, and hence Theorem \ref{T3.3} applies. \end{proof}
\section{Discussion}
\label{S5}
The Kishon's Poker algorithm is a trivial short algorithm for two processes that execute concurrently. It serves here as a platform to
introduce an event-based {\em model theoretic} approach to the problem of proving the correctness of distributed systems
and to
compare this approach to the {\em standard}
approach that is based on the notions of global state, history, and invariance.
Invariance is certainly an extremely important key concept, but there are situations
in which it is natural to argue about the events and their temporal interrelations, whereas finding the invariance and its
proof is quite difficult.
This observation is well-known, as is the well-documented observation that proofs that rely on these events and temporal interrelations
often lead to grave errors. For example Lamport writes in \cite{Ass}:
\begin{quote}
Most computer scientists find it natural to reason about a concurrent program in terms of its behavior--the sequence of events generated by its execution. Experience has taught us that such reasoning is not reliable; we have seen too many convincing proofs of incorrect algorithms. This has led to assertional proof methods, in which one reasons about the program's state instead of its behavior.
\end{quote}
The approach outlined in this paper to the problem of proving properties of distributed systems is guided by the desire
to preserve the naturalness of the behavioral approach and to combine it with reliable mathematical rigor. This approach is characterized by the following features.
\begin{enumerate}
\item
The interleaving semantics and its global states and histories are not used. Instead of global states and interleaving of actions by the different processes, only local states and histories are used in
order to express what each process does in a way that is detached from the properties of the communication devices.
Local states of an
individual process are needed in order to define how the code is executed by a process without any commitment to any specification of
the shared communication devices.
The resulting structures are called ``non-restricted Tarskian system executions'' and the manifold of all of these
structures is called the ``unrestricted system''. These structures are unrestricted in the sense
that the values of inter-process communication objects of one process are not connected to values of other processes, simply because the other processes and
the common communication objects are not represented in the local states of a process.
The properties of these unrestricted
system executions are properties that refer to each of the processes separately, as if that process lives in a world
by itself. Figure \ref{P1} is an example of such properties. Note that
no concurrency is involved so far, and only local analysis is involved.
\item The correct specification of the communication devices is formulated (again in some first-order language)
in a way that is not connected to any specific system of programs. For example, regularity is a property of registers
that is not connected to the Kishon's Poker algorithm or any other algorithm that uses them.
\item If from the
class of non-restricted system executions we take only those
system executions in which the communication devices work properly (i.e. satisfy the communication device specifications)
then we get the system $\cal M$ that represents the manifold of all possible executions of the algorithm under the required assumptions (such as regularity)
on the
communication devices. See for example Definition \ref{D5.6}.
\item The correctness of the algorithm is expressed wih a certain condition $\tau$, and to prove it one has to prove that any system
execution $M$ in $\cal M$ satisfies $\tau$.
\item The proof of correctness is thus composed of two main stages: establishing properties of the non-restricted
system executions (obtained by analyzing serial processes with their local states and histories), and then using
these properties in conjunction with assumed properties of the communication devices in order to prove the correctness condition $\tau$.
This separation of concerns is a main feature of the event-based approach outlined here, and the following slogan expresses this.
\begin{quote} \em The specification of a distributed system depends on
properties of sequential programs that work in isolation and on generally formulated properties of the communication devices that the processes employ.
\end{quote}
\end{enumerate}
There is an obvious price to pay for such a model-theoretic framework--it requires a certain (minimal) familiarity with basic notions in
logic and model theory. Some may say that this price is a barrier that cannot be accepted for a framework that claims to be
intuitive and natural. But there is plenty of historical evidence to show that what was once considered as difficult
becomes with time standard material when
better ways of explaining and presenting complex issues are developed.
It is interesting to note that that the event-based correctness proof framework, exemplified here with the
Kishon's Poker algorithm, has its origin, to some extent, in Lamport's earlier work \cite{SE} when he presented system-executions
(and the notion of
higher-level events). This is my reason to continue to use the term system-execution, to acknowledge the
connection with Lamport's earlier articles. Two important features however were missing from that earlier work which limited its degree of mathematical formality
and restricted its range of applications. Firstly, those system-executions were not full-fledge Tarskian structures
in the model-theoretic sense (they were not interpretations of a first-order language), and
secondly the notion of local state was missing from these proofs, and thus Lamport's system-executions were not mathematically
connected to the programs that their processes employ: there remained a gap between the software and the system-executions that represent executions of that software.
A bridge was missing between these two aspects (code and execution), and perhaps it is
because of this
gap that the earlier system-executions were forsaken, and Lamport himself concentrated in his later work on the invariant approach.
The bridge that I describe here using the Kishon's Poker algorithm as an example is quite simple. It relies on the notion of local states and on non-restricted semantics of single-process protocols. The advantage of non-restricted semantics is the separation of
the two issues: the execution of the code by each of the individual processes, and the specification
of the communication devices. These ideas were first presented in \cite{A} and developed in \cite{abraham} and in some later publications,
and surely much work is still needed in order to transform them into a well-developed useful framework for proving properties of
distributed systems.
|
\section{Introduction}
\noindent Item non-response may affect the quality of the estimates when the respondents and the non-respondents exhibit different characteristics with respect to the variables of interest. Item non-response in surveys is usually handled by single imputation, whose main objective is to reduce the non-response bias. Two approaches are commonly used in sample surveys to motivate imputation. Under the non-response model approach (NM), the response mechanism is explicitly modeled, whereas under the imputation model approach (IM), the variable under study is explicitly modeled. \\
\noindent Single imputation consists of replacing a missing value with an artificial one. It leads to a single imputed data set, constructed so that it is possible to apply complete data estimation procedures for obtaining point estimates. The response indicators are therefore not required. On the other hand, multiple imputation methods \citep{rub:87,lit:rub:87} consist in building $M>1$ imputed datasets, and in estimating the parameters under study for each of them. The $M$ analyses are then combined for inference. Multiple imputation has been extensively studied in the literature, some recent references include \citet{iac:por:07}, \citet{whi:car:10} and \citet{tem:kow:fil:11}. However, multiple imputation is not commonly used in sample surveys. Under the NM approach, multiple imputation needs to be proper for valid inference. Some sufficient conditions are given in \cite{rub:87}, pp. 118-119, but they are usually difficult to check for complex sampling designs, see \citet{bin:sun:96}, \citet{fay:92}, \citet{fay:96} and \citet{nie:03}. Also, under the IM approach, the multiple imputation variance estimator does not track the variance correctly, and can be considerably biased, see \citet{kot:95}, \citet{kim:bri:ful:kal:06}, \citet{bri:kal:kim:06} and \citet{bea:haz:boc:11}. Therefore, we focus in this paper on single imputation methods. \\
\noindent The Imputation Model (IM) approach is of common use to treat item non-response in surveys. The imputation methods are then motivated by a modeling of the relationship between the variable of interest and the available auxiliary variables. Both the imputation model and the imputation methods need to be adapted to the study variable. For instance, in business surveys, the interest variables often contain a large number of zeros. In the Capital Expenditure Survey conducted at Statistics Canada, approximately 70\% of businesses reported a value of zero to Capital Machinery and 50\% reported a value of zero to Capital Construction \citep{haz:cha:nam:14}. In case of some interest variable containing a large amount of zeroes, \citet{haz:cha:nam:14} propose imputation methods based on a mixture regression model. They prove that these methods lead to doubly robust estimators of the population mean, i.e. the imputed estimator of the mean is consistent whether the interest variable or the non-response mechanism is adequately modeled. However, these methods are not appropriate when estimating more complex parameters such as the population distribution function. \\
\noindent In this work, we propose an imputation which enables to preserve the distribution function for zero inflated data. This is an important practical property if the data users are not only interested in estimating means or totals, but also parameters related to the distribution of the imputed variable, e.g. the Gini coefficient. We use the IM approach, without explicit assumptions on the non-response mechanism for the interest variable. We propose a random imputation method which leads to a $\sqrt{n}$-consistent estimator of the total, and to a mean-square consistent estimator of the distribution function. \\
\noindent As recalled in \citet{haz:cha:nam:14}, random imputation methods suffer from an additional variability due to the imputation variance. Three approaches have been proposed in survey sampling to reduce this variance. Fractional imputation is somewhat similar to multiple imputation, and consists in replacing some missing value with $M$ imputed values to which some weights are given \citep{kal:kis:81,kal:kis:84,fay:96,kim:ful:04,ful:kim:05}. The imputation variance decreases as $M$ increases. The second approach consists in using some standard imputation mechanism, and in modifying the imputed values in order to suppress the imputation variance \citep{che:rao:sit:00}. Finally, the third approach consists of directly imputing artificial values in such a way that the imputation variance is eliminated \citep{kal:kis:81,kal:kis:84,dev:06,cha:dev:haz:11,cha:haz:12,has:til:14,cha:cha:haz:sal:sol:18}. This last approach is of particular interest because it leads to a single imputed dataset, which is attractive from a data user's perspective, and it does not require any modification of the imputed values. \\
\noindent In this paper, we propose a balanced version of our imputation method, which enables to greatly reduce the imputation variance. It consists of randomly generating the imputed values while satisfying appropriate balancing constraints, by using an adaptation of the Cube algorithm \citep{dev:til:04,cha:dev:haz:11}. Our simulation results prove that the balanced imputation method succeeds in preserving the distribution function of the imputed variable, with large variance reductions as compared to the proposed non-balanced imputation method. In order to produce confidence intervals for the estimated parameters with appropriate coverage, we also propose variance estimators adapted from the linearization variance estimators proposed by \cite{kim:rao:09}. Our simulation results indicate that these estimators perform well, both in terms of relative bias and of coverage rate. \\
\noindent The paper is organized as follows. In Section \ref{sec2}, we describe the theoretical set-up and the notation used in the paper. In Section \ref{sec3}, we briefly recall the two imputation procedures proposed by \citet{haz:cha:nam:14}, and introduce our two proposed imputation methods. In Section \ref{sec4}, we prove that the proposed random imputation procedure yields a consistent estimator of the total and of the distribution function. Variance estimation for the imputed estimator of the total is discussed in Section \ref{sec5}. The results of a simulation study comparing the four procedures and evaluating the proposed variance estimator are presented in Section \ref{sec6}. An application of the proposed methodology on data modelled in the Monthly Retail Trade Survey is presented in Section \ref{sec:appli}. We conclude in Section \ref{sec7}. All the proofs are given in the Appendix. Some additional simulation results are available in the Supplementary Material.
\section{Theoretical set-up} \label{sec2}
\noindent We are interested in some finite population $U$ of size $N$, with some quantitative variable of interest $y$ taking the value $y_i$ for unit $i \in U$. We note $y_U=(y_{1}, \ldots, y_{N})^\top$ for the vector of values for the variable $y$. We are interested in estimating the total $t_y=\sum_{i \in U} y_i$, and the finite population distribution function
\begin{eqnarray} \label{sec2:eq1}
F_N(t) & = & \frac{1}{N} \sum_{i \in U} 1(y_i \leq t)
\end{eqnarray}
where $1(\cdot)$ is the indicator function.\\
\noindent A sample $s$ of size $n$ is selected according to a sampling design $p(.)$, with $\pi_i$ the first-order inclusion probability in the sample for unit $i$. We suppose that $\pi_i>0$ for any unit $i \in U$, and we note $d_i=\pi_i^{-1}$ the design weight. We note $\delta_U=(\delta_1, \ldots, \delta_N)^\top$ for the vector of sample membership indicators. In case of full response, a complete data estimator of $t_y$ is the expansion estimator or Horvitz-Thompson~(1952) \nocite{hor:tho:52} estimator
\begin{eqnarray} \label{sec2:eq2}
\hat{t}_{y\pi} & = & \sum_{i \in s} d_i y_i.
\end{eqnarray}
This estimator is design-unbiased for $t_y$, in the sense that $E_p(\hat{t}_{y\pi})=t_y$ with $E_p$ the expectation under the sampling design $p(.)$, conditionally on $y_U$. We note $V_p$ the variance under the sampling design $p(.)$. Concerning the population distribution function $F_N$, plugging into (\ref{sec2:eq1}) the expansion estimators of the involved totals yields the plug-in estimator
\begin{eqnarray} \label{sec2:eq3}
\hat{F}_N(t) = \frac{1}{\hat{N}_{\pi}} \sum_{i \in s} d_i 1(y_i \leq t) & \textrm{ with } & \hat{N}_{\pi}=\sum_{i \in s} d_i.
\end{eqnarray}
Under some mild assumptions on the variable of interest and the sampling design \citep[see][]{dev:99,car:cha:gog:lab:10}, $\hat{F}_N(t)$ is approximately unbiased and mean-square consistent for $F_N(t)$. \\
\noindent We now turn to the case when the variable of interest $y$ is subject to missingness. Let $r_{i}$ be the response indicator, such that $r_{i}=1$ if unit $i$ responded to item $y$, and $r_{i}=0$ otherwise. Let $p_i$ be the response probability of some unit $i$. We note $r_U=(r_{1}, \ldots, r_{N})^\top$ for the vector of response indicators. We assume that each unit responds independently of one another. Let $E_q$ and $V_q$ denote the expectation and variance under the non-response mechanism, conditionally on the vector $y_U$ of population values and on the vector ${\delta}_U$ of sample membership indicators. An imputation mechanism is used to replace some missing value $y_i$ by an artificial value $y_{i}^*$. An imputed estimator for $t_y$ based on observed and imputed values is
\begin{eqnarray} \label{sec2:eq4}
\hat{t}_{yI} & = & \sum_{i \in s} d_i r_i y_i + \sum_{i \in s} d_i (1-r_i) y_i^*.
\end{eqnarray}
Similarly, an imputed estimator of the distribution function based on observed and imputed values is
\begin{eqnarray} \label{sec2:eq5}
\hat{F}_I(t) & = & \frac{1}{\hat{N}_{\pi}} \left\{\sum_{i \in s} d_i r_i 1(y_i \leq t) + \sum_{i \in s} d_i (1-r_i) 1(y_i^* \leq t)\right\}.
\end{eqnarray}
In comparison with the estimators obtained in (\ref{sec2:eq2}) and (\ref{sec2:eq3}) with complete data, there are two additional random mechanisms involved in the estimators given in (\ref{sec2:eq4}) and (\ref{sec2:eq5}). First, the non-response mechanism leads to observe the values of $y$ for a part of $s$ only. Then, the imputation mechanism is used to replace missing $y_i$'s with artificial values. \\
\noindent The imputation mechanism is motivated by an imputation model, which is a set of assumptions on the variable $y$ subject to missingness. In this paper, we are interested in some quantitative variable of interest for which the imputation model may be described as a mixture regression model, see equation (\ref{imput:model}) below. The proposed imputation methods are therefore not suitable for categorical variables. An extension of the proposed imputation methods to cover such variables is discussed in our conclusion. \\
\noindent In the context of a zero-inflated variable of interest, the mixture regression model introduced in \citet{haz:cha:nam:14} is as follows:
\begin{eqnarray} \label{imput:model}
y_i & = & \eta_i \left\{z_i^{\top} \beta + \sqrt{v_i} \epsilon_i \right\},
\end{eqnarray}
where the $\eta_i$'s are independent Bernoulli random variables equal to $1$ with probability $\phi_i$, and equal to $0$ otherwise; the $\epsilon_i$'s are independent and identically distributed random variables of mean $0$, variance $\sigma^2$, and with a common distribution function $F_{\epsilon}$; the parameters $\beta$ and $\sigma$ are unknown, and $v_i$ is a known constant. The vector of auxiliary variables $z_i$ is assumed to be known on the whole sample including non-respondents. To sum up, according to the imputation model (\ref{imput:model}) the variable $y_i$ follows a regression model with a probability $\phi_i$, and is equal to $0$ otherwise. \\
\noindent Note that no assumptions are made on some specific distribution for the residuals $\epsilon_i$. We only suppose that they share a common distribution function $F_{\epsilon}$. Let $E_m$ et $V_m$ denote respectively the expectation and variance under the imputation model. We suppose that the sampling design is non-informative, in the sample that the vector $\delta_U$ of sample membership indicators is independent of $\epsilon_U = (\epsilon_1,\ldots,\epsilon_N)^{\top}$ and $\eta_U=(\eta_1,\ldots,\eta_N)^{\top}$, conditionally on a set of design variables. \\
\noindent In practice, the $\phi_i$'s are unknown and need to be estimated. We assume that they may be parametrically modeled as
\begin{eqnarray} \label{model_phi_k}
\phi_i & = & f(u_i,\gamma)
\end{eqnarray}
where $f$ is a known function, $u_i$ is a vector of variables recorded for all sampled units, and $\gamma$ is an unknown parameter. An estimator of $\phi_i$ is
\begin{eqnarray} \label{estim_phi_i}
\hat{\phi}_i & = & f(u_i,\hat{\gamma}_r)
\end{eqnarray}
with $\hat{\gamma}_r$ an estimator of $\gamma$ computed on the responding units. We assume that $\eta_i$ and $\epsilon_i$ are independent, conditionally on the vectors $z_i$ and $u_i$. \\
\noindent In this paper, we use the Imputation Model (IM) approach where the inference is made with respect to the imputation model, the sampling design, the response mechanism and the imputation mechanism. This does not require an explicit modeling of the non-response mechanism unlike the Non-response Model approach \citep{haz:09}, but we assume that the data are missing at random, which means that model (\ref{imput:model}) holds for both the respondents and the non-respondents. We note $E_I$ and $V_I$ the expectation and variance under the imputation mechanism, conditionally on the vectors $y_U$, ${\delta}_U$ and $r_U$.
\section{Imputation methods} \label{sec3}
\noindent In this Section, we first briefly recall in Sections \ref{ssec31} and \ref{ssec32} the random imputation methods proposed by \citet{haz:cha:nam:14} for zero-inflated data. We then introduce the new methods that we propose in Sections \ref{ssec33} and \ref{ssec34}.
\subsection{Haziza-Nambeu-Chauvet random imputation} \label{ssec31}
\noindent A first proposal of \citet{haz:cha:nam:14} is to use the imputation mechanism
\begin{eqnarray} \label{imput:mecha:1}
y_i^* & = & \eta_i^* \left\{z_i^{\top} \hat{B}_r \right\},
\end{eqnarray}
where the unknown regression parameter $\beta$ is estimated by
\begin{eqnarray} \label{ssec31:eq2}
\hat{B}_r = \hat{G}_r^{-1} \left( \frac{1}{N}\sum_{i \in s} \omega_i r_i v_i^{-1} z_i y_i \right) & \textrm{with} & \hat{G}_r = \frac{1}{N}\sum_{i \in s} \omega_i r_i \hat{\phi}_i v_i^{-1} z_i z_i^{\top},
\end{eqnarray}
where $\omega_i$ denotes a so called imputation weight, and $\hat{\phi}_i$ is given in (\ref{estim_phi_i}). The $\eta_i^*$'s are independently generated, and $\eta_i^*$ is equal to $1$ with the probability $\hat{\phi}_i$, and is equal to $0$ otherwise. \\
\noindent There are several possible choices for the imputation weights $\omega_i$. Using a modeling of the response mechanism for the variable $y_i$, \citet{haz:cha:nam:14} propose to choose the imputation weights so that $\hat{t}_{yI}$ is a doubly robust estimator for $t_y$. This means that the imputed estimator is approximately unbiased for $t_y$ whether the imputation model or the non-response model is adequately specified. \citet{haz:cha:nam:14} also prove that the resulting imputed estimator is consistent for $t_y$ under either approach. \\
\noindent The random imputation mechanism in (\ref{imput:mecha:1}) has three drawbacks. Firstly, it leads to an additional imputation variance due to the $\eta_i^*$'s. To overcome this problem, \citet{haz:cha:nam:14} proposed a balanced version of their imputation mechanism that is presented in Section \ref{ssec32}. Secondly, the imputation mechanism in (\ref{imput:mecha:1}) does not lead to an approximately unbiased estimator of the distribution function, as will be illustrated in the simulation study conducted in Section \ref{sec5}. Finally, the consistency of the imputed estimator $\hat{t}_{yI}$ relies on an assumption of mean square consistency for $\hat{B}_r$, which may be difficult to prove since the matrix $\hat{G}_r$ can be close to similarity for some samples. Following \citet{car:gog:lar:13} and \citet{cha:dop:18}, we introduce in Sections \ref{ssec33} and \ref{ssec34} a regularized version of $\hat{B}_r$.
\subsection{Haziza-Nambeu-Chauvet balanced imputation} \label{ssec32}
\noindent The balanced random imputation procedure of \citet{haz:cha:nam:14} consists in replacing a missing value with
\begin{eqnarray} \label{imput:mecha:2}
y_i^* & = & \tilde{\eta}_i^* \left\{z_i^{\top} \hat{B}_r \right\},
\end{eqnarray}
where the $\tilde{\eta}_i^*$'s are not independently generated, but so that the imputation variance of $\hat{t}_{yI}$ is approximately equal to zero. Indeed, the imputation variance of $\hat{t}_{yI}$ is eliminated if the $\tilde{\eta}_i^*$'s are generated so that
\begin{eqnarray} \label{ssec32:eq2}
\sum_{i \in s} d_i(1-r_i)(\tilde{\eta}_i^*-\hat{\phi}_i)(z_i^{\top} \hat{B}_r) & = & 0.
\end{eqnarray}
\citet{haz:cha:nam:14} propose a procedure adapted from the Cube method \citep{dev:til:04,cha:til:06} which enables to generate the $\tilde{\eta}_i^*$'s so that (\ref{ssec32:eq2}) is satisfied, at least approximately. As a result, the imputation variance is eliminated or at least significantly reduced. \\
\noindent This imputation procedure is called balanced random $\phi$-regression ($BRR_\phi$) imputation by \citet{haz:cha:nam:14}. They prove that under the $BRR_\phi$ imputation, an appropriate choice for the imputation weights $\omega_i$ leads to a doubly robust estimator for $t_y$. Also, their empirical results indicate that it performs well in reducing the imputation variance. A drawback of the $BRR_\phi$ imputation mechanism is that it does not preserve the distribution function of the imputed variable, because it does not take into account the error terms $\epsilon_i$ in the imputation model (\ref{imput:model}). This is empirically illustrated in section \ref{sec6}. To overcome this problem, two new imputation procedures are proposed in Sections \ref{ssec33} and \ref{ssec34}.
\subsection{Proposed random imputation} \label{ssec33}
\noindent The random imputation procedure that we propose consists in mimicking as closely as possible the imputation model (\ref{imput:model}), by replacing some missing $y_i$ with the imputed value
\begin{eqnarray} \label{imput:mecha:3}
y_i^* & = & \eta_i^* \left\{z_i^{\top} \hat{B}_{ar} + \sqrt{v_i} \epsilon_i^* \right\},
\end{eqnarray}
where $\hat{B}_{ar}$ is a regularized version of $\hat{B}_r$, and $\eta_i^*$ is a Bernoulli random variable as defined in (\ref{imput:mecha:1}). The $\epsilon_i^*$'s are selected independently and with replacement in the set of observed residuals
\begin{eqnarray} \label{set:obs:resid}
E_r = \left\{e_j~;~r_j=1 \textrm{ and } \eta_j=1 \right\} & \textrm{ where } & e_j = \frac{y_j-z_j^{\top} \hat{B}_{ar}}{\sqrt{v_j}},
\end{eqnarray}
with $Pr(\epsilon_i^*=e_{j})=\tilde{\omega}_j$ for any $j \in s$ such that $r_j=1$ and $\eta_j=1$, where
\begin{eqnarray} \label{pr:epsi:st}
\tilde{\omega}_j & = & \frac{\omega_{j}}{\sum_{k \in s} \omega_{j} r_k \eta_k}.
\end{eqnarray}
We note
\begin{eqnarray} \label{ssec33:eq0}
\bar{e}_r = \sum_{j \in s} \tilde{\omega}_j r_j \eta_j e_j & \textrm{and} & \sigma_{er}^2 = \sum_{j \in s} \tilde{\omega}_j r_j \eta_j (e_j-\bar{e}_r)^2.
\end{eqnarray}
\noindent The regularized version of $\hat{B}_{r}$ is obtained by following the approach in \citet{car:gog:lar:13} and \citet{cha:dop:18}. We first write
\begin{eqnarray} \label{ssec33:eq1a}
\hat{G}_r & = & \sum_{j=1}^p \alpha_{jr} v_{jr} v_{jr}^{\top},
\end{eqnarray}
with $\alpha_{jr} \geq \ldots \geq \alpha_{pr}$ the non-negative eigenvalues of $\hat{G}_r$, and where $v_{1r},\ldots,v_{pr}$ are the associated orthonormal vectors. For some given $a>0$, the regularized versions of $\hat{G}_r$ and $\hat{B}_r$ are
\begin{eqnarray} \label{ssec33:eq1b}
\hat{G}_{ar} = \sum_{j=1}^p \max(\alpha_{jr},a) v_{jr} v_{jr}^{\top} & \textrm{and} & \hat{B}_{ar} = \hat{G}_{ar}^{-1} \left( \frac{1}{N}\sum_{i \in s} \omega_i r_i v_i^{-1} z_i y_i \right).
\end{eqnarray}
The regularization leads to a matrix $\hat{G}_{ar}$ which is always invertible, and such that $\|\hat{G}_{ar}^{-1}\| \leq a^{-1}$ with $\|\cdot\|$ the spectral norm. \\
\noindent We prove in Section \ref{sec4} that $\hat{B}_{ar}$ is a mean-square consistent estimator of $\beta$, and that under the proposed imputation procedure the imputed estimator of the total is mean-square consistent for $t_y$. Also, we prove that the imputed estimator $\hat{F}_{I}(t)$ is $L_1$-consistent for the population distribution function. However, this imputation procedure leads to an additional variability for $\hat{t}_{yI}$ due to the imputation variance. Therefore, a balanced version is proposed in Section \ref{ssec34}.
\subsection{Proposed balanced imputation} \label{ssec34}
\noindent The balanced procedure consists in replacing a missing value with
\begin{eqnarray} \label{imput:mecha:4}
y_i^* & = & \tilde{\eta}_i^* \left\{z_i^{\top} \hat{B}_{ar} + \sqrt{v_i} \tilde{\epsilon}_i^* \right\},
\end{eqnarray}
where the $\tilde{\eta}_i^*$'s and the $\tilde{\epsilon}_i^*$'s are not independently generated, but so as to eliminate the imputation variance of $\hat{t}_{yI}$. A sufficient condition for this consists in generating the residuals $\tilde{\eta}_i^*$ and $\tilde{\epsilon}_i^*$ so that
\begin{eqnarray}
\sum_{i \in s} d_i(1-r_i)(\tilde{\eta}_i^*-\hat{\phi}_i)(z_i^{\top} \hat{B}_r^*) & = & 0, \label{ssec34:eq1} \\
\sum_{i \in s} d_i(1-r_i)\tilde{\eta}_i^* \sqrt{v_i} \tilde{\epsilon}_i^* & = & 0. \label{ssec34:eq2}
\end{eqnarray}
\noindent This is done in a two-step procedure: first, the $\tilde{\eta}_i^*$'s are generated by means of Algorithm 1 in \citet{haz:cha:nam:14}, so that (\ref{ssec34:eq1}) is approximately respected; then, the $\tilde{\epsilon}_i^*$'s are generated by using Algorithm 1 described in \citet{cha:dev:haz:11}, so that (\ref{ssec34:eq2}) is approximately respected. \\
\noindent Since the balancing equations (\ref{ssec34:eq1}) and (\ref{ssec34:eq2}) are usually only approximately respected, the imputation variance is not completely eliminated, but it may be significantly reduced: see the simulation study in Section \ref{sec6}. Though the balanced imputation procedure is expected to provide estimators with smaller variance, the asymptotic properties of these estimators are difficult to study due to intricate dependencies introduced in the imputation process. Extending the results in Section \ref{sec4} is a challenging problem for further theoretical research.
\section{Properties of the proposed methods} \label{sec4}
\noindent To study the asymptotic properties of the sampling designs and estimators, we use the asymptotic framework of \citet{isa:ful:82}. We suppose that the population $U$ belongs to a nested sequence $\{U_\tau\}$ of finite populations with increasing sizes $N_\tau$, and that the vector of values for the variable of interest $y_{U\tau}=(y_{1\tau},\ldots,y_{N\tau})^{\top}$ belongs to a nested sequence $\{y_{U\tau}\}$ with increasing sizes $N_\tau$. For simplicity, the index $\tau$ is omitted in what follows and all limits are computed when $\tau \to \infty$. \\
\noindent We consider the following regularity assumptions:
\begin{itemize}
\item[H1:] Some constants $C_1,C_2>0$ exist, s.t. $C_1 \leq N n^{-1} \pi_i \leq C_2$ for any $i \in U$.
\item[H2:] Some constant $C_3$ exists, s.t. $\sup_{i \neq j \in U} \left(n \left|1-\frac{\pi_{ij}}{\pi_i\pi_j}\right|\right) \leq C_3$.
\item[H3:] Some constants $C_4,C'_4>0$ exist, s.t. $C_4 \leq \min_{i \in U} p_i$ and $C'_4 \leq \min_{i \in U} \phi_i$.
\item[H4:] Some constants $C_5,C_6>0$ exist, s.t. $C_5 \leq N^{-1} n \omega_i \leq C_6$ for any $i \in U$.
\item[H5:] Some constants $C_7,C_8,C_9>0$ exist, s.t. $C_7 \leq v_i \leq C_8$ and $\|z_i\| \leq C_9$ for any $i \in U$. Also, the matrix
\begin{eqnarray} \label{sec4:eq1}
G & = & \frac{1}{N} \sum_{i \in U} \omega_i \pi_i p_i \phi_i v_i^{-1} z_i z_i^{\top}
\end{eqnarray}
is invertible, and the constant $a$ chosen is s.t. $\|G^{-1}\| \leq a^{-1}$.
\item[H6:] We have $E\left(\|\hat{\gamma}_r - \gamma\|^2\right)=O(n^{-1})$.
\item[H7:] Some constant $C_{11}$ exists, s.t. for any vector $\tilde{\gamma}$
\begin{eqnarray*}
\left|f(u_i,\tilde{\gamma})-f(u_i,\gamma)\right| & \leq & C_{11} \|\tilde{\gamma}-\gamma\| \textrm{ for all } i \in U.
\end{eqnarray*}
\end{itemize}
\noindent It is assumed in (H1) that the inclusion probabilities do not differ much from that obtained under simple random sampling, so that no design weight dominates the other. It is assumed in (H2) that the units in the population are not far from being independently selected: this assumption is verified for stratified simple random sampling and rejective sampling \citep{haj:64}, for example. It is assumed in (H3) that the response probabilities are bounded away from $0$, i.e. there is no hard-core non-respondents, and that the probabilities of observing a null value are also bounded away from $0$, i.e. the variable of interest is not degenerate. The assumption (H4) is related to the imputation weights, and is similar to assumption (H1). The assumption (H5) is related to the imputation model, and is necessary to control the behaviour of the regularized estimator $\hat{B}_{ar}$; see \citet{car:gog:lar:13} and \citet{cha:dop:18}. It is assumed in (H6) that the estimator $\hat{\gamma}_r$ is $\sqrt{n}$ mean-square consistent for the parameter $\gamma$. This assumption is somewhat strong, but is needed to obtain the standard rate of convergence for the imputed estimator of the total. It is assumed in (H7) that $f(\cdot,\cdot)$ is Lipschitz-Continuous in its second component. The assumptions (H5) and (H6) are also considered in \citet{haz:cha:nam:14}.
\begin{prop} \label{prop1}
Suppose that the imputation model in (\ref{imput:model}) holds and that the assumptions (H1)-(H7) are satisfied. Then we have
\begin{eqnarray}
E\left\{\|\hat{B}_{ar}-\beta\|^2\right\} & = & O(n^{-1}). \label{prop1:eq1}
\end{eqnarray}
\end{prop}
\begin{prop} \label{prop2}
Suppose that the imputation model in (\ref{imput:model}) holds and that the assumptions (H1)-(H7) are satisfied. Then under the random imputation mechanism proposed in Section \ref{ssec33}, we have
\begin{eqnarray} \label{prop2:eq1}
E\left[\left\{N^{-1}(\hat{t}_{yI}-t_{y})\right\}^2\right] & = & O(n^{-1}).
\end{eqnarray}
\end{prop}
\begin{prop} \label{prop3}
Suppose that the imputation model in (\ref{imput:model}) holds and that the assumptions (H1)-(H7) are satisfied. Also, suppose that the distribution function $F_{\epsilon}$ is absolutely continuous. Then under the random imputation mechanism proposed in Section \ref{ssec33}, we have for any $t \in \mathbb{R}$
\begin{eqnarray} \label{prop3:eq1}
E \left[\left\{\hat{F}_I(t)-F_N(t)\right\}^2\right] & = & o(1).
\end{eqnarray}
\end{prop}
\section{Variance estimation} \label{sec5}
\noindent We now consider variance estimation for the imputed estimator of the total $\hat{t}_{yI}$, under the proposed imputation procedures. The variance estimators are adapted from a linearized variance estimator proposed by \citet[][Section 2]{kim:rao:09} for deterministic/random regression imputation. They are obtained under a variance decomposition which makes use of the reverse approach \citep{fay:96,sha:ste:99}. For simplicity, we suppose that the $\phi_i$'s are modeled according to a logistic regression model and that the unknown parameter $\beta$ is the solution of the weighted estimated equation
\begin{eqnarray} \label{sec5:eq1}
\sum_{i \in s} \omega_i r_i u_i \left\{\eta_i-f(u_i,\gamma)\right\} & = & 0,
\end{eqnarray}
with $\textrm{logit} f(u_i,\gamma)=u_i^{\top} \gamma$.
\subsection{Balanced imputation procedure} \label{ssec51}
\noindent We first consider the balanced imputation procedure proposed in Section \ref{ssec34}. We do not need to account for the imputation variance, since it is approximately eliminated for the estimation of the total with the proposed imputation procedure. By following the approach of Kim and Rao~(2009), we obtain after some algebra the two-term variance estimator
\begin{eqnarray} \label{sec5:eq2}
\hat{V}_{BMRR}(\hat{t}_{yI}) & = & \hat{V}_1(\hat{t}_{yI}) + \hat{V}_2(\hat{t}_{yI}),
\end{eqnarray}
see equations (10) and (13) in \citet{kim:rao:09}. The first term in the right-hand side of (\ref{sec5:eq2}) is
\begin{eqnarray} \label{sec5:eq3}
\hat{V}_1(\hat{t}_{yI}) & = & \sum_{i,j \in s} \left(\frac{\pi_{ij}-\pi_i \pi_j}{\pi_{ij}}\right) \hat{\xi}_i \hat{\xi}_j, \nonumber \\
\textrm{with } \quad \hat{\xi}_i & = & d_i(\hat{\phi}_i z_i^{\top} \hat{B}_{ar}) + r_i \left(d_i+\omega_i \hat{\phi}_i v_i^{-1} \hat{a}^{\top} z_i \right)\left(y_i-\hat{\phi}_i z_i^{\top} \hat{B}_{ar}\right) \nonumber \\
& + & r_i \omega_i (\hat{b}-\hat{c})^{\top} u_i \left(\eta_i-\hat{\phi}_i \right),
\end{eqnarray}
with
\begin{eqnarray} \label{sec5:eq3a}
\hat{a} & = & \left( \sum_{i \in s} r_i \omega_i \hat{\phi}_i v_i^{-1} z_i z_i^{\top} \right)^{-1} \sum_{i \in s} d_i (1-r_i) \hat{\phi}_i z_i, \nonumber \\
\hat{b} & = & \left( \sum_{i \in s} r_i \omega_i \hat{\phi}_i (1-\hat{\phi}_i) u_i u_i^{\top} \right)^{-1} \sum_{i \in s} d_i (1-r_i) \hat{\phi}_i (1-\hat{\phi}_i) (z_i^{\top}\hat{B}_{ar}) u_i, \\
\hat{c} & = & \left( \sum_{i \in s} r_i \omega_i \hat{\phi}_i (1-\hat{\phi}_i) u_i u_i^{\top} \right)^{-1} \sum_{i \in s} \omega_i r_i v_i^{-1} \hat{\phi}_i (1-\hat{\phi}_i) (z_i^{\top}\hat{a}) (z_i^{\top}\hat{B}_{ar}) u_i, \nonumber
\end{eqnarray}
and with $\pi_{ij}$ the probability that units $i$ and $j$ are selected together in the sample. The second term in the right-hand side of (\ref{sec5:eq2}) is
\begin{eqnarray} \label{sec5:eq4}
\hat{V}_2(\hat{t}_{yI}) = \sum_{i \in s} r_i d_i \left\{(1+\omega_i \pi_i v_i^{-1} \hat{a}^{\top} z_i) (y_i-\hat{\phi}_i z_i^{\top} \hat{B}_{ar})
+\omega_i \pi_i (\hat{b}-\hat{c})^{\top} u_i (\eta_i-\hat{\phi}_i) \right\}^2.
\end{eqnarray}
As underlined by \citet{kim:rao:09}, $\hat{V}_2(\hat{t}_{yI})$ is not sensitive to a mis-specification of the covariance structure in model (\ref{imput:model}).
\subsection{Random imputation procedure} \label{ssec52}
\noindent We now consider the random imputation procedure proposed in Section \ref{ssec33}. We need to account for the additional variance due to the imputation process. By following once again the approach in \citet[][Section 4.1]{kim:rao:09}, we obtain the variance estimator
\begin{eqnarray} \label{sec5:eq5}
\hat{V}_{MRR}(\hat{t}_{yI}) & = & \hat{V}_{BMRR}(\hat{t}_{yI}) + \hat{V}_3(\hat{t}_{yI}),
\end{eqnarray}
where $\hat{V}_{BMRR}(\hat{t}_{yI})$ is given in equation (\ref{sec5:eq2}), and with
\begin{eqnarray} \label{sec5:eq6}
\hat{V}_3(\hat{t}_{yI}) = \sum_{i \in s} d_i^2 (1-r_i) (y_i^*-\hat{\phi}_i z_i^{\top} \hat{B}_{ar})^2,
\end{eqnarray}
with $y_i^*$ the imputed value given in equation (\ref{imput:mecha:3}).
\section{Simulation study} \label{sec6}
\noindent To evaluate the performance of the proposed imputation methods, we implement a simulation study inspired by \citet{haz:cha:nam:14}. We generate nine finite populations of size $N=10,000$ with an interest variable $y$ and four auxiliary variable $z_1,\ldots,z_4$. The values of $z_i,~i=1,\ldots,4,$ are generated according to a Gamma distribution with shift parameter $2$ and scale parameter $5$. The values of $y$ are generated according to the following mixture model:
\begin{eqnarray} \label{generatoDPRop}
y_i & = & \eta_i(a_0 + a_1 z_{1i} + a_2 z_{2i} + a_3 z_{3i} + a_4 z_{4i} + \epsilon_{i}),
\end{eqnarray}
where the $\epsilon_i$'s are generated according to a standard normal distribution with variance $\sigma^2$. We use $a_0=30$ and $a_1=a_2=a_3=a_4=0.7$. Also, we choose three different values of $\sigma^2$ so that the coefficient of determination $R^2$ equals $0.4$, $0.5$ or $0.6$ for the units $i$ such that $\eta_i=1$. \\
\noindent The $\eta_i$'s are generated according to a Bernoulli distribution with parameter $\phi_i$, and
\begin{eqnarray} \label{eq:phik}
\log\left(\frac{\phi_i}{1-\phi_i}\right) & = & b_0 + b_1 z_{1i} + b_2 z_{2i} + b_3 z_{3i} + b_4 z_{4i},
\end{eqnarray}
and different values for the parameters $b_0,\ldots,b_4$, chosen so that the proportion of non-null values is approximately equal to $0.60$, $0.70$, or $0.80$. The three different proportion of non-null values, crossed with the three different levels for the $R^2$, lead to the nine finite populations. \\
\noindent In each population, we select $R=1,000$ samples by means of rejective sampling \citep{haj:64} of size $n=500$, with inclusion probabilities proportional to the variable $z_{1i}$. In each sample, we generate a response indicator $r_i$ for unit $i$ according to a Bernoulli distribution with parameter $p_i$ such that
\begin{eqnarray}\label{eq:pk}
\log\left(\frac{p_i}{1-p_i}\right) & = & c_0 + c_1 z_{1i} + c_2 z_{2i} + c_3 z_{3i} + c_4 z_{4i}.
\end{eqnarray}
We use different values for the parameters $c_0,\ldots,c_4$, chosen so that the proportion of respondents is approximately equal to $0.30$, $0.50$ or $0.70$.
\subsection{Properties of point estimators} \label{ssec61}
\noindent In this Section, we are interested in estimating the total $t_y$, and the distribution function $F_{N}(t)$ with $t=t_{\alpha}$, the $\alpha$-th quantile. In this simulation study, we consider the values $\alpha= 0.50, 0.75$ and $0.90$. We compare four imputation methods to handle non-response: (i) random imputation ($RR_\phi$) proposed by \citet{haz:cha:nam:14}, and presented in Section \ref{ssec31}; (ii) balanced random imputation ($BRR_\phi$) proposed by \citet{haz:cha:nam:14}, and presented in Section \ref{ssec32}; (iii) proposed random imputation method ($MRR_\phi$), presented in Section \ref{ssec33}; (iv) proposed balanced random imputation method ($BMRR_\phi$), presented in Section \ref{ssec34}. For each of the four methods, we use imputation weights $\omega_i=1$, and the $\phi_i$'s and $p_i$'s are estimated by means of logistic regression modeling. In each sample, missing values are replaced by imputed values according to imputation methods (i) to (iv), and the imputed estimators $\hat{t}_{yI}$ and $\hat{F}_I(t_{\alpha})$ are computed. \\
\noindent As a measure of bias of an estimator $\hat{\theta}_I$ of a finite population parameter $\theta$, we compute the Monte Carlo percent relative bias
\begin{eqnarray} \label{biaisRelatif}
RB_{MC}(\hat{\theta}_I) & = & \frac{100}{R} \sum_{k=1}^R \frac{(\hat{\theta}_{I(k)} -\theta)}{\theta}, \label{biaisRelatif}
\end{eqnarray}
where $\hat{\theta}_{I(k)}$ denotes the imputed estimator computed in the $k$-th sample. As a measure of relative efficiency for each imputation method, using $BMRR_\phi$ as a benchmark, we computed
\begin{eqnarray*}
RE_{MC}(\hat{\theta}_I)=\frac{MSE_{MC}(\hat {\theta}_{I})}{MSE_{MC}(\hat {\theta}_{BMRR_\phi})} & \textrm{with} & MSE_{MC}(\hat{\theta}_I) = \frac{1}{R} \sum_{k=1}^R(\hat{\theta}_{I(k)} -\theta)^2, \label{erreurQuadratique}
\end{eqnarray*}
the Mean Square Error of $\hat{\theta}_I$ approximated by means of the $R$ simulations. We observed no qualitative difference according to the different response rates. For brevity, we therefore only present the simulation results with an average proportion of respondents of $0.50$. The simulation results for the two other response rates are given in the Supplementary Material. \\
\noindent We first consider the estimation of the total $t_y$, for which the simulation results are given in Table \ref{tab:total:05}. The four imputation methods lead to approximately unbiased estimators of the total, as expected. Turning to the relative efficiency (RE), we note that in all studied cases the balanced version of an imputation method outperforms its unbalanced version. Also, the two balanced imputation procedures exhibit similar efficiency, with $BRR_\phi$ performing slightly better. This is likely due to fact that the balancing equations (\ref{ssec34:eq1}) and (\ref{ssec34:eq2}) are not exactly respected due to the landing phase of the cube method \citep[see][]{dev:til:04}. \\
\begin{table}
\begin{center}
\begin{tabular}{|ll|c|c|c|c|c|c|c|c|}
\hline
& & \multicolumn{2}{|c|}{$RR_{\phi}$} & \multicolumn{2}{|c|}{$BRR_{\phi}$} & \multicolumn{2}{|c|}{$MRR_{\phi}$} & \multicolumn{2}{|c|}{$BMRR_{\phi}$} \\ \hline
$R^2$ & $\overline{\phi}$ & RB \% & RE & RB \% & RE & RB \% & RE & RB \% & RE \\ \hline
0.4 & 0.6 & 0.23 & 1.16 & 0.25 & 0.99 & 0.16 & 1.21 & 0.28 & 1.00 \\
0.4 & 0.7 & 0.11 & 1.07 & 0.26 & 0.96 & 0.09 & 1.14 & 0.32 & 1.00 \\
0.4 & 0.8 & 0.35 & 1.06 & 0.41 & 0.98 & 0.34 & 1.14 & 0.42 & 1.00 \\
\hline
0.5 & 0.6 & 0.33 & 1.09 & 0.27 & 0.99 & 0.31 & 1.12 & 0.24 & 1.00 \\
0.5 & 0.7 & 0.26 & 1.16 & 0.35 & 0.99 & 0.23 & 1.23 & 0.31 & 1.00 \\
0.5 & 0.8 & 0.44 & 1.13 & 0.44 & 0.99 & 0.43 & 1.21 & 0.45 & 1.00 \\
\hline
0.6 & 0.6 & 0.33 & 1.17 & 0.37 & 0.99 & 0.32 & 1.21 & 0.33 & 1.00 \\
0.6 & 0.7 & 0.18 & 1.13 & 0.35 & 0.99 & 0.15 & 1.18 & 0.35 & 1.00 \\
0.6 & 0.8 & 0.46 & 1.09 & 0.50 & 0.98 & 0.43 & 1.16 & 0.49 & 1.00 \\ \hline
\end{tabular}
\caption{Relative bias (RB \%) and Relative efficiency (RE) of four imputed estimators of the total with an average response probability of $50 \% $} \label{tab:total:05}
\end{center}
\end{table}
\noindent We now consider the estimation of the population distribution function, for which the simulation results are presented in Table \ref{tab:F50:F75:F90:05}. In all cases, the two proposed imputation methods $MRR_\phi$ and $BMRR_\phi$ lead to approximately unbiased estimators of the distribution function, with absolute relative biases no greater than 2 \% . On the contrary, the $RR_\phi$ and the $BRR_\phi$ imputation methods lead to biased estimators, and the absolute relative bias can be as large as 14 \% . We note that the bias is larger for the lower quantiles. Turning to the relative efficiency, we note that $MRR_\phi$ and $BMRR_\phi$ always outperform $RR_\phi$ and $BRR_\phi$, which is partly due to the bias under these latter imputation methods. Comparing the two proposed imputation methods, we note that $BMRR_\phi$ is systematically better than $MRR_\phi$ in terms of efficiency, with values of $RE$ ranging from $1.08$ to $1.32$ for $MRR_\phi$. \\
\begin{table}
\begin{center}
\begin{tabular}{|ll|c|c|c|c|c|c|c|c|}
\hline
& & \multicolumn{2}{|c|}{$RR_{\phi}$} & \multicolumn{2}{|c|}{$BRR_{\phi}$} & \multicolumn{2}{|c|}{$MRR_{\phi}$} & \multicolumn{2}{|c|}{$BMRR_{\phi}$} \\ \hline
& & RB \% & RE & RB \% & RE & RB \% & RE & RB \% & RE \\ \hline
$R^2$ & $\overline{\phi}$ & \multicolumn{8}{|c|}{$50 \% $ quartile} \\ \cline{1-10}
0.4 & 0.6 & -6.58 & 2.29 & -6.62 & 2.12 & 0.07 & 1.27 & -0.03 & 1.00 \\
0.4 & 0.7 & -12.03 & 4.37 & -12.23 & 4.36 & 0.90 & 1.24 & 0.76 & 1.00 \\
0.4 & 0.8 & -14.07 & 6.40 & -14.17 & 6.47 & 0.39 & 1.20 & 0.48 & 1.00 \\
0.5 & 0.6 & -6.78 & 2.22 & -6.71 & 2.11 & -0.07 & 1.17 & 0.00 & 1.00 \\
0.5 & 0.7 & -12.14 & 4.35 & -12.26 & 4.26 & 1.01 & 1.26 & 0.95 & 1.00 \\
0.5 & 0.8 & -12.97 & 6.38 & -12.94 & 6.24 & 0.64 & 1.32 & 0.79 & 1.00 \\
0.6 & 0.6 & -6.71 & 2.40 & -6.75 & 2.20 & 0.15 & 1.26 & 0.26 & 1.00 \\
0.6 & 0.7 & -12.06 & 4.58 & -12.26 & 4.59 & 0.98 & 1.23 & 0.66 & 1.00 \\
0.6 & 0.8 & -11.37 & 5.37 & -11.33 & 5.27 & 1.05 & 1.24 & 1.08 & 1.00 \\ \hline \hline
$R^2$ & $\overline{\phi}$ & \multicolumn{8}{|c|}{$75 \% $ quartile} \\ \cline{1-10}
0.4 & 0.6 & 6.80 & 4.13 & 6.83 & 4.15 & 1.45 & 1.18 & 1.31 & 1.00 \\
0.4 & 0.7 & 8.12 & 5.17 & 8.10 & 5.16 & 1.39 & 1.18 & 1.34 & 1.00 \\
0.4 & 0.8 & 8.07 & 5.64 & 8.06 & 5.63 & 0.60 & 1.23 & 0.72 & 1.00 \\
0.5 & 0.6 & 6.46 & 3.93 & 6.47 & 3.91 & 1.28 & 1.21 & 1.42 & 1.00 \\
0.5 & 0.7 & 7.61 & 4.81 & 7.61 & 4.81 & 1.27 & 1.18 & 1.34 & 1.00 \\
0.5 & 0.8 & 7.63 & 4.95 & 7.64 & 4.95 & 0.77 & 1.21 & 0.79 & 1.00 \\
0.6 & 0.6 & 6.12 & 3.68 & 6.12 & 3.64 & 1.39 & 1.21 & 1.52 & 1.00 \\
0.6 & 0.7 & 7.38 & 4.52 & 7.36 & 4.50 & 1.51 & 1.22 & 1.53 & 1.00 \\
0.6 & 0.8 & 7.14 & 4.49 & 7.15 & 4.48 & 0.80 & 1.17 & 0.86 & 1.00 \\ \hline \hline
$R^2$ & $\overline{\phi}$ & \multicolumn{8}{|c|}{$90 \% $ quartile} \\ \cline{1-10}
0.4 & 0.6 & 3.27 & 2.86 & 3.27 & 2.85 & 0.80 & 1.19 & 0.72 & 1.00 \\
0.4 & 0.7 & 3.55 & 2.89 & 3.55 & 2.89 & 0.98 & 1.08 & 0.91 & 1.00 \\
0.4 & 0.8 & 3.46 & 3.37 & 3.46 & 3.37 & 0.60 & 1.15 & 0.55 & 1.00 \\
0.5 & 0.6 & 3.10 & 2.64 & 3.10 & 2.63 & 0.74 & 1.21 & 0.79 & 1.00 \\
0.5 & 0.7 & 3.54 & 2.93 & 3.53 & 2.92 & 1.04 & 1.08 & 1.08 & 1.00 \\
0.5 & 0.8 & 3.43 & 3.43 & 3.43 & 3.43 & 0.69 & 1.18 & 0.64 & 1.00 \\
0.6 & 0.6 & 3.19 & 2.47 & 3.19 & 2.46 & 1.02 & 1.14 & 1.03 & 1.00 \\
0.6 & 0.7 & 3.39 & 2.87 & 3.39 & 2.86 & 1.08 & 1.16 & 1.06 & 1.00 \\
0.6 & 0.8 & 3.26 & 3.16 & 3.26 & 3.16 & 0.61 & 1.14 & 0.66 & 1.00 \\ \hline
\end{tabular}
\caption{Relative bias (RB \%) and Relative efficiency (RE) of four imputed estimators of the distribution function evaluated at the $50 \% $, $75 \% $ and $90 \% $ quartiles with an average response probability of $50 \% $} \label{tab:F50:F75:F90:05}
\end{center}
\end{table}
\noindent We also conducted additional simulations to evaluate the influence of the specific distribution used to simulate the random residuals in the imputation model. More precisely, we generated the variable of interest according to the mixture model presented in (\ref{generatoDPRop}), but with the residuals $\epsilon_i$'s generated either (a) from a gamma distribution or (b) from a log-normal distribution with variance $\sigma^2$. The simulation results are presented in the Supplementary Material. We observed no qualitative difference as compared to normally distributed residuals.
\subsection{Properties of variance estimators} \label{ssec62}
\noindent We now consider the properties of the variance estimators proposed in Section \ref{sec5}. Under the rejective sampling design used in the simulation study, we replace the component $\hat{V}_1(\hat{t}_{yI})$ given in (\ref{sec5:eq3}) with the Hajek-Rosen variance estimator
\begin{eqnarray} \label{ssec62:eq1}
\hat{V}_{HR,1}(\hat{t}_{yI}) = \frac{n}{n-1} \sum_{i \in s} (1-\pi_i) (\hat{\xi}_i-\hat{R})^2 \textrm{ with } \hat{R}=\frac{\sum_{i \in s} (1-\pi_i) \hat{\xi}_i}{\sum_{i \in s} (1-\pi_i)},
\end{eqnarray}
see also \citet{cha:dop:18}. This leads to the simplified variance estimator
\begin{eqnarray} \label{ssec62:eq2}
\tilde{V}_{BMRR}(\hat{t}_{yI}) & = & \hat{V}_{HR,1}(\hat{t}_{yI}) + \hat{V}_2(\hat{t}_{yI}),
\end{eqnarray}
for the proposed balanced imputation procedure $BMRR_{\phi}$, and to the simplified variance estimator
\begin{eqnarray} \label{ssec62:eq3}
\tilde{V}_{MRR}(\hat{t}_{yI}) & = & \tilde{V}_{BMRR}(\hat{t}_{yI}) + \hat{V}_3(\hat{t}_{yI}),
\end{eqnarray}
for the proposed random imputation procedure $MRR_{\phi}$. \\
\noindent We computed the Monte-Carlo percent relative bias of these two variance estimators, using an independent simulation-based approximation of the true mean square error of $\hat{t}_{yI}$ based on $10,000$ simulations. We also computed the coverage rates of the associated normality-based confidence intervals, with nominal error rate of $2.5 \% $ in each tail. We only consider the two cases when the average proportion of respondents is $0.50$ and $0.70$. We first consider the results for $BMRR_{\phi}$, which are presented in Table \ref{evar:bal}. In all cases, the variance estimator $\tilde{V}_{BMRR}(\hat{t}_{yI})$ has a small bias, no greater than 6 \% . The variance estimator is slightly negatively biased with $\bar{p}=0.50$. This is likely due to the fact that the imputation variance is not completely eliminated with the proposed balanced imputation procedure, due to the landing phase of the cube method. The coverage rates are approximately respected in any case. We now turn to $MRR_{\phi}$, for which the simulation results are presented in Table \ref{evar:rand}. The variance estimator $\tilde{V}_{MRR}(\hat{t}_{yI})$ is approximately unbiased with $\bar{p}=0.50$, but is slightly positively biased with $\bar{p}=0.70$. The coverage rates are approximately respected in all cases.
\begin{table}
\begin{center}
\begin{tabular}{|l|cc|cc|cc|}
\hline
& \multicolumn{6}{|c|}{Population 1} \\ \cline{2-7}
& \multicolumn{2}{|c|}{$\bar{\phi}=0.6$} & \multicolumn{2}{|c|}{$\bar{\phi}=0.7$} & \multicolumn{2}{|c|}{$\bar{\phi}=0.8$} \\
& $\bar{p}=0.5$ & $\bar{p}=0.7$ & $\bar{p}=0.5$ & $\bar{p}=0.7$ & $\bar{p}=0.5$ & $\bar{p}=0.7$ \\ \hline
RB (\% ) & -5.6 & 5.1 & -5.4 & 2.1 & -4.7 & 3.2 \\
Cov. Rate & 93.4 & 95.3 & 93.8 & 95.3 & 93.5 & 95.7 \\ \hline \hline
& \multicolumn{6}{|c|}{Population 2} \\ \cline{2-7}
& \multicolumn{2}{|c|}{$\bar{\phi}=0.6$} & \multicolumn{2}{|c|}{$\bar{\phi}=0.7$} & \multicolumn{2}{|c|}{$\bar{\phi}=0.8$} \\
& $\bar{p}=0.5$ & $\bar{p}=0.7$ & $\bar{p}=0.5$ & $\bar{p}=0.7$ & $\bar{p}=0.5$ & $\bar{p}=0.7$ \\ \hline
RB (\% ) & -4.5 & 5.3 & -3.7 & 3.2 & -3.6 & 2.7 \\
Cov. Rate & 93.4 & 95.8 & 93.0 & 95.3 & 93.6 & 95.9 \\ \hline \hline
& \multicolumn{6}{|c|}{Population 3} \\ \cline{2-7}
& \multicolumn{2}{|c|}{$\bar{\phi}=0.6$} & \multicolumn{2}{|c|}{$\bar{\phi}=0.7$} & \multicolumn{2}{|c|}{$\bar{\phi}=0.8$} \\
& $\bar{p}=0.5$ & $\bar{p}=0.7$ & $\bar{p}=0.5$ & $\bar{p}=0.7$ & $\bar{p}=0.5$ & $\bar{p}=0.7$ \\ \hline
RB (\% ) & -4.7 & 4.8 & -4.0 & 2.2 & -4.3 & 2.8 \\
Cov. Rate & 93.6 & 95.5 & 93.5 & 95.3 & 93.1 & 95.9 \\ \hline
\end{tabular}
\caption{Monte-Carlo percent relative bias of the variance estimator and coverage rate for the proposed balanced imputation procedure $BMRR_{\phi}$} \label{evar:bal}
\end{center}
\end{table}
\begin{table}
\begin{center}
\begin{tabular}{|l|cc|cc|cc|}
\hline
& \multicolumn{6}{|c|}{Population 1} \\ \cline{2-7}
& \multicolumn{2}{|c|}{$\bar{\phi}=0.6$} & \multicolumn{2}{|c|}{$\bar{\phi}=0.7$} & \multicolumn{2}{|c|}{$\bar{\phi}=0.8$} \\
& $\bar{p}=0.5$ & $\bar{p}=0.7$ & $\bar{p}=0.5$ & $\bar{p}=0.7$ & $\bar{p}=0.5$ & $\bar{p}=0.7$ \\ \hline
RB (\% ) & -2.1 & 6.5 & -2.2 & 4.6 & -1.2 & 5.0 \\
Cov. Rate & 93.5 & 95.1 & 94.3 & 96.2 & 93.7 & 96.2 \\ \hline \hline
& \multicolumn{6}{|c|}{Population 2} \\ \cline{2-7}
& \multicolumn{2}{|c|}{$\bar{\phi}=0.6$} & \multicolumn{2}{|c|}{$\bar{\phi}=0.7$} & \multicolumn{2}{|c|}{$\bar{\phi}=0.8$} \\
& $\bar{p}=0.5$ & $\bar{p}=0.7$ & $\bar{p}=0.5$ & $\bar{p}=0.7$ & $\bar{p}=0.5$ & $\bar{p}=0.7$ \\ \hline
RB (\% ) & -1.6 & 7.2 & -3.6 & 4.4 & -0.3 & 4.0 \\
Cov. Rate & 94.2 & 94.8 & 93.6 & 95.9 & 92.7 & 96.2 \\ \hline \hline
& \multicolumn{6}{|c|}{Population 3} \\ \cline{2-7}
& \multicolumn{2}{|c|}{$\bar{\phi}=0.6$} & \multicolumn{2}{|c|}{$\bar{\phi}=0.7$} & \multicolumn{2}{|c|}{$\bar{\phi}=0.8$} \\
& $\bar{p}=0.5$ & $\bar{p}=0.7$ & $\bar{p}=0.5$ & $\bar{p}=0.7$ & $\bar{p}=0.5$ & $\bar{p}=0.7$ \\ \hline
RB (\% ) & -1.5 & 7.0 & -1.3 & 5.5 & 0.1 & 3.5 \\
Cov. Rate & 94.1 & 95.8 & 94.6 & 95.2 & 93.7 & 95.8 \\ \hline
\end{tabular}
\caption{Monte-Carlo percent relative bias of the variance estimator and coverage rate for the proposed random imputation procedure $MRR_{\phi}$} \label{evar:rand}
\end{center}
\end{table}
\section{Application} \label{sec:appli}
\noindent We apply the proposed imputation methods on data modelled from the Monthly Retail Trade Survey \citep{mul:oli:kap:14,boi:cha:haz:16,cha:dop:18}, which have been created to reproduce as closely as possible the original survey data. We consider the variable giving the sales ($y_{0i}$). We create in the dataset a domain indicator, equal to $1$ if the unit is in the domain and to $0$ otherwise. The variable of interest is $y_{i} = y_{0i} 1(i \in d)$, and we are interested in estimating the total and the distribution function of this variable. This case occurs when we are interested in domain estimation, and when the domain itself is not known for all sampled units due to non-response. For imputation purpose, we use as auxiliary variables a measure of size ($z_{1i}$), the prior month sales ($z_{2i}$) and the prior month inventories ($z_{3i}$). \\
\noindent The survey data arise from a stratified simple random sampling design with $6$ strata $U_h,~h=1,\ldots,6$. In this application, we leave apart the take-all stratum, which leads to five strata with sizes $N_h$ ranging from $463$ to $9~993$, and with sample sizes $n_h$ ranging from $57$ to $145$. The number of responding units per stratum $n_{rh}$ varies from $44$ to $75$. We suppose that the response mechanism is Missing At Random (MAR), and is explained by the strata indicators. In other words, we suppose that the response mechanism is uniform within each stratum. \\
\noindent The variable $y_{i}$ suffers from item non-response. We use an imputation model similar to that in \citet{boi:cha:haz:16}, but adapted to cover zero-inflated variables. More precisely, we suppose that each stratum $U_h$ is partitioned into $G_h$ imputation cells, obtained by ranking the units with respect to $z_{1i}$. The imputation model is
\begin{eqnarray} \label{imput:model:appli}
y_{i} & = & \eta_i \left\{\beta_{hg} + \epsilon_i \right\}
\end{eqnarray}
for any unit $i$ belonging to the cell $g$ in stratum $U_h$. This is a particular case of the imputation model given in (\ref{imput:model:appli}), using for each stratum $U_h$ as auxiliary information $z_i$ the set of cell indicators. \\
\noindent We perform the imputation methods presented in Sections \ref{ssec33} and \ref{ssec34}, using equal imputation weights $\omega_i$. For any non-responding unit $i$, we obtain the estimated probability $\hat{\phi}_i$ through a logistic regression on the set $u_i=(1,z_{1i},z_{2i},z_{3i})^{\top}$ of auxiliary variables. Inside each stratum $U_h$, the estimator $\hat{B}_{arh}$ of $\beta_h$ is obtained from equation (\ref{ssec33:eq1b}), with $z_{i}$ the vector of cell indicators. We used $a=0.05$, and in this case no regularization was needed. The imputed values are then obtained from equation (\ref{imput:mecha:3}) for the proposed random imputation procedure, and from equation (\ref{imput:mecha:4}) for the proposed balanced random imputation procedure. \\
\noindent For each of the two imputation methods, we computed the imputed estimator of the total $\hat{t}_{yI}$ and the imputed estimator of the distribution function $\hat{F}_{I}(t)$ for several values of $t$. We also computed a with-replacement bootstrap variance estimator for the imputed estimators, see \cite{boi:cha:haz:16} and \cite{cha:dop:18}. The bootstrap is performed as if the samples were selected with replacement, which is reasonable in view of the small sampling rates inside strata. To compare the efficiency of the imputed estimators, we compute
\begin{eqnarray} \label{appli:eq6}
re & = & \frac{v_{boot}\{\hat{\theta}_{EBRI}\}}{v_{boot}\{\hat{\theta}_{BRI}\}}.
\end{eqnarray}
\noindent The results are shown in Table \ref{tab:appli:results}. From the imputed data set, both imputation methods give similar results in terms of point estimation. Turning to relative efficiency, we note that the proposed exact balanced random imputation procedure yields more efficient estimations, with values of $re$ ranging from $0.87$ to $0.94$.
\renewcommand{\arraystretch}{1.0}
\begin{table}[htb!]
\begin{center}
\begin{tabular}{lcccccccccc} \hline
& & $\hat{t}_{yI}$ & & \multicolumn{7}{c}{$\hat{F}_{yI}(t)$ with $t (\times 1,000)$} \\ \cline{5-11}
& & $(\times 10^9)$ & & 300 & 700 & 1,000 & 2,000 & 5,000 & 8,000 & 10,000 \\ \hline \hline
EBRI & & $29.94$ & & $0.38$ & $0.51$ & $0.60$ & $0.75$ & $0.94$ & $0.98$ & $0.99$ \\
BRI & & $30.44$ & & $0.37$ & $0.50$ & $0.60$ & $0.74$ & $0.94$ & $0.98$ & $0.99$ \\
re & & $0.91$ & & $0.88$ & $0.88$ & $0.89$ & $0.87$ & $0.94$ & $0.92$ & $0.92$ \\ \hline
\end{tabular}
\caption{Imputed estimator of the total and of the distribution function, and estimated related efficiency with two imputation methods \label{tab:appli:results}}\end{center}
\end{table}
\section{Conclusion} \label{sec7}
\noindent In this paper, we considered imputation for zero-inflated data. We proposed two imputation methods which enable to respect the nature of the data, and which preserve the finite population distribution function. In particular, we proposed a balanced imputation method which enables to preserve the distribution of the imputed variable while being fully efficient for the estimation of a total. \\
\noindent Our imputation methods rely upon the mixture regression imputation model proposed by \citet{haz:cha:nam:14}. As mentioned by these authors, the proposed methods could be extended to more general mixture regression models, for example to handle count data. \\
\noindent In practice, we may not be interested in the distribution function in itself, but rather in complex parameters such as quantiles. Establishing the theoretical properties of estimators of such parameters under the proposed imputation procedures is a challenging task, and is currently under investigation.
\bibliographystyle{apalike}
|
\section*{Supplementary material}
\label{sec:appendix}
The experiment of section~\ref{sec:sparse_graph_interpretation} is also done with the LeNet-5 network. But the sparse graph interpretation is only applied to the dense layers. The convolution layers are not manipulated. To use the trained weights of the convolution layers and set them as untrainable is a closer approach to the experiment in section~\ref{sec:sparse_graph_interpretation}. Table~\ref{tab:pruning_test2} and \ref{tab:pruning_test2_2} show the results for this case. Table~\ref{tab:pruning_test3} and \ref{tab:pruning_test3_2} show the result of the case that the convolution layers are trainable and initialized with the original weights of the unpruned network.
\begin{table}[h]
\caption{Changing rate of active connections for a pruned network based on sparse graph initialization for different remaining network sizes. The origin LeNet-5 network reaches a validation accuracy of $\unit[75.44]{\%}$ in 200 epochs for CIFAR10 (setup see section~\ref{sub:cifar_10}) with the initialization by \citet{He2015}. The pruned networks are trained with a learning rate of $10^{-3}$.}
\label{tab:pruning_test2}
\centering
\begin{tabular}{rrrrrr}
\toprule
active net & \multicolumn{4}{c}{changed connections} & accuracy \\
& hidden 0 & hidden 1 & output & over all & \\
\toprule
$\unit[10.0]{\%}$ & $\unit[37.00]{\%}$ & $\unit[33.52]{\%}$ & $\unit[12.05]{\%}$ & $\unit[34.80]{\%}$ & $\unit[70.26]{\%}$ \\
\midrule
$\unit[20.0]{\%}$ & $\unit[18.41]{\%}$ & $\unit[18.13]{\%}$ & $\unit[7.80]{\%}$ & $\unit[17.98]{\%}$ & $\unit[72.71]{\%}$ \\
\midrule
$\unit[30.0]{\%}$ & $\unit[8.88]{\%}$ & $\unit[10.37]{\%}$ & $\unit[6.60]{\%}$ & $\unit[9.09]{\%}$ & $\unit[73.69]{\%}$ \\
\midrule
$\unit[40.0]{\%}$ & $\unit[4.96]{\%}$ & $\unit[5.97]{\%}$ & $\unit[6.01]{\%}$ & $\unit[5.16]{\%}$ & $\unit[73.54]{\%}$ \\
\midrule
$\unit[50.0]{\%}$ & $\unit[2.86]{\%}$ & $\unit[5.01]{\%}$ & $\unit[6.04]{\%}$ & $\unit[3.28]{\%}$ & $\unit[73.41]{\%}$ \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[h]
\caption{Pearson correlation coefficient of the active edges between the trained parent and child net for a pruned network based on sparse graph initialization for different remaining network sizes. The origin LeNet-5 network reaches a validation accuracy of $\unit[75.44]{\%}$ in 200 epochs for CIFAR10 (setup see section~\ref{sub:cifar_10}) with a uniform initialization by \citet{He2015}. For all given Pearson correlation coefficient the p-value is $0$ numerically. The pruned networks are trained with a learning rate of $10^{-3}$.}
\label{tab:pruning_test2_2}
\centering
\begin{tabular}{rllllr}
\toprule
active net & \multicolumn{4}{c}{Pearson correlation coefficient} & accuracy \\
& hidden 0 & hidden 1 & output & over all & \\
\toprule
$\unit[10.0]{\%}$ & $0.793$ & $0.812$ & $0.936$ & $0.806$ & $\unit[70.26]{\%}$ \\
\midrule
$\unit[20.0]{\%}$ & $0.903$ & $0.905$ & $0.954$ & $0.905$ & $\unit[72.71]{\%}$ \\
\midrule
$\unit[30.0]{\%}$ & $0.954$ & $0.946$ & $0.965$ & $0.953$ & $\unit[73.69]{\%}$ \\
\midrule
$\unit[40.0]{\%}$ & $0.975$ & $0.969$ & $0.968$ & $0.974$ & $\unit[73.54]{\%}$ \\
\midrule
$\unit[50.0]{\%}$ & $0.985$ & $0.971$ & $0.966$ & $0.983$ & $\unit[73.41]{\%}$ \\
\bottomrule
\end{tabular}
\end{table}
Even that the changing rates are higher than for the test from section~\ref{sec:sparse_graph_interpretation}, the network shows the same behavior like the LeNet-300-100 experiment. Only the reproducing effect is weaker. Probably this happens, because the development process of the convolution layers and the dense layers are not the same. In the original network both learn and develop together. The result of one sector influences the result of the other. Even though the convolution layers are fixed to a useful state and the dense sector use this, its learning process differs.
By using the convolution layers as a trainable part of the network, which is originally initialized, the difference between the original network and the sparse graph variants increases. This can be expected, because the additional parameters of the convolution layers increase the amount of opportunities to solve the problem. On the other hand, the starting point of the dense is radically changed. Both sectors interact through the learning. Thus it is clear that this variant behaves worse than with fixed convolution layers.
\begin{table}[h]
\caption{Changing rate of active connections for a pruned network based on sparse graph initialization for different remaining network sizes and trainable convolution layers. The origin LeNet-5 network reaches a validation accuracy of $\unit[75.44]{\%}$ in 200 epochs for CIFAR10 (setup see section~\ref{sub:cifar_10}) with the initialization by \citet{He2015}. The pruned networks are trained with a learning rate of $10^{-3}$.}
\label{tab:pruning_test3}
\centering
\begin{tabular}{rrrrrr}
\toprule
active net & \multicolumn{4}{c}{changed connections} & accuracy \\
& hidden 0 & hidden 1 & output & over all & \\
\toprule
$\unit[10.0]{\%}$ & $\unit[44.59]{\%}$ & $\unit[39.23]{\%}$ & $\unit[13.36]{\%}$ & $\unit[41.58]{\%}$ & $\unit[66.46]{\%}$ \\
\midrule
$\unit[20.0]{\%}$ & $\unit[27.73]{\%}$ & $\unit[23.84]{\%}$ & $\unit[6.34]{\%}$ & $\unit[26.16]{\%}$ & $\unit[68.33]{\%}$ \\
\midrule
$\unit[30.0]{\%}$ & $\unit[19.17]{\%}$ & $\unit[14.52]{\%}$ & $\unit[4.74]{\%}$ & $\unit[17.90]{\%}$ & $\unit[70.36]{\%}$ \\
\midrule
$\unit[40.0]{\%}$ & $\unit[11.45]{\%}$ & $\unit[12.78]{\%}$ & $\unit[6.01]{\%}$ & $\unit[11.55]{\%}$ & $\unit[71.77]{\%}$ \\
\midrule
$\unit[50.0]{\%}$ & $\unit[10.14]{\%}$ & $\unit[12.85]{\%}$ & $\unit[6.03]{\%}$ & $\unit[10.50]{\%}$ & $\unit[70.25]{\%}$ \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[h]
\caption{Pearson correlation coefficient of the active edges between the trained parent and child net for a pruned network based on sparse graph initialization for different remaining network sizes and trainable convolution layers. The origin LeNet-5 network reaches a validation accuracy of $\unit[75.44]{\%}$ in 200 epochs for CIFAR10 (setup see section~\ref{sub:cifar_10}) with a uniform initialization by \citet{He2015}. For all given Pearson correlation coefficient the p-value is $0$ numerically. The pruned networks are trained with a learning rate of $10^{-3}$.
}
\label{tab:pruning_test3_2}
\centering
\begin{tabular}{rllllr}
\toprule
active net & \multicolumn{4}{c}{Pearson correlation coefficient} & accuracy \\
& hidden 0 & hidden 1 & output & over all & \\
\toprule
$\unit[10.0]{\%}$ & $0.737$ & $0.772$ & $0.929$ & $0.758$ & $\unit[66.46]{\%}$ \\
\midrule
$\unit[20.0]{\%}$ & $0.846$ & $0.872$ & $0.964$ & $0.856$ & $\unit[68.33]{\%}$ \\
\midrule
$\unit[30.0]{\%}$ & $0.896$ & $0.923$ & $0.975$ & $0.903$ & $\unit[70.36]{\%}$ \\
\midrule
$\unit[40.0]{\%}$ & $0.938$ & $0.923$ & $0.958$ & $0.936$ & $\unit[71.77]{\%}$ \\
\midrule
$\unit[50.0]{\%}$ & $0.937$ & $0.908$ & $0.955$ & $0.933$ & $\unit[70.25]{\%}$ \\
\bottomrule
\end{tabular}
\end{table}
\clearpage
\begin{figure}[t]
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics{figure8a.pdf}
\subcaption{LeNet 300-100}
\label{fig:lenet-300-100}
\end{subfigure}\hfill
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics{figure8b.pdf}
\subcaption{LeNet-5}
\label{fig:lenet-5}
\end{subfigure}
\caption{Used LeNet networks.}
\label{fig:lenet-schematic}
\end{figure}
\begin{figure}[t]
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics{figure9a.pdf}
\subcaption{ResNet14d}
\label{fig:resnet14d}
\end{subfigure}\hfill
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics{figure9b.pdf}
\subcaption{ResNet20 by \citet{keras_resnet}}
\label{fig:resnet20}
\end{subfigure}
\caption{Used ResNet networks.}
\label{fig:ResNet-architecure}
\end{figure}
The figures~\ref{fig:lenet-schematic} and \ref{fig:ResNet-architecure} show the used networks in detail. The following annotation is used:
\begin{center}
\nntensor{\text{kernel~shape,~strides}}{{\color{cnnDimension}\text{output~shape}}}{{\color{cnnActivation}\text{activation and additional steps}}}{{\color{cnnFeature}\text{features}}}{\text{layer~type}}
\end{center}
The layer types are:
\begin{description}
\item[i] input
\item[d] dense
\item[c2] 2D convolution
\item[mP] max pooling
\item[aP] average pooling
\item[f] flatten
\end{description}
Possible activations and additional steps are:
\begin{description}
\item[bn] batch normalization
\item[relu] ReLu activation
\item[softmax] softmax activation
\item[0.3] 30\% dropout
\end{description}
\subsection{Networks with additional layer types}
\label{sub:cifar10_tests}
To show the behavior of the lightning initialization on more complex networks, it is tested on the CIFAR-10 dataset with two different models. The LeNet5 network consists in the upper layers of convolutions and max pooling. The tail is build by three dense layers, which is similar to the LeNet 300-100 of the pure dense test in section~\ref{sub:mnist_tests}. Figure~\ref{fig:cifar10}\subref{fig:lenet5} shows the comparison of the LeNet5 network solving the CIFAR-10 problem with a classical initialization based on \citet{He2015-2} and the lightning initialization with different parameters for the dense layers. Similar to the results in section~\ref{sub:mnist_tests}, the lightning variant shows a better and faster learning. But the effect is less effective and seems only observable at lower learning rates. On the other hand, this shows that the alternative initialization retains its effect through the epochs.
\begin{figure}[t]
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics{figure7a.pdf}
\subcaption{LeNet5}
\label{fig:lenet5}
\end{subfigure}\hfill
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics{figure7b.pdf}
\subcaption{ResNet}
\label{fig:resnet}
\end{subfigure}
\caption{Comparison of a classical and the lightning initialization for the CIFAR10 problem.}
\label{fig:cifar10}
\end{figure}
Table~\ref{tab:cifar10}\subref{tab:lenet-cifar-10} shows the best accuracies of a little parameter study for the LeNet5 network solving the CIFAR10 problem with lightning initialization for the dense layer. In opposition to the results in section~\ref{sub:mnist_tests} the lightning initialization performs better with more and stronger lightnings. It is predictable that the parameters have borders in both directions. The parameter study in section~\ref{sub:mnist_tests} is located only on the one end. It seems that the parameter for the strength of the LeNet5 are less robust in comparison to the LeNet test in section~\ref{sub:mnist_tests}. But this has to be tested on a larger parameter study.
\begin{table}[h]
\caption{Best accuracies in 200 epochs for lightning initialized dense subnets for the dataset cifar-10 with learning rate scheduler, data preprocessing and data augmentation based on subsection~\ref{sub:cifar_10}.}
\label{tab:cifar10}
\begin{subtable}[t]{0.475\textwidth}
\captionsetup{skip=6pt}
\subcaption{Values for the LeNet 5 network. With a \citet{He2015-2} initialization the network reached an accuracy of $\unit[74.76]{\%}$.}
\label{tab:lenet-cifar-10}
\centering
\begin{tabular}{r|ccc}
\toprule
& \multicolumn{3}{c}{strength} \\
lightnings & 0.1 & 0.25 & 0.5 \\
\toprule
100 & $\unit[69.11]{\%}$ & $\unit[70.57]{\%}$ & $\unit[69.32]{\%}$ \\
\midrule
1000 & $\unit[72.03]{\%}$ & $\unit[74.47]{\%}$ & $\unit[74.39]{\%}$ \\
\midrule
5000 & $\unit[72.76]{\%}$ & $\unit[75.17]{\%}$ & $\unit[74.78]{\%}$ \\
\bottomrule
\end{tabular}
\end{subtable}\hfill
\begin{subtable}[t]{0.475\textwidth}
\subcaption{Values for the ReNet 14d network. With a \citet{He2015-2} initialization the ResNet 8d reached $\unit[87.75]{\%}$, ResNet14d $\unit[90.79]{\%}$ and ResNet20 $\unit[91.54]{\%}$ accuracy.}
\label{tab:resnet-cifar-10}
\centering
\begin{tabular}{r|ccc}
\toprule
& \multicolumn{3}{c}{strength} \\
lightnings & 0.1 & 0.25 & 0.5 \\
\toprule
100 & $\unit[90.55]{\%}$ & $\unit[90.93]{\%}$ & $\unit[90.75]{\%}$ \\
\midrule
1000 & $\unit[90.59]{\%}$ & $\unit[90.92]{\%}$ & $\unit[90.75]{\%}$ \\
\midrule
5000 & $\unit[90.63]{\%}$ & $\unit[90.81]{\%}$ & $\unit[90.33]{\%}$ \\
\bottomrule
\end{tabular}
\end{subtable}
\end{table}
The LeNet5 network is relatively simple. To show that the lightning initialization works as well with deeper networks, it is tested with a ResNet14d network described in section~\ref{sub:cifar_10}. The classical ResNet has only one dense layer at the end, which is not usable with lightning because every edge creates a complete path. Therefore, the ResNet14d network is modified that the end contains three dense layers. Figure~\ref{fig:cifar10}\subref{fig:resnet} shows the learning curve of a ResNet14d with a classical initialization based on \citet{He2015-2}, a lightning initialized variant and the result of the ResNet20 to compare the benefit of lightning against a much more complex ResNet. Equal to the LeNet5 experiment, the lightning initialization improves the learning primarily for lower learning rates.
Table~\ref{tab:cifar10}\subref{tab:resnet-cifar-10} shows the best accuracies for the same parameters as for the LeNet5 study in table~\ref{tab:cifar10}\subref{tab:lenet-cifar-10}. The dependency against the strength parameter seems to be less than for the LeNet5 network. The number of lightnings shows the opposite behavior, because the optimum is below 100 for the ResNet14d tests and above 5000 for the LeNet5. The best result for the ResNet14d was achieved for 90 lightnings with a strength of 0.25 an reached an accuracy of $\unit[91.12]{\%}$, which is shown in figure~\ref{fig:cifar10}\subref{fig:resnet}.
\section{Conclusion}
\label{sec:conclusion}
It has been shown that a dense network can be interpreted as sparse graph. By using a threshold, the network can be transfered into a sparse graph. A sparse graph only consists from the information about an existing connection between two nodes and if this connection is activating or inhibiting. A LeNet 300-100, which solved the MNIST problem, can be interpreted as sparse graph and can be reconstructed only based on this information. More complex networks with additional layers are more difficult to reconstruct, probably because the range of possible solutions is larger and thus the learning process gains influence.
The sparse graph interpretation results in the lightning initialization thus the network is initialized with complete paths. This assists the development of the solution. The experiments show that several network architectures for different problems benefit from the lightning initialization. The learning process is faster and the resulting accuracies are better. Parameter studies demonstrates that the two parameters, number and strength of lightning, are robust against the LeNet 300-100 network solving the MNIST problem. Both networks, which are tested against the CIFAR10 problem, are more sensitive against this parameters. But this needs to be investigated more closely in order to be able to make a clear statement about the robustness against complex networks.
\section{Experimental setup}
\label{sec:experimental_setup}
The concepts are tested on different network topologies and problems. This section describes the used models for different problems and the used optimization parameters. The tests are categorized in two parts. First, networks that only consists of dense layers. Second, networks that only have dense layers at the tail.
\subsection{MNIST}
\label{sub:mnist}
The MNIST dataset (\url{http://yann.lecun.com/exdb/mnist/}) has been chosen to demonstrate the behavior for pure dense networks. To solve the MNIST problem a LeNet 300-100 by \citet{LeCun1998} and a SGD optimizer with a constant learning-rate of 0.05 is used. The batch size is set to 100 and the data is preprocessed that the target range is between 1 and a minimum of 0.
The LeNet 300-100 is build with "relu" activation functions in all layers except the output layer, which is build with "softmax" as activation function. Figure~\ref{fig:lenet-schematic}\subref{fig:lenet-300-100} shows the schematic of the used LeNet 300-100 network.
\subsection{CIFAR-10}
\label{sub:cifar_10}
The CIFAR-10 (\url{https://www.cs.toronto.edu/~kriz/cifar.html}) dataset is used in combination with two different network architectures. One is the LeNet-5 by \citet{LeCun1998} based on the code of \citet{keras_lenet}. Figure~\ref{fig:lenet-schematic}\subref{fig:lenet-5} shows the schematic of this network. This net is solved with the Adam solver with a scheduled learning rate, shown in table~\ref{tab:learningrate}. A batch size of 32, data augmentation and data preprocessing is used. The data is scaled that it has a range of 1 and a mean of 0.
The ResNet is used as second network architecture. Therefor two structures were optimized. The reference is a ResNet20 based on the code of \citet{keras_resnet}. The second structure is the custom ResNet14d, which has a three layer dense network as finish. The schematic of the ResNet20 is shown in figure~\ref{fig:ResNet-architecure}\subref{fig:resnet20} and the ResNet14d schematic in figure~\ref{fig:ResNet-architecure}\subref{fig:resnet14d}. The networks are optimized by an Adam solver with a scheduled learning rate, shown in table~\ref{tab:learningrate}. Batch size, data augmentation and data preprocessing are equal to the LeNet-5 networks.
\begin{table}[h]
\caption{Learning rate}
\label{tab:learningrate}
\centering
\begin{tabular}{cc}
\toprule
Epochs & learning rate \\
\toprule
$1 \ldots 80$ & $1 \cdot 10^{-3}$ \\
\midrule
$81 \ldots 120$ & $1 \cdot 10^{-4}$ \\
\midrule
$121 \ldots 160$ & $1 \cdot 10^{-5}$ \\
\midrule
$161 \ldots 180$ & $1 \cdot 10^{-6}$ \\
\midrule
$> 180$ & $0.5 \cdot 10^{-6}$ \\
\bottomrule
\end{tabular}
\end{table}
\section{Introduction}
\label{sec:introduction}
The development of the neural networks trends to even more complex structures. These deep networks are build by the combination of different types (\cite{Krizhevsky2012}, \cite{Szegedy2014}, \cite{He2015}). Usually the finish is a fully connected layer or a dense network. Several studies show that dense networks can be pruned to significantly sparse nets (\cite{Setiono1996}, \cite{Denil2013}, \cite{Han2015}). In opposite to this, these sparse networks can not be trained successfully. \citet{Frankle2018} recently showed that the difference of extreme pruning and incapacity to train sparse dense networks are initially based on the lottery hypotheses. Shortly, within a big network the probability of a valid subset of weights is so high that a random initialization results in a successful training.
Common initialization is based on normal or uniform random numbers. Different researchers have shown that the variance of weights at each layer should be scaled based on the number of neurons enclosed (\cite{LeCun2012}, \cite{Glorot2010}, \cite{He2015-2}). The weight initialization supports the training through the scaled variance, so that the optimization converges faster. The initialization with random numbers does not consider the information transport between the input and the output. Therefore, we consider that it is possible to improve the initialization for a better learning performance.
\section{Lightning initialization}
\label{sec:lightning_initialization}
The common initializations do not take the sparse graph theory into account. A random initialization produces for increasing deepness only weakly connected subgraphs, similar to figure~\ref{fig:simple_net_c}. The information transport is interrupted. Based on the sparse graph interpretation, an initialization that builds paths between the input and output layer should improve the learning of the network. If paths between input and output exists, the information in the backpropagation should be better distributed to the input near layers.
\begin{figure}[t]
\centering
\includegraphics{figure2.pdf}
\caption{Proportion of edges on complete paths between input and output in relation to the remaining connections for a LeNet 300-100 network with 784 inputs and 10 outputs based on a Monte Carlo simulation with 250 tries.}
\label{fig:sparse_connection}
\end{figure}
Figure~\ref{fig:sparse_connection} shows the result of an experiment that demonstrates the relation between the number of remaining edges in a network and the portion of edges on complete paths between input and output. The different initializations lead to very similar curves with only little noise. The transition zone between no existing path to all edges be on complete paths is with $\approx \unit[0.75]{\%}$ very sharp, in respect to the total number of possible edges. For a LeNet 300-100 with 784 inputs and 10 outputs the strongest 10000 edges do not build a sparse graph that connects the input and output. Figure~\ref{fig:paths_vs_accuracy} shows the relation between the portion of edges that are located on complete paths and the validation accuracy of the unpruned network for different training epochs. The experiment is repeated five times, which is represented in the different node shapes. Figure~\ref{fig:paths_vs_accuracy}\subref{fig:path1000} shows this relation for the 1000 and figure~\ref{fig:paths_vs_accuracy}\subref{fig:path10000} for the 10000 strongest edges. Both, figure~\ref{fig:paths_vs_accuracy}\subref{fig:path1000} and figure~\ref{fig:paths_vs_accuracy}\subref{fig:path10000} show that the solver arranges the sparse graph edges to form complete paths.
\begin{figure}[t]
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics{figure3a.pdf}
\subcaption{Largest 1000 weights.}
\label{fig:path1000}
\end{subfigure}
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics{figure3b.pdf}
\subcaption{Largest 10000 weights.}
\label{fig:path10000}
\end{subfigure}\hfill
\caption{Proportion of edges on complete paths between input and output for a certain amount of weights in relation to the validation accuracy for a LeNet 300-100 network with 784 inputs and 10 outputs for five independent \citet{Glorot2010} initializations. Each point represents an epoch. The network was trained for 30 epochs.}
\label{fig:paths_vs_accuracy}
\end{figure}
The lightning initialization has the goal to build a sparse graph, which consists only of complete paths. For this, a specific number of "lightnings" (complete paths) are randomly chosen. Edges on these paths get non-zero value. This value is symmetrically distributed to positive and negative. The strength of these edges can be equal or noisy. It is possible that an edge is part of more than a single path. In this case the strengths are not summed up. The strengths of edges on multiple paths are equally distributed for edges that are only on a single a path. The used lightning initializations in this paper have all constant strengths and no noise.
\section{Results}
\label{sec:results}
\subsection{Pure dense networks}
\label{sub:mnist_tests}
The LeNet 300-100 network, as described in section~\ref{sub:mnist}, was initialized by different methods and trained for 30 epochs. Figure~\ref{fig:mnist} shows the result of this experiment. After the first epoch all variants reached an accuracy of $> \unit[90]{\%}$. The lightning initialization performs this first training much better. The wrong answer probability of the lightning initialization is $\approx \unit[32]{\%}$ lower in respect to the mean of the other initializations. After 30 epochs the lightning initialized model performs $\approx \unit[13]{\%}$ better then the other initializations. So, the LeNet 300-100 network with lightning initialization learns faster and better compared to the other initializations.
\begin{figure}[t]
\centering
\includegraphics{figure4.pdf}
\caption{Comparison of different initialization for the LeNet 300-100 and the MNIST problem. The truncated initialization has a standard deviation of $0.1$ and the lightning initialization was used with 1000 lightnings with a strength of $0.5$.}
\label{fig:mnist}
\end{figure}
As assumed in section~\ref{sec:lightning_initialization}, the optimizer uses the initialized paths and performs a better learning. Over the training in every epoch all largest 10000 edges are at complete paths between input and output. So, the lightning initialization produces an alternative implementation of the paths and results in a different solution. This is shown in figure~\ref{fig:mnist_cfd} by the cumulative distribution of the absolute weight values separated by layers. The lightning initialization produces two categories of absolute weight values. Especially in the "hidden 2" and "output" layer these two categories are clearly shown. In the "hidden 1" layer the amount of active edges is so small that the active category nearly vanishes by the resolution.
\begin{figure}[t]
\centering
\includegraphics{figure5.pdf}
\caption{Cumulative distribution of the absolute weight values for different initialization for the LeNet 300-100 and the MNIST problem separated by the layers. The truncated initialization has a standard deviation of $0.1$ and the lightning initialization was used with 1000 lightnings with a strength of $0.5$.}
\label{fig:mnist_cfd}
\end{figure}
The other initializations show a different behavior. The most values are nearly uniform distributed, which is represented by the approximately straight lines for the most of the values. Only the biggest weights strive for even lager values. This can be seen in the significant asymptotic behavior against $\unit[100]{\%}$. By these initializations, the weights are uniformly distributed (by the truncated initialization normal distributed). The optimizer keeps this distribution for most values. Only the largest values do not correspond to this distribution. It can be assumed that these are the values that are mainly responsible for the solution.
\begin{figure}[t]
\centering
\includegraphics{figure6.pdf}
\caption{Parameter study of the lightning initialization against the mean of the best accuracy in 100 epochs for five repeats for the LeNet network in different sizes and the MNIST problem. The wrong answer probability values are truncated to $\unit[10]{\%}$ to show the behavior in the relevant range.}
\label{fig:mnist_parameter_study}
\end{figure}
The lightning initialization has two parameters, the amount of lightnings and the strength of them. The parameters are robust against the LeNet network in different sizes for the MNIST problem. Figure~\ref{fig:mnist_parameter_study} shows the best accuracy for various combinations of lightning amounts and strengths. The lightning initialization is robust against its parameters for the LeNet networks and the MNIST problem. Parameters in a range of factor 100 for both parameters produce usable and similar results. The larger networks give an idea that for an optimum lightning strength and number depend reciprocally on each other.
\section{Sparse graph interpretation}
\label{sec:sparse_graph_interpretation}
A fully connected feed forward network with non-equal weights does not transport information on every path. The optimizer distributes the weights over a value range based on the initialization. For common dense nets this is nearly a continuous distribution. Small weights represent a weak connection, because the output of all neurons is in a comparable range. Relatively, small weights do not have a larger impact for the activation of the neuron. We approximate this behavior of small values as a non existent connection.
To interpret the weights as a sparse graph, the weights have to be categorized into inhibiting ($\rightfootline$), inactive ($\nrightarrow$) and activating ($\rightarrow$) connections. By choosing a threshold for the weight magnitude, the connections will be separated in active and inactive. Positive weights are activating and negative weights are inhibiting edges. Figure~\ref{fig:simple_net} shows this on a very simple example. The threshold is chosen that the five strongest edges are active. So, the edge $\mathrm{A}~\nrightarrow~\mathrm{B}$ is inactive in this example.
\begin{figure}[t]
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics{figure0a.pdf}
\subcaption{Dense network with weights}
\label{fig:simple_net_a}
\end{subfigure}
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics{figure0b.pdf}
\subcaption{Interpretation of \subref{fig:simple_net_a} as sparse graph with five remaining edges}
\label{fig:simple_net_b}
\end{subfigure}
\caption{Schematic of converting a dense network in a sparse graph by choosing the five strongest edges. Weak edges are interpreted as a not existing connection ($\mathrm{A}~\nrightarrow~\mathrm{B}$). Positive weights are activating ($\mathrm{A}~\rightarrow~\mathrm{C}$, $\mathrm{C}~\rightarrow~\mathrm{E}$, $\mathrm{D}~\rightarrow~\mathrm{E}$) and negative weights are inhibiting ($\mathrm{A}~\rightfootline~\mathrm{D}$, $\mathrm{B}~\rightfootline~\mathrm{E}$) connections.}
\label{fig:simple_net}
\end{figure}
A result of the sparse graph interpretation is the behavior that a neuron could lose all its inputs, outputs or both. In figure~\ref{fig:simple_net}\subref{fig:simple_net_b} node $\mathrm{B}$ has no input. This neuron produces no changing output and even the output $\mathrm{B}~\rightarrow~\mathrm{E}$ has a big weight, there is no information about the input in A. The neuron $\mathrm{B}$ does not influence the result of the node E, because the output is constant. So, the missing connection $\mathrm{A}~\nrightarrow~\mathrm{B}$ produces the dead path $\mathrm{A}~\nrightarrow~\mathrm{B}~\rightfootline~\mathrm{E}$. These dead paths could build dead areas in bigger networks.
\begin{table}[h]
\caption{Changing rate of active connections for a pruned network based on sparse graph initialization for different remaining network sizes. The origin network reaches a validation accuracy of $\unit[98.12]{\%}$ in 30 epochs for MNIST (setup see section~\ref{sub:mnist}) with truncated random initialization with a standard deviation of $0.01$.}
\label{tab:pruning_test}
\centering
\begin{tabular}{rrrrrr}
\toprule
active net & \multicolumn{4}{c}{changed connections} & accuracy \\
& hidden 0 & hidden 1 & output & over all & \\
\toprule
$\unit[10.0]{\%}$ & $\unit[6.30]{\%}$ & $\unit[3.40]{\%}$ & $\unit[3.39]{\%}$ & $\unit[4.82]{\%}$ & $\unit[97.71]{\%}$ \\
\midrule
$\unit[20.0]{\%}$ & $\unit[0.31]{\%}$ & $\unit[0.27]{\%}$ & $\unit[1.15]{\%}$ & $\unit[0.32]{\%}$ & $\unit[98.01]{\%}$ \\
\midrule
$\unit[30.0]{\%}$ & $\unit[0.00]{\%}$ & $\unit[0.02]{\%}$ & $\unit[1.00]{\%}$ & $\unit[0.02]{\%}$ & $\unit[98.00]{\%}$ \\
\midrule
$\unit[40.0]{\%}$ & $\unit[0.00]{\%}$ & $\unit[0.00]{\%}$ & $\unit[0.22]{\%}$ & $\unit[0.00]{\%}$ & $\unit[97.98]{\%}$ \\
\midrule
$\unit[50.0]{\%}$ & $\unit[0.00]{\%}$ & $\unit[0.00]{\%}$ & $\unit[0.22]{\%}$ & $\unit[0.00]{\%}$ & $\unit[97.98]{\%}$ \\
\bottomrule
\end{tabular}
\end{table}
Networks have to be oversized to be trained. This indicates the fact of pruning, which has been shown for example by \citet{Han2015} and \citet{Frankle2018}. Equivalent to this pruning effect, the sparse graph interpretation of the network is insensitive against the remaining active network size. Table~\ref{tab:pruning_test} shows an example network with different remaining active network sizes and the resulting connection type changes of the active edges. A LeNet~300-100 \cite{LeCun1998} with a uniform initialization by \citet{Glorot2010} is trained for 30 epochs. All other parameters are equal to section~\ref{sub:mnist}. This training reached a maximum accuracy of $\unit[98.12]{\%}$. The value based sparse graph interpretation is used to reinitialize the network. The weight of active edges is set to $0.1$, for inactive edges to $0$ and for inhibiting edges to $-0.1$. Then, this reinitialized network is trained with the same setup as the parent state. The only change is the different initialization and that inactive edges of the parent graph are pruned in the child network.
The achieved maximum accuracy is comparable for all chosen thresholds. The rate of changed connection types decreases to more active edges. If $\unit[80]{\%}$ of the network is inactive the change of all remaining active edges is only $\unit[0.32]{\%}$. $\unit[99.68]{\%}$ of all active connections holds their initial type based on the previous step. With $\unit[90]{\%}$ inactive connections, the network is still trainable and reached in 10 epochs a maximum accuracy of $\unit[97.71]{\%}$. The remaining net is large enough to solve the task, but the structure of the network has to be changed more compared to the parent structure.
\begin{table}[h]
\caption{Pearson correlation coefficient of the active edges between the trained parent and child net for a pruned network based on sparse graph initialization for different remaining network sizes. The origin network reaches a validation accuracy of $\unit[98.12]{\%}$ in 30 epochs for MNIST (setup see section~\ref{sub:mnist}) with a uniform initialization by \citet{Glorot2010}. For all given Pearson correlation coefficient the p-value is $0$ numerically.}
\label{tab:pruning_test_2}
\centering
\begin{tabular}{rllllr}
\toprule
active net & \multicolumn{4}{c}{Pearson correlation coefficient} & accuracy \\
& hidden 0 & hidden 1 & output & over all & \\
\toprule
$\unit[10.0]{\%}$ & $0.968$ & $0.983$ & $0.983$ & $0.976$ & $\unit[97.71]{\%}$ \\
\midrule
$\unit[20.0]{\%}$ & $0.998$ & $0.999$ & $0.994$ & $0.998$ & $\unit[98.01]{\%}$ \\
\midrule
$\unit[30.0]{\%}$ & $1 $ & $1 $ & $0.995$ & $1 $ & $\unit[98.00]{\%}$ \\
\midrule
$\unit[40.0]{\%}$ & $1 $ & $1 $ & $0.999$ & $1 $ & $\unit[97.98]{\%}$ \\
\midrule
$\unit[50.0]{\%}$ & $1 $ & $1 $ & $0.999$ & $1 $ & $\unit[97.98]{\%}$ \\
\bottomrule
\end{tabular}
\end{table}
The Pearson correlation coefficient indicates the linear correlation between two variables. Table~\ref{tab:pruning_test_2} shows the Pearson correlation coefficient between the weights of the parent and the child network. A coefficient near $1$ indicates a direct linear correlation. This shows that not even the structure remains also the weights get similar values through the learning.
The results of table~\ref{tab:pruning_test} and \ref{tab:pruning_test_2} show that the sparse graph structure is a valid approximation for the network. The sparse graph holds the necessary information to rebuild a very similar copy of the parent net.
\begin{figure}[t]
\centering
\includegraphics{figure1.pdf}
\caption{Interpretation of \ref{fig:simple_net}\subref{fig:simple_net_a} as sparse graph with three remaining edges}
\label{fig:simple_net_c}
\end{figure}
|
\section{Introduction}\label{Sec-1}
One of the fundamental and important problems in differential geometry
is to establish structure theorems or classifications
for varieties satisfying certain curvature conditions.
The Frankel conjecture,
which has been solved by Siu-Yau in \cite{SY80} and by Mori in \cite{Mor79},
states that a smooth projective variety with {\textit{positive bisectional curvature}}
is isomorphic to the projective space (see \cite{Mor79} for the Hartshorne conjecture).
As one of the extensions of the Frankel conjecture,
it is a significant problem to consider the geometry
of {\textit{semi-positive bisectional curvature}}
(or more generally the geometry of nef tangent bundles).
In their paper \cite{HSW81},
Howard-Smyth-Wu studied a structure theorem for a compact K\"ahler manifold $M$
with semi-positive bisectional curvature
and they showed that $M$ can be decomposed into
a \lq \lq flat" manifold $B$ and a \lq \lq quasi-positively curved" manifold $M'$
(see also \cite{CG71} and \cite{CG72}).
Precisely speaking, they proved that $M$ admits a locally trivial morphism $f : M \to B$
to a flat K\"ahler manifold $B$ such that
the fiber $M'$ of $f$ is a smooth projective variety with quasi-positive bisectional curvature.
The flat manifold $B$ is a complex torus up to finite \'etale covers,
and thus the geometry of $M$ can be reduced to
the smooth projective variety $M'$ with quasi-positive bisectional curvature,
thanks to their structure theorem.
Further, it can also be proven that
the universal cover of $M$ is isomorphic to
the product $M' \times \mathbb{C}^m$
(see \cite{DPS94} for compact K\"ahler manifolds with nef tangent bundle).
After the work of Howard-Smyth-Wu, in his paper \cite{Mok88},
Mok studied smooth projective varieties with quasi-positive bisectional curvature and
he showed that the fiber $M'$ of $f$ is isomorphic to the product of
projective spaces and compact Hermitian symmetric manifolds
(see \cite{CP91} for the Campana-Peternell conjecture).
This paper is devoted to studies of compact K\"ahler manifolds
whose {\textit{holomorphic sectional curvature}} is semi-positive or quasi-positive,
motivated by generalizing Howard-Smyth-Wu's structure theorem for manifolds with semi-positive bisectional curvature
and Mok's result for the geometry of quasi-positive bisectional curvature.
\vspace{0.2cm}
The first contribution of this paper is
concerned with the solution and its generalization
of the following conjecture posed by Yau in \cite{Yau82},
which gives a relation between the \lq \lq strict" positivity of holomorphic sectional curvature
and the geometry of $X$ (rationally connectedness).
Yau's conjecture can be seen as an analogy of Mok's result
in the studies of holomorphic sectional curvature
and it corresponds to the geometry of the fiber $M'$
appearing in Howard-Smyth-Wu's structure theorem.
\begin{conj}[Yau's conjecture for projective varieties]\label{conj-Yau}
If a smooth projective variety $X$ admits a K\"ahler metric with positive holomorphic sectional curvature,
then $X$ is rationally connected $($that is, two arbitrary points can be connected by a rational curve$)$.
\end{conj}
In their paper \cite{HW15}, Heier-Wong considered Yau's conjecture for projective varieties
under the weaker assumption that the holomorphic sectional curvature is quasi-positive
(that is, it is semi-positive everywhere and positive at some point).
We emphasize that it is essentially important to consider Yau's conjecture for quasi-positive holomorphic sectional curvature
from the viewpoint of structure theorems,
since the bisectional curvature of the fiber $M'$ is quasi-positive, but not necessarily positive everywhere.
Yang affirmatively solved Yau's conjecture even for the case of compact K\"ahler manifolds
by introducing the notion of RC positivity in \cite{Yan17a}
(see \cite{Yan18c} and references therein for recent progress of RC positivity),
but it seems to be quite difficult to apply his method to the case of quasi-positive holomorphic sectional curvature.
In this paper,
we obtain a generalization of Yau's conjecture, Heier-Wong's result, and Yang's result
(see Theorem \ref{thm-mai}),
by using an idea in \cite{HW15} and by developing techniques for a partial positivity and certain flatness.
This theorem can be seen as a version of Mok's result for holomorphic sectional curvature,
and further it gives a more precise relation
between the positivity of holomorphic sectional curvature
and the dimension of images of maximal rationally connected (MRC for short) fibrations
(which measures how far $X$ is from rationally connectedness).
See \cite{Cam92} and \cite{KoMM92} for MRC fibrations.
\begin{theo}\label{thm-mai}
Let $(X, g)$ be a compact K\"ahler manifold such that $X$ is projective and
the holomorphic sectional curvature is semi-positive.
Let $\phi: X \dashrightarrow Y$ be a MRC fibration of $X$.
Then we have
$$
\dim X -\dim Y \geq n_{{\rm{tf}}}{(X, g)}.
$$
In particular, the manifold $X$ is rationally connected
if $n_{{\rm{tf}}}{(X, g)}=\dim X$ $($which is satisfied if the holomorphic sectional curvature is quasi-positive$)$.
\end{theo}
Here the invariant $n_{{\rm{tf}}}{(X, g)}$ is defined by
$$
n_{{\rm{tf}}}{(X, g)}:=\dim X - \inf_{p \in X} \dim V_{{\rm{flat}},p},
$$
where $V_{{\rm{flat}},p}$ is the subspace of the tangent space $T_{X,p}$ at $p$
consisting of all the truly flat tangent vectors (see subsection \ref{subsec-2-2} for the precise definition).
The invariant $n_{{\rm{tf}}}{(X, g)}$ can be seen as an analogue of the numerical Kodaira dimension
in terms of truly flat tangent vectors introduced in \cite{HLWZ}.
The condition of $n_{{\rm{tf}}}{(X, g)}=\dim X$
(that is, there is no truly flat tangent vectors at some point)
is a weaker assumption than the quasi-positivity,
but it works in a more flexible manner from the viewpoint of our argument.
\vspace{0.2cm}
The second contribution of this paper
is a partial answer for the following structure conjecture.
The following conjecture, which is a revised version of \cite[Conjecture 1.1]{Mat18},
asks a structure theorem for compact K\"ahler manifolds
with semi-positive holomorphic sectional curvature and
it can be seen as a natural generalization of Howard-Smyth-Wu's structure theorem.
In this paper, we affirmatively solve Conjecture \ref{conj-str}
under the assumption that a MRC fibration of a smooth projective variety $X$
can be chosen to be a morphism without indeterminacy locus (see Theorem \ref{thm-main}).
This assumption is satisfied when $X$ has the nef anti-canonical bundle
by the deep result of \cite{CH17} (see Corollary \ref{cor-CH}).
Moreover we solve Conjecture \ref{conj-str} for compact K\"ahler surfaces without any assumptions
(see Corollary \ref{cor-main}).
\begin{conj}[{cf. \cite{HSW81} and \cite[Conjecture 1.1]{Mat18}}]\label{conj-str}
Let $X$ be a compact K\"ahler manifold with semi-positive holomorphic sectional curvature.
\vspace{0.2cm}\\
$(1)$
Then there exists a smooth morphism $X \to Y$ with the following properties\,$:$
\begin{itemize}
\item The morphism $X \to Y$ is locally trivial $($that it, all the fibers $F$ are isomorphic$)$.
\item The fiber $F$ is projective and rationally connected.
\item $Y$ is a compact K\"ahler manifold with flat metric.
\end{itemize}
In particular, there exist a complex torus $T$ and a finite \'etale cover $T \to Y$ such that
the fiber product $X^*:=X \times_Y T $ admits
a locally trivial morphism $X^*=X \times_Y T \to T$ to the complex torus $T$
with the rationally connected fiber $F$ and that it satisfies the following commutative diagram\,$:$
\begin{equation*}
\xymatrix@C=40pt@R=30pt{
X^{*}=X \times_Y T \ar[d] \ar[r]^{} & T\ar[d]^{} \\
X \ar[r]^{} & Y.\\
}
\vspace{0.1cm}
\end{equation*}
$(2)$ Moreover we have the decomposition
$$
X_{\rm{univ}} \cong \mathbb{C}^m \times F,
$$
where $X_{\rm{univ}}$ is the universal cover of $X$ and
$F$ is the rationally connect fiber.
In particular, the fundamental group of $X$ is an extension of a finite group by $\mathbb{Z}^{2m}$.
\end{conj}
When we approach to the above conjecture in the case of $X$ being projective,
it seems to be the right direction to study
a MRC fibration
$\phi: X \dashrightarrow Y$ of $X$,
based on the strategy explained in \cite{Mat18}.
We remark that MRC fibrations are almost holomorphic maps
(that is, dominant rational maps with compact general fibers) and
they are uniquely determined up to birational equivalence.
It can be seen that we can always choose a MRC fibration $\phi: X \dashrightarrow Y$
such that the image $Y$ is smooth by taking a resolution of singularities and
that the image $Y$ has the pseudo-effective canonical bundle by \cite[Theorem 1.1]{GHS03} and \cite{BDPP}.
The following theorem, which is one of the main results of this paper,
reveals a detailed structure of morphisms whose domain has semi-positive holomorphic sectional curvature.
Theorem \ref{thm-main} is formulated for MRC fibrations of projective varieties and
Albanese maps of K\"ahler manifolds.
By applying Theorem \ref{thm-main} to MRC fibrations,
we can affirmatively solve Conjecture \ref{conj-str} for compact K\"ahler surfaces
and (1) of Conjecture \ref{conj-str}
in the case where a MRC fibration can be chosen to be a morphism
(see Corollary \ref{cor-main}).
Further, by applying Theorem \ref{thm-main} to Albanese maps,
we can obtain a vanishing theorem for the global holomorphic $1$-forms
(see Corollary \ref{cor-mainnnn}).
This vanishing theorem is an extension of \cite[Theorem 1.7]{Yan17a}.
\begin{theo}\label{thm-main}
Let $(X, g)$ be a compact K\"ahler manifold
with semi-positive holomorphic sectional curvature and
let $Y$ be a compact K\"ahler manifold with pseudo-effective canonical bundle.
Further let $\phi: X \to Y$ be a morphism from $X$ to $Y$.
Then the following statements hold\vspace{0.1cm}\,$:$
\begin{itemize}
\item[(1)] $\phi$ is a smooth morphism $($that is, a submersion$)$. \vspace{0.2cm}
\item[(2)] The standard exact sequence of vector bundles obtained from $(1)$
$$
0 \xrightarrow{\quad \quad} T_{X/Y}:=\Ker d\phi_* \xrightarrow{\quad \quad}
T_X \xrightarrow{\quad d\phi_* \quad}
\phi^{*} T_Y
\xrightarrow{\quad \quad} 0
$$
splits.
Moreover its holomorphic splitting
$$
T_X = T_{X/Y} \oplus \phi^{*} T_Y
$$
coincides with the orthogonal decomposition of $T_{X}$ with respect to $g$.
Here $T_X$ $($resp. $T_Y$$)$ is the $($holomorphic$)$ tangent bundle of $X$
$($resp. $Y$$)$.
\vspace{0.2cm}
\item[(3)] Let $g_Q$ be the hermitian metric on $\phi^{*} T_Y$ induced by
the above exact sequence and the given metric $g$.
Then there exists a hermitian metric $g_Y$ on $T_Y$ with the following properties\,$:$
\vspace{0.1cm}
\begin{itemize}
\item[$\bullet$] $g_Q$ is obtained from the pull-back of $g_Y$ $($namely, $g_Q=\phi^* g_Y$$)$.
\vspace{0.1cm}
\item[$\bullet$] The holomorphic sectional curvature of $(Y, g_Y)$ is identically zero.
In particular, the image $Y$ is flat and it admits a finite \'etale cover $T \to Y$
by a complex torus $T$.
\end{itemize}
\vspace{0.2cm}
\item[(4)] $\phi$ is locally trivial if we further assume that $X$ is projective.
\end{itemize}
\end{theo}
\begin{cor}\label{cor-main}
Let $X$ be a compact K\"ahler manifold with semi-positive holomorphic sectional curvature.
Then the followings hold\,$:$
\vspace{0.2cm}\\
\quad $\bullet$ All the statements of Conjecture \ref{conj-str} hold
in the case of $X$ being a surface.
\vspace{0.1cm}\\
\quad $\bullet$ The statement $(1)$ of Conjecture \ref{conj-str} holds
if $X$ is projective and a MRC fibration of $X$ can be chosen to be a morphism.
\end{cor}
\begin{cor}\label{cor-mainnnn}
Let $(X, g)$ be a compact K\"ahler manifold
with semi-positive holomorphic sectional curvature.
Then we have
$$
h^{0}(X, \Omega_X)\leq \dim X - n_{{\rm{tf}}}{(X, g)}.
$$
In particular, we obtain $h^{0}(X, \Omega_X)=0$
if $n_{{\rm{tf}}}{(X, g)}=\dim X$ $($which is satisfied if the holomorphic sectional curvature is quasi-positive$)$.
\end{cor}
For the proof of Theorem \ref{thm-main},
we will carefully observe the curvature current and its integration of
an induced \lq \lq singular" hermitian metric on $\phi^* K_Y$,
and further we investigate the scalar curvature of the K\"ahler form $g$,
based on the idea in \cite{HW15}.
The main difficulty of Theorem \ref{thm-main} is that the given metric $g$ has no a priori relation with the morphism $\phi$.
To overcome this difficulty,
we will show
that all the tangent vector in the horizontal direction of $\phi:X\to Y$ are truly flat,
which produces a relation (for example, the statements (2) and (3)) between the metric $g$ and the morphism $\phi$.
The key point here is
to construct a suitably chosen orthonormal basis of $T_X$
by using an argument on a partial positivity developed in \cite{Mat18}.
By modifying the above techniques for a general MRC fibration $\phi: X \dashrightarrow Y$
(which is not necessarily a morphism),
we can prove that the numerical dimension of the image $Y$ is equal to zero.
Moreover we can obtain the same conclusions as in (1), (2), (3) of Theorem \ref{thm-main}
over the smooth locus of $\phi$.
\begin{theo}\label{thm-mainn}
Let $(X, g)$ be a compact K\"ahler manifold
with semi-positive holomorphic sectional curvature and
$Y$ be a compact K\"ahler manifold with pseudo-effective canonical bundle.
Let $\phi: X \dashrightarrow Y$ be an almost holomorphic map from $X$ to $Y$.
Then the numerical dimension $\nu(Y)$ of $Y$ is equal to zero.
Moreover, the same conclusions as in $(1)$, $(2)$, $(3)$ of Theorem \ref{thm-main} hold
if we replace $X$ and $Y$ in the statements of Theorem \ref{thm-main}
with $X_1$ and $Y_1$ $($see Theorem \ref{r-thm-mainn} for the precise statement$)$.
Here $X_1$ and $Y_1$ are Zariski open sets such that
$\phi :X_1:=\phi^{-1}(Y_{1}) \to Y_1$ is a morphism.
\end{theo}
In Section \ref{Sec-2}, we will recall some basic results on curvature and truly flat tangent vectors.
In Section \ref{Sec-3}, we will prove all the theorems and corollaries.
In Section \ref{Sec-4}, we will discuss open problems related to the geometry of semi-positive holomorphic sectional curvature.
\subsection*{Acknowledgements}
The author wishes to thank Professor Xiaokui Yang
for stimulating discussions on RC positivity and related open problems,
and he also wishes to thank
Professor Masaaki Murakami for helpful comments on the classifications of surfaces.
He would like to thank the members of
Morningside Center of Mathematics, Chinese Academy of Sciences
for their hospitality during my stay.
He is supported by the Grant-in-Aid
for Young Scientists (A) $\sharp$17H04821 from JSPS.
\section{Preliminaries}\label{Sec-2}
For reader's convenience, we summarize some formulas and properties of
curvature tensors, holomorphic sectional curvature, and truly flat tangent vectors
in this section.
\subsection{Curvature and exact sequences of vector bundles}\label{subsec-2-1}
In this subsection, we recall several formulas of curvature of induced hermitian metrics
and properties of exact sequences of vector bundles.
Let $(E, g)$ be a (holomorphic) vector bundle on a complex manifold $X$
equipped with a (smooth) hermitian metric $g$.
The Chern curvature of $(E, g)$
$$
\sqrt{-1}\Theta_{g}:=\sqrt{-1}\Theta_{g}(E)
\in C^{\infty}(X, \Lambda^{1,1}\otimes \End (E)),
$$
defines the curvature tensor
$$
R_g:=R_{(E,g)} \in
C^{\infty}(X, \Lambda^{1,1}\otimes E^\vee \otimes \bar E^{\vee})
$$
to be
$$
R_g(v, \bar w,e, \bar f):=
\big \langle \sqrt{-1}\Theta_{g}(v, \bar w)(e), f \big \rangle_g
$$
for tangent vectors $v, w \in T_X$ and vectors $e, f \in E$.
We denote the dual vector bundle of $E$ by the notation $E^\vee$ and
the inner product with respect to $g$ by the notation $\langle \bullet, \bullet \rangle_g$
throughout this paper.
The metric $g$ induces the hermitian metric $\Lambda^m g$
on the vector bundle $\Lambda^m E$ of the $m$-th exterior product.
Then it follows that
\begin{align}\label{eq-8}
&\big \langle \sqrt{-1}\Theta_{\Lambda^m g}(v, \bar v)
(e_{1}\wedge e_{2}\wedge \dots \wedge e_m), e_{1}\wedge e_{2}\wedge \dots \wedge e_m
\big \rangle_{\Lambda^m g}\\ \notag
=&\sum_{k=1}^{m}\big \langle \sqrt{-1}\Theta_{g}(v, \bar v)(e_k), e_k \big \rangle_g
\end{align}
for a tangent vector $v \in T_X$ and
vectors $\{e_k\}_{k=1}^{m}$ in $E$ with $\langle e_{i}, e_{j} \rangle_g = \delta_{ij}$
since the curvature $\sqrt{-1}\Theta_{\Lambda^m g}$ associated to $\Lambda^m g$ satisfies that
$$
\sqrt{-1}\Theta_{\Lambda^m g}(v, \bar v)(e_{1}\wedge e_{2}\wedge \dots \wedge e_m)=
\sum_{k=1}^{m} e_{1}\wedge \cdots \wedge e_{k-1}\wedge \sqrt{-1}\Theta_{g}(v, \bar v)(e_{k}) \wedge e_{k+1} \wedge \cdots \wedge e_m.
$$
In particular, the curvature $\sqrt{-1}\Theta_{\det g}$ of the determinant bundle
$\det E:=\Lambda^{{\rm{rk}}E} E$ with the induced metric $\det g:=\Lambda^{{\rm{rk}}E} g$ satisfies that
\begin{align*}
\sqrt{-1}\Theta_{\det g}(v, \bar v)
=&\big \langle \sqrt{-1}\Theta_{\det g}(v, \bar v)
(e_{1}\wedge e_{2}\wedge \dots \wedge e_{{{\rm{rk}}E}} ), e_{1}\wedge e_{2}\wedge \dots \wedge e_{{{\rm{rk}}E}}
\big \rangle_{\det g}\\ \notag
=&\sum_{k=1}^{{\rm{rk}}E}\big \langle \sqrt{-1}\Theta_{g}(v, \bar v)(e_k), e_k \big \rangle_g
\end{align*}
for an orthonormal basis $\{e_k\}_{k=1}^{{\rm{rk}}E}$ of $E$.
For a subbundle $S$ of $E$ and its quotient vector bundle $Q:=E/S$,
we consider the hermitian metric $g_S$ (resp. $g_Q$) on $S$ (resp. $Q$)
induced by the metric $g$ and the exact sequence
$$
0 \xrightarrow{\quad \quad} (S, g_S) \xrightarrow{\quad \quad}
(E, g) \xrightarrow{\quad \quad}(Q, g_Q)
\xrightarrow{\quad \quad} 0.
$$
The quotient bundle $Q$ is isomorphic to the orthogonal complement
$S^{\bot}$ of $S$ in $(E, g)$ as $C^\infty$-vector bundles.
By this isomorphism,
the quotient bundle $Q$ can be identified with the $C^{\infty}$-vector bundle $S^{\bot}$ and
the metric $g_Q$ can be regarded as the hermitian metric on $S^{\bot}$.
Also, it can be proven that there exist smooth sections (which are called fundamental forms)
$$A\in C^{\infty}(X, \Lambda^{1,0}\otimes \Hom(S, S^{\bot}))
\text { and }
B\in C^{\infty}(X, \Lambda^{0,1}\otimes \Hom(S^{\bot}, S))
$$
satisfying that
\begin{align}
\big \langle \sqrt{-1}\Theta_{g}(v, \bar v)(e), e \big \rangle_g +
\big \langle B_{\bar v} (e), B_{\bar v} (e) \big \rangle_{g_S} &=
\big \langle \sqrt{-1}\Theta_{g_Q}(v, \bar v)(e), e \big \rangle_{g_Q}, \label{eq-1} \\
\big \langle \sqrt{-1}\Theta_{g}(v, \bar v)(f), f \big \rangle_g -
\big \langle A_v (f), A_v (f) \big \rangle_{g_Q} &=
\big \langle \sqrt{-1}\Theta_{g_S}(v, \bar v)(f), f \big \rangle_{g_S}, \notag\\
\big \langle A_v (f), e \big \rangle_{g_Q}+\big \langle f, B_{\bar v} (e) \big \rangle_{g_S} &=0 \notag
\end{align}
for a tangent vector $v \in T_X$, a vector $e \in S^{\bot}$, and a vector $f \in S$.
In particular, we have
\begin{align}
\big \langle \sqrt{-1}\Theta_{g}(v, \bar v)(e), e \big \rangle_g &\leq
\big \langle \sqrt{-1}\Theta_{g_Q}(v, \bar v)(e), e \big \rangle_{g_Q}, \label{eq-1a} \\
\big \langle \sqrt{-1}\Theta_{g}(v, \bar v)(f), f \big \rangle_g &\geq
\big \langle \sqrt{-1}\Theta_{g_S}(v, \bar v)(f), f \big \rangle_{g_S}. \label{eq-2a}
\end{align}
Moreover it can be shown that the above exact sequence determines
the holomorphic orthogonal decomposition
$E=S \oplus Q$ (that is, $S^{\bot}$ is a holomorphic vector bundle and it is isomorphic to $Q$)
if and only if $A$ (equivalently $B$) is identically zero.
In the rest of this subsection,
we summarize the notion of singular hermitian metrics on a line bundle $L$ (see \cite{Dem} for more details).
A hermitian metric $h$ on $L$ is said to be
a {\textit{singular hermitian metric}}, if
$\log |e|_h$ is an $L^{1}_{\rm{loc}}$-function for any local frame $e$ of $L$.
Then the curvature current $\sqrt{-1}\Theta_h$ of $(L,h)$ is defined by
\begin{equation*}
\sqrt{-1} \Theta_{h}:=\sqrt{-1} \Theta_{h}(L) :=-\sqrt{-1}\partial \overline{\partial} \log |e|^2_h
\end{equation*}
in the sense of distributions.
The singular hermitian metric $h$ is said to have
neat {\textit{analytic singularities}},
if there exists an ideal sheaf $\mathcal{I} \subset \mathcal{O}_{X}$ such that
the function $ - \log |e|^2_h$ can be locally written as
\begin{equation*}
- \log |e|^2_h = c \log \big(
|f_{1}|^{2} + |f_{2}|^{2} + \cdots + |f_{k}|^{2}\big) +\text{smooth function},
\end{equation*}
where $c $ is a positive real number and
$f_{1}, \dots, f_{k}$ are local generators of $\mathcal{I}$.
We say that $h$ has {\textit{divisorial singularities}}
when the ideal sheaf $\mathcal{I}$ is defined by an effective divisor.
The dual singular hermitian metric $h^{\vee}$
on the dual line bundle $L^\vee$ can be defined to be
$|e^\vee|_{h^{\vee}}:=|e|_{h}^{-1}$ for the dual local frame $e^\vee$.
Further, for a morphism $f: Z \to X$, the singular hermitian metric $f^{*}h$ on the pull-back $f^*L$ can also be defined to be
$|f^* e|_{f^{*}h}:=f^*(| e|_{h})$ for the local frame $f^* e$ of $f^*L$.
Then we have
$$
\sqrt{-1}\Theta_{h}=-\sqrt{-1}\Theta_{h^\vee}=\sqrt{-1}\partial \overline{\partial} \log |e^\vee|^2_{h^{\vee}} \text{\quad and \quad }
f^* \sqrt{-1}\Theta_{ h}:=\sqrt{-1}\Theta_{f^* h}.
$$
\subsection{Holomorphic sectional curvature and truly flat tangent vectors}\label{subsec-2-2}
In this subsection, we summarize some properties of holomorphic sectional curvature and truly flat tangent vectors.
For a hermitian metric $g$ on the (holomorphic) tangent bundle $T_X$,
the holomorphic sectional curvature $H_g$ is defined to be
$$
H_g([v]):=\frac{\Rur{g}{v}{v}}{|v|_g^{4}}=\frac{\big \langle \sqrt{-1}\Theta_{g}(v, \bar v)(v), v \big \rangle_g}{|v|_g^{4}}
$$
for a non-zero tangent vector $v \in T_X$.
The holomorphic sectional curvature $H_g$ is said to be {\textit{positive}}
(resp. {\textit{semi-positive}})
if $H_{g}([v]) > 0$ (resp. $H_{g}([v]) \geq 0$) holds
for any non-zero tangent vector $v \in T_X$.
Also $H_g$ is said to be {\textit{quasi-positive}}
if it is semi-positive everywhere and it is positive at some point in $X$.
In this paper, we handle only the case of $g$ being a K\"ahler metric
(that is, the associated $(1,1)$-form $\omega_g$ is $d$-closed).
In this case, the following symmetries hold\,$:$
$$
R_{g}(e_{i}, \bar e_{j}, e_{k}, \bar e_{\ell})
=R_{g}(e_{k}, \bar e_{\ell}, e_{i}, \bar e_{j})
=R_{g}(e_{k}, \bar e_{j}, e_{i}, \bar e_{\ell}).
$$
The above symmetries lead to the following lemmas.
\begin{lemm}[{\cite[Lemma 4.1]{Yan17c}, \cite[Lemma 2.2]{Mat18} cf. \cite{Bre}, \cite{BKT13}}]\label{lem-ineq}
Let $g$ be a K\"ahler metric of $X$ and $V$ be a subspace of $T_{X,p}$ at a point $p \in X$.
If a unit vector $v \in V$ is a minimizer of
the holomorphic sectional curvature $H_g$ on $V$,
that it, it satisfies
$$
H_g([v])=\min\{H_g([x])\, |\, {0 \not = x \in V}\},
$$
then we have
$$
2\Rur{g}{v}{x} \geq (1+|\langle v, x \rangle_g|^2) \Rur{g}{v}{v}
$$
for any unit vector $x \in V$.
In particular, if the holomorphic sectional curvature $H_g$ is semi-positive,
a minimizer $v$ of $H_g$ on $V$ satisfies that
$$
\Rur{g}{v}{x} \geq 0
$$
for any tangent vector $x \in V$.
\end{lemm}
The above lemma was proved in \cite[Lemma 4.1]{Yan17c}
when the subspace $V$ in Lemma \ref{lem-ineq} coincides with the tangent space $T_{X,p}$.
It is easy to see that the same argument as in \cite[Lemma 4.1]{Yan17c}
works even in the case of $V$ being a subspace of $T_{X,p}$,
and thus we omit the proof of Lemma \ref{lem-ineq}.
Note that we can always take the minimizer of $H_{g}$
on a given subspace $V$ of $T_{X,p}$ at a point $p \in X$,
since the holomorphic sectional curvature can be regarded as a smooth function
on the projective space bundle $\mathbb{P}(T_X ^\vee)$
(that is, the set of all complex lines $[v]$ in $T_X$) and
$\mathbb{P}(V^\vee) \subset \mathbb{P}(T_{X,p} ^\vee)$ is compact.
Now we define truly flat tangent vectors and
the invariant $n_{{\rm{tf}}}{(X, g)}$ introduced in \cite{HLWZ}.
We remark that the invariant $n_{{\rm{tf}}}{(X, g)}$
was denoted by the different notation $r_{{\rm{tf}}}^-$ in \cite{HLWZ}.
\begin{defi}[Truly flat tangent vectors and the invariant $n_{{\rm{tf}}}{(X, g)}$]\label{def-flat}
Let $(X, g)$ be a K\"ahler manifold.
\ \\
\quad $\bullet$ A tangent vector $v \in T_X$ at $p$ is said to be {\textit{truly flat}} with respect to $g$
if $v$ satisfies that
$$\text{
$R_{g}(v, \bar x, y,\bar z)=0$
for any tangent vectors $x, y, z \in T_{X,p}$.
}
$$
\quad $\bullet$ We define the subspace $V_{{\rm{flat}},p}$ of $T_{X,p}$ at $p$ by
$$
V_{{\rm{flat}},p}:=\{v \in T_{X,p} \, |\, v \text{ is a truly flat tangent vector in }T_{X,p} \}.
$$
\quad $\bullet$ We define the invariants $n_{{\rm{tf}}}{(X, g)}_p$ and $n_{{\rm{tf}}}{(X, g)}$ by
$$
n_{{\rm{tf}}}{(X, g)}_p:=\dim X - \dim V_{{\rm{flat}},p} \text{\quad and \quad}
n_{{\rm{tf}}}{(X, g)}:=\dim X - \inf_{p \in X} \dim V_{{\rm{flat}},p}.
$$
\end{defi}
It is easy to see that the invariant $n_{{\rm{tf}}}{(X, g)}_p$ is
lower semi-continuous with respect to $p \in X$ in the classical topology.
In particular, if we have the equality $n_{{\rm{tf}}}{(X, g)}_p=n_{{\rm{tf}}}{(X, g)}$ at $p$,
the same equality holds on a neighborhood of $p$.
The following lemma gives a characterization of truly flat tangent vectors
in terms of holomorphic sectional curvature and bisectional curvature.
\begin{lemm}[cf. {\cite{HLWZ}}]\label{lemm-flat}
Let $g$ be a K\"ahler metric of $X$ with semi-positive holomorphic sectional curvature
and $V$ be a subspace of $T_{X,p}$ at a point $p \in X$.
If a tangent vector $v \in T_X$ satisfies that
$$
H_g([v])=0 \quad \text{ and } \quad \Rur{g}{v}{w}=0 \text{ for any tangent vector $w \in V$},
$$
then $v$ satisfies that
$$
R_{g}(v, \bar x, y, \bar z) =0\text{ for any tangent vectors } x,y,z \in V.
$$
In particular, if $v$ satisfies the above assumptions for any tangent vector $w \in T_{X,p}$,
then $v$ is a truly flat tangent vector at $p$.
\end{lemm}
\begin{proof}
When the holomorphic sectional curvature is semi-negative and the subspace $V$ coincides with
the tangent space $T_{X,p}$,
the same conclusion was proved in \cite[Lemma 2.1]{HLWZ}.
For reader's convenience, we will give a sketch of the proof.
For an arbitrary complex number $re^{\sqrt{-1}\theta}$,
we obtain that
\begin{align*}
0 &\leq H([v+re^{\sqrt{-1}\theta}w])|v+re^{\sqrt{-1}\theta}w|_g^4\\
&=2 \Re(e^{\sqrt{-1}\theta}R_g(v,\bar v,v,\bar w) ) r^3
+ 2 \Re(e^{\sqrt{-1}\theta}R_g(v,\bar w,w,\bar w) ) r + R_g(w,\bar w,w,\bar w)
\end{align*}
from the assumptions $\Rur{g}{v}{v}=0$ and $\Rur{g}{v}{w}=0$.
Here we used the symmetries obtained from K\"ahler metrics.
If $R_g(v,\bar v,v,\bar w)$ is not zero,
we have a contradiction by suitably choosing $\theta$ such that
$\Re(e^{\sqrt{-1}\theta}R_g(v,\bar v,v,\bar w) ) <0$ and by taking a sufficiently large $r>0$.
Hence we obtain $R_g(v,\bar v,v,\bar w)=0$.
By repeating the same argument as above for $e^{\sqrt{-1}\theta}R_g(v,\bar w,w,\bar w) $,
we can see that $R_g(v,\bar w,w,\bar w)=0$ for any tangent vector $w \in T_X$.
Then we can easily check the desired equality by the standard polarization argument.
\end{proof}
\section{Proof of the results}\label{Sec-3}
\subsection{Proof of Theorem \ref{thm-main}}\label{subsec-3-1}
In this subsection, we give a proof of Theorem \ref{thm-main}.
The arguments in this subsection will be modified
to handle almost holomorphic maps in the proof of Theorem \ref{thm-mainn}.
This subsection is the core of this paper.
\begin{theo}[=Theorem \ref{thm-main}]\label{r-thm-main}
Let $(X, g)$ be a compact K\"ahler manifold
with semi-positive holomorphic sectional curvature and
let $Y$ be a compact K\"ahler manifold with pseudo-effective canonical bundle.
Further let $\phi: X \to Y$ be a morphism from $X$ to $Y$.
Then the following statements hold\vspace{0.1cm}\,$:$
\begin{itemize}
\item[(1)] $\phi$ is a smooth morphism $($that is, a submersion$)$. \vspace{0.2cm}
\item[(2)] The standard exact sequence of vector bundles obtained from $(1)$
$$
0 \xrightarrow{\quad \quad} T_{X/Y}:=\Ker d\phi_* \xrightarrow{\quad \quad}
T_X \xrightarrow{\quad d\phi_* \quad}
\phi^{*} T_Y
\xrightarrow{\quad \quad} 0
$$
splits.
Moreover its holomorphic splitting
$$
T_X = T_{X/Y} \oplus \phi^{*} T_Y
$$
coincides with the orthogonal decomposition of $T_{X}$ with respect to $g$.
Here $T_X$ $($resp. $T_Y$$)$ is the $($holomorphic$)$ tangent bundle of $X$
$($resp. $Y$$)$.
\vspace{0.2cm}
\item[(3)] Let $g_Q$ be the hermitian metric on $\phi^{*} T_Y$ induced by
the above exact sequence and the given metric $g$.
Then there exists a hermitian metric $g_Y$ on $T_Y$ with the following properties\,$:$
\vspace{0.1cm}
\begin{itemize}
\item[$\bullet$] $g_Q$ is obtained from the pull-back of $g_Y$ $($namely, $g_Q=\phi^* g_Y$$)$.
\vspace{0.1cm}
\item[$\bullet$] The holomorphic sectional curvature of $(Y, g_Y)$ is identically zero.
In particular, the image $Y$ is flat and it admits a finite \'etale cover $T \to Y$
by a complex torus $T$.
\end{itemize}
\vspace{0.2cm}
\item[(4)] $\phi$ is locally trivial if we further assume that $X$ is projective.
\end{itemize}
\end{theo}
\begin{proof}
Throughout this proof,
let $(X, g)$ be a compact K\"ahler manifold
with the semi-positive holomorphic sectional curvature $H_g$ and
let $\phi: X \to Y$ be a morphism (that is, a surjective holomorphic map)
to a compact K\"ahler manifold $Y$ with the pseudo-effective canonical bundle $K_Y$.
For simplicity, we put $n:=\dim X$ and $m:= \dim Y$.
We will divide the proof into five steps to refer later.
\begin{step}[Singularities of induced singular hermitian metrics]\label{step1}
Our first purpose is to prove that $\phi$ is actually a smooth morphism.
For this purpose, in this step, we first construct a possibly \lq \lq singular" hermitian metric $G$ on the line bundle $\phi^* K_Y^{\vee}$
from the given K\"ahler metric $g$ of $X$ such that the singularities of $G$ corresponds to the non-smooth locus of $\phi$.
This enables us to reduce our first purpose to observe the singularities of $G$.
Moreover, in this step, we show that the pull-back $\pi^* G$ has divisorial singularities and
its curvature current can be decomposed into a smooth part and a divisorial part,
after we take a suitable modification $\pi : \bar X \to X$.
Now we have the injective sheaf morphism
\begin{equation*}
(\phi^{*} K_Y, H)
\xrightarrow{\quad d\phi^* \quad} (\Lambda^m \Omega_X, \Lambda^m h)
\end{equation*}
between the vector bundle $\Lambda^m \Omega_X:=\Lambda^m T_X^\vee$ of the $m$-th exterior product
and the line bundle $\phi^{*} K_Y:=\phi^{*} \Lambda^m \Omega_Y$.
We interchangeably use the words \lq \lq line bundles" and \lq \lq invertible sheaves"
(also \lq \lq vector bundles" and \lq \lq locally free sheaves") throughout this paper.
Note that the above morphism is not a bundle morphism
since the rank drops on the non-smooth locus of $\phi$,
but it is an injective morphism as sheaf morphisms.
Let $h$ be the dual hermitian metric of $g$
on the (holomorphic) cotangent bundle $\Omega_X=T_X^{\vee}$ and
let $\Lambda^m h$ be the induced metric on $\Lambda^m \Omega_X$.
Then, from the above morphism, we can construct a possibly singular hermitian metric $H$ on $\phi^{*} K_Y$ to be
$$
|e|_H:=|d\phi^* (e)|_{\Lambda^m h} \text{ for a local frame $e$ of $\phi^{*} K_Y$}.
$$
From now on, we mainly consider the dual singular hermitian metric $G:=H^{\vee}=H^{-1}$ on $\phi^{*} K_Y^{^\vee}$.
For a local coordinate $(t_1, t_2, \dots, t_m)$ of $Y$,
the $m$-form $dt:=dt_1\wedge dt_2 \wedge \cdots \wedge dt_m$ naturally determines
the local frame of $\phi^{*} K_Y$, which we denote by the same notation $dt$.
By the definitions of the curvature and the dual metric,
the curvature current of $(\phi^{*} K_Y^{^\vee}, G)$ can be locally written as
$$
\sqrt{-1}\Theta_G:=\sqrt{-1}\Theta_G(\phi^{*} K_Y^{^\vee})=
\sqrt{-1}\partial \overline{\partial} \log |\phi^* dt|^2_{\Lambda^m h},
$$
where $\phi^* dt$ is the pull-back of the $m$-form $dt$ by $\phi$.
We remark that the pull-back $\phi^* dt$ coincides with
the image $d\phi^* (dt)$ of the section $dt$ of $\phi^*K_Y$ by $d\phi^{*}$.
By the above expression,
it can be shown that the singular locus of $G$
(that is, the polar set of the quasi-psh function $\log |\phi^* dt|_{\Lambda^m h}$)
coincides with the non-smooth locus of $\phi$,
since the zero locus of the section $\phi^* dt$ of
$\Lambda^m \Omega_X$ is equal to the non-smooth locus of $\phi$.
Therefore it is sufficient for our first purpose (that is, the proof of the smoothness of $\phi$) to prove
that $G$ is actually a smooth hermitian metric.
We take a (log) resolution $\pi: \bar X \to X$ of the degenerate ideal $\mathcal{I}$ of the above sheaf morphism.
The degenerate ideal $\mathcal{I}$ is the ideal sheaf
generated by the coefficients of $\phi^* dt$ with respect to local frames of $\Lambda^m \Omega_X$.
Then we obtain the following claim\,$:$
\begin{claim}\label{claim-div}
Let $Z$ be the non-smooth locus of $\phi$.
Then the following statements hold\,$:$
\ \vspace{0.1cm}\\
\quad $\bullet$ $ \pi^{-1}(Z) $ has codimension one.
\vspace{0.1cm}\\
\quad $\bullet$ $\pi: \bar X \setminus \pi^{-1}(Z) \cong X \setminus Z $.
\vspace{0.1cm}\\
\quad $\bullet$ $\phi^* G$ has divisorial singularities along $\pi^{-1}(Z) $.
\vspace{0.1cm}\\
More precisely, the pull-back $\pi^{*}\sqrt{-1}\Theta_G$ of
the curvature current $\sqrt{-1}\Theta_G$ can be written as
\begin{align*}
\pi^{*}\sqrt{-1}\Theta_G:=\sqrt{-1}\partial \overline{\partial} \log \pi^{*} (|\phi^* dt|^2_{\Lambda^m h})=\gamma + [E],
\end{align*}
where $\gamma$ is a smooth $(1,1)$-form on $\bar X$ and $[E]$ is the integration current
defined by an effective divisor $E$.
\end{claim}
\begin{proof}[Proof of Claim \ref{claim-div}]
The subvariety $Z$ coincides with the support of the cokernel $\mathcal{O}_X/\mathcal{I}$,
and thus the first and second statements are obvious by the choice of $\pi$.
However the third statement seems to be a subtle problem,
since we do not know whether or not the metric $G$ itself has neat analytic singularities
(see Remark \ref{rem-div} for more details).
To check the third statement, we fix an arbitrary point $p \in \bar X$.
When $p$ is outside $\pi^{-1}(Z)$,
the metric $G$ is smooth on a neighborhood of $\pi(p)$
since $Z$ also coincides with the zero locus of $\phi^* dt$.
The third statement is obvious in this case,
and thus we may assume that $p \in \pi^{-1}(Z)$.
We take a local frame $\{s_{i}\}_{i=1}^{N}$ of $\Lambda^m \Omega_X$ on a neighborhood of $\pi(p)$.
Here we put $N:=\binom{n}{m}$ for simplicity.
The holomorphic $m$-form $\phi^* dt$ can be locally written as
$$\phi^* dt=\sum_{i=1}^{N} f_i s_i \text{ on a neighborhood of $\pi(p)$ }$$
for some holomorphic functions $\{f_{i}\}_{i=1}^{N}$.
The degenerate ideal $\mathcal{I}$ is generated by $\{f_i\}_{i=1}^{N}$ and
$\pi^{-1}\mathcal{I}=\mathcal{I} \cdot \mathcal{O}_{\bar X}$ is
the ideal sheaf associated to an effective divisor $E$.
Let $t$ be a local holomorphic function such that $t$ determines the effective divisor $E$.
Then it follows that $g_i:=\pi^{*} f_{i}/t$ is a holomorphic function and the common zero locus $\cap_{i=1}^N g_i^{-1}(0)$ is empty
from the choice of $\pi$.
Therefore a simple computation yields
$$
\log \pi^{*} (|\phi^* dt|^2_{\Lambda^m h})=\log |t|^2+\log \sum_{i, j=1}^{N} g_{i}
\bar{g_{j}}\pi^{*}\langle s_i, s_j \rangle_{\Lambda^m h}.
$$
It can be proven that the Levi form of the first term is equal to the integration current $[E]$
by the Poincar\'e-Lelong formula.
On the other hand, it follows that the Levi form of the second term determines a smooth $(1,1)$-form $\gamma$,
since it is easy to see that the function
$$
\sum_{i, j=1}^{N} g_{i} \bar{g_{j}}\pi^{*}\langle s_i, s_j \rangle_{\Lambda^m h}
$$
is a non-vanishing smooth function.
\end{proof}
\begin{rem}\label{rem-div}
$\bullet$ It follows that the smooth form $\gamma$ can be identified with the curvature $\sqrt{-1}\Theta_G$
under the isomorphism $\pi: \bar X \setminus \pi^{-1}(Z) \cong X \setminus Z$ from the second and third properties.
\vspace{0.1cm}\\
$\bullet$ The metric $G$ itself may not have neat analytic singularities
although the pull-back $\pi^* G$ by $\pi$ has divisorial singularities.
For example, in the case of $n=2$ and $m=1$,
we consider the following situation\,$:$
\begin{align*}
\phi^{*} dt=z_1 s_1 +z_2 s_2 \text{ and } h=\left[
\begin{array}{cc}
2 & 1 \\
1 & 2 \\
\end{array}
\right]
\text{with respect to a local frame $(s_1, s_2)$ of $\Omega_X$}.
\end{align*}
Here $(z_1, z_2)$ is a local coordinate of $X$.
Then we can see that the function
$$
\frac{|\phi^{*} dt|^2_h}{|z_1|^2+|z_2|^2}=
\frac{2|z_1|^2+\bar z_1 z_2 + z_1 \bar z_2 + 2|z_2|^2}{|z_1|^2+|z_2|^2}
$$
can not be extended to a smooth function defined at the origin.
Of course, when we take a resolution of the degenerate ideal
(which is just one point blow-up in this case),
we can easily check that the pull-back of the above function is a non-vanishing smooth function.
\end{rem}
\end{step}
\begin{step}[Construction of orthonormal basis in the horizontal direction]\label{step2}
In this step, by using the argument in \cite[Lemma 3.5]{Mat18},
we will choose a suitable orthonormal basis of $T_{X}$
at a smooth point $p$ of $\phi$,
in order to obtain a partial positivity of $\sqrt{-1}\Theta_{G}$ and $\gamma$ in the horizontal direction.
For a given point $p \in X$ at which $\phi$ is smooth,
we consider
the standard exact sequence
\begin{equation*}\label{eq-stan}
0 \xrightarrow{\quad \quad} T_{X/Y}:=\Ker d\phi_* \xrightarrow{\quad \quad}
T_X \xrightarrow{\quad d\phi_* \quad}
\phi^{*} T_Y
\xrightarrow{\quad \quad} 0 \text{ at } p.
\end{equation*}
In the proof, we say that a tangent vector $v \in T_X$ is in the {\textit{horizontal direction}} (resp. in the {\textit{vertical direction}})
in the case of $v \in (T_{X/Y})^{\bot} $ (resp. $v\in T_{X/Y}$).
Here $ (T_{X/Y})^{\bot}$ is the orthogonal complement of $T_{X/Y}$ in $T_X$ with respect to $g$
and it is identified with $\phi^{*}T_Y$ at $p$.
Then we obtain the following claim\,$:$
\begin{claim}\label{claim-semi}
For a smooth point $p$ of $\phi$,
there exists an orthonormal basis $\{e_k \}_{k=1}^{n}$ of $T_X$ at $p$ with the following properties\,$:$
\ \vspace{0.1cm}\\
\quad $\bullet$ $\{e_i \}_{i=1}^{m}$ is an orthonormal basis of $(T_{X/Y})^{\bot}$ at $p$.
\vspace{0.1cm}\\
\quad $\bullet$
$\Rur{g}{e_i}{e_j} \geq 0$ for any $1 \leq i, j \leq m$.
\vspace{0.1cm}\\
\quad $\bullet$ $\sqrt{-1}\Theta_{G}(e_i, \bar e_i) \geq 0 \text{ for any $i=1,2,\dots, m$. }$
\end{claim}
\begin{proof}[Proof of Claim \ref{claim-semi}]
We first take an arbitrary orthonormal basis $\{e_k \}_{k=1}^{n}$ of $T_X$ at $p$
such that
$$
(T_{X/Y})^{\bot}=\Spn\langle \{e_i \}_{i=1}^{m} \rangle
\quad \text{ and } \quad
T_{X/Y} = \Spn\langle \{e_j \}_{j=m+1}^{n} \rangle.
$$
By choosing a new orthonormal basis $\{e_i \}_{i=1}^{m}$ of $(T_{X/Y})^{\bot}$,
we may assume that $e_{1}$ is the minimizer of $H_{g}$ on
$(T_{X/Y})^{\bot}=\Spn\langle \{e_k \}_{k=1}^{m} \rangle$,
that is, the unit tangent vector $e_{1}$ satisfies that
$$
H_g([e_1])=\min\{ H_g([v])\, |\, 0 \not = v \in \Spn \langle \{e_k \}_{k=1}^{m} \rangle \}.
$$
After we fix the tangent vector $e_1$ chosen as above,
we choose an orthonormal basis $\{e_i \}_{i=2}^{m}$ of
$$
(T_{X/Y} \oplus \Spn \langle e_1 \rangle )^{\bot}=\Spn\langle \{e_k \}_{k=2}^{m} \rangle
$$ such that
$e_{2}$ is the minimizer of $H_{g}$ on $\Spn\langle \{e_k \}_{k=2}^{m} \rangle$.
By repeating this process,
we can construct an orthonormal basis $\{e_i \}_{i=1}^{m}$ of $(T_{X/Y})^{\bot}$
satisfying that
\begin{align*}
H_g([e_i])=\min\{H_g([v])\, |\, 0 \not = v \in \Spn \langle \{e_k \}_{k=i}^{m} \rangle\}.
\end{align*}
for any $i = 1,2,\dots, m$.
Then, for this orthonormal basis, we can prove that
$$\text{
$\Rur{g}{e_i}{e_j} \geq 0$ for any $1 \leq i, j \leq m$.
}
$$
Indeed, we may assume that $ i \leq j $
by the symmetry
$$
\Rur{g}{e_i}{e_j}=\Rur{g}{e_j}{e_i}.
$$
Further, for $ i \leq j $,
the tangent vector $e_i$ is the minimizer of $H_g$ on the subspace $\Spn \langle \{e_k \}_{k=i}^{m} \rangle$ which contains $e_j$.
Therefore it follows that $\Rur{g}{e_i}{e_j}$ is non-negative
from Lemma \ref{lem-ineq}.
By applying the formulas (\ref{eq-8}) and (\ref{eq-1a}) to
the exact sequence
\begin{equation*}\label{eq-surj}
0 \xrightarrow{\quad \quad} \Ker d\phi_*
\xrightarrow{\quad \quad}
(\Lambda^m T_X, \Lambda^m g)
\xrightarrow{\quad d\phi_* \quad} (\phi^{*} K_Y^{\vee}, G) \xrightarrow{\quad \quad} 0
\end{equation*}
on a neighborhood of $p$,
we obtain that
\begin{align}\label{eq-4}
\sum_{k=1}^{m}\Rur{g}{v}{e_k}&= \big \langle \sqrt{-1}\Theta_{\Lambda^m g}(v, \bar v)
(e_1\wedge e_2 \wedge \dots \wedge e_m), e_1\wedge e_2 \wedge \dots \wedge e_m
\big \rangle_{\Lambda^m g} \\ \notag
&\leq \big \langle \sqrt{-1}\Theta_{G}(v, \bar v)
(e_1\wedge e_2 \wedge \dots \wedge e_m), e_1\wedge e_2 \wedge \dots \wedge e_m
\big \rangle_{G}
\\
&= \sqrt{-1}\Theta_{G}(v, \bar v) |e_1\wedge e_2 \wedge \dots \wedge e_m|^2_{G}
\notag \\&= \sqrt{-1}\Theta_{G}(v, \bar v) \notag
\end{align}
for any tangent vector $v \in T_X$.
Note that $G$ (defined by the dual metric of $H$) is equal to
the quotient metric induced by $\Lambda^m g$ since $p$ is a smooth point of $\phi$.
When we consider the above formula in the case of $v=e_{i}$,
we can see that the left hand side is non-negative by the second statement in Claim \ref{claim-semi}.
Therefore we can conclude that
$\sqrt{-1}\Theta_{G}(e_i, \bar e_i)$ is non-negative
for any $i=1,2,\dots, m$.
\end{proof}
\begin{rem}\label{rem-semi}
At the end of the proof, we can conclude
that all the tangent vectors in the horizontal direction are actually truly flat and that
the curvature $\sqrt{-1}\Theta_{G}$ is flat,
but a further argument is needed for these conclusions.
\end{rem}
\end{step}
\begin{step}[Positivity of scalar curvature and its integration]\label{step3}
In this step, we will consider the scalar curvature of $g$ and its integration,
based on the idea in \cite{HW15}.
Let $\omega$ be the K\"ahler form associated to the K\"ahler metric $g$.
The first Chern class of $\pi^*\phi^{*} K^\vee_Y$ can be represented
by the curvature current $\pi^{*} \sqrt{-1}\Theta_{G}/2\pi$.
Hence, by taking the wedge product of the equality in Claim \ref{claim-div}
with $ \pi^* \omega^{n-1}$ and by considering the integration over $\bar X$,
we obtain
\begin{align}\label{eq-5}
2 \pi \int_{\bar X} c_1(\pi^*\phi^{*} K^\vee_Y) \wedge \pi^* \omega^{n-1}
=\int_{\bar X} \gamma \wedge \pi^* \omega^{n-1} + \int_E \pi^* \omega^{n-1}.
\end{align}
The purpose of this step is to prove that the first term of the right hand side is non-negative.
If it is proven, all the terms can be shown to be zero.
Indeed, the left hand side is non-positive since $K_Y$ is pseudo-effective by the assumption and
the second term of the right hand side is non-negative.
We will show that this observation leads to a certain flatness of $K_Y$ and the smoothness of $\phi$ in Step \ref{step4}.
We first decompose the first term into the vertical part and the horizontal part.
The integration of $\gamma \wedge \pi^* \omega^{n-1}$ on $\bar X$ is equal to the integration on a Zariski open set
since $\gamma$ and $\pi^* \omega$ are smooth differential forms.
Further $\bar X \setminus \pi^{-1}(Z)$ is isomorphic to $X \setminus Z$ by the morphism $\pi$
and the equality $\gamma=\pi^{*}\sqrt{-1}\Theta_G$ holds on the Zariski open set $\bar X \setminus \pi^{-1}(Z) $
(cf. Remark \ref{rem-div}).
Therefore we can obtain that
\begin{align*}
\int_{\bar X} \gamma \wedge \pi^* \omega^{n-1}&=
\int_{\bar X \setminus \pi^{-1}(Z)} \gamma \wedge \pi^* \omega^{n-1}\\
&=\int_{\bar X \setminus \pi^{-1}(Z)} \pi^{*}(\sqrt{-1}\Theta_G \wedge \omega^{n-1})\\
&=\int_{X \setminus Z} \sqrt{-1}\Theta_G \wedge \omega^{n-1}\\
&=\int_{X_0} \sqrt{-1}\Theta_G \wedge \omega^{n-1}.
\end{align*}
Here $X_0$ is the Zariski open set defined by $X_0:=\phi^{-1}(Y_0)$ and
$Y_{0}$ is the maximal Zariski open set of $Y$ such that
the restriction $\phi : X_{0}=\phi^{-1}(Y_{0}) \to Y_{0}$ is a smooth morphism over $Y_0$.
On the other hand, for a given point $p \in X_{0}$,
we take an orthonormal basis $\{e_{k}\}_{k=1}^n$
of $T_X$ at $p$ satisfying the properties in Claim \ref{claim-semi}.
Then we have $\omega= (\sqrt{-1}/2)\sum_{k=1}^{n}e_k^{\vee} \wedge \bar e_k^{\vee}$ at $p$,
and thus we obtain
\begin{align}\label{eq-7}
&\frac{n}{2}\int_{X_0} \sqrt{-1}\Theta_G \wedge \omega^{n-1}\\
=&\int_{X_0} \sum_{i=1}^{m} \sqrt{-1}\Theta_{G}(e_i, \bar e_i) \, \omega^{n} +
\int_{X_0} \sum_{j=m+1}^{n} \sqrt{-1}\Theta_{G}(e_j, \bar e_j)\, \omega^{n} \notag
\end{align}
from straightforward computations of the scalar curvature.
The integrand of the first term (which measures positivity of the scalar curvature in the horizontal direction) is non-negative by Claim \ref{claim-semi}.
We will show that the second term (that is, the vertical part)
is also non-negative
by using Stokes's theorem and Fubini's theorem (see Claim \ref{claim-key}).
Note that the integrand of the second term can be shown to be non-negative
later (cf. Remark \ref{rem-semi}).
However it seems to be quite difficult to directly check this fact.
For this reason, we need to handle the integration instead of the integrand.
\begin{claim}\label{claim-key}
Under the above situation, the second term is non-negative, namely, we have
$$\int_{X_0} \sum_{j=m+1}^{n} \sqrt{-1}\Theta_{G}(e_j, \bar e_j)\, \omega^{n} \geq 0.
$$
\end{claim}
\begin{proof}[Proof of Claim \ref{claim-key}]
Let $\omega_Y$ be a K\"ahler form on $Y$.
Then, for a given local coordinate $(t_1, t_2, \dots, t_m)$ of $Y_0$,
there exists a smooth positive function $f$ defined on an open set in $Y_0$ such that
$$
\omega^n= \frac{1}{\phi^{*}f \cdot |\phi^* dt|^2_{\Lambda^m h}} \,
\phi^{*} \omega_Y^{m} \wedge \omega^{n-m},
$$
where $dt:=dt_{1}\wedge dt_2 \wedge \cdots \wedge dt_m$.
We remark that $\phi^{*}f$ and $|\phi^* dt|^2_{\Lambda^m h}$ depend on the choice of local coordinates,
but the product is independent of the coordinates and it is globally defined on $X_0$.
Indeed, it can be seen that
\begin{align*}
\langle \phi^* dt_{\ell}, e_j^{\vee} \rangle_h =\langle \phi^*dt_{\ell}, e_j \rangle_{\rm{pairing}}=
\langle dt_{\ell}, \phi_*e_j \rangle_{\rm{pairing}}=0
\end{align*}
for any $j=m+1,\dots, n$ since $e_j$ is in the kernel of $d\phi_*$.
Therefore we obtain
\begin{align*}
\phi^* dt_\ell&=\sum_{k=1}^{n}\langle \phi^* dt_{\ell}, e_k^{\vee} \rangle_h \, e_{k}^{\vee}
=\sum_{i=1}^{m}\langle \phi^* dt_{\ell}, e_i^{\vee} \rangle_h \, e_{i}^{\vee}.
\end{align*}
Further we obtain \begin{align*}
\phi^* dt &= \det [\langle \phi^{*}dt_{\ell}, e_i^{\vee} \rangle_h]\, e_1^\vee \wedge e_2^\vee \wedge \cdots \wedge e_m^\vee
\text{ \quad and \quad}|\phi^* dt|^2_{\Lambda^m h}=\big| \det [\langle \phi^* dt_{\ell}, e_i^{\vee} \rangle_h ] \big|^2
\end{align*}
by straightforward computations.
On the other hand, the K\"ahler form $\omega_Y$ can be locally written as
$$
\omega_Y=\sqrt{-1}\sum_{i,j=1}^{m} f_{ij} dt_i \wedge d\bar t_{j}
$$
in terms of the given local coordinate $(t_1, t_2, \dots, t_m)$.
From this local expression, we can easily show that
\begin{align*}
\phi^{*} \omega_Y^m \wedge \omega^{n-m} &=c_{n, m}\, \phi^{*}(\det[f_{ij}])\, \phi^{*}(dt \wedge \bar dt)\wedge \omega^{n-m}\\
&=d_{n, m}\, \phi^{*}(\det[f_{ij}])\, \big| \det [\langle \phi^{*}dt_{\ell}, e_i^{\vee} \rangle_h ] \big|^2 \,\omega^{n},
\end{align*}
where $c_{n,m}$ and $d_{n,m}$ are the universal constants depending only on $n$ and $m$.
Therefore it can be seen that $f:=d_{n,m} \det[f_{ij}]$ satisfies the desired equality.
By Fubini's theorem,
we have
\begin{align*}
\int_{X_0} \sum_{j=m+1}^{n} \sqrt{-1}\Theta_{G}(e_j, \bar e_j)\, \omega^{n}&=
\int_{Y_0} \frac{1}{f} \, \omega_Y^m \, \int_{X_y} \frac{1}{|\phi^* dt|^2_{\Lambda^m h}}
\sum_{j=m+1}^{n} \sqrt{-1}\Theta_{G}(e_j, \bar e_j) \, \omega^{n-m}\\
&=\frac{n-m}{2}\int_{Y_0} \frac{1}{f} \, \omega_Y^m \, \int_{X_y} \frac{1}{|\phi^* dt|^2_{\Lambda^m h}}
\sqrt{-1}\Theta_{G} \wedge \omega^{n-m-1},
\end{align*}
where $X_y$ is the fiber of $\phi$ at $y \in Y_0$.
Here we used the equality
$$
\sum_{j=m+1}^{n} \sqrt{-1}\Theta_{G}(e_j, \bar e_j) \, \omega^{n-m}
= \frac{n-m}{2}
\sqrt{-1}\Theta_{G} \wedge \omega^{n-m-1}
$$
of the scalar curvature on the fiber $X_y$.
We finally prove that the fiber integral in the above equality is non-negative.
For simplicity, we put $F:=|\phi^* dt|^2_{\Lambda^m h}$.
Then, by the definition of the curvature $\sqrt{-1}\Theta_{G}$,
we can show that
\begin{align*}
&\int_{X_y} \frac{1}{|\phi^* dt|^2_{\Lambda^m h}}\sqrt{-1}\Theta_{G} \wedge \omega^{n-m-1}\\
=&
\int_{X_y} \frac{1}{F} \sqrt{-1}\partial \overline{\partial} \log F
\wedge \omega^{n-m-1}\\
=&\sqrt{-1} \int_{X_y} \partial \Big(
\frac{1}{F}\, \overline{\partial} \log F \wedge \omega^{n-m-1} \Big)
-\sqrt{-1} \int_{X_y} \partial \Big( \frac{1}{F} \Big) \wedge \overline{\partial}
\log F \wedge \omega^{n-m-1}\\
=&
\int_{X_y} \frac{1}{F^3} \sqrt{-1}\partial F \wedge \overline{\partial} F \wedge \omega^{n-m-1}.
\end{align*}
The last equality follows from Stokes's theorem.
The integrand of the right hand side is non-negative,
and thus the desired inequality can be obtained.
\end{proof}
\end{step}
\begin{step}[Curvature of the canonical bundle $K_Y$]\label{step4}
In this step, from the assumption that $K_Y$ is pseudo-effective,
we will show that the curvature $\sqrt{-1}\Theta_G$ is flat and
$\phi$ is a smooth morphism.
The key point here is
the observation on the flatness of curvature in the horizontal direction
obtained from Claim \ref{claim-semi} and Claim \ref{claim-key}.
\begin{claim}\label{claim-flat}
The following statements hold\,$:$
\vspace{0.1cm} \\
\quad $\bullet$ The canonical bundle $K_Y$ is numerically zero $($that is, $c_{1}(K_Y)=0$$)$.
\vspace{0.1cm} \\
\quad $\bullet$ $H_g([e_i])=\Rur{g}{e_i}{e_i}=0$ for any $i=1,2,\dots, m$.
\vspace{0.1cm} \\
\quad $\bullet$ $\Rur{g}{v}{e_i}\geq 0$ for any tangent vector $v \in T_X$ and any $i=1,2,\dots, m$.
\vspace{0.1cm} \\
\quad $\bullet$ The curvature $\sqrt{-1}\Theta_G$ is flat on $X$.
In particular, the effective divisor $E$ is actually the zero divisor,
and thus the morphism $\phi$ is smooth.
\end{claim}
\begin{proof}[Proof of Claim \ref{claim-flat}]
The left hand side of the equality (\ref{eq-5}) is non-positive
since $\pi^*\phi^{*} K_Y$ is pseudo-effective by the assumption,
and further each term of the right hand side is non-negative
by Claim \ref{claim-key}.
Hence we obtain
\begin{align*}
\int_{ X} c_1(\phi^{*} K^\vee_Y) \wedge \omega^{n-1}&=\int_{\bar X} c_1(\pi^*\phi^{*} K^\vee_Y) \wedge \pi^* \omega^{n-1} =0.
\end{align*}
In general,
if a pseudo-effective line bundle $L$ satisfies $c_1(L) \cdot \{\omega^{n-1} \}=0$ for some K\"ahler form $\omega$,
then $L$ is numerically zero (for example see \cite{Mat13}).
Indeed, for an arbitrary $d$-closed $(n-1, n-1)$ form $\eta$,
we can take a positive constant $C$ such that
$$
\frac{1}{C}\, \omega^{n-1} \leq \eta \leq C \omega^{n-1}.
$$
Then we obtain $c_1(L) \cdot \{\eta\}=0$ by the assumption $c_1(L) \cdot \{\omega^{n-1} \}=0$.
This leads to $c_{1}(L)=0$ by the duality.
Therefore it can be seen that $\phi^{*} K_Y$ is numerically zero.
On the other hand, by the equalities (\ref{eq-5}) and (\ref{eq-7}),
we have
$$
\int_{X_0} \sum_{i=1}^{m} \sqrt{-1}\Theta_{G}(e_i, \bar e_i) \, \omega^{n}=0.
$$
It follows that $\sqrt{-1}\Theta_{G}(e_i, \bar e_i)=0$
for any $i=1,2,\dots, m$ at a point $p \in X_0$
since the integrand $\sqrt{-1}\Theta_{G}(e_i, \bar e_i) $ is
non-negative by Claim \ref{claim-semi}.
By applying the formula (\ref{eq-4}) to the case of $v=e_i$,
we obtain
\begin{align*}
0 \leq \sum_{k=1}^{m}\Rur{g}{e_i}{e_k} \leq \sqrt{-1}\Theta_{G}(e_i, \bar e_i)=0.
\end{align*}
The left inequality follows from Claim \ref{claim-semi}.
In particular, we can see that
$$ \text{
$H_g([e_i])=\Rur{g}{e_i}{e_i}=0$ for any $i=1,2,\dots, m$.
}
$$
This implies that $e_i$ is the minimizer of
the semi-positive holomorphic sectional curvature $H_g$ on $T_X$,
and thus it can be shown that $\Rur{g}{v}{e_i}$ is non-negative
for any tangent vector $v \in T_X$
by Lemma \ref{lem-ineq}.
By applying the formula (\ref{eq-4}) to an arbitrary tangent vector $v \in T_X$ again,
we obtain
\begin{align*}
0 \leq \sum_{i=1}^{m}\Rur{g}{v}{e_i} \leq \sqrt{-1}\Theta_{G}(v, \bar v).
\end{align*}
This means that the curvature $\sqrt{-1}\Theta_{G}$ is semi-positive on $X_0$.
We can see that $\gamma \geq 0$ holds on the Zariski open set $\pi^{-1}(X_0)$,
since $\gamma=\pi^{*}\sqrt{-1}\Theta_{G}$ holds on $\bar X \setminus \pi^{-1}(Z)$
and we have $X_0 \subset X \setminus Z$.
Hence it follows that $\gamma \geq 0 $ on the ambient space $\bar X$ since $\gamma $ is a smooth form.
By the above arguments, the first Chern class $c_{1}(\pi^* \phi^* K_Y^\vee)$
(which is numerically zero) is represented by
the sum of the semi-positive form $\gamma$ and the positive current $[E]$.
Therefore we can conclude that $\gamma=0$ and $E=0$ (namely, $\sqrt{-1}\Theta_{G}=0$).
In particular, we can see that $G$ is a smooth metric (that is, $\phi $ is a smooth morphism).
\end{proof}
\end{step}
\begin{step}[Truly flatness in the horizontal direction and its applications]\label{step5}
In this step, we first show that the statement (2) in Theorem \ref{thm-main} holds and
all the tangent vectors in the horizontal direction are truly flat.
We will prove the statement (3) as an application of the truly flatness.
Further we finally obtain the statement (4) from the theory of foliations.
Now we have the exact sequence of vector bundles
\begin{equation}\label{eq-restan}
0 \xrightarrow{\quad \quad} (T_{X/Y}, g_S) \xrightarrow{\quad \quad}
(T_X, g) \xrightarrow{\quad d\phi_* \quad}
(\phi^{*} T_Y, g_Q)
\xrightarrow{\quad \quad} 0
\end{equation}
on the ambient space $X$ since $\phi$ is a smooth morphism by Claim \ref{claim-flat}.
Let $g_Q$ (resp. $g_S$) be the induced hermitian metric on $\phi^{*} T_Y$ (resp. $T_{X/Y}$).
Then we prove the following claim\,$:$
\begin{claim}\label{claim-truly}
The following statements hold\,$:$
\vspace{0.1cm}\\
\quad $\bullet$ $e_{i}$ is a truly flat tangent vector for any $i=1,2,\dots, m$.
\vspace{0.1cm}\\
\quad $\bullet$ The exact sequence $($\ref{eq-restan}$)$ splits,
and its splitting $T_X= T_{X/Y} \oplus \phi^{*} T_Y$ coincides with the orthogonal decomposition of $(T_X, g)$.
\vspace{0.1cm}\\
\quad $\bullet$ There exists a hermitian metric $g_Y$ on $T_Y$
such that $g_Q=\phi^{*} g_Y$ and $H_{g_Y}\equiv 0$.
In particular, the image $Y$ admits a finite \'etale cover $T \to Y$ by a complex torus $T$.
\end{claim}
\begin{proof}[Proof of Claim \ref{claim-truly}]
For any $i=1,2,\dots, m$, the tangent vector $e_i$ is the minimizer of the holomorphic sectional curvature $H_g$,
and further $\langle \sqrt{-1}\Theta_{g}(v, \bar v)(e_i), e_i \rangle_g $ is non-negative
for any tangent vector $v \in T_X$ by Claim \ref{claim-flat}.
By the formula (\ref{eq-1}), we obtain that
\begin{align}\label{eq-6}
0 \leq \big \langle \sqrt{-1}\Theta_{g}(v, \bar v)(e_i), e_i \big \rangle_g +
\big \langle B_{\bar v} (e_i), B_{\bar v} (e_i) \big \rangle_{g_S}
= \big \langle \sqrt{-1}\Theta_{g_Q}(v, \bar v)(e_i), e_i \big \rangle_{g_Q}
\end{align}
for a tangent vector $v \in T_X$.
On the other hand, the induced metric $\det g_Q$ on $\phi^{*}K_Y^{\vee}=\det \phi^{*} T_Y$
coincides with the metric $G$ constructed in Step \ref{step1}
and the curvature of $\det g_Q=G$ is flat by Claim \ref{claim-flat}.
Therefore we obtain that
\begin{align*}
&\sum_{i=1}^{m}\big \langle \sqrt{-1}\Theta_{g_Q}(v, \bar v)(e_i), e_i \big \rangle_{g_Q}\\
=&\big \langle \sqrt{-1}\Theta_{\det g_Q}(v, \bar v)
(e_1\wedge e_2 \wedge \cdots \wedge e_m ), (e_1\wedge e_2 \wedge \cdots \wedge e_m)
\big \rangle_{\det g_Q} \\
=&0
\end{align*}
by the equality (\ref{eq-8}).
By combining with the inequality (\ref{eq-6}),
we can obtain that
\begin{align*}
\Rur{g}{v}{e_i} = \big \langle \sqrt{-1}\Theta_{g}(v, \bar v)(e_i), e_i \big \rangle_g =0
\quad \text{ and } \quad
\big \langle B_{\bar v} (e_i), B_{\bar v} (e_i) \big \rangle_{g_S} =0
\end{align*}
for any tangent vector $v \in T_X$ and $i=1,2,\dots, m$.
Here we used the fact that
$\langle B_{\bar v} (e_i), B_{\bar v} (e_i) \rangle_{g_S}$ is non-negative.
Then, by Lemma \ref{lemm-flat}, we can see that $e_i$ is truly flat
since $e_i$ satisfies that $\Rur{g}{v}{e_i}=0 $ and $H_g([e_i])=0$.
Further it follows that
$B\in C^{\infty}(X, \Lambda^{0,1}\otimes \Hom(S^{\bot}, S))$ is identically zero
since $\langle B_{\bar v} (\bullet), B_{\bar v} (\bullet) \rangle_{g_S}$
is a semi-positive definite quadratic form on $S^{\bot}$ and
its trace
$\sum_{i=1}^{m}\langle B_{\bar v} (e_i), B_{\bar v} (e_i) \rangle_{g_S}$ is zero
by the above argument.
Hence we obtain the holomorphic orthogonal decomposition $T_X= T_{X/Y} \oplus \phi^{*} T_Y$ (see subsection \ref{subsec-2-1}).
Now we prove the last statement.
For a (local) vector field $v $ of $ T_{Y}$ defined on an open set $U$ in $Y$,
we consider
the section $\phi^{*}v \in H^{0}(\phi^{-1}(U), \phi^{*}T_Y)$
defined by
$$
\phi^{-1}(U) \ni p \to v(\phi(p)) \in T_{Y, \phi(p)}={(\phi^{*}T_Y)}_p,
$$
which we will denote by the notation $\phi^{*}v$.
If the function $|\phi^{*} v|_{g_Q}$ is a constant on the fiber $X_y$,
we can define the hermitian metric $g_Y$ of $Y$ by $|v|_{g_Y}:=|\phi^{*} v|_{g_Q}$.
Then we have $g=\phi^{*} g_Y$ by the definition.
If we can show that the restriction of $\sqrt{-1}\partial \overline{\partial} \log |\phi^{*} v|_{g_Q} $
to the fiber is a semi-positive $(1,1)$-form,
it should be a constant by the maximal principle,
since $|\phi^{*} v|_{g_Q}$ is a psh function globally defined on the compact fiber.
For this purpose, we consider the sub-line bundle $L$ of $\phi^{*}T_Y$
$$
(L:=\Spn \langle \phi^{*} v \rangle, g_L) \subset (\phi^{*}T_Y, g_Q)
$$
spanned by $\phi^{*} v$.
Let $g_L$ be the induced metric on $L$.
By the definition of the curvature and the induced metric,
we obtain
$$
\sqrt{-1}\Theta_{g_L}: = \sqrt{-1}\Theta_{g_L}(L)=-\sqrt{-1}\partial \overline{\partial} \log |\phi^{*} v|^2_{g_L}=-\sqrt{-1}\partial \overline{\partial} \log |\phi^{*} v|^2_{g_Q}.
$$
By applying the formula (\ref{eq-2a}) to the above injective bundle morphism, we obtain that
$$
\sqrt{-1}\Theta_{g_L}(w, \bar w) |\phi^{*} v|^2_{g_L}
=\big \langle \sqrt{-1}\Theta_{g_L}(w, \bar w)(\phi^{*} v), \phi^{*} v \big \rangle_{g_L}
\leq \big \langle \sqrt{-1}\Theta_{g_Q}(w, \bar w)(\phi^{*} v), \phi^{*} v \big \rangle_{g_Q}
$$
for a tangent vector $w \in T_X$.
We have already shown that the tangent vectors $\{e_{i}\}_{i=1}^{m}$ are truly flat by the above argument.
The vector $\phi^{*} v$ can be written as a linear combination of $\{e_{i}\}_{i=1}^{m}$,
and thus it is also truly flat.
On the other hand, by the holomorphic orthogonal decomposition $T_X=T_{X/Y}\oplus \phi^{*} T_Y$,
the section $\phi^{*} v$ of $\phi^{*} T_Y$ determines the section of $T_X$,
which we denote by the same notation $\phi^{*} v$.
Then we obtain
$$
\big \langle \sqrt{-1}\Theta_{g_Q}(w, \bar w)(\phi^{*} v), \phi^{*} v \big \rangle_{g_Q}
=
\big \langle \sqrt{-1}\Theta_{g}(w, \bar w)(\phi^{*} v), \phi^{*} v \big \rangle_{g}
=
0.
$$
The right equality follows from the truly flatness of $\phi^{*} v$.
Therefore we can see that $\sqrt{-1}\Theta_{g_L}$ is semi-negative (in particular $|\phi^{*} v|_{g_Q}$ is a constant).
We finally check that the holomorphic sectional curvature of $g_Y$ is identically zero.
Note that, in general, a compact K\"ahler manifold is a complex torus up to finite \'etale covers
when the holomorphic sectional curvature is identically zero
(see \cite{Igu54}, \cite[Proposition 2.2]{HLW16}, \cite{Ber66}, \cite{Igu54}).
For a given tangent vector $v \in T_Y$,
the vector $\phi^* v$ satisfies $d\phi_* (\phi^* v)=v$.
Hence we can easily see that
\begin{align*}
0=\big \langle \sqrt{-1}\Theta_{g}(\phi^* v, \bar \phi^* v)(\phi^{*} v),
\phi^{*} v \big \rangle_{g}
&=\big \langle \sqrt{-1}\Theta_{g_Q}(\phi^* v, \bar \phi^* v)(\phi^{*} v), \phi^{*} v \big \rangle_{g_Q}\\
&=\big \langle \sqrt{-1}\Theta_{g_Y}(v, \bar v)(v), v \big \rangle_{g_Y}
\end{align*}
by $d\phi_* (\phi^* v)=v$, $g_Q=\phi^{*}g_Y$, and the truly flatness of $\phi^{*} v$.
\end{proof}
We check the statement (4) in Theorem \ref{thm-main}.
When $X$ is projective,
it can be shown that the morphism $\phi: X \to Y$ is a (holomorphic) fiber bundle
(in particular, all the fibers are isomorphic)
by the classical Ehresmann theorem and \cite[Lemma 3.19]{Hor07}.
\end{step}
We finish the proof of Theorem \ref{thm-main}.
\end{proof}
\begin{rem}\label{rem-main}
If the foliation $\phi^*T_Y \subset T_X$ obtained from Theorem \ref{thm-main} is integrable
(that is, it is closed under the Lie bracket),
then it can be shown that $\phi$ is locally trivial and
we have the decomposition
$$
X_{\rm{univ}} \cong Y_{\rm{univ}} \times F_{\rm{univ}}=\mathbb{C}^m \times F_{\rm{univ}}
$$
by the Ehresmann theorem (for example see \cite[Theorem 3.17]{Hor07}).
The integrability of $\phi^*T_Y \subset T_X$ is satisfied
when the dimension of $Y$ is one.
In this case, the image $Y$ is automatically an elliptic curve by the statement (3) in Theorem \ref{thm-main}.
See \cite{Hor07} and references therein for more details.
\end{rem}
\subsection{Proof of Theorem \ref{thm-mainn}}\label{subsec-3-2}
In this subsection, we will prove Theorem \ref{thm-mainn}
by modifying the arguments in the proof of Theorem \ref{thm-main} for almost holomorphic maps.
\begin{theo}[=Theorem \ref{thm-mainn}]\label{r-thm-mainn}
Let $(X, g)$ be a compact K\"ahler manifold
with semi-positive holomorphic sectional curvature and
$\phi: X \dashrightarrow Y$ be an almost holomorphic map to
a compact K\"ahler manifold $Y$ with pseudo-effective canonical bundle.
Let $X_1$ and $Y_1$ be Zariski open sets such that
$\phi :X_1:=\phi^{-1}(Y_{1}) \to Y_1$ is a morphism.
Then we have the followings\,$:$
\begin{itemize}
\item[(0)] The numerical dimension $\nu(Y)$ of $Y$ is equal to zero.
\vspace{0.1cm}
\item[(1)] $\phi$ is a smooth morphism on $X_1$.
\vspace{0.1cm}
\item[(2)] The standard exact sequence of vector bundles on $X_1$
$$
0 \xrightarrow{\quad \quad} T_{{X_1}/{Y_1}}:=\Ker d\phi_* \xrightarrow{\quad \quad}
T_{X_1} \xrightarrow{\quad d\phi_* \quad}
\phi^{*} T_{Y_1}
\xrightarrow{\quad \quad} 0
$$
gives the holomorphic orthogonal decomposition
$$
T_{X_1} = T_{X_1/Y_1} \oplus \phi^{*} T_{Y_1}.
$$
Moreover, all the tangent vectors in $\phi^{*} T_{Y_1} \subset T_{X_1}$ are truly flat.
\vspace{0.1cm}
\item[(3)] Let $g_Q$ be the hermitian metric on $\phi^{*} T_{Y_1}$ induced by
the above exact sequence and the given metric $g$.
Then there exists a hermitian metric $g_Y$ on $T_{Y_1}$ with the following properties\,$:$
\vspace{0.1cm}
\begin{itemize}
\item[$\bullet$] $g_Q$ is obtained from the pull-back of $g_Y$ $($namely, $g_Q=\phi^* g_Y$$)$.
\vspace{0.1cm}
\item[$\bullet$] The holomorphic sectional curvature of $(Y_1, g_Y)$ is identically zero.
\end{itemize}
\end{itemize}
\end{theo}
\begin{proof}[Proof of Theorem \ref{thm-mainn}]
The strategy of the proof is essentially the same as that of Theorem \ref{thm-main}.
We will only explain how to revise the proof of Theorem \ref{thm-main} to avoid repeating the same arguments.
We use the same notations as in the proof of Theorem \ref{thm-main}.
For an almost holomorphic map $\phi: X \dashrightarrow Y$,
we take a resolution $\tau: \Gamma \to X$ of the indeterminacy locus of $\phi$.
We denote, by the notation $\bar \phi: \Gamma \to Y$,
the morphism with the following commutative diagram\,$:$
\begin{equation*}
\xymatrix@C=40pt@R=30pt{
& \Gamma \ar[d]_\tau \ar[rd]^{\bar{\phi}\ \ } & \\
& X \ar@{-->}[r]^{\phi \ \ \ } & Y.\\
}
\end{equation*}
Then we have the injective sheaf morphism
$$
\bar \phi^* K_Y \xrightarrow{\quad d \bar \phi^{*} \quad} \Lambda^m \Omega_{\Gamma}.
$$
By taking the push-forward by the modification $\tau$,
we obtain the injective sheaf morphism
$$
(L:= \tau_* \bar \phi^* K_Y, H) \xrightarrow{\quad f \quad}
(\Lambda^m \Omega_{X}, \Lambda^m h).
$$
Here we used the formula $\tau_{*}\Lambda^m \Omega_{\Gamma} =\Lambda^m \Omega_{X}$.
For simplicity,
we denote the line bundle $\tau_* \bar \phi^* K_Y$
by the notation $L$ and the above sheaf morphism by the notation $f$.
We can take a (non-empty) Zariski open set $Y_1$ such that
the restriction $\phi :X_1:=\phi^{-1}(Y_{1}) \to Y_1$
is a morphism without the indeterminacy locus
since $\phi$ is an almost holomorphic map.
Also, we take the maximal Zariski open set $Y_0 \subset Y_1$ such that
$\phi :X_0:=\phi^{-1}(Y_{0}) \to Y_0$ is smooth.
One of our purposes is to prove that $Y_0=Y_1$.
From now on, we will check that the same arguments as in each step in the proof of Theorem \ref{thm-main} work
by replacing $\phi^{*}K_Y$ with $L$ and $\phi :X \to Y$ with $\phi :X_1 \to Y_1$.
By the same way as in Step \ref{step1},
we can construct a singular hermitian metric $H$ on $L$ and
its dual metric $G$ on $L^{\vee}$.
The line bundle $L=\tau_* \bar \phi^* K_Y$
coincides with the usual pull-back $\phi^{*}K_Y$ on $X_1$
(that is, $L$ can be seen as the extension of the pull-back
$\phi^{*}K_Y$ defined on $X_1$ to $X$).
Let $\mathcal{I}$ be the degenerate ideal of $f$ and let $Z'$ be the support of the cokernel $\mathcal{O}_X/\mathcal{I}$.
We do not know whether $Z'$ coincides with the non-smooth locus $Z$ of $\phi$,
but we have $Z' \cap X_1=Z \cap X_1$
since $f$ is just the morphism $d\phi^*$ of the pull-back on $X_1$.
For a resolution $\pi: \bar{X} \to X$ of the degenerate ideal $\mathcal{I}$,
we can easily check the same statements as in Claim \ref{claim-div}
by replacing $Z$ with $Z'$.
In particular, we have
\begin{align*}
2\pi c_1(\pi^{*}L^\vee)
\ni
\pi^{*}\sqrt{-1}\Theta_G=\gamma + [E],
\end{align*}
for some smooth $(1,1)$-form $\gamma$ and
the integration current $[E]$ of an effective divisor $E$.
In Step \ref{step2}, we only considered tangent vectors at a smooth point of $\phi$.
Hence there is no difficulty to obtain Claim \ref{claim-semi}.
In Step \ref{step3}, we essentially discussed local problems in $Y_1$.
Therefore we can obtain the equality (\ref{eq-5}), the equality (\ref{eq-7}), and Claim \ref{claim-key}
by replacing $\phi^*K_Y$ with $L$.
In Step \ref{step4}, we used the global condition that $\phi^{*}K_Y$ is pseudo-effective.
However we can see that the line bundle $L$ is pseudo-effective by the definition,
and thus we can repeat the same argument as in Claim \ref{claim-flat}.
As a result, we can conclude that $L$ is numerically zero and $\sqrt{-1}\Theta_{G}$ is flat on $X$.
This implies that $\bar \phi^* K_{Y}$ is numerically equivalent
to some exceptional divisor by the definition $L=\tau_* \bar \phi^* K_Y$.
Hence the numerical dimension of $K_{Y}$ is zero.
Further the morphism $f$ is an injective bundle morphism
since the effective divisor $E$ is the zero divisor by $\sqrt{-1}\Theta_{G}$ is flat.
In particular, since the morphism $f$ is equal to the morphism $d\phi^*$ of the pull-back over $Y_1$,
the morphism $\phi$ is smooth over $Y_1$ (namely, $Y_1=Y_0$).
The rest arguments in Step \ref{step5} are local in $Y_1=Y_0$,
and thus we can easily check the same conclusions as in Theorem \ref{thm-main} over $Y_{0}$
by replacing $X$ and $Y$ with $X_0$ and $Y_0$.
\end{proof}
\subsection{Proof of Theorem \ref{thm-mai} and Corollary \ref{cor-mainnnn}}
In this subsection, we will prove the following theorem.
Theorem \ref{thm-mai} and Corollary \ref{cor-mainnnn} can be directly obtained from the following theorem.
\begin{theo}\label{thm-sum}
Let $(X, g)$ be a compact K\"ahler manifold
with semi-positive holomorphic sectional curvature and
$Y$ be a compact K\"ahler manifold with pseudo-effective canonical bundle.
Let $\phi: X \dashrightarrow Y$ be an almost holomorphic map from $X$ to $Y$.
Then we have
$$
\dim X - \dim Y \geq n_{{\rm{tf}}}{(X, g)}.
$$
\end{theo}
\begin{proof}[Proof of Theorem \ref{thm-sum}]
For the proof, we will use the arguments in the proof of Theorem \ref{thm-main} and Theorem \ref{thm-mainn}.
For simplicity, we put $k:=n_{{\rm{tf}}}{(X, g)}$.
We take a Zariski open set $Y_1$ in $Y$ such that $\phi: X_1=\phi^{-1}(Y_1) \to Y_1$ is a morphism
(over which $\phi$ is actually smooth by Theorem \ref{thm-mainn}).
The invariant $n_{{\rm{tf}}}{(X, g)}_p$ is
lower semi-continuous with respect to $p \in X$ in the classical topology (see Definition \ref{def-flat}).
In particular, the condition $n_{{\rm{tf}}}{(X, g)}_p=k$ is an open condition.
Hence we can find a point $p$ such that
$$
n_{{\rm{tf}}}{(X, g)}_p=k=\max_{p \in X}n_{{\rm{tf}}}{(X, g)}_p \text{\quad and \quad } p \in X_1.
$$
It follows that tangent vectors $\{e_{i}\}_{i=1}^{m}$ in the horizontal direction are truly flat by (2) in Theorem \ref{r-thm-mainn} (see also Claim \ref{claim-truly}).
In particular, the vector space $\phi^{*}T_Y=\Spn \langle \{e_{i}\}_{i=1}^{m} \rangle$ at $p$
is contained in $V_{{\rm{flat}},p}$.
Therefore we obtain the desired inequality $m \leq n-k$.
\end{proof}
In the rest of this subsection,
we will check Theorem \ref{thm-mai} and Corollary \ref{cor-mainnnn}.
\begin{theo}[=Theorem \ref{thm-mai}]\label{r-thm-mai}
Let $(X, g)$ be a compact K\"ahler manifold such that $X$ is projective and
the holomorphic sectional curvature is semi-positive.
Let $\phi: X \dashrightarrow Y$ be a MRC fibration of $X$.
Then we have
$$
\dim X -\dim Y \geq n_{{\rm{tf}}}{(X, g)}.
$$
In particular, the manifold $X$ is rationally connected
if $n_{{\rm{tf}}}{(X, g)}=\dim X$ $($which is satisfied if the holomorphic sectional curvature is quasi-positive$)$.
\end{theo}
\begin{proof}
We obtain the desired inequality by applying Theorem \ref{thm-sum} to MRC fibrations.
The latter conclusion is obvious.
Indeed, when $n_{{\rm{tf}}}{(X, g)}=\dim X$,
the image $Y$ should be one point.
This implies that $X$ is rationally connected.
\end{proof}
\begin{cor}[=Corollary \ref{cor-mainnnn}]\label{r-cor-mainnnn}
Let $(X, g)$ be a compact K\"ahler manifold
with semi-positive holomorphic sectional curvature.
Then we have
$$
h^{0}(X, \Omega_X)\leq \dim X - n_{{\rm{tf}}}{(X, g)}.
$$
In particular, we obtain $h^{0}(X, \Omega_X)=0$
if $n_{{\rm{tf}}}{(X, g)}=\dim X$ $($which is satisfied if the holomorphic sectional curvature is quasi-positive$)$.
\end{cor}
\begin{proof}
We consider the Albanese map $\alpha: X \to \Alb(X)$ of $X$.
Then the canonical bundle $K_{\Alb(X)}$ is trivial,
and thus the assumptions in Theorem \ref{thm-sum} are satisfied.
We obtain the desired conclusion by $\dim Y =\dim \Alb(X)=h^{0}(X, \Omega_X)$.
\end{proof}
\subsection{Proof of Corollary \ref{cor-main}}
In this subsection, we will obtain Corollary \ref{cor-main} as an application of Theorem \ref{thm-main}.
\begin{cor}[=Corollary \ref{cor-main}]\label{r-cor-main}
Let $X$ be a compact K\"ahler manifold with semi-positive holomorphic sectional curvature.
Then the followings hold\,$:$
\vspace{0.2cm}\\
\quad $\bullet$ All the statements of Conjecture \ref{conj-str} hold
in the case of $X$ being a surface.
\vspace{0.1cm}\\
\quad $\bullet$ The statement $(1)$ of Conjecture \ref{conj-str} holds
if $X$ is projective and a MRC fibration of $X$ can be chosen to be a morphism.
\end{cor}
\begin{proof}[Proof of Corollary \ref{cor-main}]
We consider a compact K\"ahler manifold $X$
with semi-positive sectional curvature.
If the holomorphic sectional curvature is identically zero,
then $X$ itself admits a finite \'etale cover by a complex torus by \cite{Igu54}
(see also \cite[Proposition 2.2]{HLW16} and \cite{Ber66}).
Then there is nothing to prove.
From now on, we consider the case where $H_g$ is semi-positive,
but not identically zero.
In this case, we can conclude that the canonical bundle $K_X$ is not pseudo-effective.
Indeed, it follows that the scalar curvature $S$ of the K\"ahler metric $g$ is positive
since the scalar curvature $S$ can be expressed
as the integral of the holomorphic sectional curvature on
the projective space $\mathbb{P}(T_{X,p}^\vee)$ (for example, see \cite{Ber66}).
Then we can see that the canonical bundle $K_X$ is not pseudo-effective
by the formula
$$
\int_{X} c_1(K_X) \wedge \omega^{n-1} = - \frac{1}{n\pi}\int_{X} S \, \omega^n < 0,
$$
where $\omega$ is the K\"ahler form associated to $g$.
To check the first statement, we assume that $X$ is a compact K\"ahler surface.
By the classification of compact complex surfaces,
it can be seen that a K\"ahler surface such that $K_X$ is not pseudo-effective
is a minimal rational surface or a ruled surface over a curve of genus $\geq 1$.
It is sufficient to consider the case of $X$ being a ruled surface over
a curve of genus $\geq 1$ since a minimal rational surface is rationally connected.
In this case, we can conclude that the ruling $X \to B$ is minimal
(that is, a submersion) and the base is elliptic curve,
by applying Theorem \ref{thm-main}.
The direct summand $\phi^* T_B $ is integrable
since the rank of $\phi^* T_B $ is one (see Remark \ref{rem-main}).
Hence the universal cover can be shown to be
the product of $\mathbb{C} \times \mathbb{P}^1$
by the classical Ehresmann theorem.
To check the second statement, we consider a smooth projective variety
whose holomorphic sectional curvature is semi-positive, but not identically zero.
Then, since $K_Y$ not pseudo-effective by the first half argument,
a MRC fibration $\phi: X \dashrightarrow Y$ is non-trivial.
Then the image $Y$ is not uniruled by \cite[Theorem 1.1]{GHS03} and
the canonical bundle $K_Y$ of $Y$ is pseudo-effective by \cite{BDPP}.
Therefore we can directly apply Theorem \ref{thm-main} if a MRC fibration can be chosen to be a morphism.
Then the statement $(1)$ of Conjecture \ref{conj-str} is obvious.
\end{proof}
In the rest of subsection, we give a remark on smooth projective varieties with nef anti-canonical bundle.
Even if a compact K\"ahler manifold $X$ has the semi-positive holomorphic sectional curvature,
the anti-canonical bundle $K_X^{\vee}$ is not necessarily nef (for example, see \cite[Example 3.6]{Yan16}).
However it is worth to mention that we can confirm that Conjecture \ref{conj-str} holds
when $X$ is projective and $X$ has the nef anti-canonical bundle,
by using Theorem \ref{thm-main} and the deep structure theorem proved by Cao-H\"oring in \cite{CH17}.
\begin{cor}\label{cor-CH}
Let $(X, g)$ be a compact K\"ahler manifold such that $X$ is projective and the holomorphic sectional curvature is semi-positive.
Further we assume that the anti-canonical bundle $K_X^\vee$ is nef.
Then Conjecture \ref{conj-str} can be affirmatively solved.
\end{cor}
\begin{proof}
By the structure theorem of \cite{CH17}, we can choose a MRC fibration $\phi: X \to Y$ to be a (locally trivial) morphism.
Further we have the decomposition $X_{\rm{univ}}\cong F \times Y_{\rm{univ}}$,
where $F$ is the rationally connected fiber of $\phi$.
By applying Theorem \ref{thm-main} to this MRC fibration $\phi: X \to Y$,
we can see that $Y$ admits a finite \'etale cover $T \to Y$ by an abelian variety $T$.
This finishes the proof.
\end{proof}
\section{Open problems related to semi-positive sectional curvature}\label{Sec-4}
In this section, we discuss open problems related to the geometry of semi-positive sectional curvature.
The first problem is concerned with Conjecture \ref{conj-str}.
If (1) and (2) in the following problem are affirmatively solved,
then Conjecture \ref{conj-str} for smooth projective varieties
can be obtained from Theorem \ref{thm-main}.
\begin{prob}\label{prob-main}
Let $(X, g)$ be a compact K\"ahler manifold such that $X$ is projective and
the holomorphic sectional curvature $H_g$ is semi-positive.
Let $\phi : X \dashrightarrow Y$ be a MRC fibration of $X$.
\begin{itemize}
\item[(1)] Can we choose a MRC fibration of $X$ to be a morphism?
\item[(2)] Does $\phi^{*}T_Y$ determine an integrable foliation?
\item[(3)] Does the fiber $F$ admit a K\"ahler metric $g_F$ such that $n_{{\rm{tf}}}(F, g_F)=\dim F$?
\item[(4)] Does the equality $\dim X = \dim Y + n_{{\rm{tf}}}(X, g)$ hold?
\end{itemize}
\end{prob}
The following problem seems to give a strategy to study Conjecture \ref{conj-str} for compact K\"ahler manifolds.
If the following problem can be solved,
we can apply Theorem \ref{thm-main} and Conjecture \ref{conj-str} to MRC fibrations of the fiber $Z$.
\begin{prob}\label{prob-mainn}
Let $(X, g)$ be a compact K\"ahler manifold
with the semi-positive holomorphic sectional curvature $H_g$.
After we take a suitable finite \'etale cover $X^* \to X$,
we consider the Albanese map $\alpha: X^* \to \Alb(X^*)$.
\begin{itemize}
\item[(1)] Is the fiber $Z$ projective?
\item[(2)] Is the holomorphic sectional curvature
$H_{g_Z}$ of the induced metric $g_Z$ semi-positive?
\end{itemize}
\end{prob}
When $X$ admits a K\"ahler metric with quasi-positive holomorphic sectional curvature,
it seems to be natural to expect that $X$ is automatically projective
(cf. \cite[Theorem 1.7]{Yan17a}).
The following problem, which was posed by Yang in a private discussion,
gives a generalization of this expectation.
Also, it is an interesting problem to consider
rationally connectedness or holomorphic sectional curvature
from the viewpoint of (uniform) RC positivity introduced by Yang.
See \cite{Yan18b} for vanishing theorems and
\cite[Theorem 1.7, Conjecture 1.9]{Yan18c} for rationally connectedness.
\begin{prob}\label{prob-proj}
Let $(X, g)$ be a compact K\"ahler manifold $($or more generally a hermitian manifold$)$
with the semi-positive holomorphic sectional curvature $H_g$.
Assume that $X$ has no truly flat vector at some point of $X$
$($or $H_g$ is quasi-positive$)$.
\begin{itemize}
\item[(1)] Can we obtain $h^{i}(X, \mathcal{O}_X)=0$ for any $i>0$?
\item[(2)] Is $X$ automatically projective and rationally connected?
\end{itemize}
\end{prob}
|
\section{Related Work}\label{ch:related-work}
This paper contributes to the research area Learning Analytics in MOOCs.
Recent work investigates learner motivation and activity, finding and labelling characteristic groups.
\subsection{Learner Activity}
Part of Learning Analytics research skips motivational factors and starts at learner activity, which is represented by website usage in form of clickstream events.
Whitehill et al.~\cite{whitehill2015beyond} have the goal to react automatically to student stopout.
In order to predict a good time to intervene, they include metrics such as the time since the last student activity, a measure for the regularity of the events, and the total number of different event types produced by a user.
Taylor et al.\cite{taylor2014likely} analyze event stream data from edX\footnote{\url{https://www.edx.org/}} courses and aim to predict stopout one week in advance.
Among their metrics are the total number and the average length of forum posts, the total time spent on all resources, and a correctness percentage for homework assignments.
Another approach by Halawa et al. \cite{halawa2014dropout} uses binary features to predict student dropout.
Their features include whether an assignment or a video was skipped, whether a student is lagging behind by more than two weeks, as well as whether the average quiz score fell below 50\%.
\noindent Kizilcec et al. \cite{kizilcec2013deconstructing} examine learner disengagement.
They use engagement trajectories of students, based on assignment completion and video consumption.
They argue, that for their feature choice, favoring trends of engagement over student scores was a deciding factor in finding meaningful groups.
As a second step, they test these activity groups for correlations with another set of features.
Among these features are survey results, such as enrollment intentions and overall course experience, and the number of forum posts per student.
\subsection{Characteristic Groups}\label{subsec:rel:characteristic-groups}
Previous work has derived several characteristic groups from student activity.
All of the regarded works describe their groups as mutually exclusive for a given point in time, but students may move between groups during the time of the course.
Wilkowski et al. \cite{wilkowski2014student} identify four groups of students based on their stated intentions.
Hill \cite{hill2013archetypes} describes five types of student activity.
Regarding groups with less activity, they both agree on: \emph{No-shows}, who enroll for the course, but never log in or engage with the content; \emph{Observers}, who drop in, only to see how the course is taught.
Within more active learners, Wilkowski et al. describe two groups:
\emph{Casual learners}, who engage with the content to learn a few new things related to school, work, or simply curiosity, and \emph{Completers}, who complete all necessary tasks and earn a certificate of completion.
Here, Hill sees three additional groups, \emph{Drop-ins}, who watch videos for selected topics, browse or participate in the forum, but do not attempt to complete the course;
\emph{Passive Participants}, who view the course as content to consume, participate, but do not engage with the assignments;
\emph{Active Participants}, who take part in discussion forums and finish the majority of the assignments.
Knowing these findings, similarities can be observed in the 4 groups found by Kizilcec et al. \cite{kizilcec2013deconstructing} through clustering engagement trajectories.
\emph{Auditing} and \emph{Completing} users seem to closely resemble the \emph{Passive Participants} and \emph{Active Participants} by Hill. The group of \emph{Sampling} users is similar to \emph{Observers}.
Kizilcec's \emph{Disengaging} group is not examined extensively in the descriptions of Wilkowski et al. and Hill, probably because they did not focus on changes in activity over time.
Coffrin et al. \cite{coffrin2014visualizing} define three groups on a weekly basis depending on student participation and success:
\emph{Auditors} watched videos but did not participate in assessments for a particular week;
\emph{Active} participated in assessments for a particular week;
\emph{Qualified} watched a video or participated in an assessment for a particular week and obtained marks above the 60\textsuperscript{th} percentile, leading to the assumption that these students have the capabilities to complete the course.
\emph{Auditors} and \emph{Active} are similarly defined to Kizilcec's \emph{Auditing} and \emph{Completing} groups.
In their work, Coffrin et al. also consider visualizing state changes between these groups and argue that those visualizations may benefit course instructors.
Lingras et al. \cite{lingras2004interval} analyze data from an online course offered internally to students of a particular university (non-MOOC).
They define three student groups:
\emph{Studious} download current reading material for a week as they usually study using class notes;
\emph{Crammers} download a large set of reading material, indicating their plan for a pre-test cramming;
\emph{Workers} continiously work on assignments and access the discussion forum.
Some efforts focus on single specific groups.
\hbox{Beaudoin} \cite{beaudoin2002learning} suggests that learning often happens on course absence and names this student group \emph{Invisible}, who do not show visibility on the platform in form of written contributions in the forums. In contrast, Huang et al. \cite{huang2014superposter} look at characteristics and influence of very active forum participants and label those \emph{Superposters}, who are among the top 5\% of students based on the number of forum contributions.
The various terms coined in the different works all describe related behavior and are based on similar observations and metrics.
However, there are no standardized definitions yet, thus it is helpful to compare and align the existing terms.
Clustering users should be done starting from the actual platform data instead of an artificial metric defined beforehand.
Thus we describe the verbs used on our platform and the actions they reflect in fine granularity, in order to ease portability and reproducability of the underlying concepts.
To illustrate the relation between the cited groups, the diagram in Figure \ref{fig:related-groups} was created.
Non-overlapping ovals indicate groups that were distinctly or differently defined, overlapping ovals signal similar definitions, and groups in the same oval were likely to be merely differently labeled.
While the width of the ovals has no further meaning, the Y-axis describes the degree of course activeness for a group.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\linewidth]{gfx/related-groups}
\caption{Approximate relations between the characteristic groups discovered by previous works.}
\label{fig:related-groups}
\end{figure}
Previous work supports course instructors and researchers in finding previously described and partly defined student groups in MOOCs.
Currently lacking is support and tooling to tailor and discover previously unknown groups.
This work presents concepts as well as an implementation to fill this current gap.
\section{Evaluation}
We evaluated our concepts and tool with two interview series, ensuring we covered instructors' needs and provided a helpful tool.
The first interviews conducted were aimed to validate the acceptance of our tool and used an early stage prototype of the Cluster Viewer, allowing us to adapt the software and process if necessary.
For this series, we interviewed eight instructors from five different MOOCs, about 30 minutes each.
Regardless of knowing the cited publications concerning the different student groups, most instructors had encountered and were able to name groups as ``no-shows'', students who finish but don't show up in the forum (we coined them ``private passers'') and individuals that contribute extensive and helpful forum posts (``thoughtful thread starters'').
When offering instructors the possibility to uncover and react to students with specific behavior, they were most interested in students who: are likely to drop out (stopouts), behave malicious in the forum (trolls), are most active (actives), have questions but don't ask in the forum (reluctants), are active but don't perform well (effortlers) and those who perform best (high-performers).
A second series of interviews was conducted with five different instructors of a german course about internet security at course mid to evaluate the final implementation of the prototype.
Four of five rated the tool as helpful, while the one expressing that it did not help was confident that his existing experience was enough to support and steer a MOOC.
The four instructors were able to find interesting student groups and wanted to react to them:
Students endangered of stopout after they performed worse in graded quizzes than in ungraded quizzes should be encouraged to ask their questions in the forum prior to the next assignments.
Students performing below average and learning in few very long session should be encouraged to try another studying schedule consisting of more but shorter sessions.
\section{Conclusion}
This paper presented a concept to take informed action on student activity in MOOCs.
We related our work to recent research and contributed a holistic overview of characteristics groups discovered in previous works.
Based on individual events and 17 combined metrics that may be used in any online learning platform, our prototype Cluster Viewer allows to explore student behavior within courses.
We showcased parts of our visualization and explained how instructors can send targeted emails to groups, based on their findings.
The acceptance and perceived usefulness of our tool was validated with several interviews.
A first test in a live course revealed an effect on our conducted informed action, but requires further re-evaluations in order to show statistic significance.
\section{Introduction}
The most striking differences when comparing MOOCs with in-class courses are the mere amount of participants enrolled in MOOCs and the absence of direct personal communication.
These differences make it difficult to gain an intuitive perception of how well a MOOC is currently running.
While a holistic view might present that the overall quiz scores are at about 80\% and the number of support tickets is on average, it would not show that there is a specific group that issued support tickets and achieved low scores due, for example due to wording problems.
Uncovering specific groups in MOOCs is difficult. Several previous works have labelled student groups \cite{wilkowski2014student, hill2013archetypes, kizilcec2013deconstructing, coffrin2014visualizing, lingras2004interval}, but there are no best practices yet on how to separate the participants.
Additionally, it is highly debatable whether there is a common separation criteria that holds for all MOOCs with respect to their strongly varying topics, requirements and settings.
Necessary steps are therefore the distinction of potential events to separate users, combine those to relevant metrics and to provide a framework that allows real-time exploration of the course status, progress and interaction.
\section{Problem Statement}
\section{Implementation}
Prior to implementing ``Cluster Viewer'', two requirements were defined:
exploration of possible groups in a course should be enabled directly on the platform;
instructors shall be given the possibility to interactively analyze activity along the metrics based on their intuition and experience.
Given those prerequisites, we decided to integrate parts of the software directly within the micro-services of our MOOC platform (xikolo), and to attach an R-server for the clustering and evaluation tasks.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{gfx/architecture-overview}
\caption{The openHPI architecture with a focus on the analytics service for the task of cluster analysis.}
\label{fig:architecture-overview}
\end{figure}
Figure \ref{fig:architecture-overview} shows the most important services that contribute to fulfill the posed requirements.
The \emph{web} service builds the user frontend and aggregates data from other services to allow for a coherent and rich presentation of the potential clusters.
The actual calculation is done inside the analytics service.
Whenever a clustering request is issued by the user via the frontend (i.e. the web service), the \emph{QueryController} class takes part in handling this request.
It interacts with the \emph{Dimensions} class to aggregate the desired metrics, which itself retrieves the required data from the underlying postgres database.
The Dimensions class holds specific queries for the pre-defined metrics that combine several of the base events.
Whenever a new metric should be added, this class can be easily adapted.
Aggregated data from the Dimensions class is passed to the Clusterer class that feeds it into the R-Server, carrying out the actual calculations.
While our current setup stores all necessary event data in a postgres database, the analytics service also supports to query neo4J or ElasticSearch data sources.
\subsection{Metric computation}
The event data used for the clustering has a defined obligatory minimal structure and additional optional elements.
Each element comes with the user ID of the participant issuing the event, the verb (or name of the event), and a resource ID.
Optional elements might include additional context and result information.
As mentioned in the Section \ref{sec:concept}, metrics combine several events to a more general or better graspable value.
Discovery metrics, for example, subsume visits of distinct resources, while performance metrics average the results of submitted assignments.
The metrics thus rely on manually written SQL queries that simply join the participating tables in order to create the outcome as a single combined numerical value per metric and user.
To speed up response times and to better utilize our system resources, we combine all requested metrics to one concatenated query using SQL joins.
Thus we need only one round trip to the data source.
The extracted metrics are all numeric and returned as a vector per user $u = \{v_{1}, \dots ,v_{m} \}$ with $m$ values, one per computed metric.
The result set of $n$ valid users thus forms a matrix $D_{n,m} $, containing the value of each metric for each valid user and being the starting point for the cluster analysis.
In order to simplify the results, we only expose those rows, where at least one metric has a value differing from zero.
\subsection{Clustering}
\label{sec:clustering}
The analytics service, written in Ruby on Rails, communicates with the R server using \emph{Rserve}\footnote{\url{https://rforge.net/Rserve/}} and the Ruby gem \emph{rserve-client}\footnote{\url{https://rubygems.org/gems/rserve-client/versions/0.3.1}}.
Once the data has been transferred by the \emph{QueryController}, we employ K-Means \cite{hartigan1979algorithm} to create clusters representing potentially meaningful student groups.
Other distance-baseed clustering approaches like hierarchical clustering or density based clustering were evaluated but rejected due to their increased runtime (>10 seconds) for participant volumes of over 15,000 users per course preventing for interactive exploration.
The steps to build the clusters consist of the following:
\begin{itemize}
\item data normalization: data from different input dimensions needs to be normalized to a common domain in order not to distort the results of the distance-based K-Means clustering.
For that, we use R's scale method.
\item cluster estimation: K-Means needs the desired number of clusters as input.
Thus, we use partitioning around medoids (R's pamk-method) to perform a pre-clustering on sample data to retrieve a suited number if there is none pre-defined by the user.
\item clustering: the actual clustering is carried out by K-Means. The clusters, represented as a vector containing the cluster assignments, is added to our existing matrix $D$ as an additional column. Further information, such as cluster sizes, centers, and densities are used for quality evaluation. We use the sums of squared distances of entries to the center within the cluster (withinss), the sums of squared distances between the cluster centers (betweenss) and the total sums of squares (totalss) to judge the quality of our clusters.
\end{itemize}
\section{Future Work}\label{ch:future-work}
The methods targeting both aspects of our initial motivation - understanding students and taking informed action - can be improved individually.
The metrics could be expanded to cover aspects such as peer assessments to determine social behavior and learning styles. They could also be computed for specific weeks to reveal student trajectories.
Incorporating user optional profile data (age, gender, educational background) also adds further potential.
The exploratory data analysis can be extended to allow filtering on discovered groups and to add additional metrics to be tracked.
To further improve performance, sampling and selective rendering could provide faster feedback for the instructors.
To reproduce the findings of related work, a direct next step is to use our metrics and attempt to find previously discovered groups by others (see Subsection \ref{subsec:rel:characteristic-groups}).
As soon as experience with typical student activity has been gathered across several courses, it will be possible to highlight activity out of the norm in a running course and suggest actions for course instructors.
\section{Concept}\label{sec:concept}
This paper presents a concept that enables instructors of online courses to take informed action based on student activity.
On the most abstract level, our concept consists of two motivations:
(1) To find a reason for action, instructors need to be able to understand their students.
(2) After gaining a detailed understanding, instructors need to be able to take action that is as targeted and as measurable as possible.
We refer to this kind of action as \emph{informed action}.
With the first goal in mind - understanding students - we introduce metrics that reflect student activity, some of which have also been found in previous work.
These metrics are sorted into five categories and set in relation to each other.
To approach the second goal - taking informed action - we define three categories of instructor actions and allow the instructors to \emph{encourage} students on an informed basis.
\subsection{Metrics}\label{sec:metrics}
In order to understand characteristic student activity, we collect platform usage data in the form of events triggered when users perform tracked actions.
The structure of the gathered events is similar to the definitions in the Experience API\footnote{\url{https://github.com/adlnet/xAPI-Spec/blob/master/xAPI.md}} \cite{del2013learning}.
On the gathered events, aggregations can be performed.
Yet, combining information inherent in events into possibly more meaningful metrics could provide a more abstract understanding of the underlying activity.
Therefore, we derive 17 metrics from the events and grouped them into five categories (Fig. \ref{fig:metrics-overview}).
All metrics are computed for each individual user and specific to a particular course.
In addition to the verb counts, we allow these metrics to be used as the basis for discovering characteristic groups of students.
\begin{figure}[h]
\includegraphics[width=\linewidth]{gfx/metrics-overview}
\caption{Overview of the metrics and their relation to higher-level platform usage categories.}
\label{fig:metrics-overview}
\end{figure}
\begin{description}
\item[Platform Exploration] (PE) measures the number of distinct verbs per user.
Since most actions are possible in various courses, this metric expresses the experience a user has with the platform.
\item[Sessions] (S) is the number of consecutive event streams (events per user have no wider gap than 30 minutes).
\item[Total Session Duration] (TSD) is the duration of all sessions.
\item[Average Session Duration] (ASD) is the total duration of all sessions divided by the amount of sessions.
\item[Forum Activity] (FA) represents the sum of textual forum contribution (TFC, questions, comments, and answers) and forum observation (FO, visits and subscriptions).
\item[Video Player Activity] (VPA) represents the sum of video player-related events (video played, paused, resized, fullscreen triggered, speed changed).
\item[Download Activity] (DA) represents the sum of downloads.
\item[Item Discovery] (ID) measures the share of visited items (quizzes (QD) and videos (VD)).
\item[Quiz Performance] (QP) measures the average percentage of correct answers over all graded (GQP) and ungraded (UQP) quizzes taken. Graded quizzes are further divided into main quizzes (MQP, mandatory) and bonus quizzes (BQP, optional).
\end{description}
\subsection{Group Discovery}\label{sec:group-discovery}
Once computed, the previously described metrics can be used to create characteristic student groups.
The aim is to minimize the number of groups to be able to digest the clustering results, such as group sizes, coherencies and attributes, but to maximize a group's expressivity.
We suggest a group is very expressive when teachers can easily understand who is part of the group and are able to assign the group a label that describes their activity.
Simply assigning and labelling groups based on multiples of standard deviations from the mean for a particular metric, is not expressive, while groupings with labels like \emph{frequent video downloaders} or \emph{moderate quiz performers} would be.
The task of finding groups can be approached by classification algorithms.
While many machine learning algorithms for classification need a ground truth or other prior knowledge about expected classes (supervised learning), cluster analysis is one way to perform classification when there is little known about the data or the resulting groups (unsupervised learning).
As we do not have prior understanding about distributions or possible correlations of our metrics we decided to use clustering algorithms.
\subsection{Informed Action}\label{subsec:actions}
When course instructors have gained enough understanding about individual learner activity, they should be able to react to specific activity groups and take interventions to increase student success.
Given all options that came up, we found three action categories: \emph{Encouragement} (i.e. personalized emails), \emph{Rewarding} (i.e. badges) and \emph{Material Improvement} (i.e. add reading material, re-record videos).
\noindent From the three action categories, we focus on enabling the \emph{Encouragement} actions, since we consider them most promising to potentially influence student success.
Thus, we enable teachers to save discovered student groups and send targeted emails to members of a specific group.
\subsection{Visualization}\label{subsec:visualization}
Instructors are free to choose any metrics and events the system has to offer to be included into the clustering process.
To give them a starting point, the interface also suggests several common clustering tasks to have a look into.
After the clustering has finished, several visualizations are presented, including a representation of the cluster sizes and qualities (Figure \ref{fig:cluster-centers}), the centroids, several scatter plots (Figure \ref{fig:scatter-plots}) and distribution charts.
Colors, indicating the clusters, are coherent across all visualizations, so that the graphs support each other and instructors can get a quick glimpse whether their chosen input parameters result in a meaningful distribution.
If they are confident with their findings and believe they understand the students, they can assign names to the found clusters and perform an informed action on users in the clusters afterwards.
\begin{figure}[h]
\includegraphics[width=\linewidth]{gfx/cluster-centers2}
\caption{Cluster centers for quiz metrics measured in our MOOC.}
\label{fig:cluster-centers}
\end{figure}
\begin{figure}[h]
\includegraphics[width=\linewidth]{gfx/scatter-plots}
\caption{Scatter plot visualization.}
\label{fig:scatter-plots}
\end{figure}
\subsection{Performing Actions}
From the mentioned possibilities to react to student behavior, our tool Cluster Viewer allows to send targeted emails to found clusters, or half of the cluster for A/B testing.
Changes in the metrics can be compared on the respective group pages to track the effects.
|
\section{Introduction}
The path integral formalism of Liouville Conformal Field Theory (LCFT hereafter) was proposed in the seminal paper of Polyakov \cite{polyakov1981quantum} and can been seen as a probabilistic theory of $2d$ Riemannian metrics. In a series of recent works \cite{david2016liouville,kupiainen2015local,kupiainen2017integrability} by David, Kupiainen, Rhodes and Vargas, a rigorous mathematical construction of Polyakov's path integral formalism is carried out, which defines the Liouville correlation functions in a probabilistic setting, and consequently provides rigorous mathematical proofs of fundamental formulas in Conformal Field Theory such as the BPZ equations \cite{belavin1984infinite} and the DOZZ formula \cite{dorn1992correlation,zamolodchikov1996conformal} on the 3-point structure constant. Their construction is based on Gaussian Free Field (GFF hereafter) and its exponential which is defined using the theory of Gaussian Multiplicative Chaos (GMC hereafter) of Kahane \cite{kahane1985chaos}. The goal of this article is to investigate extension of this path integral construction to Liouville correlation functions with complex parameters: this problem is known in the physics literature as analytic continuation of Liouville theory \cite{harlow2011analytic}. We refer the reader to the introductory reviews \cite{vargas2017lecture,kupiainen2018dozz,harlow2011analytic,ribault2014conformal} for background and references both in mathematics and physics on this subject.
\subsection{Main result}
The goal of this paper is to prove that the path integral formalism proposed in \cite{david2016liouville} directly defines a natual analytic function on a much larger region for general $n$-point Liouville correlation functions. In the language of LCFT, our main statement is the following (definitions and a precise statement will be given later):
\begin{theo}[Main theorem]\label{th:MainTheorem}
Let $\alpha_i$ be real parameters satisfying the Seiberg bounds
\begin{equation}
\forall i,\quad \alpha_i<Q\quad\text{and}\quad \sum\limits_{i}\alpha_i>2Q
\end{equation}
on the Riemann sphere. The $n$-point Liouville correlation function with $\alpha_j\in\mathbb{R}$, $z_j\in\mathbb{C}$ and $\mu>0$ defined by the path integral formalism in \cite{david2016liouville} on the Riemann sphere
\begin{equation}
\left<\prod_{j=1}^{n}V_{\alpha_j}(z_j)\right>\coloneqq\mathbb{E}\left[\int_{\mathbb{R}}e^{(\sum_j\alpha_j-2Q)c}\prod_j e^{\alpha_j X(z_j)}e^{-\mu e^{\gamma c}\int_{\mathbb{C}}e^{\gamma X}}dc\right]
\end{equation}
also defines a natural analytic continuation to complex valued insertions
\begin{equation*}
\alpha_j+i\beta_j\in\mathbb{C}
\end{equation*}
in the region (which we refer to as the pencil region)
\begin{equation}
\mathcal{R}\coloneqq\cap_j\{|\beta_j|<Q-\alpha_j\}.
\end{equation}
More precisely, the n-point correlation function with complex parameters $(\alpha_j+i\beta_j)_{1\leq j\leq n}\in(\mathbb{C})^n$:
\begin{equation}
\left<\prod_{j=1}^{n}V_{\alpha_j+i\beta_j}(z_j)\right>\coloneqq\mathbb{E}\left[\int_{\mathbb{R}}e^{(\sum_j(\alpha_j+i\beta_j)-2Q)c}\prod_j e^{(\alpha_j+i\beta_j)X(z_j)}e^{-\mu e^{\gamma c}\int_{\mathbb{C}}e^{\gamma X}}dc\right]
\end{equation}
can be defined using the standard regularization procedure similar to the one used in the \cite{david2016liouville,kupiainen2017integrability} and is analytic in all $\alpha_j+i\beta_j$ in the pencil region $\mathcal{R}$.
\end{theo}
The novelty in this paper is to apply various techniques from stochastic calculus to LCFT in order to obtain non-trivial estimates in the case with complex parameters. In particular, two different approaches are proposed to study the convergence in the region $\mathcal{R}$: either by applying Itô calculus or by using the renewal theory to a certain Browian motion associated with the vertex operators.
\subsection{Relation to the work of Kupiainen-Rhodes-Vargas}
Analytic continuation of LCFT via GMC approach was first studied in \cite[Section~4]{kupiainen2017integrability} and was essential for rigorously proving the DOZZ formula using probabilistic approach. The authors of \cite{kupiainen2017integrability} applied radial decomposition (see Section~\ref{sec:Radial} below for a brief review) to some GFF locally and obtained an exponential convergence of Liouville correlation functions along a subsequence of times. As a result, they were able to define analytic continuation of Liouville correlation functions in a small neighborhood near each real parameter $\alpha_j$.
Our work strengthens their result. We identify explicitly a large region $\mathcal{R}$ for which Liouville analytic continuation holds within this probabilistic framework; this is coherent with the physics literature \cite{harlow2011analytic} by Harlow-Maltz-Witten on analytic continuation of Liouville theory, although we are not able to identify our pencil region with the physical region of \cite{harlow2011analytic}. Our method yields in some particular cases explicit formulas and precise estimations on Liouville correlation functions, and we expect to recover or obtain exact formulas using methods based on these new observations.
\subsection{Motivations from conformal bootstrap program}
The problem of investigating analytic continuation of Liouville correlation functions in the parameters $\{\alpha_i\}$ fits naturally within the ongoing program in constructive LCFT of Kupiainen-Rhodes-Vargas, of which the goal is to unify two different approaches to LCFT in physics: the path-integral approach and the conformal bootstrap approach. Without going into too much details, let us cite for example one formula from the conformal bootstrap picture that remains mathematically conjectural, see \cite{teschner2001liouville} for example for a detailed discussion:
\begin{equation}
\left<V_{\alpha_1}(z)V_{\alpha_2}(0)V_{\alpha_3}(1)V_{\alpha_4}(\infty)\right>=\int_{\mathbb{R}_+}C_\gamma(\alpha_1,\alpha_2,Q+iP)C_\gamma(Q-iP,\alpha_3,\alpha_4)\left|\mathcal{F}_{Q+iP,\{\alpha_i\}}(z)\right|^2dP
\end{equation}
where $C_\gamma(\alpha_1,\alpha_2,Q+iP)$ is the three-point correlation function: its value is explicitly known since the DOZZ formula. On the other hand, $\mathcal{F}_{\alpha,\{\alpha_i\}}(z)$ are called universal conformal blocks and are explicit meromorphic functions depending only on $Q+iP$ and $\{\alpha_i\}$. The line $Q+i\mathbb{R}_+$ over which we integrate should correspond to the spectrum of LCFT.
It is tempting to give probabilitic interpretation to this decomposition over the spectrum $Q+iP$ for $P\geq 0$, and to deduce for instance an probabilistic expression for computing the conformal blocks that appear above. Recent progress has been made towards this goal (see \cite{kupiainen2016constructive,baverez2018modular,baverez2018fusion}) and the current article is motivated by defining the vertex operators over the critical line, i.e. with parameter $Q+iP$ for $P\in\mathbb{R}$. In other words, we investigate the question of whether the general $n$-point Liouville correlation function $C_\gamma(\{\alpha_i\},Q+iP)$ can be defined directly using the probabilistic approach of \cite{david2016liouville}. While this idea is directly inspired by the physicists \cite{harlow2011analytic} but based on probabilistic constructions \cite{david2016liouville,kupiainen2017integrability}, we are not able to get very close to the critical line at this moment: the regularization procedure used in the current version diverges when we get out of of the pencil region that we identify explicitly, where same kind of phenomenon has been observed in \cite{harlow2011analytic}. We intend to continue the study of this problem in an upcoming work.
\subsection{Acknowledgements}
We gratefully appreciate stimulating advices from Antti Kupiainen, Rémi Rhodes and Vincent Vargas. We are indebted to Rémi Rhodes in particular for communicating some crucial ideas. We also thank Guillaume Baverez, Linxiao Chen and Joona Oikarinen for discussions. The work is finished at the Isaac Newton Institute during the program \emph{Scaling limits, rough paths, quantum field theory} and started at the Institut Mittag-Leffler during the program \emph{Fractal geometry and dynamics}: we wish to thank both institutes for their very warm hospitality.
\section{Preliminaries}\label{sec:Preliminaries}
\emph{Notations.} Throughout this article, $X$ will denote a Gaussian random field in some region of the Riemann sphere $\widehat{\mathbb{C}}=\mathbb{C}\cup\{\infty\}$. Parameters $\alpha_j+i\beta_j\in\mathbb{C}$ will be complex numbers, and $\{z_i\}$ are points on the Riemann sphere $\widehat{\mathbb{C}}$. We denote by $\bm{z}$, $\bm{\alpha+i\beta}$, $\bm{r}$ some $n$-dimensional vectors of resp. $z,\alpha+i\beta,r$. We also use classical notations $\gamma\in(0,2)$, $Q=\frac{2}{\gamma}+\frac{\gamma}{2}$ and $\mu>0$ to denote parameters in the Liouville action as in \cite{david2016liouville}.
\subsection{Geometric setup}\label{sec:metric}
Throughout the rest of this article we will use a fixed metric $\mathbf{g}(z)d^2 z$ on the Riemann sphere $\widehat{\mathbb{C}}=\mathbb{C}\cup\{\infty\}$ of the form
\begin{equation}
\mathbf{g}(z)\coloneqq |z|_+^{-4}
\end{equation}
where $|z|_+=|z|\vee 1$. This metric has scalar curvature
\begin{equation}
R_\mathbf{g}(z)=-4\mathbf{g}^{-1}\partial_z\partial_{\overline{z}}\ln\mathbf{g}(z)=4\nu
\end{equation}
with $\nu$ the uniforme measure on the circle $\partial B(0,1)$.
\subsection{Gaussian Free Field}\label{sec:GFFSetup}
The GFF (for mathematical backgrounds, see \cite{sheffield2007gaussian,dubedat2009sle,david2016liouville}) with zero average in the uniforme measure on the circle $\nu$ has covariance kernel
\begin{equation}
K(x,y)\coloneqq\mathbb{E}[X(x)X(y)]=-\ln|x-y|+\ln|x|_{+}+\ln|y|_{+}
\end{equation}
for $x,y\in\mathbb{C}\cup\{\infty\}$. Notice that inside the unit disk $\mathbb{D}=B(0,1)$,
\begin{equation}
K(x,y)=K_{\mathbb{D}}(x,y)\coloneqq-\ln|x-y|
\end{equation}
where $K_{\mathbb{D}}$ is the (Neumann boundary condition) Green function inside the unit disk $\mathbb{D}$.
\subsection{Local radial decomposition}\label{sec:Radial}
Let $X$ be the centered log-correlated Gaussian field on the unit disk $\mathbb{D}=B(0,1)$ of covariance kernel
\begin{equation}\label{eq:G_D}
K_\mathbb{D}(x,y)\coloneqq\mathbb{E}[X(x)X(y)]=\ln\frac{1}{|x-y|}.
\end{equation}
Recall that $X(x)$ can be defined as a distribution (in the sense of Schwartz) and admits the following decomposition:
\begin{lemm}[Radial decomposition for log-correlated Gaussian field]\label{lem:Radial}
For every $x\in\mathbb{D}\backslash\{0\}$, $X(x)$ can be written as
\begin{equation}
X(x)=X_{|x|}(0)+N(x)
\end{equation}
where $X_{|x|}(0)$ is the circle-average on center $0$ and radius $|x|$ defined for $r>0$ as
\begin{equation}
X_{r}(0)=\frac{1}{2\pi}\int_{0}^{2\pi}X(r e^{i\theta})d\theta
\end{equation}
and $N(x)$ is the lateral noise distribution defined as
\begin{equation}
N(x)=X(x)-X_{|x|}(0).
\end{equation}
We have the following properties:
\begin{enumerate}
\item The Gaussian fields $\{X_r(0), 0<r<1\}$ and $\{N(x), x\in\mathbb{D}\backslash\{0\}\}$ are independent.
\item The process $B_t\coloneqq\{X_{e^{-t}}(0)\}$ is a standard Brownian motion.
\end{enumerate}
\end{lemm}
This decomposition is well-known in the literature: for example see \cite{sheffield2007gaussian,duplantier2014liouville,kupiainen2017integrability}. One can also verify this lemma directly on the specific kernel $K_\mathbb{D}$ of equation~\eqref{eq:G_D} by calculating explicitly the covariance of each Gaussian field.
\begin{rema}[Independence property]\label{rem:Independence}
We record here a useful property of the radial decomposition. On can extend the radial decomposition procedure to balls of the form $B(z,1)$ for $z\in\mathbb{C}$ by conformal mapping and in particular, the process $\widetilde{B}_t\coloneqq\{X_{e^{-t}}(z)-X_{0}(z)\}$ is also a Brownian motion starting at $0$. Then for disjoint balls $B(z_j,1)\subset B(0,1)^{c}$, these Brownian motions are mutually independent and independent of the $\sigma$-algebra generated by $\{X(z);z\notin\cup_j B(z_j,1)\}$.
\end{rema}
\subsection{Gaussian Multiplicative Chaos}\label{sec:Chaos}
The study of GMC measures started with the seminal work \cite{kahane1985chaos} of Kahane. The theory of GMC allows one to define mathematically exponentials of log-correlated Gaussian field in any dimension and in particular, the exponential of the GFF above in $2d$. We refer the reader to \cite{kahane1985chaos,robert2010gaussian,berestycki2017elementary} for more materials on the definition and convergence of GMC measure and \cite{rhodes2014gaussian} for more applications; let us briefly recall the idea and gather some notations here (we restrict ourselves to minimal setting, namely dimension $2$ and the metric $\mathbf{g}$):
\begin{defi}
Let $X$ be a $\log$-correlated field on a subdomain $\Omega\subset\mathbb{R}^2$ equipped with the metric $\mathbf{g}$. We can define the GMC measure associated to $X$ with parameter $\gamma\in(0,2)$,
\begin{equation}
M_{\gamma}(d^2z)\coloneqq e^{\gamma X(z)}\mathbf{g}(z)d^2z
\end{equation}
to be the limit of the family of random measures
\begin{equation}
M_{\gamma,\epsilon}(d^2z)\coloneqq e^{\gamma X_\epsilon(z)-\frac{\gamma^2}{2}\mathbb{E}[X_\epsilon(z)^2]}\mathbf{g}(z)d^2z
\end{equation}
as $\epsilon$ goes to $0$. Here, $X_\epsilon(z)$ denotes a regularization by a smooth mollifier of the field $X$ (another common regularization is by circle average similar to the previous section).
\end{defi}
Applying the radial decomposition to the GMC measures, we have the following:
\begin{prop}[Radial decomposition for GMC measures]
Let $X$ be the GFF as in Lemma~\ref{lem:Radial} and consider the domain $\Omega=\mathbb{D}\backslash B(0,e^{-t})$. The GMC random measure $M_\gamma(\Omega)$ associated to $X$ on the domain $\Omega$ has the following equivalent expression
\begin{equation}
M_\gamma(\Omega)\overset{(law)}{=}\int_{0}^{t}e^{\gamma(B_s-Qs)}Z_sds
\end{equation}
where
\begin{equation}
Z_s=\frac{1}{2\pi}\int_{0}^{2\pi}e^{\gamma Y(s,\theta)-\frac{\gamma^2}{2}\mathbb{E}[Y(s,\theta)^2]}d\theta
\end{equation}
with $Y(s,\theta)$ a Gaussian field (seen as a distribution) independent of $B_t$ with covariance kernel
\begin{equation}\label{eq:NoiseCorr}
\mathbb{E}[Y(s,\theta)Y(t,\theta')]=\ln\frac{e^{-s}\vee e^{-t}}{|e^{-s}e^{i\theta}-e^{-t}e^{i\theta'}|}.
\end{equation}
\end{prop}
\begin{proof}
For more details and discussions, see \cite[Section 2.7]{kupiainen2017integrability}.
\end{proof}
\subsection{Liouville correlation functions}
One main feature in the path integral formalism of LCFT as defined in \cite{david2016liouville} is that one can express Liouville correlation functions by means of GMC measures with $\log$-singularities. More precisely, by a change of variables, one can express the $n$-point correlation functions (with real parameters $\bm{\alpha}$) in the metric $\mathbf{g}$ (defined as in Section~\ref{sec:metric}) on the Riemann sphere $\mathbb{C}\cup\{\infty\}$ in the following manner:
\begin{equation}
\left<\prod_{j=1}^{n}V_{\alpha_j}(z_j)\right>=\frac{2}{\gamma}\mu^{-s}\Gamma(s)\prod_{k<l}\frac{1}{|z_k-z_l|^{\alpha_k\alpha_l}}\mathbb{E}\left[\left(\int_{\mathbb{C}}F(x,\mathbf{z})M_{\gamma}(d^2 x)\right)^{-s}\right]
\end{equation}
with
\begin{equation}
s=\frac{\sum_j\alpha_j-2Q}{\gamma}
\end{equation}
and
\begin{equation}
F(x,\mathbf{z})=\prod_{j=1}^{n}\left(\frac{|x|_+}{|x-z_j|}\right)^{\gamma\alpha_j}.
\end{equation}
In particular, the correlation function is well-defined if one can make sense of the negative moment and shows that it is positive: this can be done using GMC techniques (see \cite{david2016liouville,huang2018liouville}). One sufficient condition is known as the Seiberg bound \cite{seiberg1990notes}:
\begin{equation}
\forall i,\quad \alpha_i<Q\quad\text{and}\quad \sum\limits_{i}\alpha_i>2Q.
\end{equation}
We refer the reader to \cite{david2016liouville,kupiainen2017integrability} for proof and details on this expression. In the following, we will consequently study the analyticity of moments with negative real parts of GMC measure by the same regularization procedure as in \cite{kupiainen2017integrability}. Namely, we will study the convergence of the regularized $n$-point negative GMC measure:
\begin{equation}
G(\bm{\alpha+i\beta},\mathbf{z};\mathbf{t})\coloneqq\mathbb{E}\left[\prod_{j=1}^{n}e^{(\alpha_j+i\beta_j)X_{r_j}(z_j)-\frac{(\alpha_j+i\beta_j)^2}{2}t_j}M_{\gamma}(C_\mathbf{t})^{-s}\right]
\end{equation}
where $r_j=e^{-t_j}$, $C_{\mathbf{t}}$ is the regularized complex plane
\begin{equation}
C_{\mathbf{t}}\coloneqq \mathbb{C}-\bigcup_{j}B(z_j,e^{-t_j}),
\end{equation}
and the $r_j$-regularization procedure is the one described in Subsection~\ref{sec:Chaos}. Consequence of Remark~\ref{rem:Independence}, we identify $X_{r_j}(z_j)$ as independent Brownian motions and denote them by $B_j(t_j)$. Notice that the function $G(\bm{\alpha+i\beta},\mathbf{z};\mathbf{t})$ is complex differentiable in all components of $\mathbf{t}$, hence defines an multivariate entire function in the $\bm{\alpha}$.
Our goal is to establish conditions on local uniform convergence of $G_{\mathbf{t}}$ as $\mathbf{t}$ goes to infinity: this will yield analyticity of the limit function $G(\bm{\alpha+i\beta};\mathbf{z})$.
\subsection{Freezing estimate}
We use frequently in this paper an estimate on integrals of GMC measure with singularities known as the freezing estimate in the literature \cite{fyodorov2008freezing}:
\begin{lemm}\label{lem:Freezing}
For $\alpha>Q$ and $p>0$,
\begin{equation}
\mathbb{E}\left[\left(\int_{|x|>\epsilon}\frac{1}{|x|^{\gamma\alpha}}M_\gamma(d^2x)\right)^{-p}\right]\leq C\epsilon^{\frac{1}{2}(\alpha-Q)^2}
\end{equation}
and if $\mu>0$,
\begin{equation}
\mathbb{E}\left[\exp\left(-\mu\int_{|x|>\epsilon}\frac{1}{|x|^{\gamma\alpha}}M_\gamma(d^2x)\right)\right]\leq C\epsilon^{\frac{1}{2}(\alpha-Q)^2}.
\end{equation}
with some constant $C$ locally uniform in $\alpha$ and independent of $\epsilon$ when $\epsilon$ is small enough.
\end{lemm}
\begin{proof}
See \cite[Section~6]{kupiainen2015local}.
\end{proof}
In Appendix~\ref{app:freezing} we provide a variant that slightly generalizes the above estimate.
\subsection{Stopping time of drifted Brownian motion}
Let $\alpha<Q$ and consider a negatively drifted Brownian motion $B_t-(Q-\alpha)t$. We define the following stopping times for $n\in\mathbb{N}$:
\begin{equation}
T_n=\inf\{s:B_s-(Q-\alpha)s=-n\}.
\end{equation}
Recall several elementary facts that we will use in the following.
\begin{prop}[Elementary facts on stopping time of drifted Brownian motion]\label{prop:renewalprop}
We recall some basic facts on drifted Brownian motion:
\begin{enumerate}
\item The law of $T_1$ follows an inverse Gaussian with parameter $IG((Q-\alpha)^{-1},1)$, i.e. its probability density function is
\begin{equation}\label{eq:T1density}
\left(\frac{1}{2\pi x^3}\right)^{\frac{1}{2}}\exp\left(-\frac{((Q-\alpha)x-1)^2}{2x}\right)dx.
\end{equation}
In particular, from the exponential tail of the above density function,
\begin{equation}
\mathbb{E}\left[e^{\frac{\beta^2}{2}T_1}\right]<\infty
\end{equation}
if and only if $|\beta|<Q-\alpha$.
\item The sequence $(T_{i+1}-T_{i})_{i\in\mathbb{N}}$ is an i.i.d. sequence distributed as $T_1$.
\item Let $t>0$. Define the residual time of the sequence $(T_{i})_{i\in\mathbb{N}}$ at time $t$ as
\begin{equation}
R_T(t)=\inf\{T_n;T_n>t\}-t.
\end{equation}
Then for all fixed $t>0$ and $|\beta|<Q-\alpha$,
\begin{equation}
\mathbb{E}\left[e^{\frac{\beta^2}{2}R_T(t)}\right]<\infty.
\end{equation}
\end{enumerate}
\end{prop}
We provide in Appendix~\ref{app:renewal} a proof for the last claim.
\subsection{Girsanov theorem}
We will apply Girsanov theorem to some exponential functionals of a GFF (or a Brownian motion) in the following form:
\begin{lemm}[Girsanov theorem]\label{lem:Girsanov}
Let $Y$ be some Gaussian variable measurable with respect to a Gaussian free field $X$ and $F$ some bounded functional. Then
\begin{equation}
\mathbb{E}[e^{Y-\frac{\mathbb{E}[Y]^2}{2}}F(X(\cdot))]=\mathbb{E}[F(X(\cdot)+\mathbb{E}[X(\cdot)Y])].
\end{equation}
\end{lemm}
\subsection{Kahane's inequality}
We record two versions of Kahane's Gaussian comparaison inequality.
\begin{lemm}[Kahane's convexity inequality]\label{lem:KahaneConvexity}
Let $X,Y$ be two centered Gaussian field indexed by $T$ such that
\begin{equation}
\mathbb{E}\left[X(i)X(j)\right]\leq\mathbb{E}\left[Y(i)Y(j)\right],\quad \forall (i,j)\in T\times T.
\end{equation}
Then for all non-negative weights $(p_i)_{i\in T}$ and all convex function $F$ with at most polynomial growth at infinity,
\begin{equation}
\mathbb{E}\left[F\left(\sum\limits_{i\in T}p_i e^{X_i-\frac{1}{2}\mathbb{E}[X_i^2]}\right)\right]\leq\mathbb{E}\left[F\left(\sum\limits_{i\in T}p_i e^{Y_i-\frac{1}{2}\mathbb{E}[Y_i^2]}\right)\right].
\end{equation}
\end{lemm}
\begin{proof}
See \cite[Theorem~2.1]{rhodes2014gaussian} for references in English.
\end{proof}
\begin{lemm}[Kahane-Slepian diagonal inequality]\label{lem:KahaneDiagonal}
Let $X,Y$ be two centered Gaussian fields indexed by $T$ such that there exist subsets $A,B\subset T$ on which
\begin{align}
\mathbb{E}\left[X(i)X(j)\right]\leq\mathbb{E}\left[Y(i)Y(j)\right],&\quad \forall(i,j)\in A;\\
\mathbb{E}\left[X(i)X(j)\right]\geq\mathbb{E}\left[Y(i)Y(j)\right],&\quad \forall(i,j)\in B;\\
\mathbb{E}\left[X(i)X(j)\right]=\mathbb{E}\left[Y(i)Y(j)\right],&\quad \forall(i,j)\notin A\cup B.
\end{align}
Suppose $F:\mathbb{R}^{T}\to\mathbb{R}$ is some smooth real functional with appropriate growth at infinity in both its first and second derivatives and such that
\begin{align}
\partial_{ij}F\geq 0,&\quad \forall(i,j)\in A;\\
\partial_{ij}F\leq 0,&\quad \forall(i,j)\in B.
\end{align}
Then we have
\begin{equation}
\mathbb{E}\left[F(X)\right]\leq\mathbb{E}\left[F(Y)\right].
\end{equation}
\end{lemm}
\begin{proof}
See \cite{zeitouniGaussian}, Theorem 3.
\end{proof}
\section{Local study: setup and regularization}\label{sec:LocalSetup}
To prove the Main Theorem~\ref{th:MainTheorem}, it is instumental to study its local version, namely the regularization and convergence near only $1$ insertion point at $z=0$ with complex parameter $\alpha+i\beta$. We will use study two different regularization procedures that yields probabilistically the correct analytic continuation of (local) Liouville correlation functions in the region
\begin{equation}
\mathcal{R}_{loc}=\{\alpha+i\beta\in\mathbb{C}; |\beta|<Q-\alpha\}.
\end{equation}
More precisely, we will seperate two different regimes:
\begin{itemize}
\item In the region \begin{equation}\mathcal{R}^{M}_{loc}=\mathcal{R}_{loc}\cap\{\alpha>Q-\gamma\}\end{equation} we apply the so-called martingale method, Theorem~\ref{th:LocalMartingale};
\item In the region \begin{equation}\mathcal{R}^{T}_{loc}=\mathcal{R}_{loc}\cap\{\alpha<Q-\frac{\gamma}{2}\}\end{equation} we apply the so-called stopping time method, Theorem~\ref{th:LocalStoppingTime}.
\end{itemize}
Each of these methods will yield in a probabilistic way a natural analytic continuation of (local) Liouville correlation functions. Since their intersection contains a non empty open set, together they extend the Liouville correlation functions analytically to the whole region $\mathcal{R}_{loc}$.
\subsection{Setup and notations}
We first define the local version of Liouville correlation function which reflects the regularization procedure near one insertion point at $z=0$.
\begin{defi}[Liouville correlation function: local version and regularization]\label{def:LocalCorrelation}
Consider the unit disk $\mathbb{D}\subset\mathbb{C}$ parametrized by $(s,\theta)\in\mathbb{R}_{+}\times[0,2\pi]$ by the following map:
\begin{equation}
(s,\theta)\mapsto e^{-s}e^{i\theta}\in\mathbb{D}.
\end{equation}
Let $X$ be the centered log-correlated Gaussian field on $\mathbb{D}$ of covariance kernel
\begin{equation}
K_\mathbb{D}(x,y)\coloneqq\mathbb{E}[X(x)X(y)]=-\ln|x-y|.
\end{equation}
Then following the radial decomposition of GFF (Lemma~\ref{lem:Radial}) one can decompose $X$ into two independent Gaussian components:
\begin{itemize}
\item The radial part with can be expressed in terms of a time-changed Brownian motion \begin{equation}B_t=X_{e^{-t}(0)};\end{equation}
\item The lateral noise part $Y(s,\theta)$ which has covariance kernel \begin{equation}\mathbb{E}[Y(s,\theta)Y(t,\theta')]=\ln\frac{e^{-s}\vee e^{-t}}{|e^{-s}e^{i\theta}-e^{-t}e^{i\theta'}|}.\end{equation} We also use the notation $Z_s$ for the GMC measure associated with the lateral noise $Y$.
\end{itemize}
The local regularized Liouville correlation function with parameter $\alpha+i\beta\in\mathbb{C}$ is defined as
\begin{equation}
G(\alpha+i\beta;T)\coloneqq\mathbb{E}\left[e^{i\beta B_{T}+\frac{\beta^2}{2}T}e^{-\mu\int_{0}^{T}e^{\gamma(B_r-(Q-\alpha)r)}Z_r dr}\right]
\end{equation}
where $T$ denotes some (possibly random) positive time, as long as this expectation can be well defined (e.g. finite).
\end{defi}
\subsection{Convergence and analyticity: martingale method}
In this section, we consider the local regularized Liouville correlation function
\begin{equation}
G(\alpha+i\beta;t)=\mathbb{E}\left[e^{i\beta B_t+\frac{\beta^2}{2}t}e^{-\mu\int_{0}^{t}e^{\gamma(B_r-(Q-\alpha)r)}Z_r dr}\right]
\end{equation}
with fixed deterministic time $t\geq 0$. It is the same regularization as in \cite[Section~4]{kupiainen2017integrability} and it is readily seen that $G$ defines an entire function in $\alpha+i\beta$ for every fixed $t$. We thus study its convergence at $t$ goes to infinity.
As announced before we focus on the region \begin{equation}\mathcal{R}^{M}_{loc}=\mathcal{R}_{loc}\cap\{\alpha>Q-\gamma\}.\end{equation}
\begin{theo}[Local version of the main theorem with fixed time]\label{th:LocalMartingale}
Consider the function
\begin{equation}
G(\alpha+i\beta;t)=\mathbb{E}\left[e^{i\beta B_t+\frac{\beta^2}{2}t}e^{-\mu\int_{0}^{t}e^{\gamma(B_r-(Q-\alpha)r)}Z_r dr}\right].
\end{equation}
We claim that:
\begin{enumerate}
\item For every $\alpha+i\beta\in\mathbb{C}$, $G(\alpha+i\beta;t)$ is well defined and analytic in $(\alpha,\beta)$ for every finite $t\geq 0$;
\item For fixed $\alpha+i\beta\in\mathcal{R}^{M}_{loc}$, the limit $G^M(\alpha+i\beta)$ of $G(\alpha+i\beta;t)$ as $t\to\infty$ is well-defined;
\item The limit $G^M(\alpha+i\beta)$ as a function of $\alpha+i\beta\in\mathbb{C}$ is analytic in $\mathcal{R}^{M}_{loc}$.
\end{enumerate}
\end{theo}
Since for real parameter $\alpha$ this regularization is exactly the original regularization of Liouville correlation function using path integral formalism as defined in \cite{david2016liouville}, this theorem proves that the limit function
\begin{equation}
G^M(\alpha+i\beta)
\end{equation}
is the analytic continuation of the local Liouville correlation function from $\alpha\in\mathbb{R}$ to $\alpha+i\beta\in\mathcal{R}^{M}_{loc}$.
\subsection{Convergence and analyticity: stopping time method}
In this section, we consider the local regularized Liouville correlation function
\begin{equation}
G(\alpha+i\beta;T_N)=\mathbb{E}\left[e^{i\beta B_{T_N}+\frac{\beta^2}{2}T_N}e^{-\mu\int_{0}^{T_N}e^{\gamma(B_r-(Q-\alpha)r)}Z_r dr}\right]
\end{equation}
for $N\in\mathbb{N}$, where $T_N$ is defined as the stopping time for the drifted Brownian motion at level $-N$:
\begin{equation}
T_N=\inf\{t; B_t-(Q-\alpha)t=-N\}.
\end{equation}
We study its convergence as $N$ goes to infinity.
As announced before we focus on the region \begin{equation}\mathcal{R}^{T}_{loc}=\mathcal{R}_{loc}\cap\{\alpha<Q-\frac{\gamma}{2}\}.\end{equation}
\begin{theo}[Local version of the main theorem with stopping time]\label{th:LocalStoppingTime}
Consider the function
\begin{equation}
G(\alpha+i\beta;T_N)=\mathbb{E}\left[e^{i\beta B_{T_N}+\frac{\beta^2}{2}T_N}e^{-\mu\int_{0}^{T_N}e^{\gamma(B_r-(Q-\alpha)r)}Z_r dr}\right].
\end{equation}
We claim that:
\begin{enumerate}
\item If $|\beta|<Q-\alpha$, then $G(\alpha+i\beta;T_N)$ is well defined for every $N\in\mathbb{N}$;
\item For fixed $\alpha+i\beta\in\mathcal{R}^{T}_{loc}$, $|G(\alpha+i\beta;T_N)|<C$ for some finite constant $C=C_{\alpha+i\beta}$ independent of $N$ and locally uniform in $\alpha+i\beta$;
\item For fixed $\alpha+i\beta\in\mathcal{R}^{T}_{loc}$, the limit $G^T(\alpha+i\beta)$ of $G(\alpha+i\beta;T_N)$ as $N\to\infty$ is well-defined;
\item The limit $G^T(\alpha+i\beta)$ as a function of $\alpha+i\beta\in\mathbb{C}$ is analytic in $\mathcal{R}^{T}_{loc}$.
\end{enumerate}
\end{theo}
It is not readily seen that $G(\alpha+i\beta;T_N)$ is analytic in $\alpha+i\beta\in\mathbb{C}$ for fixed $N$, thus in the proof of the above theorem, it shall suffer from some minor modification which disappears in the limit. Lastly, in the region with non-trivial interior $\mathcal{R}^{M}_{loc}\cap\mathcal{R}^{T}_{loc}$ the two functions $G^{T}(\alpha+i\beta)$ and $G^{M}(\alpha+i\beta)$ coincide (since one is the sequential limit of the other), so $G^{T}(\alpha+i\beta)$ is the correct analytic continuation of the local Liouville correlation function from $\mathcal{R}^{M}_{loc}$ to $\mathcal{R}_{loc}=\mathcal{R}^{M}_{loc}\cup\mathcal{R}^{T}_{loc}$.
The last observation defines the analytic continuation of local Liouville correlation function $G(\alpha)$ for $\alpha<Q$ to $G(\alpha+i\beta)$ with $\alpha+i\beta\in\mathcal{R}_{loc}$, i.e. $\{|\beta|<Q-\alpha\}$, where
\begin{itemize}
\item $G(\alpha+i\beta)\coloneqq G^{M}(\alpha+i\beta)$ when $\alpha+i\beta\in\mathcal{R}^{M}_{loc}$;
\item $G(\alpha+i\beta)\coloneqq G^{T}(\alpha+i\beta)$ when $\alpha+i\beta\in\mathcal{R}^{T}_{loc}$.
\end{itemize}
\section{Local study I: martingale method}\label{sec:MartingaleLocal}
We follow notations from Section~\ref{sec:LocalSetup}. In this section we focus on the region
\begin{equation}
\mathcal{R}^{M}_{loc}=\mathcal{R}_{loc}\cap\{\alpha>Q-\gamma\}.
\end{equation}
The main property that we need in this regime is the following:
\begin{prop}[Property: martingale region]\label{prop:MartingaleLocal}
One has the following upper bound on the first derivative of $G(\alpha+i\beta;t)$ with respect to $t$:
\begin{equation}
\left|\frac{\partial G(\alpha+i\beta;t)}{\partial t}\right|\leq Ce^{-\frac{(Q-\alpha)^2}{2}t+\frac{\beta^2}{2}t}
\end{equation}
where $C$ is some finite constant locally uniform in $(\alpha,\beta)$ independent of $t$ when $t\geq 1$.
\end{prop}
\begin{proof}
This is essentially a consequence of the generalized freezing estimate, Appendix~\ref{app:freezing}. It is more convenient to work with the geometry of $\mathbb{D}$, so let us use the following equivalent representation for $G(\alpha+i\beta;t)$ (see Definition~\ref{def:LocalCorrelation} for the conformal change of domain):
\begin{equation}
G(\alpha+i\beta;t)=\mathbb{E}\left[e^{i\beta B_t+\frac{\beta^2}{2}t}e^{-\mu\int_{\mathbb{D}\backslash B(0,e^{-t})}\frac{1}{|x|^{\gamma\alpha}}M_\gamma(d^2x)}\right].
\end{equation}
By Itô calculus on the Brownian motion $B_t$ one can write
\begin{equation}\label{eq:MartingaleIto}
\frac{\partial G(\alpha+i\beta;t)}{\partial t}=\frac{1}{2\pi}\int_{0}^{2\pi}-\mu e^{(\gamma\alpha-2)t}e^{i\gamma\beta t}\mathbb{E}\left[e^{i\beta B_t+\frac{\beta^2}{2}t}e^{-\mu\int_{\mathbb{D}\backslash B(0,e^{-t})}\frac{1}{|x-e^{-t}e^{i\theta}|^{\gamma^2}}\frac{M_\gamma(d^2x)}{|x|^{\gamma\alpha}}}\right]d\theta
\end{equation}
where we recognize a GMC measure with two (real) insertions of respective parameters $\alpha$ and $\gamma$ all contained in $B(0,e^{-t})$. Since by assumption $\alpha+\gamma>Q$, applying Lemma~\ref{lem:GeneralFreezing} yields
\begin{equation}
\left|\frac{\partial G(\alpha+i\beta;t)}{\partial t}\right|\leq\mu Ce^{(\gamma\alpha-2)t}e^{-\frac{(\alpha+\gamma-Q)^2}{2}t}e^{\frac{\beta^2}{2}t}\leq Ce^{-\frac{(Q-\alpha)^2}{2}t+\frac{\beta^2}{2}t}
\end{equation}
with $C$ finite constant locally uniform in $(\alpha,\beta)$, independent of $t$ when $t\geq 1$.
\end{proof}
\subsection{Convergence and analyticity of the limit}
With Proposition~\ref{prop:MartingaleLocal} we are ready to give analytic continuation of local Liouville correlation function in the martinagle region.
\begin{proof}[Proof of Theorem~\ref{th:LocalMartingale}]
It is readily seen in \cite[Section~4]{kupiainen2017integrability} that for every fixed $t$, $G(\alpha+i\beta;t)$ is well defined and analytic in $(\alpha,\beta)$ since it is complex differentiable in $\alpha+i\beta$.
To prove claim (2), notice that if $|\beta|<Q-\alpha$ and $\alpha>Q-\gamma$, then using Proposition~\ref{prop:MartingaleLocal},
\begin{equation}
\left|\frac{\partial G(\alpha+i\beta;t)}{\partial t}\right|
\end{equation}
decays exponentially as $t$ goes to infinity, thus proving the convergence.
To prove claim (3), notice that the exponential convergence rate above is locally uniform in $(\alpha,\beta)$. Since for every $t$, $G(\alpha+i\beta;t)$ is analytic in $(\alpha,\beta)$, local uniform convergence of analytic functions yields analyticity of the limit. As the limit is the same when $\beta=0$, the limit thus defined is indeed the only analytic continuation of local Liouville correlation function from real parameter $\alpha$ to the region
\begin{equation}
\mathcal{R}^{M}_{loc}=\mathcal{R}_{loc}\cap\{\alpha>Q-\gamma\}.
\end{equation}
\end{proof}
\begin{rema}[Extention of martingale method]
One can ask if we can define probabilistically the analytic continuation of local Liouville correlation function to $\mathcal{R}_{loc}$ only using this martingale method. We make several comments on this question: it is possible but the actual proof is much more involved than the current article. In fact, one can look at higher derivatives of the function $G(\alpha+i\beta;t)$ in $\mathcal{R}_{loc}$ but they are in general ill-defined. To renormalize this explosion phenomenon one has to analyse carefully the situation and do a precise Taylor expansion. Getting rid of the first terms in this Taylor expansion appropriately will allow us to push the analytic continuation beyond the region $\mathcal{R}^{M}_{loc}$ and eventually cover the whole region $\mathcal{R}_{loc}$ recursively. However obtaining this Taylor expansion is technically much more involved and will be investigated seperately.
\end{rema}
\section{Local study II: stopping time method}\label{sec:StoppingLocal}
We follow notations from Section~\ref{sec:LocalSetup}. In this section we focus on the region
\begin{equation}
\mathcal{R}^{T}_{loc}=\mathcal{R}_{loc}\cap\{\alpha<Q-\frac{\gamma}{2}\}.
\end{equation}
The main property that we need in this regime is the following:
\begin{prop}[Property: stopping time region]\label{prop:StoppingLocal}
If $\alpha<Q-\frac{\gamma}{2}$, then the positive random measure
\begin{equation}
M_{1}=\int_{0}^{T_1}e^{\gamma(B_r-(Q-\alpha)r)}Z_rdr
\end{equation}
is bounded in $L^1$, i.e.
\begin{equation}
\mathbb{E}\left[\int_{0}^{T_1}e^{\gamma(B_r-(Q-\alpha)r)}Z_rdr\right]<\infty.
\end{equation}
\end{prop}
\begin{proof}
By positivity of the measure we have always the following bound
\begin{equation}
M_{1}\leq\int_{0}^{\infty}e^{\gamma(B_r-(Q-\alpha)r)}Z_rdr
\end{equation}
and when we map the right hand side conformally to the disk, we obtain
\begin{equation}
\int_{B(0,1)}\frac{1}{|z|^{\alpha\gamma}}M_\gamma(dz)
\end{equation}
where $M_\gamma(dz)$ denotes the GMC measure associated to some log-correlated field $X$ in $B(0,2)$.
By the study of generalized Seiberg bound (see \cite[Section~3.4]{david2016liouville}), the last quantity has finite positive moment up to
\begin{equation}
p_c(\alpha)=\frac{4}{\gamma^2}\wedge\frac{2}{\gamma}(Q-\alpha)
\end{equation}
which is strictly bigger than $1$ with our assumption on $\alpha$.
\end{proof}
\subsection{Notations and Brownian motion decorrelation}
Now we proceed towards the study of the function
\begin{equation}
G(\alpha+i\beta;T_N)=\mathbb{E}\left[e^{i\beta B_{T_N}+\frac{\beta^2}{2}T_N}e^{-\mu\int_{0}^{T_N}e^{\gamma(B_r-(Q-\alpha)r)}Z_r dr}\right].
\end{equation}
First we want to use Markov property to cut the time interval $[0,T_N]$ into subintervals with disjoint interiors
\begin{equation}
[0,T_N]=\coprod\limits_{i=0}^{N-1}[T_i,T_{i+1}]
\end{equation}
such that by definition of stopping time for a drifted Brownian motion, we can decorrelate the Brownian motion into independent components on each interval. More precisely, if for each $i$, $B^{i}(s)$ denotes the fluctuation of the Brownian motion $B$ from time $T_i$ until $T_{i+1}$, i.e.
\begin{equation}
B^{i}(s)\coloneqq B(s+T_i)-B(T_i),\quad s\leq T_{i+1}-T_i,
\end{equation}
then if $i\neq j$, the Brownian paths $B^{i}$ and $B^{j}$ are mutually independent.
By Markov property we consequently write $G(\alpha+i\beta;T_N)$ as
\begin{equation}
G(\alpha+i\beta;T_N)=\mathbb{E}\left[\prod_{i=0}^{N-1}\left(e^{i\beta (B_{T_{i+1}}-B_{T_{i}})+\frac{\beta^2}{2}(T_{i+1}-T_{i})}e^{-\mu e^{-\gamma i}\int_{0}^{T_{i+1}-T_{i}}e^{\gamma(B^i_r-(Q-\alpha)r)}Z_{r+T_{i}}dr}\right)\right].
\end{equation}
If we let $M_{i}$ denote the positive random measure
\begin{equation}
M_{i}\coloneqq \int_{0}^{T^{i}_1}e^{\gamma(B^i_r-(Q-\alpha)r)}Z_{r+T_{i}}dr
\end{equation}
with $T^{i}_{1}=T_{i+1}-T_i$, the stopping time associated to the drifted Brownain motion $B^{i}_r-(Q-\alpha)r$ at level $-1$, then the above expression can be condensed into
\begin{equation}
G(\alpha+i\beta;T_N)=\mathbb{E}\left[\prod_{i=0}^{N-1}\left(e^{i\beta (B_{T_{i+1}}-B_{T_{i}})+\frac{\beta^2}{2}(T_{i+1}-T_{i})}e^{-\mu e^{-\gamma i}M_{i}}\right)\right].
\end{equation}
Notice that for every $i$, $M_{i}$ is distributed as $M_{1}$ in Proposition~\ref{prop:StoppingLocal}, but they are correlated (only in the lateral noise part $Z_r$). The correlation between $M_{i},M_{j}$ decays as the corresponding time intervals $[T_{i},T_{i+1}],[T_{j},T_{j+1}]$ separate apart.
\subsection{Preliminary observation}
The first observation we will make is that terms where the measure $M_{i}$ is not too big can be bounded uniformly by some global constant $C$ which is locally uniform in $(\alpha,\beta)$.
More precisely, let $N$ be arbitrary and $J\subset[|0,N-1|]$ any subset of the set
\begin{equation}
[|0,N-1|]\coloneqq \{0,1,\dots,N-1\}
\end{equation}
and consider the function, where the measure $M_i$ is truncated at level $e^{\frac{\gamma i}{2}}$,
\begin{equation}
g^{J}(\alpha+i\beta;T_N)=\mathbb{E}\left[\prod_{i\in J}\left(e^{i\beta (B_{T_{i+1}}-B_{T_{i}})+\frac{\beta^2}{2}(T_{i+1}-T_{i})}e^{-\mu e^{-\gamma i}(M_{i}\wedge e^{\frac{\gamma i}{2}})}\right)\right].
\end{equation}
We claim that
\begin{prop}\label{prop:SmallIndicator}
With the above notation, there exists some constant $C_0$ independent of $N$ and $J$ and locally uniform in $(\alpha,\beta)$ such that the following bound holds:
\begin{equation}
\left|g^{J}(\alpha+i\beta;T_N)\right|\leq C_0.
\end{equation}
Nonetheless the bound $C_0$ depends on $\mu$ and we will give an estimate in a corollary.
\end{prop}
\begin{proof}
Write
\begin{equation}
e^{-\mu e^{-\gamma i}(M_{i}\wedge e^{\frac{\gamma i}{2}})}=1-\left(1-e^{-\mu e^{-\gamma i}(M_{i}\wedge e^{\frac{\gamma i}{2}})}\right)
\end{equation}
and expand all the product over subsets $K\subset J$. By Markov property, for any subset $K\subset J$ the corresponding term in the product expansion
\begin{equation}
\mathbb{E}\left[\prod_{i\in J}\left(e^{i\beta (B_{T_{i+1}}-B_{T_{i}})+\frac{\beta^2}{2}(T_{i+1}-T_{i})}\right)\prod_{i\in K}\left(1-e^{-\mu e^{-\gamma i}(M_{i}\wedge e^{\frac{\gamma i}{2}})}\right)\right]
\end{equation}
can be bounded in absolute value by (with $C$ independent of $\mu$)
\begin{equation}
\mathbb{E}\left[\prod_{i\in K}\left(e^{\frac{\beta^2}{2}(T_{i+1}-T_{i})}\left(1-e^{-\mu e^{-\frac{\gamma i}{2}}}\right)\right)\right]=C^{\#(K)}\prod_{i\in K}\left(1-e^{-\mu e^{-\frac{\gamma i}{2}}}\right)
\end{equation}
and summing up we get an upper bound for $g^{J}(\alpha+i\beta;T_N)$ in absolute value:
\begin{equation}\label{eq:SmallIndicatorUpperBound}
\prod_{i\in J}\left(1+C\left(1-e^{-\mu e^{-\frac{\gamma i}{2}}}\right)\right)\leq\prod_{i\in\mathbb{N}}\left(1+C\left(1-e^{-\mu e^{-\frac{\gamma i}{2}}}\right)\right).
\end{equation}
To see that this is bounded independent of $I$, using $1-e^{-x}\leq x$ for $x\geq 0$ and write
\begin{equation}
\prod_{i\in\mathbb{N}}\left(1+C\left(1-e^{-\mu e^{-\frac{\gamma i}{2}}}\right)\right)\leq\prod_{i\in\mathbb{N}}\left(1+C\mu e^{-\frac{\gamma i}{2}}\right)<\infty.
\end{equation}
This finishes the proof.
\end{proof}
We study the upper bound when $\mu$ varies and record the following corollary:
\begin{coro}\label{cor:SmallIndicatorMu}
Let $c>0$ and write $\mu=\mu_0 e^{\gamma c}$ for $\mu_0>0$. The bound $C_0(\mu)$ in Proposition~\ref{prop:SmallIndicator} can be bounded above by
\begin{equation}
(1+C)^{\frac{c+1}{2}} C_0(\mu_0)
\end{equation}
\end{coro}
\begin{proof}
Let us restart from the Equation~\eqref{eq:SmallIndicatorUpperBound} and use the slighly better bound
\begin{equation}
1-e^{-x}\leq x\wedge 1
\end{equation}
for $x\geq 0$. Now if $2c=k\in\mathbb{N}$ then
\begin{equation}
\prod_{i\in\mathbb{N}}\left(1+C\left(1-e^{-\mu e^{-\frac{\gamma i}{2}}}\right)\right)=\prod_{i=0}^{k-1}\left(1+C\left(1-e^{-\mu e^{-\frac{\gamma i}{2}}}\right)\right)\prod_{i\in\mathbb{N}}\left(1+C\left(1-e^{-\mu_0 e^{-\frac{\gamma i}{2}}}\right)\right)
\end{equation}
where we recognize the upper bound
\begin{equation}
(1+C)^{k} C_0(\mu_0).
\end{equation}
For general $c$, a similar argument yields the upper bound
\begin{equation}
(1+C)^{\frac{c+1}{2}} C_0(\mu_0)
\end{equation}
as desired.
\end{proof}
Consequently, we will decompose the function $G$ with respect to the subset $J\in[|0,N-1|]$ where the measures $M_i$ are extremely large. Indeed, we can write
\begin{equation}
G(\alpha+i\beta;T_N)=\sum\limits_{J\subset[|0,N-1|]}G^{J}(\alpha+i\beta;T_N)
\end{equation}
where $G(J)$ denotes the configuration where measures with indices in $J$ are extremely large:
\begin{align}
G^{J}(\alpha+i\beta;T_N)\coloneqq\mathbb{E}\bigg[&\prod_{i\in J}\left(e^{i\beta (B_{T_{i+1}}-B_{T_{i}})+\frac{\beta^2}{2}(T_{i+1}-T_{i})}\left(e^{-\mu e^{-\gamma i}M_{i}}-e^{-\mu e^{-\frac{\gamma i}{2}}}\right)\mathbf{1}_{\{M_{i}>e^{\frac{\gamma i}{2}}\}}\right)\\
\times&\prod_{i\in J^c}\left(e^{i\beta (B_{T_{i+1}}-B_{T_{i}})+\frac{\beta^2}{2}(T_{i+1}-T_{i})}e^{-\mu e^{-\gamma i}(M_{i}\wedge e^{\frac{\gamma i}{2}})}\right)\bigg]\nonumber
\end{align}
with abused notation $J^c$ denoting the complement of $J$ in $[|0,N-1|]$.
Finding a proper way to translate the large deviation phenomenon
\begin{equation}
\prod_{i\in J}\mathbf{1}_{\{M_{i}>e^{\frac{\gamma i}{2}}\}}
\end{equation}
for multiple indices simulteneously will be the focus of the following sections.
\subsection{Cutting annuli and Kahane's inequality}
One crucial step in the proof using stopping time method is to decorrelate the random measures $\{M_i\}_{i\in\mathbb{N}}$ in a suitable way. The classical Gaussian comparaison inequalities allow one to do so to some extend, usually under some weak correlation assumption. We can obtain such assumption if we force the time intervals of $M_i$ to be far away mutually. Indeed,
\begin{prop}\label{prop:WeakCorrelation}
Let $\epsilon$ be arbitraty positive constant. There exists some constant $\delta$ only depending on $\epsilon$ such that if $|t-s|>\delta$, for all $(\theta,\theta')\in[0,2\pi]^2$,
\begin{equation}
\mathbb{E}[Y(s,\theta)Y(t,\theta')]<\epsilon.
\end{equation}
\end{prop}
\begin{proof}
Recall the lateral noise correlation function
\begin{equation}
\mathbb{E}[Y(s,\theta)Y(t,\theta')]=\ln\frac{e^{-s}\vee e^{-t}}{|e^{-s}e^{i\theta}-e^{-t}e^{i\theta'}|}.
\end{equation}
Without loss of generosity, let us suppose $s<t$, and the above expression equals
\begin{equation}
\mathbb{E}[Y(t,\theta)Y(s,\theta')]=\ln\frac{1}{|1-e^{-(t-s)e^{i(\theta-\theta')}}|}\leq\ln\frac{1}{1-e^{-(t-s)}}.
\end{equation}
In particular, any $\delta$ such that
\begin{equation}
\delta>\ln\frac{1}{1-e^{-\epsilon}}
\end{equation}
will satisfy the claim.
\end{proof}
To apply this observation to our case, we have to specify a certain kind of subset of $[|0,N-1|]$ (measurable with respect to the Brownian motion $B$) where the above weak correlation assumption is met. We will then apply Kahane's inequality to decorrelate the random measures on this subset.
\begin{defi}[Cutting annuli]\label{def:CuttingAnnuli}
Let $B$ be the Brownian motion corresponding to the radial part of the GFF $X$ as defined above. We call a subset $I=\{i_1<i_2<\dots<i_k\}\subset[|0,N-1|]$ a $(B,\delta)$-cutting annuli if the following condition is met:
\begin{equation}
\forall i_j\in I,\quad |T_{i_{j+1}}-T_{i_j+1}|>\delta.
\end{equation}
Otherwise put, for all different indices $i\neq j$, the intervals $[i_i,i_{i+1}]$ and $[i_j,i_{j+1}]$ are seperated by distance $\delta$ if $(i_i,i_j)\in I^2$.
\end{defi}
It is useful to reformulate the above definition in another way. For any indice $i\in\mathbb{N}$, define the $(B,\delta)$-precutting image of $i$, denoted by $\overline{i^{(B,\delta)}}$, in the following way:
\begin{defi}[Precutting image]\label{def:PreCuttingImage}
Let $i\in\mathbb{N}$. We define the $(B,\delta)$-precutting image of $i$ to be the only indice $\overline{i^{(B,\delta)}}$ such that
\begin{equation}
T_{i}-1\in[T_{\overline{i^{(B,\delta)}}}, T_{\overline{i^{(B,\delta)}}+1}).
\end{equation}
If such indice does not exist, then we define $\overline{i^{(B,\delta)}}$ to be $-1$. We drop the indices on $(B,\delta)$ and simply write $\overline{i}$ if there is no ambiguity.
\end{defi}
Then for every ordered set $I=\{i_1<i_2<\dots\}\subset\mathbb{N}$, $I$ is a $(B,\delta)$-cutting annuli if and only if
\begin{equation}
i_1<\overline{i_2}<i_2<\overline{i_3}<\dots
\end{equation}
is satisfied. If this is the case, we define the an extended version of $I$ in the following manner:
\begin{defi}[Extended cutting annuli]\label{def:ExtendedCuttingAnnuli}
Let $I=\{i_1<i_2<\dots\}\subset\mathbb{N}$ be an ordered set. We will use the notation
\begin{equation}
I\leftrightarrow(B,\delta)
\end{equation}
to denote that $I$ is a $(B,\delta)$-cutting annuli. If it is the case, define
\begin{equation}
\overline{I^{(B,\delta)}}\coloneqq\{\overline{i_1}<i_1<\overline{i_2}<i_2<\overline{i_3}<\dots\}
\end{equation}
to be the $(B,\delta)$-extended version of $I$. We drop the indice $(B,\delta)$ when there is no ambiguity.
\end{defi}
Now we are able to state the Gaussian decorrelation inequality that we will be depending on.
\begin{lemm}[Kahane's decorrelation]\label{lem:KahaneDecorrelation}
Let
\begin{equation}
(\alpha,\beta)\in\mathcal{R}^{T}_{loc}=\mathcal{R}_{loc}\cap\{\alpha<Q-\frac{\gamma}{2}\}.
\end{equation}
Consider the function
\begin{equation}
\overline{G^{I}}(\alpha+i\beta;T_N)\coloneqq\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{i\in I}\left(e^{\frac{\beta^2}{2}(T_{i+1}-T_{i})}\mathbf{1}_{\{M_{i}>e^{\frac{\gamma i}{2}}\}}\right)\right].
\end{equation}
Then when $\delta>\ln\frac{1}{1-e^{-\frac{1}{8\gamma}}}$, we have the following bound for any $q>\frac{(Q-\alpha)^2}{(Q-\alpha)^2-\beta^2}$:
\begin{equation}
|\overline{G^{I}}(\alpha+i\beta;T_N)|\leq C_0\times C^{\#(I)}\prod\limits_{i\in I}e^{-\frac{\gamma i}{4q}}
\end{equation}
where $C_0, C$ are constants independent of $N$ and $I$ and locally uniform in $(\alpha,\beta)$.
\end{lemm}
\begin{proof}
If
\begin{equation}
\mathbf{1}_{I\leftrightarrow(B,\delta)}=0
\end{equation}
then the claim is trivial. We thus assume that
\begin{equation}
I\leftrightarrow(B,\delta)
\end{equation}
holds. With this assumption in mind, if $Y_i(t,\theta)$ denotes the underlying log-correlated fields of the measures $M_{i}$, i.e. if
\begin{equation}
M_{i}=\int_{0}^{T^{i}_1}e^{\gamma(B^{i}_r-(Q-\alpha)r)}\left(\frac{1}{2\pi}\int_{0}^{2\pi}e^{\gamma Y_i(r,\theta)}d\theta\right)dr
\end{equation}
then by Proposition~\ref{prop:WeakCorrelation}, if $i,j\in I$ and $i\neq j$ then for all $s,t\in[0,T^{i}_1]\times[0,T^{j}_1]$,
\begin{equation}
\mathbb{E}[Y_i(s,\theta)Y_j(t,\theta')]<\epsilon
\end{equation}
where $\epsilon>0$ is such that
\begin{equation}
e^{-\delta}+e^{-\epsilon}=1.
\end{equation}
We want to use this information and apply the diagonal decorrelation inequality of Kahane-Slepian, Lemma~\ref{lem:KahaneDiagonal}. Consider the following modified GMC measures with a collection of i.i.d. standard Gaussian variables denoted by $\widetilde{N},(N_i)_{i\in\mathbb{N}}$.:
\begin{equation}
\overline{M_i}=\int_{0}^{T^{i}_1}e^{\gamma(B^{i}_r-(Q-\alpha)r)}\left(\frac{1}{2\pi}\int_{0}^{2\pi}e^{\gamma (Y_i(r,\theta)+\sqrt{\epsilon}N_i)}d\theta\right)dr=e^{\gamma\sqrt{\epsilon}N_i}M_i
\end{equation}
\begin{equation}
\widehat{M_i}=\int_{0}^{T^{i}_1}e^{\gamma(B^{i}_r-(Q-\alpha)r)}\left(\frac{1}{2\pi}\int_{0}^{2\pi}e^{\gamma (\widetilde{Y_i}(r,\theta)+\sqrt{\epsilon}\widetilde{N})}d\theta\right)dr=e^{\gamma\sqrt{\epsilon}\widetilde{N}}\widetilde{M_i}
\end{equation}
where $(\widetilde{Y_i})_{i\in\mathbb{N}}$ are mutually independent copies of $(Y_i)_{i\in\mathbb{N}}$ (similarly for $\widetilde{M_i}$). Since for all $(i,j)$,
\begin{equation}
\mathbb{E}[(Y_i(s,\theta)+\sqrt{\epsilon}N_i)(Y_j(t,\theta')+\sqrt{\epsilon}N_j)]\leq\mathbb{E}[(\widetilde{Y_i}(s,\theta)+\sqrt{\epsilon}\widetilde{N})(\widetilde{Y_j}(t,\theta')+\sqrt{\epsilon}\widetilde{N})]
\end{equation}
and $Y+\sqrt{\epsilon}N,\widetilde{Y}+\sqrt{\epsilon}\widetilde{N}$ have the same variance, Kahane-Slepian inequality Lemma~\ref{lem:KahaneDiagonal} yields, for every set $I\leftrightarrow(B,\delta)$
\begin{equation}
\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{i\in I}\overline{M_i}\right]\leq\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{i\in I}\widehat{M_i}\right]
\end{equation}
which is equivalent to
\begin{equation}
\mathbb{E}\left[\prod_{i\in I}e^{\gamma\sqrt{\epsilon}N_i}\right]\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{i\in I}M_i\right]\leq\mathbb{E}\left[\prod_{i\in I}e^{\gamma\sqrt{\epsilon}\widetilde{N}}\right]\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{i\in I}\widetilde{M_i}\right].
\end{equation}
We retain the weaker relation
\begin{equation}
\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{i\in I}M_i\right]\leq e^{\#(I)^2\frac{\gamma^2}{2}\epsilon}\prod_{i\in I}\mathbb{E}\left[M_i\right].
\end{equation}
To finish the prove we use Markov inequality on the indicator and the property in the stopping time region from Proposition~\ref{prop:StoppingLocal}: remember
\begin{equation}\label{eq:StoppingTimeFirstBound}
\left|\overline{G^{I}}(\alpha+i\beta;T_N)\right|=\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{i\in I}\left(e^{\frac{\beta^2}{2}(T_{i+1}-T_{i})}\mathbf{1}_{\{M_{i}>e^{\frac{\gamma i}{2}}\}}\right)\right]
\end{equation}
and use Hölder inequality with a pair $\frac{1}{p}+\frac{1}{q}=1$, $0<p<\frac{(Q-\alpha)^2}{\beta^2}$, i.e.
\begin{equation}
q>\frac{(Q-\alpha)^2}{(Q-\alpha)^2-\beta^2}
\end{equation}
to have
\begin{equation}
\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{i\in I}\left(e^{\frac{\beta^2}{2}(T_{i+1}-T_{i})}\mathbf{1}_{\{M_{i}>e^{\frac{\gamma i}{2}}\}}\right)\right]\leq\mathbb{E}\left[\prod_{i\in I}e^{\frac{p\beta^2}{2}(T_{i+1}-T_{i})}\right]^{\frac{1}{p}}\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{i\in I}\mathbf{1}_{\{M_{i}>e^{\frac{\gamma i}{2}}\}}\right]^{\frac{1}{q}}
\end{equation}
By the assumption on $p$, for each $i\in I$ the expectation is bounded uniformly
\begin{equation}
\mathbb{E}\left[e^{\frac{p\beta^2}{2}(T_{i+1}-T_{i})}\right]<C
\end{equation}
with some constant $C$ locally uniform in $(\alpha,\beta)$. Next, use
\begin{equation}
\mathbf{1}_{\{M_{i}>e^{\frac{\gamma i}{2}}\}}\leq e^{-\frac{\gamma i}{2}}M_i
\end{equation}
and the above decorrelation inequality yields
\begin{equation}\label{eq:StoppingTimeHolder}
\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{i\in I}\mathbf{1}_{\{M_{i}>e^{\frac{\gamma i}{2}}\}}\right]^{\frac{1}{q}}\leq\prod_{i\in I}e^{-\frac{\gamma i}{2}}\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{i\in I}M_i\right]^{\frac{1}{q}}\leq e^{\#(I)^2\frac{\gamma^2}{2}\epsilon}\prod_{i\in I}e^{-\frac{\gamma i}{2}}\prod_{i\in I}\mathbb{E}\left[M_i\right].
\end{equation}
Now by the assumption on the stopping time region Proposition~\ref{prop:StoppingLocal}, we can bound
\begin{equation}
\prod_{i\in I}\mathbb{E}\left[M_i\right]<C^{\#(I)}
\end{equation}
with some constant $C$ locally uniform in $(\alpha,\beta)$. Also, for
\begin{equation}
\epsilon<\frac{1}{8\gamma}\quad \text{or equivalently}\quad\delta>\ln\frac{1}{1-e^{-\frac{1}{8\gamma}}}
\end{equation}
we have (for $\#(I)\geq 2$, the case $\#(I)<2$ can be treated similarly)
\begin{equation}
e^{\#(I)^2\frac{\gamma^2}{2}\epsilon}\prod_{i\in I}e^{-\frac{\gamma i}{2}}\leq \prod_{i\in I}e^{-\frac{\gamma i}{4}}
\end{equation}
since the indices in $I$ are positive and distinct.
Finally Equations~\eqref{eq:StoppingTimeFirstBound} and \eqref{eq:StoppingTimeHolder} yield, with any $\delta>\ln\frac{1}{1-e^{-\frac{1}{8\gamma}}}$ and $q>\frac{(Q-\alpha)^2}{(Q-\alpha)^2-\beta^2}$,
\begin{equation}
\left|\overline{G^{I}}(\alpha+i\beta;T_N)\right|\leq C^{\#(I)}\prod_{i\in I}e^{-\frac{\gamma i}{4q}}
\end{equation}
with $C$ locally uniform in $(\alpha,\beta)$: this is the claim.
\end{proof}
\begin{rema}
Notice that this upper bound does not depend on $\mu$, since this parameter does not appear in the definition of $\overline{G^{I}}(\alpha+i\beta;T_N)$.
\end{rema}
\subsection{Renewal theory and boundedness of local Liouville correlation function}
In this subsection we will prove another estimation:
\begin{lemm}\label{lem:RenewalBound}
We use the same notations as above and let
\begin{equation}
(\alpha,\beta)\in\mathcal{R}^{T}_{loc}=\mathcal{R}_{loc}\cap\{\alpha<Q-\frac{\gamma}{2}\}.
\end{equation}
We claim that for any fixed $\delta$,
\begin{equation}
|G(\alpha+i\beta;T_N)|\leq C_0\times\sum\limits_{J\in[|0,N-1|]}C^{\#(J)}\left|\overline{G^{J}}(\alpha+i\beta;T_N)\right|
\end{equation}
where $C_0, C$ are constants independent of $N$ and $J$ and locally uniform in $(\alpha,\beta)$.
\end{lemm}
As a corollary of the two previous lemmas we get the following consequence:
\begin{lemm}[Boundedness of local Liouville correlation function]\label{lem:Boundedness}
Let
\begin{equation}
(\alpha,\beta)\in\mathcal{R}^{T}_{loc}=\mathcal{R}_{loc}\cap\{\alpha<Q-\frac{\gamma}{2}\}.
\end{equation}
There exists some constant $C$ independent of $N$ and locally uniform in $(\alpha,\beta)$ such that
\begin{equation}
|G(\alpha+i\beta;T_N)|<C.
\end{equation}
\end{lemm}
\begin{proof}
Combining the two previous lemmas, for any $\delta>\ln\frac{1}{1-e^{-\frac{1}{8\gamma}}}$ and $q>\frac{(Q-\alpha)^2}{(Q-\alpha)^2-\beta^2}$,
\begin{equation}
|G(\alpha+i\beta;T_N)|\leq C_0\times\sum\limits_{J\in[|0,N-1|]}C^{\#(J)}\prod_{j\in J}e^{-\frac{j\gamma}{4q}}\leq C_0\times\prod\limits_{j\in\mathbb{N}}(1+Ce^{-\frac{j\gamma}{4q}})<C
\end{equation}
and one checks that all constants are locally uniform in $(\alpha,\beta)$.
\end{proof}
The rest of this subsection will be devoted to the proof of Lemma~\ref{lem:RenewalBound}.
\begin{proof}[Proof of Lemma~\ref{lem:RenewalBound}]
The proof depends heavily on the definition of cutting annuli.
We first define a map $\Phi_{(B,\delta)}$, depending on $(B,\delta)$, from the subsets of $\mathbb{N}$ to the sets of cutting annuli.
\begin{defi}[Reduction map to cutting annuli]
Let $J=\{j_1<\dots<j_k\}$ be any finite ordered subset of $\mathbb{N}$. The reduction map $\Phi_{(B,\delta)}$ maps $J$ to another ordered subset $I=\{i_1<\dots<i_l\}$ of $\mathbb{N}$ in the following way (from the largest element downward):
\begin{itemize}
\item $i_l=j_k$;
\item For $j=l-1$, $i_j=\sup\{a\in J; T_{a+1}+\delta<T_{i_{j+1}}=T_{i_l}\}$.
\item Recursively so for $j=l-2$, $j=l-3$ until the algorithm stops.
\end{itemize}
Then $l$ is defined as the number of elements in $I$. Observe the two following properties of this map:
\begin{enumerate}
\item $I$ is a subset of $J$;
\item $I$ is a $(B,\delta)$-cutting annuli.
\end{enumerate}
\end{defi}
Intuitively speaking, this is a greedy way of choosing elements from $J$ downward in such a way that the time intervals they represent are seperated apart by distance $\delta$. Although $I$ depends on $(B,\delta)$, we don't keep that in the notation when there is no ambiguity.
We will use an important observation on the preimage by $\Phi_{(B,\delta)}$ of a cutting annuli. Recall the Definition~\ref{def:ExtendedCuttingAnnuli} that couples any $(B,\delta)$-cutting annuli $I$ with an extended version $\overline{I}$.
\begin{prop}[Preimage by reduction map]\label{prop:ReductionPreimage}
Let $I$ be a $(B,\delta)$-cutting annuli an $\overline{I}$ the $(B,\delta)$-extended version of $I$. Suppose that $i_0=-\infty$ and
\begin{equation}
\overline{I}=\{\overline{i_1}<i_1<\overline{i_2}<i_2<\overline{i_3}<\dots<\overline{i_l}<i_l\}.
\end{equation}
Then $J\in\Phi_{(B,\delta)}^{-1}(I)$ if and only if:
\begin{enumerate}
\item $I\subset J$, i.e. for all $i_j\in I$, $i_j\in J$.
\item Let $i_k\in I$, if $i_{k}<j<\overline{i_{k+1}}$, then $j\neq J$.
\end{enumerate}
For $\overline{i_{k}}\leq j<i_{k}$, $j$ can be in $J$ or not: this gives exactly the cardinal of $\Phi_{(B,\delta)}^{-1}(I)$, although we don't need that
number later.
\end{prop}
We see from this observation that $\Phi_{(B,\delta)}$ is surjective, say from the set of subsets of $[|0,N-1|]$ to the set of cutting annuli of $[|0,N-1|]$: this is because for any cutting annuli $I$, $I$ itself is contained in the preimage $\Phi_{(B,\delta)}^{-1}(I)$.
Now let us return to the study of the function $G$. Remember
\begin{equation}
G(\alpha+i\beta;T_N)=\sum\limits_{J\subset[|0,N-1|]}G^{J}(\alpha+i\beta;T_N)
\end{equation}
with $G(J)$ denotes the configuration where measures with indices in $J$ are extremely large
\begin{align}
G^{J}(\alpha+i\beta;T_N)\coloneqq\mathbb{E}\bigg[&\prod_{i\in J}\left(e^{i\beta (B_{T_{i+1}}-B_{T_{i}})+\frac{\beta^2}{2}(T_{i+1}-T_{i})}\left(e^{-\mu e^{-\gamma i}M_{i}}-e^{-\mu e^{-\frac{\gamma i}{2}}}\right)\mathbf{1}_{\{M_{i}>e^{\frac{\gamma i}{2}}\}}\right)\\
\times&\prod_{i\in J^c}\left(e^{i\beta (B_{T_{i+1}}-B_{T_{i}})+\frac{\beta^2}{2}(T_{i+1}-T_{i})}e^{-\mu e^{-\gamma i}(M_{i}\wedge e^{\frac{\gamma i}{2}})}\right)\bigg]\nonumber
\end{align}
Now rewrite the sum in the following way, depending on the cutting annuli $I=\Phi_{(B,\delta)}(J)$:
\begin{equation}\label{eq:ReductionSumAnnuli}
G(\alpha+i\beta;T_N)=\sum\limits_{I\subset[|0,N-1|]}\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\sum\limits_{J\in\Phi_{(B,\delta)}^{-1}(I)}h_J\right]
\end{equation}
where the notation $h_J$ stands for
{\small
\begin{equation}
\prod_{i\in J}\left(e^{i\beta (B_{T_{i+1}}-B_{T_{i}})+\frac{\beta^2}{2}(T_{i+1}-T_{i})}\left(e^{-\mu e^{-\gamma i}M_{i}}-e^{-\mu e^{-\frac{\gamma i}{2}}}\right)\mathbf{1}_{\{M_{i}>e^{\frac{\gamma i}{2}}\}}\right)\prod_{i\in J^c}\left(e^{i\beta (B_{T_{i+1}}-B_{T_{i}})+\frac{\beta^2}{2}(T_{i+1}-T_{i})}e^{-\mu e^{-\gamma i}(M_{i}\wedge e^{\frac{\gamma i}{2}})}\right).
\end{equation}
}
Let us fix an ordered cutting annuli $I=\{i_1<i_2<\dots<i_l\}\subset[|0,N-1|]$ and focus on controlling
\begin{equation}
\left|\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\sum\limits_{J\in\Phi_{(B,\delta)}^{-1}(I)}h_J\right]\right|.
\end{equation}
By Proposition~\ref{prop:ReductionPreimage} on the preimage $\Phi_{(B,\delta)}^{-1}(I)$ above, this quantity is equal to
\begin{align}
\bigg|\mathbb{E}\bigg[\mathbf{1}_{I\leftrightarrow(B,\delta)}
&\prod_{k=1}^{l}\left(e^{i\beta (B_{T_{i_k+1}}-B_{T_{i_k}})+\frac{\beta^2}{2}(T_{i_k+1}-T_{i_k})}\left(e^{-\mu e^{-\gamma i_k}M_{i_k}}-e^{-\mu e^{-\frac{\gamma i_k}{2}}}\right)\mathbf{1}_{\{M_{i_k}>e^{\frac{\gamma i_k}{2}}\}}\right)\\
&\prod_{k=1}^{l}\prod_{i_{k-1}<j<\overline{i_k}}\left(e^{i\beta (B_{T_{j+1}}-B_{T_{j}})+\frac{\beta^2}{2}(T_{j+1}-T_{j})}e^{-\mu e^{-\gamma j}(M_{j}\wedge e^{\frac{\gamma j}{2}})}\right)\nonumber\\
&\prod_{k=1}^{l}\prod_{\overline{i_k}\leq j<i_k}\left(e^{i\beta (B_{T_{j+1}}-B_{T_{j}})+\frac{\beta^2}{2}(T_{j+1}-T_{j})}e^{-\mu e^{-\gamma j}M_{j}}\right)\bigg]\bigg|\nonumber
\end{align}
and with the argument of Proposition~\ref{prop:SmallIndicator} applied to the products of terms with small indicators, we can majorize this by
\begin{equation}
C_0\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{k=1}^{l}\left(e^{\frac{\beta^2}{2}(T_{i_k+1}-T_{i_k})}\mathbf{1}_{\{M_{i_k}>e^{\frac{\gamma i_k}{2}}\}}\right)\prod_{k=1}^{l}\prod_{\overline{i_k}\leq j<i_k}\left(e^{\frac{\beta^2}{2}(T_{j+1}-T_{j})}\right)\right]
\end{equation}
with $C_0$ locally uniform in $(\alpha,\beta)$, which is
\begin{equation}
C_0\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{k=1}^{l}\left(e^{\frac{\beta^2}{2}(T_{i_k+1}-T_{i_k})}\mathbf{1}_{\{M_{i_k}>e^{\frac{\gamma i_k}{2}}\}}\right)\prod_{k=1}^{l}\left(e^{\frac{\beta^2}{2}(T_{i_k}-T_{\overline{i_k}})}\right)\right].
\end{equation}
Now we are not far from being able to apply Kahane's decorrelation Lemma~\ref{lem:KahaneDecorrelation}: it suffices to get rid of the last product term in the above expression. By a standard argument in renewal theory (see Appendix~\ref{app:renewal}), we know that
\begin{equation}
\mathbb{E}\left[e^{\frac{\beta^2}{2}(T_{i_k}-T_{\overline{i_k}})}\right]<C
\end{equation}
for some $C$ locally uniform in $(\alpha,\beta)$, so that by Markov property of the Brownian motion one arrives at the following bound
\begin{equation}
\left|\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\sum\limits_{J\in\Phi_{(B,\delta)}^{-1}(I)}h_J\right]\right|\leq C_0\times C^{\#(I)}\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\prod_{k=1}^{l}\left(e^{\frac{\beta^2}{2}(T_{i_k+1}-T_{i_k})}\mathbf{1}_{\{M_{i_k}>e^{\frac{\gamma i_k}{2}}\}}\right)\right]
\end{equation}
which is with the previous notations
\begin{equation}
\left|\mathbb{E}\left[\mathbf{1}_{I\leftrightarrow(B,\delta)}\sum\limits_{J\in\Phi_{(B,\delta)}^{-1}(I)}h_J\right]\right|\leq C_0\times C^{\#(I)}\left|\overline{G^{I}}(\alpha+i\beta;T_N)\right|.
\end{equation}
Summing up over all possibilities $I\subset[|0,N-1|]$ by Equation~\eqref{eq:ReductionSumAnnuli}, we have finally
\begin{equation}
|G(\alpha+i\beta;T_N)|\leq C_0\times\sum\limits_{I\in[|0,N-1|]}C^{\#(I)}\left|\overline{G^{I}}(\alpha+i\beta;T_N)\right|.
\end{equation}
This finishes the proof of Lemma~\ref{lem:RenewalBound}.
\end{proof}
\subsection{Convergence and analyticity of the limit}
In order to prove the analyticity of the limit, it is convenient to consider a slightly modified version of regularization of local Liouville correlation function in the following way:
\begin{equation}
\mathcal{G}(\alpha+i\beta;T_N)=\mathbb{E}\left[e^{(\alpha+i\beta)B_{T_N}-\frac{(\alpha+i\beta)^2}{2}T_N}e^{-\mu\int_{0}^{\infty}e^{\gamma(B_r-Qr)}Z_r dr}\right].
\end{equation}
This function is well-defined and analytic in $(\alpha,\beta)$ for all $N$ as long as $|\beta|<Q-\alpha$.
We will show that $\mathcal{G}$ and $G$ are exponentially close as $N$ goes to infinity. More precisely:
\begin{lemm}[Modified regularization of local Liouville correlation function]\label{lem:ModifiedG}
We claim the following properties on $\mathcal{G}(\alpha+i\beta;T_N)$ and $G(\alpha+i\beta;T_N)$ in the region
\begin{equation}
\mathcal{R}^{T}_{loc}=\mathcal{R}_{loc}\cap\{\alpha<Q-\frac{\gamma}{2}\}.
\end{equation}
\begin{enumerate}
\item There exists some constants $C,\eta>0$ locally uniform in $(\alpha,\beta)$ such that
\begin{equation}
\left|\mathcal{G}(\alpha+i\beta;T_N)-G(\alpha+i\beta;T_N)\right|\leq Ce^{-\eta N};
\end{equation}
\item There exists some constants $C,\eta>0$ locally uniform in $(\alpha,\beta)$ such that
\begin{equation}
\left|G(\alpha+i\beta;T_{N+1})-G(\alpha+i\beta;T_N)\right|\leq Ce^{-\eta N};
\end{equation}
\item There exists some constants $C,\eta>0$ locally uniform in $(\alpha,\beta)$ such that
\begin{equation}
\left|\mathcal{G}(\alpha+i\beta;T_{N+1})-\mathcal{G}(\alpha+i\beta;T_N)\right|\leq Ce^{-\eta N}.
\end{equation}
\end{enumerate}
\end{lemm}
Let us first show how this can be used to finish the proof of Theorem~\ref{th:LocalStoppingTime}.
\begin{proof}[Proof of Theorem~\ref{th:LocalStoppingTime}]
By property (3) in Lemma~\ref{lem:ModifiedG}, local uniform convergence of $\mathcal{G}(\alpha+i\beta;T_N)$ yields that the limit of $\mathcal{G}(\alpha+i\beta;T_N)$ as $N$ goes to infinity is analytic in $(\alpha,\beta)$ in the region
\begin{equation}
\mathcal{R}^{T}_{loc}=\mathcal{R}_{loc}\cap\{\alpha<Q-\frac{\gamma}{2}\}.
\end{equation}
Furthurmore, by (1) of Lemma~\ref{lem:ModifiedG}, the limit is the same as the limit of $G(\alpha+i\beta;T_N)$ when $N$ goes to infinity. This proves that the limit
\begin{equation}
G^{T}(\alpha+i\beta)\coloneqq \lim_{N\to\infty}G(\alpha+i\beta;T_N)
\end{equation}
is well-defined and analytic in $\mathcal{R}^{T}_{loc}$.
\end{proof}
We now study Lemma~\ref{lem:ModifiedG}.
\begin{proof}[Proof of Lemma~\ref{lem:ModifiedG}]
The proof uses the same techniques as in the boundedness arguments.
Let us start with (2). We apply the same technique as before using the cutting annuli. The proof is almost identical with Lemma~\ref{lem:RenewalBound}, except that the set of cutting annuli over which we sum must contain the time segment with indice $N$, which gives rise to an exponential factor $e^{-\frac{N\gamma}{4q}}$.
More precisely, let us consider
\begin{align}
&G(\alpha+i\beta;T_{N+1})-G(\alpha+i\beta;T_N)\\
={}&\mathbb{E}\left[\prod_{i=0}^{N-1}\left(e^{i\beta (B_{T_{i+1}}-B_{T_{i}})+\frac{\beta^2}{2}(T_{i+1}-T_{i})}e^{-\mu e^{-\gamma i}M_{i}}\right)\times e^{i\beta (B_{T_{N+1}}-B_{T_{N}})+\frac{\beta^2}{2}(T_{N+1}-T_{N})}\left(1-e^{-\mu e^{-\gamma N}M_{N}}\right)\right]\nonumber.
\end{align}
We now apply the same argument with the indicators: in particular we add the indicator
\begin{equation}
e^{-\mu e^{-\gamma N}M_{N}}=e^{-\mu e^{-\gamma N}(M_N\wedge e^{\frac{\gamma N}{2}})}+\left(e^{-\mu e^{-\gamma N}M_N}-e^{-\mu e^{-\frac{\gamma N}{2}}}\right)\mathbf{1}_{\{M_N> e^{\frac{\gamma N}{2}}\}}
\end{equation}
to control the term with $M_N$.
For the part with the small indicator, using the same method as in boundedness Lemma~\ref{lem:Boundedness} we obtain the bound
\begin{equation}
Ce^{-\frac{\gamma N}{2}}
\end{equation}
which is of the desired form.
Now we control the part with the other indicator. That is
\begin{equation}
\left|\mathbb{E}\left[\prod_{i=0}^{N-1}\left(e^{i\beta (B_{T_{i+1}}-B_{T_{i}})+\frac{\beta^2}{2}(T_{i+1}-T_{i})}e^{-\mu e^{-\gamma i}M_{i}}\right)\mathbf{1}_{\{M_N> e^{\frac{\gamma N}{2}}\}}\left(e^{-\mu e^{-\gamma N}M_N}-e^{-\mu e^{-\frac{\gamma N}{2}}}\right)\right]\right|.
\end{equation}
This quantity has been studied before using cutting annuli, it contributes at most
\begin{equation}
Ce^{-\frac{\gamma N}{4q}},
\end{equation}
thus the proof of claim (2).
For (1), the proof is similar to (2) except that we replace the time segment $[T_N,T_{N+1}]$ by $[T_N,\infty]$, on which the drift is $-Q$ instead of $-(Q-\alpha)$. However, one can check that the rest of the proof goes in the same way as in (2).
Finally, (3) is a direct consequence of (1) and (2).
\end{proof}
\begin{rema}[Extention of stopping time method]
One can ask if we can define probabilistically the analytic continuation of local Liouville correlation function to $\mathcal{R}_{loc}$ only using this stopping time method. It is indeed possible modulo a slight modification by looking at lower-than-1 fractional moment of the GMC measure for real $\alpha$ in the disk $\mathbb{D}$. By the study of the generalized Seiberg bound (see \cite{david2016liouville,huang2018liouville}) some small moments of this GMC measure always exist as long as $\alpha<Q$, and one can adapt easily appropriately elements in the above proof to this case. Nonetheless, the nature of the stopping time method yields convergence in a somewhat weaker sense than the martingale method, although their limits coincide and yield the correct analytic continuation of (local) Liouville correlation function in $\mathcal{R}^{T}_{loc}\cap\mathcal{R}^{M}_{loc}$.
\end{rema}
\section{Proof of the main theorem}\label{sec:ProofGeneral}
To pass from the local versions Theorem~\ref{th:LocalMartingale} and Theorem~\ref{th:LocalStoppingTime} to the main Theorem~\ref{th:MainTheorem}, one uses the independence property of radial decomposition Remark~\ref{rem:Independence}. Recall that we are interested in studying the limit as $\mathbf{t}$ goes to infinity of
\begin{equation}
G(\bm{\alpha+i\beta},\mathbf{z};\mathbf{t})\coloneqq\mathbb{E}\left[\prod_{j=1}^{n}e^{(\alpha_j+i\beta_j)X_{r_j}(z_j)-\frac{(\alpha_j+i\beta_j)^2}{2}t_j}M_{\gamma}(C_t)^{-s}\right]
\end{equation}
where $r_j=e^{-t_j}$, $s=\frac{\sum_j(\alpha_j+i\beta_j)-2Q}{\gamma}$ and $C_t=\mathbb{C}-B(0,e^{-t})$. Remember that under the Seiberg bound, we have
$\alpha_j<Q$ for all $j$ and $\Re(s)>0$.
Without lost of generosity, we can suppose $z_1=0$ and $|z_j|\geq 2$ for $j\neq 1$: one can get this configuration by applying a deterministic conformal map if necessary. Now we fix all $t_j$ for $j\neq 1$ and study the limit as $t_1$ goes to infinity: we write
\begin{equation}
M_\gamma(\mathbb{C})=M_\gamma(\mathbb{D})+M_\gamma(\mathbb{C}\backslash\mathbb{D})
\end{equation}
and remark that the second summand is independent of $\{X_{r_1}(0)\}_{r_1\leq 1}$. In view of Remark~\ref{rem:Independence}, one verifies that all the calculations in the local case can be done in this general case (since other Brownian motions are also independent of $\{X_{r_1}(0)\}_{r_1\leq 1}$ and do not enter in the calculation) as soon as $\Re(s)>0$ (which implies $\Re(s)+1>0$). This is similar to the local case with some technical modifications, we provide a detailed sketch of proof in Appendix~\ref{app:NonLocal}.
Thus, the limit when $t_1$ goes to infinity when $\{|\beta_1|<Q-\alpha_1\}$ is well-defined and analytic. One can successively apply the same procedure to $t_2,\dots,t_n$ and define the limit function $G(\bm{\alpha+i\beta};\mathbf{z})$ in this manner. This limit is analytic in the region $\mathcal{R}$ since it is the uniform local limit of analytic functions by Hartog's theorem: this completes our proof of Theorem~\ref{th:MainTheorem}.
\begin{rema}[Beyond the region $\mathcal{R}$]
It is natural to study the same question beyond the region $\mathcal{R}$. In view of the DOZZ formula \cite{dorn1992correlation,zamolodchikov1996conformal,kupiainen2017integrability}, it is reasonable to believe that the $n$-point correlation function should admit an analytic continuation to the whole complex set $\mathbb{C}^n$ as a meromorphic function, i.e. analytic modulo some poles.
It seems however that our approaches are limited to the region $\mathcal{R}$. With the martingale method, this limitation stems from the fact that, although we use a strong freezing estimate for the real part $\alpha$, our control on the complex argument part is still poor. Indeed, we always perform at some stage of our proof the following crude estimate
\begin{equation}
\forall t\in\mathbb{R}_+,\quad \left|e^{i\beta B_t}\right|\leq 1
\end{equation}
which does not take into account of the angular compensation between different phases of $B_t$. The freezing estimate for the real part translates to the condition $|\beta|<Q-\alpha$.
In the same way, with the stopping-time method, we need the condition
\begin{equation}
\mathbb{E}\left[e^{\frac{\beta^2}{2}T_1}\right]<\infty
\end{equation}
which translates again to $|\beta|<Q-\alpha$.
It is curious that both methods yield the same constraint and indicate that the region $\mathcal{R}$ is the maximal set on which analytic continuation of Liouville correlation function is possible with the current techniques for attacking this problem. It seems that at this stage that a rigorous justification for a purely probabilistic definition of the analytic continuation of $n$-point Liouville correlation functions on the Riemann sphere beyond the region $\mathcal{R}$, using directly the original approach of \cite{david2016liouville}, remains a mathematical challenge.
\end{rema}
|
\section{Introduction}\label{sec:intro}
To understand complex biological systems such as tissues and cells, extensive knowledge of molecular interactions and mechanisms is necessary. However, an important part of understanding biological complexity is also mathematical modeling, which allows researchers to investigate connections between cellular processes and to develop hypotheses for the design of new experiments.
Jacob and Monod \cite{jacobmonod} were the first to present a model of the regulation of the synthesis of a structural protein. In this model enzyme levels are regulated at the level of transcription. Specific proteins are produced which repress the transcription of the DNA to its product (mRNA -- messenger ribonucleic acid), which is translated into $\beta$-galactosidase,
an enzyme for degradation of galactose into simple sugars.
Shortly after Jacob and Monod, Goodwin \cite{goodwin} proposed the first mathematical model of a more complex biological system, a genetic oscillator. The simplest formulation of the Goodwin model involves a single gene that represses its own transcription via a negative feedback loop and uses three variables, $ x, y$ and $z$, where
$x$ denotes the quantity of mRNA, $y$ stands for the quantity of the repressor protein, and $z$ is the quantity of the product, which acts as a corepressor and generates the feedback loop by negative control of mRNA production:
\begin{equation}\label{eq:goodwin}
\deriv{x}=\frac{k_1}{k_2+z^n}-k_3x\quad
\deriv{y}=k_4x-k_5y\quad
\deriv{z}=k_6y-k_7z
\end{equation}
All synthesis and degradation rates in the model (represented by coefficients $k_1$ to $k_7$) are linear, with the exception of the repression, which takes the form of a sigmoidal Hill curve.
Here $ n$ denotes the Hill exponent, which may be interpreted in biological systems as the number of ligand molecules that a receptor can bind. At the level of transcriptional regulation, this can be explained by cooperative binding of the repressor protein to DNA (formation of protein-DNA complexes). It has been demonstrated by Griffith \cite{griffith} that limit cycle oscillations can only be obtained when $n>8$, which is unrealistic in terms of transcriptional regulation, where Hill exponents are rarely higher than 3 or 4.
A repressilator is a network of several genes and can be thought of as an extension of the Goodwin oscillator, which is a one-gene repressilator linked by mutual repression in a cyclic topology. Models of cycles of 2--5 genes have first been studied by Fraser and Tiwari \cite{frasertiwari}, while the first experimental implementation of a 3-gene repressilator in a biological system along with a refined model was demonstrated by Elowitz and Leibler \cite{elowitzleibler}. Let $X_i$ denote the quantity of mRNA and $Y_i$ the quantity of the repressor protein and
let $ \alpha_{0}, \alpha$ and $ \beta$
represent the transcription rate of a repressed promoter, the maximal transcription rate of a free promoter and the ratio of protein and mRNA decay rate, respectively.
Then the model is given by the equations:
\be \label{eq:leoleib}
\begin{aligned}
\deriv{X_{i}}=& \alpha_0 + \frac \alpha {1 + Y_{i-1}^n} - X_{i}\\
\deriv{Y_{i}}=& -\beta(Y_{i}-X_{i}) \quad(i=1,2,3),
\end{aligned}
\ee
where the indices 0 and 3 are identified. (Let us note that
Elowitz and Leibler write $j$ instead of $i-1$, still they speak about 6 equations.
However, if $i$ and $j$ run independently, then we have $3\times 3+3=12$ equations.
Our modification is also in accordance with our model below.)
In the paper mentioned above Elowitz and Leibler also determine the unique positive stationary point, and the parameters when the stationary point looses its stability.
They map part of the parameter space, and find oscillations \emph{numerically}.
In the Goodwin model, undamped oscillations can only occur when repression is accomplished by the co-repressor $Z$ and never directly by the protein $Y$ \cite{griffith}, probably due to the increased time delay. In the cyclic repressilator by Elowitz and Leibler, oscillations can occur without co-repressors and for Hill exponents $n$ as low as 2, which is more applicable to biological systems. It also takes into consideration the production of mRNA with a constant rate.
A theoretical solution for the introduction of non-linearity to non-co\-op\-er\-a\-tive biological systems by using transcription factors, where the same proteins are able to repress one gene and activate another gene has been proposed by M\"uller et al. \cite{mullerhofbauerendlerflammwidderschuster} and Widder et al. \cite{widdermaciasole}. Tyler et al. \cite{tylershiuwalton} continue the work by \cite{mullerhofbauerendlerflammwidderschuster} with biologically less restrictive assumptions.
However, such transcription factors are extremely rare in nature and would also be hard to design by directed evolution.
Recently, Lebar et al. \cite{lebarbezeljakgolobjeralakaduncpirsstrazarvuckozupancicbencinaforstnericgaberlonzaricmajerleoblaksmolejerala} have shown that non-linearity can be introduced into a biological system, by introduction of non-cooperative repressors in combination with activators, competing for binding to the same DNA sequence, thus creating a positive feedback loop. In principle, positive feedback loops could be introduced---based on the same DNA binding domain---to build functional repressilator circuits, consisting of non-cooperative repressors.
The above described oscillator circuit was experimentally constructed using three natural repressor proteins, the TetR, LacI and CI repressors. However, construction of functional biological circuits using such natural repressors requires fine-tuning due to their diverse biochemical properties. Furthermore, the low number of well-characterized natural repressor proteins does not enable construction of multiple circuits in a single cell, a fact that may
support the use of stochastic models, cf. e.g. \cite{aranyitoth,erdilente,tothnagypapp}. With the developments in the field of synthetic biology in the recent years, the use of designable repressors has become more and more frequent
\cite{qilarsongilbertdoudnaweissmanarkinlim
kianibealebrahimkhanihuhhallxieliweiss
lohmuellerarmelsilve
garglohmuellersilverarmel
congzhoukuocunniffzhang
gaberlebarmajerlesterdobnikarbencinajerala
lebarbezeljakgolobjeralakaduncpirsstrazarvuckozupancicbencinaforstnericgaberlonzaricmajerleoblaksmolejerala
lebarjerala,nissimprtlifridkinperezpineralu}. Such repressors can be designed to bind any DNA sequence due to their modular structure, which can be exploited to eliminate interactions with the cells' genome. Furthermore, they can be designed in almost unlimited numbers and the biochemical properties of individual repressors are very similar, making construction and modeling of synthetic circuits easier. However, the main disadvantage of designable repressors is that they are monomeric, meaning that their binding to DNA is non-cooperative and the Hill exponent $n$ is equal to 1. Under those conditions, the above described models are not expected to produce oscillations. This poses a challenge of introducing non-linearity in complex biological systems, consisting of such repressors.
Equations describing the model of the repressilator by Elowitz and Leibler with only two variables are easy to handle. However, addition of activators to the model increases the number of variables, thus expanding the complexity of the model. Mathematical analysis of systems of equations with a large number of variables is harder,
and can be investigated using deterministic approach based on ordinary differential equations (ODEs) with kinetics which can be either of the mass action type or other, and use the qualitative theory of ordinary differential equations to find bistability, oscillation etc., or calculating solutions numerically. The stochastic description \cite[Chapter 5]{erditoth},\cite[Chapter 10]{tothnagypapp} or \cite{erdilente} usually does not allow to make symbolic calculations because of the complexity of the model. However, in this case one also may turn to the computer to do simulations \cite{sipostotherdi,nagypapptoth}.
In this work, we compare deterministic mathematical models of three different repressilator topologies based on non-cooperative repressors, which can be implemented in biological systems based on designed DNA binding domains such as zinc fingers, TALEs or dCas9/CRISPR fused to activation or repression domains. The models are simplified and consider reactions only at the protein level. The concentration of each repressor and activator over time is described in a separate equation in a system of equations. In the 3D model, we perform the singular point analysis of
the 3-variable equation system for the basic repressilator topology, consisting of 3 repressors. In the 6D models we expand the complexity by addition of 3 variables, representing activators.
The study of the system is non-trivial since there are no efficient methods
for determining singular points of polynomial or rational systems
of ODEs of high dimension and depending on parameters. We perform
our analysis using the combinations of modern symbolic algorithms of computer
algebra systems \Mathematica\ \cite{mathematica} and \Singular\ \cite{deckergreuelpfisterschonemann}, which has not yet been covered in the literature and represents a novel approach in analysis of biological circuits.
Extensive theoretical studies have already been done on the 3D repressilator circuit.
\cite{kuznetsovafraimovich} only treat the special case $\alpha_0=0$ of our model.
In a nonlinear model such a seemingly slight difference may cause qualitative differences.
They also treat saturable degradation, i.e. cases when instead of
$-k_{\mathrm{deg}}x$ one has a term $\frac{-k_{\mathrm{deg}}x}{1+x}.$
They have shown the connection between the evolution of the oscillatory solution and formation of a heteroclinic cycle at infinity.
\cite{dilao} also deals with the $\alpha_0=0$ case, but he derives the usual nonlinear term starting from a mass action model, and using the Michaelis--Menten type approximation.
That author is mainly interested in models with delay.
\cite{guantespoyatos} again assumes $\alpha_0=0,$ and the rational functions are such that both the denominator and the numerator are second degree polynomials. The paper contains no general mathematical statements, only numerical simulations. On the other hand, the mathematically correct paper \cite{mullerhofbauerendlerflammwidderschuster} treats a large class of models including the model by Elowitz and Leibler (but not our models) and give a detailed description of the attractors. Summarizing, none of the models in the literature cover the classes of models we are interested in, and also, the present approach seems to be a novel one from the mathematical point of view and uses models based on recent experiments in synthetic biology.
Note also that \cite{tiggesmarquezlagostellingfussenegger} consider a much more complicated
process, no formulae can be found in the paper itself.
However, its Supplement contains models, delay, stochastic effects, and no qualitative analysis at all. They estimate the parameters of the model.
\cite{thieffrythomas} use the heuristic ideas (\textit{kinetic logic}) of Thomas without a
mathematical treatment.
\section{A 3D model}\label{sec:3D}
First we model the basic repressilator circuit based on non-cooperative repressors, similar to the Elowitz repressilator. The difference compared to the original repressilator model is that here the Hill exponent $n$ is always equal to 1, due to the non-cooperative nature of the repressors. We consider a symmetrical system, where the biochemical properties of all repressors are similar, as expected with designed transcription factors (and \emph{not} to simplify mathematics). We simplify the system to only consider reactions on the protein level. The variables $x,$ $ y$ and $z$ represent the concentrations of each of the repressors, while the parameters
$ \alpha_{0}, \alpha$ and $k_{\mathrm{deg}}$ represent the rate of protein synthesis when the promoter is repressed, the maximal rate of protein synthesis from the free promoter and protein degradation rate, respectively (Figure \ref{fig:RepTop}).
\begin{figure}[!h]\centering
\includegraphics[width=84mm]{repressilatorCorrected.eps}\\
\caption{The 3D repressilator topology} \label{fig:RepTop}
\end{figure}
We assume equal rates of synthesis and degradation for all three repressor proteins. Then the concentration of each repressor over time is described by the following equations:
\be \label{ss1}
\begin{aligned}
\frac{d x}{dt}=& \alpha_0 + \frac \alpha {1 + z} - k_{\mathrm{deg}} x\\
\frac{d y}{dt}=& \alpha_0 + \frac \alpha {1 + x} - k_{\mathrm{deg}} y\\
\frac{d z}{dt}=& \alpha_0 + \frac \alpha {1 + y} - k_{\mathrm{deg}} z,
\end{aligned}
\ee
To simplify the notation we denote
$$
\ s=\alpha_0,\ b=\alpha, \ g=k_{\mathrm{deg}},
$$
where the parameters $s$, $b$ and $g$ are positive real numbers, and the dot denotes the derivative with respect to time.
With this notation system \eqref{ss1} is written as
\be \label{s1}
\begin{aligned}
\dot x= & s +\frac b{1 + z} - g x \\
\dot y=& s + \frac b{1 + x} - g y \\
\dot z= & s +\frac {b}{1 + y} - g z.
\end{aligned}
\ee
We are interested in the behavior of trajectories of system
\eqref{s1} in the region
$$
D=\{ (x,y,z) : x>0, \ y>0, \ z>0\}.
$$
System \eqref{s1} has two singular points whose coordinates
contain the expression
$u=\sqrt{4 b g + (g + s)^2}.$ With this we have
\be\label{eq:bandineq}
b= \frac{u^2-(g + s)^2}{4 g}{\quad{\rm and}\quad u>g+s.}
\ee
Then the steady states of the system are
$$ A=(x_0,y_0,z_0)=\left(\frac{s- g - u}{ 2 g}, \frac{s- g - u}{ 2 g}, \frac{s- g - u}{ 2 g}\right)
$$
and
$$
B=(x_1,y_1,z_1)=\left( \frac{s+ u-g}{ 2 g},\frac{s+ u-g}{ 2 g},\frac{s+ u-g}{ 2 g}
\right).
$$
The eigenvalues of the Jacobian matrix of system \eqref{s1} at $A$ are
$$
\kappa_1=\frac{2 g u}{g+s-u}, \quad \kappa_{2,3}=-\frac{g (3 g+3 s-u)}{2 (g+s-u)}\pm i \frac{\sqrt{3} g (g+s+u)}{2 (g+s-u)}
$$
and the eigenvalues at $B$ are given by
\be \label{lam}
\lambda_1=
-\frac{2 g u}{g+s+u},\quad \lambda_{2,3}= -\frac{g (3 g+3 s+u)}{2 (g+s+u)}
\pm i \frac{\sqrt{3} g \left( g+s-u\right) }{2 (g+s+u)}.
\ee
We can expect chemically relevant non-trivial behavior of trajectories in the domain $D$
if both singular points of the system are located in $D$.
The necessary and sufficient condition
for this is
\be \label{sem1}
x_0> 0, \ x_1 > 0.
\ee
\delete{
However, $x_0> 0$ alone implies that $s>u+g,$
and this together with $u-g>s$ (see \eqref{eq:bandineq})
shows that these inequalities cannot be fulfilled simultaneously,
thus $A\in D$ can never occur.
Similarly, $x_1<0$ implies $u<g-s,$
and this together with $u>g+s$ (see \eqref{eq:bandineq})
shows that these inequalities cannot be fulfilled simultaneously,
thus $B\in -D$ can never occur.
The only possibility left is that $x_0<0$ and $x_1>0,$
which happens if and only if $s+g<u,$
(which is always the case)
because this inequality automatically implies $s-g<u$ and $g-s<u.$
From $u=\sqrt{4bg + (g+s)^2}$, one gets both $u>g$ and $u>s$ (since $b,g$ and $s$ are positive).
As a consequence, $s-g-u = (s-u) - g < 0$ since $s<u.$ Thus, $A$ can be discarded.
Moreover, $s+u-g = s + (u-g) > 0$ since $u>g,$ so $B$ is always in the domain $D$.
Thus, in this case $B$ is in $D$ and $A$ has negative coordinates.
For the eigenvalues \eqref{lam} of the matrix of the linear approximation
of \eqref{s1} at $B$ we have
$\lambda_1<0$, $\mathrm{Re}\, \lambda_{2,3}<0$, that is,
$B$ is asymptotically stable.
To conclude, in the domain $D=\{ (x,y,z) : x>0, \ y>0, \ z>0\}$ the system can have
only one steady state (point $B$), which is a (locally) asymptotically stable attractor and the trajectories (exponentially) fast approach a neighborhood of the steady state.
In a small neighborhood of it there are damping oscillations, however the amplitude of
oscillations is very small. Why? Because
to obtain oscillations with a large amplitude we need to have at the point $B$ in
$D$ the eigenvalues with $\mathrm{Abs(Re}\, \lambda_{2,3})$ small and $\mathrm{Abs(Im}\, \lambda_{2,3})$
large. However, it can be shown easily that $\mathrm{Abs(Re}\, \lambda_{2,3})<\mathrm{Abs(Im}\, \lambda_{2,3})$ cannot occur.
Thus, this is difficult to achieve in
our system, while it would probably be facilitated in the system with high Hill exponent $n$.
In Fig. \ref{fig:damposc3d} we have chosen the parameters so as to make the difference between
$\mathrm{Abs(Re}\, \lambda_{2,3})$ and $\mathrm{Abs(Im}\, \lambda_{2,3})$ as small as possible.
Fig. \ref{fig:damposc3d} shows the behaviour of the model for a single, specific set of the parameters, but the argument above is symbolic, i.e. valid for all sets of the parameters.
\begin{figure}[!ht]\centering
\includegraphics[width=84mm]{DampOsc3D}\\
\caption{ Damping oscillation (overshoot) in the 3D model. $s=0.3, b=4, g=0.6,$ initial concentrations: $(1,2,2).$}
\label{fig:damposc3d}
\end{figure}
Our calculations above provided an alternative proof of a part of the statement by Allwright \cite{allwright} who has obtained stronger results: he has shown for a class of more general class of models including our one the existence, uniqueness and \emph{global} asymptotic stability of the stationary point.
In order to apply Allwright's results to our model one has to calculate a few quantities, this we will do in the Appendix \ref{subsec:D}.
\section{The forward feedback repressilator 6D model
By a similar principle that was demonstrated to introduce a non-linear response into a non-cooperative system \cite{lebarbezeljakgolobjeralakaduncpirsstrazarvuckozupancicbencinaforstnericgaberlonzaricmajerleoblaksmolejerala}, we devise a more complex repressilator topology (Figure \ref{fig:doscill2}).
The new system consists of the same repressor topology as the 3D model, but also includes three transcriptional activators, binding to the same DNA targets as the repressors. Each of the activators drives the synthesis of itself and of the next repressor in the cycle. This topology can be implemented in biological systems using a set of three DNA binding domains (X, Y, Z), their combination with an activator (a) or a repressor (r) domain and appropriate binding sites within the three operons.
\begin{figure}[!ht]\centering
\includegraphics[width=84mm]{forwardloopCorrected}\\
\caption{A repressilator topology, involving activators, driving the synthesis of the next repressor in the cycle.}
\label{fig:doscill2}
\end{figure}
The new topology therefore includes 6 variables: the concentration---denoted by the corresponding lowercase letters---of the
3 repressors ($X_{r}, Y_{r}$ and $Z_{r}$) and 3 activators ($X_{a}, Y_{a}$ and $ Z_{a}$).
The Hill exponent $n$ is always equal to 1, the parameters $\alpha_{0}, \alpha$ and $k_{\mathrm{deg}}$ represent the rate of protein synthesis when the promoter is repressed, the rate of protein synthesis from the free promoter and protein degradation rate, respectively. We assume equal rates of synthesis and degradation for all repressor and activator proteins. In this case, the protein synthesis rate is considered maximal when the activator is bound to the promoter, so concentration of repressors and activators over time is given as:
\be \label{s6o}
\begin{aligned}
\frac{dx_r}{dt} = & \alpha_0 + \alpha z_a/(1 + z_r + z_a ) - k_{\mathrm{deg}} x_r, \\
\frac{dz_a}{dt} = & \alpha_0 + \alpha z_a/(1 + z_r + z_a ) - k_{\mathrm{deg}} z_a,\\
\frac {dy_r }{dt} = & \alpha_0 + \alpha x_a/(1 + x_r + x_a) - k_{\mathrm{deg}} y_r \\
\frac{dx_a}{dt} = & \alpha_0 + \alpha x_a/(1 + x_r + x_a) - k_{\mathrm{deg}} x_a, \\
\frac {dz_r }{dt} = & \alpha_0 + \alpha y_a/(1 + y_r + y_a) - k_{\mathrm{deg}} z_r,\\
\frac{ dy_a } {dt} = & \alpha_0 + \alpha y_a/(1 + y_r + y_a) - k_{\mathrm{deg}} y_a.
\end{aligned}
\ee
Introducing the notation
$$
x_1=x_{r},\ x_2=
z_{a}, x_3= y_{r}, \ x_4= x_{a},\ x_5= z_{r}, \ x_6= y_{a},
\ s= \alpha_{0}, \ b=\alpha, \ g = k_{\mathrm{deg}}
$$
where $b, g, s >0$ we rewrite system \eqref{s6o} in the form
\be \label{s6}
\begin{aligned}
\dot x_1= &s - g x_1 + \frac{b x_2}{1 + x_2 + x_5}=X(x_1,x_2,x_5)
\\
\dot x_2= & s - g x_2 + \frac {b x_2}{1 + x_2 + x_5}=X(x_2,x_2,x_5)
\\
\dot x_3= & s - g x_3 + \frac{b x_4}{1 + x_1 + x_4}=X(x_3,x_4,x_1)
\\
\dot x_4=&s - g x_4 + \frac{b x_4}{1 + x_1 + x_4}=X(x_4,x_4,x_1)
\\
\dot x_5=& s - g x_5 + \frac{b x_6}{1 + x_3 + x_6}=X(x_5,x_6,x_3)
\\
\dot x_6= & s - g x_6 + \frac{b x_6}{1 + x_3 + x_6}=X(x_6,x_6,x_3)
\end{aligned}
\ee
with
\be\label{X}
X(u,v,w):=s-gu+\frac{bv}{1+v+w}.
\ee
From the first two equations of \eqref{s6} we obtain that for any
steady state $(x_1,x_2,\dots, x_6)$ of the system it should be
that $x_1=x_2$. Similarly, two other pairs of equations
\eqref{s6} yield
that $x_3=x_4, x_5=x_6.$ Thus, the simplified stationary point
equations are:
\begin{eqnarray}
0&= &s - g x_1 + \frac{b x_1}{1 + x_1 + x_5}=X(x_1,x_1,x_5)\label{eq:st1}
\\
0&= & s - g x_3 + \frac{b x_3}{1 + x_1 + x_3}=X(x_3,x_3,x_1)\label{eq:st2}
\\
0&=& s - g x_5 + \frac{b x_5}{1 + x_3 + x_5}=X(x_5,x_5,x_3).\label{eq:st3}
\end{eqnarray}
We first look for steady states of system \eqref{s6}
using the routine \texttt{Solve} of \Mathematica\ and
we find 8 steady states.
Two of them are
\be \label{F}
F=(f,f,f,f,f,f), \quad {\rm where \quad } f= \frac{\sqrt{(b-g+2 s)^2+8 g s}+b-g+2 s}{4
g}
\ee
and
\be
\label{H}
H=(h,h,h,h,h,h), \quad {\rm where \quad } h = -\frac{\sqrt{(b-g+2 s)^2+8 g s}-b+g-2 s}{4
g}.
\ee
However, coordinates of the other steady states are given by long cumbersome expressions which are not convenient to analyse. (If one applies \texttt{Simplify} or even \texttt{FullSimplify} the result of \texttt{LeafCount} is more than thirteen thousand.) Thus, we choose another approach to finish.
Chemically relevant steady states should satisfy the conditions
\begin{multline} \label{6semi}
X(x_1,x_1,x_5)= X(x_3,x_3,x_1)=X(x_5,x_5,x_3)=0,\\ s>0,\ g>0,\
b>0,\ x_1>0,\ x_3>0, \ x_5>0 .
\end{multline}
System \eqref{6semi} is a so-called semi-algebraic system
(since it contains not only algebraic equations $X(x_1,x_1,x_5)= X(x_3,x_3,x_1)=X(x_5,x_5,x_3)=0$, but also inequalities).
Nowadays powerful algorithms to solve such systems have been developed
and implemented in many computer algebra systems.
In particular, in \Mathematica\
the routine
\texttt{ Reduce } can be applied to finding solutions of
semi-algebraic systems.
For algebraic functions
\texttt{Reduce} constructs equivalent purely polynomial systems
and then uses cylindrical algebraic decomposition (CAD) introduced by Collins in \cite{collins} for real domains and Gr\"{o}bner basis methods for complex domains.
To simplify computations we first clear
the denominators on the right hand side of \eqref{eq:st1}--\eqref{eq:st3} obtaining the polynomials
\begin{eqnarray*}
f_1&:=&s + b x_1 - g x_1 + s x_1 - g x_1^2 + s x_5 - g x_1 x_5 \\
f_1&:=&s + s x_1 + b x_3 - g x_3 + s x_3 - g x_1 x_3 - g x_3^2 \\
f_1&:=&s + s x_3 + b x_5 - g x_5 + s x_5 - g x_3 x_5 - g x_5^2 0
\end{eqnarray*}
Solving with \texttt{Reduce} of \Mathematica\
the semi-algebraic system
\begin{equation}
f_1 = f_3 = f_5 = 0, x_1 > 0,
x_3 > 0, x_5 > 0, s > 0, g > 0, b > 0,\label{semiin}
\end{equation}
with respect to $
x_1, x_3, x_5, s, b,
g$
we obtain the solution
\begin{equation}
x_1 > 0, x_1 =x_3 = x_5, s > 0, b > 0,
g =\frac s{x_1} +\frac { b}{1 + x_1 + x_5}.\label{semiout}
\end{equation}
The input command and the output are given in Appendix \ref{subsec:B}. The exact result may slightly differ depending on the version you use, but nevertheless, it always implies the essential relation that $x_1 =x_3 = x_5.$
Solving the last equation for $x_1$ we obtain two solutions:
$$
\frac{\sqrt{(b-g+2 s)^2+8 g s}+b-g+2 s}{4
g}
\mathrm{ and }
\frac{-\sqrt{(b-g+2 s)^2+8 g s}+b-g+2 s}{4
g}.
$$
However in the second case $x_1$ is negative, so the only
steady state whose coordinates satisfy \eqref{6semi} is the point
$F$ defined by \eqref{F}.
Computing the eigenvalues of the Jacobian matrix of system
\eqref{s6} at $F$ we find that they are
\begin{eqnarray*}
\kappa_{1,2,3}&=&-g, \quad \kappa_4 = -g+\frac{b}{(1 + 2 f)^2} \\
\kappa_{5,6}&=&-g+\frac{b(2 + 3f)}{2 (1 + 2 f)^2}\pm i \frac{\sqrt{3} b f}{2 (1 + 2 f)^2},
\end{eqnarray*}
where $f$ is defined by \eqref{F}.
A short calculation shows that all eigenvalues of the Jacobian matrix have negative real parts yielding
that $F$ is asymptotically stable.
Thus, we have proven the following result.
\begin{theorem}
Point $F$ is the only positive stationary point of system \eqref{s6}
and it is asymptotically stable.
\end{theorem}
\section{The backward feedback repressilator 6D model
Due to the absence of oscillations in the above described model we next consider a repressilator topology with activators wired to activate transcription of the previous repressor in the cycle (Figure 4). The notations of the variables and the constants are the same as in the previous 6D model. Therefore, the concentrations of repressors and activators over time are as follows:
\be \label{s6sec}
\begin{aligned}
\dot x_1= & s - g x_1 + b x_4/(1 + x_4 + x_5)
=X(x_1,x_4,x_5)
\\
\dot x_2= & s - g x_2 + b x_4/(1 + x_4 + x_5)
=X(x_2,x_4,x_5)
\\
\dot x_3= & s - g x_3 + b x_6/(1 + x_1 + x_6)
=X(x_3,x_6,x_1)
\\
\dot x_4=&s - g x_4 + b x_6/(1 + x_1 + x_6)
=X(x_4,x_6,x_1)
\\
\dot x_5=& s - g x_5 + b x_2/(1 + x_2 + x_3)
=X(x_5,x_2,x_3)
\\
\dot x_6= & s - g x_6 + b x_2/(1 + x_2 + x_3)
=X(x_6,x_2,x_3)
\end{aligned}
\ee
with $X(u,v,w):=s-gu+\frac{bv}{1+v+w}$,
that is, $X$ is again defined by \eqref{X}, but the right-hand-sides do depend on three variables, differently form the previous case.
\begin{figure}\centering
\includegraphics[width=154mm]{repressilatorrevloopd}\\
\caption{The backward feedback repressilator 6D model: repressilator topology including transcriptional activators, driving the synthesis of the previous repressor in the cycle.}
\label{fig:doscill1}
\end{figure}
\subsection{Steady states of the model}
From \eqref{s6sec} it is easily seen that any stationary point
of \eqref{s6sec} should fulfil
$x_2=x_1, x_4=x_3, x_6=x_5$. Then, similarly as in the case of system \eqref{s6}, computing
with \Mathematica\ we find that the system has singular points $F$ and $H$
defined by \eqref{F} and \eqref{H} and a few other singular
points whose coordinates are given by cumbersome expressions,
which are not suitable for further analysis.
Therefore, again we proceed using the previous ideas.
The chemically relevant
steady states of system \eqref{s6sec} are solutions to the semi-algebraic system
\be \label{6sec_semi}
f_1=f_2 =f_3=0,\ s>0,\ g>0,\ b>0,\ x_1>0, x_3>0, x_5>0
\ee
where
$$
\begin{aligned}
f_1= & s - g x_1 + b x_3 + s x_3 - g x_1 x_3 + s x_5 -
g x_1 x_5 ,\\
f_2= & s + s x_1 - g x_3 - g x_1 x_3 + b x_5 + s x_5 -
g x_3 x_5 , \\
f_3= & s + b x_1 + s x_1 + s x_3 - g x_5 - g x_1 x_5 -
g x_3 x_5
\end{aligned}
$$
(that is, $f_1= X(x_1,x_3,x_5 )(1 + x_3 + x_5),
f_2 = X(x_3,x_5,x_1) (1 + x_1 + x_5), \ f_3 =X(x_5,x_1,x_3) (1 + x_1 + x_3)
$).
But unlike the case of the previous model, we were
able to solve system \eqref{6sec_semi} neither with \texttt{Reduce} nor \texttt{Solve}
of \Mathematica. (\texttt{Solve} provides five roots, most of them in uselessly complicated form.)
It appears that the reason is
that in the previous model the steady states were determined
from the system
$$
X(x_1,x_1,x_5)= X(x_3,x_3,x_1)=X(x_5,x_5,x_3)=0,
$$
where each equation depended only on two variables,
whereas in the present case they are to be determined from the system
$$
X(x_1,x_3,x_5)=X(x_3,x_5,x_1)=X(x_5,x_1,x_3)=0,
$$
where each equation depends on three variables,
so the latter system is more complicated.
To find the steady states of system \eqref{s6sec} we
use the computer algebra system \Singular\ \cite{deckerlaplagnepfisterschonemann,deckergreuelpfisterschonemann}.
We look for solutions of system
\be\label{fss}
f_1(x_1,x_3,x_5,s,g,b)= f_2(x_1,x_3,x_5,s,g,b)=f_3(x_1,x_3,x_5,s,g,b)=0.
\ee
The polynomials $f_1,f_2,f_3$ are polynomials of six variables
with rational coefficients, that is, they are polynomials
of the ring $\Q[ s,b,g,x_1,x_3,x_5]$. In \Singular\ the ring of such
polynomials can be declared as
\centerline{
\texttt{ring r=0,(s,b,g,x1,x3,x5),(lp)}),}
\noindent where \texttt{r} is the name of the ring, $0$ is the characteristic
of the field of rational numbers $\Q$, and \texttt{lp}
means that Gr\"{o}bner basis calculations should
be performed using the lexicographic ordering.
Let $I$ be the ideal generated by $f_1,f_2,f_3$
in $\Q[ s,b,g,x_1,x_3,x_5]$, that is,
\be \label{I}
I=\la f_1,f_2, f_3 \ra.
\ee
The set of solutions of system \eqref{fss} is the variety $V(I)$
of $I$ (the zero set of all polynomials from $I$).
(We give definitions and some facts about polynomial ideals and their varieties in Appendix \ref{subsec:A}.)
Then, applying the routine \texttt{minAssGTZ} of
\cite{deckergreuelpfisterschonemann}, which computes
minimal associate primes of polynomial ideals using
the algorithm of \cite{giannitragerzacharias}, we find that the variety of
$I$ consists of three components,
\begin{equation}
\vv(I)=\vv(I_1)\cup \vv(I_2)\cup \vv(I_3),\label{star}
\end{equation}
where $I_1,I_2,I_3$ are the ideals written under
[1]:, [2]: and [3]:, respectively, in Appendix \ref{subsec:C}.
Since $I_1=\la
x_3-x_5,
x_1-x_5,
2 s x_5+s+b x_5-2 g x_5^2-g x_5\ra
$
it is easily seen that the variety $\vv(I_1)$ consists
of two points $F$ and $H$ defined by \eqref{F} and \eqref{H}, respectively.
From the equations for the third component we have
$s=g=b=0$, so the system degenerates.
However, the polynomials defining the second component are complicated and difficult to analyse, so we are unable to extract useful description of the component
from these polynomials.
Fortunately, there is a slightly different way to treat
the problem of solving system \eqref{fss}. Namely, we can treat
polynomials
$$ f_1(x_1,x_3,x_5,s,g,b), \
f_2(x_1,x_3,x_5,s,g,b), \ f_3(x_1,x_3,x_5,s,g,b)
$$
as polynomials of $ x_1, x_2, x_3$ depending on parameters
$s,g,b$ (which is in agreement with the meaning of $s,g,b$ in differential system \eqref{s6sec}).
To do so, we declare the ring as
\centerline{\texttt{ring r=(0,s,b,g),(x1,x3,x5),(lp)},}
\noindent where \texttt{r} is the name of the ring, \texttt{(0,s,b,g)}
means that the computations should be performed in the field
of characteristic \texttt{0} and \texttt{s,b,g} should be treated as parameters, and,
as above, \texttt{lp} means that Gr\"{o}bner basis calculations should
be performed using the lexicographic ordering.
Computing with \texttt{minAssGTZ} the minimal associate primes of the ideal
$
J=\la f_1,f_2, f_3 \ra
$ (which looks as $I$ but now it is considered as the ideal of the ring
$
\Q(s,b,g)[x_1,x_3,x_5]
$)
we obtain that they are
$$
J_1=\la h_1, h_2, h_3\rangle
$$
with
\be \label{h123}
\begin{aligned}
h_1=& (2 s g^3+b g^3+g^4) x_5^3+(2 s^2 g^2+2 s b g^2+5 s g^3+2 b^2 g^2+2 b g^3+2 g^4) x_5^2\\
&+(-2 s^3 g-3 s^2 b g-s^2 g^2-3 s b^2 g-2 s b g^2+2 s g^3-b^3 g+g^4) x_5\\
&+(-2 s^4-4 s^3 b-5 s^3 g-5 s^2 b^2-8 s^2 b g-4 s^2 g^2-3 s b^3-7 s b^2 g\\
&-5 s b g^2-s g^3-b^4-2 b^3 g-2 b^2 g^2-b g^3),\\
h_2= &(2 s b g+b^2 g+b g^2) x_3+(-2 s g^2-b g^2-g^3) x_5^2\\
&+(s b g-2 s g^2-b^2 g-g^3) x_5\\
&+(2 s^3+4 s^2 b+3 s^2 g+4 s b^2+6 s b g+s g^2+2 b^3+2 b^2 g+2 b g^2)\\
h_3= &(2 s b g+b^2 g+b g^2) x_1+(2 s g^2+b g^2+g^3) x_5^2\\
&+(s b g+2 s g^2+2 b^2 g+b g^2+g^3) x_5\\
&+(-2 s^3-2 s^2 b-3 s^2 g-2 s b^2-s b g-s g^2)
\end{aligned}
\ee
and
$$ J_2=\la
2 g x_5^2+(g-2 s-b) x_5-s,
x_1-x_5,
x_3-x_5\ra.
$$
So the variety of the ideal consists of two components
$$
{\bf V}(J)={\bf V}(J_1)\cup {\bf V}(J_2).
$$
Clearly, the variety ${\bf V}(J_2) $ considered as a variety in
$\mathbb{R}^3$ consists of two points
$F$ and $H$ defined by \eqref{F} and \eqref{H}.
Chemically relevant steady states in the component ${\bf V}(J_1)$ are determined from the
semi-algebraic system
\be \label{sas6}
b>0, \ g>0, \ s>0,x_1>0, x_3>0 \ x_5>0, h_1=0, h_2=0, h_3=0.
\ee
Solving system \eqref{sas6} with \texttt{Reduce} we find that it has no solution
(the command \texttt{Reduce} returns \texttt{False} as the output).
Using the analysis performed above we can prove the following result.
\begin{theorem}
The only steady state of system \eqref{s6} with positive coordinates is the point $F$ defined by \eqref{F}.
\end{theorem}
\begin{proof}
As we have shown above the only point from the variety
$\vv (J)$ satisfying the condition
\be \label{csas6}
b>0, \ g>0, \ s>0,x_1>0, x_3>0 \ x_5>0
\ee
is the point $F$ defined by \eqref{F}.
However, the complete set of steady states of system
\eqref{s6sec} is determined from the variety $\vv(I)$
of the ideal $I$ defined by \eqref{I}.
Thus to prove the theorem it is sufficient to show that
$\vv (I)$ is a subset of $\vv(J)$.
The first components of $\vv (I)$ and the second component of $\vv(J)$ are the same,
the third component of $\vv (I)$ is the variety $\vv(I_3)$ of the ideal
$I_3=\la s, b , g\ra$. Obviously,
if
$s=b=g=0$ then all polynomials $h_1, h_2, h_3$ vanish,
that means, $\vv (I_3)$ is subset of $\vv (J)$.
So, we have to compare the second
components of the decompositions of $\vv (I)$ and $\vv(J)$,
that is, $\vv(H)$ and $\vv(G)$,
where $H=\la h_1, h_2, h_3 \ra $ with
$h_1, h_2, h_3$ defined by \eqref{h123} and
$G=\la g_1, \dots, g_{11}\ra$
where by $g_1, \dots, g_{11}$ we denote polynomials
of the second minimal associate prime given in Appendix \ref{subsec:C}.
First, with the command \texttt{std} of \Singular\
we compute Gr\"{o}bner bases of $H$ and $G$, denoting them
$H_s$ and $G_s$, respectively.
Then with \texttt{reduce} of \Singular\ we check
that $H\subset G$ (since \texttt{reduce($H_s$,$G_s$)}
returns $0$) yielding $\vv(H)\subset \vv(G)$.
$\square$
\end{proof}
\begin{remark}
Applying the command \texttt{reduce($G_s,H_s$)} we obtain
that $H\subsetneq G$ yielding $\vv(H)\subset \vv(G)$,
and $\vv(H)$ is a strict subset of $\vv(G)$ (as varieties in $\mathbb{C}^6$).
We also can find the precise difference of $\vv(H)$ and $\vv(G)$,
the set $\vv(H)\setminus \vv(G)$.
To this end, we use the fact that
$$\vv(H)\setminus \vv(G)= \vv(H:G), $$
where $H:G$ is the quotient of ideals $H$ and $G$ (see e.g. \cite{coxlittleshea} or \cite{romanovskishafer}). In \Singular\
we compute
the ideal $H:G$ with the command \texttt{quotient(H,G)}
and then with \texttt{minAssGTZ} we compute the minimal associate primes
of $H:G$ finding that the variety of $H:G$ consists of 5 components:
1) $g=
s^2+s b+b^2= 0 $
2)
$b
=2 s+g=0$
3)
$3 b-g
=3 s+2 g=0$
4) $
b
=g x_5-s=0$,
5)
$b
=g x_5+s+g=0$
Thus, we see that the varieties $\vv(H)$ and $\vv(G)$ differ only for
the set of parameters which are not relevant for our study: $g=0$ in case 1),
$b=0$ in cases 2), 4), 5) and in case 3) $s=-2/3 g$ which is impossible
since $s$ and $g$ are positive.
\end{remark}
\subsection{Stability of the positive steady state}
To study the stability properties of system \eqref{s6sec} near the point $F$ we compute the characteristic polynomial $p$ of the Jacobian matrix of system
\eqref{s6sec} at $F$ and we find that it is given as
\begin{eqnarray*}
p(u) &=&
\frac{(g + u)^3 }{(1 + 2 f)^6}
\left(-b+g(1+2f)^2 + u (1+2f)^2\right) \\
&& \left(u^2 (1 + 2 f)^4 + u (1 + 2 f)^2 (b + 2 g (1 + 2 f)^2)\right. \\
&+& \left.g^2 (1 + 2 f)^4 + b g (1 + 2 f)^2 + b^2 (1 + 3 f + 3 f^2)\right).
\end{eqnarray*}
where $f$ is defined by \eqref{F}. In order to prove that all the roots of the characteristic polynomial have a negative real part it is enough to show that
$-b+g(1+2f)^2>0,$ which can be easily proven, e.g.\ using \texttt{Reduce}.
To sum up, for any $s,b,g >0$ all roots of $p$ have negative real parts.
Therefore, we have proven the following statement.
\begin{theorem}
The only positive steady state $F$ of system \eqref{s6sec} is asymptotically stable.
\end{theorem}
We can get a more precise conclusion about the eigenvalues
of $F$. Computing the discriminant of the second degree factor of the above polynomial we find that it is
$-3 b^2 (1 + 2 f)^6<0,$
which means that the polynomial $p$ always has a pair of complex conjugate eigenvalues.
Thus, the matrix of the linear approximation of
\eqref{s6sec} at $F$ always has four negative real eigenvalues
and a pair of complex conjugate eigenvalues with negative real parts.
Consequently Hopf bifurcation is not possible in the system.
We can expect to observe strong damping oscillations near the steady
states if the absolute value of the real parts of the complex
eigenvalues are much less than their imaginary parts.
However our numerical experiments show that the situation appears to be
just the opposite: the real parts of the complex
eigenvalues are much larger than their imaginary parts.
So we can observe only oscillations which quickly goes to the steady state
(see Fig. \ref{fig:7}).
\begin{figure}[!ht]\centering
\includegraphics[width=84mm]{rs6new}\\
\caption{Damping oscillation in the 6D model. $s=1, b=10, g=0.2,$ initial concentrations: $(25, 23, 25, 30.5, 21, 30)$}
\label{fig:7}
\end{figure}
\section{Excluding Hopf Bifurcations by Fully Algorithmic Methods}
We also looked for Hopf bifurcations in the 3D and 6D models using the software
package \QeHopf\ which uses the method of the semi-algebraic
characterization of Hopf bifurcation described in \cite{elkahouiweber}
(the package is available by request to the authors).
To detect Hopf bifurcation in the models we first generate from the symbolic description of the respective ordinary differential equation a first-order formula
in the language of ordered fields, where our domain is the real numbers.
Specifically,
for a parametrized vector field $f(u,x)$ and the autonomous
ordinary differential system associated with it this semi-algebraic description
can be expressed by the following first-order
formula:
\begin{eqnarray}
\lefteqn{\exists x (f_{1}(u,x)=0 \logicaland f_{2}(u,x)=0
\logicaland \cdots \logicaland f_{n}(u,x)=0}
\nonumber \\
& & \logicaland a_{n}>0 \logicaland \Delta_{n-1}(u,x)=0 \logicaland
\Delta_{n-2}(u,x)>0 \logicaland \cdots \logicaland \Delta_{1}(u,x)>0).
\end{eqnarray}
In this formula $a_n$ is $(-1)^{n}$ times the determinant
of the Jacobian matrix $Df(u,x)$, and $\Delta_{i}(u,x)$ is the $i^{\rm
th}$ Hurwitz determinant of the characteristic polynomial of the
same matrix $Df(u,x)$.
Constraints on parameters are added, and for the rational systems we are considering
one is using the common numerators (adding the condition of non-vanishing denominators).
\QeHopf\ is implemented in \Maple, and the input for the 3D model is as follows:
\begin{footnotesize}
\begin{verbatim}
PP:=diff(x(t),t)= s-g*x(t)+b/(1+z(t)) ;
QQ:=diff(y(t),t)= s-g*y(t) +b/(1+x(t));
RR:=diff(z(t),t)= s-g*z(t)+b/(1+y(t));
fcns:={x(t), y(t) ,z(t)};
params:=[s, g, b];
paramcondlist:=[s>0, g>0, b>0];
funccondlist:=[x(t)>0, y(t)>0, z(t)>0];
DEHopfexistence({PP,QQ,RR}, fcns, params, funccondlist, paramcondlist);
\end{verbatim}
\end{footnotesize}
For the 3D model the generated first-order formula is as follows
\begin{footnotesize}
\begin{verbatim}
informula :=
ex (vv3, ex (vv2, ex (vv1, ( ( ( 0 < vv1 and 0 < vv2 ) and 0 < vv3 ) and
( ( ( ( ( ( ( s > 0 and b > 0 and g > 0 and
-g*vv1*vv3-g*vv1+s*vv3+b+s = 0 ) and
1+vv3 <> 0 ) and
-g*vv1*vv2-g*vv2+s*vv1+b+s = 0 ) and
1+vv1 <> 0 ) and
-g*vv2*vv3-g*vv3+s*vv2+b+s = 0 ) and
1+vv2 <> 0 ) and
( ( ( 0 < g^3*vv1^2*vv2^2*vv3^2+2*g^3*vv1^2*vv2^2*vv3+2*g^3*vv1^2*vv2*vv3^2
+2*g^3*vv1*vv2^2*vv3^2+g^3*vv1^2*vv2^2
+4*g^3*vv1^2*vv2*vv3+g^3*vv1^2*vv3^2+4*g^3*vv1*vv2^2*vv3
+4*g^3*vv1*vv2*vv3^2+g^3*vv2^2*vv3^2
+2*g^3*vv1^2*vv2+2*g^3*vv1^2*vv3+2*g^3*vv1*vv2^2+8*g^3*vv1*vv2*vv3
+2*g^3*vv1*vv3^2
+2*g^3*vv2^2*vv3+2*g^3*vv2*vv3^2+g^3*vv1^2+4*g^3*vv1*vv2+4*g^3*vv1*vv3
+g^3*vv2^2 +4*g^3*vv2*vv3+g^3*vv3^2+2*g^3*vv1
+2*g^3*vv2+2*g^3*vv3+b^3+g^3 and
0 < (1+vv2)^2*(1+vv3)^2*(1+vv1)^2 ) and
8*g^3*vv1^2*vv2^2*vv3^2+16*g^3*vv1^2*vv2^2*vv3+16*g^3*vv1^2*vv2*vv3^2
+16*g^3*vv1*vv2^2*vv3^2+8*g^3*vv1^2*vv2^2+32*g^3*vv1^2*vv2*vv3
+8*g^3*vv1^2*vv3^2+32*g^3*vv1*vv2^2*vv3+32*g^3*vv1*vv2*vv3^2
+8*g^3*vv2^2*vv3^2+16*g^3*vv1^2*vv2+16*g^3*vv1^2*vv3+16*g^3*vv1*vv2^2
+64*g^3*vv1*vv2*vv3+16*g^3*vv1*vv3^2+16*g^3*vv2^2*vv3+16*g^3*vv2*vv3^2
+8*g^3*vv1^2+32*g^3*vv1*vv2+32*g^3*vv1*vv3+8*g^3*vv2^2
+32*g^3*vv2*vv3+8*g^3*vv3^2
+16*g^3*vv1+16*g^3*vv2+16*g^3*vv3-b^3+8*g^3 = 0 ) and
(1+vv2)^2*(1+vv3)^2*(1+vv1)^2 <> 0 ) ) ) ) ) ) ;
\end{verbatim}
\end{footnotesize}
The system variables became quantified variables and have been renamed to \verb!vv1!, \verb!vv2!, and \verb!vv3!,
and the existential quantification is expressed using the syntax of the package \Redlog\ \cite{dolzmannsturm,sturmredlog},
which had been originally driven by the efficient implementation of quantifier elimination based on virtual substitution methods.
Applying quantifier elimination to the formula yields in principle a quantifier-free semi-algebraic description
of the parameters for which Hopf bifurcation fixed points exist.
If one suspects that there is no Hopf bifurcation fixed point or one just wants to assert that there is one,
then one can apply quantifier elimination to the existential closure of our generated formula.
If all variables and parameters are known to be positive, the technique of positive quantifier elimination
can be used \cite{sturmweberabdelrahmanelkahoui}.
\QeHopf\ uses for the quantifier elimination \Redlog, which can use \QEPCADB\ \cite{brown}
for formula simplification and as \textit{fallback method}.
However, for the 3D model already \Redlog\ reduces this formula to the equivalent formula \verb!false!,
i.e. for no parameters
(obeying the positivity condition) a Hopf bifurcation fixed point exists (for positive values).
The needed computation time was less
than 20\,ms.
For the 6D model the fully algorithmic method was not successful, as already the generation
of the formula using Maple failed.
\section{Discussion}
Synthetic biology is one of the most rapidly developing fields of biology. Synthetic genetic circuits are of high interest due to their possible applications in biosensing, bioremediation, diagnostics, therapeutics, etc.
Genetic oscillators are some of the most studied circuits due to their complexity and the possibility of many different topologies.
Building synthetic genetic oscillators with controllable periods and amplitudes would be of great interest to the synthetic biology field as they could for example potentially be used for treatment of diseases related to the circadian cycle.
The experimental validation of complex systems, such as oscillators, can be technically demanding and time consuming. To this day, there has been only few experimental implementations of synthetic oscillators (\cite{elowitzleibler,tiggesmarquezlagostellingfussenegger}). Hence, mathematical modeling of such systems is highly desirable to reduce the experimental workload. Here, we focus on mathematical modeling of 3-cycle genetic repressilators, which have been extensively studied before. However, our study is focused on models based on non-cooperative transcriptional repressors, meaning that all Hill coefficients are always equal to 1. Different studies have already demonstrated that cooperative binding is necessary to obtain oscillations in repressilator systems (\cite{elowitzleibler,bratsunvolfsontsimringhasty
mullerhofbauerendlerflammwidderschuster
wangjingchen}). Our 3D model confirms that oscillations in such a system are indeed absent. However, a theoretical study by \cite{tsaichoimapomereningtangferrell} has shown that the range of parameters in which the system produces oscillations can be expanded by including positive interactions, facilitated by transcriptional activators. We additionally model two repressilator topologies, involving 3 transcriptional activators, driving transcription of either the next or the previous repressor in the cycle. (Let us mention that Allwright's results cannot be applied for our 6D models.)
What do offer the general results of formal reaction kinetics for the treatment our models? The differential equations of each of the models can be considered as induced kinetic differential equations of a reversible reaction, therefore existence of the positive stationary state follows from general results \cite{boroswrexistence, tothnagypapp}, see the details in \ref{subsec:frk}.
To summarize our mathematical results, we have shown that for all positive values of parameters $b, g, s$ system
\eqref{s1} has a single positive stationary point which is a globally asymptotically stable attractor.
Furthermore, \eqref{s6} and \eqref{s6sec} have a single stationary state (point $F$ defined by \eqref{F}) in
the domain $x_i>0 , \ (i=1,\dots, 6)$,
which is a locally asymptotically stable attractor.
Comparing the 3D and 6D models we see that the properties
of solutions in the domains, where all phase variables are
positive, are similar. For all the three systems in these domains there
is a unique singular point which is a strong attractor.
In the 3D system, a small overshoot is possible near the steady state,
whereas no oscillations appear in the first 6D model near the steady state.
In both 6D models the steady state is an attractor:
in both cases all eigenvalues of the steady state have negative real parts,
however two eigenvalues are always
complex conjugate, so it is possible
to observe damping oscillations near the steady state, see Fig. \ref{fig:7}.
Thus, the 6D models demonstrate richer dynamics than the 3D models, including the possibility of damped oscillations.
We can also note that these models, as many others arising
in the studying of biochemical phenomena, exhibit rather simple dynamics.
It was somewhat surprising because the models are given by systems
of differential equations depending on few parameters, and there are systems
which look simpler, but exhibit rather complicate, even chaotic,
dynamics. It can be a challenging problem to understand the reasons
for such simple dynamics. One source of argument may originate in the fact
the models' stationary states are so closely related to stationary states of one linkage class reversible reactions as described in \ref{subsec:frk}.
From the biochemical point of view, the probable reason for the absence of oscillations in the first 6D model is the strength of the activator feedback, which forms a negative feedback loop despite the positive interaction. Nevertheless, different combinations of activators and repressors could result in topologies that produce regular oscillations. Due to the stochasticity of biological systems, stochastic modeling and algorithms could be used to further analyze these topologies.
As to the computational methods: they are based on recent mathematical and algorithmic developments, and can be applied to many different similar problems frequently arising in biochemical studies. Note that theory makes it possible to turn to simpler polynomials than those at the beginning, and also that it is not the same to have a six variable polynomial and to have a three variable polynomial with three parameters.
\section{Appendix}\label{sec:appendix}
\subsection{On the nonlinear term}\label{subsec:nonlinear}
The term $\frac{k_1}{k_2+z^n}$ in \eqref{eq:goodwin} is (from the point of calculations) similar to the one obtained when the Michaelis--Menten kinetics is approximated by Tikhonov method, or to the Holling type kinetics which is often used in population dynamics \cite{kisstoth}. Therefore the methods used above may have applications in reaction kinetics and population biology, as well. The main difference between this term and the reaction rates usually used is that although this rate is always positive, it is not zero if $z$ or $x$ is zero, a general requirement quite often assumed, \cite[p. 613]{volperthudjaev}.
\subsection{Solving systems of polynomial equations}\label{subsec:A}
We give a short summary on the topics of solving polynomial systems.
The interested reader may consult \cite{coxlittleshea,romanovskishafer} for more details.
Let $k[x_1, \ldots, x_n]$ denote the ring of polynomials in $n$ indeterminates with coefficients in the field $k$,
which is typically the set $\R$ of real numbers or $\mathbb{C}$ of complex numbers.
The problem of finding solutions to a system of polynomials
\begin{eqnarray}\label{e:poly.sys}
f_1(x_1,\dots,x_n)&=&0,\nonumber\\
\qquad\quad\ \vdots&&\\
f_m(x_1,\dots,x_n)&=&0\nonumber
\end{eqnarray}
is a challenging mathematical problem.
Such systems often have infinitely many solutions, and it is simply impossible to find them all numerically. Even if system (\ref{e:poly.sys}) has a finite number of solutions, it is still very difficult and often impossible to find all of them
numerically without applying methods of computational algebra.
In fact, no regular methods for solving system (\ref{e:poly.sys}) were known until the mid-sixties of the last century
when Bruno Buchberger \cite{buchberger} invented the theory of Gr\"{o}bner bases, which is now the cornerstone
of modern computational algebra. We shall recall briefly the notion of a Gr\"{o}bner basis.
Let $I=\la f_1,f_2, \dots, f_s\ra$ denote the ideal generated by polynomials $f_1(x_1,\dots,x_n)$, $\ldots$,
$f_m(x_1,\dots,x_n)$, that is,
the set of all sums
$
\{h_1 f_1+h_2 f_2+\dots + h_s f_s\},
$
where $f_k, h_k$ are polynomials.
A Gr\"{o}bner basis of a given ideal $I$ depends on a term ordering of monomials of $k[x_1,\dots, x_n]$. The two most commonly used term
orders are lexicographic order (lex) and degree reverse lexicographic order
(degrev), defined as follows.
Let $\alphab = (\alpha_1, \dots,\ \alpha_n)$ and
$\betab = (\beta_1, \dots, \beta_n)$ be elements of $\np^n$ ($\np=\N\cup 0$).
We say that
$\alphab >_{\rm lex} \betab$ with respect to lexicographic order if and only if, reading from left to right,
the first nonzero entry in the $n$-tuple
$\alphab - \betab \in \Z^n$ is positive; we say that
$\alphab >_{\rm degrev} \betab$
\ with respect to degree reverse lexicographic order
if and only if
$
|\alphab| = \sum_{j=1}^n \alpha_j > |\betab| = \sum_{j=1}^n \beta_j
$
or
$
|\alphab| = |\betab| \
$
and, reading from right to left, the first nonzero entry in the $n$-tuple
$\alphab - \betab \in \Z^n$ is negative.
For $\gamma\in \np$ let $\xb^{\gamma}$ denote the monomial $x_1^{\gamma_1}x_2^{\gamma_2}\cdots
x_n^{\gamma_n}$.
Fixing a term order on $\kxn$, any $f \in \kxn$ may be reordered
in the \emph{standard form} with respect
to the order, that is,
\begin{equation}\label{standard}
f = a_1 \xb^{\alpha_1} + a_2 \xb^{\alpha_2} + \dots + a_s \xb^{\alpha_s},
\end{equation}
where $\alpha_i \ne \alpha_j$ for $i \ne j$
and $1 \le i,j \le s$, and where, with respect to the specified term order,
$\alpha_1 > \alpha_2 > \cdots > \alpha_s$.
The \emph{leading term}\index{term!leading} $LT(f)$ of $f$ is the term
$LT(f) = a_1 \xb^{\alpha_1}$.
Let $f$ and $g$ be from $\kxn$ with $LT(f) = a \xb^\alphab$ and
$LT(g) = b \xb^\betab$.
The \emph{least common multiple}
of $\xb^\alphab$ and $\xb^\betab$, denoted
$LCM(\xb^\alphab,\xb^\betab)$, is the monomial
$\xb^\gamma = x_1^{\gamma_1} \cdots x_n^{\gamma_n}$ such that
$\gamma_j = \max(\alpha_j, \beta_j)$, $1 \le j \le n$, and the
\emph{$S$-polynomial} of $f$ and $g$ is the polynomial
\[
S(f,g)=\frac{\xb^\gamma}{LT(f)}f -\frac{\xb^\gamma}{LT(g)} g.
\]
The following algorithm due to Buchberger \cite{buchberger} produces a Gr\"{o}bner basis for
the ideal $I=\la\fs \ra \in
\kxn $.
\begin{itemize}
\item[Step 1.] $G := \{ \fs \}$.
\item[Step 2.] For each pair $g_i, g_j \in G$, $i \ne j$, compute the
$S$-polynomial $S(g_i, g_j)$ and compute the remainder $r_{ij}$
of the division $S(g_i, g_j)$ by $ G$.
\item[Step 3.] If
all $r_{ij}$ are equal to zero, output $G$,
else
add
all nonzero $r_{ij}$ to $G$ and return to Step 2.
\end{itemize}
Nowadays, all major computer algebra systems
(\Mathematica, \Maple, \Reduce, \Singular, \Macaulay\ and many others) have routines to compute Gr\"{o}bner bases.
A Gr\"{o}bner basis $G = \{ g_1, \dots, g_m \}$ is called
\emph{reduced}\index{Gr\"obner basis!reduced} if for all $i$, $1 \le i \le m$,
the coefficient of the leading term is 1 and
no term of $g_i$ is divisible by any $LT(g_j)$ for $j \ne i$.
It is well-known (see e.g. $\!$\cite{coxlittleshea}) that system (\ref{e:poly.sys})
has a solution over $\mathbb{C}$ if and only if
the reduced Gr\"obner basis $G$ for $\la \fs \ra$ with respect to any term order on
$\mathbb{C}[x_1, \dots, x_n]$ is different from $\{ 1 \}$.
The Gr\"obner basis theory allows to find all solutions of system
(\ref{e:poly.sys}) when the system has only finitely many solutions.
In such case a Gr\"obner basis with respect to
the lexicographic order is always in a ``triangular" form
(like the Gauss row-echelon form in the case of linear systems)
which means that one has an equation in a single variable, and having solved it one can substitute the roots into an equation in two variables, solve it, etc.
For a field $k$
an \emph{affine variety} is a subset of $k^n$ that is the solution set of a system of equations of the form
\eqref{e:poly.sys}, where $f_i$ are polynomials with coefficients
in $k$. It is denoted by $\vv(I)$, where $I$ is the ideal
generated by $f_1, \ldots, f_m$, $I:=\la f_1,f_2,\dots,f_m\ra$.
A variety is \emph{irreducible} if it is not the union of
finitely many proper subsets, each of which is itself a variety.
Every affine variety $V$ can
be decomposed into finitely many irreducible components, that is $V$ is expressible as
\begin{equation}\label{gevd}
V = V_1 \cup \dots \cup V_s,
\end{equation}
where each $V_j$ is irreducible and $V_j \not\subset V_k$ if $j \ne k$, and in fact this decomposition is unique up to the ordering of the components $V_j$. Thus to solve \eqref{e:poly.sys} we have
to find the decomposition \eqref{gevd} for
$V = \vv(I)$.
A radical of the ideal $I$ is the set of polynomials
$\sqrt{I}:=\{f:f^p\in I\text{ for some }p\in\N\}.$
An ideal $I \subset \kxn$ is called a \emph{primary ideal} if
for any pair $f, g \in \kxn$, $f g \in I$ only if either $f \in I$ or $g^p \in I$ for some $p \in \N$. An ideal $I$ is
primary if and only if $\sqrt{I}$ is prime; $\sqrt{I}$ is called the \emph{associated prime ideal of $I$}. A
\emph{primary decomposition} of an ideal $I \subset \kxn$ is a representation of $I$ as a finite intersection of
primary ideals $Q_j$, $I = \cap_{j = 1}^s Q_j$\,.
The decomposition is called a \emph{minimal} primary decomposition if
the associated prime ideals $\sqrt{Q_j}$ are all distinct and $\cap_{i \ne j} Q_i \not \subset Q_j$ for any $j$.
A minimal primary decomposition of a polynomial ideal always exists, but it is not necessarily unique.
Every ideal $I$ in $\kxn$ has a minimal primary decomposition according to the
Lasker--Noether Decomposition Theorem.
All such decompositions have the
same number $m$ of primary ideals and the same collection of associated prime ideals.
Minimal associate primes of a polynomial ideal $I=\la f_1,f_2,\dots,f_m\ra$
can be computed using the algorithm proposed by \cite{giannitragerzacharias}, and the varieties of the minimal associate primes give then the irreducible decomposition of the variety $V(I)$ (so give the "solution" to the system $f_1=f_2=\dots f_m=0$).
\subsection{Solving Eq. \eqref{semiout}}\label{subsec:B
Input is system \eqref{semiin}, and the output is its solution \eqref{semiout}.
\begin{verbatim}
In[20]:= Reduce[{f1 == 0 && f3 == 0 && f5 == 0 && x1 > 0 &&
x3 > 0 && x5 > 0 && s > 0 && g > 0 && b > 0},
{x1, x3, x5, s, b, g}] // FullSimplify
Out[20]= {x1 > 0 && x1 == x3 && x3 == x5 && s > 0 && b > 0 &&
g == s/x1 + b/(1 + x1 + x5)}
\end{verbatim}
\subsection{Minimal associate primes}\label{subsec:C
Minimal associate primes of ideal \eqref{I} defining the ideals $J_1, J_2, J_3$
of the decomposition \eqref{star} are:
\begin{verbatim}
[1]:
_[1]=x3-x5
_[2]=x1-x5
_[3]=2*s*x5+s+b*x5-2*g*x5^2-g*x5
[2]:
_[1]=x1^3*x3+x1^3*x5+x1^3-2*x1^2*x3*x5+x1^2*x5+x1^2+x1*x3^3
-2*x1*x3^2*x5+x1*x3^2-2*x1*x3*x5^2-6*x1*x3*x5-x1*x3+x1*x5^3
-x1*x5+x3^3*x5+x3^3+x3^2+x3*x5^3+x3*x5^2-x3*x5+x5^3+x5^2
_[2]=b*x3^3+b*x3^2+b*x3*x5+b*x3-b*x5^3-2*b*x5^2-b*x5
-g*x1^2*x3^2-2*g*x1^2*x3*x5-2*g*x1^2*x3-g*x1^2*x5^2
-2*g*x1^2*x5-g*x1^2-g*x1*x3^3+g*x1*x3^2*x5-g*x1*x3^2
+g*x1*x3*x5^2-g*x1*x3-g*x1*x5^3-3*g*x1*x5^2-3*g*x1*x5
-g*x1+g*x3^3*x5+2*g*x3^2*x5^2+4*g*x3^2*x5+g*x3^2+g*x3*x5^3
+4*g*x3*x5^2+4*g*x3*x5+g*x3
_[3]=b*x1*x5+b*x1-b*x3^2-b*x3+g*x1^2*x3+g*x1^2*x5+g*x1^2
+g*x1*x3^2+2*g*x1*x3-g*x1*x5^2+g*x1-g*x3^2*x5-g*x3*x5^2
-2*g*x3*x5-g*x5^2-g*x5
_[4]=b*x1*x3+b*x3-b*x5^2-b*x5-g*x1^2*x3-g*x1^2*x5-g*x1^2
+g*x1*x3^2-g*x1*x5^2-2*g*x1*x5-g*x1+g*x3^2*x5+g*x3^2
+g*x3*x5^2+2*g*x3*x5+g*x3
_[5]=b*x1^2+b*x1-b*x3*x5-b*x5+g*x1^2*x3-g*x1^2*x5+g*x1*x3^2
+2*g*x1*x3-g*x1*x5^2-2*g*x1*x5+g*x3^2*x5+g*x3^2-g*x3*x5^2
+g*x3-g*x5^2-g*x5
_[6]=b^2*x3^2+b^2*x3*x5+b^2*x3+b^2*x5^2+2*b^2*x5+b^2
+b*g*x3^2*x5+2*b*g*x3^2-b*g*x3*x5^2+b*g*x3*x5+2*b*g*x3
+b*g*x5^2+2*b*g*x5+b*g+2*g^2*x1*x3^3+2*g^2*x1*x3^2*x5
+3*g^2*x1*x3^2+2*g^2*x1*x3*x5^2+4*g^2*x1*x3*x5+2*g^2*x1*x3
+2*g^2*x1*x5^3+5*g^2*x1*x5^2+4*g^2*x1*x5+g^2*x1
+2*g^2*x3^3*x5+2*g^2*x3^3+4*g^2*x3^2*x5^2+8*g^2*x3^2*x5
+4*g^2*x3^2+2*g^2*x3*x5^3+8*g^2*x3*x5^2+9*g^2*x3*x5
+3*g^2*x3+2*g^2*x5^3+5*g^2*x5^2+4*g^2*x5+g^2
_[7]=2*s*x5+s-b*x1+b*x3+b*x5-2*g*x1*x3-g*x1-g*x3+g*x5
_[8]=2*s*x3+s+b*x1+b*x3-b*x5-2*g*x1*x5-g*x1+g*x3-g*x5
_[9]=2*s*x1+s+b*x1-b*x3+b*x5+g*x1-2*g*x3*x5-g*x3-g*x5
_[10]=s*b+b^2*x1+b^2*x3+b^2*x5+2*b^2+b*g*x1+b*g*x3+b*g*x5
+2*b*g+2*g^2*x1^2*x3+2*g^2*x1^2*x5+2*g^2*x1^2+2*g^2*x1*x3^2
+4*g^2*x1*x3*x5+6*g^2*x1*x3+2*g^2*x1*x5^2+6*g^2*x1*x5
+4*g^2*x1+2*g^2*x3^2*x5+2*g^2*x3^2+2*g^2*x3*x5^2
+6*g^2*x3*x5+4*g^2*x3+2*g^2*x5^2+4*g^2*x5+2*g^2
_[11]=s^2+s*g-b^2*x1-b^2*x3-b^2*x5-b^2-b*g*x1-b*g*x3
-b*g*x5-b*g-2*g^2*x1^2*x3-2*g^2*x1^2*x5-2*g^2*x1^2
-2*g^2*x1*x3^2-4*g^2*x1*x3*x5-5*g^2*x1*x3-2*g^2*x1*x5^2
-5*g^2*x1*x5-3*g^2*x1-2*g^2*x3^2*x5-2*g^2*x3^2
-2*g^2*x3*x5^2-5*g^2*x3*x5-3*g^2*x3-2*g^2*x5^2-3*g^2*x5-g^2
[3]:
_[1]=g
_[2]=b
_[3]=s
\end{verbatim}
\subsection{Checking the conditions of Allwright's theorem in the 3D case}\label{subsec:D
Here we strongly rely on the paper \cite{allwright}: we use the definitions and notations of that paper.
His equations (5) specialize into our Eq. \eqref{s1}
with the following cast: $n=3$, and for $j=1,2,3: T_j=0, h_j(x)=s+\frac{b}{1+x}, k_j(x)=-gx.$ The quantities and functions defined in this way fulfil conditions (6)--(8) in his paper. As the inverse of $k$ is
$y\mapsto-y/g$ the function $\Phi$ in (9) can be calculated as
\begin{equation*}
\frac{b^2 g+u \left(-b g^2+2 b g s-g^2 s+2 g s^2-s^3\right)-b g^2+3 b g s-b s^2-g^2 s+2 g s^2-s^3}{g u \left(-b g+g^2-2 g s+s^2\right)+g \left(-2 b g+b s+g^2-2 g s+s^2\right)}.
\end{equation*}
The derivative of $\Phi$ is negative for nonnegative arguments $u$ in accordance with the fact that the function is decreasing. Thus we have Case I with the notation of the paper. Further---lengthy---calculations show that the equation $\Phi(\Phi(u))=u$ has one positive (and one negative) real root:
\begin{equation}
u_1=\frac{-g+s+\sqrt{(g+s)^2+4bg}}{2g}>0,\quad u_2=\frac{-g+s-\sqrt{(g+s)^2+4bg}}{2g}<0,
\end{equation}
therefore case (i) of Theorem 1 of the paper applies stating the \emph{global} asymptotic stability of the unique stationary point.
\subsection{Realizations with reversible reactions}\label{subsec:frk}
Consider the equation \eqref{eq:st6D1} for the stationary points of the first 6D model:
\begin{eqnarray}
s + b x_1 - g x_1 + s x_1 - g x_1^2 + s x_5 - g x_1 x_5 &=& 0 \nonumber\\
s + s x_1 + b x_3 - g x_3 + s x_3 - g x_1 x_3 - g x_3^2 &=& 0 \label{eq:st6D1}\\
s + s x_3 + b x_5 - g x_5 + s x_5 - g x_3 x_5 - g x_5^2 &=& 0.\nonumber
\end{eqnarray}
Let us note that the mass action type induced kinetic differential equation
of the reaction in Fig. \ref{fig:sixd1} is has exactly the right hand side equal to the left hand sides of the sbove equations if the reaction rate coefficients have appropriately been chosen. Therefore, based on the results by Orlov and Rozonoer \cite{orlovrozonoer2,tothnagypapp} (or using the recent generalization by Boros \cite{boroswrexistence}) one can conclude that there exists a positive stationary point of the reaction, and thus, of the original (first) 6D model also has one.
\begin{figure}
\centering
\includegraphics[width=300pt]{sixd1}
\caption{Feinberg--Horn--Jackson graph of a reaction leading to the stationary point of th first 6D model}\label{fig:sixd1}
\end{figure}
The same argument can be applied in the case of the other two models.
|
\section{Introduction}
Motivated by the works \cite{bcs1, bcs2}, we introduce the following class of systems
\begin{equation}
\label{BBM-KdV}
\left\{
\begin{aligned}
&F_1 \equiv u_t + (a + b)vu_x + (au + c)v_x + \epsilon u_{txx} + \kappa v_{xxx} = 0\\
&F_2 \equiv v_t + (bu + c)u_x + (a + b)vv_x + \lambda u_{xxx} + \sigma v_{txx} = 0
\end{aligned}
\right.,
\end{equation}
henceforth simply referred to as BBM-KdV system, a two-component generalization\footnote{If $u = v$, the system (\ref{BBM-KdV}) is reduced to equations $u_t + [(2a + b)u + c]u_x + \epsilon u_{txx} + \kappa u_{xxx} = 0$ and $u_t + [(a + 2b)u + c]u_x + \sigma u_{txx} + \lambda u_{xxx} = 0$. Both contain (\ref{BBM}) and (\ref{KdV}) as special cases.} of the classic equations \cite{bbm}
\begin{equation}
\label{BBM}
BBM\!\!: \ u_t + (u + 1)u_x - u_{txx} = 0
\end{equation}
and
\begin{equation}
\label{KdV}
KdV\!\!: \ u_t + (u + 1)u_x + u_{xxx} = 0,
\end{equation}
with the objective of studying it from the point of view of the group analysis.
In (\ref{BBM-KdV}), the constants are such that $(a + b)c \neq 0$ and $\{\epsilon, \kappa, \lambda, \sigma\} \neq \{0\}$. Particularly when $a = c = 1$ and $b = 0$, we obtain the already widely investigated systems of Boussinesq ($\epsilon = \kappa = \lambda = 0$, $\sigma = -1/3$), Kaup ($\epsilon = \lambda = \sigma = 0$, $\kappa = 1/3$) and Bona-Smith ($\epsilon = \sigma = \lambda/2 - 1/6$, $\kappa = 0$, $\lambda < 0$), all of them first-order approximations to the Euler equations in the framework of hydrodynamics. Useful in situations where dissipative effects are not significant, these models provide a good description for the two-dimensional motion of small-amplitude long waves on the surface of an ideal fluid. In this context, the independent variable $x$ represents the distance traveled along a fixed depth channel and $t$ the time. The quantities $u(t, x)$ and $v(t, x)$ are related to the deviation of the surface from its undisturbed level and to the horizontal velocity of the fluid, respectively. For more information, see \cite{bcs1,bcs2} and references therein. Relevant results, including exact solutions, can be found in \cite{adm,c,ddlm}.
It's well known that evolution equations don't possess an usual Lagrangian. Therefore this paper is thus organized: first we determine the Lie point symmetries (Section 2) of the BBM-KdV system and establish its nonlinear self-adjointness (Section 3); we then construct conservation laws via Ibragimov's Theorem (Section 4), an extension of the celebrated Noether's Theorem to problems with no variational structure. In the next sections, unless otherwise stated, $c_i$'s are arbitrary constants. All functions are smooth.
We consider that the reader is familiar with the fundamental concepts of group analysis. The basic literature used is \cite{bk,gan,i1,i3,i2,i4,i5,ol}.
\section{Lie Point Symmetries Classification}
Without many details, applying the standard algorithm presented in \cite{bk} and \cite{ol}, a differential operator
\begin{equation}
X = \mathcal{T}(t, x, u, v)\frac{\partial}{\partial t} + \mathcal{X}(t, x, u, v)\frac{\partial}{\partial x} + \mathcal{U}(t, x, u, v)\frac{\partial}{\partial u} + \mathcal{V}(t, x, u, v)\frac{\partial}{\partial v}\nonumber
\end{equation}
generates the Lie point symmetries of the system (\ref{BBM-KdV}) if the conditions of invariance (the so-called determining equations)
\begin{equation}
\label{determining}
\begin{gathered}
\mathcal{T}_x = \mathcal{T}_u = \mathcal{T}_v = \mathcal{X}_u = \mathcal{X}_v = 0,\\
\epsilon\mathcal{X}_t = \epsilon\mathcal{X}_x = \sigma\mathcal{X}_t = \sigma\mathcal{X}_x = 0,\\
\mathcal{U}_t = \mathcal{U}_x = \mathcal{U}_v = \mathcal{V}_t = \mathcal{V}_x = a\mathcal{U} - (au + c)\mathcal{U}_u = 0,\\
b\mathcal{U} + (bu + c)[\mathcal{U}_u + 2(\mathcal{T}_t - \mathcal{X}_x)] = 0,\\
\kappa(\mathcal{U}_u + 2\mathcal{X}_x) = \lambda[\mathcal{U}_u + 2(\mathcal{T}_t - 2\mathcal{X}_x)] = 0,\\
(a + b)[\mathcal{V} + (\mathcal{T}_t - \mathcal{X}_x)v] - \mathcal{X}_t = 0
\end{gathered}
\end{equation}
are satisfied. From (\ref{determining}), it's easy to see that
\begin{equation}
\begin{gathered}
\mathcal{T} = (a + b)c_1t + c_2, \ \mathcal{X} = (a + b)(c_3x + c_4t) + c_5,\\
\mathcal{U} = 2(c_3 - c_1)(au + c), \ \mathcal{V} = (a + b)(c_3 - c_1)v + c_4
\end{gathered}
\nonumber
\end{equation}
with
\begin{equation}
\begin{gathered}
b(a - b)(c_1 - c_3) = 0, \ \epsilon c_3 = \epsilon c_4 = \kappa[ac_1 - (2a + b)c_3] = 0,\\
\lambda[bc_1 - (a + 2b)c_3]= \sigma c_3 = \sigma c_4 = 0.
\end{gathered}
\nonumber
\end{equation}
\
\noindent{{\bf Proposition 1.} \textit{The Lie point symmetries of the BBM-KdV system are summarized in Table 1, where}
\begin{equation}
\begin{gathered}
X_1 = (a + b)\left(t\frac{\partial}{\partial t} - v\frac{\partial}{\partial v}\right) - 2(au + c)\frac{\partial}{\partial u},\\
X_2 = \frac{\partial}{\partial t}, \ X_3 = (a + b)\left(x\frac{\partial}{\partial x} + v\frac{\partial}{\partial v}\right) + 2(au + c)\frac{\partial}{\partial u},\\
X_4 = (a + b)t\frac{\partial}{\partial x} + \frac{\partial}{\partial v}, \ X_5 = \frac{\partial}{\partial x}.
\end{gathered}
\nonumber
\end{equation}
\begin{table}[h!]
\centering
\begin{tabular}{ccccl}
\cline{2-4}
\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{$b = 0$} & \multicolumn{1}{c|}{$a = b$} & \multicolumn{1}{c|}{$b(a - b) \neq 0$} & \\ \cline{1-4}
\multicolumn{1}{|c|}{$\{\epsilon, \sigma\} = \{0\}$} & \multicolumn{1}{c|}{\begin{tabular}{clll}
$X_1$ ($\kappa = 0)$ \\
$2X_1 + X_3$ ($\lambda = 0$) \\
$X_2$, $X_4$, $X_5$
\end{tabular}} & \multicolumn{1}{c|}{\begin{tabular}{cllll}
$3X_1 + X_3$ \\
$X_2$, $X_4$, $X_5$
\end{tabular}} & \multicolumn{1}{c|}{$X_2$, $X_4$, $X_5$} & \\ \cline{1-4}
\multicolumn{1}{|c|}{$\{\epsilon, \sigma\} \neq \{0\}$} & \multicolumn{1}{c|}{\begin{tabular}{cllll}
$X_1$ ($\kappa = 0$), $X_2$, $X_5$
\end{tabular}} & \multicolumn{1}{c|}{\begin{tabular}{cllll}
$X_1$ ($\kappa = \lambda = 0$) \\
$X_2$, $X_5$
\end{tabular}} & \multicolumn{1}{c|}{$X_2$, $X_5$} & \\ \cline{1-4}
& \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} &
\end{tabular}
\\
\tablename{ 1.}
\end{table}
\section{Self-Adjointness Classification}
To begin with, let $\bar{u}$ and $\bar{v}$ be the new dependent variables. The formal Lagrangian of the system (\ref{BBM-KdV}) is
\begin{equation}
\mathcal{L} = \bar{u}F_1 + \bar{v}F_2.
\nonumber
\end{equation}
Calculated the adjoint equations
\begin{equation}
\left\{
\begin{aligned}
&F_1^* \equiv -\frac{\delta\mathcal{L}}{\delta u} = \bar{u}_t + (a + b)v\bar{u}_x + \ \! (bu + c)\bar{v}_x + \ \! b\bar{u}v_x + \epsilon\bar{u}_{txx} + \lambda\bar{v}_{xxx} = 0\\
&F_2^* \equiv -\frac{\delta\mathcal{L}}{\delta v} = \bar{v}_t + (au + c)\bar{u}_x + (a + b)v\bar{v}_x - b\bar{u}u_x + \kappa\bar{u}_{xxx} + \sigma\bar{v}_{txx} = 0
\end{aligned}
\right.,\nonumber
\end{equation}
where $\delta/\delta u$ and $\delta/\delta v$ are Euler-Lagrange operators, we assume that
\begin{equation}
F_1^*|_{(\bar{u}, \bar{v}) = (\varphi, \psi)} = MF_1 + NF_2, \quad F_2^*|_{(\bar{u}, \bar{v}) = (\varphi, \psi)} = PF_1 + QF_2.\label{self}
\end{equation}
Here $M$, $N$, $P$ and $Q$ is a set of coefficients to be determined and
\begin{equation}
\label{sub}
\varphi = \varphi(t, x, u, v), \quad \psi = \psi(t, x, u, v)
\end{equation}
two functions that not vanish simultaneously. As
\begin{equation}
F_1^*|_{(\bar{u}, \bar{v}) = (\varphi, \psi)} = D_t\varphi + (a + b)vD_x\varphi + (bu + c)D_x\psi + b\varphi v_x + \epsilon D_tD_x^2\varphi + \lambda D_x^3\psi
\nonumber
\end{equation}
and
\begin{equation}
F_2^*|_{(\bar{u}, \bar{v}) = (\varphi, \psi)} = D_t\psi + (au + c)D_x\varphi + (a + b)vD_x\psi - b\varphi u_x + \kappa D_x^3\varphi + \sigma D_tD_x^2\psi,
\nonumber
\end{equation}
from (\ref{self}) it's possible to conclude that $M = \varphi_u$, $N = \varphi_v$, $P = \psi_u$, $Q = \psi_v$ and
\begin{equation}
\label{self-determining}
\begin{gathered}
\varphi_t + (a + b)v\varphi_x = \epsilon\varphi_x = 0,\\
\psi_t + (au + c)\varphi_x = \psi_x = 0,\\
b\varphi = (au + c)\varphi_u - (bu + c)\psi_v,\\
\varphi_v - \psi_u = (\epsilon - \sigma)\varphi_v = \kappa\varphi_u - \lambda\psi_v = 0,\\
\epsilon\varphi_{uu} = \varphi_{uv} = \varphi_{vv} = 0.
\end{gathered}
\nonumber
\end{equation}
Hence
\begin{equation}
\varphi = (c_1t + c_2)av + f(u) - c_1x, \quad \psi = c_3v + (c_1t + c_2)au + c_1ct + c_4\nonumber
\end{equation}
with
\begin{equation}
\begin{gathered}
bc_1 = bc_2 = \epsilon c_1 = \sigma c_1 = (\epsilon - \sigma)c_2 = 0,\\
\epsilon f''(u) = \kappa f'(u) - \lambda c_3 = 0,\\
bf(u) = (au + c)f'(u) - (bu + c)c_3.
\end{gathered}
\nonumber
\end{equation}
\
\noindent{{\bf Proposition 2.}} \textit{The BBM-KdV system is nonlinearly self-adjoint. The substitutions $(\ref{sub})$ are as follows.}
\
\noindent{\textbf{i)}} \textit{If $b = 0$,}
\begin{equation}
\varphi = (c_1t + c_2)av + c_3c\ln(au + c) - c_1x + c_4, \quad \psi = c_3av + (c_1t + c_2)(au + c) + c_5\nonumber
\end{equation}
\textit{where}
\begin{equation}
\left\{
\begin{aligned}
&c_1 = 0, \ \mathrm{to} \ \{\epsilon, \sigma\} \neq \{0\},\\
&c_2 = 0, \ \mathrm{to} \ \epsilon \neq \sigma,\\
&c_3 = 0, \ \mathrm{to} \ \{\epsilon, \kappa, \lambda\} \neq \{0\}.
\end{aligned}
\right.\nonumber
\end{equation}
\noindent{\textbf{ii)}} \textit{If $a = b$,}
\begin{equation}
\varphi = (au + c)[c_1\ln(au + c) + c_2], \quad \psi = c_1av + c_3\nonumber
\end{equation}
\textit{where}
\begin{equation}
\left\{
\begin{aligned}
&c_1 = 0, \ \mathrm{to} \ \{\epsilon, \kappa, \lambda\} \neq \{0\},\\
&c_2 = 0, \ \mathrm{to} \ \kappa \neq 0.\\
\end{aligned}
\right.\nonumber
\end{equation}
\noindent{\textbf{iii)}} \textit{Let $b(a - b) \neq 0$.}
\textbf{iii.a)} \textit{If $a = 0$,}
\begin{equation}
\varphi = c_1e^{bu/c} - c_2(bu + 2c), \quad \psi = c_2bv + c_3\nonumber
\end{equation}
\textit{where}
\begin{equation}
\left\{
\begin{aligned}
&c_1 = 0, \ \mathrm{to} \ \{\epsilon, \kappa\} \neq \{0\},\\
&c_2 = 0, \ \mathrm{to} \ \kappa \neq -\lambda.
\end{aligned}
\right.\nonumber
\end{equation}
\textbf{iii.b)} \textit{If $a \neq 0$,}
\begin{equation}
\varphi = c_1(au + c)^{b/a} + c_2[b^2u + (2b - a)c], \quad \psi = c_2(a - b)bv + c_3\nonumber
\end{equation}
\textit{where}
\begin{equation}
\left\{
\begin{aligned}
&c_1 = 0, \ \mathrm{to} \ \{\epsilon, \kappa\} \neq \{0\},\\
&c_2 = 0, \ \mathrm{to} \ \lambda a \neq (\kappa + \lambda)b.
\end{aligned}
\right.\nonumber
\end{equation}
\
\noindent{\textbf{Remark.} Actually, the system (\ref{BBM-KdV}) is quasi self-adjoint. It becomes strictly self-adjoint in only two circumstances: $a = 2b$ and $\kappa = \lambda$; or $b = 0$ and $\epsilon = \sigma$.}
\section{Conservation Laws}
In view of Proposition 2, the components of the conserved vector $C = (C^t, C^x)$ associated to $X$, a Lie point symmetry admitted by the system (\ref{BBM-KdV}), are according to Ibragimov's Theorem given by
\begin{equation}
\label{componente-t}
C^t = (\varphi - \epsilon D_x\varphi D_x)W^u + (\psi - \sigma D_x\psi D_x)W^v\nonumber
\end{equation}
and
\begin{equation}
\begin{split}
C^x = [(a &+ b)v\varphi + (bu + c)\psi + \epsilon(\varphi D_tD_x + D_tD_x\varphi) + \lambda(\psi D_x^2 - D_x\psi D_x + D_x^2\psi)]W^u +\\
&+ [(au + c)\varphi + (a + b)v\psi + \sigma(\psi D_tD_x + D_tD_x\psi) + \kappa(\varphi D_x^2 - D_x\varphi D_x + D_x^2\varphi)]W^v,
\end{split}
\nonumber
\end{equation}
with
\begin{equation}
W^u = \mathcal{U} - \mathcal{T}u_t - \mathcal{X}u_x, \quad W^v = \mathcal{V} - \mathcal{T}v_t - \mathcal{X}v_x.\nonumber
\end{equation}
We find the conservation laws corresponding to each generator of Table 1. In most cases, however, we are led to trivial vectors or the vectors
\begin{equation}
C^t = u + \epsilon u_{xx}, \quad C^x = (au + c)v + \kappa v_{xx}\nonumber
\end{equation}
and
\begin{equation}
C^t = 2(v + \sigma v_{xx}), \quad C^x = (a + b)v^2 + (bu + 2c)u + 2\lambda u_{xx}\nonumber
\end{equation}
that can be obtained from the first (when b = 0) and second equation of the BBM-KdV system by simple integration (obvious conservation laws). The really interesting cases we list below.
\
\noindent{{\bf Proposition 3.} \textbf{i)} {\it Let $b = 0$.}
\
\textbf{i.a)} {\it From $X_1$, $2X_1 + X_3$ and $X_2$, we obtain}
\begin{equation}
\begin{gathered}
C^t = 2(uv - \epsilon u_xv_x),\\
C^x = cu^2 + (2au + c)v^2 - (\lambda u_x^2 + \kappa v_x^2) + 2[u(\lambda u_{x} + \epsilon v_{t})_x + v(\epsilon u_{t} + \kappa v_{x})_x]
\end{gathered}
\nonumber
\end{equation}
{\it when $\epsilon = \sigma$.}
\
\textbf{i.b)} {\it For $\epsilon = \kappa = 0$, $X_1$ also provides}
\begin{equation}
\begin{gathered}
C^t = \frac{1}{a}(au + c)\ln(au + c) + \frac{a}{2c}(v^2 - \sigma v_x^2),\\
C^x = (au + c)[\ln(au + c) + 1]v + \frac{av}{c}\!\left(\frac{av^2}{3} + \sigma v_{tx}\right)
\end{gathered}
\nonumber
\end{equation}
{\it when $\lambda = 0$ and}
\begin{equation}
\begin{gathered}
C^t = 2[t(au + c)v - xu],\\
C^x = t[c(au + 2c)u - a\lambda u_x^2] + 2(au + c)[(atv - x)v + \lambda tu_{xx}]
\end{gathered}
\nonumber
\end{equation}
{\it when $\sigma = 0$.}
\
\noindent{\textbf{ii)} {\it Let $a = b$.}}
\
\textbf{ii.a)} {\it From $X_1$ and $3X_1 + X_3$, we obtain}
\begin{equation}
\begin{gathered}
C^t = (au + 2c)u - a\epsilon u_x^2,\\
C^x = 2(au + c)[(au + c)v + \epsilon u_{tx}]
\end{gathered}
\nonumber
\end{equation}
{\it when $\kappa = 0$.}
\
\textbf{ii.b)} {\it $X_1$ also provides}
\begin{equation}
\begin{gathered}
C^t = \frac{1}{a}(au + c)^2\ln(au + c) + a(v^2 - \sigma v_x^2),\\
C^x = (au + c)^2[2\ln(au + c) + 1]v + 2av\!\left(\frac{2av^2}{3} + \sigma v_{tx}\right)
\end{gathered}
\nonumber
\end{equation}
{\it when $\epsilon = \kappa = \lambda = 0$.}
|
\section{Introduction}
\label{intro}
\input{sections/1_introduction.tex}
\section{Related Work}
\label{related_work}
\input{sections/2_related_work.tex}
\section{Proposed Approach}
\label{proposed_approach}
\input{sections/3_proposed_approach.tex}
\section{Results}
\label{results}
\input{sections/4_results.tex}
\section{Interfacing with Twitter}
\label{interface}
\input{sections/5_interface.tex}
\section{Conclusion}
\label{conclusion}
\input{sections/6_conclusion.tex}
\subsection{Data}
The dataset that we have generated is a combination of three different datasets. The first dataset is publicly available on Crowdflower\footnote{\url{https://data.world/crowdflower/hate-speech-identification}}, which was used in \cite{7} and \cite{15}. This dataset contains tweets that have been manually classified into one of the following classes: ``Hateful'', ``Offensive'' and ``Clean''. The second dataset is also publicly available on Crowdflower\footnote{\url{https://data.world/ml-research/automated-hate-speech-detection-data}}, which consists the tweets with same classes as described previously. The third dataset is published on Github\footnote{\url{https://github.com/ZeerakW/hatespeech}} and used in the work \cite{7} and \cite{16}. It consists of two columns: tweet-ID and class. In this dataset, tweets corresponding to the tweet-ID are classified into one of the following three classes: ``Sexism'', ``Racism'' and ``Neither''.
\subsection{Data Preprocessing}
In the data preprocessing stage, we combine the three datasets used for this work. The tasks involves removal of unnecessary columns from the datasets and enumerating the classes. For the third dataset, we retrieve the tweets corresponding to the tweet-ID present in the dataset. We use Twitter API for this purpose. The classes ``Sexism'' and ``Racism'' in this dataset are both considered as hate speech according to the definition.
We convert the tweets to lowercase and remove the following unnecessary contents from the tweets:
\begin{itemize}
\item Space Pattern
\item URLs
\item Twitter Mentions
\item Retweet Symbols
\item Stopwords
\end{itemize}
We use the Porter Stemmer algorithm to reduce the inflectional forms of the words.
After combining the dataset in proper format, we randomly shuffle and split the dataset into two parts: train dataset containing 70\% of the samples and test dataset containing 30\% of the samples.
\subsection{Feature Extraction}
We extract the n-gram features from the tweets and weight them according to their TFIDF values. The goal of using TFIDF is to reduce the effect of less informative tokens that appear very frequently in the data corpus. Experiments are performed on values of $n$ ranging from one to three. Thus, we consider unigram, bigram and trigram features. The formula that is used to compute the TFIDF of term $t$ present in document $d$ is: \[tfidf(d,t) = tf(t)\; * \;idf(d,t)\] Also, both L1 and L2 (Euclidean) normalization of TFIDF is considered while performing experiments. L1 normalization is defined as: \[v_{norm} = \frac{v}{|v_1|+|v_2|+...+|v_n|}\] where $n$ in the total number of documents. Similarly, L2 normalization is defined as: \[v_{norm} = \frac{v}{\sqrt{v_1^2+v_2^2+...+v_n^2}}\] We feed these features to machine learning models.
\subsection{Model}
We consider three prominent machine learning algorithms used for text classification: Logistic Regression, Naive Bayes and Support Vector Machines. We train each model on training dataset by performing grid search for all the combinations of feature parameters and perform 10-fold cross-validation. The performance of each algorithm is analyzed based on the average score of the cross-validation for each combination of feature parameters. The performance of these three algorithms is compared.
Further, the hyperparameters of two algorithms giving best results are tuned for their respective feature parameters, which gives the best result. Again, 10-fold cross validation is performed to measure the results for each combination of hyperparameters for that model. The model giving the highest cross-validation accuracy is evaluated against the test data. We have used scikit-learn in Python for the purpose of implementation.
\section{Introduction}
\label{intro}
\input{sections/1_introduction.tex}
\section{Related Work}
\label{related_work}
\input{sections/2_related_work.tex}
\section{Proposed Approach}
\label{proposed_approach}
\input{sections/3_proposed_approach.tex}
\section{Results}
\label{results}
\input{sections/4_results.tex}
\section{Interfacing with Twitter}
\label{interface}
\input{sections/5_interface.tex}
\section{Conclusion}
\label{conclusion}
\input{sections/6_conclusion.tex}
\subsection{Data}
The dataset that we have generated is a combination of three different datasets. The first dataset is publicly available on Crowdflower\footnote{\url{https://data.world/crowdflower/hate-speech-identification}}, which was used in \cite{7} and \cite{15}. This dataset contains tweets that have been manually classified into one of the following classes: ``Hateful'', ``Offensive'' and ``Clean''. The second dataset is also publicly available on Crowdflower\footnote{\url{https://data.world/ml-research/automated-hate-speech-detection-data}}, which consists the tweets with same classes as described previously. The third dataset is published on Github\footnote{\url{https://github.com/ZeerakW/hatespeech}} and used in the work \cite{7} and \cite{16}. It consists of two columns: tweet-ID and class. In this dataset, tweets corresponding to the tweet-ID are classified into one of the following three classes: ``Sexism'', ``Racism'' and ``Neither''.
\subsection{Data Preprocessing}
In the data preprocessing stage, we combine the three datasets used for this work. The tasks involves removal of unnecessary columns from the datasets and enumerating the classes. For the third dataset, we retrieve the tweets corresponding to the tweet-ID present in the dataset. We use Twitter API for this purpose. The classes ``Sexism'' and ``Racism'' in this dataset are both considered as hate speech according to the definition.
We convert the tweets to lowercase and remove the following unnecessary contents from the tweets:
\begin{itemize}
\item Space Pattern
\item URLs
\item Twitter Mentions
\item Retweet Symbols
\item Stopwords
\end{itemize}
We use the Porter Stemmer algorithm to reduce the inflectional forms of the words.
After combining the dataset in proper format, we randomly shuffle and split the dataset into two parts: train dataset containing 70\% of the samples and test dataset containing 30\% of the samples.
\subsection{Feature Extraction}
We extract the n-gram features from the tweets and weight them according to their TFIDF values. The goal of using TFIDF is to reduce the effect of less informative tokens that appear very frequently in the data corpus. Experiments are performed on values of $n$ ranging from one to three. Thus, we consider unigram, bigram and trigram features. The formula that is used to compute the TFIDF of term $t$ present in document $d$ is: \[tfidf(d,t) = tf(t)\; * \;idf(d,t)\] Also, both L1 and L2 (Euclidean) normalization of TFIDF is considered while performing experiments. L1 normalization is defined as: \[v_{norm} = \frac{v}{|v_1|+|v_2|+...+|v_n|}\] where $n$ in the total number of documents. Similarly, L2 normalization is defined as: \[v_{norm} = \frac{v}{\sqrt{v_1^2+v_2^2+...+v_n^2}}\] We feed these features to machine learning models.
\subsection{Model}
We consider three prominent machine learning algorithms used for text classification: Logistic Regression, Naive Bayes and Support Vector Machines. We train each model on training dataset by performing grid search for all the combinations of feature parameters and perform 10-fold cross-validation. The performance of each algorithm is analyzed based on the average score of the cross-validation for each combination of feature parameters. The performance of these three algorithms is compared.
Further, the hyperparameters of two algorithms giving best results are tuned for their respective feature parameters, which gives the best result. Again, 10-fold cross validation is performed to measure the results for each combination of hyperparameters for that model. The model giving the highest cross-validation accuracy is evaluated against the test data. We have used scikit-learn in Python for the purpose of implementation. |
\section{Introduction}
\hspace*{4ex} Since the works of Sheppard and Wilson \cite{sheppard},
Brittingham \cite{brit} and Durnin \cite{durnin}, the subject of
non-diffracting beams and pulses, also known as Localized Waves,
has been of interest to many researches. It is now a well
established fact that non-diffracting waves \cite{livro1,livro2}
such as the Bessel beams can resist diffraction effects for long
distances, when compared to the ordinary waves. Moreover, in
Ref.\cite{fw1} it was shown that it is possible to model the
longitudinal intensity pattern of the non-diffracting beams at
will, in an approach dubbed \emph{Frozen Waves} (FWs).
Such intensity pattern, chosen a priori, can be constructed on
axis ($\rho = 0$) or on the surface of a cylinder \cite{fw3} of radius
$\rho = \rho_{\nu}$, where the resulting beams -- called Frozen Waves
(FWs) -- were experimentally demonstrated in Refs. \cite{tarcio1,tarcio2}. Later, the FW
method was extended in \cite{fw2,ahmed2}, allowing the spatial shaping of
non-diffracting beams to take place in absorbing materials. That was an
important step forward in the Localized Wave theory, as it
provided beams not only resistant to the diffraction but
also to the attenuation. The theoretical finite energy version of
diffraction-attenuation resistant beams was developed in \cite{fwbg,fwmicro}, and later on it was shown (theoretically and experimentally) that FW beams can be used
to control the orbital angular momentum (OAM) \cite{ahmed1} and the polarization \cite{mateus1,mateus2} of a beam
along its axis of propagation. In short, today, FWs can be understood as a class of structured light in which the beam's intensity pattern, polarization, OAM (magnitude and sign), and wavelength can be engineered, almost at will, along its axis of propagation.
\hspace*{4ex} As related to pulses and their resistance to diffraction, they have been discussed in Refs. [2,17-30], where the theory was later extended to include pulses resistant to both diffraction and dispersion [31-38]. However, in order to adorn the above mentioned pulses with a degree of resistance toward dispersion, a complicated space-time
coupling in their spectra was needed. But, with the advent of Airy pulses the situation became relatively simpler since such pulses only required a suitable cubic phase in the frequency spectrum [39-44]
\hspace*{4ex} In summary, today, within the Localized Wave theory, we have: i)
beams that can resist diffraction and attenuation effects when
propagating in unguided absorbing media, and ii) pulses resistant
to the diffraction and dispersion effects in unguided dispersive
media. In light of our previous discussion regarding FWs and structured light,
naturally, a question can be asked: is it possible to engineer a vortex pulse
that is concurrently resistant to
the three concomitant effects of diffraction, dispersion and
attenuation in unguided dispersive absorbing media, while at same time --
similar to the case of FWs -- the pulse intensity can display priorly chosen
values at different locations (regions) along its direction of propagation?
Undoubtedly, demonstrating such capability can open new venues for other
possibilities such as engineering the pulse polarization, (local) OAM and central
wavelength along the propagation. However, these particular considerations will
be postponed to future communications.
\hspace*{4ex} In this paper, we show that it is possible to construct pulses that, over a
finite distance, are immune to diffraction, dispersion and attenuation while, at
same time, their intensities can assume arbitrary values and patterns along the
propagation. The paper is organized as the following. In section 2, we develop
the required theoretical frame work, whereas section 3 describes some of the
applicable examples. Section 4 contains our final remarks and conclusions.
We expect that the theoretical formulation presented here will have an important
impact in multiple areas of research in optical sciences such as optical
communications, non-linear optics and optical manipulations.
\section{The Method}
\hspace*{4ex} Let us first present the basic equations that describe the
evolution of a pulse in a linear media. A pulse, $\Psi(\mathbf{r},t)$, with slowly varying envelope, propagating in a dispersive absorbing medium with complex index of refraction,
$n(\omega) = n_R(\omega) + i n_I(\omega)$, can be
described as:
\begin{equation} \Psi(\mathbf{r},t) \; = \; e^{i k(\omega_0) z}e^{-i \omega_0 t}
A(\mathbf{r},t), \label{psi} \end{equation}
with
\begin{equation} k(\omega_0) \; = \; n_R(\omega_0)\frac{\omega_0}{c} + i
n_I(\omega_0)\frac{\omega_0}{c} = k_R(\omega_0) + ik_I(\omega_0) \equiv k_{R
0} + i k_{I 0}\ . \end{equation}
The pulse envelope, $A(\mathbf{r},t)$, obeys
\begin{equation}
\frac{1}{2k(\omega_0)}\nabla^2_{\perp} A + i\left(\frac{\partial}{\partial z} +
\beta_1\frac{\partial}{\partial t}\right)A -
\frac{\beta_{2}}{2}\frac{\partial^{2}A}{\partial t^{2}} \; = \; 0,
\label{eq3D}
\end{equation}
where
\begin{equation} \beta_1 \; = \; \frac{\partial k_R}{\partial \omega}|_{\omega_0} \,\,\,\,\, {\rm
and} \,\,\,\,\, \beta_2 \; = \; \frac{\partial^2 k_R}{\partial \omega^2}|_{\omega_0}
\,\,. \end{equation}
\
\hspace*{4ex} By making the usual transformations, $z=z$ and $T = t-z/v_g$,
with $v_g = 1/\beta_1$, we can rewrite Eq.(\ref{eq3D}) as
\begin{equation}
\frac{1}{2k(\omega_0)}\nabla^2_{\perp} A + i\frac{\partial}{\partial z}A -
\frac{\beta_{2}}{2}\frac{\partial^{2}A}{\partial T^{2}} \; = \; 0\ .
\label{eq3D2}
\end{equation}
\
\hspace*{4ex} Now, let us consider the pulse envelope, $A(\mathbf{r},t)$, having the
following form
\begin{equation} A(\mathbf{r},t) \; = \; W(\rho,\phi,z)P(z,T) \,\, . \label{A}\end{equation}
By substituting Eq.(\ref{A}) in (\ref{eq3D2}), we have
\begin{equation} \frac{1}{2 k(\omega_0)}\nabla^2_{\perp} W + i\, \frac{\partial W}{\partial z} \; = \; 0, \label{W} \end{equation}
and
\begin{equation} i\,\frac{\partial P}{\partial z} - \frac{\beta_2}{2}\frac{\partial^2 P}{\partial T^2} \; = \;
0\ . \label{P} \end{equation}
At this point it is important to note that $W(\rho,\phi,z)$, governed by Eq.(\ref{W}),
is the partial differential equation for the envelope of a
paraxial beam; while $P(z,T)$, governed by Eq.(\ref{P}), is
the partial differential equation for the envelope of a 1D pulse propagating in a
dispersive medium.
\hspace*{4ex} Having established the differential equations governing the behavior of $W(\rho,\phi,z)$ and $P(z,T)$, in the next subsection we will develop a space-time modeling of vortex pulses (i.e. pulses carrying OAM) propagating in unguided, dispersive, and absorbing medium. Consequently, we shall see that it is possible to construct vortex pulses which resist the effects of dispersion, diffraction, and attenuation as they propagate in an unbounded medium. Such construction and formulation can be viewed as the generalization of the Localized Wave theory in the paraxial regime.
\subsection{Space-time Modeling of Diffraction, Dispersion, and Attenuation
Resistant Vortex Pulses}
\hspace*{4ex} The basic approach in developing vortex pulses, $A(\mathbf{r},t)$, that are immune to diffraction, attenuation, and dispersion is to enforce resistance to diffraction and attenuation (along with control of the OAM) through the spatial function $W(\rho,\phi,z)$ and its corresponding differential equation, Eq.(\ref{W}); whereas to make the vortex pulses resistant to dispersion through the temporal function, $P(z,T)$, and its corresponding differential equation, Eq.(\ref{P}).
\hspace*{4ex} Similar to the approach in \cite{fwbg}, we choose $W(\rho,\phi,z)$ to be a superposition of $2N+1$ copropagating $\nu$-order Bessel-Gauss beams given by
\begin{equation}
W(\rho,\phi,z) = \frac{\exp\left(\displaystyle{-q^2\frac{\rho^2}{\mu}}\right)}{\mu}
\exp\left(\displaystyle{- i k(\omega_0) \frac{z}{\mu}}\right)\sum^{N}_{m=-N}A_{m}
J_{\nu}\left(\eta_m\frac{\rho}{\mu}\right)e^{i\nu\phi}\,
\exp\left(\displaystyle{i \zeta_m \frac{z}{\mu}}\right), \label{FWbg}
\end{equation}
where
\begin{equation} \mu \; = \; 1 + i 2 \frac{q^2}{k(\omega_0)} z \,\,. \label{mu} \end{equation}
In Eq.(\ref{FWbg}), $q$ is a constant, $A_m$ are coefficients of expansions to be determined, and $\eta_m$ and $\zeta_m$ are
the transverse and longitudinal wavenumbers (respectively) of the
mth Bessel-Gauss beam in the superposition that must satisfy
$\zeta_m = k(\omega_0) - \eta_m^2/2k(\omega_0)$.
\hspace*{4ex} The solution (\ref{FWbg}) can be used
for obtaining a light beam, given by $ \exp(-k_{I 0}z)\exp(i
k_{R 0} z - i \omega_0 t)W(\rho,\phi,z)$, resistant to the diffraction
and attenuation effects in an absorbing medium, with a longitudinal
intensity pattern that can be chosen \emph{a priori}. This intensity
pattern is given by a function, $|F(z)|^2$, of our choice,
and can be concentrated over the z axis (in the case $\nu=0$), with a
beam spot radius $r_0$, or over a cylindrical surface (in the case $|\nu| \geq 1$) of radius
$\rho_{\nu}$, both also of our choice, as we are going to explain soon. More specifically, one can have\footnote{Here, $\rho_0 \equiv 0$.} $\exp(-2k_{I
0}z)|W(\rho=\rho_{\nu},\phi,z)|^2 \approx |F(z)|^2$ within a
predefined longitudinal range $0 \leq z \leq L/2$, and
$\exp(-2k_{I 0}z)|W(\rho=\rho_{\nu},\phi,z)|^2 \approx 0$ for $0>z>L/2$.
\hspace*{4ex} It is important to say that while in the FW method $W$ shapes the beam envelope, here, due to Eqs.(\ref{psi},\ref{A}), such modeling takes place on the pulse, i.e., $|F(z)|^ 2$ becomes the pulse's peak intensity pattern along the propagation, the same occurring for the transverse features, i.e., $ r_0 $ becomes the pulse's spot radius (when $ \nu = 0 $) and $ \rho_ {\nu} $ becomes the radius of the vortex pulse, which has a donut-shaped profile.
For such modelling, the finite energy FW method \cite{fwbg} requires the following choices
for $\eta_m$ in the solution (\ref{FWbg}):
\begin{equation} \eta_m \; = \; \sqrt{2}\displaystyle{\sqrt{1 - \frac{1}{k_{R
0}}\left(Q+\frac{2\pi m}{L} \right) }} \, |k(\omega_0)| \,\, ,
\label{etan} \end{equation}
which implies $\zeta_m = \zeta_{r\,m} + i\, \zeta_{i\,m}$, with
\begin{equation} \left\{\begin{array}{l}
\zeta_{r\,m} \; = \; Q + \displaystyle{\frac{2 \pi m}{L}} \\
\\
\zeta_{i\,m} \; = \; k_{I 0} \displaystyle{\left(2 - \frac{\zeta_{r\,m}}{k_{R 0}}
\right)} \,\, ,
\end{array} \right. \label{zetam}
\end{equation}
where $Q$ is a positive constant, obeying $ 0 \leq Q + 2\pi N/L \leq k_{R 0} $
and being related
to the transverse dimensions of the beam. Actually, when $\nu=0$, the pulse spot
radius, $r_0$, will be approximately given by $r_0 \approx 2.4/\eta_0$ and,
when $|\nu| \geq 1$, the radius of the donut-shaped vortex pulse will be
approximately given by $\rho_{\nu}$, which corresponds to the first positive
root of $[(d/d\rho)J_{\nu}(\eta_0 \rho)]|_{\rho=\rho_{\nu}} = 0$.
Yet, according to \cite{fwbg}, the values of $q$ and of the coefficients $A_m$
have to be given by
\begin{equation} q \; = \; \frac{2 k_{R 0}}{L \, \eta_0} \label{q2} \end{equation}
and
\begin{equation} A_{m} = \frac{1}{L} \int^{L}_{0}\frac{F(z)}{G(z)}\,
e^{-i\frac{2\pi}{L}m z} dz \,\, , \label{An2} \end{equation}
with
\begin{equation} G(z) \; = \; e^{- k_{I 0} z}\exp\left(\frac{-\eta_0^2(k_{I 0} + 2
q^2 z)z}{2 [k_{R 0}^2 + (k_{I 0} + 2 q^2 z)^2]}\right) I(z)
\label{G} \end{equation}
where $I(z)$\footnote{The function $I(z)$ does not appear in the work of the
finite energy FWs \cite{fwbg} because there the method deals only with beams,
that is, it considers $P (z, T) = 1$.} is a function that is related, as we
are going to
see, with a possible intensity decay of $P(z,T)$.
\hspace*{4ex} So, by considering $W(\rho,\phi,z)$ given by Eq.(\ref{FWbg}), with
Eqs.(\ref{mu}-\ref{G}), we are ensuring that our 3D pulse solution
given by Eqs.(\ref{psi},\ref{A}): a) can be spatially modelled on demand,
presenting,
as a subproduct, resistance to the diffraction and attenuation
effects for long distances; b) it is endowed with OAM, being its topological
charge equal
to $\nu$. The distances of diffraction and attenuation resistance, that we
call $Z_{diff}$ and $Z_{att}$,
respectively, will be set by the morphological function $F(z)$ and so, as
that function is set to be null
for $z>L/2$, they can be considered as $Z_{diff} = Z_{att} = L/2$.
\hspace*{4ex} Here, we have to stress an important point. While the diffraction resistance
distance can be made arbitrarily large (even experimentally it can be made
hundred of times greater than the Rayleigh distance, $L_{diff}\approx \sqrt{3}\,k_{R0} r_0^2$),
the same does not occur with the attenuation resistance distance which, according
to the FW method \cite{fw2}, should not exceed about 10 times the penetration
depth $L_{att} = 1/\alpha$ (where $\alpha = 2 n_I \omega /c $ is the absorption
coefficient), otherwise the lateral lobules can acquire very high intensity
levels. Due to this, the parameter $L$, used to set up the range $0 \leq z \leq L/2$
where the morphological function $F(z)$ is non-null, is limited to values about $10/\alpha$.
\hspace*{4ex} Now, let us move on and make a choice to $P(z,T)$, a solution of
(\ref{P}) which, we expect, shall provide resistance to the
dispersion effects to the resulting 3D pulse envelope.
For this purpose, a natural choice is a 1D Airy-type pulse,
as the finite energy Airy-exponential one given by
\begin{eqnarray}
P\left(z,T\right) & = &
\exp\left[\frac{6a\left(2\varepsilon\tau-Z^{2}\right)+
iZ\left(-6a^{2}-6\varepsilon\tau+Z^{2}\right)}{12}\right]\nonumber \\
& \times & \mathrm{Ai}\left(\varepsilon\tau-\frac{Z^{2}}{4}-iaZ\right)\,\, , \label{airyfinito}
\end{eqnarray}
where $\mathrm{Ai}(.)$ is the Airy function, $a>0$ is a constant related
with the time exponential apodization and $\varepsilon = \pm 1$ determines the
direction of the Airy pulse envelope \cite{airycausal}, which can have the smaller peaks preceding
the main lobe ($\varepsilon = 1$) or the main lobe preceding the smaller
peaks ($\varepsilon=-1$). In Eq.(\ref{airyfinito}), we have used the normalized variables $\tau=T/T_{0}$ and
$Z=z\beta_{2}/T_{0}^{2}$, with $T_0$ a constant equal to the
initial time width of the main peak of the 1D pulse $P(z,T)$
which, depending on the value of the parameter $a$, can resist to
the dispersion effects for distances much longer than the usual
dispersion distance, $L_{disp} = T_0^2 / |\beta_2|$, of ordinary
gaussian pulses. A conservative estimation of the field
depth, $Z_{disp}$, of the Airy-Exponential pulse (\ref{airyfinito})
is given by
\begin{equation} Z_{disp} \; = \; \sqrt{\frac{2}{a}}\,\frac{T_0^2}{|\beta_2|} \; = \;
\sqrt{\frac{2}{a}} L_{disp} \,\,\, . \label{a}\end{equation}
\hspace*{4ex} Although the finite energy 1D pulse solution exhibits resistance to the
dispersion effects, it also exhibits a continuous intensity decay from $z=0$.
For distances smaller than $Z_{disp}$, it is not difficult to show that the
pulse's peak intensity decays approximately according to the function
\begin{equation} I(z) \; = \; \exp\left(-\frac{a\beta_2^2}{4T_0^4}z^2 \right) \,\,. \label{I} \end{equation}
The intensity decay given by (\ref{I}) is taken into account (i.e., it is
compensated) by $W(\rho,\phi,z)$, Eq.(\ref{FWbg}), through the coefficients
$A_m$ given by Eqs.(\ref{An2},\ref{G}).
\hspace*{4ex} In this way, with $P(z,T)$ given by Eq.(\ref{airyfinito}), we are ensuring
that our 3D pulse, Eqs.(\ref{psi},\ref{A}), is also resistant to the dispersion
effects for long distances.
\hspace*{4ex} At this point, an important observation has to be made about the 1D pulse
solution given by Eq.(\ref{airyfinito}). For causality reasons, which are very
well explained in \cite{airycausal}, in the case $\varepsilon = -1$, expression
(\ref{airyfinito}) should not be considered for $z \geq 2T_0^3/\beta_2^2 v_g$.
In our case, however, such limitation is not alarming at all because, according
to our method, the morphological function $F(z)$ ensures that the resulting pulse
will possess negligible intensities for $z \geq L/2$ which, in general, will be
set to values much smaller than $2T_0^3/\beta_2^2 v_g$.
\hspace*{4ex} The final result is that the resulting 3D pulse,
$\Psi(\rho,\phi,z,t)$, is given by
\begin{eqnarray}
\Psi(\rho,\phi,z,t) & = & e^{-k_{I 0} z}e^{ik_{R 0} z}e^{-i \omega_0
t} \nonumber \\
& \times & \left[
\frac{\exp\left(\displaystyle{-q^2\frac{\rho^2}{\mu}}\right)}{\mu}
\exp\left(\displaystyle{- i k(\omega_0) \frac{z}{\mu}}\right)\sum^{N}_{m=-N}A_{m}
J_{\nu}\left(\eta_m\frac{\rho}{\mu}\right)e^{i\nu\phi}\,
\exp\left(\displaystyle{i \zeta_m \frac{z}{\mu}}\right) \right] \nonumber \\
& \times & \left[ \exp\left[\frac{6a\left(2\varepsilon\tau-Z^{2}\right)+
iZ\left(-6a^{2}-6\varepsilon\tau+Z^{2}\right)}{12}\right]
\mathrm{Ai}\left(\varepsilon\tau-\frac{Z^{2}}{4}-iaZ\right) \right] \,\, ,
\label{psi2}
\end{eqnarray}
with $\mu$, $q$, $\eta_m$, $\zeta_m$ and $A_m$ given by
Eqs.(\ref{mu},\ref{q2},\ref{etan},\ref{zetam},\ref{An2}), being
that the parameter $a$ can be estimated from Eq.(\ref{a}) once we
have chosen the desired distance ($Z_{disp}$) of dispersion
resistance for the pulse.
\hspace*{4ex} The new pulse solutions represented by Eq.(\ref{psi2}),
propagating in dispersive, absorbing and unguided media and
carrying OAM, are resistant to the concomitant effects of
diffraction, dispersion and attenuation for long distances.
Actually, we can construct such pulses in such a way that we can
choose where and how intense their peaks will be within the
longitudinal spatial range $0 \leq z \leq L/2$.
\hspace*{4ex} A natural question that arises is about the generation process of these new
pulses. Interestingly, this can be done in a relatively simple way, through a
combination of the FW-type beams generation techniques with those for the generation
of Airy-type pulses. In this sense, through a Gaussian frequency spectrum with a
cubic phase \cite{nature}, an Airy-Gaussian pulse is created and, thereafter, it
is addressed onto a spatial light modulator, which encodes on it the FW's
hologram transmission function. The desired resulting pulse can then be obtained
after a 4f optical system and an iris.
\section{Examples}
\hspace*{4ex} In this section we will present two examples obtained from our theoretical method.
The material medium considered here is a SF10 glass and the central (free space)
wavelength used here is $\lambda_0 = 454.6$nm, which corresponds to an angular
frequency $\omega_0 = 4.1464 \times 10^{15}$Hz.
\hspace*{4ex} At this frequency, the SF10 glass presents the following optical constants:
refractive index $n = n_R + in_I$, with $n_R=1.7554$ and $n_I= 1.5051 \times 10^{-7}$;
absorption coefficient $\alpha = 0.0416 {\rm cm}^{-1}$, which implies an
attenuation length $L_{att} = 1/\alpha = 24$cm; $\beta_1 = 6.3383 \times 10^{-9}$s/m,
which implies $ v_g=1.5777 \times 10^8 $m/s; dispersion coefficient
$\beta_2 = 459.73 {\rm fs}^2/$mm.
\hspace*{4ex} The following two examples deal with a vortex and a non-vortex pulse,
respectively, both with $\varepsilon=-1$ and a main lobe of duration $T_0 = 200$fs,
which implies a dispersion length $L_{disp}=T_0^2/|\beta_2| = 8.7$cm. In the first
case, the vortex pulse of topological charge $\nu=4$ has a donut shape of radius
approximately $\rho_4 =49\mu$m, which implies a Rayleigh distance (diffraction
length) $L_{diff} \approx 10$cm; in the second case, the non-vortex pulse has a
spot radius $r_0 = 22.1 \mu$m, which implies a Rayleigh distance
$L_{diff} = \sqrt{3}k_{R0}r_0^2 \approx 2$cm. The resulting pulses will be
designed to resist the effects of diffraction, dispersion and attenuation
till $z = 87$cm.
\subsection{First case}
\hspace*{4ex} Let us use our solution given by Eq.(\ref{psi2}), with $\varepsilon=-1$, to
construct a diffraction-dispersion-attenuation
resistant vortex pulse with topological charge $\nu = 4$, radius of the
donut-shaped-pulse $\rho_4 \approx 49 \mu$m, main lobe of duration $T_0 = 200$fs
and whose peak's intensity pattern along the propagation is given by a on-off-on
pattern, i.e, it is dictated by $|F(z)|^2$ with:
\begin{equation} F(z) \; = \; [H(z - \ell_1)-H(z-\ell_2)] + [H(z-\ell_3)-H(z-\ell_4)] \,\,\, , \label{F1} \end{equation}
where $H(.)$ is the Heaviside function and $\ell_1=0$, $\ell_2=29$cm, $\ell_3=58$cm
and $\ell_4=87$cm. To this case, we can choose $L=1.74$m and, according to the
desired donut-shaped-pulse's radius, the value of the parameter $Q$ results to
be $Q = 0.99999 n_R\omega/c$. The morphological function $F(z)$, besides pre-defining
the pulse's peak intensity behavior, also defines the diffraction-attenuation
resistance distance\footnote{According to our method, such characteristics are
acquired by the resulting pulse, $\Psi(\rho,\phi,z,t) = W(\rho,\phi,z)\,P(z,T)$,
through the function $W(\rho,\phi,z)$.}, which in this case is $87$cm. To get
dispersion resistance for such a distance, we have to set $a=0.02$. Now, the
resulting pulse (\ref{psi2}) can be completely characterized through
Eqs.(\ref{q2}-\ref{G}). In this case, we use $N=60$.
\hspace*{4ex} Figure 1 shows the comparison between the desired intensity pattern for the
pulse's peak along the propagation, given by
$|F(z)|^2$, and $|\Psi(\rho=\rho_4, \phi, z, t = t(z)) |^2$, which is the evolution
of the resulting pulse's peak intensity, which occurs, approximately,
on $ \rho = \rho_4 = 49 \mu$m and at times given by
$t = -(\beta_2^2/4T_0^3)z^2 + z/v_g + 1.019 T_0$ for the a given position
$z$ of the pulse's peak. We can see a good agreement between them.
\begin{figure}[!h]
\begin{center}
\scalebox{.35}{\includegraphics{Fig1.jpg}}
\end{center}
\caption{Comparison between the desired intensity pattern for the pulse's peak
along the propagation (continuous red line), given by $|F(z)|^2$,
and $|\Psi(\rho=\rho_4, \phi, z, t = t(z)) |^2$, which is the evolution
of the resulting pulse's peak intensity (doted line), which occurs,
approximately, on $ \rho = \rho_4 = 49 \mu$m and at
times $t = -(\beta_2^2/4T_0^3)z^2 + z/v_g + 1.019 T_0$ for the a given position
$z$ of the pulse's peak. We can see a good agreement between them.
} \label{fig1}
\end{figure}
\hspace*{4ex} Figure 2 shows the temporal evolution of the resulting pulse at different
distances (i.e., values of $z$) and considering $\rho = \rho_4 = 49 \mu$m.
It is evident that the pulse intensity obeys the required on-off-on behavior
and also that it preserves the temporal width of its main lobe till the distance
pre-defined by the morphological function $F(z)$.
\begin{figure}[!h]
\begin{center}
\scalebox{.35}{\includegraphics{Fig2.jpg}}
\end{center}
\caption{Temporal evolution of the resulting pulse at different distances and
considering $\rho = \rho_4 = 49 \mu$m. One can see that the pulse intensity
obeys the required on-off-on behavior and also that it preserves the temporal
width of its main lobe till a distance 10 times greater than the dispersion
length of an ordinary pulse with the same temporal width.} \label{fig2}
\end{figure}
\hspace*{4ex} Finally, Fig.3 shows the $3D$ pulse intensity, $|\Psi(\rho,\phi,z,t)|^2$,
at nine
different instants of time. The first, second and third lines of the subfigures
show the pulse evolution within
the ranges $\ell_1<z<\ell_2$, $\ell_2<z<\ell_3$ and $\ell_3<z<\ell_4$, respectively.
We can see that, in addition to the vortex pulse having the desired space-time
evolution, it is resistant to the effects of the diffraction, dispersion and absorption.
\begin{figure}[!h]
\begin{center}
\scalebox{.37}{\includegraphics{Fig3.jpg}}
\end{center}
\caption{The $3D$ pulse intensity, $|\Psi(\rho,\phi,z,t)|^2$, of the first example
at nine
different instants of time. The first, second and third lines of the subfigures
show the vortex pulse evolution within
the ranges $\ell_1<z<\ell_2$, $\ell_2<z<\ell_3$ and $\ell_3<z<\ell_4$, respectively.
We can see that, in addition to the vortex pulse having the desired space-time
evolution, it is resistant to the effects of the diffraction, dispersion and
absorption.} \label{fig3}
\end{figure}
\subsection{Second case}
\hspace*{4ex} Here, we use our solution, Eq.(\ref{psi2}), to construct a
diffraction-dispersion-attenuation
resistant pulse with null topological charge, i.e. $\nu = 0$, spot radius
$r_0 \approx 22 \mu$m, main lobe of duration $\tau_0 = 200$fs and whose
peak's intensity pattern along the propagation is given by a ladder
pattern, i.e, it is dictated by
$|F(z)|^2$ with:
\begin{equation} \begin{array}{clr} F(z) &=
\sqrt{1}\,[H(z-\el1)-H(z-\el2)]\\
\\
&+ \sqrt{2}\,[H(z-\el2)-H(z-\el3)] \\
\\
&+ \sqrt{3}\,[H(z-\el3)-H(z-\el4)]
\,\, ,
\end{array} \label{F2} \end{equation}
where $H(.)$ is the Heaviside function and, as before, $\ell_1=0$, $\ell_2=29$cm,
$\ell_3=58$cm and $\ell_4=87$cm. To this case, we can again choose $L=1.74$m and,
according to the desired pulse spot radius, the value of the parameter $Q$
results to be $Q = 0.99999 n_R \omega/c$. The morphological function $F(z)$,
besides pre-defining the pulse's peak intensity behavior, also defines the
diffraction-attenuation resistance distance, which in this case is $87$cm.
For obtaining dispersion resistance for such a distance, we have to set
$a=0.02$ and the resulting pulse (\ref{psi2}) can be completely characterized
through Eqs.(\ref{q2}-\ref{G}).
\hspace*{4ex} Figure 4 compares the desired intensity pattern for the pulse's peak along
the propagation, given by $|F(z)|^2$, with
$|\Psi(\rho=0,z, t = t(z))|^2$, the evolution of the resulting pulse's peak
intensity, which occurs on $\rho = 0$ and at
times $t = -(\beta_2^2/4T_0^3)z^2 + z/v_g + 1.019 T_0$ for the a given
position $z$ of the pulse's peak. We can see a good agreement between them.
\begin{figure}[!h]
\begin{center}
\scalebox{.55}{\includegraphics{Fig4.jpg}}
\end{center}
\caption{Comparison between the desired intensity pattern for the pulse's peak
along the propagation (continuous red line), given by $|F(z)|^2$,
with $|\Psi(\rho=0,z, t = t(z))|^2$, the evolution of the resulting pulse's
peak intensity (doted line), which occurs on $\rho = 0$ and at times
$t = -(\beta_2^2/4T_0^3)z^2 + z/v_g + 1.019 T_0$ for the a given position
$z$ of the pulse's peak. We can see a good agreement between them} \label{fig4}
\end{figure}
\hspace*{4ex} The on-axis ($\rho=0$) temporal evolution of the resulting pulse at
different distances is shown in Fig.5. It is very clear that the pulse
intensity obeys the required ladder behavior, preserving the temporal
width of its main lobe till the distance pre-defined by the morphological
function $F(z)$.
\begin{figure}[!h]
\begin{center}
\scalebox{.35}{\includegraphics{Fig5.jpg}}
\end{center}
\caption{On-axis ($\rho=0$) temporal evolution of the resulting pulse at
different distances. One can see that the pulse intensity obeys the required
ladder behavior, preserving its temporal width till a distance 10 times
greater than the dispersion length of an ordinary pulse with the same temporal
width.} \label{fig5}
\end{figure}
\begin{figure}[!h]
\begin{center}
\scalebox{.37}{\includegraphics{Fig6.jpg}}
\end{center}
\caption{The $3D$ pulse intensity, $|\Psi(\rho,z,t)|^2$, at nine
different instants of time. The first, second and third lines of the subfigures
show the pulse evolution within
the ranges $\ell_1<z<\ell_2$, $\ell_2<z<\ell_3$ and $\ell_3<z<\ell_4$, respectively.
In addition to the desired pulse's space-time evolution, one can see the pulse
is resistant to the concomitant effects of diffraction, dispersion and
absorption.} \label{fig6}
\end{figure}
\hspace*{4ex} Finally, Fig.6 depicts the $3D$ pulse intensity, $|\Psi(\rho,z,t)|^2$, at nine
different instants of time. The first, second and third lines of the subfigures
show the pulse evolution within
the ranges $\ell_1<z<\ell_2$, $\ell_2<z<\ell_3$ and $\ell_3<z<\ell_4$, respectively.
We can see, in addition to the desired pulse's space-time evolution, it is
resistant to the concomitant effects of diffraction, dispersion and absorption.
\section{Conclusions}
In this work we have developed a method capable of providing, in unguided dispersive and absorbing media, vortex pulses resistant to the three concomitant effects of diffraction, dispersion and attenuation. As a matter of fact, with our approach it is possible to perform a space-time modelling on such new pulses, i.e., it allows the choice of multiple spatial ranges where the pulse intensities can be chosen a priori.
Such approach is a result of a fusion between two important theoretical methodologies, one related to the so called Frozen-Wave-beams, which are non-diffracting beams whose spatial intensity pattern can be chosen \emph{a priori} in
absorbing media, the other related to the Airy-Type pulses, which
are pulses resistant to the dispersion effects in material dispersive media.
The new kind of pulses can have potential applications in
different fields as photonics, nonlinear optics, optical
communications, optical tweezers, optical atom guiding, medicine,
etc..
\section*{Acknowledgements}
Thanks are due to partial support from FAPESP (under grant 2015/26444-8) and from CNPq
(under grant 304718/2016-5).
|
\section{Introduction}
Ion implantation is a mature technology \cite{Hamm12}. In most implanters, a single ion beam is
formed by extraction of ions from an ion source. Ions are then accelerated and a desired species
is selected for implantation. The achievable ion current density is limited by space charge forces
and the total ion current is limited by the size of the extraction aperture from which ions
can be extracted to form a beam with low enough emittance for efficient transport in the beam line
\cite{Brown2004}. The concept of multi-beam ion accelerators was developed in the late 1970s by
Maschke et al. with the concept of a Multiple Electrostatic Quadrupole Linear Accelerator (MEQALAC)
\cite{Maschke_1979b}. MEQALACs are RF-driven linear accelerators where the total ion current can be
scaled by adding more beams and the ion kinetic energy can be increased by adding accelerator modules.
In the first implementations, MEQALACs used RF-cavities to achieve ion acceleration with high
voltages driven at frequencies in the \SI{25}{\MHz} range \cite{URBANUS1989508}. Multi-beam
RF-linacs can boost the total current for applications where beams of mass analyzed single species
are required. We have recently reported on the development of multi-beam RF accelerators that we
assemble from stacks of low cost wafers \cite{Persaud:RSI-2016,Seidl2018}. We form RF-acceleration
structures and electrostatic quadrupole (ESQ) focusing elements on printed circuit board and silicon
wafers with 10-cm diameter using standard microfabrication techniques \cite{Vinaya2018}. In this article
we report on the use of compact GaN based RF amplifiers to accelerate ions in an array of $3\times3$
beams with energy gains of up to \SI{5.1}{\kilo\volt} per RF-acceleration unit. This method for
generating RF-acceleration voltages leads to an effective acceleration gradient of about
\SI{0.3}{\mega\volt\per\m} in our multi-beam RF linac. While this is much lower than for
conventional RF accelerators based on cavities, our implementation is very compact with high
current density and is formed from low cost components in a modular architecture.
Figure~\ref{fig:concept} shows a schematic and a photo of the wafer based multi-beam RF-linac.
Ions are extracted from a $3\times3$ array of apertures (\SI{0.5}{\mm} diameter) from a filament driven
multi-cusp ion source and accelerated to 5 to \SI{10}{\keV}. We use argon and helium ions
for beam tests and the ion source can also provide ions of common dopant species. Following
a matching section comprised of a series of ESQ focusing elements, the ion beams reach the
lattice of RF acceleration units and ESQs. We determine the energy gain by scanning the
bias on a retarding grid while monitoring the ion current in a Faraday cup.
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{2-unit-matching-setup.pdf}
\includegraphics[width=0.6\linewidth]{pic-wafer.jpg}
\caption{Schematic (top) and photo (bottom) of the multi-beam RF linac with an array of
$3\times3$ beam apertures.}
\label{fig:concept}
\end{figure}
The RF acceleration units are formed from four wafers, with the outside wafers held at ground and the
central two wafers driven by high voltage pulses at the selected RF frequency. Ions entering an
RF acceleration unit are accelerated by the RF voltage and then drift in the field-free gap
between the two central wafers before further acceleration in the gap between the third wafer
and the grounded last wafer in the RF acceleration unit. The drift distance is set to match the
phase advance of the RF at a given frequency according to $x=0.5\,\, v/f$, where $x$ is the gap distance,
$v$ the ion velocity and $f$ the RF frequency. For operation at \SI{13.5}{\MHz} with argon ions injected
at \SI{10}{\keV} ($v_{Ar}=\SI{2.2e5}{\meter\per\second}$) the drift distances are \SI{8}{\mm}. After several stages of acceleration the RF frequency
can be increased to compensate for increasing ion velocities and to keep drift distances small.
In our first proof-of-concept of this approach of multi-beam ion acceleration we used an
external LC-tank circuit to generate RF acceleration voltages of 600 V/gap \cite{Persaud:RSI-2016}.
We have now implemented compact GaN based RF sources \cite{Airity} that are placed near the
acceleration wafers inside the vacuum chamber. The high voltage pulses are delivered through
low capacitance wires to the RF acceleration wafers. One RF source can drive ion acceleration
in several modules with up to \SI{2.6}{\kV} per acceleration gap.
\section{RF source}
A custom RF power source was developed by Airity technologies to generate the high
voltage AC required to drive the wafer array. By leveraging the intrinsic inductive and
capacitive elements of the wafers and incorporating them into the source design, a compact and efficient vacuum compatible RF source was built that enables placement close to the wafer array to minimize cable effects. The
specifications of the source are summarized in Figure ~\ref{fig:RFamplifier}.
Initial tests of an updated RF source design suggest that voltages exceeding
\SI{5}{\kV} are possible. Future work will further investigate such designs with
the goal of scaling array voltage to \SI{10}{\kV}.
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{AirityRFamplifier3.pdf}
\caption{Photo and specifications of the RF amplifier}
\label{fig:RFamplifier}
\end{figure}
\section{Modeling}
To simulate the system and investigate its scaling behaviour an equivalent circuit was developed.
The RF source was modeled as a simple $RLC$ circuit shown in Figure~\ref{fig:source-schematic}.
Since the RF source has been developed to drive a \SI{120}{\pico\farad} load, the capacitance
in the model has been split into an internal \SI{50}{\pico\farad} capacitance, $C_1$, and
the external load. In our case our load of a single RF unit is smaller than the load for which the unit
was designed and we therefore augmented the load capacitance with a parallel capacitor, $C_2$.
When driving more units this capacitor should be adjusted accordingly. The resistor $R$ in
the source model varies with the driving voltage. This was added to compensate for the observed
behaviour of lower gains at higher voltages, possibly due to heating of the RF driver.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.7\linewidth]{source-schematic.pdf}
\caption{The equivalent circuit used to model the RF source. The resistor
$R=(1.212\times\frac{V_{drive}}{\SI{103}{V}})\SI{}{\ohm}$ captures variations in regards to
the source voltage. The capacitor $C_1$ and inductor $L$ represent
internal values of the RF source. The load and $C_2$ represent the external load.}
\label{fig:source-schematic}
\end{figure}
The RF units were modeled as a series of cables with inductance and capacitance. The acceleration
gaps were modeled as purely capacitive and since the wafers are connected with a short wire loop, a
inductance was added between them. The complete circuit is shown in Figure~\ref{fig:wafer-schematic}.
The circuit model includes a capacitance to represent the BNC connector, a cable resistance (the cable
resistance was placed to the left of the capacitor due to requirements of the simulations software),
a test resistor (see below), a network of capacitors and inductors to represent the cable from the
RF source to the wafers, and the capacitance of the acceleration gap including an inductance that
represents the ground connection between the wafers.
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{wafer-schematic.pdf}
\caption{The equivalent circuit used to model a RF unit including the cables. The two marked areas
on the left represent wires and connectors for the two RF-acceleration gaps modeled as two capacitors
(right box) and an inductive cable. Several of these circuits in parallel were used to model
scaling behaviour when driven by the RF source. The capacitor $R_{test}$ was only added during
some measurements to get better results from a vector network analyzer.}
\label{fig:wafer-schematic}
\end{figure}
To fit the parameters of the models, a RF unit was analyzed using a vector network analyzer (VNA),
Keysight Technology N9923A, in the frequency range of \SI{5}{\mega\hertz}-\SI{30}{\mega\hertz}
and also in the range \SI{5}{\mega\hertz}-\SI{150}{\mega\hertz}. Furthermore, the capacitance was
measured directly. The VNA measurement ($S_{11}$ mode) showed a resonance behaviour around \SI{90}{\mega\hertz}
which allowed us to calculate the inductance using the measured capacitance. Furthermore, since
we know the source impedance of the VNA (\SI{50}{\ohm}), we can calculate the load impedance at,
for example, \SI{15}{\mega\hertz} and from there estimate the resistance. However, due to the small
resistance of the system, it is difficult to measure this value directly. We therefore added
a known resistor, $R_{test}$, to the system, which made it easier to estimate the total resistance
using the VNA. The estimated value is in agreement with direct measurements and with calculated
values when taking the skin effect into account.
Simulating the modelled circuits using LTspice~\cite{LTspice} and Matlab RF toolbox gave
good agreement in the frequency range of operation (around \SI{15}{\mega\hertz}) with an
error between measurement and simulation smaller than $1\%$.
Furthermore, the RF source allowed measurement of the output voltage using a voltage
monitor output and we were also able to measure the voltage directly using a high
voltage probe (Tektronix P6015A). Figure~\ref{fig:model} compares the simulated and measured results when one to three RF units are used. The figure also shows that there is a slight shift in resonance frequency. This is due to the fact that we did not change the capacitor $C_2$ to optimally balance the load impedance of the RF source. This can also explain the slight decrease in gain between the different setups.
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{freq-response.pdf}
\caption{Frequency response of the modelled and measured circuits using one to three RF units.}
\label{fig:model}
\end{figure}
\section{Beam Experiment}
The beam energy was measured using the same setup and method as described previously\cite{Persaud:RSI-2016}.
Ions were produced using a filament-driven ion
source operated with argon at a pressure of \SI{1}{\pascal} (\SI{7.5}{\milli\torr}). The source was
floated at a high voltage of \SI{6}{\kilo\volt} and ions were extracted from a $3\times3$~array of extraction
apertures using an extraction voltage of \SI{400}{\volt}. The ions are then accelerated to
ground potential where they enter the RF units. Here, the ions are either accelerated or decelerated
depending on their arrival time with respect to the phase of the RF voltage in the acceleration gaps.
In our experiments we utilized a long pulse (\SI{300}{\micro\second}) compared to the
RF frequency whereas in a final implementation of the accelerator one would bunch the beam and
only inject beam packages at the correct phase. We then select ions above a certain beam energy using a
deceleration grid and measure the beam current in a Faraday cup. Due to the almost continuous beam injection,
we expect a distribution of beam energies that we can model using a 1D simulation code. The length of the
drift sections were chosen for optimal transport of the accelerated ions. We can therefore either fit the
beam energy using the simulation code or read of the maximal beam energy gain and divide this by the
number of acceleration gaps to get the applied RF voltage. Experiments were carried out using one
to three RF units.
Figure~\ref{fig:scan} shows the result of such a voltage scan on the retarding grid for a system using
two RF units (four acceleration gaps). The initial beam energy was \SI{6}{\keV} and one can clearly
see that due to the continuous injection we measure all beam energies from 0-\SI{16}{\keV}. The
maximum energy gain was \SI{10.2}{\keV} equivalent to an applied voltage of \SI{2.56}{\keV} per
acceleration gap. The signal agrees well with our expectation from a 1d model \cite{Persaud:RSI-2016}.
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{single-scan.pdf}
\caption{Measured result of a grid voltage scan for a system using two RF units driving
the RF source with \SI{65}{\volt}. The measurement shows that \SI{2.6}{\kilo\volt} per
acceleration gap was achieved. The vertical line shows the beam energy before extraction (\SI{6}{\keV}).}
\label{fig:scan}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{V_gap.pdf}
\caption{Results from beam measurements showing the RF-output voltage per acceleration
gap compared to the results from the expected result from the circuit model.}
\label{fig:result}
\end{figure}
\section{Outlook}
Using the circuit model we estimate that with the present RF source we will be able to drive
nine RF-units in parallel by placing the RF source inside the vacuum system nearby the
acceleration gaps. Simulations show that to achieve this one has to take care to minimize
the capacitance in the cables connecting the RF units to the RF source. This is due to the
fact that the capacitance of the cables for nine RF units will be a considerable part of
the load that we need to drive. To minimize phase shift and cable length, we assume that
three commercial coax cables will be used in parallel from the RF source
which will then split into individual cables to each board. Our simulation was
based on coax cables with an impedance of \SI{125}{\ohm} and a low capacitance.
Furthermore, several possible optimization opportunities in our current design of the
RF units became clear. First of all, we should be able to further reduce the capacitance
of the acceleration gaps by reducing the amount of copper used and secondly higher-rated high-voltage connectors should be used to withstand the higher anticipated voltages per gap.
Custom RF cables could further lower the total load capacitance.
\section{conclusion}
Voltages of \SI{2.6}{\kilo\volt} per acceleration gap were achieved using a new, compact, near-board
RF driver. We believe that optimizing the RF driver specifically for this application will allow us
to double or triple the achieved voltage in the near future, and RF acceleration voltages of about
\SI{10}{\kV}/gap might be achievable in the future.
Furthermore, a model was developed and compared to measurements that allowed us to model scaling
behaviour of the system to higher numbers of RF units. The simulations indicated that up to nine
RF units using the current setup can be driven by a single RF source in parallel. By further
optimizing the system we believe that this number can also be increased.
The demonstration of RF acceleration voltages of several kV/gap and the near-board RF driver is an
important step in the development of a compact and low cost multi-beam accelerator technology
that can be scaled to very high beam currents. We estimate that we can accelerate ions in an
array of at least $15\times15$ beams with a 4" wafer platform for a total current \SI{\sim 5}{\mA}
(from an ion source with current density of \SI{10}{\milli\ampere\per\square\cm}). With
optimization of the RF drivers, acceleration gradients of over about \SI{1}{\mega\volt\per\meter}
can be implemented, resulting in a compact, high power accelerator for applications in
materials processing.
\section{acknowledgements}
We thank Michael Current for his encouragement and advice on the evolving field of high-energy ion implantation.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\IEEEPARstart{N}{on}-Bayesian parameter estimation is usually based on the assumption that an observation vector (defined by $\underline{x}$) is available, and that the Probability Density Function (PDF) $f_{\underline{X}}(\underline{x};\underline{\theta})$ is known and is dependent on the parameter vector $\underline{\theta}$, to be estimated \cite{Key}. When dealing with the case where the observations are identically and independent distributed (iid), knowing the PDF $f_{X_i}(x_i;\underline{\theta})$ of each observation $x_i$ is sufficient. However, there are certain situations in which the full observation vector is unavailable or unobservable. Specifically, in this paper, we focus on situations where from a set of observations, only the extreme values, i.e., the minimum and/or the maximum observations are reported.
We deal with a sequence of $N\cdot{K}$ iid observations. Define a sub-sequence of $K$ iid observations which constitute the $i^{th}$ group (i.e., interval), by the vector $\underline{x}_i$, and the minimal and the maximal observed values in the $i^{th}$ interval by $y_{min_i}\equiv{min(\underline{x}_i)}$ and $y_{max_i}\equiv{max(\underline{x}_i)}$, respectively. Thus, the minimum-observation vector of $N$ non-overlapping intervals is defined by $\underline{y}_{min}=[y_{{min}_1},y_{{min}_2},\cdots,y_{{min}_N}]^T$ and the maximum-observation vector of the same $N$ intervals is defined by $\underline{y}_{max}=[y_{{mx}_1},y_{{max}_2},\cdots,y_{{max}_N}]^T$. The complete (unobservable) data is the $N\cdot{K}$-dimension observation vector which can be formulated as: $\underline{x}=[\underline{x}^T_1,\underline{x}^T_2,\cdots,\underline{x}^T_N]^T$.
Such scenarios, where only $\underline{y}_{min}$ and/or $\underline{y}_{max}$ are given rather than $\underline{x}$, can be found in numerous fields, such as earthquake recurrence estimation \cite{wang2011}, wind research \cite{walshaw1994getting}, heat accumulation \cite{MaxTemp1}, and precipitation monitoring \cite{ExtremeRainEx1}, to name a few. Furthermore, there are practical reasons which may force compressing the entire dataset to the $\underline{y}_{min}$ and/or $\underline{y}_{max}$ observations, such as energy constraints in sensor networks \cite{EnvSN}. E.g., most Network Management Systems (NMS) which monitor the backhaul of the cellular networks, although sample the network microwave links channels Received Signal Levels (RSL) at a high frequency (of up to 10Hz), usually report only the minimum and the maximum observed values per 15-minute intervals \cite{RemkoCountry2,Eric10S,yonidiss}. Furthermore, reporting only the extreme values from a given set of observations is especially attractive in the emerging field of the Internet-of-Things (IoT), since, apart from transmission costs, it has been shown that the extraction of the minimum and the maximum values from a sequence can be done extremely efficient with respect to processing power, delay, energy, and memory requirements \cite{lemire2006streaming}. Indeed, there are scenarios where it could be more beneficial to treat other types of information rather than the minimum and/or the maximum (e.g., when dealing with the Normal distribution, calculating the \emph{mean} and the \emph{standard deviation} from the original observations may be preferable, since they are sufficient statistics). However, one still needs to consider the fact that under extreme energy and cost constraints, the additional hardware needed for these kinds of calculations may not be available. The minimum and the maximum values, on the other hand, as they are the extremes of the original observation vector, can be identified using very basic circuits, without the need for extra processing power \cite{sedraandsmith}.
Thus, the motivation to understand the estimation accuracy that can be achieved by using only extreme measurements (in comparison with the full original sample set) is strong, especially as giving access only to extreme values may result in much more efficient designs of future applications, if the achievable estimation accuracy is sufficient.
In this paper we show that it is possible to evaluate the accuracy of the estimate of a parameter vector $\underline{\theta}$ given only the minimum, the maximum, or the minimum \emph{and} the maximum observation vectors ($\underline{y}_{min}$, $\underline{y}_{max}$, and \{$\underline{y}_{min}$,$\underline{y}_{max}$\}, respectively). We describe the corresponding Fisher Information Matrices (FIM) of the (asymptotically) optimal estimators in the Minimum Mean Squared Error sense (MMSE), using a novel approximation which simplify the resulted expressions so that they become analytically solvable, and, by means of the approximated Cramer-Rao Lower Bound (CRLB), we present an analysis of the achievable estimation performance.
\subsection{Related Studies}
Many previous studies have dealt with the problem of parameter estimation involving $\underline{y}_{min}$ or $\underline{y}_{max}$. The subject of most of these studies, however, was to characterise the extreme values themselves. I.e., these studies aimed to estimate the properties of extreme events (such as catastrophic earthquakes probabilities \cite{wang2011}, variations in climate \cite{ClimateExtreme92}, extreme floods \cite{Flood2007,Katz3}, long precipitation events \cite{ExtremeRainEx1}, daily rainfall \cite{coles2003anticipating}, etc.), usually by taking advantage of the Extreme Value Theory (EVT). The EVT had started to attract interest in the last century, and was formalized by \emph{Gumbel} in 1958 \cite{Gumbel}. In its base, the EVT states that under some regularity conditions, the PDF of maximum (or minimum, sometimes under certain transformations) values converges asymptotically to the Generalized Extreme Value (GEV) PDF, $f_{\underline{Y}}(\underline{y};\underline{\psi})$ \cite{FisherT,walshaw2013}. Thus, it is possible to estimate the parameter vector $\underline{\psi}$ \cite{GEVEstimator} (or other properties such as moments \cite{hosking1985,Dekkers}), directly from $\underline{y}_{min}$,$\underline{y}_{max}$.
In recent years, the estimation of the parameter vector of the original PDF, $\underline\theta$, from extreme values has been partially covered by the \emph{Record Theory}, which deals with ordered data \cite{records}. And indeed, numerous studies discussed the estimation of $\underline\theta$ from maximum or minimum values \cite{recordsparam1986}, \cite{records}, as well as the FIM properties of those estimates \cite{ahmadi2001fisher,hofmann2003,recordsreview2009}. However, as described in those studies, the presented FIM expressions, apart for some specific examples such as for the exponential distribution case \cite{hofmann2003}, are implicit and may not hold a simple solution \cite{recordsreview2009}. Recently, we have shown that an estimate $\hat{\underline{\theta}}$ of the original parameter vector $\underline{\theta}$ can be evaluated directly from the GEV estimated parameter vector $\hat{\underline{\psi}}$ \cite{SAM2014}. This method, however, relies on the asymptotic convergence of the PDF of the maximum values to the GEV, and thus, cannot guarantee optimal performance in non-asymptotic conditions. Several other studies presented approaches for the estimation of $\underline{\theta}$ from incomplete observations-set, such as the Expectation-Maximization (EM) algorithm \cite{EM}, for which $\underline{x}$ is the complete information and $\underline{y}_{min}$,$\underline{y}_{max}$ are the incomplete information. However, the estimator proposed in these studies basically extant approximation of the full observation vector $\hat{\underline{x}}$ from the set of the incomplete data (i.e., the extremes $\underline{y}_{min}$,$\underline{y}_{max}$), and estimate $\underline{\theta}$ from $\hat{\underline{x}}$. The EM, as other methods which first recover $\hat{\underline{x}}$ (or its properties) and then estimate $\underline{\theta}$ assuming $\hat{\underline{x}}$ is the observation vector \cite{MaxTemp1,LibermanTh}, are usually case specific, often require pre-calibration stages \cite{RemkoCountry,Yoni1}, and are sub-optimal due to the approximation of $\underline{x}$.
Different from these (among many) past studies, in this paper we are interested in evaluating the achievable performance of the estimation of the parameter vector $\underline{\theta}$ of the \emph{original} PDF, directly from the available extreme measurements. We present a tool which approximates the cumbersome expressions of the relevant FIM into a simple, practical, and solvable form, without a significant loss of accuracy. Furthermore, we extend previous results and discuss the case where both the minimum \emph{and} the maximum values are being used in the same estimation process.
\subsection{Summary of the Results}
We consider the (asymptotically) optimal estimates of the parameter vector ,$\underline\theta$ based on the set of the extremes $\underline{y}_{min}$,$\underline{y}_{max}$, and we present a novel and simple performance analysis tools. We consider three cases:
\begin{itemize}
\item[1.]{Only the minimum-observation vector, $\underline{y}_{min}$, is available.}
\item[2.]{Only the maximum-observation vector, $\underline{y}_{max}$, is available.}
\item[3.]{Both the minimum-observation vector, $\underline{y}_{min}$, \emph{and} the maximum-observation vector, $\underline{y}_{max}$, are available.}
\end{itemize}
For each of the three cases we derive the appropriate FIM, and present simplified approximated expressions, to be used in the comparative performance analysis. The optimal estimator, $\hat{\underline{\theta}}_{opt}$, makes use of the entire range of $K\cdot{N}$ samples, once the observation vector $\underline{x}$ is available. Using only a part of the available samples results in a sub-optimal estimation, and thus, decreases the estimation accuracy \cite{Key}. While it is obvious that any estimation based on $\underline{y}_{min}$ and/or $\underline{y}_{max}$ is suboptimal, our aim is to quantify the relative performance loss of the estimators based on $\underline{y}_{min}$ and/or $\underline{y}_{max}$, relative to the performance of $\hat{\underline{\theta}}_{opt}$.
We show that whereas using both the minimum and the maximum measurements in the estimation is preferred (regarding the estimates accuracy), for certain distributions, most of the information about $\underline{\theta}$ is stored in one of the extremes (i.e., only the minimum or only the maximum values contain most of the information needed for the estimation). And, by using the tools we present, it is possible to pre-determine which of the extremes is preferable, and how much information it contains. We validate our findings by a series of simulations, compare the results with previously published studies for the exponential case, where analytical solutions for the FIM expressions exist \cite{records,hofmann2003}, and show that our proposed approximations are valid.
Thus, the main contributions of this paper are:
\begin{enumerate}
\item{Presentation of a novel approximation for the FIM of estimates based on extreme measurements (minimum, maximum, \emph{and} minimum and maximum combined) by using the Characteristic Values of the Extremes, in order to express an approximate CRLB analytically;
\item{Analysis of extreme measurements in terms of the information they hold (with respect to parameter estimation) compared with the original set of samples.}
\end{enumerate}
\subsection{Organization}
The rest of the paper is organized as follows: In \emph{Section II} the methodology, tools, and our analysis are presented. \emph{Section III} presents an explicit development of our tools, followed by a simulation which is used in order to validate the results. A comparison with previously published results is also included and discussed in this section. Lastly, in \emph{Section V} we conclude this paper.
\section{Methodology and Tools}
The asymptotically \emph{optimal} (in the MMSE sense) estimator $\hat{\underline{\theta}}_{opt}$ is given by:
\begin{IEEEeqnarray}{cCl}
\hat{\underline{\theta}}_{opt}=\arg\max\limits_{\underline{\theta}}\left\{\sum_{i=1}^{N\cdot{}K}f_{{X}}({x_i};\underline{\theta})\right\} \IEEEyesnumber \label{MLEOpt}
\IEEEyesnumber
\end{IEEEeqnarray}
where ${x}_i$ are the original iid observations (i.e., measurements), and $\underline{\theta}$ is the parameter vector, to be estimated. And, under mild regularity conditions, $\hat{\underline{\theta}}_{opt}$ achieves the corresponding Cramer-Rao Lower Bound (CRLB) \cite{Key}, defined by $\text{CRLB}_{opt}$:
\begin{IEEEeqnarray}{cCl} \label{CRLBOpt}
\text{CRLB}_{opt}=J^{opt}(\underline{\theta})^{-1} \IEEEyesnumber \IEEEyessubnumber \\
{J_{m,n}^{opt}(\underline{\theta})}=-\sum_{i=1}^{N\cdot{}K}{E\left\{\frac{\partial^2}{\partial{\theta}_m\partial{\theta}_n}\log\left[f_{{X}}({x_i};\underline{\theta})\right]\right\}} \IEEEyessubnumber
\end{IEEEeqnarray}
where $J^{opt}(\underline{\theta})$ is the FIM consisting of the entries $J_{m,n}^{opt}(\underline{\theta})$.
In addition to the optimal estimator, we define a sub-optimal estimator, $\hat{\underline{\theta}}_L$, based on partial samples $L\in{\mathbb{N}}$, s.t. $1\leq{L}\leq{K}$ observations from each of the $N$ non overlapping sub-intervals:
\begin{IEEEeqnarray}{cCl}
\hat{\underline{\theta}}_{L}=\arg\max\limits_{\underline{\theta}}\left\{\sum_{i=1}^{N}\sum_{j=1}^{{L}}f_{{X}}({x_{i,j}};\underline{\theta})\right\} \IEEEyesnumber \label{MLEl} \IEEEyesnumber
\end{IEEEeqnarray}
which achieves
\begin{IEEEeqnarray}{cCl}
\text{CRLB}_{L}=J^{L}(\underline{\theta})^{-1} \IEEEyesnumber \label{CRLBl} \IEEEyessubnumber \label{CRLBla} \\
{J_{m,n}^{L}(\underline{\theta})}=-\sum_{i=1}^{N}\sum_{j=1}^{{L}}{E\left\{\frac{\partial^2\log\left[f_{{X}}({x}_{i,j};\underline{\theta})\right]}{\partial{\theta}_m\partial{\theta}_n}\right\}} \IEEEyessubnumber \label{CRLBlb}
\end{IEEEeqnarray}
where ${x}_{i,j}$ represents the partial measurements, and $J^{L}(\underline{\theta})$ is the FIM (of the entries $J_{m,n}^{L}(\underline{\theta})$). This sub-optimal estimator $\hat{\underline{\theta}}_L$ and its corresponding CRLB ($\text{CRLB}_L$) will be used as a performance benchmark tool in the sequel.
\subsection{$\underline{y}_{min}$ and/or $\underline{y}_{max}$ based Estimation}
The PDF of an extreme value taken from the $K$ measurements which constitute the $i^{th}$ interval, $y_{min_i}$ and $y_{max_i}$, and the joint PDF of $y_{min_i}$ and $y_{max_i}$ are given by \cite{Gumbel}:
\begin{IEEEeqnarray}{cCl} \label{fgen}
f_{Y_{min}}(y_{min};\underline{\theta})= \IEEEyesnumber \IEEEnonumber
K\left[1-F_{X}(y_{min};\underline{\theta})\right]^{K-1}f_{X}(y_{min};\underline{\theta}) \IEEEyessubnumber \label{fmin} \\
f_{Y_{max}}(y_{max};\underline{\theta})=K\left[F_{X}(y_{max};\underline{\theta})\right]^{K-1}f_{X}(y_{max};\underline{\theta}) \IEEEyessubnumber \label{fmax} \\
f_{Y_{min},Y_{max}}(y_{min},y_{max};\underline{\theta})=
K(K-1)\left[F_{X}(y_{max};\underline{\theta})-F_{X}(y_{min};\underline{\theta})\right]^{K-2}\cdot{} \IEEEyessubnumber
f_{X}(y_{min};\underline{\theta})f_{X}(y_{max};\underline{\theta})
\end{IEEEeqnarray}
where $F(\cdot)$ represents a Cumulative Density Function (CDF).
Due to the iid properties of $\underline{x}$, which guarantee that the extreme values are also iid\footnote{Meaning, that the set of the maximum values are iid, and the set of the minimum values are iid. However, for each sub-interval, the minimum value and the maximum value are dependant for any $K<\infty$ \cite{coles1999dependence}.}, the PDF of $\underline{y}_{min}$, $\underline{y}_{max}$ and the joint PDF of $\underline{y}_{min}$ and $\underline{y}_{max}$ can be easily expressed:
\begin{IEEEeqnarray}{cCl} \label{fgeni}
f_{\underline{Y}_{min}}(\underline{y}_{min};\underline{\theta})= \IEEEyesnumber \IEEEnonumber
{K}^N\prod_{i=1}^{N}\left[1-F_{X}(y_{min_i};\underline{\theta})\right]^{K-1}f_{X}(y_{min_i};\underline{\theta}) \IEEEyessubnumber \\
\label{fmini}
f_{\underline{Y}_{max}}(\underline{y}_{max};\underline{\theta})= \IEEEnonumber
{K}^N\prod_{i=1}^{N}\left[F_{X}(y_{max_i};\underline{\theta})\right]^{K-1}f_{X}(y_{max_i};\underline{\theta}) \IEEEyessubnumber \\
\label{fmaxi}
f_{\underline{Y}_{min},\underline{Y}_{max}}(\underline{y}_{min},\underline{y}_{max};\underline{\theta})= \left[{K(K-1)}\right]^N\cdot{}
\prod_{i=1}^{N}\left[F_{X}(y_{max_i};\underline{\theta})-F_{X}(y_{min_i};\underline{\theta})\right]^{K-2}\cdot{} \IEEEyessubnumber
f_{X}(y_{min_i};\underline{\theta})f_{X}(y_{max_i};\underline{\theta}) \label{fminmaxi}
\end{IEEEeqnarray}
From which, the desired estimators, can be directly formulated. For the sequence, we define
$\hat{\underline{\theta}}_{min}$, $\hat{\underline{\theta}}_{max}$, and $\hat{\underline{\theta}}_{mix}$ as the estimates based on $\underline{y}_{min}$, $\underline{y}_{max}$, and \{$\underline{y}_{min}$,$\underline{y}_{max}$\}, respectively.
The corresponding FIM of these estimates is given by:
\begin{IEEEeqnarray}{cCl}
{J^{min}_{m,n}(\underline{\theta})}=-{E\left\{\frac{\partial^2}{\partial{\theta}_m\partial{\theta}_n}\log\left[f_{\underline{Y}_{min}}(\underline{y}_{min};\underline{\theta})\right]\right\}}
=-\sum_{i=1}^{N}{E\left\{\frac{\partial^2}{\partial{\theta}_m\partial{\theta}_n}\log\left[f_{{Y}_{min_i}}({y_{min_i}};\underline{\theta})\right]\right\}} \IEEEyesnumber \label{FIMmixt}\IEEEyessubnumber \label{CRLBmi} \\
{J^{max}_{m,n}(\underline{\theta})}=-{E\left\{\frac{\partial^2}{\partial{\theta}_m\partial{\theta}_n}\log\left[f_{\underline{Y}_{max}}(\underline{y}_{max};\underline{\theta})\right]\right\}}
=-\sum_{i=1}^{N}{E\left\{\frac{\partial^2}{\partial{\theta}_m\partial{\theta}_n}\log\left[f_{{Y}_{max_i}}({y}_{max_i};\underline{\theta})\right]\right\}} \IEEEyessubnumber \label{CRLBmx} \\
{J^{mix}_{m,n}(\underline{\theta})}=-{E\left\{\frac{\partial^2\log\left[f_{\underline{Y}_{{min}},\underline{Y}_{{max}}}(\underline{y}_{min},\underline{y}_{max};\underline{\theta})\right]}{\partial{\theta}_m\partial{\theta}_n}\right\}}
=-\sum_{i=1}^{N}{E\left\{\frac{\partial^2\log\left[f_{{Y}_{min_i},Y_{max_i}}({y}_{{min_i}},y_{max_i};\underline{\theta})\right]}{\partial{\theta}_m\partial{\theta}_n}\right\}} \IEEEyessubnumber \label{CRLBmix}
\end{IEEEeqnarray}
And, as in \eqref{CRLBOpt}, under mild regularity conditions, the asymptotic covariance matrices of the estimates $\hat{\underline{\theta}}_{min}$, $\hat{\underline{\theta}}_{max}$, and $\hat{\underline{\theta}}_{mix}$ achieve the corresponding CRLBs:
\begin{IEEEeqnarray}{cCl} \label{CRLBmixt}
\text{CRLB}_{min}=J^{min}(\underline{\theta})^{-1} \IEEEyesnumber \IEEEyessubnumber \\
\text{CRLB}_{max}=J^{max}(\underline{\theta})^{-1} \IEEEyessubnumber \\
\text{CRLB}_{mix}=J^{mix}(\underline{\theta})^{-1} \IEEEyessubnumber
\end{IEEEeqnarray}
where $\text{CRLB}_{min}$, $\text{CRLB}_{max}$, and $\text{CRLB}_{mix}$, are the corresponding asymptotic covariances of $\hat{\underline{\theta}}_{min}$, $\hat{\underline{\theta}}_{max}$, and $\hat{\underline{\theta}}_{mix}$, respectively.
In order to simplify the presentation of our approach, the parameter vector $\underline\theta$ is reduced to a single parameter (i.e., $\underline\theta\equiv\theta$). The same development stages and conclusions presented in the sequel can be generalized for the multi-parameter case.
\subsection{The Relationship Between $J^{min}(\theta)$ and $J^{max}(\theta)$}
While it is obvious that the following relationship stands for any distribution:
\begin{IEEEeqnarray}{cCl}
\left\{J^{min}(\theta),{J}^{max}(\theta)\right\} \label{MajorG}
\leq{J}^{mix}(\theta)\leq{J}^{opt}(\theta) \IEEEyesnumber
\end{IEEEeqnarray}
the specific order of $J^{min}(\theta)$ and $J^{max}(\theta)$ is distribution dependent.
Next, we introduce the following:
\begin{lemma}
\label{Lemma1}
Let $\underline{z}$ be a ${K}$-dimensional observation vector, the entries of which are iid with PDF $f_Z(z;\underline\theta)$, of \textbf{finite} sample space of $\Omega=[C_1,C_2]$, meaning that the entries of $\underline{z}$ are bounded so that $\forall{i}:C_1\leq{z_i}\leq{C_2}$. Define two random variables, $Y_{min}$ and $Y_{max}$, so that ${Y}_{min}$=$min(\underline{z})$ and ${Y}_{max}$=$max(\underline{z})$. For any given function $g(\cdot)$ that satisfies a regularity condition so that $\forall \alpha: \left|\frac{\partial{^i}{g(\alpha)}}{\partial{\alpha}{^i}}\right|<\infty;i\in{\mathbb{N}}$, the following hold:
\begin{IEEEeqnarray}{cCl}
g(Y_{min})\xrightarrow[K\rightarrow\infty]{{w.p.1}}g(E[Y_{min}]) \IEEEyesnumber \IEEEyessubnumber \label{eqtaylormindeducted} \\
g(Y_{max})\xrightarrow[K\rightarrow\infty]{{w.p.1}}g(E[Y_{max}]) \IEEEyessubnumber \label{eqtaylormaxdeducted}
\end{IEEEeqnarray}
\end{lemma}
\begin{proof}
Eqs. \eqref{fmin} and \eqref{fmax}, in combination with the finite sample space $\Omega$, yield
\begin{IEEEeqnarray}{cCl}
Y_{min}\xrightarrow[K\rightarrow\infty]{{w.p.1}}C_1 \IEEEyesnumber \IEEEyessubnumber \label{eqyminc1} \\
Y_{max}\xrightarrow[K\rightarrow\infty]{{w.p.1}}C_2 \IEEEyessubnumber \label{eqymaxc2}
\end{IEEEeqnarray}
The Taylor expansions of $g(Y_{min})$ about $E[Y_{min}]$ and $g(Y_{max})$ about $E[Y_{max}]$ are
\begin{IEEEeqnarray}{cCl}
g(Y_{min})=g(E[Y_{min}])+ \IEEEyesnumber
\sum_{i=1}^{\infty}\frac{{(Y_{min}-E[Y_{min}])}^i}{i!}\cdot{}\left.\frac{\partial^i{g}(Y_{min})}{\partial{Y^i_{min}}}\right|_{E[Y_{min}]} \IEEEyessubnumber \label{eqtaylormin} \\
g(Y_{max})=g(E[Y_{max}])+
\sum_{i=1}^{\infty}\frac{{(Y_{max}-E[Y_{max}])}^i}{i!}\cdot{}\left.\frac{\partial^i{g}(Y_{max})}{\partial{Y^i_{max}}}\right|_{E[Y_{max}]} \IEEEyessubnumber \label{eqtaylormax}
\end{IEEEeqnarray}
Applying the \emph{Expected Value} operator on Eqs. \eqref{eqyminc1} and \eqref{eqymaxc2} yields
\begin{IEEEeqnarray}{cCl}
E[Y_{min}]\xrightarrow[K\rightarrow\infty]{{w.p.1}}C_1 \IEEEyesnumber \IEEEyessubnumber \label{eqexpmin} \\
E[Y_{max}]\xrightarrow[K\rightarrow\infty]{{w.p.1}}C_2 \IEEEyessubnumber \label{eqexpmax}
\end{IEEEeqnarray}
which dictates that the sums of Eqs. \eqref{eqtaylormin} and \eqref{eqtaylormax} converge to zero:
\begin{IEEEeqnarray}{cCl}
\sum_{i=1}^{\infty}\frac{{(Y_{min}-E[Y_{min}])}^i}{i!}\cdot{} \IEEEyesnumber
\left.\frac{\partial^i{g}(Y_{min})}{\partial{Y^i_{min}}}\right|_{E[Y_{min}]}\xrightarrow[K\rightarrow\infty]{{w.p.1}}0 \IEEEyesnumber \IEEEyessubnumber \\
\sum_{i=1}^{\infty}\frac{{(Y_{max}-E[Y_{max}])}^i}{i!}\cdot{}
\left.\frac{\partial^i{g}(Y_{max})}{\partial{Y^i_{max}}}\right|_{E[Y_{max}]}\xrightarrow[K\rightarrow\infty]{{w.p.1}}0 \IEEEyessubnumber
\end{IEEEeqnarray}
This completes the proof.
\end{proof}
\begin{lemma}
\label{Lemma2}
Let $\underline{z}$ be a ${K}$-dimensional observation vector, the entries of which are iid with PDF $f_Z(z;\underline\theta)$, of \textbf{infinite} sample space of $\Omega=(-\infty,\infty)$, meaning that the entries of $\underline{z}$ are unbounded so that $\forall{i}:-\infty{}\leq{z_i}\leq{\infty}$. Define two random variables, $Y_{min}$ and $Y_{max}$, so that ${Y}_{min}$=$min(\underline{z})$ and ${Y}_{max}$=$max(\underline{z})$ each converges in distribution to an asymptotic GEV PDF with shape parameter $\epsilon{<}{\frac{1}{2}}$. Then, the following hold:
\begin{IEEEeqnarray}{cCl}
\frac{Y_{min}}{E[Y_{min}]}\xrightarrow[K\rightarrow\infty]{{a.s.}}1 \IEEEyesnumber \IEEEyessubnumber \\ \label{eqlemma2min}
\frac{Y_{max}}{E[Y_{max}]}\xrightarrow[K\rightarrow\infty]{{a.s.}}1 \IEEEyessubnumber \label{eqlemma2max}
\end{IEEEeqnarray}
\end{lemma}
\begin{proof}
Define $err_{min}\equiv{Y_{min}}-E[Y_{min}]$ and $err_{max}\equiv{Y_{max}}-E[Y_{max}]$. Based on the properties of the GEV \cite{Gumbel}, two finite constants exist (defined by $D_1$ and $D_2$), so that
\begin{IEEEeqnarray}{cCl}
\lim_{K \to \infty} \left\{var(err_{min})\right\}=D_1 \IEEEyesnumber \IEEEyessubnumber \label{errvarmin} \\
\lim_{K \to \infty} \left\{var(err_{max})\right\}=D_2 \IEEEyessubnumber \label{errvarmax}
\end{IEEEeqnarray}
which, based on Chebyshev's inequality \cite{papoulis2002probability}, in combination with the fact that
\begin{IEEEeqnarray}{cCl}
E[Y_{min}]\xrightarrow[K\rightarrow\infty]{{}}{-\infty} \IEEEyesnumber \IEEEyessubnumber \label{errmin} \\
E[Y_{max}]\xrightarrow[K\rightarrow\infty]{{}}\infty \IEEEyessubnumber \label{errmax}
\end{IEEEeqnarray}
completes this proof.
\end{proof}
Under the conditions of \emph{Lemma} \ref{Lemma1}, the variances of $Y_{min}$ and $Y_{max}$ converge to zero (as $K$ increases). Thus,
\begin{IEEEeqnarray}{cCl}
E[g(Y_{min})]={g}(E[Y_{min}])+\mathcal{R}^{min} \IEEEyesnumber \IEEEyessubnumber \\
E[g(Y_{max})]={g}(E[Y_{max}])+\mathcal{R}^{max} \IEEEyessubnumber
\end{IEEEeqnarray}
where $\mathcal{R}^{min}$ and $\mathcal{R}^{max}$ are the residues, which converge to zero as $K$ increases.
\emph{Lemma} \ref{Lemma2} is weaker than \emph{Lemma} \ref{Lemma1} in the sense that under the conditions of \emph{Lemma} \ref{Lemma2}, the variances of $Y_{min}$ and $Y_{max}$ converge to known constants (as $K$ increases), and not to zero. Thus, under the conditions of \emph{Lemma} \ref{Lemma2}, the residuals $\mathcal{R}^{min}$ and $\mathcal{R}^{max}$ would not converge to zero. On the other hand, these residues do not depend on $K$, whereas the expected values of the extremes do \cite{Gumbel}. Thus, given a function $g(\cdot)$ that satisfies a regularity condition so that $\forall \alpha: \left|\frac{\partial{^i}{g(\alpha)}}{\partial{\alpha}{^i}}\right|<\infty;i\in{\mathbb{N}}$
\begin{IEEEeqnarray}{cCl}
E[g(Y_{min})]=\underbrace{{g}(E[Y_{min}])}_{\mathcal{O}\left((ln(K)\right)}+\underbrace{\mathcal{R}^{min}}_{\mathcal{O}\left(1\right)} \IEEEyesnumber \IEEEyessubnumber \\
E[g(Y_{max})]=\underbrace{{g}(E[Y_{max}])}_{\mathcal{O}\left((ln(K)\right)}+\underbrace{\mathcal{R}^{max}}_{\mathcal{O}\left(1\right)} \IEEEyessubnumber \label{resmaxappdx1}
\end{IEEEeqnarray}
and for sufficiently large values of $K$, the following conclusions can be expressed:
\begin{IEEEeqnarray}{cCl}
E[g(Y_{min})]={g}(E[Y_{min}])\cdot{}\left(1+\mathcal{O}\left(\frac{1}{\ln{(K)}}\right)\right) \IEEEyesnumber \IEEEyessubnumber \label{ConclusionGminAp} \\
E[g(Y_{max})]={g}(E[Y_{max}])\cdot{}\left(1+\mathcal{O}\left(\frac{1}{\ln{(K)}}\right)\right) \IEEEyessubnumber \label{ConclusionGmaxAp}
\end{IEEEeqnarray}
which results in our main approximation:
\begin{IEEEeqnarray}{cCl}
E[g(Y_{min})]\approx{g}(E[Y_{min}]) \IEEEyesnumber \IEEEyessubnumber \label{ConclusionGminAp} \\
E[g(Y_{max})]\approx{g}(E[Y_{max}]) \IEEEyessubnumber \label{ConclusionGmaxAp}
\end{IEEEeqnarray}
for both finite and infinite sample spaces. Furthermore, the same conclusions hold in cases where the sample space is infinite, but is one-side bounded, e.g., the \emph{Gamma} distribution, the sample space of which is $\Omega=[0,\infty)$.
In addition, these conclusions can be easily expanded for the two-variables case, resulting in, assuming $g(\cdot)$ satisfies the same regularity condition of \emph{Lemma} \ref{Lemma1} and/or \emph{Lemma} \ref{Lemma2} for every variable, and sufficiently large $K$
\begin{equation} \label{ConclusionGmixApp}
E[g(Y_{min},Y_{max})]\approx{g}(E[Y_{min}],E[Y_{max}])
\end{equation}
as for $K\rightarrow\infty$, the minimum and the maximum values taken from the same group become independent \cite{coles1999dependence}.
\subsection{Characteristic Values of Extremes}
In order to specify Eq. \eqref{MajorG} for a given distribution, we suggest to use the Characteristic Largest (or Smallest) Value \cite{Gumbel}.
The Characteristic Largest Value (CLV) and the Characteristic Smallest Value (CSV) were first introduced by \emph{Gumbel} in 1958 \cite{Gumbel}, as averages of extremes that are analogous to quantiles: the CLV is defined as $\mu_K$, so that from $K$ iid observations, exactly one observation is expected to be equal to or larger than $\mu_K$. In a similar manner, the CSV is defined as $\mu_1$, so that from $K$ iid observations, exactly one observation is expected to be equal to or smaller than $\mu_1$. The CSV and the CLV definitions lead to the relations
\begin{IEEEeqnarray}{cCl}
F_X(\mu{_1};\underline{\theta},K)=\frac{1}{K} \IEEEyesnumber \label{CLVG} \IEEEyessubnumber \label{CSVch2} \\
F_X(\mu{_K};\underline{\theta},K)=1-\frac{1}{K} \IEEEyessubnumber \label{CLVch2}
\end{IEEEeqnarray}
This means that, given a CDF of the original observation, $F_X({x;{\underline\theta}})$, which is analytically representable with respect to $K$, $F_X({x;\underline\theta,K})$, the explicit expressions of the CSV, $\mu_1$, and the CLV, $\mu_K$, can be directly extracted.
The characteristic largest and smallest values are closely related to the expected values of the extremes. In cases where the sample space is finite, the CLV and/or the CSV
converge to the boundaries of the sample space as $K$ increases, and thus, converge to the expected value of the extreme. Hence, for any given function $g(\cdot)$ that satisfies the regularity condition $\forall \alpha: \left|\frac{\partial{^i}{g(\alpha)}}{\partial{\alpha}{^i}}\right|<\infty;i\in{\mathbb{N}}$, the following hold:
\begin{IEEEeqnarray}{cCl} \label{ConclusionULem1}
E[g(Y_{min})]\xrightarrow[K\rightarrow\infty]{}{g}(\mu_1) \IEEEyesnumber \IEEEyessubnumber \label{ConclusionUmin} \\
E[g(Y_{max})]\xrightarrow[K\rightarrow\infty]{}{g}(\mu_K) \IEEEyessubnumber \label{ConclusionUmax} \\
E[g(Y_{min},Y_{max})]\xrightarrow[K\rightarrow\infty]{}{g}(\mu_1,\mu_K) \IEEEyessubnumber \label{ConclusionUmix}
\end{IEEEeqnarray}
However, in cases where the sample space is infinite and the conditions of \emph{Lemma} \ref{Lemma2} apply, the CLV and/or the CSV converge to the \emph{mode}, i.e., the most probable value, of the extreme, as $K$ increases \cite{Gumbel}, and, as the variance of the GEV under these conditions is finite and independent of $K$, the following hold:
\begin{IEEEeqnarray}{cCl}
\frac{\mu_1}{E[Y_{min}]}\xrightarrow[K\rightarrow\infty]{}1 \IEEEyesnumber \IEEEyessubnumber \\ \label{eqlemma2minrep}
\frac{\mu_K}{E[Y_{max}]}\xrightarrow[K\rightarrow\infty]{}1 \IEEEyessubnumber \label{eqlemma2maxrep}
\end{IEEEeqnarray}
which, under the assumption $\forall \alpha: \left|\frac{\partial{^i}{g(\alpha)}}{\partial{\alpha}{^i}}\right|<\infty;i\in{\mathbb{N}}$, for sufficiently large $K$, yields
\begin{IEEEeqnarray}{cCl} \label{ConclusionULem2}
E[g(Y_{min})]\approx{g}(\mu_1) \IEEEyesnumber \IEEEyessubnumber \label{ConclusionUmin2} \\
E[g(Y_{max})]\approx{g}(\mu_K) \IEEEyessubnumber \label{ConclusionUmax2} \\
E[g(Y_{min},Y_{max})]\approx{g}(\mu_1,\mu_K) \IEEEyessubnumber \label{ConclusionUmix2}
\end{IEEEeqnarray}
Next, by implementing the set of Eqs. \eqref{ConclusionULem1} and \eqref{ConclusionULem2} directly on the different FIM expressions, the following approximations can be presented.
\begin{IEEEeqnarray}{cCl}
\widetilde{J}^{min}_{m,n}(\underline\theta)=N\cdot{g_{m,n}(\mu_1,\underline\theta)}\approx \IEEEyesnumber
\sum_{i=1}^{N}E\left[g_{m,n}(z_i,\underline\theta)\right]=J^{min}_{m,n}(\underline\theta) \label{Japprx} \IEEEyessubnumber \label{Japprxminch2} \\
\widetilde{J}^{max}_{m,n}(\underline\theta)=N\cdot{g_{m,n}(\mu_K,\underline\theta)}\approx
\sum_{i=1}^{N}E\left[g_{m,n}(z_i,\underline\theta)\right]=J^{max}_{m,n}(\underline\theta) \IEEEyessubnumber \label{Japprxmaxch2} \\
\widetilde{J}^{mix}_{m,n}(\underline\theta)=N\cdot{g_{m,n}(\mu_1,\mu_K,\underline\theta)}\approx\sum_{i=1}^{N}E\left[g_{m,n}(z_i,q_i,\underline\theta)\right]=J^{mix}_{m,n}(\underline\theta) \IEEEyessubnumber \label{Japprxmixch2}
\end{IEEEeqnarray}
which, given that $g(\cdot)$ is the known log-likelihood function, can be analytically expressed and solved.
Next, we define a set of matrices, $\mathcal{A}^i$, as follows.
\begin{IEEEeqnarray}{cCl}
J_{m,n}^{max}(\underline\theta)-J_{m,n}^{min}(\underline\theta)=\sum^{N}_{i=1}\mathcal{A}^i_{m,n} \IEEEyesnumber \label{Adef}
\end{IEEEeqnarray}
where
\begin{IEEEeqnarray}{cCl}
\mathcal{A}^i_{m,n}= E\left\{\frac{\partial^2\log\left[f_{Y_{min_i}}(z_{i};{\theta})\right]}{\partial{\theta{_m}\theta{_n}}}\right.
\left.-\frac{\partial^2\log\left[f_{Y_{max_i}}(q_{i};{\theta})\right]}{\partial{\theta{_m}\theta{_n}}}\right\} \IEEEyesnumber
\end{IEEEeqnarray}
which, based on the approximations $\widetilde{J}^{min}(\underline\theta)$, $\widetilde{J}^{max}(\underline\theta)$, and $\widetilde{J}^{mix}(\underline\theta)$ (of Eq. \eqref{Japprx}), yields
{{
\begin{IEEEeqnarray}{cCl}
\mathcal{A}^i_{m,n}\underset{\underset{\forall{i}}{\uparrow}}{\approx}\mathcal{A}_{m,n}=\left.\frac{\partial^2\log\left[f_{Y_{min}}(z_i;{\theta})\right]}{\partial{\theta{_m}}\partial\theta{_n}}\right|_{z_i=\mu_1} - \IEEEyesnumber
\left.\frac{\partial^2\log\left[f_{Y_{max}}(q_i;{\theta})\right]}{\partial{\theta{_m}}\partial\theta{_n}}\right|_{q_i=\mu_K} \label{Agen}
\end{IEEEeqnarray}
which equals
\begin{IEEEeqnarray}{cCl}
\mathcal{A}_{m,n}=(K-1)\cdot\left[\frac{\partial{F_m}}{\partial\theta_n}\cdot\frac{\partial{F_m}}{\partial\theta_m}\cdot\frac{1}{(1-F_m)^2} \right. \label{AAA}
\left.-\frac{\partial^2{F_m}}{\partial\theta{_n}\partial\theta{_m}}\cdot\frac{1}{1-F_m}+\frac{\partial{F_M}}{\partial\theta_n}\cdot\frac{\partial{F_M}}{\partial\theta_m}\cdot\frac{1}{(F_M)^2}-\frac{\partial^2F_M}{\partial\theta_{n}\partial\theta_m}\cdot\frac{1}{F_M}\right]+ \IEEEnonumber \\
+\frac{\partial^2f_m}{\partial\theta_{n}\partial\theta_{m}}\cdot\frac{1}{f_m}-\frac{\partial^2f_M}{\partial\theta_{n}\partial\theta_{m}}\cdot\frac{1}{f_M}
+{\frac{\partial{f_M}}{\partial\theta_n}\cdot\frac{\partial{f_M}}{\partial\theta_m}\cdot\frac{1}{(f_M)^2}}-{\frac{\partial{f_m}}{\partial\theta_n}\cdot\frac{\partial{f_m}}{\partial\theta_m}\cdot\frac{1}{(f_m)^2}} \IEEEyesnumber
\end{IEEEeqnarray}}}
where $F_m\equiv{F}_{X}(\mu_1;\theta)$, $F_M\equiv{F}_{X}(\mu_K;\theta)$, $f_m\equiv{f}_{X}(\mu_1;\theta)$, and $f_M\equiv{f}_{X}(\mu_K;\theta)$ (of Eq. \eqref{fgen}), which are known. Thus, the value of $\mathcal{A}$ can be directly calculated.
The value of $\mathcal{A}$ holds important information:
\begin{equation} \label{CasesA}
\begin{cases}
\mathcal{A} > 0, {\text{~~}} \Rightarrow {\text{~~}} \widetilde{J}^{max}(\underline\theta)>\widetilde{J}^{min}(\underline\theta) \\
\mathcal{A} < 0, {\text{~~}} \Rightarrow {\text{~~}} \widetilde{J}^{max}(\underline\theta)<\widetilde{J}^{min}(\underline\theta) \\
\mathcal{A} \approx 0, {\text{~~}} \Rightarrow {\text{~~}} \widetilde{J}^{min}(\underline\theta)\approx\widetilde{J}^{max}(\underline\theta) \\
\end{cases}
\end{equation}
Based on this analysis, the questioned inequalities of Eq. \eqref{MajorG} can be established for any relevant distribution of interest. Some specific cases are worth noting:
\begin{itemize}
\item{{Symmetric Distributions}: In the special case where the original observations, $\{x_i\}$ ; $\forall{i}$, follow a \emph{symmetric PDF}, in the sense that a value of $y$ exists, s.t. $f_X(y+\delta)=f_X(y-\delta)$ ; $\forall{\delta}\in{\mathcal{R}}$, either single-parameter (e.g., the \emph{uniform} distribution) or multi-parameter (e.g., the \emph{normal} distribution), the calculation of $\mathcal{A}$ yields $\mathcal{A}=0$. This is to be expected, since it can be proven\footnote{By substituting the variables s.t. $x_i\equiv-y_i$ ; $\forall{i}$, it can be shown that the general expressions of $J^{min}(\underline{\theta})$ and $J^{max}(\underline{\theta})$ are identical.} that for the symmetric case, $J^{min}(\underline{\theta})=J^{max}(\underline{\theta})$.}
\item{Non-Negative Distributions: Many real-world physical and socio-economical scenarios are modelled using PDFs of the form: $f_X(x<C)=0; C\geq{0}$. Examples of such cases are measurements of rain intensity, earthquakes magnitudes, wind speeds, yearly income, and sun-solar flare intensity, among many other. In those cases, by implementing the methodology presented in \emph{Lemma 1} and \emph{Lemma 2}, it can be shown that for large value of $K$, (i.e., $K\rightarrow\infty$), $\mathcal{A}>0$. Thus, for many practical uses, given that $K$ is sufficiently large, one can consider the maxima to hold most of the information (with respect to parameter estimation). Furthermore, in cases where the distribution is non-positive, $f_X(x>C)=0; C\leq{0}$, such as when modeling the volume of water pending evaporation, the speed of free electrons in metals, etc., similar conclusions apply, where for sufficiently large $K$, $\mathcal{A}<0$.}
\end{itemize}
In order to demonstrate and analyze the proposed methodology, the original observations distribution will be assumed to be of the \emph{exponential} type. The exponential distribution was chosen due to various reasons: First, the exponential distribution is considered to be a good and solid model for various phenomena in many fields of interest (from rain-rate intensity \cite{salisu2010modeling} to income in the USA \cite{ExpEcon}). Second,
the PDF of the maximum value taken from groups of $K$ samples of an iid distributed population that follows a distribution belonging to the exponential family, e.g., \emph{exponential}, \emph{Gamma}, $\chi^2$, or \emph{Normal}, will converge asymptotically to the \emph{Gumbel} distribution \cite{Gumbel}, as $K\rightarrow\infty$. Thus, the behavior of $\hat{\theta}_{max}$ and $\widetilde{J}^{max}(\theta)$ regarding the \emph{exponential} distribution may be used on other distributions of the \emph{exponential} family, once $K$ is sufficiently large.
\section{Example: The Case of Exponential Distribution}
$z$ is said to follow the exponential distribution, with the parameter ${\theta}$, if $f_Z(z;\theta)=\frac{1}{\theta}e^{-\frac{1}{\theta}\cdot{z}}$, or, equivalently, $F_Z(z;\theta)=1-e^{-\frac{1}{\theta}\cdot{z}}$, so that $Pr\{Z\leq{z}\}=F_Z(z;\theta)$). The asymptotically optimal estimator $\hat{\theta}_{opt}$ can be implicitly expressed:
\begin{equation} \label{MLoptexp}
{\hat{\theta}}_{opt} = \frac{1}{N\cdot{K}}\sum\limits_{i=1}^{N\cdot{K}}{x_{i}}\IEEEyessubnumber
\end{equation}
where $x_i$ is the $i^{th}$ entry of the observation vector $\underline{x}$.
Similarly, the sub-optimal estimator $\hat{{\theta}}_L$ (of Eq. \eqref{MLEl}) can be expressed:
\begin{equation}
{\hat{\theta}}_{L} = \frac{1}{N\cdot{L}}\sum\limits_{p=1}^{N}\sum\limits_{l=1}^{L}{x_{p,l}} \label{lexp}
\end{equation}
where $x_{p,l}$ represents the $l^{th}$ sample within the $p^{th}$ group. Note that if $L=K$, then $\hat{\theta}_L\equiv\hat{\theta}_{opt}$, as expected.
Next, the extreme-based estimates for the exponential case can be directly formalized:
\begin{IEEEeqnarray}{cCl} \label{extremeexpML}
\hat{{\theta}}_{min}=\arg\max\limits_{{\theta}}\left[\left(\frac{K}{\theta}\right)^N{e^{-\frac{K}{\theta}\cdot\sum\limits_{i=1}^{N}{y_{min_i}}}}\right]= \IEEEnonumber
\frac{\sum\limits_{i=1}^{N}y_{min_i}}{N} \IEEEyesnumber \label{minexpml} \IEEEyessubnumber \label{MLEminch3} \\
\hat{{\theta}}_{max}=\arg\max\limits_{{\theta}}\left[\left(\frac{K}{\theta}\right)^N{e^{-\sum\limits_{i=1}^{N}\frac{y_{max_i}}{\theta}}}\right. \left.\cdot\right.
\left.\prod_{i=1}^{N}\left(1-e^{-\frac{{y_{max_i}}}{\theta}}\right)^{K-1}\right] \IEEEyessubnumber \label{MLEmaxch3} \\ \label{maxexpml}
\hat{{\theta}}_{mix}=\arg\max\limits_{{\theta}}\left[\left(\frac{K(K-1)}{\theta^2}\right)^N{e^{\frac{-\sum\limits_{i=1}^{N}\left(y_{min_i}+y_{max_i}\right)}{\theta}}}\right. \left.\right.
\left.\cdot\prod_{i=1}^{N}\left(e^{-\frac{{y_{min_i}}}{\theta}}-e^{-\frac{{y_{max_i}}}{\theta}}\right)^{K-2}\right] \IEEEyessubnumber \label{MLEmixch3} \label{mixexpml}
\end{IEEEeqnarray}
where $y_{min_i}$ and $y_{max_i}$ are the minimum and maximum values of the $i^{th}$ ($1\leq{i}\leq{N}$) interval, respectively. It is noteworthy here that $\hat{\theta}_{min}$, which has a closed analytical expression, is equal to $\hat{\theta}_{L=1}$. This property of $\hat{\theta}_{min}$ is explained via the exponential minima properties, the distribution of which remains \emph{exponential}, with the parameter $\theta/K$ \cite{Gumbel}.
The corresponding FIMs of the estimators $\hat{\theta}_{opt}$, $\hat{\theta}_L$, and $\hat{\theta}_{min}$, $\hat{\theta}_{max}$, and $\hat{\theta}_{mix}$ (of Eqs. \eqref{MLoptexp},\eqref{lexp}, and \eqref{extremeexpML}), can now be formalized:
{\small{
\begin{IEEEeqnarray}{cCl}
{J^{opt}({\theta})}= \frac{N\cdot{K}}{\theta^2} \IEEEyesnumber \label{FIMexp} \IEEEyessubnumber \\ \label{FIMexpopt}
{J^{L}({\theta})}= \frac{N\cdot{L}}{\theta^2} \IEEEyessubnumber \label{FIMexpl} \\
{J^{min}({\theta})}= \frac{N}{\theta^2} \IEEEyessubnumber \label{FIMexpmin} \\
{J^{max}({\theta})}= -\frac{N}{\theta^2} + \frac{2}{\theta^3}\sum\limits_{i=1}^NE[y_{max_i}]+\frac{(K-1)}{\theta^3} \IEEEyessubnumber \label{FIMexpmax}
\cdot\sum\limits_{i=1}^NE\left[\frac{\left(\frac{y_{max_i}^2}{\theta}-2y_{max_i}\right)\cdot{}e^{\frac{-y_{max_i}}{\theta}}}{\left(1-e^{\frac{-y_{max_i}}{\theta}}\right)}\right. \left.+\frac{\frac{y_{max_i}^2}{\theta}e^{\frac{-2y_{max_i}}{\theta}}}{\left(1-e^{\frac{-y_{max_i}}{\theta}}\right)^2}\right] \\
{J^{mix}({\theta})}= -\frac{2N}{\theta^2} + \frac{2}{\theta^3}\sum\limits_{i=1}^NE[y_{min_i}+y_{max_i}]+ \IEEEyessubnumber \label{FIMexpminmax}
+\frac{(K-2)}{\theta^3}\sum\limits_{i=1}^NE\left[\frac{\left({y_{min_i}}e^{\frac{-y_{min_i}}{\theta}}-{y_{max_i}}e^{\frac{-y_{max_i}}{\theta}}\right)^2}{\theta\cdot\left(e^{\frac{-y_{min_i}}{\theta}}-e^{\frac{-y_{max_i}}{\theta}}\right)^2}\right. + \IEEEnonumber \\
\left. +\frac{\left({2y_{min_i}}-\frac{y_{min_i}^2}{\theta}\right)\cdot{e}^{\frac{-y_{min_i}}{\theta}}}{e^{\frac{-y_{min_i}}{\theta}}-e^{\frac{-y_{max_i}}{\theta}}}\right.
-\left.\frac{\left({2y_{max_i}}-\frac{y_{max_i}^2}{\theta}\right)\cdot{e}^{\frac{-y_{max_i}}{\theta}}}{e^{\frac{-y_{min_i}}{\theta}}-e^{\frac{-y_{max_i}}{\theta}}}\right]
\end{IEEEeqnarray}}}
and, while $J^{opt}(\theta)$, $J^{L}(\theta)$, and $J^{min}(\theta)$ are explicitly presented, in order to simplify the expressions of $J^{max}(\theta)$ and $J^{mix}(\theta)$, we use the CLV (of Eq. \eqref{CLVch2}) and the CSV (of Eq. \eqref{CSVch2}) approximations. For the exponential distribution with the parameter $\theta$, the CSV, $\mu^{exp}_1$, and the CLV, $\mu^{exp}_K$, take the forms (for $K>1$)
\begin{IEEEeqnarray}{cCl} \label{CLV}
\mu^{exp}_{1}= \theta\cdot{ln\left({\frac{K}{K-1}}\right)} \IEEEyesnumber \IEEEyessubnumber \\
\mu^{exp}_{K}= \theta\cdot{ln({K})} \IEEEyessubnumber
\end{IEEEeqnarray}
and therefore, substituting $\mu^{exp}_1$ and $\mu^{exp}_K$ into Eqs. \eqref{FIMexpmax} and \eqref{FIMexpminmax} yields (for $K>2$)
\begin{IEEEeqnarray}{cCl}
\widetilde{J}^{max}(\theta)=\frac{N}{\theta^2}\cdot{}\left[\frac{K\cdot{ln}^2(K)}{K-1}-1\right] \IEEEyesnumber \label{FIMmaxapp} \IEEEyessubnumber \label{FIMapprxmaxch3} \\
\widetilde{J}^{mix}(\theta)=\frac{2N}{\theta^2}\cdot{}\left[\frac{(K-1)\cdot\left({ln}(K)-{ln}\left(\frac{K}{K-1}\right)\right)^2}{2(K-2)}
+K\cdot{ln}\left(\frac{K}{K-1}\right)-1\right] \IEEEyessubnumber \label{FIMapprxmixch3}
\end{IEEEeqnarray}
It is worth noting that, although unnecessary, a comparison of the exact (analytical solvable) expression of $J^{min}(\theta)$, which is available for the exponential case, and the approximated expression, $\widetilde{J}^{min}(\theta)$ (via Eq. \eqref{Japprxminch2}), demonstrates the accuracy of the approximation. The approximation accuracy improves as $K$ increases, since $\widetilde{J}^{min}(\theta)\xrightarrow{K\rightarrow\infty}J^{min}(\theta)$). However, even for $K$ as small as $K=10$, the difference between $\widetilde{J}^{min}(\theta)$ and ${J}^{min}(\theta)$ is only $\approx{5\%}$. For $K=100$, the difference already drops to $\approx{0.5\%}$.
Finally, by combining Eq. \eqref{MajorG} with Eqs. \eqref{FIMexp} and \eqref{FIMmaxapp}, the main relationship can be presented for the exponential case:
\begin{IEEEeqnarray}{cCl}
J^{L=1}(\theta){=}J^{min}(\theta)<\widetilde{J}^{max}(\theta)
<\widetilde{J}^{mix}(\theta)<J^{opt}(\theta){=}J^{{L=K}}(\theta) \IEEEyesnumber \label{Main}
\end{IEEEeqnarray}
\subsection[Relative Contribution of the Min/Max Values for Estimation \\ Purposes]{The Relative Contribution of the Min/Max for Estimation \\ Purposes} \label{ContSimch2}
The relationship between the different FIMs (of Eq. \eqref{Main}) provides powerful insights. First, as can be seen in Eqs. \eqref{FIMexpl} and \eqref{FIMexpmin}, the minimum values-based estimator, $\hat{\theta}_{min}$, is identical to the estimator $\hat\theta_{L=1}$. This fact, which may be counter-intuitive, shows that fetching the minimum values themselves does not give an advantage over a random selection of a single value in the group. In other words, knowing the observed minimum value of each group for a given number of groups,
each of which constitutes $K$ iid exponential distributed observations, is identical (with respect to one parameter-estimation performance) to knowing a single randomly selected observation from each group. The situation is different regarding the maxima. Knowing the maximum observed values yields the maximum based estimator, $\hat\theta_{max}$, which achieves a better estimation performance than using only one value from the sequence $\left\{x_i\right\}^K_{i=1}$. Thus, when access to the observations is constrained such that only a single observation per group is allowed, which can be the minimum observed value, the maximum observed value, or a single randomly selected observation within the group, using the maximum observed values is the preferred choice from the parameter estimation accuracy point of view. Second, as expected, access to both the minimum \emph{and} the maximum values further enhances the estimation performance. The optimal estimator, $\hat\theta_{opt}$, obviously achieves the best performance.
Rewriting $J^{min}(\theta)$, $\widetilde{J}^{max}(\theta)$, and $\widetilde{J}^{mix}(\theta)$ in a more general form, with respect to $K$, yields
\begin{IEEEeqnarray}{cCl}
J^{min}(\theta)=\frac{N}{\theta^2}\cdot\mathcal{O}(1) \IEEEyesnumber \label{orderofmag} \IEEEyessubnumber \label{FIMtrueminch3} \\
\widetilde{J}^{max}(\theta)=\frac{N}{\theta^2}\cdot\mathcal{O}(ln^2(K)) \IEEEyessubnumber \label{FIMtruemaxch3} \\
\widetilde{J}^{mix}(\theta)=\frac{N}{\theta^2}\cdot\mathcal{O}(ln^2(K)) \IEEEyessubnumber \label{FIMtruemixch3}
\end{IEEEeqnarray}
where $\mathcal{O}(\cdot)$ stands for order of magnitude. While $J^{min}(\theta)$ is independent of $K$, both $\widetilde{J}^{max}(\theta)$ and $\widetilde{J}^{mix}(\theta)$ depend on $K$ in a similar manner. Furthermore, it can be shown that:
\begin{IEEEeqnarray}{cCl}
\widetilde{J}^{mix}(\theta)-\widetilde{J}^{max}(\theta) \leq J^{L=1}(\theta)=\frac{N}{\theta^2} \IEEEyesnumber \label{Converg} \IEEEyessubnumber \\
\text{where~~~} \widetilde{J}^{mix}(\theta)-\widetilde{J}^{max}(\theta) \underset{\underset{K\rightarrow\infty}{\uparrow}}{\rightarrow} \frac{N}{\theta^2} \IEEEyessubnumber
\end{IEEEeqnarray}
from which, it can be seen that the information stored in the maximum observed values increases as the group size $K$ increases. When the maximum observed values are available, the additional information from the minimum observed values is relatively small; indeed, as $K$ increases, the additional information from the minimum observed values increases, but is bounded by $N/\theta^2$, since, as $K\rightarrow\infty$, the added information converges to $N/\theta^2=J^{min}(\theta)$. This result is expected, since it has been shown that for $K\rightarrow\infty$, the extremes, i.e., the minimum and the maximum values, become independent \cite{coles1999dependence}, and therefore,
\begin{equation} \label{conto1ch3}
J^{mix}(\theta)\underset{\underset{K\rightarrow\infty}{\uparrow}}{\rightarrow}J^{min}(\theta)+J^{max}(\theta)
\end{equation}
TABLE \ref{Jcomp} describes the behavior of the different (approximated) FIMs as a function of the group size $K$. The normalized values of the FIMs, expressed in Eqs. \eqref{FIMexp} and \eqref{FIMmaxapp}, correspond to the equivalent value of $L$ out of the $K$ samples in each group, defined as the \emph{L-equivalent} value (see Eqs. \eqref{CRLBla} and \eqref{CRLBlb}). As Eq. \eqref{Main} suggests $\forall{K}:J^{min}(\theta){=}J^{L=1}$, in Table \ref{Jcomp} the value of $J^{min}(\theta)$ is constant for all $K$. On the other hand, the estimates $\hat{\theta}_{max}$ and $\hat{\theta}_{mix}$ become more accurate as $K$ increases, with $\hat{\theta}_{mix}$ always being better. However, the difference between the performance of the estimates $\hat{\theta}_{max}$ and $\hat{\theta}_{mix}$ converges to $N/\theta^2$, which translates into adding only one sample, as can be seen in the column of $\Delta\equiv{\widetilde{J}}^{mix}(\theta)-\widetilde{J}^{max}(\theta)$. This is again, to be expected, as seen in Eq. \eqref{conto1ch3}.
\begin{table}[!ht]
\centering
\center{\caption{\emph{L-equivalent} values of $J^{opt}(\theta)$, $J^{min}(\theta)$, $\widetilde{J}^{max}(\theta)$, $\widetilde{j}^{mix}(\theta)$, $J^{opt}(\theta)$, and $\Delta\equiv{\widetilde{J}}^{mix}(\theta)-\widetilde{J}^{max}(\theta)$, for selected values of $K$.} \label{Jcomp}}
\centering
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{c c c c c c}
\hline \hline
$K$ & $\left[\frac{{J}^{\text{opt}}(\theta)}{N/\theta^2}\right]$ & $\left[\frac{J^{\text{min}}(\theta)}{N/\theta^2}\right]$ & $\left[\frac{\widetilde{J}^{\text{max}}(\theta)}{N/\theta^2}\right]$ & $\left[\frac{\widetilde{J}^{\text{mix}}(\theta)}{N/\theta^2}\right]$ & $\left[\frac{\Delta}{N/\theta^2}\right]$ \\
\hline \hline
5 & $5$ & $1$ & $2.238$ & $2.794$ & $0.556$\\
\hline
10 & $10$ & $1$ & $4.891$ & $5.539$ & $0.648$\\
\hline
25 & $25$ & $1$ & $9.793$ & $10.580$ & $0.787$\\
\hline
50 & $50$ & $1$ & $14.616$ & $15.482$ & $0.866$\\
\hline
100 & $100$ & $1$ & $20.422$ & $21.341$ & $0.919$\\
\hline
1000 & $1000$ & $1$ & $46.765$ & $47.752$ & $0.987$\\
\hline \hline
\end{tabular}
\end{table}
\subsection{Simulation Results} \label{simm}
In order to validate the relationship of Eq. \eqref{Main} and the insights presented in Table \ref{Jcomp}, the following simulation was designed.
\begin{itemize}
\item[(i)]{One hundred groups of $K$ iid exponentially distributed variables, with parameter $\theta=1$, were generated, i.e., $N=100$.}
\item[(ii)]{For each group, the minimum and the maximum values were logged. Thus, a vector containing 100 minimum observed values and a vector containing 100 maximum observed values were created.}
\item[(iii)]{Using these two vectors as inputs, the estimation processes of Eqs. \eqref{MLEminch3}, \eqref{MLEmaxch3}, and \eqref{MLEmixch3} were performed, in order to obtain the three estimates $\hat\theta_{min}$, $\hat\theta_{max}$, and $\hat\theta_{mix}$. In addition, for comparison, the optimal estimate, $\hat\theta_{opt}$, was calculated from the full dataset.}
\item[(iv)]{Steps (i)-(iii) were repeated 10000 times.}
\item[(v)]{Steps (i)-(iv) were repeated for $K$ ranging from $K=5$ to $K=100$.}
\end{itemize}
In addition to the simulation, the analytical variances, i.e., the CRLBs, of $\hat\theta_{min}$ (exact), $\hat\theta_{max}$ (approximated), and $\hat\theta_{mix}$ (approximated) were calculated using Eqs. \eqref{FIMexp}, and \eqref{FIMmaxapp}.
The simulation results are depicted in {\textbf{Figure} \ref{b15to540}}. The calculated variances of $\hat\theta_{min}$, $\hat\theta_{max}$, and $\hat\theta_{mix}$, and their corresponding asymptotic variances, i.e., CRLBs, as a function of $K$, for $\theta=1$ and $N=100$ are shown. In addition, the calculated variance of $\hat\theta_{opt}$ and its corresponding asymptotic CRLB is drawn. As expected, the variance of ${\hat{\theta}}_{min}$ is independent of $K$, and equals $\theta^2/N=10^{-2}$), whereas the variances of ${\hat{\theta}}_{max}$ and ${\hat{\theta}}_{mix}$ are $K$-dependent, and decrease as $K$ increases. Note that for small values of $K$ ($\approx{K}<15$), the accuracy of the approximated CRLBs drops (due to Eq. \eqref{eqlemma2maxrep}). For larger values of $K$, the approximated CRLBs are rather accurate, and describe well the values of $var({\hat{\theta}}_{max})$ and $var({\hat{\theta}}_{mix})$. The simulated results are encouraging. First, note that the variance of $\hat\theta_{min}$ is constant, and follows the expression suggested in Eq. \eqref{FIMexpmin}. Second, as expected, it can be seen that, regardless of the value of $K$ (for $\approx{K}>15$), the performance of $\hat\theta_{max}$ and $\hat\theta_{mix}$ is described by the approximated expressions of Eqs. \eqref{FIMapprxmaxch3} and \eqref{FIMapprxmixch3}. Furthermore, it can be seen that both $\hat\theta_{max}$ and $\hat\theta_{mix}$ perform significantly better than $\hat\theta_{min}$, which agrees with Eqs. \eqref{FIMtrueminch3}, \eqref{FIMtruemaxch3}, and \eqref{FIMtruemixch3}. Thus, we conclude that the results of the simulation validate the suggested analysis. In addition, note that, even for relatively small values of $K\approx10$, the suggested approximations of $\widetilde{J}^{max}(\theta)$ and $\widetilde{J}^{mix}(\theta)$, although not as accurate as for higher values of $K$, are still usable, as their values are still relatively close to the simulated-based variances.
\begin{figure}[ht!]
\centering
\includegraphics[width=150mm]{fig2.pdf}
\caption{Calculated variances of $\hat\theta_{min}$, $\hat\theta_{max}$, $\hat\theta_{mix}$, and their corresponding asymptotic variances, i.e., CRLBs, as a function of $K$, for $N=100$.} \label{b15to540}
\end{figure}
\subsection{Demonstration of the accuracy of the Fisher information matrix approximations}
The problem of estimating the original PDF parameter vector from a set of the upper or lower record values has been addressed in the field of record theory \cite{records,recordsparam1986}. Therefore, it is interesting to compare the proposed approximations with previously presented results.
The FIM properties, using only the minimum or only the maximum values of an exponentially distributed population, were discussed in \cite{hofmann2003}, and we now compare our findings with these properties.
Specifically, in \cite{hofmann2003} it was shown that the minimum-based estimate (for the exponential distribution) can be solved analytically, and has the \emph{L-equivalent} value of 1, regardless of $K$. This matches our findings, as we concluded that $J^{min}(\theta)=N/\theta^2$, and strengthens our CSV approximation, as we showed that $\widetilde{J}^{min}(\theta)\approx{J}^{min}(\theta)$.
Regarding the maximum based estimation, in \cite{hofmann2003} it was suggested that exponential order statistics properties be used to approximate $J^{max}(\theta)$ (see \cite{arnold1992first}, \emph{example 7.3.4}). The values of this approximation of $J^{max}(\theta)$, defined by ${J}^{max}_{2003}(\theta)$, for selected values of $K$, is presented, in an \emph{L-equivalent} normalized form, in Table \ref{Rcomp}, alongside the \emph{L-equivalent} normalized values of the proposed approximated CLV-based $\widetilde{J}^{max}(\theta)$, and the empirical variance of $\hat\theta_{max}$ calculated from the simulation described in this section, the results of which presented are in {\textbf{Figure} \ref{b15to540}.
As can be seen in TABLE \ref{Rcomp}, although the values of ${J}^{max}_{2003}(\theta)$ are closer to the simulation results, the difference in the \emph{L-equivalent} values of $\widetilde{J}^{max}(\theta)$ and ${J}^{max}_{2003}(\theta)$ is less than $1.6$ for the entire range of $5\leq{K}\leq{200}$. This is encouraging, especially as the $\widetilde{J}^{max}(\theta)$ approximation is not confined to the exponential case and can be utilized for every distribution of interest.
Finally, it is worth noting that we could not compare the approximated results of $\widetilde{J}^{mix}(\theta)$, since the analysis of situations in which both of the extremes, i.e., the minimum \emph{and} the maximum values, are used at the same time has attracted less interest in the past, and no comparable publications that address the values of $\widetilde{J}^{mix}(\theta)$ are known to us. However, based on the presented comparison for both $\widetilde{J}^{min}(\theta)$ and $\widetilde{J}^{max}(\theta)$, in combination with the agreement of $\widetilde{J}^{mix}(\theta)$ with the simulation results (see Table \ref{Jcomp}),
the validity of the CSV- and the CLV-based approximations is promising.
\begin{table}[!ht]
\caption{Normalized values of $\widetilde{J}^{max}(\theta)$, ${J}^{max}_{2003}(\theta)$), and the simulation results, $var^{-1}(\hat\theta_{max})$, for selected values of $K$.} \label{Rcomp}
\centering
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{c c c c}
\hline \hline
& Approximated & Asymp. Exact & Empirical \\
$K$ & $\left[\frac{\widetilde{J}^{\text{max}}(\theta)}{N/\theta^2}\right]$ & $\left[\frac{{J}^{\text{max}}_{2003}(\theta)}{N/\theta^2}\right]$ & $\left[\frac{var^{-1}(\hat\theta_{max})}{N/\theta^2}\right]$ \\
\hline \hline
5 & $2.24$ & $3.66$ & $3.66$ \\
\hline
10 & $4.89$ & $5.86$ & $5.89$ \\
\hline
20 & $8.45$ & $8.87$ & $8.86$ \\
\hline
30 & $10.97$ & $11.05$ & $11.05$ \\
\hline
40 & $12.96$ & $12.79$ & $12.83$ \\
\hline
50 & $14.62$ & $14.26$ & $14.13$ \\
\hline
100 & $20.42$ & $19.45$ & $19.21$ \\
\hline
200 & $27.21$ & $25.63$ & $25.71$ \\
\hline \hline
\end{tabular}
\end{table}
It is worth noting that the approximations of $\widetilde{J}^{max}(\theta)$ and $\widetilde{J}^{mix}(\theta)$ are not the bounds of the actual values of the FIMs. For very large values of $K$, the approximations $\widetilde{J}^{max}(\theta)$ and $\widetilde{J}^{mix}(\theta)$ seem to be less accurate than for moderate values of $K$. This phenomenon is to be expected. In order to find an analytical solution to the presented problem, the expected values of the extremes were replaced with the CSV and the CLV, based on the approximations presented in Eqs. \eqref{ConclusionUmin}, \eqref{ConclusionUmax2}, and \eqref{ConclusionUmix2}. However, the CSV and the CLV converge to the \emph{modes} of the extremes (under the conditions of \emph{Lemma} \ref{Lemma2}), and not to the expected values. The difference between the mode and the expected value of the maximum observed value, is finite, and independent of $K$, whereas both the CLV and the expected value of $Y_{max}$, $E[Y_{max}]$, increase as $K$ increases. Thus, the difference between the CLV, $\mu^{exp}_K$, and the expected value, $E[Y_{max}]$, can be neglected for most values of $K$. However, as the \emph{Gumbel} distribution, to which $Y_{max}$ converges in distribution, is positively skewed, this difference introduces a small bias into the expressions of the FIM. This bias, although small, becomes non-negligible for very large values of $K$, as the variance of $\theta$ decreases. This fact is interesting, but unimportant for practical purposes, as it starts to affect the accuracy of the FIM approximations only for very large values of $K$ ($K>>100$), which is generally non-realistic for real-world usage, as for such large $K$, the information held in the extremes is only a fraction of the information held in the entire set of measurements (e.g., as presented in TABLE \ref{Jcomp}, for $K=1000$, estimating $\theta$ using the set of the maximum values is equivalent (performance wise) to an estimate based on 47 measurements per interval, which is a merely $4.7\%$ of the available original measurements).
\section{Conclusion}
This paper deals with the information in extreme values with respect to parameter estimation. We presented a new approach which uses the Characteristic Values of the extremes, by which we were able to establish simple and analytical solvable approximations to the expressions of the Fisher Information Matrices of the estimates based on either the minimum, the maximum, or the minimum \emph{and} the maximum measurements.
Based on these approximations, we designed a new tool to evaluate the accuracy of estimates which use only the minimum or the maximum measurements, relative to the optimal estimation, which uses the entire dataset of measurements. We showed that the presented methodology gives simple and practical expressions, which are solvable, yet still capable to approximate the performance of the estimates. Furthermore, we expanded the proposed practical expressions and were able to analyse the performance of the estimate based on both extremes (i.e., the minimum \emph{and} the maximum values, combined).
We demonstrated our tool and performed an analysis for the case of the exponential distribution, which is an important example, as it is used to model many naturally occurring phenomena, especially for environmental monitoring. This fact makes the exponential case important for real-world applications. In addition, the exponential case is one of the few examples that were studied previously, and allowed us to perform a detailed comparison of our approximations with past results. Moreover, as the exponential distribution share commonality with many other distributions (in which, the extreme behaviour is similar), such as the log-normal and the gamma, our performed analysis on the exponential case, which demonstrates the importance of the maximum values rather than the minimum values (with respect to the accuracy of the outcome estimates) can be directly applied to other distribution families.
To conclude, we believe that the new approach and tools we have presented in this paper hold the potential to be used in many current and future applications, especially in the emerging field of IoT sensing, where the logging and transmitting of less observations is desired, for instance, where the cost per additional measurement is high.
\singlespacing
|
\section{Appendix}
\section{Survey Material}
\subsection{Ante Hoc Demographic Questionnaire}
\label{app:antehoc}
\begin{enumerate}
\item What is your age? (18-24, 25-34, 35-44, 45-54, +65, NA)?
\item What is your identified gender?
\item Do you have any physical conditions that might prevent you from observing authentication gestures performed on a mobile phone?
\item Do you use a smartphone currently? If so, what is its operating system?
\item Why did you select that phone and OS?
\item If you currently use an authentication method to lock your phone, what is the method (i.e. PIN, TouchID, grid, etc.), and why did you select it?
\item What types of mobile phone authentication have you used? (i.e. PIN, grid pattern, password, fingerprint, face, voice, other)
\item Without telling me your current passcode, how do you select the passcodes you use?
\item How concerned are you with keeping your phone secure (1, not at all concerned, to 5, highly concerned)?
\item What experiences can you recall involving people either trying to steal or use your phone without permission?
\item What experiences can you recall involving people trying to observe your passcodes without permission?
\item How concerned are you with the threat of someone watching you authenticate and collecting your passcodes (1, not at all concerned, to 5, highly concerned)?
\item If you had any of these experiences, how did it affect your behavior?
\item Have any other experiences or concerns affected your authentication?
\end{enumerate}
\subsection{Post Hoc Participant Strategies Questionnaire Questions}
\label{app:posthoc}
\begin{enumerate}
\item What strategies did you employ to collect the passcodes?
\item Do you have any ideas for additional strategies?
\item How challenging was it to collect PIN passcodes (1, not at all challenging, to 5, very challenging)?
\item How challenging was it to collect grid passcodes (1, not at all challenging, to 5, very challenging)?
\item What features of the passcodes made it easier or more difficult to collect the passcodes you saw?
\item How did the number of views you were given make a difference?
\item How did which side you stood on make any difference?
\end{enumerate}
\subsection{Observation Forms}
\label{fig:obvforms}
\begin{center}
\fbox{\includegraphics[width=0.8\linewidth]{forms/obvforms_pg1.eps}}
\fbox{\includegraphics[width=0.8\linewidth]{forms/obvforms_pg2.eps}}
\fbox{\includegraphics[width=0.8\linewidth]{forms/obvforms_pg4.eps}}
\end{center}
\subsection{Guide/Script for Administering Study}
\label{app:studyguide}
\begin{enumerate}
\item Verify current participant number, exp (1-4, order (a-c). Record this.
\item Introduction - "Welcome, thanks for participating. Our study deals with the security of different types of passcodes for mobile phones. Your help today will be pretty straightforward. We will record some basic demographic information about you. You will observe someone entering in passcodes from a few feet away, and write down your best guess of what you have seen. We will go over the steps required to make sure you are comfortable and understand your role. We will record the session to verify the results. All data collected will be anonymized for publication. Your part in the study should take about 20-30 minutes."
\item Payment - "The study pays 5 USD."
\item Observer disclosure - "Are there any issues, such as corrective glasses or contacts, which might interfere with performing the role I have described?"
\item IRB Introduction - "This study has been reviewed by the University's review board, the IRB, and approved as safe and ethical. Here is a copy of that form that describes the study that you can read. Please ask any questions you may have, and sign the form if you would like to participate."
\item Demographic questionnaire - "Please fill out the demographic questionnaire."
\item Training: Overview - "Here is the process your role as an observer. You will stand to the left or right behind our researcher, who will sit in the same position entering in passcodes. We will specify where to stand for each attempt. You will watch each attempt, and then draw on the form we provide your best guess of the passcode you just saw being entered on the phone. Passcodes may vary in length. We will repeat this ten times."
\item Training: Filling out the form diagrams - "Look at the form we have provided. It has blank PIN and grid pattern diagrams for you to enter your guesses. For the grid patterns passcodes, draw the shape you saw entered, and circle the starting point of the shape. For PIN passcodes, write out the sequence, like "1234", and draw the shape you saw entered."
[Demonstrate drawing a diagram, then allow the participant to practice drawing 2-3 times, based on a practice code for their prescribed passcode condition (PIN, grid, no-line grid) that you show them slowly, up close, on a phone. Confirm for grid shapes that they are circling the starting point. Correct any issues that appear, and repeat until ready.]
\item Training: Taking position for each attempt - "We will call out a position, LEFT or RIGHT, for you to stand in for each attempt. Move to the corresponding marker on the floor, figure out which diagram you are going to fill in, and when set to begin, say "Ready." Sometimes, we may also call out MULTIPLE if you are allowed to view the passcode being entered twice before making your final guess."
"Any questions?"
\item Post hoc questionnaire - [Conduct post hoc interview]
\end{enumerate}
\section*{List of Provided Material}
\subsection{Patterns and PINs Visualized}
\label{sec:viz}
\input{pattern_viz}
\subsection{Pre-Survey Questions}
\begin{itemize}
\item What is your age (18-24, 25-34, 35-44, 45-54, +65, NA)?
\item What is your identified gender?
\item Do you have any physical conditions that might affect your ability to
enter authentication passcodes on a mobile phone?
\item Do you use a smartphone currently? What is its operating system? Why did
you select that phone?
\item Do you use an authentication method to lock your phone, and if so which
method, and why (i.e. PIN, grid, TouchID, etc.)?
\item Without telling me your current passcode, how do you select the passcodes
you use to lock your phone (i.e. familiar number, or visual pattern)?
\item How concerned are you with keeping your phone secure (1, not at all
concerned, to 5, highly concerned)? item What experiences can you recall
involving people either trying to steal or use your phone without permission?
\item What experiences can you recall involving people trying to observe your
passcodes without permission?
\item How concerned are you, typically, in a public space, with the threat of
someone watching you authenticate and collecting your passcodes (1, not at all
concerned, to 5, highly concerned)?
\item If you had any of these experiences, how did it affect your behavior?
\item Have any other experiences or concerns indirectly affected your
authentication behavior (news articles, stories about friends, etc.)?
\item If you do authentication, how do you typically hold your phone for that?
\end{itemize}
\subsection{Post-Survey Questions}
\begin{itemize}
\item On a scale from 1-5, how difficult was entering passcode this way (1, very
easy, to 5, very hard)? How so?
\item On a scale from 1-5, how easy was the grid pattern tactile app to learn
(1, very easy, to 5, very hard)? How so?
\item On a scale from 1-5, how easy was the grid pattern tactile app to use (1,
very easy, to 5, very hard)? How so?
\item On a scale from 1-5, how easy was the PIN tactile app to learn (1, very
easy, to 5, very hard)? How so?
\item On a scale from 1-5, how easy was the PIN tactile app to use (1, very
easy, to 5, very hard)? How so?
\item Can you see yourself using the grid pattern tactile aid to help
authenticate on your phone in your actual daily life? Why or why not?
\item Can you see yourself using the PIN tactile aid to help authenticate on
your phone in your actual daily life? Why or why not?
\item How is this approach similar or different from how you enter passcodes on
your phone now?
\item Do you think the grid tactile aid would help protect you from someone
shoulder surfing you? Why or why not?
\item Do you think the PIN tactile aid would help protect you from someone
shoulder surfing you? Why or why not?
\end{itemize}
\section{Conclusions}
In this paper, we have described a study comparing video recreations of shoulder
surfing to live simulation. We recreated a subset of the factors explored in the
video study and attempted to confirm prior findings in this setting.
We were able to confirm many of the prior
claims regarding the video study, that authentication type, repeated viewings,
observation angle, and passcode properties can affect attacker performance. We
were also able to confirm that video study does form a baseline for the live
simulation; however, this baseline may be much less than desired, as much as
1.9x difference. From these findings we suggest, for researchers conducting
shoulder surfing studies with video components, that data can form a baseline
and be representative, in many situations, of what would occur in a live
simulation. However, when possible, those results should be compared to a live simulation to get a fuller picture of the data and results.
\section{Introduction}
Biometric authentication mechanisms offer considerable promise to smartphone
users. However, the protection of unlock authentication still relies on choosing
hard to guess passcodes (e.g., PINs and unlock patterns), while not revealing
those passcodes to untrusted parties. A common means of attack for gaining access to the
passcode is via {\em shoulder surfing}. In a shoulder surfing attack, an
observer attempts to view a victim in the process of entering his/her passcode
with the intention of recreating that passcode after gaining possession of the
device~\cite{wiedenbeck2006design}.
The area of shoulder surfing has been the subject of a great deal of work~\cite{egelman2014areyouready, forget2010eyegaze, man2003shoulder,
deluca2014nowyouseeme, deluca2009look, deluca2012touchme, deluca2010colorpin,
kumar2007reducing,gao2010new,aviv2017shoulder}, for both understanding the
threat and proposing mechanisms to prevent it. Of particular relevance to this
study (termed "current study"), is the work conducted by Aviv et al.~\cite{aviv2017shoulder} (termed: "prior study"). The prior study examined the shoulder surfing susceptibility of three commonly used unlock authentication mechanisms: 4- and 6-digit PINs, 4- and 6-length Android
graphical patterns, and 4- and 6-length Android graphical patterns with the
feedback display turned off (lines rendered by the interface between grid points as they are touched by the user). Due to the difficult nature of evaluating shoulder
surfing attacks in the field, the goal of the prior study was to establish
baselines for shoulder surfing vulnerability in controlled settings that can be
used to compare across authentication types and used as baseline for evaluating
authentication systems that are designed to defend against such attacks.
To control the analysis, the prior study was conducted using a video-based
methodology where the researchers recorded a set of videos with highly
controlled factors and then asked participants to view these videos as a
simulated shoulder surfing scenario. The data was analyzed to determine
shoulder-surfing susceptibility under each condition. The attack rate (how effectively the participant could recall the passcode entered in the video) was the primary metric.
In this paper, we seek to compare the video-based methodology to a similarly
controlled live setting. In particular, we are interested in assessing the prior
work's following findings relating to the attack success rate.
\begin{itemize}
\item Longer authentication lengths (e.g, 4-digit vs. 6-digit PINs) are less
vulnerable.
\item PIN authentication is less vulnerable to the attack compared to patterns
with and without feedback lines.
\item Removing the feedback lines from patterns decreases the vulnerability to
shoulder surfing.
\item Multiple observations increases vulnerability.
\item Video based evaluation provides a baseline for live, in-person shoulder
surfing vulnerability.
\end{itemize}
Using the raw results of the prior study, we compare the attacker success rates
of the live setting to a comparable subset of the video study data. Testing for
differences in proportionality, we are {\em unable} to reject the null
hypothesis that the attacker success rate are the same for Android patterns as
well as in many of the settings with patterns without feedback lines. This
suggests that there is consistency between the results of the video and live
simulations. However, the advantage of removing feedback lines previously
observed in video simulation is considerably lessened in the live setting. For
PINs, we observe significant difference between the video and the live settings,
where live attackers performed up to 1.9x better in some scenarios. Stereo
vision seems to greatly improve the reliability of recalling the more complex
motions of entering a PIN. Despite this discrepancy, the claim of Aviv et al. of
these results forming a baseline is still supported: we never observed a
situation by which the live simulation performed worse than a video study when
significant differences exist.
We conclude that video studies do provide a reasonable approximation for live
simulation of shoulder surfing in settings that involve graphical passwords (but
not PINs), like the Android password pattern, and at least a lower-bound on the
attack success rate for all tested authentication types (including PINs).
However, researchers should consider that this lower-bound may be a significant
underestimation compared to the true attack rate in live simulations.
\vspace{5 mm}
\section{Realism and Limitations}
\label{sec:limits}
As described in the previous section, we attempted, as best as possible, to
recreate the settings of the prior video study in live simulation. Due to the complexities of performing such a process, the study described in this paper had its own set of limitations.
\paragraph{Viewing angles} While we use a similar lab environment for the live
simulation to that used in the video study, the participants could not stand in
exactly the same position as the cameras due to height differences and the
relatively close proximity of the {\em near} and {\em far} angles from a given
side. We thus reduced the observations to simply {\em left} and {\em right} and
relied on the fact that our participants naturally vary in height to compensate
for the {\em near} and {\em far} setting of camera height placement in the prior
study.
\paragraph{Victim entry speed}
Another recreation challenge is that our victim (a proctor)
must enter the authentication sequence many times over at a consistent speed. Clearly, a
video ensures consistency here, and so we trained the victim-proctor on the original
videos to maintain consistent timings of authentication entry. While there is no guarantee that every
participant viewed the authentication at the same rate, we believe this
training, and the total number of entries performed by the victim, ensures
consistency. Further, the same victim-proctor was used in all data collection.
\paragraph{Subset of conditions} As summarized in Section~\ref{sec:methods}, a
subset of the original conditions were used in the live simulation. We kept
factors that were shown to be significant in the video study, but also had to
remove some that posed usability challenges for the proctor acting as the
victim. While the selection process was done carefully to address conditions
likely to be important, it was also done for a practical nature of conducting a
study with live participants as compared to online. To ensure that we made a
fair comparison, we selected a similar subset of the data from the prior study.
In particular, we used results from the previous study from participants who
had viewing screens of at least 1800px across, who viewed authentication attempts via the Nexus 5 phone with
thumb input from the left or right side.
\paragraph{Pen-and-paper attacker recordings}
As participants were using pen-and-paper to record their observations
during the shoulder surfing attack, some participants were able to use this as
an added aid to support recall of the passcodes. For example, some participants were
viewed by the proctor mimicking the movements made by the victim-proctor between multiple-view conditions
prior to writing down their final observation. While we directed participants to
{\em not} do this during training, it was difficult to stop due to the nature of
the task. In the video study, participants were also directed not to use
additional aids, such as writing down observations while observing the passcodes,
and were required to attest to this. However, it is possible that the attestations
were not fully truthful, nor could the researchers verify this as the study was
conducted online. As such, as neither study could fully control for this we
believe that this provides for a fair comparison.
\paragraph{Ecological validity}
Low levels of ecological validity are known to be commonplace among lab-based studies for mobile interactions \cite{kjeldskov2014worth}. Although the method and setting selected for our study cannot approximate the conditions by which shoulder surfing may take place in-the-wild, we designed the study to provide a sense of realism even in a lab-based environment (e.g. victim in seated position similar to attacks taking place while seated on public transport, while seated in a classroom, etc.). However, due to time constraints, conditions such as providing multiple attempts to observe and/or recreate entry, could not be examined. Further study would be needed to widen the range of factors examined, and to identify the applicability of these findings to other types of tasks (e.g. authenticating while ambulatory) or other types of settings (e.g. field-based).
\section*{Acknowledgments}
This work was supported by the Office of Naval Research. The authors wish to thank Chukwuemeka KC Marume and John T. Davin for their assistance conducting the study.
\clearpage
\bibliographystyle{acm}
\section{Methodology}
\label{sec:methods}
To investigate the efficacy of video-based recreations for evaluating
observation attacks, we recreated the study conducted by Aviv et
al.~\cite{aviv2017shoulder} with live participants in a controlled lab
environment. We asked participants to position themselves in similar locations
to where the cameras were positioned in the prior study. They then attempted to
shoulder-surf a victim (played by a proctor). We varied the type and length of
authentication sequences, observation angle, and number of repeated viewing
attempts, to determine if these factors impact the success of the attacker. The
results were then compared with Aviv et al.'s findings using a comparable subset
of the prior data. For simplicity of discussion, we refer to the prior work of
Aviv et al. as the {\em video study} and the results here as the {\em live
study}.
\paragraph{Hypotheses} In particular, we are interested in testing the following
hypothesis related to the efficacy of video based shoulder surfing experiments
as compared to live settings.
\begin{itemize}
\item {\bf H1-r}: Live shoulder surfing confirms accepting prior hypotheses:
\begin{itemize}
\item {\bf H1-p}: The authentication type affects shoulder surfing vulnerability
\item {\bf H2-p}: Repeated viewing affects shoulder surfing vulnerability
\item {\bf H4-p}: The angle of observation affects shoulder surfing vulnerability
\item {\bf H5-p}: The properties of the passcodes affects should surfing vulnerability
\end{itemize}
\item {\bf H2-r}: Video simulation forms a baseline of performance compared to
live settings.
\end{itemize}
\subsection{Study Design and Materials}
\paragraph{Treatments}
The study followed a mixed factorial design, similar to the video study.
Independent variables included authentication type (PIN vs pattern) on the Nexus
5 device using the same hand posture/interaction style (one-handed, right thumb
input). For dependent variables, we reduced the observation angle to two (left
or right) as opposed to the five angles used in prior work. The video study used
the variety of angles to simulate different heights, but height variation is
naturally present in a live study. We kept the same variables for observations
(single observation from one angle, two observations from the same angle, or two
observations from different angles), and we used a lab environment for our live study very similar to the set-up to capture videos for the video study (Aviv et al.) (see
Figure~\ref{fig:setup}).
There were two notable differences between factors in the video study and the
live study. First, we only allowed each participant a single attempt at
recreating the passcode. This choice was motivated by results of the video study
whereby participants, knowing they would have multiple attempts in advance,
actually did worse at the tasks than those that knowingly had one attempt. It
was conjectured that participants attempted to ``game'' the task knowing that
they would have multiple attempts at recreating the passcode. As such, we only
allowed participants to make one recreation attempt, and this fact was
communicated during training.
Another difference in the live study was that passcode recreation occurred
using pen-and-paper, as opposed to a simulation of the device used in the video study. This choice was made to simplify the data collection
procedures for both proctors and participants.
Finally, as we only tested a subset of the treatments of the prior video study,
we only performed our analytic comparisons on a relevant subset of the video
study data. In particular, we removed data that included a top angle and reduced
the two side angles into a single {\em left} or {\em right}
setting. Additionally, as the video cannot control for monitor display size,
which was a large factor in the prior results, we only used the most ideal
viewing conditions, where the reported y-axis pixels were greater than 1800. We
believe this restriction provided the {\em most fair} comparisons possible given
the potential uncontrolled factors. We discuss limitations and realism further
in Section~\ref{sec:limits}.
\paragraph{Authentication types}
We analyzed three authentication types with two different length settings, as
used in the video study. These included:
\begin{itemize}
\item {\bf PIN}: 4- or 6-length PINs consisting of a set of numbers.
\item {\bf PAT}: Android unlock patterns consisting of 4 or 6 contact points {\em with} the feedback lines present.
\item {\bf NPAT}: Android unlock patterns consisting of 4 or 6 contact points
{\em without} the feedback lines present.
\end{itemize}
While the PIN interaction display is as one expects, the presence or absence of grid pattern feedback lines is less well known. When a pattern is entered with feedback lines (PAT), the display will show connecting lines on the screen between grid points touched by the user while entering their passcode shape. Alternatively, the connecting lines are not rendered on screen during passcode entry in the without feedback lines (NPAT) pattern display, although the user must still contact the appropriate points in the correct order. As identified by Aviv et al.~\cite{aviv2017shoulder} and von
Zezschwitz et al.~\cite{vzw2015easy}, the absence of feedback lines can make it more
difficult for an observer to recreate the patterns. As part of {\bf H1-r}, we
will make a similar evaluation.
\begin{table}[t]
\centering
\small
\begin{tabular}{ c | l | l c}
{\bf Auth. id} & {\bf Patterns} & {\bf PINs} \\
\cline{1-3}
0 & 0145 & 1328 & \multirow{10}{*}{\includegraphics[width=0.2\linewidth]{imgs/pattern_labeled.eps}}\\
1 & 014763 & 153525 &\\
2 & 1346 & 159428 &\\
3 & 136785 & 1955 &\\
4 & 3157 & 366792 &\\
5 & 4572 & 441791 &\\
6 & 642580 & 458090 &\\
7 & 6745 & 5962 &\\
8 & 743521 & 6702 &\\
9 & 841257 & 7272 &\\
\end{tabular}
\hfill
\caption{Authentication identifiers for patterns and PINs. To the right, the
numeric labeling for patterns to contact points.}
\label{tab:passcodes}
\end{table}
To maintain consistency, we used the same set of patterns and PINs as in prior
work (Table~\ref{tab:passcodes} and Appendix \ref{fig:patterns}). The patterns
were selected from an online study of self-reported
patterns~\cite{aviv2015isbigger}, and the PINs were obtained from sequences of
digits in leaked password sets, similar to the analysis by Bonneau et
al.~\cite{bonneau2012birthday}. Further, the set of passcodes were selected for
physical properties, as the layout and sequence of gestures in entry may affect
shoulder surfing attack rate. The patterns' spatial properties might affect surfing attacks because an attacker's view from some viewing angles might be obscured for some parts of the touchscreen.
\paragraph{Randomization and counterbalancing}
One of the restrictions for performing the study using live participants as
compared to video recreation is that the same level of randomization is nearly
impractical for the target recruitment size and the set of factors being
considered. As such, we designed a two stage randomization procedure, one for
ordering the passcodes and one for ordering the observation angles.
In particular, Table~\ref{tab:orders} contains three different randomized
orders across the passcode. These are labeled Order a, b, and c. Note that the
authentication identifiers refer to Table~\ref{tab:passcodes}. In
Table~\ref{tab:angleorder} are four randomized orders for observation angles (i,
ii, iii, and iv). For each participant, we randomly assigned them a
passcode order and an observation angle, producing 12 different
randomizations.
At this point, it is important to consider counterbalancing. Selecting
randomized orders for passcodes or observations can weight the data
improperly. This leads to an optimization problem, and we used a utility
function to find a set of randomized orders that would provide (1) sufficient
data in each factor for us to perform statistical tests, (2) a roughly equal ratio of data within each factor being compared (4- vs 6-length,
auth-type, angle), (3) that each passcode only appears once per viewing, and (4) that within
each viewing sequence, per participant, there are roughly an equal number of
single and multiple observations. We found a case that nearly met these
criteria, as displayed in Table~\ref{tab:orders} and~\ref{tab:angleorder}. The
weighting is then displayed based on 12 participants in
Table~\ref{tab:counterbalance}, leaving us with 72 single-view observations and
48 multi-view observations, 24 from the same angle twice and 24 from two different
angles. Additionally, there is equal weighting across angles and viewing
(Table~\ref{tab:angleorder}), and nearly equal weighting across passcodes.
We acknowledge that this counterbalancing is not a perfect weighting, and
solving this particular optimization problem is challenging and may
not have a solution. However, the resulting counterbalancing compares
favorably to the subset of relevant video study data. For PINs, there is nearly an
equal number of observations in the one-view and two-view conditions. For
PAT/NPAT, there is 50\% less observations in one-view condition with a
significant proportion necessary for statistical testing, and the two-view
conditions for PAT/NPAT are of the same magnitude as the video study (see
Table~\ref{tab:views}).
In total, we were able to run complete trials for 18 participants each for PAT and
NPAT, and all of those 36 participants also completed a PIN viewing. The order
between PIN and PAT/NPAT for participants was randomized, so that half of the
participants completed a PIN trial before doing a PAT/NPAT trial, and the other
half completed the protocol in the reverse order, PAT/NPAT then PIN.
\begin{table}[t]
\centering
\small
\begin{tabular}{ c | c c c c c c c c c c}
{\bf Order } & \multicolumn{10}{c}{{\bf Auth. id}}\\
\hline
a & 8& 1& 0& 7& 9& 2& 6& 5& 4& 3\\
b & 0& 6& 3& 8& 2& 4& 9& 7& 1& 5\\
c & 6& 0& 9& 4& 8& 3& 5& 1& 7& 2\\
\end{tabular}
\caption{Orderings of the patterns and PINs in the experiments.}
\label{tab:orders}
\end{table}
\begin{table}[t]
\centering
\footnotesize
\begin{tabular}{ c | c c c c c c c c c c}
{\bf Exp. } & \multicolumn{10}{c}{{\bf Angle(s)}}\\
\hline
i & L & R & R & RR & L& LR& RL& R& L& LL \\
ii & RL& L& LR& R& LL& R& R& L& L& RR \\
iii &LL& RR& L& R& R& L& R& LR& RL& L \\
iv & LR& R& R& L& L& RL& RR& L& LL& R \\
\end{tabular}%
\caption{Angles used within each experiment, including multiple views with two angles indicated. L=view from left side, R=view from right side.}
\label{tab:angleorder}
\end{table}
\begin{table}[t]
\centering
\small
\begin{tabular}{ c | c c c c c c | c c c}
{\bf Auth. id} & {\bf L} & {\bf R} & {\bf LL} & {\bf RR} & {\bf LR} & {\bf RL} & {\bf one} & {\bf two-same} & {\bf two-different}\\
\hline
0 & 3 & 4 & 1 & 1 & 2 & 1 & 7 & 2 & 3\\
1 & 5 & 3 & 1 & 1 & 1 & 1 & 8 & 2 & 2\\
2 & 4 & 3 & 2 & 1 & 1 & 1 & 7 & 3 & 2\\
3 & 3 & 4 & 1 & 1 & 2 & 1 & 7 & 2 & 3 \\
4 & 4 & 3 & 1 & 1 & 1 & 2 & 7 & 2 & 3\\
5 & 3 & 4 & 1 & 2 & 1 & 1 & 7 & 3 & 2\\
6 & 2 & 4 & 1 & 2 & 1 & 2 & 6 & 3 & 3\\
7 & 5 & 3 & 1 & 1 & 1 & 1 & 8 & 2 & 2\\
8 & 4 & 3 & 2 & 1 & 1 & 1 & 7 & 3 & 2\\
9 & 3 & 5 & 1 & 1 & 1 & 1 & 8 & 2 & 2\\
\hline
total &36&36 & 12 & 12 & 12 & 12 & 72 & 24 & 24
\end{tabular}%
\caption{Balancing of observation angles, number of views, for each authentication after 12 participants, $\text{Order}\times\text{Exp}$.}
\label{tab:counterbalance}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=0.45\linewidth]{imgs/inPerson_setup_front_view_03_bothBlurred.eps}
\caption{Experimental setup with an observer attacking a victim, a member of a research team. Note the Google Glass displaying the passcode to enter.}
\label{fig:setup}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.2\linewidth]{imgs/Grid-screenshot.eps}
\hspace{.2in}
\includegraphics[width=0.2\linewidth]{imgs/PIN-screenshot.eps}
\caption{Screenshots of the web-based applications used by the victim entering the passcode. Note, that for the pattern without feedback lines setting, the white trace lines would {\em not} appear.}
\label{fig:apps}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.45\linewidth]{imgs/glass-pat.eps}
\hfill
\includegraphics[width=0.45\linewidth]{imgs/glass-pin.eps}
\caption{Screenshots of the Google Glass application as viewed by the victim and enter the correct PIN or pattern.}
\label{fig:glass}
\end{figure}
\subsection{Live Simulation Setup and Coordination}
We sought to recreate nearly the same scenario for shoulder surfing as the video study. Namely, we had our victim placed in a sitting position with the participant observer behind the victim, either standing to the right or the
left, directed by one of two proctors. These were the same positions where the cameras were located (near left and near right views) in the study by Aviv et al.~\cite{aviv2017shoulder}. See Figure~\ref{fig:setup} for a visual of this
arrangement for the live study.
Additionally, for the phone application used to enter the passcodes, we used the
same mobile applications as in the prior study, which includes a web-based
platform for entering PINs and patterns. Screenshots of those applications are
provided in Figure~\ref{fig:apps}.
For patterns with feedback lines, the white tracing lines would follow the user
gesture, and once the pattern was entered, it would remain visible on-screen for
a half a second before disappearing. The same would be true for the patterns
without feedback lines, however, neither the tracing lines nor the
contact points of the grid would be rendered on the screen. For PINs, the layout
allowed for numeric entry as expected. Once digit keys were selected, the
corresponding digits were presented on the interface. These would then fade to
a * after a half a second, similar to most mobile PIN entry interfaces.
For the participant observer to record their pattern entry, we used pen and
paper. Examples of the observer forms are provided in the
Appendix (\ref{fig:obvforms}). The forms had text boxes and mini-diagrams of the application interfaces, so
the participants could easily record the observed entry. Participants were asked not to write down the passcodes observed until directed following all observations, which was important for the multiple viewing scenario.
As shown in
Figure~\ref{fig:setup}, two pre-marked spots were placed on the floor to direct
participants where to stand on the left or right side. The second proctor,
following the randomized treatment order, would call out directions to the
participant; for example, ``one view, from the left'' or ``two views, first from
left and then right.'' Once the participant was in place for each view, the
second proctor would cue the first proctor (playing the victim) to enter a passcode.
At this point, a significant challenge we had to overcome was how to prompt the
victim-proctor with the correct passcode to enter without tipping off the
participant-observer. Due to the randomization procedures, requiring the
victim-proctor to memorize the numerous orderings was not realistic. As a
solution, we developed a Google Glass application to guide the victim-proctor
through the various passcode orders. Google Glass is a wearable eyeglass display unit that runs on a modified Android OS. It enables one to scroll interactively
through images projected onto a viewing screen built into the right
eyepiece. Moreover, the small display screen on the Google Glass was not visible to the
participant. A screenshot of the Google Glass application is provided in
Figure~\ref{fig:glass}. \vspace{5 mm}
\subsection{Procedure}
The replication experiment proceeded in four stages:
\begin{enumerate}
\item Informed Consent and Ante Hoc Questionnaire: All participants were
properly informed and consented, as we conducted an IRB approved
experiment. Following consent, we asked participants to complete
an ante hoc questionnaire that covered basic demographic questions, such as
age and gender, as well as questions regarding the participants experience
with smartphones, mobile authentication, and sense of risk from shoulder
surfing.
The subjective response questions were largely intended to orient participants to physical security issues related to the study. The ante hoc questions are found in the Appendix (\ref{app:antehoc}).
\item Training: Depending on the set of authentications being observed in the
trial run, a training session would include two basic passcodes, the L shape
for patterns and the 1234 PIN, to help familiarize the participant with the
procedures, how to record on the observation sheets, and where to stand
for the trials (similar training was performed in the video study). Additional
training on how to fill out the observation form was also provided, which is
included in the Appendix (\ref{fig:obvforms}).
\item Trial: Under the direction of the proctor, the participant conducted 10 observations of either the PAT or NPAT pattern entry, and 10 observations of PIN entry.
\item Post Hoc Questionnaire: Following the trials, the participant answered a
series of post hoc questions related to the challenge of the task and his/her
perceived performance thereon. See Appendix~\ref{app:posthoc} for the set of
post hoc questions.
\end{enumerate}
As each participant completed two trials, one for either PAT or
NPAT and another for PIN, once the trial stage was over for the first
authentication we would return to training for the second authentication. As a way to control for training effects, whereby observing PINs first could increase
or decrease performance on observing PAT/NPAT, we ensured that there was an even
ratio between the order of the trials. A guide was also followed to ensure that
the researchers followed the same steps in the protocol (see
Appendix~\ref{app:studyguide}).
\begin{table*}[t]
\centering
\small
\begin{tabular}{ c |r | c | c | c || c | c | c | c}
& & \multicolumn{3}{c||}{\em Live} & \multicolumn{4}{c}{\em Video} \\
& & {\bf Male} & {\bf Female} & {\bf Total} & {\bf Male} & {\bf Female} & {\bf Neither} &{\bf Total}\\
\hline
\multirow{6}{*}{\rotatebox{90}{\em Age}}
&18-24 & 16 & 4 & 10 (27.8\%) & 30 & 9 & 0 & 39 (39.4\%)\\
&25-34 & 8 & 5 & 13 (36.1\%) & 24 & 10 & 1 & 34 (34.3\%)\\
&35-44 & 1 & 1 & 2 (5.6\%) & 10 & 4 & 0 & 14 (14.1\%)\\
&45-54 & 0 & 0 & 0 (0.0\%) & 4 & 3 & 0 & 7 (7.1\%)\\
&55-64 & 1 & 0 & 1 (2.8\%) & 1 & 0 & 0 & 1 (1.0\%)\\
& 65+ & 0 & 0 & 0 (0.0\%) & 1 & 2 & 0 & 3 (3.0\%)\\
\cline{2-9}
&{\bf total} & 26 (72.2\%) & 10 (27.8\%) & {\bf 36} & 70 (70.7\%) & 28 (28.2\%) & 1 (1.0\%) & {\bf 99}\\
\hline
\multirow{3}{*}{\rotatebox{90}{\em Phone}}& iOS & 15 & 7 & 22\\
&Android & 9 & 2 & 11\\
&Windows & 2 & 0 & 2\\
\hline
\multirow{5}{*}{\rotatebox{90}{\em Unlock}}& Fingerprint & 15 & 6 & 21\\
&PIN-6 & 8 & 5 & 13\\
&PIN-4 & 10 & 2 & 12\\
&Pattern & 5 & 1 & 6\\
& None & 3 & 1 & 4
\end{tabular
\caption{Demographic, phone usage, and unlock authentication types of participants. For the video study, the subset of comparable data that includes participants in both the ``in-person'' and ``online'' settings that had screen resolution greater than 1800px and observed patterns on the Nexus 5 phone.}
\label{tab:dem}
\end{table*}
\subsection{Recruitment}
Participants were recruited from university student mailing lists, and paid \$5
(USD). In total, we recruited 36 participants, including 10 females. The cohort was predominately aged between 18 to 24 years old. Almost two-thirds of participants used iOS mobile devices. 21 used a fingerprint reader to unlock their phones, and 6 used patterns (we did not ask if feedback lines were turned off).
The demographic breakdown, as well as their choice in mobile device and
authentication are presented in Table~\ref{tab:dem}.
Additionally presented in Table~\ref{tab:dem} are the demographics of a
comparable set of participants from the prior video study; these participants
observed authentication on the Nexus 5 phone in the ``in-person'' lab setup or
the on-line MTurk setup with a screen resolution of at least 1800px in the
y-axis, the most realistic setting of the prior work. The breakdown of these two
groups are similar, slightly younger overall with about 70/30 gender breakdown.
\section{Visualization of Authentication}
Below are the set of patterns and PINs as visualized in prior
work~\cite{aviv2017shoulder}, as well as the description of the visual
properties.
\vspace{50 mm}
\subsection{Patterns}
\label{fig:patterns}
\begin{center}
\begin{tabular}{c c}
\fbox{\includegraphics[width=0.1\linewidth]{imgs/patterns/0-1-4-5.eps}} &
\fbox{\includegraphics[width=0.1\linewidth]{imgs/patterns/6-7-4-5.eps}} \\
0145 & 6745 \\
up & down\\\\
\end{tabular}
\begin{tabular}{c c}
\fbox{\includegraphics[width=0.1\linewidth]{imgs/patterns/3-1-5-7.eps}} &
\fbox{\includegraphics[width=0.1\linewidth]{imgs/patterns/1-3-4-6.eps}} \\
3157 & 1346 \\
neutral & left \\\\
\end{tabular}
\begin{tabular}{c c}
\fbox{\includegraphics[width=0.1\linewidth]{imgs/patterns/4-5-7-2.eps}} & \fbox{\includegraphics[width=0.1\linewidth]{imgs/patterns/7-4-3-5-2-1.eps}} \\
4572 & 743521 \\
right/cross & up/non-adj \\ \\
\end{tabular}
\begin{tabular}{c c}
\fbox{\includegraphics[width=0.1\linewidth]{imgs/patterns/1-3-6-7-8-5.eps}} &
\fbox{\includegraphics[width=0.1\linewidth]{imgs/patterns/6-4-2-5-8-0.eps}} \\
136785 & 642580\\
down & neutral/cross \\ \\
\end{tabular}
\begin{tabular}{c c}
\fbox{\includegraphics[width=0.1\linewidth]{imgs/patterns/0-1-4-6-7-3.eps}} &
\fbox{\includegraphics[width=0.1\linewidth]{imgs/patterns/8-4-1-2-5-7.eps}} \\
014673 & 841257 \\
left & right/kmove/cross \\
\end{tabular}
\end{center}
\subsection{PINs}
\label{fig:pins}
Note that filled circle is the start point, multiple circles on a number indicate multiple touches.
\bigskip
\begin{center}
\begin{tabular}{c c}
\fbox{\includegraphics[width=0.1\linewidth]{imgs/pins/1-3-2-8.eps}}&
\fbox{\includegraphics[width=0.1\linewidth]{imgs/pins/6-7-0-2.eps}}\\
1328 & 6702 \\
up/non-adj & down/kmove/cross\\\\
\end{tabular}
\begin{tabular}{c c}
\fbox{\includegraphics[width=0.1\linewidth]{imgs/pins/1-9-5-5.eps}}&
\fbox{\includegraphics[width=0.1\linewidth]{imgs/pins/7-2-7-2.eps}}\\
1955 & 7272\\
neutral/non-adj/repeats & left/kmoves/repeats\\\\
\end{tabular}
\begin{tabular}{c c}
\fbox{\includegraphics[width=0.1\linewidth]{imgs/pins/5-9-6-2.eps}}&
\fbox{\includegraphics[width=0.1\linewidth]{imgs/pins/1-5-3-5-2-5.eps}}\\
5962 & 152525\\
right & up/repat \\\\
\end{tabular}
\begin{tabular}{c c}
\fbox{\includegraphics[width=0.1\linewidth]{imgs/pins/4-5-8-0-9-0.eps}}&
\fbox{\includegraphics[width=0.1\linewidth]{imgs/pins/1-5-9-4-2-8.eps}}\\
458090 & 159428 \\
down/repeat & neutral/cross/non-adj\\\\
\end{tabular}
\begin{tabular}{c c}
\fbox{\includegraphics[width=0.1\linewidth]{imgs/pins/4-4-1-7-9-1.eps}}&
\fbox{\includegraphics[width=0.1\linewidth]{imgs/pins/3-6-6-7-9-2.eps}}\\
441791 & 366792 \\
left/kmove/repeat & right/repeat/kmove/cross\\
\end{tabular}
\end{center}
\subsection{Post-Hoc Participant Feedback}
One advantage of the live study is that the researchers can directly observe the
strategies of the participants and the relative difficulties encountered, as
well as via post hoc questions (the precise questions are found in the
Appendix~\ref{app:posthoc}). There is no direct comparison to the Aviv et al. prior work
here, but we believe that the strategies likely mirror those used by
participants in the video study, to some extent.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\linewidth]{imgs/pattern_crossing_examples_01.eps}
\hfill
\includegraphics[width=0.45\linewidth]{imgs/patterns_symbol_examples_02.eps}
\caption{Grid patterns with crossing and knightmove (4572) features that
challenged some observers, and patterns that were deemed more memorable by
some observers because they offered easy symbolic associations.}
\label{fig:knightmove_shapes}
\label{fig:shape_associations}
\end{figure}
The most commonly reported strategy for the observation task (n=16) was simply
focusing on memorizing the passcode as it appeared and then, after it was
completely entered, writing it down immediately without delay. Only three
participants reported strategies involving writing or physically mirroring the
input gesture while it was happening. Other participants (n=2) described
"chunking" PINs into larger numbers (e.g. "seventeen" versus "one-seven") in
their first languages (Farsi and Chinese) to make quick memorization
easier. Five participants mentioned that they watched the readout field in the
PIN conditions, while others preferred to watch only the finger gesture as it
was performed.
Participants mentioned several factors that could make PIN and grid passcodes
challenging to accurately record. These included grid pattern shapes that
crossed over themselves or contained knightmoves (n=11, e.g. 743521 and 4572,
Figure~\ref{fig:knightmove_shapes}), as well as both long physical jumps between
sequential PIN digits (n=3) and sequential digits physically close together
(n=7). Ten participants reported that viewing from the right was harder because
their view of the phone screen was partially blocked by the victim-proctor's
thumb in his right-handed grip, which is supported in the data, particularly for
NPAT results. Six participants also felt that glare from overhead lighting was
sometimes an issue.
Other passcode features and conditions were described as helpful by
observers. Four participants mentioned that it was easier to memorize shapes
that they could easily associate with a visual image, such as 136785 as a house,
or 842157 as a picnic table (Figure ~\ref{fig:shape_associations}).
Finally, multiple observations of the same passcode were commonly deemed helpful
for confirming or piecing together sequences, although one participant stated
that it was easier to do this if both observations were made from the same
side. This is supported by the quantitative data.
\section{Related Work}
\paragraph{Mobile authentication and observation attacks}
Threats such as shoulder-surfing attacks
have been well documented by researchers \cite{wiese2015pitfalls,aviv2017shoulder}. Studies
have been conducted examining experiences of users who had encountered
observation attacks~\cite{eiband2017understanding} where shoulder surfing was
found to be ``casual'' and ``opportunistic.'' Harbach et
al. \cite{harbach2014hardlock} found that participants only very rarely
reported shoulder surfing (0.3\% of 1134 sampled events) as an immediate high
risk threat when authenticating.
In order to minimize the risk associated with observation attacks, users are
known to modify their own usage behaviors when using a mobile device, hiding the
device from sight and performing mobile interactions in the pocket or bag, or
even shielding the screen \cite{abdolrahmani2016empirical}. Solutions also exist
to obscure screens from third parties \cite{deluca2014nowyouseeme}, to detect
the presence of shoulder surfers in a nearby vicinity \cite{ryu2017} or to
deceive onlookers from data being entered \cite{von2015swipin,
krombholz2016force}. Attacks have also been simulated by having observers watch video
footage of victims entering authentication sequences. Examples include \cite{khan2018CHI} where attacks took place from top and side views. A range of solutions have also been proposed to minimize the likelihood of shoulder-surfing when entering authentication
sequences \cite{ali2016developing}. However, as highlighted by Wiese and Roth
\cite{wiese2015pitfalls}, it can be difficult to compare the efficacy of these
solutions, as the ways in which these systems are studied varies. Furthermore,
the outcomes can be difficult to compare and interpret.
\paragraph{Evaluating resistance from shoulder surfing}
Many evaluation studies have focused on observing unlock screen interactions
where PINs and patterns are entered
\cite{schaub2012password,aviv2017shoulder,khan2018CHI}. Wiese and Roth
\cite{wiese2015pitfalls} suggest that conducting such studies are challenging
because real-world adversaries are not available for study and must be simulated in one way
or another. In contrast to live studies where participants and
actors/researchers perform tasks together in person, video simulations have been
used to identify susceptibility of on-screen
threats~\cite{sahami2012assessing,ali2016developing}. Video recordings offer
consistency when presented to multiple users \cite{wiese2015pitfalls}, and can
also be accessed independent of location. However, research indicates that that
the success of adversaries is lower when performing video observations compared to live
settings \cite{schaub2013exploring, wiese2015pitfalls}; we make a similar
observation here. Prior research also recommends that shoulder surfing attackers
should be allowed a number of observations~\cite{wiese2015pitfalls} as well as
viewing interactions from a range of
views~\cite{sahami2012assessing,aviv2017shoulder} and different properties of
passcodes~\cite{aviv2017shoulder}. Additionally, the hand position
\cite{schaub2012password} and interaction style when entering data into the
device \cite{aviv2017shoulder} should also be considered. We tested scenarios found to be significant in Aviv et al., following similar procedures.
\paragraph{Overview of Aviv et al.~\cite{aviv2017shoulder}}
Aviv et al. considered the lack of a baseline for comparing common unlock
authentication mechanisms under the threat of shoulder surfing. As a method of
creating such a baseline, the authors used a series of controlled video
simulations of a victim entering unlock authentications using several methods. These methods were PINs and
Android's graphical pattern unlock, with and without feedback lines
present. Additional factors were considered, including the angle of observation,
number of observations, the number of recreation attempts by the observer, the
hand posture of the victim, phone size, and spatial layout of the passcodes.
The methodology of that experiment was multi-factorial. Participants
were selected into one of a number of independent factors (phone type, passcode
choice, authentication type, hand posture) and then a set of randomized
dependent factors (passcodes, observation angles, number of views, and
attempts). For recruitment, the primary results were based off participants
on Amazon Mechanical Turk ($n=1173$) and participants recruited locally
($n=91$), with both groups completing a web survey whereby they viewed videos of
authentication and attempted to recreate the passcodes observed.
Using the results, the authors tested the following hypotheses (the {\bf -p}
indicates a prior work hypothesis):
\begin{itemize}
\item {\bf H1-p}: The type of unlock authentication, PIN pattern with lines,
patterns without lines, affects the shoulder surfing vulnerability.
\item {\bf H2-p}: Repeated viewing of user input increases the likelihood of a
shoulder surfing vulnerability.
\item {\bf H3-p}: Multiple attempts to recreate the input affects the likelihood
of a shoulder surfing vulnerability.
\item {\bf H4-p}: The angle of observations affects shoulder surfing vulnerability.
\item {\bf H5-p}: The properties of the unlock authentication, such as length and
visual features, affect shoulder surfing vulnerability.
\item {\bf H6-p}: The phone size affects shoulder surfing vulnerability.
\item {\bf H7-p}: The hand position used to hold and interact with a device
affects shoulder surfing vulnerability.
\end{itemize}
Of those hypotheses, {\bf H1-p}, {\bf H2-p}, {\bf H3-p}, {\bf H4-p}, and {\bf H6-p} were accepted, while {\bf H5-p} was partially accepted, and {\bf H7-p} was
rejected. The authors claim that the video studies, generally, can form a
reasonable replacement for live simulation, and that at the very least a video
study could provide a baseline for shoulder-surfing vulnerability.
\section{Results}
As the live simulation used a subset of the variables in the video study (see
prior section), we in turn performed comparisons on an appropriate subset of
the video study data. We used video data that met the following criteria:
one-handed/thumb-input on the Nexus 5 (red) phone, viewed from the left or right
angle, and a single recreation attempt. Additionally, we only included video
data that was collected with a screen resolution $>1800$ pixels, which was
identified as the most ideal viewing condition in the prior study
\cite{aviv2017shoulder}. With these reductions, we compared 720 shoulder surfing
attempts for the live simulation to a comparable 1,171 attempts in the video
study.
\subsection{Comparing Attack Rates Across Video and Live Studies}
\paragraph{Statistical Procedures}
As the results of the experiments for both the live and video study are
proportional, either the participant succeeded in recreating the passcode or did
not, we compare the results using a {\em proportionality test} for equality of
proportions, which follows a $\chi^2$ distribution. That is, we compare the attacker success rate for the video study to that of the live study
using the same conditions, reporting the $\chi^2$ statistic, the two-tailed $p$
value, and the 95\% confidence interval ($\delta_{95}$) for the difference
between proportions.
In the cases where $p\le0.05$, we can conclude that the live study was {\em not}
well modeled by the video study because the proportions of attacker success are
significantly different. Similarly if $p>0.05$ we cannot reject the null
hypothesis that the two proportions are the same and thus must conclude that the
proportions are more likely measuring the same effect. The confidence interval
reports the most likely range of difference between the attacker success rate
for the video and live results, but is only relevant when a significant
difference is found.
When comparing data across factors with greater than two conditions, we used a
$\chi^2$ test for goodness of fit to determine significant differences in attack
success rates. Post-hoc analysis is conducted using pairwise comparisons with a
Bonferonni correction.
Across tests, while the data is overlapping for some of the factors being
examined, we do not normalize/correct $p$ values as we are not attempting to
control for type-1 errors across {\em all} tests. Instead, we are performing
exploratory analysis and interested in determining if significant differences
may exist and from where they may arise. In post-hoc analysis, as described
above, we do correct $p$ values as appropriate as this occurs within a single
test with directly overlapping hypothesis.
\paragraph{Authentication Types (H1-r/H1-p)}
In the prior study, a key finding was that a statistical difference was
identified in attacker performance across authentication type. We can perform
the same tests by comparing vulnerability to shoulder surfing for the single
view conditions; see the first column of Table~\ref{tab:views}.
We first compare each of the authentications between the video and the live study,
irrespective of the authentication length. For patterns with feedback lines (termed: PAT) ($\chi^2=0.0,p=1$), there is
strong statistical similarity. However, for patterns without lines (termed: NPAT) ($\chi^2=4.54,p=0.03$) we do
see a significant difference between the live and video study, and an even more
prominent difference for PINs ($\chi^2=37.76,p=0.00$). Statistical differences for NPAT can be accounted for by an increase in the 4-length performance for attackers in the live setting (see
Table~\ref{tab:views}), and for PINs, we consistently see performance increases
for the live setting compared to the video setting. In this case, the success
rate for PINs in the video setting is 65/208=32.0\% compared to 135/216=62.5\%
for the live setting, an increase of 1.95x; however, the video study does
provide a baseline.
We can also compare authentication types within collection method, as related to
{\bf H1-p}. Using a three-way $\chi^2$ tests with pairwise comparisons,
there are statistically significant differences between each of the success rates
for each of the authentications for both the live ($\chi^2=24.8,p=0.00$) and
video ($\chi^2=133.4,p=0.00$) settings. The residuals suggest the leading cause
of this difference is the increased difficulty of shoulder surfing PINs, for
both the video and live setting, but post-hoc, pairwise-analysis (with
Bonferroni correction) suggest the benefits of removing feedback lines in NPAT is
not consistent across studies. While there are statistical differences between
PAT and NPAT in the video study, this effect disappears in the live study with
$p=.147$ (under the correction). {\em This provides further evidence that
removing traceback lines from pattern entry provides limited protection, and
perhaps less than what was previously considered~\cite{vzw2015easy}.}
Despite seeing a reduced benefit from NPAT as compared to PAT, we can {\bf
confirm} {\bf H1-p} in the live setting. The authentication type has an
impact on shoulder surfing performance as evident in the differences in attacker
success rate for different authentication types, particularly for PINs.
\tableangles{}
\paragraph{Repeated Viewings (H1-r/H2-p)}
An important result of the video study was the finding that repeated viewings
have significant impact on attacker performance ({\bf H2-p}). By expanding our
view of Table~\ref{tab:views} to the {\em Two-Same} and {\em Two-Different}
column, we can test for similar effects resulting from repeated viewings. As
before, we observe the most consistency in the PAT and NPAT settings for the
live and video study, and strong differences in the PIN setting. However, where
we do see significant difference the confidence interval suggest that the video
study does provide a baseline to the live setting.
We can further directly measure the impact of multiple viewings by performing
within collection method $\chi^2$ tests across viewing methods. For PAT, no
effect could be identified for multiple views in both the video and live
settings. There is an effect for NPAT in the live ($\chi^2=12.0,p<0.01$) but not
in the video setting ($\chi^2=5.1,p=0.08$). Post-hoc analysis revealed that, for
NPAT in the live setting, having the same viewing angles twice compared to a
single viewing angle or two difference angles drives this difference ($p=0.03$,
corrected), {\em suggesting that two-different viewing conditions for NPAT is
most advantageous to an attacker.}
The case is similar for PINs. In the live setting, a statistically significant
difference occurs for conditions of repeated views ($\chi^2=23.1,p=0.00$).
However, this was not the case for the video setting ($\chi^2=4.1,p=0.14$). {\em
Post-hoc analysis suggests that gaining any repeated viewing, the same angle
twice or two different, benefits the attacker significantly in the live
setting.} The lack of significance for the video setting may be due to using
this particular subset of video data, but we conjecture that it more likely
reflects the high difficulty of shoulder surfing PINs, generally, which was
further exacerbated by the video observation setting without stereo vision.
Overall, we can {\bf confirm} {\bf H2-p} in the live setting, that repeated
viewings have an impact on performance. Where there were previous significant
differences in the video study, these persisted in the live setting, except for
NPAT. While there is consistency in viewing the same angle twice, observing the
entry from multiple angles seems to play a larger role in the live setting
compared to the video setting. However, the larger hypothesis that repeated
views impacts performance of shoulder surfing is confirmed.
\paragraph{Observation Angle (H1-r/H4-p)}
To assess the impact of observation angle, we use only single-view conditions so as
not to conflate the results with the impact of multiple observations. These results are
presented in Table~\ref{tab:angles} with pairwise comparisons between the live
and video study for different passcode lengths.
While we continue to see significant differences for PIN and a lack thereof for
PAT, we see significant improvements in the live setting for NPAT viewed from the
right angle. We conjecture that this improved attacker performance relates to
being able to stereoscopically determine touch locations that are more
challenging to see from the same angle via video simulation. However, depth of touch
events continue to be more challenging when viewed from the left angle. The
difference between the observations angles here may also explain other
statistical differences in the previously presented results for NPAT.
However, overall, we {\em do not} see significant differences when comparing
within a collection method and authentication when comparing left vs. right
angle. This is in conflict with prior work; however, recall that the top
observation angle was removed and the two near and far angles were reduced to a
single side angle (L or R). As the two comparable subsets are consistent, we can
{\bf confirm} that under {\bf H4-p} the live settings are well predicted by a
comparable subset of video data.
\tablecodes{}
\paragraph{Passcode Properties (H1-r/H5-p)}
In Table~\ref{tab:codes}, again using single view data, a direct comparison
between each of the passcodes used in the study is displayed, with findings from
proportionality tests between the live and video setting. We find that
no significant differences exist for the PAT and NPAT codes, and only
three of the PIN codes show differences. These include the following PINs: 5962, 159428, and 366792 with the live setting attacker performance being significantly better in each case. The spatial
properties of these codes (see Appendix~\ref{fig:pins}) does not suggest that a
single factor played a role. Although both 5962 and 3669722 are both right
shifted PINs, there are too many other features at play to draw conclusions.
We can perform a within-collection method analysis across the passcodes using a
$\chi^2$ test, and we find that significant differences exist for the
attacker success rate within both the live and video study, for all
authentication types. However, post-hoc analysis suggest that none of the NPAT
pairwise comparisons are significant, and only one set of PAT pairwise comparisons are
significant (743521 vs. 3157) --- 743521 was the most difficult of the patterns
to shoulder surf. For PINs in post-hoc analysis, again 159428 and 366792 have
significant comparisons, particularly with PINs 7272 and 1955, which were two of
the easiest PINs to shoulder surf in comparison to 159428 and 366792, two of the
most difficult to shoulder surf.
Finally, we can compare the impacts of length. For PAT, we do not see
significant differences between success rate for 4- vs. 6-length patterns
($\chi^2=2.9,p=0.09$), but we do for the video study
($\chi^2=12.83,p<0.001$). We find the reverse for NPAT, where there is a
significant difference in length for the live setting ($\chi^2=5.7,p=0.02$) and
not for the video study ($\chi^2=3.64,p=0.06$). Finally, we see significant
differences for PIN for both live ($\chi^2=28.9,p=0.00$) and video
($\chi^2=27.6,p=0.00$). This suggests that, yes, the length of the passcode can
have an impact, {\bf confirming} {\bf H1-r} for the {\bf H5-p} condition;
however, other properties of the passcode were not significant, but were so in
similar ways between the two studies under the subset being evaluated.
\paragraph{Hypothesis {\bf H1-r}}
Based on the results presented previously, in each case we are able to find
confirmation of each of the previous hypotheses, although, we also find that
PINs are the least consistent. This suggests that researchers should be more
skeptical of results related to PIN based authentication in the video setting. In particular, the true values may be much higher. Additionally, we find
strong evidence that the differences between PAT and NPAT may be greatly
dimensioned (although still different) in the live setting.
\paragraph{Hypothesis {\bf H2-r}}
We can confirm {\bf H2-r} that video based recreations do provide a baseline for
live simulation. Observe that in all cases where there is significant
differences between a video and live measurement in
Tables~\ref{tab:views},~\ref{tab:angles}, and~\ref{tab:codes}, the confidence
interval suggests that the live setting has {\em higher} proportionality than
that of video setting. In essence, yes, the video study provides a baseline, but
the baseline may be much lower than one may expect, as much as 1.7x.
\section{Implications}
\paragraph{Importance of evaluating in appropriate settings}
Researchers often favor performing studies examining observational attacks with video-based stimuli presented to participants. While likely simpler to coordinate and easier to control compared to studies conducted in live settings, video studies can lack realism and are considered a methodological substitute only when necessary \cite{wiese2015pitfalls}. While findings from video studies can be helpful to determine attack rate, our findings suggest that researchers evaluating authentication interfaces should be aware that there is no substitute for testing in live settings, as the video baseline may greatly underestimate the threat of an attacker. The video baseline may serve as a method for a preliminary assessment.
\paragraph{Factors which should be taken into account when performing observational attack studies}
While factors such as authentication type and repeated views can impact attack rate, as evidenced through our study, other factors are worthy of further investigation. Examples include examination of the impact of observational angle and spatial properties of passcodes and device screen sizes. While significant differences in some of these factors could not always be detected, subjective feedback gathered from participant observers suggested that these factors could make a difference to attacker success. Examining these in more detail, alongside gathering subjective data for purposes of identifying reasoning, is suggested to researchers, as these may play a greater role than once thought.
\paragraph{Care in selection of passcode}
Our results suggest that specific types and properties of passcode may be more susceptible to observational attack, as identified through the comparison with live settings. As a result, users should be aware that removing the feedback lines from pattern unlock interfaces may not provide the security benefits that users expect. Secondly, PINs are more susceptible to attack than previously identified by researchers performing video-based studies. This is also supported in our qualitative feedback where participants noted that PINs with larger jumps were harder to attack, and for PAT/NPAT, those that are less ``shape like'' (e.g. resembling a house-like shape) are harder for participants.
\paragraph{Need for training}
As our findings have highlighted that observational attacks are more successful under specific conditions, security training for mobile device users can be developed to better understand the nature of observational threats, encouraging them to make better security choices. Some users may need to better understand what methods and parameters would provide resilience against high-probability multiple-view observation attacks mounted by ``insider threats'' \cite{wiese2016see}. Others might want those authentication factors tilted towards greater ease of use if they perceive less risk of observational attack. Better informing these choices could come in the form of interactive guidance/prompting when setting-up devices.
|
\section{Introduction}
Ising-like models and their countless variations have been used throughout the last decades to describe data or model systems with the most diverse nature \cite{bury2013market,Bouchaud2013,tanaka1977model,cocco2017functional,kadirvelu2017inferring} and to increase our understanding of how natural, artificial, social and economic systems work.\\
On the one hand these models, studied in their original physical formulation, can be manipulated to generate a wide range of behaviours mimicking the features of these systems \cite{Bouchaud2013,bornholdt2001expectation}, and use a deductive approach to explain the stylized properties of data we observe in the real world. On the other hand one can use these models in the fashion of descriptive and forecasting models \cite{bury2013market,cocco2017functional,ibuki2013statistical,kadirvelu2017inferring}, by using
Maximum Likelihood (ML) and Maximum A Posteriori (MAP) techniques to fit the model to the data, inductively working towards an explanation of the observations. This is typically regarded to as the inverse formulation of the model, while the former is the direct formulation.\\
A model of this family has recently been revamped for time-series data, the non-equilibrium or Kinetic Ising Model \cite{derrida1987exactly,crisanti1988dynamics}, describing a set of binary units - named ``spins" in the physics literature - that influence each other through time. The simplicity of the model makes it extremely flexible in the kinds of systems it can represent, ranging from networks of neurons in the brain \cite{capone2015inferring} all the way to traders in a financial market \cite{bornholdt2001expectation,SornetteReview}. Recent work on the inverse Kinetic Ising Model has led to the development of exact \cite{sakellariou2013inverse} and Mean Field (MF) \cite{roudi2011dynamical} techniques for the inference of the parameters, and the latter have been used to work with partially observed systems linking to the realm of (Semi-) Restricted Boltzmann Machines \cite{dunn2013learning}. \\
This latest stream of literature sparked our interest for the model applied to time series of financial data at high frequency, where we typically encounter problems related to the lack of homogeneously frequent and synchronized observations \cite{ait2010high,buccheri2017score,Corsi2012}.
The literature on Kinetic Ising Model has previously considered mainly the inference problem in the presence of hidden nodes \cite{dunn2013learning}, i.e. part of the spins are {\it never} observed, but it is known that they exist and interact with the visible nodes (i.e. spins). This setting is of particular interest in neuroscience where an experiment typically monitors the firing activity of a subset of neurons. In other domains, such as in economics, finance, and social sciences, another type of missing data is often present, namely the case where even for the visible agents (nodes), observations are missing a significant fraction of the times. Moreover in these cases there is a strong heterogeneity of the frequency of observations, i.e. some nodes are frequently observed while other are rarely observed. There are different sources for this lack of data: in some cases, it might be due to the fact the observation is costly for the experimenter, whereas in other cases it is intrinsic to the given problem. Consider, for example, the problem of inferring the opinion of investors from their trading activity. When an investor buys (sells) it is reasonable to assume that she believes the price will increase (decrease), but in many circumstances the investor will not trade leading to missing observations for her belief. Using a suitable inference model, as the one proposed in this paper, it is possible to estimate her belief from the inferred structure of interaction among investors and the observed state of the set of visible ones. We will also include external fields (for example the market price in the previous example) that can influence spins (investors' opinion).
Moving our steps from the work by Dunn et al. \cite{dunn2013learning}, we extend the formulation of the inference procedure to cases where the missing observations are unevenly cross-sectionally distributed, meaning that time series are sampled at a constant rate and whenever no observations are found between two timestamps a missing value is recorded. The result is an algorithm closely related to an Expectation-Maximization (EM) method \cite{expmax1977}, iteratively alternating a step of log-likelihood gradient ascent \cite{Nesterov2008} and the self-consistent resolution of TAP equations \cite{roudi2011dynamical}, that gives as output both a coupling matrix and a maximum-likelihood estimate of the missing values. \\
To evaluate the algorithm performance we devise a series of tests stressing on different characteristics of the input, simulating synthetic datasets with several regimes of intrinsic noise, observation frequency, heterogeneity of variables and model misspecification. We thus define some performance standards that can be expected given the quality of data fed to the method, giving an overview of how flexible the approach is.\\
The paper is organized as follows: in Section 2 we define the considered Kinetic Ising Model, we explain the inference method in detail and describe the approximations needed to make the algorithm converge in feasible time; in Section 3 we present results on synthetic data and give an overview of the performance that can be expected with different data specifications; Section 4 concludes the article.
\section{Solving the Inverse Problem with missing values}
The Kinetic Ising Model (or non-equilibrium Ising Model) \cite{derrida1987exactly} is defined on a set of spins $y \in \lbrace -1, +1 \rbrace^N$, whose dynamics is described by the transition probability mass function
\begin{multline}
p[y(t+1) \vert y(t) ] = Z^{-1}(t) \exp \Bigg[\sum_{\langle i,j \rangle} y_i (t+1) J_{ij}y_j(t) +\\
+ \sum_i y_i(t+1) h_i\Bigg]
\end{multline}
where $\langle i,j \rangle$ is a sum over neighbouring pairs on an underlying network, $J_{ij}$ are independent and identically distributed couplings, $h$ is the vector of spin-specific fields and $Z(t)$ is a normalizing constant also known as the partition function.\\
In our treatment of the problem we will adopt a Mean Field (MF) approximation, which relies on the assumption that the dynamics of a spin $i$ depends only on an effective field locally ``sensed" by the spin rather than on the sum of the single specific interactions with others. The result of this picture is that the topology of the underlying network is considered irrelevant and assumed fully connected - although the goal of the inference would be the reconstruction of the network nonetheless - thus the sum on neighbours is substituted by a sum on all the other spins. This recasts the transition probability into the following form
\begin{equation}
p[y(t+1) \vert y(t) ] = Z^{-1}(t) \exp \left[\sum_{i=1}^N y_i (t+1) \tilde{g_i}(t)\right]
\end{equation}
where $\tilde{g_i}(t) = \sum_{j=1}^N J_{ij}y_j(t) + h_i$ is the local effective field of spin $i$ and $J$ is now a square and fully asymmetric matrix with normally distributed entries $J_{ij} \sim \mathcal{N}(0, J_1^2/N)$, where the assumption on the distribution and the scaling of the variance with $N^{-1}$ will be necessary in the forthcoming calculations.\\
Consider observing only a fraction $M(t)/N$ of spins at each time step, and define $G(t)$ as the $M(t) \times N$ matrix mapping the configuration $y(t)$ into the observed vector $s(t) \in \lbrace -1,1 \rbrace^{M(t)}$. Also define $F(t)$ as the $(N-M(t)) \times N$ matrix mapping $y(t)$ into the unobserved spins vector $\sigma(t) \in \lbrace -1,1 \rbrace^{N-M(t)}$. We require that both matrices are right-invertible at all $t$, thus they must have full rank, that implies that observations are not linear combinations of the underlying variables as our interest is in a partially observed system rather than a low-dimensional observation of a high-dimensional system. For the sake of simplicity we assume that the entries are either $0$ or $1$, meaning observation is not noisy or distorted and the right-inverse matrices will coincide with the transpose. \\
In the upcoming calculations we will use some simplifying custom notation in order to reduce what can be some cumbersome equations. We will thus denote $\sideset{}{^\prime}\sum_i$ the sum over indices $i$ at time $t+1$, while the regular $\sideset{}{}\sum_i$ indicates a sum over indices $i$ at time $t$ and $\sideset{}{^-}\sum_i$ a sum at time $t-1$. Accordingly, we will indicate with $s_i$ spin $i$ at time $t$, with $s^-_i$ at time $t-1$ and with $s_i^\prime$ at time $t+1$, and the same applies for $g$, $\sigma$ and any other variable. Also indices $i,j,k,l$ are used for observed variables, whereas indices $a,b,c,d$ will identify unobserved variables.\\
In this notation, the probability mass function is rewritten as
\begin{align}
p[\lbrace s^\prime, \sigma^\prime \rbrace \vert \lbrace s, \sigma \rbrace ] = Z^{-1} \exp \left[ \sideset{}{^\prime}\sum_{i} s_i^\prime g_i^\prime + \sideset{}{^\prime}\sum_{a} \sigma_a^\prime g_a^\prime \right] \label{eq:pmf}
\end{align}
Defining the matrices $J^{oo}(t+1) = G(t+1)JG^T(t)$, $J^{oh}(t+1) = G(t+1)JF^T(t)$, $J^{ho}(t+1) = F(t+1)JG^T(t)$ and $J^{hh}(t+1) = F(t+1)JF^T(t)$ the local fields are
\begin{align}
g_i = \sum_{j} J_{ij}^{oo} s_j^- + \sum_{b}J^{oh}_{ib} \sigma_b^- +h_i \nonumber \\
g_a = \sum_{j} J_{aj}^{ho} s_j^- + \sum_{b}J^{hh}_{ab} \sigma_b^- +h_a \label{eq:fielddef}
\end{align}
and the partition function or normalization constant is
\begin{equation*}
Z = \sideset{}{^\prime}\prod_{i,a} 2 \cosh (g_i^\prime) 2 \cosh (g_a^\prime)
\end{equation*}
The ultimate purpose of this work is to devise a method to obtain Maximum Likelihood Estimates (MLE) for the parameters $J,h$ and the unobserved spins $\sigma$. The likelihood function is just the product through time of the independent transition probabilities expressed in Eq. \ref{eq:pmf}, taking the trace over the missing values
\begin{equation}
p[\lbrace s \rbrace ] = \mathrm{Tr}_{\sigma} \prod_t p[\lbrace s^\prime, \sigma^\prime \rbrace \vert \lbrace s, \sigma \rbrace ] \label{eq:likelihood}
\end{equation}
To solve the problem, our approach is closely related to the one developed by Dunn et al. \cite{dunn2013learning}, where the authors investigate on a system where only a subset of spins is observable. The extension to our case is presented below.
The trace of Eq. \ref{eq:likelihood} is non-trivial to be done. However the Martin-Siggia-Rose path integral formulation \cite{msr1973} allows to decouple spins and perform the trace at the cost of computing a high dimensional integral. Define the functional
\begin{equation}
\mathcal{L}[\psi] = \log \mathrm{Tr}_{\sigma} \prod_t \exp \left[ \sum_a \psi_a \sigma_a \right] p[\lbrace s^\prime, \sigma^\prime \rbrace \vert \lbrace s, \sigma \rbrace ]
\end{equation}
Notice that this is equivalent to the log-likelihood if $\psi_a(t) = 0$ $\forall a,t$, thus the goal of the calculation will be to efficiently maximise $\mathcal{L}[\psi]$ in the $J, h$ coordinates considering the limit when $\psi \rightarrow 0$. As will become clear in the next steps, the introduction of these so-called ``auxiliary fields" is necessary to switch from the unknown values $\sigma$ to their posterior expectations $m$, thus smoothing the log-likelihood function eliminating unknown binary variables from its formula.
Call
\begin{align*}
Q[s, \sigma] = \sum_t \sum_{i}&s_i g_i + \sum_t \sum_{a} \sigma_a g_a + \\
- \sum_t \sum_{i}&\log 2\cosh(g_i) - \sum_t \sum_{a} \log 2 \cosh(g_a) \\
\Delta = \sum_t \sum_{i}& i \hat{g}_i \left[ g_i - \sum_{j} J^{oo}_{ij}s_j^- - \sum_{b} J^{oh}_{ib} \sigma_b^- - h_i \right] + \\
+ \sum_t \sum_{a}& i \hat{g}_a \left[ g_a - \sum_{j} J^{ho}_{aj} s_j^- - \sum_{b} J^{hh}_{ab} \sigma_b^- - h_a \right]
\end{align*}
where $e^\Delta$, integrated over the $\hat{g}$s is the integral representation of the Dirac delta function. Then one obtains
\begin{equation}\label{lagrangian}
\mathcal{L}[\psi] = \log \int \mathcal{DG} \exp [\Phi]
\end{equation}
where $\mathcal{G} = \lbrace g_i, g_a, \hat{g}_i, \hat{g}_a \rbrace_t$ and
\begin{equation}
\Phi = \log \mathrm{Tr}_{\sigma} \exp \left[Q + \Delta + \sum_t \sum_{a} \psi_a \sigma_a \right]
\end{equation}
Now the trace can be easily computed since the introduction of the delta function has decoupled the $\sigma$s by fixing the value of the local fields $g$.
As mentioned, the cost is computing the integral of Eq. \ref{lagrangian}, which can be solved via the saddle-point approximation, where the saddle-point is obtained by the extremization of $\Phi$ with respect to the coordinates in $\mathcal{G}$.\\
The missing part of the puzzle is the posterior mean $\mathbb{E}\left[\sigma_a (t) \right]$, for which $\mathcal{L}$ acts as the generating functional
\begin{equation*}\mathbb{E}\left[\sigma_a (t)\right]= m_a(t) = \lim_{\psi_a(t) \rightarrow 0} \mu_a(t) = \lim_{\psi_a(t) \rightarrow 0}\frac{\partial \mathcal{L}}{\partial \psi_a(t)}
\end{equation*}
where the expectation is performed under the posterior measure $p[\lbrace \sigma \rbrace \vert \lbrace s, J, h \rbrace]$. \\
This zero-order approximation is rather rough, nonetheless the saddle-point method can be solved at higher orders of approximation.\\
The second-order (\textit{i.e.} Gaussian) correction to the saddle point solution of the integral in Eq. \ref{lagrangian} is
\begin{equation*}
\delta \mathcal{L} = - \frac{1}{2}\log \det [\nabla^2_{\mathcal{G}} \mathcal{L}]
\end{equation*}
where $\nabla^2_{\mathcal{G}} \mathcal{L}$ is the Hessian matrix in the $\mathcal{G}$ space of $\mathcal{L}$ evaluated at the saddle point. The resulting structure of the matrix, shown in the Supplementary Material for the sake of space, is sparse and almost block-diagonal.
We are interested in the determinant, and in particular its logarithm. Dividing the Hessian in the matrices $\alpha$ containing block-diagonal elements and $\beta$ containing the rest, we find
\begin{multline}
\log\det(\alpha+\beta)=\log\det(\alpha) + \log\det[\mathbb{I}+\alpha^{-1}\beta] =\\
= \log\det(\alpha) + \mathrm{Tr}\log[\mathbb{I} + \alpha^{-1}\beta] \approx \\
\approx \log\det(\alpha) + \mathrm{Tr}[\alpha^{-1}\beta] + \frac{1}{2} \mathrm{Tr}\lbrace [\alpha^{-1} \beta]^2 \rbrace + ... \label{eq:approx}
\end{multline}
Given that $\alpha$ is block-diagonal, so will be $\alpha^{-1}$, then $\mathrm{Tr}[\alpha^{-1}\beta]=0$ and we ignore higher order terms assuming the off-diagonal part of the Hessian matrix is small compared to the diagonal one. In our initial assumption, the couplings $J_{ij}$ are Gaussian random variables with mean of order $1/N$ and variance of order $J^2_1/N$, which means $\log \det (\alpha)$ is quadratic in $J_1$ (see Supplementary Material). The determinant now can be computed and a weak couplings expansion (i.e. $J_1 \rightarrow 0$) can be made to eliminate the logarithm, leading to the final approximate form of the correction
\begin{align*}
\delta \mathcal{L} \approx & -\frac{1}{2} \sum_t \sideset{}{'}\sum_{i} \left[ \left(1-\tanh^2(g_i^\prime) \right) \sum_{b} \left[J^{oh \prime}_{ib}\right]^2(1-\mu_b^2)\right] + \\
&- \frac{1}{2} \sum_t \sideset{}{'}\sum_{a} \left[ \left(\mu_a^{\prime \, 2} - \tanh^2(g_a^\prime) \right) \sum_{b} \left[ J^{hh \prime}_{ab} \right]^2 (1-\mu_b^2)\right]
\end{align*}
Given the new form of $\mathcal{L}_1 = \mathcal{L}_0 + \delta \mathcal{L}$, we need to recalculate the self-consistency relation for $m_a(t)$ and the learning rule for $J$. As for $m_a(t)$, we can easily see that it is going to coincide with $m_a(t) = \lim_{\psi_a(t) \rightarrow 0} \mu_a(t) + l_a(t)$, where
\begin{equation}
l_a(t) = \frac{\partial( \delta \mathcal{L})}{ \partial \psi_a(t)}
\label{eq:corrselfcon}
\end{equation}
Implementing the MSR method has introduced an explicit dependence of the $\mathcal{L}$ functional from the auxiliary fields $\hat{g}$ and $\psi$, which however make little sense in terms of the model itself. Now that we have solved the integral at the saddle-point and in its immediate neighbourhood the auxiliary fields can be absorbed back into the original variables by performing a Legendre transform of $\mathcal{L}$, exploiting the fact that $\mathcal{L}$ is convex and that we would rather have it depend on the conjugate field of $\psi$, that is $\mu$. The transform is
\begin{equation}
\Gamma [\mu] = \mathcal{L} - \sum_t \sum_{a} \psi_a(t) \mu_a(t) \; s.t. \; -\psi_a(t) = \frac{\partial \Gamma[\mu]}{\partial \mu_a(t)}
\end{equation}
and so we can adopt $\Gamma$ as the functional to be maximised in the learning process instead. At zero-order, this is easily found to be
\begin{multline}
\Gamma_0 [\mu] = \sum_t \Bigg[ \sideset{}{'}\sum_{i} \left[ s_i^\prime g_i^{0 \, \prime} - \log 2\cosh (g_i^{0 \, \prime}) \right] + \\ + \sideset{}{'}\sum_{a} \left[ \mu_a^\prime g_a^{0 \, \prime} - \log 2 \cosh(g_a^{0 \, \prime}) \right] + \sum_{a} S[\mu_a] \Bigg]
\end{multline}
where $S[x] = -\frac{1+x}{2} \log (\frac{1+x}{2}) - \frac{1-x}{2} \log(\frac{1-x}{2})$ is the entropy of an uncoupled spin with magnetization $x$. It is relevant to mention that so far the functional is expressed in terms of $\mu$, while we have already highlighted that after the Gaussian correction a new term $l$ is introduced in the formula for $m$. However, since we are restricting to second order in $J$, the terms containing $l$ in $\Gamma$ are all of superior order and are thus negligible in this approximation, then $\Gamma_0[m] \approx \Gamma_0[\mu]\vert_{\mu = m}$. Performing the exact same steps on the correction term $\delta \mathcal{L}$ one finds the corrected functional
\begin{equation*}
\Gamma_1[m] = \Gamma_0 [m] + \delta \mathcal{L}[m]
\end{equation*}
$\Gamma_1$ is the functional to be optimized through an Expectation-Maximization-like algorithm, recursively computing the self-consistent magnetizations $m$ given $J,h$ and then climbing the gradient $\nabla_{J,h} \Gamma_1$ to obtain a new $J$ matrix and $h$ vector. \\
Once the log-likelihood is maximized and the final iteration of the expectation part of the algorithm is finished, the result is a Maximum Likelihood Estimate of the couplings as well as a Maximum A Posteriori estimate of the hidden spins $\sigma$, given by $\hat{\sigma} (t) = \mathrm{sign} (m_t)$.\\
Summarizing, the procedure is the following:
\begin{tcolorbox}[title=\textbf{Algorithm},colbacktitle=white,coltitle=black]
\begin{itemize}
\item Initialize $J$, $h$, $m(t)$
\item Until convergence is reached
\begin{itemize}
\item compute the self-consistent magnetizations $m(t)$
\item compute the gradient $\nabla_{J,h} \Gamma_1$
\item apply Gradient Ascent step, in our case Nesterov's II method proximal gradient ascent with backtracking line search
\end{itemize}
\item Possibly involve LASSO $\ell_1$-norm regularization or pruning techniques to obtain a sparse model.
\end{itemize}
\end{tcolorbox}
\section{Tests on synthetic data}
We perform a series of tests on the algorithm in order to assess its performance in several diverse conditions of data availability. We particularly focus on how we select the observed spins and on the structure of the coupling matrix $J$ in the data generating model. To construct the $G(t)$ and $F(t)$ matrices, we assign to each spin a probability $p_i$ of being observed, meaning that $y_i(t)$ is observed with probability $p_i$ for all $t$.\\
We explore how the performance of the inference depends on the following model specifications:
\begin{enumerate}
\item[0.] The average observation frequency, taking the Bernoulli probabilities $p_i = p$, $\forall i=1,\ldots, N$;
\item The heterogeneity of the Bernoulli probabilities $p_i$, which we choose to be distributed according to a Beta distribution $B(a(K), b(K))$ with given mean $K$ and shape parameters $a$ and $b$;
\item The scale $J_1$ of the $J$ entries, which are distributed as $J_{ij} \sim \mathcal{N}(0,J_1^2/N)$;
\item The structure of the $J$ matrix, specifically whether the underlying network is fully connected or an Erd\H{o}s-R\'{e}nyi random network of varying density, adopting either the LASSO $\ell_1$ regularization \cite{tibshirani1996regression} or the decimation procedure \cite{decelle2015inference} to select the links;
\item The asymmetry of the $J$ matrix. One of the key assumptions in the calculation is that $J_{ij} \neq J_{ji}$ and that they are independent and identically distributed, and we investigate how far one can violate it up to the case of a symmetric $J$ matrix;
\item The dependency on the length of the time series relative to the number of units involved, $T/N$, to check the estimate asymptotic efficiency.
\end{enumerate}
In Test 0 we study the performance of the algorithm in a very simple setting of missing information, where each variable has the same probability of being observed and the generating model is a fully-connected Kinetic Ising model. This is intended to study the effect the average amount of missing information in the sample has on the inference, without considering the possibility of having heterogeneous types of nodes. In this setting we also introduce a procedure we call Recursive E-M: by properly iterating the algorithm multiple times it allows to boost data artificially thus achieving good performances even when the fraction of missing values is particularly high. \\
In Test 1 we explore the possibility that spins have heterogeneous observational properties. We sample the $\lbrace p_i\rbrace$ from a Beta distribution varying parameters to probe different levels of heterogeneity. The Beta distribution allows to range from a sharply peaked unimodal distribution to a sharply peaked bimodal distribution tuning the shape parameters $\alpha$ and $\beta$, while keeping the mean $K$ constant: the former case is a situation of perfect homogeneity in the frequency of observations calling back to Test 0, while the latter is the extreme heterogeneity of having some units that are (almost) always hidden while the others are (almost) always observed. We select some intermediate cases to characterize how heterogeneity in observation frequency affects the identification of the model parameters. \\
Test 2 aims at assessing whether there is a minimal interaction strength to have the inferential process converging and how the approximations necessary to develop the method impact the accuracy of the inference. Indeed while $J_1$ in the physical model is proportional to the ratio between the strength of the magnetic coupling interaction and the temperature at which the system is observed, from a modelling perspective it is inversely proportional to the impact of the noise on the dynamics. Given the approximation of Eq. \ref{eq:approx}, if $J_1$ gets too large, the precision with which the parameters are identified should get worse. We thus expect to find an optimal region for the inference to be accurate, bounded from below by an identifiability threshold and from above by the limit of validity of the expansion.\\
In Test 3 we pursue the goal of making the methodology useful for real world scenarios, where it is highly unlikely that all spins interact among themselves and the underlying network is probably sparse. We compare the performance of two well established techniques, the LASSO $\ell_1$ regularization and the decimation procedure, and explore how these two methods perform paired with our algorithm by simulating data on a set of Erd\H{o}s-R\'{e}nyi random networks with different densities. \\
In a similar spirit, in Test 4 we study how the i.i.d. assumption made in Eq. \ref{eq:approx} affects the performance in situations where coupling coefficients are pairwise correlated or even symmetric, a condition we envision to be more realistic in social and economic environments \cite{squartini2013reciprocity}. We vary the correlation parameter $\mathrm{Cor}(J_{ij}, J_{ji})=\rho$ for $i\neq j$ between 0 and 1, with the symmetric case being also of special interest because the model transforms into a dynamical form of the Sherrington-Kirkpatrick model, thus connecting to the extensive literature on the topic.\\
Finally, a sanity check is made in Test 5 by looking at the dependency of performance metrics on the ratio $T/N$, that is the ratio between the number of observations and the number of spins, to characterize the convergence rate of the estimator towards the true value and its consistency. \\
We will test the algorithm and evaluate the performance using mainly two metrics, one relative to the reconstruction of the couplings and one to the reconstruction of missing values:
\begin{enumerate}
\item The Root Mean Square Error (RMSE) on the elements of the matrix $J$, $\mathrm{RMSE} = \sqrt{\langle (\hat{J}_{ij} - J_{ij})^2 \rangle_{ij}}$, suitably rescaled when comparing experiments with different $J_1$;
\item The ``Reconstruction Efficiency" (RE), namely the fraction of spins that are correctly guessed among the hidden ones averaged throughout the time series, or $\mathrm{RE} = \langle \frac{1}{N-M(t)}\sum_a\delta_{\hat{\sigma}_a (t), \sigma_a (t)} \rangle_t $
\end{enumerate}
\subsection{Test 0: dependency on a homogeneous $p_i$}
\begin{figure}[h]
\includegraphics[width=1\linewidth]{figure/Recursive/Jbias_wide.pdf}
\includegraphics[width=1\linewidth]{figure/Recursive/JMSEvP_rec_wide.pdf}
\includegraphics[width=1\linewidth]{figure/Recursive/CSvP_rec_wide.pdf}
\caption{(top) Angular coefficient of the linear fit $\hat{J}_{ij} = a J_{ij} + c$ before and after R-EM varying the average observation density $p$; (middle) Root Mean Squared Error on the couplings; (bottom) Reconstruction Efficiency.}\label{fig:test0}
\end{figure}
The algorithm is outstandingly resilient to cases with few observations available. We simulate a system of $N=100$ spins, for $T=10000$ time steps, with $J_{ij} \overset{iid}{\sim} \mathcal{N}(0,1/N)$ lying on a fully connected network and we give a probability of observation to each variable $p_i=p$, with $p$ ranging from $0.1$ to $0.9$. As can be seen from the top panel of Figure \ref{fig:test0}, showing the linear regression coefficient $a$ of $\hat{J}_{ij} = a J_{ij} + c$, with one iteration of the method we get a very reliable result for the couplings for $p \geq 0.8$, although below this value the lack of data reduces the quality of the estimation and moves the estimates towards $0$. To overcome this issue, we propose the aforementioned R-EM procedure as a further enhancement of our algorithm: once a maximum of the likelihood has been reached, a fraction of hidden spins is substituted with their maximum likelihood estimates $\hat{\sigma}_a = \mathrm{sign} (m_a)$ and the inference is run again on the new, artificially boosted data. Since $m$ is proportional to the probability of the spin being up, we choose the missing values to be substituted at every $t$ as the ones with the most polarized magnetization, \textit{i.e.} for which $m$ is closer to $\pm 1$. This artificial boosting on the data shows promising results since with a few recursions the performance is noticeably better even in cases with severe lack of observations, as is also reflected in the middle and bottom panels of Figure \ref{fig:test0}. We defer a more rigorous treatment of this recursive method to future work, while still proposing it here as we find it surprisingly accurate.\\
The bottom panel of Figure \ref{fig:test0} shows the Reconstruction Efficiency, which gets worse almost linearly as the number of observations decreases and on which the R-EM has a smaller effect, albeit still being a clear improvement. It is evident from all panels that when a large fraction of data is missing ($p \leq 0.2$) the inference fails to identify any of the parameters and the model is no better than a coin flip at reconstructing configurations.
\subsection{Test 1: heterogeneous $p_i$}
In Test 1 we want to highlight how our model is a generalization of the one studied extensively by Dunn et al. \cite{dunn2013learning} and to characterize the impact of heterogeneity on the inference performance. To give a better comparison with the aforementioned paper, we realize simulations morphing from our initial specification of $p_i = p \; \forall i$, studied in Test 0, to a case very close to the one of Dunn et al. where $p_i \in \lbrace 0, 1 \rbrace$, that is some variables are always observed and some are always hidden. We choose to take the probabilities distributed according to a Beta distribution, $p_i \sim B(a(K),b(K))$, giving us the possibility of leaving the average number of observations constant while skewing the distribution between a fully bimodal (small $b(K)$) and a sharp quasi-delta function (large $b(K)$). We choose the parameters $a$ and $b$ such that the mean $\mathbb{E}[p_i] = K$ is constant, so that different tests can be compared and the role of heterogeneity is highlighted. This binds the values of $a$ and $b$ through $a = \frac{K b}{1-K}$.
\begin{figure}
\includegraphics[width=.49\linewidth]{figure/Beta/betaCS_nomarg.pdf}
\includegraphics[width=.49\linewidth]{figure/Beta/betaMSE_nomarg.pdf}
\caption{(left) Reconstruction efficiency as a function of $K$ with different Beta parameters. Inset: the pdf of the adopted Beta distributions with $K=0.5$ (color coding is the same as in the main panel) (right) Root Mean Square Error on the couplings as a function of $K$ with different Beta parameters.}\label{betadep}
\end{figure}
The results of Figure \ref{betadep} clearly show that when the distribution is bimodal, that is when some variables are very rarely observed, the performance of the algorithm is worse. With a sample size of $T=10^4$ and $N=40$, the Dunn et al. model approximated by $B(a(K), 0.1)$ is identified with reasonable performance only when $K\geq 0.8$. This is extremely mitigated when the observations are more homogeneously distributed, particularly in the case of the coupling coefficients whose estimation seem to require a rather homogeneous distribution of observations among variables to be reliable. On the other hand, the reconstruction efficiency is far less demanding in terms of data quality and a reasonable performance is achieved even with sparse data and heterogeneous observations.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{figure/Beta/betaRMSE_vPlink_bpar_K7.pdf}
\includegraphics[width=1\linewidth]{figure/Beta/betaFit_vPlink_bpar_K7.pdf}
\caption{Quality of inference varying the probability of observing the end nodes at subsequent times. (top) RMSE for different values of the Beta $b$ parameter with mean $K=0.7$; (bottom) Linear fit coefficient for different values of the $b$ parameter, $K=0.7$.}
\label{fig:test1b}
\end{figure}
In Figure \ref{fig:test1b} we plot the Root Mean Square Error on couplings conditional on the probability of observing subsequently the spins at their ends. This probability is simply given by $p_{ij} = p_i p_j$ since observations are independently sampled, and the RMSE is
\begin{equation*}
\mathrm{RMSE}(p) = \sqrt{\langle (\hat{J}_{ij} - J_{ij})^2 \rangle_{p_{ij} = p}}
\end{equation*}
where the mean is taken on links that have (close to) the same joint observation probability. The plots highlight how the least observed the pair, the worse the precision of the fit, however it is also clear that the error grows for the more frequently observed couplings too. This is partially mitigated when one looks at the linear fit between the inferred $J$s and the true ones, meaning that the error is mostly affected by the variance component rather than the bias one.\\
The overall effect of heterogeneity is thus a decrease in the quality of the inference, with a stronger effect on couplings that are between the least observed pairs of spins and an important loss in accuracy, but with a bias component that is mitigated for the most frequently observed pairs.
\subsection{Test 2: dependency on $J_1$}
So far we have dealt with elements of $J$ drawn i.i.d. from a $\mathcal{N}(0,1/N)$ distribution. We want to relax this hypothesis and, while changing the mean value of the distribution would not be particularly meaningful in that it would just shift the correlation patterns between variables, it makes sense to investigate the behaviour as one changes the variance and thus the strength of the interactions. While there is no phase transition in the underlying model as long as the $J_{ij}$ are i.i.d., we want to check how weak can the couplings be in order to be correctly inferred and give a reliable reconstruction of the data. In other words, we are trying to identify a threshold in the interaction strength below which the algorithm is unable to converge. \\
We report results for an experiment with $N=100$, $T=10000$, $p_i=p=0.8$ and $J_1$ ranging from $0.05$ to $13$. We see from Figure \ref{fig:jdep} that increasing the typical size of couplings positively affects the quality of the inference, as should be expected since the dynamics is less affected by randomness. In the top panel we plot the reconstruction efficiency which has a steady increase and saturates towards $1$ after $J_1 \simeq 5$. The bottom panel shows the relative RMSE, that is $\mathrm{RMSE}/J_1$, and we see that it drops below $5\%$ for $J_1 > 0.5$. It is rather surprising to see how, regardless of the small couplings expansion we utilize in Eq. \ref{eq:approx}, the algorithm seems to work efficiently even in cases where the variance of the couplings $J_1^2/N$ is of order $1$, albeit a region of optimality for the inference of the couplings seems to lie within $0.5 \leq J_1 \leq 7$.
\begin{figure}[h]
\includegraphics[width=1\linewidth]{figure/Jdep/CSvJ_wide.pdf}
\includegraphics[width=1\linewidth]{figure/Jdep/JMSEvJnorm_wide.pdf}
\caption{(top) Reconstruction Efficiency as a function of $J_1$. (bottom) Rescaled RMSE (by $J_1$) on the couplings as a function of $J_1$.}\label{fig:jdep}
\end{figure}
\subsection{Test 3: impact of network structure}
\begin{figure}
\centering
\includegraphics[width=.49\linewidth]{figure/Lasso/JMSEvlasso.pdf}
\includegraphics[width=.49\linewidth]{figure/Lasso/Lasso_ROC.pdf}
\includegraphics[width=.49\linewidth]{figure/Deciall/tiltlik_nomarg_fix.pdf}
\includegraphics[width=.49\linewidth]{figure/Deciall/Deci_ROC_nomarg_fix.pdf}
\caption{(top) Results from the LASSO with 80\% observations: (left) RMSE on couplings as a function of the LASSO parameter; (right) ROC curves. (bottom) Results from the decimation procedure with 80\% observations: (left) Tilted likelihood evolution through the decimation process, vertical lines show the correct number of null elements; (right) ROC curves through the decimation process with different network densities. The circle identifies the point at which the Tilted Likelihood is maximized.} \label{fig:test3}
\end{figure}
We test the algorithm performance on some more realistic network structure than the fully connected one. It is indeed known that real networks, and particularly social networks, are typically sparse and thus network models have to implement some pruning mechanism permitting to discriminate between noise, spurious correlations and actual causal relations. We generate our data simulating the Kinetic Ising model on one of the simplest random network models, the Erd\H{o}s-R\'{e}nyi model, with edges that have weights $J_{ij}$ normally distributed with variance $1/N$, $N=100$ and $T=10000$ and with a probability of observing the variables of $p \in \lbrace 0.8, 0.6, 0.4 \rbrace$. One then needs to adjust the algorithm to give sparse solutions, as the mean field approximation will tend to return fully connected $J$ matrices. The adjustments we make are the LASSO regularization and the decimation procedure of Decelle et al. \cite{decelle2015inference}. The first is the well known $\ell_1$ norm regularization of the objective function, which projects the maximum likelihood fully connected solution on a symplex of dimensions determined by a free parameter $\lambda$ (which has to be validated out of sample).\\
The second is a recently proposed technique that selects parameters starting to decimate them from the least significant ones and repeating the process until a so-called Tilted log-Likelihood function shows a discontinuity in the first derivative.\\
To briefly describe the procedure, call $\mathcal{L}_{max}$ the value of the log-likelihood provided by the maximum likelihood algorithm without any constraint and then call $x$ the fraction of parameters $J_{ij}$ that are being set to $0$. Finally call $\mathcal{L}(x)$ the log-likelihood of the model with the fraction $x$ of decimated parameters and $\mathcal{L}_1$ the log-likelihood of a model with no couplings that is, in case $h_i =0 \, \forall \, i$, $\mathcal{L}_1 = - \sum_t M(t) \log 2$. The Tilted log-Likelihood takes the form
\begin{equation*}
\mathcal{L}^{tilted}(x) = \mathcal{L}(x) - \left((1-x) \mathcal{L}_{max} + x \mathcal{L}_1 \right)
\end{equation*}
that is, the difference between a convex combination of the original log-likelihood with the log-likelihood of a system with no parameters and the log-likelihood of the decimated model. This function is strictly positive and is $0$ only for $x=0,1$, since $\mathcal{L}(0)=\mathcal{L}_{max}$ and $\mathcal{L}(1) = \mathcal{L}_1$, thus there has to be a maximum. The decimation process thus consists in gradually increasing the fraction of pruned parameters $x$ until the maximum of the Tilted log-Likelihood is found, giving the optimal set of parameters of the model.\\
We show in Figure \ref{fig:test3} and \ref{fig:test3b} the results of the test. We observe how the ROC curves seem to lean strongly in favor of the decimation approach, which tends to score perfectly on the False Positives Ratio (FPR) - True Negatives Ratio (TNR) plane. However the maximum of the Tilted Likelihood does not always correspond to the optimal score in the ROC diagram, both in the case of a non-sparse network and when the data has a large number of missing values. While the former case is not particularly interesting in that a dense network model fitted on real data would be prone to overfitting and of disputable use, the latter is much more of a concern, albeit the process is still surprisingly efficient even when data is extremely sparse.\\
\begin{figure}
\centering
\includegraphics[width=.49\linewidth]{figure/Deciall/tiltlik_vPobs.pdf}
\includegraphics[width=.49\linewidth]{figure/Deciall/Deci_ROC_vPobs.pdf}
\includegraphics[width=.49\linewidth]{figure/Deciall/tiltlik_nomarg_cut.pdf}
\includegraphics[width=.49\linewidth]{figure/Deciall/Deci_ROC_nomarg.pdf}
\caption{(top) Results from the decimation procedure with 80\%, 60\% and 40\% observations available and a network density of $0.05$: (left) Tilted Likelihood evolution through the decimation, vertical line shows the correct number of null elements; (right) ROC curves through the decimation process with different observation densities. (bottom) Results from the decimation introducing local fields $h$: (left) Tilted likelihood, vertical lines show the correct number of null elements; (right) ROC curves. The introduction of local fields makes the tilted likelihood non-convex and seriously affects the performance.}
\label{fig:test3b}
\end{figure}
Even if the decimation procedure is consistently outperforming the LASSO, there is reason to still hold the $\ell_1$ regularization as a viable option. Indeed when one introduces local fields $h$ of non-negligible entity, the decimation procedure is not anymore reliable in that the Tilted Likelihood becomes non-convex as shown in Figure \ref{fig:test3b} and the maximum is not in the correct position. This is due to the underestimation of the $h$ parameters during the log-likelihood maximization of the fully connected model, where part of the role of the local fields is absorbed in couplings that should be pruned. However these couplings are still relevant to the model since they compensate for the underestimated $h$ parameters, giving the Tilted likelihood a non-convex form and shifting its maximum towards a more dense network model. This situation does not occur with the LASSO regularization as the pruning is performed at the same time as the maximization, giving the LASSO the advantage of a much more reliable fit of the local fields albeit with an overall worse performance in the inference of the nonzero couplings.
\subsection{Test 4: Impact of asymmetricity assumption}
Another assumption we made to perform the calculations in Equation \ref{eq:approx} was that the $J_{ij}$ are iid Gaussian random variables. In the case of social networks and trade networks reciprocity, that is the correlation between $J_{ij}$ and $J_{ji}$, is often found to be much higher than what would be expected in an iid setting \cite{squartini2013reciprocity}. We ask ourselves how impactful is this assumption on the outcome of the inference and we test the algorithm on data generated from a model with $N=100$, $T=10000$, $p_i=p=0.8$, $J_1=1$ and such that $\mathrm{Cor}(J_{ij}, J_{ji}) = \rho, \; i \neq j$. We show the results for this series of tests in Figure \ref{fig:rho}. What we find is that the $\rho$ parameter barely affects the performance and even makes it easier to infer the hidden variables, albeit marginally. Indeed we only used the assumption to approximate the determinant of the Hessian in the second order correction to the saddle-point solution, and letting the couplings not be reciprocally independent should affect the approximation slightly by having some elements of $J^2$ that vanish slower than others in the sums. It is possible that having a large enough $N$ facilitates the inference then, since the amount of those slowly vanishing terms grows with $N$ while the number of entries of $J$ grows with $N^2$.\\
\begin{figure}
\includegraphics[width=1\linewidth]{figure/Asymm/CSvRho_wide.pdf}
\includegraphics[width=1\linewidth]{figure/Asymm/JMSEvRho_wide.pdf}
\caption{(top) Reconstruction Efficiency varying the correlation between symmetric elements of $J$; (bottom) RMSE on the couplings.}\label{fig:rho}
\end{figure}
We then turn our attention to the extreme case of $\rho=1$, corresponding to the well known Sherrington-Kirkpatrick (SK) model \cite{kirkpatrick1978infinite}, one of the first and most studied spin glass models in the literature.
The SK model has the peculiarity of undergoing a phase transition at $J_1=2$ in our notation for the Hamiltonian (since we have not included a factor $1/2$ to remove double counting), where for $J_1>2$ the spin glass phase arises and multiple equilibrium states appear such that the model is not easy to infer anymore. It is thus interesting to see whether this affects the inference from dynamical configurations and how the identifiability transition is reached. We perform the experiment of varying $J_1$ in this framework and show the results in Fig. \ref{fig:SK}. We find the expected increase in rescaled error (that is, $\mathrm{RMSE}/J_1$) marking the transition, surrounded by a finite-size scaling noisy region, while the reconstruction efficiency of the configurations remains very good. This fits in the narrative of the phase transition of the SK model, since in the spin glass phase an equilibrium configuration of the model can be generated by multiple - and in principle undistinguishable - choices of parameters which we indeed struggle to identify with our methodology.
\begin{figure}
\includegraphics[width=1\linewidth]{figure/Asymm/SGCSvJext_wide.pdf}
\includegraphics[width=1\linewidth]{figure/Asymm/SGJMSEvJnorm_wide.pdf}
\caption{(top) Reconstruction Efficiency as a function of $J_1$ in the SK model; (bottom) Rescaled RMSE on couplings as a function of $J_1$.}\label{fig:SK}
\end{figure}
\subsection{Test 5: sample size and convergence}
We finally devolve our attention to the convergence properties of our estimator and how they are affected by finite sample sizes. The relevant parameter to be varied is the ratio between the length of the time series $T$ and the number of units that are modelled, $N$. We run simulations with $N=100$, $J_1 = 1$, $p_i = p = 0.8$ and varying $T$ between $100$ and $25000$, and report the results in Figure \ref{fig:tdep}. It can be seen that the RMSE on $J_{ij}$ diminishes, after $T/N=20$, with what might look like a power law behaviour with exponent close to $0.5$, although we do not provide an exact law for the convergence. The RMSE is below $5\%$ of $J_1$ when $T/N$ is larger than $20$ and is steadily converging towards $0$. Regarding the reconstruction efficiency we see that it saturates quickly towards $90\%$ and then it keeps increasing towards $100\%$. This evidence is an heuristic proof that the estimator is converging and is important to estimate how reliable a result might be given the $T/N$ ratio of the data. Although a more rigorous law would be much more appealing for the task, it would require being able to write the posterior of $J,\sigma$ given $s$, which to the best of our knowledge is not a feasible calculation in this setting.
\begin{figure}
\includegraphics[width=1\linewidth]{figure/Tdep/CSvTmax_wide.pdf}
\includegraphics[width=1\linewidth]{figure/Tdep/MSEJvTmax_wide.pdf}
\caption{(top) Reconstruction Efficiency as a function of the T/N ratio; (bottom) RMSE as a function of the T/N ratio. Area in green is 1 standard deviation from the mean over 30 repetitions.}\label{fig:tdep}
\end{figure}
\subsection{Additional parameters: exogenous drivers}
The model can be easily extended to a version in which an exogenous driver (or multiple ones), observed at all times, affects the dynamics of the variables. In a financial setting the first external driver would be given by the log-returns $r_t$ and the associated parameter would be the typical reaction of a trader to price changes, typically categorized between contrarians and chartists whether they go ``against" the flow (i.e. sell when the price rises and viceversa) or follow the trend. In the model, this is introduced by adding a set of linear parameters $\beta$ in the local fields that couple the variables to the driver
\begin{equation*}
g_k(t) = \sum_l y_l (t) + h_k + \beta_k r_t
\end{equation*}
The introduction of the parameter does not complicate the inference process at all and is particularly important if one wants to use the model to describe and possibly forecast order flows in financial markets. We omit the results for this section for the sake of space and because no significant dependency on the size of the $\beta_k$ parameters is found for our performance metrics.
\section{Conclusions}
In this article we develop a methodology to perform inference of Kinetic Ising Models on datasets with missing observations. We successfully adapt a known approximation from the Mean Field literature to the presence of missing values in the sample and devise several performance tests to characterize the algorithm and show its potential. We also propose a recursive methodology, R-EM, that gradually reconstructs the dataset with inferred quantities and tries to refine the inference, and show its efficacy on synthetic data.\\
The main results are that it is indeed possible to infer Kinetic Ising Models from incomplete datasets and that our procedure is resilient to noise, heterogeneity in the nature of data and in the frequency of missing values, and overall quantity of missing data. We make the algorithm ready for real-world applications by implementing pruning techniques in the form of LASSO and decimation, and give a brief overview of what we think are the better uses for each.\\
The methodology lends itself to applications on many diverse datasets, but our main focus for future research will be on opinion spreading in financial markets where transactions occur at high frequency, such as the FX or the cryptocurrency markets. We indeed envision our algorithm can identify significant structures of lagged correlations between traders, that in turn can be mapped to a network of lead-lag relations. Such a network would be particularly useful to get a quantitative picture of how possible speculative or irrational price movements can occur due to voluntary or involuntary coordination between traders and to devise appropriate strategies to counteract them.
\vfill
\section*{Acknowledgements}
The authors are grateful to prof. Matteo Marsili and to the participants of the 2018 Spring College on the Physics of Complex Systems (Trieste) for insightful comments and discussions. DT acknowledges GNFM-Indam and SNS for financial support of the project SNS18\_A\_TANTARI.
|
\section{Introduction}\label{intro}
The goal of data assimilation is to optimally combine known
information about the dynamics of a solution with low-resolution
observational measurements of that solution over time to create better
and better approximations of the current state.
While model error in the dynamics and measurement error in the
observations are significant issues with practical data assimilation,
we consider here the error-free case in order to study the
role played by spatial filtering.
In particular, even if the observations are error free,
in certain geophysical models
they can contain high-frequency spillover and gravity waves
which need to be controlled in order for the data assimilation
to perform well.
Additional issues arise because commonly used filtering
techniques can lead to non-orthogonal interpolants.
These issues are the focus of the current paper.
Our results extend the work of Hayden, Olson and Titi
\cite{Hayden2011} on discrete-in-time data assimilation
from the case where the low-resolution observations are given by
projection onto the low Fourier modes
to both the first and second type of general interpolant
observables that appear in Azouani, Olson and Titi \cite{Azouani2014},
see also Bessaih, Olson and Titi~\cite{Bessaih2015}.
To make this extension, we apply a spectral filter based on the
Stokes operator to the interpolant observables and call the
new method spectrally-filtered discrete-in-time downscaling data assimilation.
It is worth noting that much of advances in the accuracy of present
day weather forecasting have come from better filtering techniques,
see for example Budd, Freitag and Nichols \cite{Budd2011}.
From this point of view, the analytic results presented here for
spectral filtering may be seen as a first step towards a
rigorous analysis of more complicated methods.
An alternative algorithm for discrete-in-time data assimilation
based on nudging was recently studied by Foias, Mondaini and Titi
in \cite{Foias2016}.
In that work it was shown that nudging works for interpolants of
what is known by now as type-I, such as those which are based
on local coarse spatial scale volume elements measurements without any additional
filtering---the dissipation provided by the Navier--Stokes equations
themselves is sufficient; however,
a similar treatment for type-II interpolant observables is missing.
The algorithm studied here is based on the classical technique
of inserting the observational data directly into the model as it
is evolved forward in time, see for example Daley \cite{Daley1991}
and references therein.
When inserting the data directly into the model,
the need for filtering becomes more evident.
Moreover, by developing a spectrally-filtered algorithm we are
also able to handle type-II interpolant observable.
Although
it is likely a similar technique could be applied to a
nudging algorithm to handle type-II interpolant observables,
we do not pursue that line of analysis here,
but will be reported in future work.
The two-dimensional incompressible Navier--Stokes equations are given by
\begin{plain}\begin{equation}\label{navierstokes}
{\partial U\over\partial t}-\nu\Delta U
+\nabla P + (U\cdot\nabla) U = f,\qquad
\nabla\cdot U = 0.
\end{equation}\end{plain
Following Constantin and Foias~\cite{Constantin1988},
Foias, Manley, Rosa and Temam~\cite{Foias2001},
Robinson~\cite{Robinson2001} and
Temam~\cite{Temam1983}, and in order to simplifying our presentation and fix ideas,
we consider flows on the domain $\Omega=[0,L]^2$ equipped
with periodic boundary conditions.
Let ${\cal V}$ be the set of all divergence-free $L$-periodic
trigonometric polynomials with zero spatial averages,
$V$ be the closure of ${\cal V}$ in $H^1(\Omega,\R^2)$,
$V^*$ be the dual of $V$, and
$P_\sigma$ be the orthogonal projection of $L^2(\Omega;\R^2)$
onto $H$, where
$H$ is the closure of ${\cal V}$ in $L^2(\Omega,\R^2)$.
Define $A\colon V\to V^*$ and $B\colon V\times V\to V^*$
to be the unique continuous extensions
for $u,v\in{\cal V}$
of the operators given by
$$
Au=-P_\sigma \Delta u\qquad\hbox{and}\qquad
B(u,v)=P_\sigma (u\cdot\nabla v).
$$
Remark that in periodic case $A=-\Delta$, thus, the two-dimensional incompressible
Navier--Stokes equations may be written as
\begin{plain}\begin{equation}\label{2dns}
{dU\over dt}+\nu AU+B(U,U)=f
\end{equation}\end{plain
with initial condition $U_0\in V$, at time $t=t_0$.
Here
$\nu>0$ is the kinematic viscosity, and the body force
$f\in L^\infty([t_0,\infty);H)$
is taken to be divergence free, but possibly time dependent.
When the force is time independent, as shown in any of the aforementioned
references, equations (\ref{2dns}) are well posed with
unique regular solutions depending continuously on the initial
conditions and which exist for all time, $t\ge t_0$.
The case when the force depends on time is somewhat more delicate
and we shall place further assumptions on $f$ in Section \ref{secprelim},
see also Appendix \ref{apriori},
to ensure the resulting solutions have enough regularity
for the subsequent analysis.
In either case, we define the semi-process $S$ as the
solution operator
that maps initial conditions into their subsequent time
evolution by $S(t,t_0;U_0)=U(t)$ for all $t\ge t_0$.
We now describe the general interpolant observables to which our
results will apply.
These interpolants are inspired by the modes, nodes and
volume elements of Jones and Titi \cite{Jones1993},
see also Foias and Titi \cite{Foias1991},
and are equivalent to
the first and second types of general interpolant observables
that appear in \cite{Azouani2014}, see also \cite{Bessaih2015} and the
general framework presented in Cockburn, Jones and Titi \cite{Cockburn1997}.
In particular, we state
\begin{definition}\label{interpolants}
A linear operator ${I_h}\colon V\to L^2$ is said
to be a {\it type-I interpolant observable\/} if there
exists $c_1>0$ such that
\beq\label{typeone}
\|U-{I_h}U\|_{L^2}^2\le c_1h^2\|U\|^2
\wwords{for all} U\in V.
\eeq
A linear operator ${I_h}\colon \DA\to L^2$ is said to be a
{\it type-II interpolant observable\/} if
\beq\label{typetwo}
\|U-{I_h}U\|_{L^2}^2 \le c_1h^2\big(\|U\|^2+h^2|AU|^2\big)
\wwords{for all} U\in \DA.
\eeq
\end{definition}
Here $\DA=H^2(\Omega)\cap V$ is the domain of $A$ viewed as an operator into $L^2$.
Specifically, in terms of Fourier modes, let
$$
H = \left\{\,
\sum_{k\in\J}\widehat U_ke^{ik\cdot x}
\words{:}
\widehat U_k\in\C^2,\quad
\widehat U_k^* = \widehat U_{-k},\quad k\cdot \widehat U_k=0
\words{and}
\sum_{k\in\J}|\widehat U_k|^2 <\infty
\,\right\},
$$
where
$$
\J=\left\{ \frac{2\pi}{L}(n_1,n_2):
n=(n_1,n_2)\in\mathbb{Z}^2\backslash\{(0,0)\} \right\}.
$$
For notational convenience assume $\widehat U_0=0$ even
though this coefficient doesn't enter into the above characterization
of $H$.
We employ the notations
$$
|U|=\|U\|_0,\qquad
\|U\|=\|U\|_1
\wwords{and}
|AU|=\|U\|_2$$
where
\begin{equation}\label{alphanorm}
\|U\|_\alpha^2=L^2\sum_{k\in\J}|k|^{2\alpha}|\widehat U_k|^2
\wwords{when}
U=\sum_{k\in\J} \widehat U_k e^{ik\cdot x}.
\end{equation}
Further define
$V_\alpha= \{\, U\in H : \|U\|_\alpha<\infty\,\}$. Consequently
$\DA=V_2$ and $V=V_1$.
In Definition \ref{interpolants} we note that
$h$ is a length scale corresponding to the observation resolution
and $c_1$ is a dimensionless constant.
For example, suppose nodal measurements of the velocity are given by
$$
\big(U(x_1),U(x_2),\ldots,U(x_d)\big),
$$
where $x_i\in\Omega$ have been chosen in such a way that
$$
\sup_{x\in\Omega} \inf \big\{\, \|x-x_j\|: j=1,2,\ldots,d\,\big\}
\le h.
$$
Then
\begin{plain}$$
{I_h(U)(x)
\sum_{j=1}^{d} U(x_i) \widetilde\chi_j(x)}
\wwords{where}
{\widetilde\chi_j(x)= \chi_j(x)
- {1\over |\Omega|}\int_\Omega \chi_j}
$$\end{plain
with
\begin{plain}
$$
\chi_j(x)=
\cases{
1 & if $\|x-x_j\|<\|x-x_i\|$ for all $i\ne j$\cr
0 & otherwise
}
$$\end{plain
is a type-II interpolant observable.
It is worth reflecting that the type-II interpolant observable
described above naturally results in a piecewise constant vector
field which is discontinuous.
Although $I_h(U)\in L^2$ as required, the Fourier transform of
the resulting vector field possesses a significant high-frequency
component due to the discontinuities.
A similar interpolant was considered in \cite{Gesho2016}
for numerical simulations of a data-assimilation method based on nudging.
Those computations show that the adverse effects of the
high-frequency spill over which result from the spatial discontinuities
can be mitigated by appropriate convolution with a smoothing kernel.
The spectral filtering considered in this work also removes
the high-frequency component in spatial Fourier representation while
enjoying additional approximation properties useful for the
analysis of the resulting data assimilation algorithm.
We now introduce the spectrally-filtered discrete-in-time data
assimilation algorithm which forms the focus of our study.
Let $P_\lambda\colon H\to H$ be the orthogonal projection
onto the Fourier modes with wave numbers $k$ such
that $|k|^2\le \lambda$ given by
$$
P_\lambda U = \sum_{|k|^2\le \lambda} \widehat U_k e^{ik\cdot x}.
$$
and let $Q_\lambda=I-P_\lambda$ be the
orthogonal complement of $P_\lambda$.
Now, given $\lambda>0$ and $I_h$ define
\begin{equation}\label{filtered}
{J=P_\lambda P_\sigma I_h}
\wwords{and}
E=I-J.
\end{equation}
Note, although no additional orthogonality or
regularity properties other than those appearing in Definition
\ref{interpolants} have been assumed
on $I_h$, the above spectral filtering yields an
operator~$J$ which is nearly orthogonal and has
a range contained in $\DA$.
The downscaling data assimilation algorithm studied in this
paper may now be stated as
\begin{definition}\label{algorithm}
Let $U$ be an exact solution of \eqref{2dns} which evolves according
to dynamics given by the semi-process $S$.
Let $t_n=t_0+n\delta$ be a sequence of times for which
partial observations of $U$ are interpolated by $I_h$.
Then the approximating solution $u$ given by
\begin{plain}
$$
\left\{
\eqalign{
u_0&=J U(t_0)\cr
u_{n+1}&= E S(t_{n+1},t_n;u_n)+J U(t_{n+1})\cr
u(t)&=S(t,t_n;u_n)\qquad\hbox{for}\qquad t\in[t_n,t_{n+1})\cr
}\right.
$$
is what we shall call
{\it spectrally-filtered discrete-in-time downscaling data assimilation}.
\end{plain
\end{definition}
We stress that only the spectrally filtered low-resolution observations
of the exact solution represented by $J U(t_n)$ for $t_n\le t$ are
used to construct the approximating solution $u$ at time~$t$.
Since we assume the dynamics governing the evolution of
$U$ to be known, then exact knowledge of the initial condition $U(t_0)=U_0$
would, in theory, obviate the need for data assimilation at subsequent times.
Of course, knowing the exact dynamics and being able to practically
compute with them are two different things.
Although not the focus of the present research,
the algorithm stated above may also be used
to stabilize the growth of numerical error.
Putting such numerical considerations aside,
we view the data assimilation algorithm given in
Definition~\ref{algorithm} as a way of improving
estimates of the unknown state of $U$ at time $t$ by means of known
dynamics and a time-series of low-resolution observations.
Intuitively, at each time $t_{n+1}$ a new measurement is used to
kick the approximating solution towards the exact solution by constructing
of an improved approximation of the current state $u_{n+1}$ which may be
seen as a combination of a prediction
based on the previous approximation
and a correction based on the observation.
This improved approximation then serves as an initial condition
from which to further evolve the approximating solution.
Since $JU(t_{n+1})$ is supported on a finite number of Fourier modes,
the regularity of $u_{n+1}$ is determined by
$ES(t_{n+1},t_n;u_n)$.
For type-I interpolant observables our working assumptions described
after Proposition \ref{attbnd} in Section \ref{intro} shall
imply that $u_{n+1}\in V$ and for type-II that $u_{n+1}\in\DA$.
Although we have taken the sequence of observation times $t_n$ to
be equally spaced, intuitively one might imagine
for a suitably small value of $\delta$ that
it would be sufficient for
\begin{equation}\label{imagine}0<t_{n+1}-t_n\le\delta,
\wwords{with} t_n\to\infty,\words{as} n\to\infty.
\end{equation}
Our analysis, however, makes use of a minimum distance between
$t_{n+1}$ and $t_n$ as well as the maximum.
Measurements need to be inserted frequently enough to overcome
the tendency for two nearby solutions to drift apart, while
at the same time the possible lack of orthogonality in our
general interpolant observables means measurements should not
be inserted too frequently.
Specifically, we need to have enough time to elapse between each insertion to allow enough time for the use of the dynamics of the equation, i.e. integrating the Navier-Stokes equations for long enough time to correct the high modes. Our algorithm consists of two steps:
\textit{Step 1.} Inserting the coarse spatial scale measurements.
\textit{Step 2.} Integrating the Navier-Stokes equations for short time, but not too short, to recover and correct the missing high modes, i.e. the fine spatial scales of the solutions.
Preliminary numerical simulations further indicate this requirement
is likely physical and not merely a technical condition used
by our analysis.
Given times $t_n$ that satisfy (\ref{imagine})
it would
be straightforward to construct a subsequence of observations $t'_n$
such that $\delta/2 < t'_{n+1}-t'_n \le 2\delta$
and obtain results similar to the ones presented here.
We leave such a refinement to the reader.
Note that the algorithm described above reduces to the discrete
data assimilation method studied in \cite{Hayden2011}
by taking $I_h=P_\lambda$.
In particular, when the interpolant observable itself is
given by an orthogonal projection onto the large-scale
Fourier modes.
In this work $I_h$ can by any interpolant operator satisfying
Definition \ref{interpolants}.
Carefully adjusting
the relationship between $h$ and $\lambda$ then allows us to
prove our main result, stated as
\begin{theorem}\label{mainresult}
Let $U$ be a solution to the two-dimensional incompressible
Navier--Stokes equations (\ref{2dns}) and $u(t)$, for $t\ge t_0$
be the process given by
Definition \ref{algorithm}.
Then, for every $\delta>0$ there exists $h>0$ and $\lambda>0$
depending only on $c_1$, $f$, $\nu$
such that
$$
|u(t)-U(t)|\to 0,\words{exponentially in time, as} t\to\infty.
$$
Here $c_1$ is the constant
in Definition \ref{interpolants}
given by the general interpolant observables.
\end{theorem}
\noindent
Since we have assumed the observational measurements to be noise-free
and that the exact solution evolves according to known dynamics, it is
natural to obtain a result in which the difference between the exact
solution $U$ and the approximation $u$ converges to zero over time.
We further remark that if by chance $u(t_n)=U(t_n)$
at any of the data assimilation steps,
then $u(t)=U(t)$ for all $t\ge t_n$.
In particular, if somehow $U_0$ is known exactly and we take $u_0=U_0$
as the first step of Definition \ref{algorithm},
then $u(t)=U(t)$ for all $t\ge t_0$.
This paper is organized as follows.
In section 2 we set our notation, recall some facts about the
Navier--Stokes equations and prove some preliminary results
regarding the spectrally-filtered interpolant observables that
will be used in our subsequent analysis.
Section 3 proves our main result for type-I
interpolant observables while section 4 treats the case
of type-II interpolant observables.
We finish with some concluding remarks concerning the
dependency of $h$ and $\lambda$ on $\delta$ and the other
physical parameters in the system.
\section{Preliminaries}\label{secprelim}
We begin by recalling some inequalities.
Writing the smallest eigenvalue of the Stokes operator $A$ as
$\lambda_1=(2\pi/L)^2$ we have the
Poincar\'e inequalities
\begin{equation}\label{poincareV}
\lambda_1 |U|^2\le \|U\|^2\wwords{for} U\in V
\end{equation}
and
\begin{equation}\label{poincareDA}
\lambda_1^2 |U|^2\le \lambda_1 \|U\|^2\le |AU|^2
\wwords{for}
U\in\mathcal{D}(A).
\end{equation}
An advantage of using the projection $P_\lambda$ in our
data assimilation algorithm,
rather than a different type of spatial filtering,
is that this directly leads to
improved Poincar\'e inequalities and reverse
inequalities which are, respectively, given by
\beq\label{pimproved}
\lambda |Q_\lambda U|^2 \le\|Q_\lambda U\|^2
\wwords{and}
\lambda^2 |Q_\lambda U|^2
\le \lambda \|Q_\lambda U\|^2\le |A Q_\lambda U|^2
\eeq
and
\beq\label{preverse}
\|P_\lambda U\|^2\le\lambda |P_\lambda U|^2
\wwords{and}
|AP_\lambda U|^2\le\lambda^2 |P_\lambda U|^2.
\eeq
All of the inequalities given in \eqref{poincareV},
\eqref{poincareDA}, \eqref{pimproved}
and \eqref{preverse} may easily be verified via Fourier series.
We also recall Agmon's
inequality \cite{Agmon2010} (see also \cite{Constantin1988}) as
\begin{equation}\label{agmon}
\|U\|_{L^\infty}\le C|U|^{1/2}|AU|^{1/2}.
\end{equation}
Here $C$ is a dimensionless constant depending only
on the domain $\Omega$.
As mentioned in the introduction, the spectrally filtered
interpolation operator $J$ given by (\ref{filtered})
possesses approximate orthogonality and regularity properties
that the original interpolant observable $I_h$ may fail to have.
We summarize these properties in
\begin{proposition}\label{Ebounds}
Let $c_1$ be the dimensionless constant appearing
in Definition \ref{interpolants}.
For type-I interpolant observables setting
$\varepsilon=c_1\lambda h^2$ yields
$$
|EU|^2\le \lambda^{-1}(1+\varepsilon)\|U\|^2
\words{and}
\|EU\|^2\le (1+\varepsilon)\|U\|^2,
\words{for every} U\in V.
$$
For type-II interpolant observables
setting $\varepsilon=c_1\lambda_1^{-1}\lambda^2h^2(1+\lambda_1h^2)$
yields
$$
|EU|^2\le(\lambda\lambda_1)^{-1}(1+\varepsilon) |AU|^2,\qquad
\|EU\|^2\le \lambda_1^{-1}(1+\varepsilon) |AU|^2
$$
and
$$|AEU|^2\le (1+\varepsilon) |AU|^2
\wwords{for every} U\in \mathcal{D}(A).$$
\end{proposition}
\begin{proof}[Proof of Proposition \ref{Ebounds}]
Estimate $|EU|$ for type-I interpolant observables as
\begin{align*}
|EU|^2&=|U-JU|^2=|U-P_\lambda U+P_\lambda U-P_\lambda P_\sigma I_h U|^2\\
&\le |Q_\lambda U|^2+|P_\lambda(U-P_\sigma I_hU)|^2
\le \lambda^{-1}\|U\|^2+ |U-P_\sigma I_hU|^2\\
&=\lambda^{-1}\|U\|^2+ |P_\sigma(U- I_hU)|^2\le
\lambda^{-1}\|U\|^2+ \|U- I_hU\|_{L^2}^2\\
&\le \lambda^{-1}\|U\|^2+c_1h^2\|U\|^2
=\left(\lambda^{-1}+c_1h^2\right)\|U\|^2.
\end{align*}
From the definition of $\varepsilon$ it follows that
\beq\label{EU}
|EU|^2\le \lambda^{-1}\big(1+\varepsilon\big)\|U\|^2.
\eeq
Similarly bound $\|EU\|$ as
\begin{align*}
\|EU\|^2 &=\|U-JU\|^2=\|Q_\lambda U\|^2+\|P_\lambda(U-P_\sigma I_hU)\|^2\\
&\le\|U\|^2+\lambda|U-P_\sigma I_hU|^2
\le\|U\|^2+\lambda c_1h^2\|U\|^2
\le (1+\varepsilon)\|U\|^2.
\end{align*}
Now, estimate $|EU|$ for type-II interpolant observables as
\beq
\begin{aligned}
|EU|^2
&=|U-P_\lambda P_\sigma I_hU|^2
=|U-P_\lambda U+P_\lambda U-P_\lambda P_\sigma I_hU|^2\\
&=|U-P_\lambda U|^2+|P_\lambda P_\sigma(U-I_hU)|^2
\le |Q_\lambda U|^2 + \|U- I_h U\|_{L^2}^2\\
&\le |Q_\lambda U|^2+c_1h^2\left(\|U\|^2+h^2|AU|^2\right)\\
&\le \left( \frac 1{\lambda}+c_1h^2 \right)\|U\|^2+c_1h^4|AU|^2\\
&\le \left[\frac{1}{\lambda_1}\left( \frac 1{\lambda}+c_1h^2 \right)
+c_1h^4\right]|AU|^2.
\end{aligned}
\eeq
Setting $\varepsilon=c_1\lambda_1^{-1}\lambda^2h^2(1+\lambda_1h^2)$
yields that
$$
|EU|^2\le (\lambda\lambda_1)^{-1}(1+\varepsilon)|AU|^2.
$$
Next, estimate $\|EU\|$ as
\beq
\begin{aligned}
\|EU\|^2
&=\|U-P_\lambda P_\sigma I_hU\|^2
=\|U-P_\lambda U+P_\lambda U-P_\lambda P_\sigma I_hU\|^2\\
&=\|U-P_\lambda U\|^2+\|P_\lambda P_\sigma (U-I_hU)\|^2\\
&=\|Q_\lambda U\|^2+\lambda|P_\sigma(U-I_hU)|^2\\
&\le\|Q_\lambda U\|^2+\lambda\|U-I_hU\|_{L^2}^2\\
&\le \lambda^{-1}|AU|^2+c_1\lambda h^2
\big(\|U\|^2+h^2|AU|^2\big)\\
&\le \lambda^{-1}|AU|^2+c_1\lambda h^2
\big(\lambda_1^{-1}|AU|^2+h^2|AU|^2\big)\\
&\le \big(\lambda^{-1}+c_1\lambda_1^{-1}\lambda
h^2(1+\lambda_1 h^2)\big)|AU|^2\\
&\le \lambda_1^{-1}\big(1+c_1 \lambda h^2(1+\lambda_1 h^2)\big)|AU|^2
\le \lambda_1^{-1}(1+\varepsilon)|AU|^2,\\
\end{aligned}
\eeq
and finally $|AEU|$ as
\beq
\begin{aligned}
|AEU|^2
&=|Q_\lambda AU|^2+|AP_\lambda (U-P_\sigma I_hU)|^2\\
&\le |A U|^2+\lambda^2|P_\sigma(U-I_hU)|^2\\
&\le |A U|^2+\lambda^2\|U-I_hU\|_{L^2}^2\\
&\le |AU|^2+c_1\lambda^2 h^2
\big(\|U\|^2+h^2|AU|^2\big)\\
&\le \big(1+c_1\lambda_1^{-1}\lambda^2
h^2(1+\lambda_1 h^2)\big)|AU|^2
= (1+\varepsilon)|AU|^2.\\
\end{aligned}
\eeq
This completes the proof of the proposition.
\end{proof}
Our analysis will make use of {\it a priori\/} bounds on
the solution $U$ of \eqref{2dns}.
If $f\in H$ is time independent,
such bounds can be inferred from
bounds on the global attractor.
For example, Propositions~12.2 and 12.4 in Robinson \cite{Robinson2001}
may be stated as
\begin{proposition}\label{attbnd}
If $f\in H$ is time independent, then there are absorbing sets
in $H$, $V$ and $\DA$ of radiuses
$\rho_H$, $\rho_V$ and $\rho_A$, respectively,
depending only on $|f|$, $\Omega$ and $\nu$ such that
for every $U_0\in H$ there is a time $t_A$ depending only
on $|U_0|$ for which
\begin{align}\label{rhoA}
| U(t) |\le \rho_H,\quad
\| U(t) \|\le \rho_V
\words{and}
|AU(t)|\le \rho_A
\words{for all} t\ge t_A.
\end{align}
Moreover,
\begin{equation}\label{Aint}
\int_t^{t+\delta} |AU|^2 \le
\Big({1\over \nu}+{\delta\lambda_1\over 2}\Big)
\rho_V^2
\words{for all} t\ge t_A.
\end{equation}
\end{proposition}
\noindent
Similar bounds may be found in
Temam \cite{Temam1983} and
Constantin and Foias \cite{Constantin1988}.
The best estimate of $\rho_A$ to date appears in \cite{Foias2015}.
Before considering the case when $f$ depends on time, we further
note when $f\in V$ is time independent that
the bounds in (\ref{rhoA}) are finite for $t>t_0$.
Moreover, (\ref{Aint}) is finite and
\begin{equation}\label{vortpose}
\int_t^{t+\delta} \|A U\|^2<\infty
\wwords{for all} t\ge t_0.
\end{equation}
We remark that estimate (\ref{vortpose}) follows as a particular case of the proof presented in Appendix \ref{apriori} for the time-dependent forcing term, see discussion below.
When $f\in L^{\infty}([t_0,\infty),H)$
depends on time, the resulting solution $U$ does not
automatically satisfy the $\rho_A$ bound in (\ref{rhoA})
nor the finiteness condition (\ref{vortpose}).
In the case of
type-I interpolant observables
the remaining bounds given by $\rho_H$ and $\rho_V$
are sufficient for our analysis.
However, for type-II interpolant observables
we need $\rho_A$ as well as the
finiteness condition (\ref{vortpose}).
These bounds may be obtained in a number of different ways.
For example, one could assume that
$f\in L^{\infty}([t_0,\infty),V)$ and
$df/dt\in L^{\infty}([t_0,\infty),V^*)$.
For details see Appendix \ref{apriori}.
Our analysis shall be made under the working
assumption that $\rho_H$, $\rho_V$ and $\rho_A$ are known when
needed and
that the unknown initial condition $U_0$ in \eqref{2dns}
comes from a long-time evolution prior to time $t_0$.
Thus, we assume $t_0\ge t_A$ and in particular that
the bounds \eqref{rhoA}, \eqref{Aint} and \eqref{vortpose}
hold, in fact, for $t\ge t_0$ regardless of whether $f$
depends on time or not.
For other initial conditions we further suppose that the norms
and time integrals appearing in all the above bounds are
at least finite when $t>t_0$.
We now state a standard result
concerning the finite-time continuous
dependence on initial conditions
for solutions
to the two-dimensional
incompressible Navier-Stokes equations.
\begin{theorem}\label{beta}
Under the working assumptions given above,
there exists $\beta>0$ depending only on {$|f|$, $L$ and
$\nu$} such that the free-running solution satisfies
\beq\label{wttn}
|U(t)-S(t,t_*;{u_*})|^2 \le
e^{\beta(t-t_*)}|U(t_*)-{u_*}|^2
\words{for} t\ge t_*\words{and}u_*\in V.
\eeq
\end{theorem}
\noindent
We remark that the above continuity result is obtained from the
first Lyapunov exponent, which reflects the instability in turbulent
flows.
Thus, the constant $\beta$ in Theorem \ref{beta}
is very large but uniform for $u_*\in V$.
The fact that $\beta$ does not depend on $u^*$ is a fact we
shall make salient use of in our subsequent analysis.
We recall that the bilinear term $B$ has the algebraic property that
\begin{equation}
\big\langle B(u,v),w\big\rangle = -\big\langle B(u,w),v\big\rangle
\label{balg}
\end{equation}
for $u,v,w\in V$, and consequently the orthogonality property that
\begin{equation}
\big\langle B(u,w),w\big\rangle =0.
\label{borth}
\end{equation}
Here the pairing $\langle \cdot,\cdot\rangle$ denotes the
dual action of $V^*$ on $V$. Details may be found, e.g., in
\cite{Constantin1988}, \cite{Foias2001}, \cite{Robinson2001}
and \cite{Temam1983}.
In the case of periodic boundary conditions the bilinear term
possesses the additional orthogonality property
\begin{equation}
\big(B(w,w),Aw\big)=0,
\qquad\hbox{for every}\quad w\in\DA;
\label{balgp}
\end{equation}
and consequently one has
\begin{equation}
\big(B(u,w),Aw\big)
+\big(B(w,u),Aw\big)
=-\big(B(w,w),Au\big), \qquad\hbox{for every}\quad u, w\in \DA.
\label{borthp}
\end{equation}
We further recall some well-known bounds on the non-linear term which
appear in \cite{Constantin1988}, \cite{Temam1983}, \cite{Temam1997} and
specifically as Proposition 9.2 in \cite{Robinson2001}.
\begin{proposition}
One has
\beq\label{Best0}
|(B(u,v),w)|\le \|u\|_{L^\infty}\|v\| |w|,
\eeq
where $u\in L^\infty, $ $v\in V$ and $w\in H$.
If $u, v, w \in V$ then
\beq\label{Best}
|(B(u,v),w)|\le c|u|^{1/2}\|u\|^{1/2}\|v\| |w|^{1/2}\|w\|^{1/2},
\eeq
and if $u\in V$ $v\in \DA$, and $w\in H$,
\beq\label{Best1}
|(B(u,v),w)|\le c|u|^{1/2}\|u\|^{1/2}\|v\|^{1/2}|Av|^{1/2} |w|.
\eeq
Here $c$ is an absolute non-dimensional constant.
\end{proposition}
\section{Type-I Interpolant Observables}
In this section we treat the case when $I_h$ is a type-I
interpolant observable.
While type-I interpolant observables are also of type II,
the bounds we obtain in treating these two cases separately
are sharper.
In addition, the proof for in the type-I case is simpler
and serves as a framework to help understand the
more complicated type-II case teated in the subsequent section.
From Definition \ref{algorithm} it follows that the
approximating solution $u$ satisfies
\beq\label{ns2}
\frac{du}{dt}+\nu Au+B(u,u)=f
\wwords{for} t\in (t_n,t_{n+1}),
\eeq
where $u(t_n)=u_n$ is the initial condition given by
$$u_0=JU_0\wwords{and} u_{n+1}=ES(t_{n+1},t_n;
u_n)+JU(t_{n+1}).$$
Note that $u_n\in\DA\subseteq V$, for $n=0,1,2,\dots$.
Consequently, the solution of \eqref{ns2}
with initial data $u(t_n)=u_n$
on the interval $(t_n, t_{n+1})$
is a strong solution of the Navier--Stokes equations.
Moreover, because of our working assumptions on $f$ we further
obtain that
$u(t)\in\DA$ for $t\in[t_n,t_{n+1}]$.
It follows that the estimates we make in the proof of
Proposition \ref{tildelemma} below, and in the results
which follow, are rigorous;
in particular, $v=U-u$ exists, is unique and $Av$ makes sense
at all times $t\ge t_0$.
The equations governing the evolution of $v$ may be written as
\beq\label{nse1}
\frac{dv}{dt}+\nu Av+B(v,U)+B(U, v)+B(v,v)=0
\eeq
for $t\in(t_n, t_{n+1})$, with $v(t_n)=U(t_n)-u_n$, for $n=0,1,2,\dots.$.
\begin{proposition}\label{tildelemma}
Let $\tilde{v}_n=U(t_n)-S(t_n,t_{n-1};u_{n-1})$.
For every $\delta>0$ there are $\lambda$, large enough,
and $h$, small enough, for which there exists $\gamma\in (0,1)$
such that
$$
\|\tilde{v}_{n+1}\|^2\le \gamma\|\tilde{v}_n\|^2,
\wwords{for all} n=1,2,\ldots.
$$
\end{proposition}
\begin{proof}
Multiplying \eqref{nse1} by $Av$ and then integrating over $\Omega$ we have
$$
\frac 12\frac{d}{dt} \|v\|^2
+\nu |Av|^2+ (B(v,U), Av)+(B(U, v), Av)+(B(v,v),Av)=0.
$$
By \eqref{balgp} and \eqref{borthp}, we have
\beq\label{me1}
\frac 12\frac{d}{dt} \|v\|^2+\nu |Av|^2=(B(v,v), AU).
\eeq
Estimate the term on the right of the previous equation.
Using \eqref{Best1} and then the interpolation inequality $\|v\|\le |v|^{1/2}|Av|^{1/2}$ yields
\begin{align*}
|(B(v,v), AU)| &\le c|v|^{1/2}\|v\|^{1/2} \|v\|^{1/2} |Av|^{1/2}|AU|\\
&= c|v|^{1/2}\|v\||Av|^{1/2}|AU|\\
&\le c|v|^{1/2}|v|^{1/2}|Av|^{1/2}|Av|^{1/2}|AU|\\
&= c|v||Av||AU|.
\end{align*}
Combining this with \eqref{me1}, we have
\beq\label{me2n}
\frac 12\frac{d}{dt} \|v\|^2+\nu |Av|^2\le c|v||Av||AU|.
\eeq
Now, apply Young's inequality to obtain
\beq\label{BwwAu}
\frac{d}{dt} \|v\|^2+\nu |Av|^2
\le \frac{c^2}{\nu}|v|^2|AU|^2.
\eeq
From Poincar\'e's inequality \eqref{poincareDA}
followed by \eqref{wttn}, we get
\beq\label{me4}
\frac{d}{dt} \|v\|^2+\lambda_1\nu \|v\|^2
\le \frac{c^2}{\nu}|AU|^2e^{\beta(t-t_n)}|v_n|^2
\le \frac{c^2}{\nu}|AU|^2e^{\beta \delta}|v_n|^2,
\eeq
where we have assumed $t\in [t_n,t_{n+1})$.
Multiply equation \eqref{me4} by
$e^{\lambda_1\nu t}$
and
then integrate in time from $t_n$ to $t$. Thus,
\begin{plain}\begin{equation}\label{me8}\eqalign{
\|v(t)\|^2
&\le e^{-\lambda_1\nu (t-t_n)}\|v_n\|^2
+\frac{c^2}{\nu}
e^{-\lambda_1\nu(t-t_n)+(\beta+\lambda_1\nu)\delta}
|v_n|^2\int_{t_n}^t|AU(s)|^2ds
}\end{equation}\end{plain
for $t\in [t_n, t_{n+1})$.
Combining \eqref{me8} with the {\it a priori\/} estimate
\eqref{Aint}, we have
\beq\label{me8n}
\|v(t)\|^2\le e^{-\lambda_1\nu (t-t_n)}\|v_n\|^2
+\frac{c^2\rho_V^2}{\nu}
\left( \frac{1}{\nu}+{\delta \lambda_1\over 2}\right)
e^{-\lambda_1\nu(t-t_n)+(\beta+\lambda_1\nu)\delta}
|v_n|^2.
\eeq
Since $n\ge 1$ then
\begin{align*}
v_n=U(t_n)-u_n&=U(t_n)-ES(t_n,t_{n-1};u_{n-1})-JU(t_n)\\
&=E\big(U(t_n)-S(t_n,t_{n-1};u_{n-1})\big)
=E(\tilde v_n),
\end{align*}
and by Proposition \ref{Ebounds},
we can estimate
\begin{align*}
|v_n|^2
\le \lambda^{-1}(1+\varepsilon)\|\tilde{v}_n\|^2
\wwords{and}
\|v_n\|^2
\le (1+\varepsilon)\|\tilde{v}_n\|^2.
\end{align*}
Hence \eqref{me8n} becomes
\beq\label{tbound}
\|v(t)\|^2\le
(1+\varepsilon)
e^{-\lambda_1\nu (t-t_n)}
\left[1+\frac{c^2\rho_V^2}{\lambda\nu}
\left( \frac{1}{\nu}+{\delta \lambda_1\over 2}\right)
e^{(\beta+\lambda_1\nu)\delta}
\right]\|\tilde{v}_n\|^2,
\eeq
for $t\in [t_n, t_{n+1})$.
Taking the limit as $t \nearrow t_{n+1}$ results in
$
\|\tilde{v}_{n+1}\|^2 \le\gamma
\|\tilde{v}_n\|^2,
$
where
$$ \gamma= (1+\varepsilon)
\left[e^{-\lambda_1\nu \delta}+\frac{c^2\rho_V^2}{\lambda\nu}
\left( \frac{1}{\nu}+{\delta \lambda_1\over 2}\right)
e^{\beta\delta}
\right].
$$
We now show for every $\delta>0$ that there exists $\lambda$
and $h$ such that $\gamma\in(0,1)$. First, since
$$
e^{-\lambda_1\nu\delta}<1\wwords{and}
\frac{c^2\rho_V^2}{\lambda\nu}
\left( \frac{1}{\nu}+{\delta \lambda_1\over 2}\right)
e^{\beta \delta}
\to 0
\words{as}\lambda\to\infty,
$$
then there is $\lambda$ large enough such that
$$
e^{-\lambda_1\nu \delta}+\frac{c^2\rho_V^2}{\lambda\nu}
\left( \frac{1}{\nu}+{\delta \lambda_1\over 2}\right)
e^{\beta\delta}
<1.
$$
Finally, since $\varepsilon\to 0$, as $h\to 0$, while holding
$\lambda$ fixed,
then there is $h$ small enough such that $1+\varepsilon$
is small enough to ensure that $\gamma<1$.
\end{proof}
Observe that by a more careful analysis one could find explicit
choices for $\lambda$ and $h$ in terms of $\beta$, $\delta$,
$\lambda_1$, $\nu$ and $\rho_V$.
Note also that there is a dependency between $\lambda$ and $h$.
Since $h$ is a physical parameter related to the resolution
of the observations
while $\lambda$ is an easily-adjusted parameter
related to our spectral filter, it would be reasonable to further
choose $\lambda$ to minimize $h$.
The resulting estimate on $h$ could then be used to compare
the sharpness of the above theoretical bounds to alternative approaches
to the analysis, to numerical results obtained from simulation and to
similar analysis for different data assimilation schemes.
Such comparisons, while interesting, are outside the scope of
the present work. We end this section with our main result on
type-I interpolant observables.
\begin{theorem}\label{mainI}
If $\delta$, $h$ and $\lambda$ are chosen appropriately as in Proposition \eqref{tildelemma},
then $\|U(t)-u(t)\|\to 0$, as $t\to\infty$.
Moreover, the rate of convergence is exponential in time.
\end{theorem}
\begin{proof}
Choose $\delta$, $h$ and $\lambda$ as in Proposition \ref{tildelemma}.
In reference to equation \eqref{tbound}, let
$$
M=(1+\varepsilon)
\left[1+\frac{c^2\rho_V^2}{\lambda\nu}
\left( \frac{1}{\nu}+{\delta \lambda_1\over 2}\right)
e^{(\beta+\lambda_1\nu)\delta}
\right].
$$
We first bound $\tilde v_1$ in terms of $v_0$.
Since
$$
v_0=U_0-u_0=U_0-JU_0=EU_0,
$$
then Proposition \ref{Ebounds} and the working
assumptions which follow Proposition \ref{attbnd}
yield that
$$
|v_0|^2=|EU_0|^2 \le
\lambda^{-1} (1+\varepsilon)\|U_0\|^2
\le
\lambda^{-1} (1+\varepsilon)\rho_V^2,
$$
and similarly that $\|v_0\|^2\le (1+\varepsilon)\rho_V^2$.
These two bounds substituted into \eqref{me8n}
for $n=0$ imply
\beq\label{me9n}
\|v(t)\|^2\le (1+\varepsilon)
e^{-\lambda_1\nu (t-t_0)}
\left[1
+\frac{c^2\rho_V^2}{\lambda\nu}
\left( \frac{1}{\nu}+{\delta \lambda_1\over 2}\right)
e^{(\beta+\lambda_1\nu)\delta}
\right]
\rho_V^2.
\eeq
for $t\in[t_0, t_1)$.
Taking the limit as
$t\nearrow t_1$ results in $\|\tilde v_1\|^2\le\gamma\rho_V^2$
where $\gamma\in(0,1)$.
Now, given $t>0$ choose $n$ such that $t\in [t_n,t_{n+1})$.
Since $n>(t-t_0)/\delta-1$,
it follows from (\ref{tbound}) that
$$\|U(t)-u(t)\|^2=\|v(t)\|^2\le M\|\tilde v_n\|^2
\le M\gamma^{n} \rho_V^2
\le M \gamma^{-1} \rho_V^2 e^{-\alpha (t-t_0)},
$$
where
$\alpha=\delta^{-1}\log(\gamma^{-1})$.
Note that $\gamma\in(0,1)$ implies $\alpha>0$.
It follows that $\|U(t)-u(t)\|$
converges to zero at an exponential rate.
\end{proof}
\section{Type-II Interpolant Observables}
In this section we treat the case when $I_h$ is a type-II
interpolant observable.
As before let $v=U-u$ where $U$ is the exact solution
to \eqref{2dns} about which we know only limited information
through the observables
and $u$ is the approximating process obtained by the
spectrally-filtered discrete data assimilation algorithm
given in Definition \ref{algorithm}.
The proof that the difference between $u$ and $U$ decays to zero
over time is complicated by the fact that the $|Av|$
norm enters into the bounds given by Proposition \ref{Ebounds}
and therefore needs to be controlled.
To do so, we shall employ an equation similar to \eqref{me1}
which governs the evolution of $|Av|^2$.
While such an equation could be obtained by formally multiplying
\eqref{nse1} by $A^2u$ and integrating over $\Omega$,
it is easier to work with the vorticity in two-dimensions.
Let $W=\curl U$, $w=\curl u$, and $g=\curl f$ where
$\curl$ has been defined such that
\begin{plain}$$
\curl \Phi=
{\partial \Phi_2(x_1,x_2)\over\partial x_1}
-{\partial \Phi_1(x_1,x_2)\over\partial x_2}
\wwords{when}
\Phi(x)=\big(\Phi_1(x_1,x_2),\Phi_2(x_1,x_2)\big).
$$\end{plain
Since $u$ is the approximating solution
described in Definition \ref{algorithm},
then $w$ is the resulting vorticity approximation of $W$.
Written in terms of vorticity, the corresponding version of Theorem
\ref{mainI} for type-II interpolant observables is given by
\begin{theorem}\label{mainII}
If $\delta$, $h$ and $\lambda$ are chosen appropriately,
then $\|W-w\|\to 0$, as $t\to\infty$.
Moreover, the rate of convergence is exponential in time.
\end{theorem}
Before proving Theorem \ref{mainII} we fix our notation
by stating a few facts
about the vorticity and proving a lemma containing
bounds for non-linear terms that will be used later.
First note, after taking the $\curl$ of \eqref{navierstokes},
that Definition \ref{algorithm} implies $W$ and $w$ satisfy
\beq\label{nset}
\frac{\partial W}{\partial t}-\nu \Delta W+(U\cdot\nabla)W=g
\wwords{and}
\frac{\partial w}{\partial t}-\nu \Delta w+(u\cdot\nabla)w=g
\eeq
on each interval $(t_n,t_{n+1})$. Our working assumptions
in the case of type-II interpolant observables ensure that
the equations (\ref{nset}) hold in the strong sense.
In particular,
$W=\curl U$ and $w=\curl u$ exist and
$|\Delta W|=|A^{3/2} U|$ and
$|\Delta w|=|A^{3/2} u|$
are finite almost everywhere.
Therefore, the equations governing the evolution through
the vorticity of the difference $\xi=W-w$ may be written as
\beq\label{nset2}
\frac{\partial\xi}{\partial t}-\nu \Delta \xi+(v\cdot\nabla)W
+(v\cdot\nabla)\xi+(U\cdot\nabla)\xi=0,
\eeq
where $\xi(t_n)=W(t_n)-\curl u_n$ and $v=\curl^{-1}\xi$.
Since $v$ is divergence-free with zero average, then
$\curl^{-1}\xi$ is well defined and may be written in terms of
Fourier series as
\begin{plain}$$
\curl^{-1}\xi=
\sum_{k\in\J} {i(k_2,-k_1)\over |k|^2} \widehat\xi_k e^{ik\cdot x}
\wwords{when}
\xi=\sum_{k\in\J} \widehat\xi_k e^{ik\cdot x}.
$$\end{plain
Recall that the divergence-free condition $k\cdot\widehat v_k=0$ implies
\begin{plain}$$\eqalign{
|\widehat\xi_k|^2&=
|ik_1\widehat v_{k,2}-ik_2\widehat v_{k,1}|^2
=
k_1^2 |\widehat v_{k,2}|^2+k_2^2 |\widehat v_{k,1}|^2
- k_1k_2 \widehat v_{k,1}\widehat v_{k,2}^*
- k_1k_2 \widehat v_{k,1}^*\widehat v_{k,2}\cr
&=
k_1^2 |\widehat v_{k,2}|^2+k_2^2 |\widehat v_{k,1}|^2
+ k_2^2 |\widehat v_{k,2}|^2
+ k_1^2 |\widehat v_{k,1}|^2
= |k|^2 |\widehat v_k|^2.
}$$\end{plain
Therefore
$$
|\xi|^2 = L^2\sum_{k\in\J} |\widehat \xi_k|^2=\|v\|^2
\wwords{and}
\|\xi\|^2 =L^2\sum_{k\in\J} |k|^2|\widehat \xi_k|^2 = |Av|^2.
$$
To keep the notation in the present section similar to the notation
appearing in the previous section, we abuse it by extending the
definitions of $B$ and $A$ to the vorticity as
$$
B(v,\xi)=(v\cdot\nabla)\xi
\wwords{and}
A\xi=-\Delta\xi.
$$
Thus, equation \eqref{nset2} may be written as
\begin{equation}\label{zetaeq}
\frac{d\xi}{dt}+\nu A \xi+B(v,W)
+B(v,\xi)+B(U,\xi)=0.
\end{equation}
Equations \eqref{zetaeq} are similar to \eqref{nse1} in structure;
however, there are no cancellations when multiplying by $A\xi$
and integrating over $\Omega$. To bound the resulting terms
we prove
\begin{lemma}\label{lemmaEyes}
Let $U$, $W$, $v$ and $\xi$ be defined as above.
The following bounds hold
\begin{plain}$$\eqalignno{
|(B(v, W), A\xi)|
&\le C\frac{4^2}{3\nu^2}|v|^2\|W\|^3+\frac{\nu}{6}|A\xi|^2,\cr
|(B(v,\xi), A\xi)|
&\le C\frac{4^2}{3\nu^2}|v|^{2} \|\xi\|^3
+ \frac{\nu}{6}|A\xi|^{2},\cr
\noalign{\medskip\noindent and\medskip}
|(B(U, \xi), A\xi)|
&\le C\frac{5^5}{\nu^5}\|U\|^6_{L^\infty}|v|^{2}
+\frac{\nu}{6}|A\xi|^2
}$$\end{plain
for almost every $t\ge t_0$.
\end{lemma}
\begin{proof}
The condition (\ref{vortpose}) applied to both $U$
and $u$ implies that $\|A v\|=|A\xi|$ is finite for almost every
$t\ge t_0$. Our working assumptions further imply that the other norms
appearing in the above bounds
exist everywhere.
For convenience denote
\begin{equation*}
I_1=|(B(v, W), A\xi)|,\qquad
I_2=|(B(v,\xi), A\xi)|\wwords{and}
I_3=|(B(U, \xi), A\xi)|.
\end{equation*}
We now estimate $I_1$, $I_2$ and $I_3$ in turn.
First, estimate $I_1$ using \eqref{Best0}
followed by Agmon's inequality to obtain
\begin{align*}
I_1\le\|v\|_{L^\infty} \|W\| |A\xi|
\le C|v|^{1/2}|Av|^{1/2}\|W\| |A\xi|
=C|v|^{1/2}|Av|^{1/2-\theta}|Av|^{\theta}\|W\| |A\xi|.
\end{align*}
Since $|Av|=\|\xi\|$, we have
\begin{align*}
I_1\le C|v|^{1/2}\|\xi\|^{1/2-\theta}|Av|^{\theta}\|W\| |A\xi|.
\end{align*}
We now use interpolation inequality on $|Av|^{\theta}$
and have $|Av|^{\theta}\le |v|^{\theta/3}|A\xi|^{2\theta/3}$.
This yields
\begin{align*}
I_1\le C|v|^{1/2+\theta/3}\|\xi\|^{1/2-\theta}
|A\xi|^{1+2\theta/3}\|W\|.
\end{align*}
Using Young's inequality with powers $3$ and $3/2$, we have
\begin{align*}
I_1\le C\frac{16}{3\nu^2}|v|^{3/2+\theta}\|\xi\|^{3/2-3\theta}
\|W\|^3+\frac{\nu}{6}|A\xi|^{3/2+\theta}.
\end{align*}
Choose $\theta=\frac{1}{2}$, then we have
\beq\label{estIone}
I_1\le C\frac{16}{3\nu^2}|v|^2\|W\|^3+\frac{\nu}{6}|A\xi|^2.
\eeq
Next, estimate $I_2$ using \eqref{Best0}
and then Agmon's inequality. We have
\begin{align*}
I_2\le C|v|^{1/2} |Av|^{1/2} \|\xi\| |A\xi|
=C|v|^{1/2} |Av|^{3/2-\theta} |Av|^{\theta}|A\xi|.
\end{align*}
Using interpolation on $|Av|^{\theta}$ it follows that
\begin{align*}
I_2\le C|v|^{1/2+\theta/3} |Av|^{3/2-\theta} |A\xi|^{1+2\theta/3}.
\end{align*}
Choosing $\theta=\frac 12$ and then by Young's inequality
with powers $3$ and $3/2$, we have
\beq\label{estItwo}
I_2\le C|v|^{4/6} |Av| |A\xi|^{4/3}
\le C\frac{4^2}{3\nu^2}|v|^{2} \|\xi\|^3
+ \frac{\nu}{6}|A\xi|^{2}.
\eeq
Finally, estimate $I_3$ using \eqref{Best0}. We have
\begin{align*}
I_3&\le \|U\|_{L^\infty}\| \xi\| |A\xi|
=\|U\|_{L^\infty}| Av| |A\xi|\\
& \le \|U\|_{L^\infty}|v|^{1/3}|A\xi|^{5/3}.
\end{align*}
Using Young's inequality with powers $6$ and $6/5$ it follows that
\beq\label{estIthree}
I_3\le C\frac{5^5}{\nu^5}\|U\|^6_{L^\infty}|v|^{2}
+\frac{\nu}{6}|A\xi|^2.
\eeq
\end{proof}
\begin{proof}[Proof of Theorem \ref{mainII}]
Multiplying equations \eqref{zetaeq} by $A\xi$
and integrating over $\Omega$ yields
\begin{equation}
\label{es1}
\frac 12\frac{d}{dt}\|\xi\|^2+\nu |A\xi|^2
+(B(v, W), A\xi)+(B(v,\xi), A\xi)+(B(U, \xi), A\xi)=0.
\end{equation}
We remark that the working assumptions for type-II interpolant
observables imply both
$U$ and $u$ and consequently their difference has the needed
regularity for
the above equation to make sense. These assumptions further provide
{\it a priori\/} bounds on $U$ which are uniform in time.
Although
the corresponding norms of $u$ are finite, we cannot at this point
assume they are uniformly bounded in time.
Under the hypotheses of this theorem, however,
uniform bounds on $u$
can be inferred
as a consequence of this proof
Now, plug the estimates given by Lemma \ref{lemmaEyes} into
\eqref{es1} to obtain
\beq\label{es2}
\frac{d}{dt}\|\xi\|^2+ \nu|A\xi|^2
\le C\left(\frac{1}{\nu^2}\|W\|^3
+\frac{1}{\nu^2}\|\xi\|^3
+ \frac{1}{\nu^5}\|U\|^6_{L^\infty}\right)|v|^{2}.
\eeq
We again point out that $C$ is a non-dimensional constant independent
of $\delta$, $\lambda$ and $h$.
By Poincar\'e's inequality \eqref{poincareDA} we have
\begin{plain}\beq\label{es3}
\frac{d}{dt}\|\xi\|^2+ \lambda_1\nu\|\xi\|^2\le
{C\over \nu^2}\left(\|\xi\|^3+K\right)|v|^{2},
\eeq\end{plain
where the assumption $\|W\|\le\rho_A$ combined
with Agmon's inequality \eqref{agmon} allows us to take
$$K=\rho_A^3(1+ c \nu^{-3}\rho_H^3).$$
Alternatively, one could write $K'=\rho_A(1+ c \nu^{-3}\rho_H^3)$
to obtain
\begin{plain}$$
\frac{d}{dt}\|\xi\|^2+ \lambda_1\nu\|\xi\|^2\le
{C\over \nu^2}\left(\|\xi\|^3+K'|AU|^2\right)|v|^{2},
$$\end{plain
and then estimate the integral of $|AU|^2$ using \eqref{Aint}
as we did in \eqref{me8n}.
Unfortunately, this improvement is dominated by
subsequent estimates on $\|\xi\|$ which are
proportional to $\rho_A$.
Therefore, as the differences are minimal,
we continue with \eqref{es3} for simplicity.
By \eqref{wttn}, we have
\begin{plain}\begin{align}
\label{es4}
\frac{d}{dt}\|\xi\|^2&+ \lambda_1\nu\|\xi\|^2
\le {C\over\nu^2}\left(\|\xi\|^3 +K\right)
e^{\beta(t-t_n)}|v_n|^2.
\end{align}\end{plain
Note that equation \eqref{es4} is
similar to \eqref{me4} except for the additional term involving
$\|\xi\|^3$ on the right.
Fortunately, this term can be controlled for times of size
$\delta$ by our choosing~$h$ small and $\lambda$ large.
This complicates the proof and is the main reason why
the type-I interpolant observables were treated separately
in the previous section.
Continue as in the type-I case. First,
multiply \eqref{es4} by $e^{\lambda_1\nu t}$,
integrate from $t_n$ to $t$
and simplify as in \eqref{me8} to obtain
\begin{align*}
\label{es5}
\|\xi\|^2&\le\|\xi_n\|^2e^{-\lambda_1\nu (t-t_n)}
+
\frac {C}{\nu^2 \beta}\Big(\sup_{s\in[t_n,t)}\|\xi(s)\|^3
+K\Big)e^{\beta(t-t_n)}|v_n|^2.
\end{align*}
When $n=0$ it follows from Proposition \ref{Ebounds} that
$$
|v_0|^2=|E U_0|^2
\le (\lambda\lambda_1)^{-1}(1+\varepsilon) |AU_0|^2
\le (\lambda\lambda_1)^{-1}(1+\varepsilon) \rho_A^2,
$$
and
$$
\|\xi_0\|^2=|A v_0|^2
=|A EU_0|^2
\le (1+\varepsilon) |AU_0|^2
\le (1+\varepsilon) \rho_A^2.
$$
Therefore when $t\in[t_0,t_1)$ we have
\begin{equation}\label{zetaineq}
\|\xi\|^2\le
(1+\varepsilon)\bigg\{
e^{-\lambda_1\nu (t-t_0)}+
\frac C{\lambda\lambda_1\nu^2\beta}
\Big(\sup_{s\in[t_0,t)}\|\xi(s)\|^3
+K\Big)e^{\beta(t-t_0)}\bigg\}
\rho_A^2.
\end{equation}
Let $\delta>0$ be arbitrary and define
$$
\gamma=
(1+\varepsilon)\bigg\{
e^{-\lambda_1\nu \delta}+
\frac C{\lambda\lambda_1\nu^2\beta}
\Big(8\rho_A^3
+K\Big)e^{\beta \delta}\bigg\}.
$$
As in the the proof of Proposition \ref{tildelemma}, since
$$
e^{-\lambda_1\nu\delta}<1\wwords{and}
\frac C{\lambda\lambda_1\nu^2\beta}
\Big(8\rho_A^3
+K\Big)e^{\beta \delta}\to 0\words{as}\lambda\to\infty,
$$
then there is $\lambda$ large enough such that
\begin{equation}\label{boundtwo}
e^{-\lambda_1\nu\delta}+
\frac C{\lambda\lambda_1\nu^2\beta}
\Big(8\rho_A^3
+K\Big)e^{\beta \delta}
<1.
\end{equation}
Furthermore, since $\varepsilon\to0$ as $h\to0$
while holding $\lambda$ fixed,
then there is $h$ small enough such that $1+\varepsilon<2$
and moreover small enough to ensure that $\gamma<1$.
For the choice of $\delta$, $h$ and $\lambda$ given above, let
$$
M= \sup_{s\in [0,\delta]}(1+\varepsilon)\bigg\{
e^{-\lambda_1\nu s}+
\frac C{\lambda\lambda_1\nu^2\beta}
\Big(8\rho_A^3
+K\Big)e^{\beta s}\bigg\},
$$
and note \eqref{boundtwo} along with the
fact that $1+\varepsilon<2$ implies $M<4$.
We claim that $\|\xi(s)\|< 2\rho_A$
for $s\in[t_0,t_1)$.
For contradiction, suppose not.
Since $\|\xi\|$ is continuous on $[t_0,t_1)$ and
$$
\|\xi(t_0)\|=\|\xi_0\|\le (1+\varepsilon)^{1/2}\rho_A <
2^{1/2}\rho_A<2\rho_A,
$$
then this would imply the existence of $t_*\in (t_0,t_1)$
such that
$$\|\xi(t_*)\|=2\rho_A\wwords{and}
\|\xi(s)\|<2\rho_A\words{for} s\in [t_0,t_*).$$
However, if this were true, then inequality \eqref{zetaineq} would imply
$$
\|\xi(t_*)\|^2\le
(1+\varepsilon)\bigg\{
e^{-\lambda_1\nu(t_*-t_0)} +
\frac C{\lambda\lambda_1\nu^2\beta}
\Big(8\rho_A^3
+K\Big)e^{\beta (t_*-t_0)}\bigg\}\rho_A^2
\le M\rho_A^2 < 4\rho_A^2,
$$
which is a contradiction.
Therefore $\|\xi(s)\|< 2\rho_A$ for $s\in[t_0,t_1)$.
Consequently
$$
\sup_{s\in [t_0,t_1)}\|\xi(s)\|^3 \le 8\rho_A^3,
$$
and taking the limit of \eqref{zetaineq} as $t\to t_1$ results in
$\|\tilde \xi_1\|^2\le\gamma \rho_A^2$.
We proceed by induction.
Let $n\ge 1$ and suppose
$$\|\tilde \xi_n\|^2\le \gamma^n\rho_A^2.$$
By Proposition \ref{Ebounds} it follows that
$$
|v_n|^2=|E\tilde v_n|^2
\le (\lambda\lambda_1)^{-1} (1+\varepsilon) |A\tilde v_n|^2
= (\lambda\lambda_1)^{-1} (1+\varepsilon) \|\tilde\xi_n\|^2,
$$
and
$$
\|\xi_n\|^2=|A v_n|^2
=|AE\tilde v_n|^2
\le (1+\varepsilon) |A\tilde v_n|^2
= (1+\varepsilon) \|\tilde\xi_n\|^2,
$$
where $\tilde \xi_n=\curl\tilde v_n$.
Since $1+\varepsilon<2$ we obtain
$$\|\xi(t_n)\|=\|\xi_n\|\le (1+\varepsilon)^{1/2}\|\tilde \xi_n\|
\le 2^{1/2} \gamma^{n/2} \rho_A < 2\rho_A.
$$
Following the same arguments as before, we obtain that
$$
\sup_{s\in[t_n,t_{n+1})} \|\xi(s)\|^3 \le 8 \rho_A^3,
$$
and taking limits as $t\nearrow t_{n+1}$ conclude that
$$\|\tilde \xi_{n+1}\|^2\le\gamma \|\tilde\xi_n\|^2
\le\gamma^{n+1} \rho_A^2,
$$
which completes the induction.
Given $t>0$ choose $n$ such that $t\in[t_n,t_{n+1})$.
It follows that
$$
\|W-w\|^2=\|\xi\|^2\le M \|\tilde \xi_n\|^2
\le M\gamma^n\rho_A^2\to 0\wwords{as} t\to\infty.
$$
Therefore, the same argument used in the proof of Theorem
\ref{mainI} now implies
$$\|W-w\|\to 0\wwords{exponentially as}t\to\infty,$$
and finishes the proof of Theorem \ref{mainII}.
\end{proof}
\section{Conclusions}
In this paper we have shown that spectrally-filtered discrete
data assimilation as described in Definition \ref{algorithm}
results in an approximating solution $u$ that converges to the
reference solution $U$ over time
for any general interpolant observable of type-I or type-II
when $\delta$, $\lambda$ and $h$ are chosen appropriately.
In particular, when observations of $U$ are made using nodal
points of the velocity field, we obtain a type-II interpolant
observable which our analysis is able to handle.
We note that this analysis relies crucially on properties of
the spectral filter and would not have been possible if the
unfiltered interpolants were used instead.
Specifically, our analysis makes use of the fact that
the filtered interpolant $E$ can be made to have norm near
unity when viewed as linear operator on the functional space
implied by the bounds on the original interpolant.
This fact is characterized by the respective inequalities
$$
\|EU\|^2\le (1+\varepsilon)\|U\|^2
\wwords{and}
|AEU|^2\le (1+\varepsilon)|AU|^2,
$$
for the type-I and type-II interpolant
observables given in Proposition \ref{Ebounds}.
Different filtering methods which satisfy similar inequalities
should also be effective.
As a number of advances in practical data assimilation have resulted
from improved filtering, we find
these analytic results to be interesting and relevant.
While it may seem anticlimactic that the technique crucial
for our analysis relies on spectrally projecting the interpolant
observable in Fourier space,
since the linear term
is responsible for the dissipation, it is natural that a spectral basis
with respect to that linearity provides a convenient framework in which
to analyze the synchronization properties of our data assimilation
algorithm. Furthermore, using this basis as a means of spatial filtering
not only has the advantage of being simple, but is intrinsically
compatible with the reliance of our analysis on the dissipation.
Note that the functional dependency of $h$ and $\lambda$ on
$\delta$ and the other physical parameters in the system
appearing in Theorem \ref{mainresult}
depend on knowing an {\it a priori\/} bound $\rho_A$ on the norm
$|AU|$ in terms of those other parameters.
While suitable theoretical bounds appear in the literature,
these bounds are, in general, not sharp compared to
{\it a posteriori\/} bounds obtained through
numerical simulation.
Moreover, the algorithm may continue to work with values of $h$ much larger
and values of $\lambda$ much smaller than required by our analysis.
For example, computational experiments performed by \cite{Gesho2016}
for a different spatially filtered continuous data assimilation method
based on nudging show that the method performs far better
than the analytical estimates suggest.
We conjecture similar numerical effectiveness for the discrete
data assimilation method described in the present paper.
Therefore, we refrain from determining an
explicit theoretical relation between $h$ and the Grashof number
in this work, though such could be obtained from our
analysis, and save such comparisons for the context of a
future numerical study.
\section*{Acknowledgments}
The work of Eric Olson was supported in part
by NSF grant DMS-1418928.
The work of Edriss S. Titi was supported
in part by ONR grant N00014-15-1-2333, the Einstein Stiftung/Foundation - Berlin, through the Einstein Visiting Fellow Program, and by the John Simon Guggenheim Memorial Foundation.
|
\section{Introduction}
A memory access is a highly critical operation. Many decisions inside a program
rely on the correct execution of a memory access. Password checks, signature
verification, or grants to a privileged function they all rely on the genuine
execution of a memory access.
Under normal operating conditions, a memory access reads/writes from/to the
desired location and random malfunctions, e.g., caused by cosmic
radiation~\cite{baumann2005radiation}, are comparably rare. However, the
situation changes dramatically as soon as intentionally induced faults, via
so-called fault attacks, are considered. Here, the attacker modifies the state
of a computing device by, e.g., inducing glitches on the voltage supply or the
clock signal~\cite{DBLP:journals/pieee/Bar-ElCNTW06} or by shooting with a laser
on the chip~\cite{DBLP:conf/cardis/SelmkeBHS15}. Such a fault attack is capable
of skipping instructions~\cite{DBLP:conf/ccs/BreierJC15}, redirecting the memory
access~\cite{derouet2007secure}, or flipping bits in registers or memory leading
to a critical attack vector~\cite{DBLP:conf/cardis/GiraudT04}. While this type
of attack requires local access to the device to induce a fault, more advanced
attacks can even induce faults remotely. For example, the Rowhammer
effect~\cite{DBLP:conf/isca/KimDKFLLWLM14}, which modifies the state of the
memory by frequently accessing neighboring memory cells, can also be induced in
software via Javascript~\cite{DBLP:conf/dimva/GrussMM16} or remotely over a
network interface~\cite{tatar2018throwhammer,lipp2018nethammer}.
While a fault may not directly reveal sensitive information, different
techniques have been developed to exploit faulty computation. For example, it
has been shown that it is possible to deduce the secret key in various
cryptographic algorithms~\cite{DBLP:journals/joc/BonehDL01,
DBLP:journals/jce/AliMT13} solely by analyzing the faulty computation output.
Subsequently, a lot of research has been performed to protect specific
cryptographic algorithms against fault attacks
~\cite{DBLP:conf/cases/BarenghiBKPR10,DBLP:conf/fdtc/RauzyG14}. However, the
hardening of general purpose software against fault attacks is a young research
area.
Two complementary subareas exist. The first subarea deals with the protection
of the executed code. The respective
techniques~\cite{DBLP:journals/compsec/ClercqGUMV17,
DBLP:journals/corr/abs-1802-06691, DBLP:journals/corr/abs-1803-08359} typically
enforce control-flow integrity~(CFI), which is also a well-known mitigation
strategy against software attacks, with fine granularity. The resulting
countermeasures ensure that executed instructions and branches are genuine and
that they are processed in the correct sequence without omission.
The second subarea mainly deals with the protection of data. There, well-known
redundancy-based techniques like arithmetic
codes~\cite{DBLP:journals/tc/Brown60,massey1964survey,DBLP:journals/tit/RaoG71,
DBLP:conf/safecomp/SchiffelSSF10} are utilized. In these schemes, the data is
encoded into a redundant domain, where faults are detectable up to a certain
number of bit flips. Interestingly, while such schemes were initially developed
to protect the data while it is stored in the memory, arithmetic codes also
support to perform certain arithmetic operations on the encoded value.
However, even when mechanisms of the two subareas are combined, i.e., a system
implements a CFI protection mechanism and redundantly encodes the data, memory
transfers from the processor via the memory subsystem are still vulnerable to
fault attacks. Namely, when a fault modifies the address on one of the buses
during the read or write operation, the data is read or written from/to the
wrong memory location, which is not trivially detectable given that the data is
unmodified. Similar effects can be triggered by injecting faults into pointers,
which are typically not prevented by these schemes.
Unfortunately, current extensions to data encoding, which aim to solve this
issue, are very costly and impose severe restrictions on the protected code.
ANB-codes~\cite{DBLP:conf/safecomp/SchiffelSSF10}, for example, introduce a
tremendous runtime overhead of 90\,\% on average on top of already expensive
AN-codes, solely to solve the memory access problem. Furthermore, they can only
protect a limited set of variables with well-known memory alignment and size.
More efficient and less restrictive approaches are needed to protect memory
accesses against address tampering.
\subsection*{Contribution}
In this work, we address the issue of unprotected memory accesses in the context
of fault attacks. We propose a practical solution to detect address tampering in
pointers and on memory buses. Our generic approach works independently of the
used code and data protection schemes and therefore can effectively be combined
with state-of-the-art techniques in the context of hardening general purpose
computing against fault attacks.
In detail, the contributions of this paper are as follows. First, we present a
new approach to protect pointers against faults with negligible overhead in
terms of runtime and storage requirements. We encode pointers using a
multi-residue arithmetic code, which allows us to detect faults on encoded
pointers during both storage and computation. The redundancy information of the
code word is hereby stored in the unused upper bits of a pointer to fully
utilize the available register space and yield zero-overhead for storing an
encoded pointer. Furthermore, by transforming the pointer arithmetic into the
encoded multi-residue domain, the protection of the pointer is maintained also
when performing arithmetic operations on the pointer; e.g., when adding an
offset to the stack pointer.
Second, we propose an efficient way to protect memory accesses from tampering by
linking the stored data in memory with the address of the access. We establish
this link whenever data is written to the memory and remove the link as soon as
the data is read back into the processor. When considering fault attacks,
countermeasures like data encoding are already necessarily employed. By linking
the redundant address information with the encoded data, faults during
addressing manifest as errors in the redundantly encoded data, where they can be
detected. As the result, data integrity checks implicitly also checks for
address tampering and make explicit addressing error checks unnecessary.
Finally, to evaluate the concept, we integrated our protection mechanism into an
FPGA hardware implementation of an open-source \mbox{RISC-V}\xspace processor. Furthermore, to
avoid tedious manual encoding of all pointers and addresses inside the program,
we integrated this concept directly into a LLVM-based C compiler, which is
capable of automatically protecting complex codebases without manual
interference. The resulting prototype induces 10\,\% code size and less than
7\,\% runtime overhead on average.
\subsection*{Outline}
The remainder of this paper is structured as follows. \autoref{sec:Background}
discusses the threat model and the attack vector, gives an introduction to
arithmetic codes, and discusses related work. In \autoref{sec:pointer}, we
describe how we protect pointers against fault attacks. The approach to link the
pointer protection with data encoding is presented in~\autoref{sec:memory}.
\autoref{sec:architecture} details how we extend the \mbox{RISC-V}\xspace instruction set to
support encoded pointers and discusses our compiler modifications. Finally,
\autoref{sec:evaluation} evaluates the overhead and \autoref{sec:conclusion}
concludes this work.
\section{State of the Art and Background}
\label{sec:Background}
In this section, we first describe the attack vector and the threat model we
consider. Furthermore, we present state-of-the-art methods of error detection
codes, which we use to protect a memory access against tampering efficiently. We
also show related concepts, which aim to secure pointers or a memory access in
general.
\subsection{Threat Model and Attack Vector}
\label{sec:fault_model}
For this work, we assume a powerful attacker, which performs fault attacks in
order to compromise a system. Faults can be induced into instructions and data
at various places like, for example, in registers, during computation in the
ALU, on buses, and in memory. Many of these attack vectors can be covered by
existing and established countermeasures, which we assume to be in place.
Namely, CFI-based fault
countermeasures~\cite{DBLP:journals/compsec/ClercqGUMV17,
DBLP:journals/corr/abs-1802-06691, DBLP:journals/corr/abs-1803-08359}, which
enforce the authenticity of instructions as well as their execution sequence,
can be used to protect code against faults. Furthermore, such a CFI scheme
already protects function pointers, which do not require further protection.
Data, on the other hand, can be protected during computation and storage using
data encoding techniques like, for example, arithmetic
codes~\cite{DBLP:journals/tc/Brown60,massey1964survey,DBLP:journals/tit/RaoG71,
DBLP:conf/safecomp/SchiffelSSF10}. However, as soon as data is transferred via
a memory bus these codes are insufficient. Namely, while the value itself is
protected via the code, the corresponding address information remains
vulnerable. Furthermore, pointers as such, typically, also remain unprotected by
the data encoding schemes considering that eventually the plain value of the
pointer is used to address the memory.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{attack_vector}
\caption{Attack vector: Modified pointers and manipulated memory accesses.}
\label{fig:attack_vector}
\end{figure}
To illustrate the problem, \autoref{fig:attack_vector} visualizes a simple
memory access. On the left side there is the pointer used for a memory access,
on the right side there is the memory, and the arrow in between denotes the
memory access. The data in the memory is redundantly encoded denoted by the
\textit{c}-subscript of the variables. Originally, the pointer $p$ points to the
address 0x1001 to readout the value false\textsubscript{c} from the memory.
However, a fault can manipulate the memory access to readout a wrong value. In
particular, there are two error sources, which can lead to a wrong memory
access. First, the attacker can modify the pointer as shown in the middle
example in \autoref{fig:attack_vector}. If a pointer gets modified, then all
subsequent memory accesses lead to a wrong location. An attacker could, e.g.,
modify two pointers used for a signature comparison to point to the same
location, which always bypasses the memory comparison. This can occur anywhere
in the program, also during pointer arithmetic. The second source of a
manipulated memory access is the memory operation itself. When assuming the
pointer is correct and not manipulated, the memory access can still be
manipulated. A fault on the address bus can redirect the memory access to a
wrong location as indicated in the third example.
Both of these attack vectors can lead to a wrong memory access. Today, there is
no efficient way to protect them leaving frequently used memory operations
completely unprotected against fault attacks.
\subsection{Error Detection Codes}
\label{sec:codes}
Error detection codes~\cite{peterson1961codes} are a well-known and well-studied
concept to detect errors during storage or transmission. However, different
types of code have been developed, which also support the computation on encoded
data. Logical operations, for example, can directly be computed in the encoded
domain when binary linear codes~\cite{hamming1950error} are used. Arithmetic
codes, on the other hand, can be used when primarily arithmetic operations have
to be performed on encoded data.
\subsubsection{AN(B)-Codes.}
AN-codes~\cite{DBLP:journals/tc/Brown60,forin1990vital} are an example for such
an arithmetic code and
are defined by multiplying the functional value $x$ with the encoding constant
$A$: $x_c = x \cdot A$. Therefore, all code words are multiples of the encoding
constant $A$ and every value in between corresponds to an invalid code word. To
check if a code word is valid, a modulo operation with the encoding constant is
performed, which must yield zero. Decoding is done by using an integer division
with the encoding constant. Because of multiplying the functional value with the
encoding constant, it cannot be separated from the redundancy part, thus the
name \textit{non-separable} code. The encoding constant $A$ defines the error
detection capabilities. Finding a good encoding constant is not easy and
currently only possible via exhaustive search~\cite{DBLP:conf/wisa/MedwedS09}.
Research already found suitable encoding constants, which maximize the error
detection capabilities, so-called \textit{Super
$A$}~\cite{DBLP:conf/hase/HoffmannUDSLS14}. To maintain the error detection
capabilities, the functional value needs to be less than the encoding constant,
which limits the general-purpose application of this code in real-world
applications. Furthermore, this type of code does not protect the address in a
memory access.
Forin and Schiffel et al.~\cite{forin1990vital,DBLP:conf/safecomp/SchiffelSSF10}
extend AN-codes by assigning a variable dependent signature $B_x$ to each
encoded variable $x_c$. This yields the encoding formula $x_c = A \cdot x + B_x$
with $B_x < A$. By adding the variable dependent signature $B_x$ to the AN-code,
the AN-code property that all encoded values are a multiple of $A$ is
intentionally destroyed. Since $B_x$ is less than $A$, decoding works the same
as for normal AN-codes using an integer division. A check is also performed
using a modulo operation with the encoding constant, which now must yield the
signature $B_x$. A compiler keeps track of all assigned signatures and is able
to insert checks for the modified ANB code words. By assigning a
variable-dependent signature to the code words, a wrong memory access can be
detected, as long as signatures do not cancel out due to arithmetic. However,
this approach turns out to be complicated in practice. On the one hand,
ANB-codes have a performance penalty of 90\,\% on average on top of AN-codes. On
the other hand, the value of the signature $B_x$ needs to be less than the
encoding constant which limits the number of variables to be encoded.
\subsubsection{Residue Codes.}
A different class of arithmetic codes are residue codes~\cite{massey1964survey}.
Here, a residue code word $x_c$ is defined by the tuple $x_c = (x, r_x = M |
x)$, where $x$ denotes the functional value and $r_x$ the residue. The residue
$r_x$ is hereby computed as the remainder $M | x$ with respect to the code
modulus $M$. Residue codes separate the redundancy part from the functional
value $x$ and therefore are called \textit{separable} codes. Although the
modulus $M$ defines the robustness of the code, ordinary residue codes only
guarantee the detection of a single bit flip, because a single bit flip on the
data and on the residue is enough to construct a new, valid code word (e.g., the
Hamming distance between the $0_c$ and $1_c$ is two, where both values denote
a residue encoding with an arbitrary modulus $M$).
In order to overcome this limitation and to scale the robustness of the code,
the redundancy part can be increased by using more than one
residue~\cite{DBLP:journals/tc/Rao70,DBLP:journals/tit/RaoG71}, yielding a
multi-residue code. The modulus $M$ is now defined by $M = \left\{m_0, \dotsc,
m_n \right\}$, where $n$ is the number of residues.
Finding the set of moduli is not easy. Although finding good moduli requires
exhaustive search~\cite{DBLP:conf/wisa/MedwedS09}, the moduli selection for
multi-residue codes can be done faster than for AN-codes. Residue codes, in
general, are arithmetic codes and therefore also support certain arithmetic
operations. Here, the operation is performed on the functional part and on the
residues independently. \autoref{eq:res_add} shows how an addition works for two
multi-residue encoded values. The addition is performed on the functional value
and on the residues independently. On the residues, the addition is performed
followed by a modular reduction using on the moduli $m_i$ for the
i\textsuperscript{th} residue.
\begin{equation}
z_c = x_c + y_c = \left(x + y , \forall i: m_i | \left(r_{i,x} + r_{i,y}\right) \right)
\label{eq:res_add}
\end{equation}
Similar to the addition, residue codes also support subtractions and
multiplications. However, in this work, we only use additions and subtractions.
\subsection{ARM Pointer Authentication}
Protecting pointers against tampering is not only relevant in the context of
fault attacks but is also used to counteract software attacks. For example, ARM
added a feature called pointer
authentication~(PAC)~\cite{arm2017pointerauthentication} to the \textit{ARM
v8.3} instruction set with the goal of protecting special pointers. Several new
instructions were added to the architecture that permits to cryptographically
authenticate special pointer values in registers, like the return address in the
link register, using a message authentication code~(MAC). PAC tags have a size
between 3 and 31 bits, depending on the processor configuration, and are, as in
our work, directly embedded into the protected pointer.
Note, however, that even though the general approach is similar to our work, the
provided capabilities and the resulting protection is vastly different. PAC aims
to only protect special pointers against software attacks. In PAC, authenticated
pointers cannot be be protected during pointer arithmetic since there is no
homomorphism for the MAC. Furthermore, PAC only aims to protect the pointer. The
memory access, which uses an authenticated pointer is completely unprotected
and there are no protection mechanisms to ensure that the accessed memory
actually originates from the correct address.
\section{Pointer Protection with Residue Codes}
\label{sec:pointer}
Manipulation of a memory access is possible by attacking two different parts of
the access. The first one is the pointer itself, which is used to perform the
memory access. This section details how we use multi-residue codes to protect
every data pointer inside a program against fault attacks. Furthermore, we present
how to integrate the multi-residue code into our pointer representation and
elaborate on the additionally needed hardware support.
\subsection{Overview}
Pointers are ubiquitous. Every memory access, e.g., accessing a variable on the
stack, uses a pointer to address the memory. However, when considering fault
attacks, pointers may be manipulated to point to a different memory location.
To counteract this threat, we encode all pointers to a redundant representation,
where faults are detectable. As presented previously, there are two classes of
suitable codes: \textit{separable} and \textit{non-separable} codes, which have
similar properties in terms of error detection capabilities and support for
computation. However, a separable code has advantages to protect a pointer.
Namely, it supports direct access to the functional value and can therefore
immediately be used to address memory. On the other hand, using a non-separable
code to protect the pointer requires to perform a potentially expensive decoding
operation before the actual address is available. AN-codes, as an example for
non-separable codes, require a costly integer division during the decoding
operation. Hence, this division would be required for every memory access.
We encode pointers using a \textit{separable} multi-residue code with a scalable
number of moduli. Here, an encoded pointer $p_c$ is denoted as a tuple $(p,
r_p)$, where $p$ is the original value of the pointer and $r_p$ denotes the
redundancy part comprising the residues of $p$ given a moduli set $M$. Using a
multi-residue code to protect the pointer gives two advantages. On the one hand,
the strength of the code, i.e., the number of bit flips which are detectable, is
scalable with the number of residues. On the other hand, residue codes are
arithmetic codes and therefore also support arithmetic instructions, like
addition and subtraction, natively. This allows us to perform pointer
arithmetic, for example, the stack pointer manipulation in function prologues
and epilogues, directly inside the encoded domain without decoding the pointer.
\subsection{Pointer Layout and Residue-Code Selection}
Adding separable redundancy to data implies that the additional information
needs to be stored somewhere in order to provide a value. In the context of
protecting a processor register, various possibilities exist to provide this
storage.
For example, an additional parallel register file can be added to the processor,
which only holds the redundancy part and gets updated in lockstep with the
actual values~\cite{DBLP:conf/date/MedwedM11}. However, this approach is quite
costly for our use case considering that only a small number of registers
typically hold pointers at a certain point in time. Alternatively, pairs of
regular registers can be used to store the data and its redundancy.
Unfortunately, doing so increases the register pressure and lowers the overall
performance. Moreover, without adding costly access ports to the register file,
multiple instructions have to be performed on every pointer operation, even for
simple ones like an increment. Finally, at least for modern RISC instruction set
architectures~(ISA)~\cite{Waterman14therisc-v}, adding additional operands into
the instruction encoding is difficult without increasing the instruction size.
In this work, we therefore went with a different approach and stored the
redundancy information directly into the upper bits of the pointer. Similar to
PAC, i.e., ARM's pointer authentication feature, this approach introduces zero
overhead in terms of storage for the redundancy at the cost of some bits of
address space. Additionally, this dense representation of an encoded pointer
allows us to add new combined residue arithmetic instructions, which operate on
the functional value and on the residues in parallel, rather than requiring
separate instructions to handle both. By storing the redundant pointer in one
register, we can, therefore, use the same instruction format as regular
instructions and do not require extensive modifications of the ISA or hardware
to maintain performance.
Considering that the directly accessible address space is limited, embedding the
residues into the pointer works best for modern 64-bit architectures. Therefore,
the following design considerations as well as our prototype, that is presented
in \autoref{sec:architecture}, is built upon such an architecture. The overall
concept can still be applied to 32-bit architectures with reduced error
detection capabilities or via a different storage option.
\paragraph{Parameter Selection.}
When selecting the parameters of an error detecting code, it is always a
trade-off between error detection capabilities and the overhead introduced by
the code. However, since the functional value including the redundancy is stored
in a single register, also the remaining address space has to be considered. For
our prototype, we focus on a 64-bit architecture and partition our pointers into
24-bit redundancy and a 40-bit functional value. The resulting pointers can
still address one terabyte of memory, which is sufficient for most applications.
As a concrete code, we instantiate a multi-residue code with the moduli set $M =
\left\{5, 7, 17, 31, 127\right\}$, which is an extension to the one presented
in~\cite{DBLP:conf/date/MedwedM11}. This moduli set yields a code with a Hamming
distance of $D=5$ and is capable of detecting up to four bit flips in the
encoded 64-bit pointer value. Storing the residues for these moduli requires a
total of 23 bits, i.e, $3,3,5,5,7$ bits, respectively. The last remaining bit is
used as a tag bit and specifies if data accessed via the pointer have to perform
data linking/unlinking as presented later in \autoref{sec:mmio}. The resulting
register layout of such an encoded pointer is shown in
\autoref{fig:encoded_ptr_mmio}.
\newcommand{\fakesixtyfourbitsWithRes}[1]{
\tiny
\ifnum#1=1234567890
#1
\fi
\ifnum#1=0
0
\fi
\ifnum#1=19
39
\fi
\ifnum#1=20
40
\fi
\ifnum#1=23
43
\fi
\ifnum#1=26
46
\fi
\ifnum#1=30
51
\fi
\ifnum#1=34
55
\fi
\ifnum#1=39
63
\fi
}
\begin{figure}
\centering
\begin{bytefield}[
bitwidth=\widthof{\rotatebox{90}{\tiny MMIO}~},
boxformatting={\centering},
bitformatting=\fakesixtyfourbitsWithRes,
endianness=big]{40}
\bitheader{39,34,30,26,23,20,19,0} \\
\bitbox{5}{$r_{4,p}$} &
\bitbox{4}{$r_{3,p}$} &
\bitbox{4}{$r_{2,p}$} &
\bitbox{3}{$r_{1,p}$} &
\bitbox{3}{$r_{0,p}$} &
\bitbox{1}{\rotatebox{90}{\tiny MMIO}} &
\bitbox{20}{$p$}
\end{bytefield}
\caption{Encoded pointer representation. The actual 40-bit pointer value $p$,
the MMIO tag bit, and 23 bits of redundancy $r_p$ comprise an encoded
64-bit pointer.}
\label{fig:encoded_ptr_mmio}
\end{figure}
\subsection{Pointer Operations}
\label{sec:pointer_operations}
Pointers are not only used to perform a memory access but also are used to
perform pointer arithmetic. To maintain good performance, it is therefore vital
that the encoded pointers support these computations as efficiently as possible.
Notably, as the term pointer arithmetic already hints, arithmetic computations,
like addition and subtraction, are the most common operations that are performed
on pointers. For example, accessing larger sequential memory chunks via a
pointer involves a large number of additions between the pointer and the access
stride in a loop. Similarly, next to every function call, the respective stack
frame size is added and subtracted to/from the stack pointer in the function's
prologue and epilogue. Precisely these types of operations are natively
supported by the used multi-residue code and can therefore be performed in the
encoded domain.
On the other hand, more work is required for operations that are not directly
supported by the multi-residue code. The simplest approach is probably to
perform the operation on the plain functional value only and restore the
encoding afterward. To ensure the correctness of the computation, then
additional measures like replication have to be used. Alternatively, such
operations can be performed by first converting the pointer to a different code,
in which the computations are straightforward, followed by converting the
differently encoded result back into multi-residue representation. Still, such
operations comprise only a very small number of pointer operations compared to
arithmetic operations.
\paragraph{Software vs. Hardware.}
In a multi-residue code, the addition operation is performed on the functional
value and on all its residues. This operation can be executed in hardware or
software. However, performing this operation in software is challenging, as it
involves a modulo reduction for each residue.
Looking only at a single modulo, there exist several options for implementing
the reduction in software:
First, a normal modulo instruction from the ISA can be used. Although such an
instruction does not have much code overhead, a modulo operation involves a
costly integer division which usually takes multiple clock cycles to finish.
Second, instead of a modulo operation, a conditional subtraction can be used for
the modular reduction. Third, there are optimized modulo algorithms
available~\cite{jones2001modulus}, but their overhead is still large. A single
modular addition with an optimized reduction with the modulus five takes at
least 18 instructions on our \mbox{RISC-V}\xspace target architecture.
Considering that the runtime of these solutions additionally has to be
multiplied with the number of used residues makes a software solution even less
attractive. Furthermore, even if the performance penalty is acceptable,
additional registers have to be reserved for implementing the reduction
functionality. Summarizing, a software-based approach to perform residue
operations, while feasible, is not very practical. Therefore, hardware based
approaches to implement the residue operations have been investigated.
In particular, in our prototype, we add new instructions that permit to perform
addition and subtraction of multi-residue encoded pointers.
\autoref{sec:hw_instructions} discusses the new instructions in detail with the
focus on the target architecture. Furthermore, an instruction for performing the
expensive encoding operation is added, which computes the modulus for each
residue. For convenience reasons, also a dedicated decoding operation is added
to the ISA.
\section{Evolved Memory Access Protection}
\label{sec:memory}
Apart from faulting the pointer, the second source to manipulate a memory access
is the memory operation itself. If the attacker is able to induce faults on the
address bus, the memory access can be redirected to a different location. In
this section, we present a method to link the data with its respective address,
where addressing errors are transformed into data errors which can subsequently
be detected using a data-protection scheme.
\subsection{Overview}
In order to be able to detect address tampering, a way to uniquely identify
incorrectly accessed memory is needed. A common approach to establish this link
between the data and the address is augmenting the data-protection scheme, which
is anyway needed to protect the data against faults.
For example, ANB-codes embed the identity of the variable, in the form of a
unique residue $B_x$, into a required underlying AN-code based data encoding.
However, this approach has several drawbacks. For example, working on variable
granularity requires concise data-flow information, which is in real-world
applications hard to acquire for arbitrary memory operations, and limits the
applicability of the approach. Furthermore, maintaining these identities during
calculation is quite costly. Finally, the approach is strongly linked with
AN-codes and cannot easily be applied to other data-protection schemes.
Our scheme, takes an entirely different approach to prevent address tampering.
Instead of constructively embedding the address of the data into the
data-protection code, our scheme destructively overlays data that is written to
the memory with the respective memory address. As a result, addressing errors
are transformed into data errors that get detectable as soon as the overlay is
removed again.
In more detail, before data is written from a register to the memory bus by the
processor, the data gets encoded with respect to the target address.
Conceptually, this kind of linking is similar to encrypting the data in an
address dependent way. However, since we do not strive for confidentiality with
our approach, the use of a cryptographically secure cipher is not needed. The
resulting encoded data is then simply stored into memory like in a regular
system.
When data is read back from memory into a processor register, the decoding with
respect to the target address is performed. Considering that the performed
decoding operation is the inverse of the encoding, a genuine data value is
restored only when the read has been performed from the correct address.
Otherwise, an incorrect data value is generated which can be detected via the
used data-protection scheme. Note that the detection of address tampering during
memory writes is possible like this as well. However, the detection is delayed
to the point where the incorrectly written value is read back into the
processor.
\subsection{The Linking Approach}
As already mentioned, the general idea behind our memory access protection
approach is to link the data that is stored in memory with its respective
address. Instead of directly writing a register value $D_{Reg}$ into memory at
the a certain address $p$ (i.e.,~$\mathtt{mem}\left[p\right] = D_{Reg}$), a
little more work has to be performed in our scheme. Namely, as shown in
\autoref{eq:correction}, the linking function $l$ has to be evaluated in order
to determine the value that is actually written to the memory at address $p$.
\begin{equation}
\mathtt{mem}\left[p\right] = l\left(p, D_{Reg}\right) = l_p\left(D_{Reg}\right)
\label{eq:correction}
\end{equation}
The purpose of this linking function is to combine the address $p$ with the data
value $D_{Reg}$. However, not every function can be used for this purpose. At
the very least, the following two requirements have to be fulfilled in this
context. First, for each address $p$, the linking function $l_p$ has to be a
permutation. Having this property means that $l_p$ performs a bijective mapping
and that an inverse function $l_p^{-1}$ exists, i.e.,~$\forall p, D_{Reg}
\rightarrow l_p^{-1}\left(l_p\left(D_{Reg}\right)\right) = D_{Reg}$.
Subsequently, memory read operations can be implemented using this inverse
function as shown in \autoref{eq:inv_correction}. As the result, from the
software perspective, encoding data when storing to memory and decoding data
when loading from memory is completely transparent, yields the expected result,
and can be performed for every memory access.
\begin{equation}
D_{Reg} = l^{-1}\left(p, \mathtt{mem}\left[p\right]\right) = l_p^{-1}\left(\mathtt{mem}\left[p\right]\right)
\label{eq:inv_correction}
\end{equation}
Second, to ensure that addressing faults are detectable, data encoded under one
address should yield a modified value when being decoded under a different
address, i.e.,~$\forall p, p', D_{Reg}: p \neq p' \rightarrow
l_{p'}^{-1}\left(l_p\left(D_{Reg}\right)\right) \neq D_{Reg}$. Note,
furthermore, that the modified value should not be a valid code word in terms of
the used data-protection code.
\paragraph{Function Selection.}
Various functions, like for example cryptographic ciphers, fulfill these
requirements and are therefore suitable to link the data and the address
information as required by the memory access protection scheme. However, given
that we aim for a low-overhead design, less resource demanding functions have
been investigated.
Interestingly, already a simply xor operation, as shown in
\autoref{eq:f_xor_enc} and \autoref{eq:f_xor_dec}, is sufficient as the linking
function for our use case. In more detail, in our scheme, addresses are encoded
using arithmetic multi-residue codes and the data encoding can be selected
arbitrarily. On the one hand, when the same multi-residue code is also used for
the data, e.g., an encoded pointer is written to memory, using the xor operation
is good choice given that that multi-residue codes are not closed under the xor
operation. Subsequently, it is also unlikely that combining multiple valid code
words yields a valid result and therefore facilitates error detection. On the
other hand, even when a data protection code which is closed under the xor
operation is used, still similar error detection capabilities are expected.
After all, combining code words from different codes is highly unlikely to yield
sensible results.
\begin{align}
\mathtt{mem}\left[p\right] &= p \oplus D_{Reg} \label{eq:f_xor_enc} \\
D_{Reg} &= p \oplus \mathtt{mem}\left[p\right] \label{eq:f_xor_dec}
\end{align}
\paragraph{Linking Granularity.}
Theoretically, the previously described linking approach can be applied with
arbitrarily granularity. Therefore, applying the technique on the
processor's native word size, e.g., 64-bit in our prototype, may appear
natural. However, performing xor-based linking on such a coarse granularity does
not yield the desired amount of diffusion. Namely, bytes that are close to each
other, i.e., with a stride of 8 bytes when operating on 64-bit, are highly
likely to have the same address pad. Furthermore, in many real-world
applications, also misaligned data accesses with arbitrary size have to be
supported efficiently. Situations like this, for example, commonly arise when
arbitrarily aligned data is copied via the \textit{memcpy} function.
Therefore, to fix the problem of the low diffusion and the arbitrarily aligned
data accesses, we perform the linking of data and address with byte-wise
granularity. Each byte, even when it is part of a larger memory transfer, is
independently linked with its respective address. Hereby, each individual
byte-address pointer is still multi-residue encoded to provide the desired
diffusion during linking. Furthermore, the actual linking is again performed via
an xor similar to \autoref{eq:f_xor_enc}. However, considering that the data and
its address have different bit sizes, an additional compression is applied on
the address before linking. Namely, each 64-bit address $p = [p_0, p_1, \ldots,
p_7]$ gets reduced to an one byte value $p'$ by xor-ing the individual address
bytes as shown in \autoref{eq:ptr_reduce}.
\begin{equation}
p' = \bigoplus_{i=0}^7 p_i \label{eq:ptr_reduce}
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[width=1\columnwidth]{ptr_reduced}
\caption{Byte-wise data linking of a 64-bit word. Each byte gets xored with
its respective xor-reduced encoded address.}
\label{fig:data_corr_ptr_reduced}
\end{figure}
Applying this approach to a full 64-bit word is visualized in
\autoref{fig:data_corr_ptr_reduced}. Considering the number of needed
multi-residue operations for such a word-sized access, using this linking scheme
effectively requires hardware support. In this work, we therefore integrated the
needed transformations directly into special load and store instructions.
From the software perspective, encoding data when storing to memory and decoding
data when loading from memory is completely transparent and can be performed for
next to every memory access.
\subsection{Memory-Mapped I/O}
\label{sec:mmio}
Memory-mapped I/O~(MMIO) is a common communication interface in embedded
processors to access peripherals. In MMIO, the peripheral registers are mapped
into the standard memory layout of the processor. This allows the CPU to use
ordinary load and store instructions to access the peripheral.
However, in order to protect the memory access, our architecture uses redundant
pointers and links them with the data before executing the memory access. Since
a standard memory-mapped peripheral is not aware of this data linking, wrong
data would be written to the device. Therefore, we cannot apply data linking
when accessing a memory-mapped peripheral. However, we still can use an encoded
pointer to access the memory-mapped peripheral as this does not influence the
data. In order to use an encoded pointer but not perform the data linking, we
would need special instructions for load and store for this purpose. We avoid
this overhead by encoding this information directly into the encoded pointer.
The load and store instructions detect this and do not perform the data linking
accordingly.
As shown in~\autoref{sec:pointer}, we redundantly encode the pointer using a
multi-residue code. In \autoref{fig:encoded_ptr_mmio}, we show the pointer
layout where the 41\textsuperscript{st} MMIO-bit indicates whether the pointer
is for an MMIO access without data linking. The residues, which form the
redundancy of the pointer, are computed over the 40-bit functional pointer value
and the MMIO-bit to protect both against tampering.
\section{Architecture}
\label{sec:architecture}
The concept of protected pointers and linked memory accesses is integrated in a
prototype implementation based on a 64-bit \mbox{RISC-V}\xspace architecture. In this section
we first discuss the new instructions, show how we integrated them into the
architecture, and finally show a compiler prototype to automatically protect all
memory accesses in a program.
\subsection{New Instructions}
\label{sec:hw_instructions}
As previously described, it requires hardware support to efficiently perform the
residue arithmetic such that the performance penalty is acceptable. In this
work, we extend the instruction set of the processor with instructions that
operate in the encoded residue domain. In particular, the following custom
instructions are added to the instruction set.
\paragraph{\textit{renc}, \textit{rdec}.}
To efficiently encode a value into the multi-residue domain, a dedicated
encoding instruction (\textit{renc}) is added. The encoding operation computes
the residues over the 41-bit functional value of the pointer, which also
includes the \textit{MMIO}-bit in the protection domain. As a second
instruction, we add support to decode a multi-residue encoded register
(\textit{rdec}). Both instructions are idempotent, meaning they can repeatedly
be executed (encoding an already encoded value does not change the encoding).
\paragraph{\textit{radd}, \textit{raddi}, \textit{rsub}.}
To support pointer arithmetic on encoded pointers, hardware support for the most
commonly used operations is added. Concretely, we support adding two
multi-residue encoded register values (\textit{radd}), adding a multi-residue
encoded value to an immediate value (\textit{raddi}), and subtracting
multi-residue encoded values (\textit{rsub}). The immediate value in the
\textit{raddi} instruction is not yet multi-residue encoded. However, these
values are part of the instruction encoding and are already protected via the
CFI code protection scheme. Note that, before the immediate can be used in a
residue operation, it gets encoded as part of the instruction execution.
\paragraph{\textit{rlxck}, \textit{rsxck}.}
Since we now use encoded pointers and require data linking/unlinking, dedicated
memory instructions are added to the ISA. Therefore, a family of new load
(\textit{rlxck}) and store (\textit{rsxck}) instructions is added. Herby, the
\textit{x} denotes the access granularity of the memory operation. Concretely,
we support byte (\textit{b}), half-word (\textit{h}), word (\textit{w}), and
double word (\textit{d}) accesses with and without sign extension, which
corresponds to the original memory access instructions in the \mbox{RISC-V}\xspace 64-bit ISA.
The new instructions have the same operand interface as the original load and
store instructions of \mbox{RISC-V}\xspace. However, they now take an encoded pointer for
addressing the memory. The memory instructions also contain a plain immediate
value to add an offset to the pointer, which is protected by the CFI code
protection. Furthermore, these instructions perform the data linking and
unlinking on a byte-wise granularity. However, if the
40\textsuperscript{th}-bit, the \textit{MMIO} bit, is set to one, no data
linking and unlinking is performed, which allows us to use a protected pointer
when accessing a memory location, which does not support data linking, e.g., a
memory-mapped peripheral.
Since every memory access is replaced with its protected counterpart, the
protection mechanism could already be implemented in the original load and store
instructions of the processor. However, for the sake of still supporting the
original \mbox{RISC-V}\xspace instructions, they are left unmodified, and new instructions are
added separately.
\subsection{Hardware}
The instruction set is only one part of our protected architecture. We also
implemented the modified instruction set in hardware. As foundation, the
open-source 32-bit \mbox{RISC-V}\xspace core \textit{RI5CY\xspace}~\cite{pulp2018riscy} is used.
This core is extended to a 64-bit processor meaning that the register file,
datapath, and load-and-store unit are modified and all necessary instructions
are added to be compliant with the \mbox{RISC-V}\xspace RV64IM instruction set.
Furthermore, we added the new instructions to deal with multi-residue encoded
pointers, as defined in \autoref{sec:hw_instructions}.
\autoref{fig:residue_pipeline} shows the modified processor pipeline, which
includes a dedicated arithmetic logical unit~(ALU) for residue operations.
Furthermore, immediate values, which are part of the instruction, get encoded
during the instruction decode stage of the processor pipeline. The
load-and-store unit is extended to support data linking and unlinking to protect
all memory accesses.
\begin{figure}[t]
\centering
\includegraphics[width=1\columnwidth]{frankenstein}
\caption{Modified processor pipeline. The instruction decode stage is extended
with a 12-bit residue encoder, the execution stage with a residue
ALU, and the write-back stage with a pointer-reduction data-linking
unit.}
\label{fig:residue_pipeline}
\end{figure}
The new residue ALU is shown in detail in \autoref{fig:residue_alu}. The ALU
supports encoding and decoding of values to/from the multi-residue domain as
well as adding and subtracting two encoded values. The design of the ALU is
optimized to require only one residue adder and one encoder in the execution
stage of the processor. Decoding is for free since it only requires rewiring,
where the upper bits are set to zero. After performing an addition, the
functional value of the adder result is re-encoded and compared with the
independently computed residues in order to perform error-checking after each
residue instruction. If the computed residues and the newly re-encoded residues
mismatch, a redundant error signal is generated to force the processor into a
safe state. Since this adder is also used for computing the final pointer
address during a memory access (the encoded immediate value is added to the
encoded base pointer), every pointer is also checked before performing a memory
access. With frequent checks for every result, we minimize the probability that
error masking occurs and errors are not detectable anymore.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{residue_alu}
\caption{Residue ALU with a 41-bit adder and a shared residue encoder. The
addition result is automatically checked after the operation by
re-encoding the result and comparing it with the computed residues
and generating a redundant error signal.}
\label{fig:residue_alu}
\end{figure}
Currently, the residue encoder uses special algorithms
from~\cite{DBLP:journals/vlsisp/PerssonB09} to encode data. However, the residue
adder is implemented without any further optimizations. By using optimized
arithmetic operations, e.g., the one from~\cite{DBLP:conf/arith/Zimmermann99},
the hardware overhead can be further reduced.
\subsection{Software}
To make the countermeasure practical and protect every memory access in the
program, the new instructions and the protection mechanism also need to be
integrated into the compiler. In the following, we integrate our countermeasure
to the LLVM-based C compiler~\cite{DBLP:conf/cgo/LattnerA04}.
An LLVM-based compiler is partitioned into three parts, the front end, the
middle end, and the back end. While the middle end optimizes
target-independently on an intermediate code representation, the back end
transforms the universal intermediate representation to a target-dependent code.
To protect every memory access in the program, the countermeasure needs to be
inserted in the back end stage of the compiler. Any earlier transformation can
potentially miss a memory access leaving some accesses possibly unprotected
(e.g., the stack is created in the target-dependent part of the back end). Even
in the back end, the protection happens right before the final instruction
scheduling.
LLVM's back end uses a directed acyclic graph~(DAG) representation, the
\textit{Selection DAG}, for the code generation. The intermediate representation
is transformed in a series of steps to finally emit the machine code. However,
the back end has no information about pointers and addresses. Therefore, this
information is created and propagated manually on the Selection DAG. Dedicated
pointer nodes are added to the Selection DAG where pointers are created, e.g.,
when creating a \textit{FrameIndex} node used for a local stack memory access.
This information is then propagated on the Selection DAG and all dependent
operations are replaced with their corresponding residue counterpart. If we
obtain an instruction, which is not supported by the residue code, the pointer
is decoded, the operation is performed in the unencoded domain, and then, the
pointer is re-encoded. However, this sequence of instructions is not used in the
majority of the transformations. Finally, protected load and store instructions
are emitted, which use an encoded pointer for addressing the memory.
If the program uses a constant address, e.g., the address of a global variable,
this information needs to be encoded to the multi-residue domain. However, the
compiler does not have this information yet. Therefore, it creates a relocation
such that the linker can fill in the correct address information. Since this
information requires multi-residue encoding, the linker is also modified. In our
work, we use a custom \mbox{RISC-V}\xspace back end of LLVM's \textit{lld} linker. In addition
to resolving regular relocations, our linker also applies multi-residual
encoding to pointers in the binary. This includes pointers synthesized in code
as well as pointers stored in the memory, which additionally get linked with
address information. Similar to that, data stored in the read-only section of
the binary is also linked with its address. As soon as these values are loaded
into a register, the unlinking operation is performed and the correct value is
restored.
\section{Evaluation}
\label{sec:evaluation}
In order to make a countermeasure usable in practice, the overhead must be
reasonable. In this section, we first show the introduced hardware overhead and
then evaluate different benchmark applications on the target architecture.
Finally, we analyze the software overhead, discuss the overhead sources, and
describe future optimization possibilities.
To quantify the hardware overhead, we synthesize the hardware architecture for a
Xilinx Artix-7 series FPGA. By adding the new instructions, a dedicated ALU for
multi-residue operations, and a modified load-and-store unit, the required
number of lookup-tables~(LUTs) increases by less than 5\,\%, and the number of
flip-flops increases by less than 1\,\%. However, this prototype design is
implemented without optimizations leaving space to further improve the design.
The custom LLVM toolchain based on LLVM 6.0 is used to compile different
benchmark applications for the \mbox{RISC-V}\xspace-based target architecture. The benchmarks
were taken from the \textit{PULPino} repository~\cite{pulp2018pulpino}, which
were used to originally evaluate the performance of the RI5CY\xspace core. Simulation
is performed using a cycle accurate HDL simulation of the target processor. As
baseline, we simulate the benchmark applications solely with enabled CFI
protection~\cite{DBLP:journals/corr/abs-1802-06691} but without an
application-specific data protection scheme. On top of that baseline, we
determine the exclusive overhead of our countermeasure in terms of code size and
runtime.
As shown in \autoref{tab:overhead}, on overage, the code overhead is 10\,\% and
the runtime overhead is less than 7\,\%. This is a comparable better performance
to ANB-codes, which have an average runtime overhead of 90\,\% compared to
AN-codes solely to provide memory access protection. Instead, our countermeasure
has a considerable lower overhead, making it attractive for many real-world
applications.
\begin{table}[t]
\centering
\caption{Code and runtime overhead for different benchmark programs from an HDL simulation.}
\label{tab:overhead}
\begin{tabular}{c|cc|cc}
\toprule
\multirow{3}{*}{Benchmark} & \multicolumn{2}{c}{Code Overhead} & \multicolumn{2}{c}{Runtime Overhead} \\
& Baseline & Overhead & Baseline & Overhead \\
& [kB] & [\%] & [kCycles] & [\%] \\
\midrule
fir & 4.26 & 8.54 & 39.22 & 6.35 \\
fft & 6.52 & 6.57 & 58.01 & 4.65 \\
keccak & 4.79 & 10.11 & 255.55 & 11.31 \\
ipm & 4.84 & 12.81 & 10.80 & 3.94 \\
aes\_cbc & 7.25 & 8.77 & 60.91 & 9.10 \\
conv2d & 3.26 & 13.12 & 5.92 & 2.70 \\
\midrule
Average & & 9.99 & & 6.34 \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Future Work}
The overhead numbers are already competitive for practical usage. Still, some
improvements regarding code size or performance have not been performed yet.
For example, pointer comparisons in the encoded domain are currently only
implemented for \textit{equal} and \textit{not equal}. Although seldomly used,
comparisons with other predicates are still performed on the functional value.
Similarly, in are rare cases, when pointer arithmetic uses unsupported logical
operations, the operations is performed on the functional value only. Adding
support for these operations further reduces the overhead and slightly increases
the protection domain.
Furthermore, our current toolchain has not been highly optimized for our
prototype architecture yet. We expect that, with a more optimized compiler, even
better results can be achieved in the future.
\section{Conclusion}
\label{sec:conclusion}
Memory accesses are frequently used operations, and many different security
policies, as well as safety mechanisms, rely on their correct execution.
However, when dealing with faults, a correctness of a memory access cannot be
guaranteed. While there are dedicated methods to protect the control-flow of a
program and to protect the data in memory and registers, there is no efficient
protection mechanism to protect the memory access against address tampering.
In this work, we closed this gap and presented a new mechanism to protect memory
accesses inside a program. The countermeasure is employed in two steps. First,
all pointers including pointer arithmetic are protected by employing a
multi-residue code. The redundancy is hereby directly stored inside the unused
upper bits of the pointer, which does not add any memory overhead. The second
step links the redundant pointer with the data. Subsequently, addressing errors
manifest as data errors and get detectable as soon the data is loaded into the
register. This linking approach is universally applicable and can be used on top
of any data protection scheme.
To demonstrate the practicability of our countermeasure, we integrated the
concept of protected memory accesses into a RISC-V processor. We extended the
instruction set to deal with multi-residue encoded pointers and added new memory
operations which perform the linking and unlinking step. Furthermore, we
extended the LLVM compiler to automatically transform all pointers of a program
to the encoded domain. Our evaluation showed an average code overhead of 10\,\%
and an average runtime overhead of less than 7\,\%, which makes this
countermeasure practical for real-life applications.
\ifauthor
\section{Acknowledgment}
\label{sec:Acknowledgement}
This project has received funding from the European Research Council (ERC) under
the European Union's Horizon 2020 research and innovation programme (grant
agreement No 681402) and by the Austrian Research Promotion Agency (FFG) via the
competence center Know-Center (grant number 844595), which is funded in
the context of COMET - Competence Centers for Excellent Technologies by BMVIT,
BMWFW, and Styria.
\fi
\balance
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
The ever-increasing demands of mobile users is the main reason for further expansion of the network in terms of capacity and throughput. According to the Ericsson Mobility Report ~\cite{ericsson2015ericsson}, mobile subscription will rise from 7.3 to 9 billion users worldwide in 5 years and per user data usage will rise form 1.4 GBs to 8.9 GBs. Similarly, Cisco Visual Networking Index ~\cite{index2013global} indicated that the total mobile traffic grew 81 percent and connection speeds doubled in the year 2013. Moreover, applications like internet of things (IoT), machine to machine communication, personalized TVs, Video streaming and video conference calls and self-driven cars require network with high bandwidth, data rate and latencies for their operation. The number of users requiring connection has gone manifold and network operators are facing resource shortage to provide services to such large number of users. The major requirements laid down by Next Generation Mobile Network (NGMN) alliance ~\cite{alliance20155g} for the 5th generation networks include data rates up to tens of Gbps, latencies as low as 1 ms, 1000-fold more connected devices and 10 times more battery efficiency. Different technologies are being developed and researched for meting these requirements and they include Massive MIMO and beamforming, full duplex communications, mmWave communications, UL-DL decoupled access, separation architecture for control plane and data plane, Device to Device communication and multitier heterogeneous networks with multiple radio access technologies (e.g. LTE, CDMA-2000, HSPA and HSPA+, WiFi etc) ~\cite{dahlman20145g}. Out of these, Device to Device communication (D2D) is a technology in which two closely spaced devices communicate directly with each other without relaying the data through base stations. Due to the increase in the multimedia and online gaming applications, the bandwidth requirement of users has increased and D2D communication is a promising technology to meet these application requirements. Moreover, D2D communication can provide higher network spectral efficiency by re-using the cellular frequency resources due to its short distance communication. Owing to the performance benefits of D2D communication, standardization agencies like 3rd Generation Partnership Program (3GPP) has recognized the importance of D2D communication for future mobile networks and laid down criteria and performance requirements for proximity-based services in its release 12 ~\cite{lin2014overview}.
\begin{figure*}
\centering
\includegraphics[width=1\linewidth]{Hetnet2}
\caption{Infrastructure of 5G Heterogeneous Network}
\label{fig:Hetnet2}
\end{figure*}
5th generation networks will have multiple radio access technologies (RAT) and multiple cells with different cell radius and transmit powers thus making it a heterogeneous network. The next generation radio access network (RAN) system which is also called as the 5G RAN system will be combination of evolved and revolved multiple cooperating radio access technologies (RATs). Therefore, the architecture of the 5G RAN system will constitute evolved versions of 2G (GPRS/EDGE), 3G (HSPA/UMTS), 4G (LTE-Advanced), WLAN (WiFi) and machine-type communications (MTCs) ~\cite{olwal2016survey}. Furthermore, due to higher capacity requirement of future networks, 5G RAN comprises of multiple tiers of heterogeneous networks (HetNets). The radio access technology of each base station with asymmetrical transmit powers will determine the cell sizes and will add to the complex interference scenario in the uplink and downlink of the network. The architecture of the 5G RAN systems will consist of macro cell and small cells (i.e., microcell, picocell, femtocell, relay) and device to device (D2D) based communication ~\cite{agrawal2009heterogeneous}. Usually the macro base stations serve the macro cells and have high transmit powers ranging from 43 dBm to 46 dBm and antenna gains close to 10 to 15 dBi ~\cite{lu2015wireless}. They are suitable for the remote and rural areas and have bigger coverage areas. Micro and pico cells are served by the micro an pico base stations with transmit powers ranging from 23 to 30 dBm and antenna gain of 0 to 5 dBi and they are suitable for short range urban areas. Similarly, Femto base stations are the user deployed base stations with transmit powers less than 23 dBm for increasing the coverage of the network. Due to this complex multitier network structure with different RAT, the resource allocation techniques employed in 4G RAN are no longer feasible therefore research community is working on new techniques including machine learning especially Artificial Intelligence (AI) based techniques to enable the high capacity and high throughput requirements of the future 5G networks. The infrastructure of a typical heterogeneous 5G network is shown in Figure \ref{fig:Hetnet2}.
D2D communication in this multitier HetNet adds to another tier thus making interference mitigation and resource allocation more complexer. D2D communication can takes place in either dedicated mode where D2D pairs uses dedicated frequency resources to communicate with each other ~\cite{mach2015band}. In this mode the interference caused to the cellular user is under control because of orthogonal frequencies being used by the D2D pair however this use of D2D communication is not spectral efficient. The user densification has given rise to the scarcity of the spectrum in order to meet Quality of Service (QoS) requirements of the users therefore D2D pairs are required to use the frequency which is already being used by the cellular users. This mode of D2D communication is known as underlay mode and it is more spectral efficient but the challenge comes in the allocation of resources to these pairs such that they do not cause significant interference to the cellular user. Researchers have already identified the problem complexity and have been developing different techniques to control interference and allocate resources in D2D enabled heterogeneous networks.
\begin{table}[h!]
\begin{center}
\scriptsize
\caption{Abbreviations}
\label{tab:Abbr}
\begin{tabular}{|c|l|}
\hline
\textbf{Abbreviation} & \textbf{Definition} \\
\hline
RAN & Radio Access Network \\
\hline
RAT & Radio Access Technology \\
\hline
RRM & Radio Resource Management \\
\hline
3GPP & $3^{rd}$ Generation Partnership Program \\
\hline
NGMN & Next Generation Mobile Networks \\
\hline
IMT & International Mobile Telecommunications \\
\hline
MTC & Machine type Communication \\
\hline
UL & Uplink \\
\hline
DL & Downlink \\
\hline
CDMA & Code Division Multiple Access \\
\hline
HSPA & High Speed Packet Access \\
\hline
LTE & Long Term Evolution \\
\hline
GPRS & General Packet Radio Service \\
\hline
UMTS & Universal Mobile Telecommunication System \\
\hline
EDGE & Enhanced Data for GSM Evolution \\
\hline
MIMO & Multiple Input Multiple Output \\
\hline
CSI & Channel State Information \\
\hline
UE & User Equipment \\
\hline
DUE & D2D User Equipment \\
\hline
CUE & Cellular User Equipment \\
\hline
MME & Mobility Management Entity \\
\hline
P-GW & Packet Data Gateway \\
\hline
TDD & Time Division Duplexing \\
\hline
OFDM & Orthogonal Frequency Division Duplexing \\
\hline
MINLP & Mixed Integer Non-Linear Programming \\
\hline
TTI & Transmission Time Interval \\
\hline
eMBB & Enhanced Mobile Broadband \\
\hline
SPM & Service Provision Management \\
\hline
URLLC & Ultra-Reliable Low Latency Communication \\
\hline
QoS & Quality of Service \\
\hline
MNO & Mobile Network Operator\\
\hline
\end{tabular}
\end{center}
\end{table}
A number of surveys related to D2D communication in heterogeneous network have been done ~\cite{mach2015band}, ~\cite{liu2015device} and network architecture supporting D2D communication, D2D communication scenarios, interference mitigation techniques and research directions have been presented in these surveys. Surveys related to resource allocation schemes for 5G Hetnets have been done ~\cite{olwal2016survey} and conventional resource allocation schemes have been discussed but no survey has been done comparing Conventional Resource Allocation techniques and Artificial Intelligence / Machine Learning based techniques. Some surveys ~\cite{jiang2017machine} ~\cite{wang2015artificial} identifying the importance of Machine Learning and Artificial Intelligence in 5G networks have been presented but they are not focused on problem of resource allocation and interference mitigation. In this paper we survey conventional interference mitigation and resource allocation techniques which have been proposed in the past few years and highlight the shortcoming in these techniques. We then survey artificial intelligence-based techniques developed in recent years for interference mitigation and resource allocation to illustrate the requirement of AI in future networks. In the end, future research directions to fulfill 5th generation network requirements as laid down by 3GPP and NGMN have also been proposed. The rest of the paper is organized as follows. Section II describes the details about D2D communication management, communication scenarios and different D2D communication modes. Section III contains the discussion on conventional resource allocation and interference mitigation techniques for D2D enabled network. Section IV contains artificial intelligence and learning based techniques for interference mitigation and resource allocation. Section V presents the challebges and future research directions in the field and section VI concludes this paper.
\section{Classification of D2D Communication}
D2D communication can be classified according to the several distinctive categories like D2D management, D2D scenarios and D2D communication mode according to use of radio resource.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{D2DControl}
\caption{D2D Management}
\label{fig:D2DControl}
\end{figure}
\subsection{D2D Management}
The network’s involvement in control of initiation of D2D communication defines its classification from management point of view. (See Figure \ref{fig:D2DControl})
\subsubsection{Full Control}
The level of control can be either full control where the network is in full control of D2D peers discovery and connection initiation, After the discovery and connection establishment, the network allocates power and radio resource to the D2D peers for communication with each other. The advantage of full control is that the network can easily coordinate and allocate power and radio resource to the users. Thus, in this way the harmful interference between D2D and cellular users can easily be avoided. Moreover, the base station can easily prioritize the individual transmissions to meet the QoS requirements of the users. However, with additional management tasks in hand by the base stations for D2D control, significant overhead is added to its processing. Base stations also have to know the channel state information (CSI) and share it with other base stations and this exchange is very demanding in terms of signaling ~\cite{tao2012qos}.
\subsubsection{Decentralized Control}
The D2D control can also be decentralized where D2D peers communicate with each other autonomously with very less intervention by the base stations ~\cite{lei2012operator}. The network or base station is only responsible for authentication of the devices during the connection setup between the devices. Afterwards, transmission power selection and radio resource selection are autonomously done by the devices themselves. Most of the functions are solely controlled by the D2D devices (DUEs) in this distributively controlled D2D communication. The disadvantage in this mode is that the DUEs cause significant interference to the conventional cellular users (CUEs) and their QoS parameters cannot be met by the network. Significant intelligence is required to be added to the DUEs to select the radio resource. Interference control techniques are also needed to be developed for meeting the QoS requirement of all the users of the network. Another solution to this approach is to use unlicensed frequency bands which are being used by WiFi and bluetooth based devices however use of such bands will cause interference for DUEs as there is no control of network over these unlicensed bands and devices operating in this band.
\subsubsection{Hybrid Control}
Another classification of D2D communication on the basis of D2D management and control is the hybrid control mode ~\cite{chen2010time}. In this mode, network (base station) controls the authentication, connection establishment and resource allocation to the DUEs while DUEs themselves can also select transmission power levels and radio resource in a decentralized manner based on the measurement of the channel state. Hybrid mode offers a good tradeoff to the network operators in terms of reduced signaling overhead and control of DUEs. It also maintains the QoS requirement of CUEs as well as DUEs in the network.
\subsection{D2D Discovery}
D2D discovery is another main aspect of D2D communication in a network and it significantly affects the interference caused to the CUEs. The purpose of this discovery process is to find the potential users that can communicate directly with each other to increase network capacity and benefit from the close distance communication in terms of low latency and higher throughput. The discovery process takes place in two stages namely the discovery initiation and discovery control.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{D2DDiscovery}
\caption{D2D Discovery}
\label{fig:D2DDiscovery}
\end{figure}
The discovery initiation can be ‘priori’ or ‘posteriori’. In priori discovery, the DUEs do not communicate before the discovery is done and it is commonly done when two devices want to share some content with each other. In posteriori discovery, the devices communicate with each other after the discovery is done. (See Figure \ref{fig:D2DDiscovery})
The discovery process is also controlled with different levels of involvement by the network. The discovery process can either be network assisted or it may be fully autonomous where devices discover the closely located other cellular devices that can be potential D2D peers. The network assisted discovery is easier as the network is aware of the devices locations and channel state information however it adds to the signaling overhead to the base station. The autonomous discovery has the advantage of low signaling overhead but the discovery process itself can drain the battery of the DUEs. (See Figure \ref{fig:D2DDiscovery})
Different D2D peers discovery schemes have been presented in literature incluidng both network assisted and autonomous discoveries. A network assisted discovery procedure is presented in ~\cite{yang2013solving}. In this procedure, the packet data network gateway (P-GW) detects the potential D2D users and then a message exchange takes place between Mobility Management Entity (MME), Base Station and the UEs participating in the discovery process. After the establishment of D2D connection, direct communication takes place between the DUEs. This procedure however adds significant overhead to the processing tasks performed by the base stations therefore authors in ~\cite{nguyen2014network} presented a D2D discovery procedure with lesser overhead. The UEs performs discovery in time slot-based manner using frequency multiplexed discovery channels. During certain time intervals, the devices search and listen to the discovery signals from other devices and establish connections. The number of discoveries is significantly increased by this discovery procedure with less overhead to the base stations.
Qualcomm has developed an autonomous discovery technique namely FlashLinQ ~\cite{wu2013flashlinq}. FlashLinQ is a synchronous Time Division Duplexing (TDD) Orthogonal Frequency Division Multiplexing (OFDM) system and it is designed to enable the discovery of UEs autonomously and continuously for D2D communication at high data rates. It is designed over licensed band at 2.586 GHz carrier frequency and bandwidth of 5 MHz. A Dynamic Source Routing (DSR) based D2D peer discovery is introduced in ~\cite{kaufman2009interference} in which network broadcast discovery packets through flooding. The information in discovery packets include transmission power of the devices, channel number and measured CSI. The receivers measure the SINR and path losses based on this information and estimates their transmit powers to be heard by the transmitters. If a bidirectional link can be established, D2D connection is made between the devices for direct communication.
\begin{figure*}
\centering
\includegraphics[width=1\linewidth]{D2DScenarios}
\caption{D2D Communication Scenarios laid down in 3GPP ~\cite{mach2015band}}
\label{fig:D2DScenarios}
\end{figure*}
\subsection{D2D Scenarios}
In this section, different D2D communication scenarios (see Figure \ref{fig:D2DScenarios}) are discussed as follows:
\subsubsection{Coverage}
According to the coverage, the D2D communication can be categorized as follows
\subitem -In Coverage: Both the D2D peers are in coverage of the cellular network. (Scenario 1C)
\subitem -Partial Coverage: One of the users in D2D pair is out of coverage and one is in coverage of the cellular network. (Scenario 1B)
\subitem -Out of Coverage: Both the D2D peers are out of coverage of the cellular network (Scenario 1A). This scenario is considered in 3GPP for public safety cases when network is temporarily disabled due to some disaster like earth quake, floods etc.
\subsubsection{Type of D2D Communication}
This classification of D2D communication expresses how many DUEs are involved in the D2D communication:
\subitem -One-to-One Communication: One D2D communication pair communicating with each other.
\subitem -One-to-many Communication: One DUE is communicating with multiple DUEs simultaneously by broadcasting or multicasting the information.
\subsubsection{Area of D2D Communication}
This classification is based on whether same cell or different cells are serving the DUEs.
\subitem -Same Cell: The participating UEs are located in the same cell and are attached to same base station.
(Scenario 1C)
\subitem -Different Cell: The D2D peers are linked to different base stations and are located in different cells.
(Scenario 1D)
\subsubsection{Relaying Functionality}
If there is a requirement of retransmitting data, the DUEs can act as relay as well for multiple purposes.
\subitem -Enhance Capacity: The D2D pair communicating with each other is in coverage of base station. (Scenario 2D)
\subitem -Extend Coverage: The out of the coverage DUE can use other DUE to reach the base station. (Scenario 2C)
Depending on these classification, several scenarios can be defined as shown in figure \ref{fig:D2DScenarios}. These scenarios are defined by 3GPP standardization group in their release 12 for proximity-based services ~\cite{lin2014overview}. 3GPP has defined scenarios considering both the relay functionality as well as without relaying functionality. The simplest scenario is the one where both DUEs lie in the same cell coverage.
\subsection{D2D Communication Modes}
D2D communication can take place in following two modes:
\subsubsection{Dedicated Mode}
The DUEs communicate directly with each without intervention of the base station in Dedicated Mode. However, the base station is still responsible to assign radio resource to the D2D pair for direct communication. The radio resources assigned to D2D pairs are orthogonal to the resources assigned to the CUEs therefore, there is no interference caused to the cellular users (CUEs) but the spectral efficiency of this mode is low because of dedicated resource usage. This mode of D2D communication is also known as overlay mode in literature. Advantage of this mode is that the base station does not need to implement interference mitigation techniques for meeting QoS requirements.
\subsubsection{Shared Mode}
In shared mode, the DUEs reuse the radio resources which are already being used by the CUEs and therefore there is strong interference caused by DUEs to the CUEs. This adds to the signaling overhead to the base station as it has to intelligently assign resources to the DUEs to avoid interference among DUEs and CUEs and meet their QoS requirements. Shared mode is also known as underlay mode or non-orthogonal mode in some literature. The spectral efficiency of SM is quite higher but it is quite complex from implementation point of view. The DUE can either use uplink (UL) radio resource or downlink (DL) radio resource however UL radio resources are usually preferred because the interference is caused to the base stations and transmission powers of UEs are quite lower than base stations therefore interference caused is also quite lower.
D2D communication can significantly increase the network capacity however it can be best achieved when spectrum is efficiently utilized and sophisticated intercell and intracell interference mitigation techniques are developed. Researchers have been developing different resource allocation and interference mitigation schemes for D2D enabled networks to fulfill the envisioned requirements of 5G network. In the next section, we will be surveying the RRM and interference mitigation techniques developed by the communication engineers and researchers in the recent years to control the complex interference between DUEs and CUEs as well as intercell interference in single tier and multitier Hetnets.
\section{Conventional rrm and interference mitigation techniques}
D2D communication in underlay mode requires intelligent selection of radio resources to minimize interference to the CUEs. There can be different interference scenarios when the radio resources are shared among DUEs and CUEs. In one scenario, DUEs can cause interference to the CUEs and affect their SINR. Similarly, CUEs can also cause interference to the DUEs and affect their performance and there can be mutual interference between DUEs and CUEs.
\subsection {Interference Mitigation through Mode Selection}
The interference control can be done by selecting the communication mode for D2D communication between the DUEs. As discussed earlier, there can be two communication modes for D2D communication namely dedicated mode and shared mode. Authors in ~\cite{xing2010investigation} considered a path loss model-based communication mode selection where the DUEs determines the path loss between them. If the path loss is greater than certain threshold path loss, D2D communication does not takes place and if its less than threshold, shared mode is selected. The basic principle of this selection is shown in Figure \ref{fig:PL}. Mode selection solely according to the path loss model is not optimal and interference control by this method is far from optimized solution.
Distance based mode selection scheme is presented in ~\cite{elsawy2014analytical}. The authors considered the distance between the base station and DUEs and mutual distance between the DUEs. The mode selection scheme accounts for both the cellular link quality as well as D2D users link quality. The mode selection is done if the D2D link quality is better than cellular user and distance of DUE from base station is greater than certain threshold. Authors have also included truncated channel inversion-based power control to control the interference between DUEs and CUEs. The author has proved that his mode selection scheme outperforms the mode selection scheme based on distance between DUEs only in terms of outage probability of CUEs. The mode selection between SM and DM is however not considered in this paper.
Authors in ~\cite{janis2009device} determined the benefit of D2D communication in terms of capacity enhancement if it is enabled and which mode is most appropriate to be selected. The sum rate maximization calculated according to the Shannon Capacity formula is taken as the criterion for mode selection. If CM mode gives higher sum rate, it is selected other SM or DM mode is selected. The selection of SM or DM is done according to the distance between the DUEs and base station. If the distance between base station and DUEs is greater i-e if D2D pair is far from base station, SM mode is selected in UL direction because interference to the base station is lower. If this distance is smaller, DM mode is selected for better efficiency. So far, the techniques discussed considered static channel conditions and does not account for the continuously varying channel conditions which is present in the practical networks.
For varying channel conditions, a dynamic mode selection scheme is required therefore authors in ~\cite{han2012subchannel} presented a partial solution where the network dynamically and opportunistically selects the mode for D2D communication. Dedicated mode and Conventional communication mode (CM) is considered in his study, utilizing either the UL or DL cellular resources. The author presented his results through simulations showing the case when DM mode is always selected regardless of the distance between the DUEs and the case in which CM mode is selected. It has been shown that the average sum rate is always highest for the proposal, investigating all the distances between the DUEs. The mobility of users is however, not considered by the author which can greatly affect the mode selection scheme.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{PL}
\caption{Mode selection based on path loss ~\cite{xing2010investigation}}
\label{fig:PL}
\end{figure}
We have observed that the mode selection for D2D communication has significant effect on the interference between DUEs and CUEs and efficient and dynamic selection can significantly improve the QoS to the network users. Moreover, the signaling overhead to the base stations can also be controlled through better selection of D2D communication mode. The literature has considered simple network scenarios where only few D2D pairs and cellular users are considered for analysis. Moreover, most of the work is done considering single tier network however, 5th generation networks are expected to be multitier with huge number of small cells and user densification. D2D communication is expected to be enabled in multitier heterogeneous networks and channel conditions and network dynamics will be a lot more complexer. Mode selection schemes in such complex and dense multitier D2D enabled Hetnets are required to be developed to fulfill requirements of 5G.
\subsection {Interference Mitigation through Power Control and RRM}
As discussed previously, the D2D communication can best pay off if it takes place in shared mode also known as underlay mode of communication. Underlay mode present complex challenges in terms of interference control between the DUEs and CUEs. Interference control is easier in dedicated mode or overlay mode but the spectral efficiency is quite lower. In this section, different interference mitigation techniques based on power control and radio resource allocation are discussed.
The interference caused by DUEs to the CUEs is most important. To control this interference, a transmission power control of the DUEs based technique is developed~\cite{yu2009performance}. The authors considered a system model with base station in center and circular coverage of radius \textit{R}. A single cellular user and one D2D pair operating inside the cell are sharing same frequency resource. D2D pair is confined to distance \textit{L} between them and D2D communication cannot take place beyond this distance \textit{L}. The interference models and SINR for cellular user and D2D pair has been constructed and power transmission of D2D pair is calculated for different distance form base station that causes only 3 dB degradation in the SINR of cellular user. The cellular communication can take place without errors if it faces 3 dB SINR degradation. The author collected results for both UL and DL frequency resource and found that SINR of D2D pair fluctuates from -30dB to -7dB in UL case and fluctuates between -20dB to -15dB in DL case. The problem with this simple power control technique is that the probability of D2D communication is very low because of low transmission power of the DUEs. A similar technique is presented in~\cite{janis2009interference} in which the power levels of DUEs is set by the base stations to achieve a target SINR but the performance benefits of this technique are not higher compared to ~\cite{yu2009performance}.
Interference mitigation is also done using radio resource allocation based on some criterion. The authors in ~\cite{wang2012distance} have used a distance-based resource sharing criterion (DRC) to allocate resource blocks to the D2D pairs and cellular user. The author considered a model with base station in center with circular coverage radius R. A D2D pair and a cellular user are considered inside the single cell coverage. The distance between D2D pair and cellular user \textit{L} is considered as the criterion for resource allocation. The D2D user and cellular user are using same frequency resources. The author assumed that GPS locations of all users are known to the base station and he assumed a minimum distance $L_{min}$ below which D2D communication cannot take place because of interference from cellular user. The author calculated SINR for D2D communication for different distances $L_{min}/R$ and the outage probability of D2D communication is plotted against distance \textit{L}. The results have shown that with distance-based resource sharing criterion (DRC), the outage probability of D2D communication can be significantly decreased thus enabling D2D communication in most of the cell area.
The authors in ~\cite{yu2011resource} have further improved the results by considering a better system model for 5G communication. The authors in ~\cite{yu2011resource} considered \textit{N} cellular users distributed uniformly in the circular cell coverage of radius \textit{R}. The author utilized location estimation algorithm and further decreased the outage probability of D2D communication by choosing the resource sharing of that cellular user that minimizes the outage probability of D2D communication. The interference models used by ~\cite{yu2011resource} are same as used by the [17] with the addition of location of cellular user taken into account. The results have shown that most of the cell area is operable for D2D communication and SINR of D2D user does not fall below the preset threshold. The authors in ~\cite{yu2011resource} also compared his results with ~\cite{wang2012distance} and shown the decrease in outage probability using his proposed algorithm. The advantage of these techniques is the reduced overhead to the base stations.
Another simple method to control interference form DUEs to CUEs is suggested in ~\cite{peng2009interference}. In this method, the base station calculates the tolerable interference levels when D2D uses the resource blocks (RB) in the UL and broadcast this information to the DUEs. The DUEs use this information to choose those RBs in the UL which cause minimal interference to the CUEs. The author presented performance improvement in terms of throughput of CUEs from 2.65 Mbps to 3.33 Mbps. However, this was achieved at the cost of throughput decrease of DUEs from 3.02 Mbps to 2.83 Mbps.
The problem of resource allocation to the DUEs is addressed in a different manner in ~\cite{zulhasnine2010efficient}. Authors have formulated the problem of resource allocation as Mixed Integer Non-Linear Programming (MINLP). MINLP is of very high complexity and its practical implementation is not feasible because the algorithm cannot be solved in 1 ms transmission time interval (TTI) considered in LTE (A). In order to make the solution practical, a heuristic greedy algorithm is proposed ~\cite{tapia2009hspa}. The resource allocation is done based on the channel quality of the CUEs. If a CUEs are experiencing good channel quality, their resource is shared with the DUEs to keep their SINRs above certain thresholds. Extensive simulations have been performed by the author considering conventional allocation schemes like proportional Fair and Round Robin. It has been shown that the network throughput is substantially increased in this D2D enabled network. The author did not prove the optimality of this heuristic algorithm-based allocation.
Joint power and resource allocation-based techniques have also been developed to control the interference between DUEs and CUEs. In ~\cite{gu2011dynamic}, authors have presented a dynamic power control and resource allocation-based technique to mitigate the interference. The base station assigns the resources to the CUEs in a prioritized manner and remaining resources are allocated to the DUEs. Afterwards, if the requirement of DUEs is not met then base station determines the resources of CUEs that can be shared with DUEs. The interference among such DUEs and CUEs sharing resources is mitigated through dynamic power control. The power control is done by the base station which determines the transmission power of the DUEs based on channel quality measurements between DUEs and CUEs and between DUEs and base station in UL direction. Author compared his results with fractional power control-based technique and demonstrated a 5.7 dB increase in the SINR of CUEs and 2.77 dB increase in SINR of the DUEs. The power control management by base stations added to the significant overhead to its processing and this aspect is not discussed in the paper.
A joint resource allocation and power control technique based on column generation method to reduce to complexity of the problem has been proposed in ~\cite{phunchongharn2013resource}. Authors have considered a network with base station in the center and users distributed uniformly in the area of 50 m around it. The DUEs are randomly distributed in the area with distance between them as uniformly distributed from 0 to 25 m. The resource allocation problem objective is to reduce the interference caused to the cellular users and maintain the QoS of D2D users. One RB is shared with multiple D2D users to increase the spectral efficiency and the author has compared the results of his technique with the technique in which RB is shared with single D2D user. The D2D user calculates interference on all RBs and selects a RB that causes minimum interference to cellular users using same RB. If the interference caused is under certain threshold and access constraints are met, it reuses the RB otherwise it searches for other RBs. The objective of the reuse of RB is to reduce the transmission time interval which in turn translates to maximizing the spectral efficiency. The author has presented his results showing that with the little increase in the transmission power of D2D links, the transmission time interval can be significantly reduced thus increasing spectral efficiency. The technique is centralized and will be running on the central base station therefore it is suited for smaller and medium sized networks in which traffic demands do not change very fast.
The mutual interference between DUEs and CUEs is solved through Fractional Frequency reuse (FFR) approach ~\cite{chae2011radio}. The frequency band available to the base station is divided into four sub frequency bands (f1, f2, f3, f4). The inner region of his base station uses the sub-band f1 and while other sub-bands are utilized by outer regions. The D2D pairs also exploit this division and utilize the frequency bands of other regions when located in inner region. For example, if a D2D pair is located in region of band f1, it will reuse the resources being used in region with band f2 or f3. In this way, the interference caused to the cellular user and other D2D users will be under control and QoS for all network users is easily met. The problem with FFR approach is that the bandwidth is not efficiently utilized and it is dependent on accurate location estimation. Error in location estimation will result in very bad performance.
Graph theory has also been employed to solve the problem of interference between DUEs and CUEs in cellular networks. An interference aware graph theory-based resource allocation scheme is presented in~\cite{zhang2013interference}. Sum rate maximization of DUEs and CUEs through resource allocation is the objective of this scheme. The interference among DUEs and CUEs are represented in the form of graph. An interference graph is constructed first based on the network topology. The graph has three main characteristics; 1) the link attribute which tells whether the vertex in a graph is for DUE or CUE; 2) SINR values for each resource block; 3) attribute representing allocation of RB to the individual vertexes. The optimal solution for allocation of resources is based on exhaustive search that tries all possible combination of allocation possibilities therefore author implemented sub optimal solution to reduce the complexity of the scheme. The results show that the performance of sub optimal solution is almost same as of optimal solution and it significantly maximizes the sum rate and is spectrally efficient compared to greedy orthogonal sharing scheme.
The interference problem is solved through game theory in ~\cite{wang2013joint}. Authors have solved two optimization problems, one related to the resource sharing and other relating to the optimal transmission power selection. This optimization was solved using Stackelberg game. The CUEs are made as the leaders and DUEs as the followers and leader is made as the owner of the radio resources. The followers are charged a certain fee if they use the same resources. The utility function based on the throughput of the leader is defined and first optimization problems requires setting up of a price to maximize this utility function. The second optimization problem is to set transmission power of the DUEs (followers) to maximize their utility function. A joint resource scheduling and power allocation is done afterwards to fairly distribute resources among the DUEs and CUEs. The author has proved in his results that the throughput of DUEs increases with increase in CUEs admitted to the network, because more resources will be available to be shared with DUEs.
The interference mitigation and resource allocation schemes developed by researchers discussed till now considered simple network scenarios and did not considered multitier heterogeneous network models. The network model considered have few D2D pairs and CUEs with only one macro base station in center which is not practical in nature therefore, researchers identified this shortcoming and have done analysis using concepts of stochastic geometry and multitier heterogeneous networks. The complexity of interference in multitier Hetnets is much more than single tier network because each tier will cause interference to the other tier (see figure \ref{fig:Hetnet2}).
A successive interference cancellation (SIC) scheme for stochastic geometry-based network model has been presented in ~\cite{ma2015performance}. The author employed concepts of stochastic geometry and considered a network model with DUEs and CUEs distributed as per Homogeneous Poisson Point Process (PPP) and base stations distributed using Stationary Point Process. The base stations are assumed to have infinite SIC capability while D2D receivers have finite SIC capability. The author has presented the stochastic equivalence of the interference, by which a two-tier network (Macro and D2D tier) can be represented by a single tier interference. The successful transmission probabilities of CUEs and DUEs are calculated for equivalent model and are validated by simulations and analytical results.
A network assisted interference mitigation scheme is presented in ~\cite{tsai2012intelligent} in a two-tier heterogeneous mobile network with macro and femto base stations. The UL of OFDMA based network model is considered and it has one macro base station and cluster of femto cells deployed inside the houses. The D2D pairs are located in the coverage of the femto cells and share the spectrum of the cellular users. The Carrier to interference plus noise ratio (CINR) for different sub carriers of OFDMA system are calculated for macro BS, femto BS and D2D users. Based on these measurement, macro base station calculates the tolerable power levels and broadcast this information to the D2D users to keep the interference levels below predefined thresholds. The minimum transmission power of DUES is also calculated to keep their SINR above the predefined $SINR_{DUE}$ threshold. The scheme is reliable as it ensures the link reliability of femto and macro cell users as well as the D2D users in the network however, this scheme requires broadcast of information to the D2D users without which this scheme renders useless.
The interference caused by DUEs to the macro and femto cells is handled through Stackelberg game in ~\cite{he2014resource}. UEs of the macro and femto cells are taken as leaders and DUEs as the followers. The leaders own the radio resources and charge fees to the DUEs for using these resources. The author assumed that the macro BS and Femto BS use dedicated channels however, a more realistic scenario must be considered in which macro and small cells share channels with each other.
\subsection {Summary of Conventional Interference Mitigation and RRM techniques}
From the papers we reviewed in this survey, we observed that most of the power control techniques are used to solve interference caused by DUEs to the CUEs. The power techniques are strongly dependent on the distance of the DUEs from CUEs if DL resource is shared and from base station if UL resource is shared with the DUEs. The decrease in transmission power of DUEs due to these distances can also affect the QoS of DUEs if the distance between them is larger. Therefore, power control techniques become useless in certain cases. In such cases, RRM techniques are used to control the interference. The resource allocation can either be done by the base stations in full control mode in which the base station needs to know the CSI of all the links involved in communication or it can be done the DUEs themselves in loose control mode which also has the advantage of reduced signaling overhead. Loose control however is not desired by the network operators because they will lose control over the network management.
Joint power control and radio resource allocation schemes are better solution to control interference among DUEs and CUEs. Other interference control techniques include massive MIMO, beamforming and Interference alignment techniques. Most of the papers considered scenario 1C (see figure \ref{fig:D2DScenarios}) in which both D2D users are in coverage of the same cell and they share resource of one CUE with only one D2D pair. Only few papers assumed multiple CUEs and DUEs in their network models. The mobility of users is not considered by the papers which will significantly change the interference scenario and require more sophisticated techniques to handle interference.
\section{Artificial intelligence and machine learning based interference mitigation and rrm techniques}
Currently 4G network is providing the seamless Internet Protocol (IP) based connectivity to all mobile devices. It took almost 30 years to successfully transform the conventional telephone based mobile communication to fully digital IP based communications. Internet of things has given rise to more connected devices and throughput requirements have increased due to high quality video streaming and entertainment applications. 5th generation networks are envisioned to meet these growing requirements as described earlier in this paper however, networks need to revolutionized with cutting edge technologies to meet these demands. The three main services provided by 5G (see figure \ref{fig:5GReq}), namely Enhanced Mobile Broadband (eMBB), Ultra Reliable Low Latency Services (URLLC) and Massive Machine Type Communications (mMTC), will be enabled by technologies like cell densification through small cells and MIMO but these technologies are cost ineffective. A cost-effective technology for enabling these services is to exploit Artificial Intelligence (AI) for network functions like Radio Resource Management (RRM), Mobility Management (MM) and Orchestration (MANO) and Service Provision Management (SPM).
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{5GReq}
\caption{Main Services of 5G}
\label{fig:5GReq}
\end{figure}
AI is the science of making the machines intelligent just like humans. It has been effectively used to solve diverse problems of nature and has been applied to communication problems as well for optimized results. AI falls into two categories, one in which the machine has predefined options of action to select from and choose the best action among them. The second category of AI is the one in which, the machine interacts (e.g. sense, mine, reason and predict) with the environment and then take actions for optimal results. In communication networks, the second category of AI is applied as the channel conditions and network parameters keep on changing and devices or base stations need to sense these changes continually to taker better actions. Owing to the effectiveness of AI, the complex interference and RRM problems have also been addressed by different researchers. Different AI based algorithms like Genetic Algorithms (GA), Ant Colony Optimization (ACO) and Reinforcement Learning (RL) have been employed to solve interference problem between DUEs and CUEs.
Genetic Algorithm is inspired by the process of Natural Selection and they generate high quality solutions to optimization problems. GA is based on Swarm Intelligence and reaches optimal solution with faster speeds and lower complexity. A Genetic Algorithm based technique is presented in ~\cite{yang2014ga} in which resource sharing between DUEs and CUEs is done through GA. The authors considered a network scenario with 30 CUEs and 30 DUEs in a single cell and 50 Resource blocks to be shared among them. Number of D2D users, number of CUEs and number of resource blocks available are considered for coding required for genetic algorithm. Afterwards, optimization is performed to increase system throughput. Results are compared with Exhaustive search algorithm which is very computational extensive, greedy heuristic algorithm, network with no DUEs and network with D2D communication enabled and using orthogonal resources. A throughput gains of 30 Mbps over greedy Heuristic algorithm is achieved while throughput achieved is 5 Mbps less than exhaustive search algorithm. Computational requirement analysis has also been done and comparison is made with Exhaustive Search algorithm. The network scenario considered however is single cell based which is not practical for 5th generation networks and user mobilities are not considered which will significantly affect the results.
ACO is also a swarm intelligence-based algorithm based on probabilistic techniques to solve computational problems which can be reduced to finding optimal paths through graphs. It is inspired by the behavior of ants who move in search of food between their colonies. ACO algorithm has been employed in ~\cite{liotou2014ant} to solve the interference problem in D2D enabled cellular network. Graph coloring is used for mapping the interference among D2D users using Interference Level Indicator (ILI) term and numerical values from 1-15 are used to quantify the interference. After that, Ant colony optimization (ACO) algorithm is used to allocate Resource Blocks among D2D pairs to increase spectrum reuse. A single cell network scenario with D2D users, Cellular users and single base station is considered. Target outage probability is selected by choosing minimum outage probability threshold and minimum SINR threshold. eNB calculates the interference levels among the D2D users using channel information and path loss and graph model is created representing the interference among D2D users. Iterative approach is used to choose the optimum number of D2D links that gives the outage probability lesser than the predefined threshold. Afterwards, ACO is used by minimizing the sum of weights (mutual interference). Network parameters as per LTE standard have been chosen and convergence of ACO algorithm in a graph representing the cost measurement and number of evaluation runs is presented. The results are compared with exhaustive search algorithm which is computationally extensive and comparable spectrum efficiency with quiet lower computational requirements is achieved. Computational requirements comparison of ACO with exhaustive search algorithm has also been done and significant improvements are observed. The author recommended to include mobility of D2D users in network model as future work.
Reinforcement learning (RL) has been greatly employed in solving the network problems in 5G heterogeneous networks. Reinforcement leaning is a machine learning based technique that does not require any model to predict the future actions or draw inferences. \textit{Q}-Learning, a sub part of reinforcement learning, has been used by many researchers to solve resource allocation, cell association and interference mitigation problems. It is a model free learning technique in which the learning agent tries to maximize its reward by taking immediate actions. Due to the uncertainty of the 5G mobile networks due to changing network conditions, channel fading and user mobilities, network cannot be modeled therefore traditional model-based learning schemes cannot be employed. Therefore, \textit{Q}-Learning has proved to be a powerful tool to solve network problems. \textit{Q}-Learning includes four parameters namely action \textit{a}, state \textit{s}, transition probability form one state to other state $P_{s,s^\prime}$ and reward $r_{s, a}$. The state is the internal phenomenon of each agent while reward reflects the quality of action taken by the agent. The objective of the \textit{Q}-Learning is to determine the optimal policy ${\pi }_{s}^{\ast }$ to choose the actions that give the maximum reward. The process works as follows; the agent chooses an action at time \textit{t} in some state \textit{s} and measure the reward \textit{r}. The agent records the reward and move to next state $s^\prime$ to choose the next action. The \textit{Q} value for each action is measured and recorded in \textit{Q} matrix of size state x action according to following equation:'
\begin{equation} \label{eq:15}
\begin{aligned}
Q(s,a) = (1-\alpha) Q(s,a) + \alpha [r(s,a)\\ +\gamma\max\limits_{b \in A}Q(s^\prime ,b)]
\end{aligned}
\end{equation}
Where $\alpha$ is the learning rate and $\gamma$ is the discount factor. An appropriate action is assigned a positive reward and hence gets a high \textit{Q}-value while inappropriate action is punished and gets lower \textit{Q}-value. It has been proved in ~\cite{watkins1992q} that the update rule of \textit{Q}-values in a two dimensional look up table converges to optimal \textit{Q}-value when state and actions are visited infinitely often. The learning comprises of two stages namely exploration and exploitation. In exploration phase, the agent explores all states and actions and record the \textit{Q}-values while in exploitation phase, only those actions are chosen whose \textit{Q}-values are higher.
A \textit{Q}-Learning based resource allocation scheme has been developed in ~\cite{luo2014dynamic}. A simple network scenario with 2 DUEs and 2 CUEs is considered. 2 channels [Ch1, Ch2] and 3 power states [P1, P2, P3] making six combinations are allocated to the DUEs. User locations, user channels and user arrivals are taken as input to the algorithm and the objective is to maximize the system capacity calculated through Shannon capacity formula. Author has considered just 2 CUEs and 2 DUEs in a single cell network model which is not practical and inputs for decision making are also too simple. System capacity through \textit{Q}-learning is compared with random resource allocation and maximum power allocation and performance gains are presented.
A multicell network model is considered in ~\cite{alqerm2016cooperative} for mitigating interference among DUEs and CUEs and intercell interference through \textit{Q}-learning based resource allocation scheme. The network model considered contained a macro cell, cluster of femto cells with one UE in each femto cell and D2D users distributed in the coverage of macro cell. The authors considered N resource blocks to be allocated with P= (1,2,....,P) power levels and M= (1,2,....,M) modulation indexes chosen intelligently to control the interference and maximize D2D users throughput and spectral efficiency. SINR of D2D receiver is formulated and that of Macro users and state-action pairs were made. The \textit{Q}-values for each pair is determined for particular allocated resource and \textit{Q}-table is made by the devices and it is shared among all users of the network to find the optimal resource sharing policy. The constraints considered include SINR threshold of macro user greater than predefined threshold, only one resource block to be used by each user with one power level and modulation index and a binary decision variable which outputs 1 with transmitter selecting resource block N, power level P and modulation index M otherwise it will be zero.
Authors have done resource allocation in two phases. In exploration phase, users select different actions and measure rewards to explore best rewarded action and in exploitation phase, the action with maximum reward (high \textit{Q}-value) is selected. Author defined exploration rate $\epsilon$ which is higher in start to find highest Q-value and learning rate $\alpha$ which is faster when higher \textit{Q}-value is not found and when it is found, $\alpha$ becomes lower. All users (agents) maximize their local \textit{Q}-values and global \textit{Q}-value is decomposed into the linear combination of local \textit{Q}-values thus if each agent maximizes its \textit{Q}-value, global \textit{Q}-value is maximized. The results of his scheme are presented showing spectral efficiency and throughput improvements in comparison with joint-Resource Allocation and Link Adaptation (RALA) scheme, Matching RM and Down SA schemes. The drawback in this scheme is that the network model considered has one UE associated with each femto cell and only one macro cell is considered. The effect of macro to macro cell interference is not addressed in this paper. Moreover, the DUEs and Femto cell users need to share their \textit{Q}-tables to find the best allocation scheme which is itself an overhead for the network.
Similar to ~\cite{alqerm2016cooperative}, authors in ~\cite{khan2017throughput} have presented a cooperative reinforcement learning technique for allocating RBs and power level to D2D users underlaying cellular users. A single tier netowrk model with macro base stations is considered and D2D and cellular users operate under the coverage of macro base stations. The authors have included cooperation between the learnign agents which are D2D users in whihc they share their value functions to jointly increase the overall throughput of the system. A comparison of increase in system throughput and fairness of allocation is made with distributed reinforcement learning based technique with no coperation~\cite{luo2014dynamic}~\cite{nie2016q} and random allocation technique. The drawback of this paper is that it considered a single tier network however, 5G network is expected to be multi tier heterogeneous network.
An expected \textit{Q}-learning technique to find optimal resource allocation scheme for optimizing user’s data rates and spectrum usage in a decoupled LTE-U network is presented in ~\cite{hu2017expected}. A game theoretic model incorporating user association, spectrum allocation and load balancing is considered for resource allocation among the LTE-U and WiFi users. The network model considered has a macro cell base station in the center and $N_s$ small cell base stations, $N_u$ LTE-U users, W wireless access points (WAP) and $N_w$ WiFi users uniformly distributed. Macro cell users utilize the licensed band in uplink and downlink while LTE-U and WiFi users utilize unlicensed spectrum for communication. WAP utilizes CSMA/CA (carrier sense multiple access with collision avoidance) protocol for spectrum usage. The resources to be allocated to the users consist of uplink and downlink bands in licensed spectrum and time slots for LTE-U and WiFi users in unlicensed band. Authors considered logarithmic function to compute utility functions which make resource allocation fair among the user with different data rates. Authors have developed a \textit{Q}-learning based allocation scheme where each base station allocates resources based on $\epsilon$-greedy mechanism (exploration) and measure \textit{Q}-values and update its state while broadcasting its \textit{Q}-value to the other base stations. The other base stations will use this information to determine its resource allocation scheme and share it with others. In this way, a global maximum \textit{Q}-value is achieved to find the optimal resource allocation for required data rates. If the data rate of some users is below the required data rate, then it will keep sending requests to base stations for connection to get better data rates. In this way, algorithm finds a mixed strategy Nash Equilibrium to optimize data rates of all the users. The authors presented their results showing increase of 12.7 \% and 51.1 \% in sum rates (UL+DL) compared to traditional \textit{Q}-learning and LTE-U nearest neighbor algorithm respectively. Authors have also shown that is algorithm takes 19 \% less time to converge to the optimal solution. The drawback of this paper is the requirement to share learnt information with neighboring agents which adds to the overhead.
An autonomous \textit{Q}-Learning based technique is presented in ~\cite{asheralieva2016autonomous} in which D2D users autonomously select resource blocks and power levels in a distributive and decentralized manner to mitigate interference between DUEs and CUEs. The authors have considered joint operation of cellular users and D2D users in a heterogeneous cellular network with multiple Base Stations and D2D users. Authors have done analysis of network in two scenarios (i) when orthogonal resource is shared among cellular and D2D users (ii) and when resource is shared among them. The goal of each D2D pair is to jointly select the wireless channel and power level to maximize its reward which is the difference between throughput and cost of power consumption with the constraint of having a minimum SINR requirement. A cooperative game-based approach is used with multiple D2D users as players who learn their best strategies based on locally observed information and developed a fully autonomous multi agent \textit{Q}-learning algorithm converging to a mixed strategy Nash Equilibrium (NE). Authors considered a heterogeneous network as per 3GPP LTE-A standard with 3 Base Stations (macro, micro and femto each) with 100 cellular and 100 D2D users distributed randomly and considered standard time division duplexing (TDD) scheme. \textit{K} orthogonal resource blocks and \textit{J} power levels are considered for allocation to the D2D users for both scenarios. The reward is measured as sum of instantaneous rewards over infinite time interval as time for which user stays in network is unknown. Therefore, a discounting factor $\gamma$ is introduced to avoid the infinite sum problem.
The authors have made a matrix comprising of all possibilities of channel and power allocation and aims to find the best pair of channel and power level that maximizes the reward. Each D2D pair selects an action containing channel-power level pair to maximize its own reward and does not know about the actions of other D2D pairs. The selection of channel and power level is done at particular instant without knowledge of previous instances therefore making it a Markov Decision Process (MDP). A Multi-Agent Q-learning based scheme is developed for such MDP in which each D2D pair determines its optimal strategy for action selection. The strategy is made to maximize the value state function \textit{V} defined by authors that maximizes the expected value of the reward function and achieves Nash Equilibrium. More than one strategy can exist for each learning agent to achieve Nash Equilibrium therefore authors select the state function with maximum value. For autonomous selection of actions by D2D users without knowledge of other D2D pair actions, each D2D pair estimates the beliefs about other player’s strategies. Reference points for the beliefs and strategies were chosen and were continuously updated based on previous time slot measurements of beliefs and strategies. Based on the estimation of belief, D2D pairs autonomously select actions to maximize their rewards.
In order to overcome the challenge of exploitation/exploration tradeoff of \textit{Q}-learning, author considered $\epsilon$-greedy selection scheme and all actions were weighted according to their action values (rewards) so that actions with higher probability are selected. Authors have made use of Boltzmann Gibbs distribution for action selection and presented results showing effect of temperature $T_B$ (used in Boltzmann Gibbs Distribution) on the convergence of autonomous channel and power level section (ACS). Authors have also presented the effect of increase in D2D users on the throughput of both cellular and D2D users and compared his results with $\epsilon$-greedy \textit{Q}-learning based action selection scheme, uniform random selection scheme, parallel fictitious play (FP), parallel best response dynamics (BRD) based action selection scheme where action is selected according to CSI information and previous actions and optimal centralized strategy (OCS) where action is selected according to global CSI information. Authors have also presented results showing $SINR_{min}$ value selection on the throughput of cellular and D2D users and shown significant improvements in the network.
Random Forest alogrithm has been employed by authors in~\cite{imtiaz2016learning} to allocate resources in a Time Division Duplexing (TDD) based Cloud Radio Access Network (CRAN). Authors have considered a CRAN system with Remote Radio Heads (RRH) deployed to provide Line of Sight (LOS) communication to a large number of users. The overhead of gathering instant CSI for high mobility users is tremendously large in ultra dense network therefore, the proposed technique exploits the position estimates of high mobility users which are somewhat predictalbe~\cite{lu2013approaching} to allocate the resources to the users. This waves off the requirement of CSI for resource allocation tasks. The robustness of the proposed scheme is tested by using accuarte position estimates in training dataset and inaccurate estimated in test datasets for random forest. The system throughput is calculated afterwards and comapared with the CSI based resource allocation scheme. Significant perfornace gains are achieved by proposed technique with overheads of 2.5\% compared to 19\% overhead in CSI based technique. The assumptions of LOS communication and requirement of accurate position estimates for training dataset limits the performance of this scheme.
An energy efficient power allocation scheme is presented in ~\cite{alqerm2017energy} in which power is allocated to the users using enhanced online learning. The allocation is done is non cooperative manner to maximize the energy efficiency of the network. The devices select the power levels in distributed and autonomous manner based on intuition about the other devices power selection strategies. The power selection strategy is determined by the devices using \textit{Q}-learning algorithm with reduced states to increase the convergence time of the algorithm. The authors have considered a two tier heterogeneous network model where the first tier comprises of the macro base stations and second tier consists of femto or pico cells and D2D users. The downlink transmission model is considered for power allocation with the aim of maximizing energy efficiency. Authors have proved that the energy efficiency can be significantly increased and spectral efficiency can be enhanced using the \textit{Q}-learning based enhanced online structure and convergence times can be reduced using intuition based power allocation to the devices.
Heterogeneous Cloud Radio Access Netowrk (H-CRAN) have become a focus in 5G Networks to leverage the benefits of both heterogeneous and CRAN advantages. Therefore, authors in~\cite{alqerm2017enhanced} have presented an enhanced machine learning scheme for energy efficient resource allocation in 5G H-CRAN. The network model considered has macro base stations and Remote Radio Heads (RRH) which serve two types of users, one with high QoS requirements and other with low QoS requirements. The Q-learning methodology is employed to exploit the low power RRHs for interference mitigation between macro tier and RRH tier while meeting QoS requirements of the cellualr users and maximizing energy efficiency. The availble resource blocks (RB) are divided into two sets of RBs, one for users with high QoS requirements located at cell edges and served by RRHs and other for sharing with RRH users and macro cell users located in the center and having low QoS requirements. The learning methodology is employed separately for both sets of RBs to reduce the convergence time of the algorithm. The resource llocation is done centrally by centralized Baseband Unit (BBU). All users report their channel state information and path losses to thier serivng base stations and RRHs, which is then sent to the centralized controller. The controller exploits this information to learn the environment and allocate power and RBs to the users in order to maximize energy efficiency and maintain QoS requirements of the served users. Significant performance gains interms of energy efficiency, spectral efficiency and data rates are achieved through this centralized online learning scheme. The limitaiton of this scheme is that if the central controller fails tooperate, the whole netowrk will go downand wuill not function.
Therefore, authors in ~\cite{alqerm2018sophisticated} have extenteded the work in~\cite{alqerm2017enhanced} to include decentralized resource allocation in the network in addition to centralized resource allocation by the BBU. In decentralized resource allocation, the macro base stations allocate resources to the RRHs and cellualr users as they have all chanel state infromation and path losses form the users and RRH operatin gunder their coverage. The learning is implemented in a distributed manner in all macro base stations in whihc they learn a common strategy $\pi$ to allocate RBs and power level to the users to maximize the system energy efficiency while meeting QoS requirements. The authors have alos implemented both cnetralized and decentralized techniques on Software Defined Radio (SDR) plaatforms to practically test the performance of the schmemes. The hardware setup comprises of GNU radio, USRP N210 from Ettus Research ~\cite{SDRForum} and two dell servers for base band processing. The numerical and practical results have shown considerable increase in energy efficiency and spectral effciency of the system while providing higher bit rates and low Bit Error Rates (BER) in the system.
Authors in~\cite{chattopadhyay2016location} have used the concept of stochastic learning for opportunistic bandwidth sharing between static and mobie users in a cellular network. The channel conditions keep on changing for the mobile users at a faster rate compared to static users and mobile users keep on changing thier serving base stations. It is quite challenging to provide higher data rates to these moving mobile users. Therefore, authors in ~\cite{chattopadhyay2016location} have proposed a location dependent bandwidth sharing and formulated the problem as a long run average reward Markov Decision Process (MDP). The reward function of such an MDP either changes with time or not known at all. To overcome these problems, a time scale stochastic approximation based learning algorithm is proposed. The authors have considered a multiple Macro Base station based netowrk model with static and mobile cellular users. The mobile users are moving along a line and keeps on changing their serving base station after fixed time slot $\sigma$ while moving at some velocity $v$. Authors have presented learning model for both constrained and unconstrained objective functions and formulated states, actions, state transitions and rewards for this MDP. The optimality and convergence of the learning algorithms have been proved and fairness of bandwidth sharing has been presented in the paper. It has been shown that significant improvements can be achieved through this stochastic learning model for mobile users in the cellular network. However, the authors have limited their analysis to the users moving with fixed velocity while in practical networks the users move with different velocities.
Thus far we have seen that AI techniques have proved to be a powerful tool to address interference and resource allocation problems in 5th generation networks. The learning capability when put into the devices will distribute the processing load of the base stations and decision process will become decentralized. This decentralized decision will ease off base stations for performing other network functions like cell association, mobility management and other control tasks etc. Moreover, network performance can also be increased as depicted in the papers discussed above. Though genetic algorithm and ACO algorithm has been employed in resource allocation and interference mitigation problem, Reinforcement learning (Q-Learning) has proved to be a better solution to these problem because of its model free learning capability. The network scenarios considered in literature to solve RRM problem include heterogeneous networks however, user mobilities are not considered which can significantly change the network parameters and hence the solutions need to be tailored. Moreover, the out of coverage D2D communication scenarios are not considered in developing these learning algorithms.
\begin{figure*}
\centering
\includegraphics[width=1\linewidth]{MLChart}
\caption{Machine Learning in 5G}
\label{fig:ML}
\end{figure*}
The Artificial intelligence and Machine Learning techniques cannot be denied when it comes to solve 5G network problems. Other than RRM and interference mitigation, AI can also be employed in Software Defined Networking (SDN) and Network Function Virtualization (NFV), Self-Healing Networks, Intrusion Detection in network, channel estimation/detection and spectrum sensing and detection in Cognitive Radios (CR). Learning algorithms like Bayesian learning, K Nearest Neighbor (KNN), K-means Clustering, Support Vector Machines (SVM), Principle Component Analysis (PCA) and Markov Decision Process (MDP) plays a vital role in solving different network problems and enable envisioned 5G requirements ~\cite{alliance20155g}. The employment of Artificial Intelligence and Machine Learning techniques addressing different problems of 5G networks is depicted in Figure \ref{fig:ML}.
\section {Future research directions and challenges}
D2D communication has huge performance benefits including low latency communication, high data rates because of close proximity communications, high spectral efficiency and high energy efficiency. These performance benefits can enable the requirements laid down by 3GPP and NGMN alliance for 5th generation network. The achievement of these performance benefits requires efficient Radio Resource allocation and interference mitigation techniques to be developed.
\subsection{Mode Selection Schemes}
The interference mitigation through mode selection scheme needs to be fully dynamic and adopt to the changing network conditions. The heterogeneity in the network will create complex interferences therefore, dynamistic mode selection will significantly impact the network performance. Moreover, user mobilities need to be considered in network models because it affects the network performance and plays important role in dynamic mode selection. Interference mitigation through resource allocation needs to be dynamic too where the network adapts to the changing network and determine allocation schemes in dynamic manner. This is not only increase the network performance in terms of throughput but will also increase the spectral efficiency without compromising the cellular user’s performance. Most of the literature considered either UL and DL link shared by DUEs with the CUEs however sharing of both UL and DL resources in dynamic manner will further increase the network performance.
\subsection{Interference Mitigation in Heterogeneous Networks}
5th generation networks will have multiple cells including Macro, Micro, Pico and Femto cells and the number of small cells will significantly increase to support huge active users in the network. D2D communication in these multi cell heterogeneous networks will make interference mitigation quite complexer therefore conventional interference mitigation scheme will add significant overhead to the base stations processing because of larger number of users. The decision making requires huge number of calculations to be performed prior to allocating radio resources to the users for communication. Owing to the complexity of the problem, artificial intelligence-based techniques have been employed in literature. GA and ACO algorithms have been employed for RRM in single cell network scenarios however, employment of these algorithms in multicell heterogeneous networks needs to be explored. RL has been greatly employed and proved very useful in RRM and interference mitigation algorithms in 5G networks. The multicell heterogeneous networks have been considered in literature and \textit{Q}-Learning based schemes have been developed for RRM in both centralized and decentralized manner. Most of the literature has considered sharing of learnt information by the DUEs with other DUEs and base stations for efficient RRM however, the mechanism to share this information has not been discussed. Autonomous decision making by DUEs has also been explored in multicell network however, single macro base station is considered in the study. The practical network will have multiple macro cells and small cells and such network has not been considered in literature for RRM through RRM technique. Moreover, user mobilities have also not been considered in the network models while developing RRM and interference mitigation techniques.
\subsection{New D2D Scenarios and D2D Communication in mmWave and Unlicensed Bands}
D2D communication scenarios have been laid down by 3GPP as shown in Fig. 3. Almost, all of the literature has considered scenario 1C in which both D2D communicating devices are located in the coverage of same cell. The other scenarios of D2D communication (see Fig. 3.) needs to be investigated for example D2D communication when both devices lie in the coverage of different cells. In addition to it, D2D communication in mmWave band needs to be exploited. This has several advantages as mmWave has relatively higher bandwidth and can provide much higher data rates and smaller distance between DUEs can ensure efficient communication over mmWave bands. The usage of mmWave bands for D2D and other bands for cellular communication will not introduce any interference and resource allocation will be quite easier for the base stations. D2D communication in unlicensed bands can also be exploited because it can ensure efficient cellular communication however, interference mitigation for D2D communication in unlicensed bands will be quite complexer because of no control of network over these bands.
\section{Conclusion}
D2D communication underlaying cellular network can provide significant performance improvements in terms of throughput and spectral efficiency however, with these performance benefits there are several challenges related to management of DUEs, interference mitigation and allocation of radio resources to the DUEs. Interference mitigation and RRM for D2D pairs is an active research area and new techniques are being researched for efficient direct communication. Artificial intelligence has been greatly exploited to solve these complex interferences in 5th generation multicell heterogeneous networks. The major weakness in the research is the consideration of single macro cell based network models , lack of user mobilities in network models and lack of realistic network scenarios with densely deployed small cells. 5G networks are envisioned to be supporting huge density of users therefore network models needs to be more practical as per 5G requirements. Moreover, the true extent of artificial intelligence needs to be exploited in 5G heterogenous networks to meet the demands of increasing users.
|
\section{Introduction}
The creation of an effective quantum interface (QI) and long-lived quantum memory (QM) is of critical importance for quantum information technologies \cite{Hammerer2010,Kurizki2015}.
In practical implementation of long-lived multiqubit QM, it is required an implementation of sufficiently strong and reversible interaction of light/microwave qubits with many long-lived information carriers \cite{Roy2017}, in particular with NV-centers in diamond \cite{Jiang2009} and rare-earth ions in inorganic crystals \cite{Zhong2015}.
The best realization of this approach provided quantum storage with efficiency up to 92 \% \cite{Cho2016,Hsiao2018}, while full-fledged multifunctional quantum computer requires at least 99.9 \%.
The solution of the problem of high efficiency in operation of QI is possible due to the using the opportunities of rich dynamics in many-particle systems \cite{Hartmann2008,Hur2016,Noh2017}.
Herein, there is a basic problem in a construction of the multi-particle systems demonstrating controlled perfect time-reversible dynamics.
The considerable improvement of high-Q resonators \cite{Gorodetsky1999,Vahala2003,Kobe2017,Toth2017,Megrant2012} and their integration into multiresonator structures \cite{Armani2003,Liu2018,Xie2018,Flurin2015,Pfaff2017,Sirois2017} makes them interesting for use in broadband efficient optical and microwave QIs \cite{EMoiseev2017,Moiseev_2017_PRA,Moiseev2018} in which the resonators can have a specified periodic frequencies.
Such multi-resonator schemes demonstrate the possibility of a significant increase in the operating spectral range of the QI.
Moreover high quality of the resonators makes it possible to considerable enhance the constant coupling with light signals and resonant atomic ensembles herewith the broadband system of resonators allows reducing the effects of relaxation and decoherence due to the transition to faster storage processes.
These properties promise getting higher QM efficiency and using these systems in circuits of the universal quantum computers \cite{Kurizki2015,Perminov2017superefficient,Perminov2017spectral,Kockum2018}.
In this paper we show that combining the system of high-Q ring resonators with long-lived spin ensembles could be a realible cascade approach to obtain superefficient long-lived QM in practically realistic conditions.
We introduce the term "superefficient" by indicating the possibility to increase the quantum efficiency closely to 100 \% due to using the system of high-Q resonators providing an efficient broadband QI between free propagating signal photons and long-lived spin ensemble situated in these resonators.
The proposed cascade scheme consists of four high-Q ring resonators forming a controllable frequency comb (CFC) and interacting with long-lived resonant spin systems and with a common broadband waveguide (Fig.~\ref{Scheme}).
The cascade regime of the quantum scheme is realized in two stages.
Firstly the input signal is transferred to the ring resonators from the external waveguide and then the signal field is stored in the system of long-lived spin ensembles.
A small number of resonators greatly facilitates the spectral-topological optimization of parameters of the CFC QI scheme \cite{Perminov2017spectral}.
The simple model of spectral optimization proposed below was based for the fitting the transfer function (TF) of the studied system describing to the spectral characteristics of the quantum storage process.
It was shown that the long-lived CFC QM scheme does not impose large restrictions on the parameters of losses in resonators.
By using this optimization method we have found the optimal values of all free parameters of the CFC QI scheme and show the possibility of achieving the super high efficiency $\sim99.99\%$ in a wide frequency range.
Finally we discuss possible experimental implementations of the proposed scheme and it's potential application.
\begin{figure}[t]
\includegraphics[width = 0.45\textwidth]{figure_1.pdf}
\caption{Elementary building block of the scheme of controllable frequency comb quantum interface based on ring resonators.}
\label{Scheme}
\end{figure}
\section{Physical background}
The initial idea of the scheme under consideration is based on the photon echo QM approach \cite{Moiseev2001,Moiseev2004,Tittel2009} in a variant using resonant atoms with a periodic spectral structure of the inhomogeneous broadening of the absorption line, which is known as the AFC protocol \cite{Riedmatten2008,Akhmedzhanov2016}, and also on the realization of this approach in the optimal resonator \cite{Moiseev2010,Afzelius2010}.
This approach has been recently extended to systems of several resonators \cite{EMoiseev2017,Moiseev_2017_PRA,Perminov2017superefficient}.
For CFC QM, we analyze the dynamics of $N$ resonators and resonant two-level spin ensembles by using the quantum optics approach to the description of light in an open cavity \cite{Walls}.
Spin ensembles exist in each ring resonator and resonant frequencies of the spins close to the frequency of its ring cavity mode.
Herein resonant transition of each spin ensemble is characterized by inhomogeneous broadening and spectrum of the input signal field covers the spectrum of ring frequency combs.
We also assume that all two-level spin ensembles are prepared in the ground states $|G_s\rangle=\prod_{j=1}^N|g_j\rangle$ before signal pulses described by the initial state $|\psi_{in,1}\rangle$ is launched into the studied system.
In the framework of this approach, we get the system of Langevin-Heisenberg equations for the spin modes $s_{n,j}(t)$, the resonator field modes $b_n(t)$ and the input-output mode of the common waveguide $a_{in,out,n}(t)$:
\begin{eqnarray}\label{eq}
\nonumber \left[\partial_{t}+i(\tilde{\Delta}_n+\delta_{n,j})\right]s_{n,j}(t)+ig_nb_n(t)=0, &\\
\nonumber \left[\partial_t +i\Delta_n'+\frac{\kappa_n}{2}\right]b_n(t)+
i \sum_jg_ns_{n,j}(t)=\\ \nonumber k_n^{\frac{1}{2}}a_{in,n}(t)+{\gamma}_n^{\frac{1}{2}}F_n(t),& \\
\nonumber a_{in,n}(t)-a_{out,n}(t) =k_n^{\frac{1}{2}}b_n(t), &\\
a_{in,n+1}(t)=e^{i\varphi_n}a_{out,n}(t+(z_{n+1}-z_n)/c),
\end{eqnarray}
where $\varphi_n=\omega_0 (z_{n+1}-z_n)/c$, $z_n$ is the spatial position of the $n$-th resonator along the waveguide, $\omega_0$ is a central frequency,
$\delta_{n,j} $ is the frequency detuning of the $j$-th spin in the $n$-th spin system (reckoned from its line center) with detuning $\tilde{\Delta}_n$ ($j\in\{1,...,N_n\}$), $\{g_n,\delta_n=1/T_{2,n}^*\}$ is the field interaction constant in the $n$-th resonator with a single spin and the inhomogeneous linewidth of the $n$-th spin system (further we shall assume that the spin systems in the resonators are identical and $\delta_n=\tilde{\delta}$),
$\{\tilde{\Delta}_n=\Delta(n-\textrm{sgn}(n)/2),\Delta_n\}$ is the frequency detuning of centers of lines of spin systems and resonators, $n\in\{-N/2,...,N/2\}\backslash\{0\}$ (in this paper we consider the case of an even $N$); $\Delta_n'=\Delta_n-i\gamma_n$, where $\Delta_n$ and $\gamma_n$ are spectral detuning and decay constant of $n$-th cavity mode with corresponding Langevin forces $F_n(t) $\cite{Scully}.
Assuming the number of spins in each resonator is a quite large $N_n\gg1$, the discrete value $\delta_{n,j}$ is replaced by the frequency detuning $\delta$ with continuous distribution of the Lorentzian line shape $G_n(\delta)=\pi^{-1}{\delta}_n[\delta^2+{\delta}_n^2]^{-1}$.
Next we will use dimensionless units for all the parameters in (\ref{sol_eq}), which is equivalent to relating them to a certain unit of frequencies, assuming without loss of generality $\Delta=1$.
We also ignored the Langevin forces $F_n(t)$ \cite{Walls} in the equation (\ref{eq}), focusing only on the searching for the quantum efficiency in the CFC scheme studied and by the case of ultimately weak decay constants $\gamma_{1,...,N} \ll\kappa_1$ and $\gamma_{1,...,N} \tau\ll 1$ (where $\tau$ is a typical time of the studied processes).
\section{Recording/reading stage}
By taking the Fourier transform of (\ref{eq}), we obtain the system of algebraic equations for recording stage
\begin{eqnarray}\label{eq_four}
\nonumber i(\tilde{\Delta}_n+\delta_{n,j}-\omega)s_{n,j}+ig_nb_n=0, &\\
\nonumber \left[i(\Delta_n'-\omega)+\frac{\kappa_n}{2}\right]b_n+i \sum_jg_ns_{n,j}=k_n^{\frac{1}{2}}a_{in,n}, &\\
\nonumber a_{in,n}-a_{out,n}(\omega)=k_n^{\frac{1}{2}}b_n,&\\
a_{in,n+1}=e^{i(\varphi_n+\omega (z_{n+1}-z_n)/c)}a_{out,n}(\omega),
\end{eqnarray}
where for all the field and spin operators the Fourier transform is defined as $u(t)=[2\pi]^{-\frac{1}{2}}\int d\omega u(\omega) e^{-i\omega t}$,
where $\omega$ is the frequency counted from the central frequency of the radiation $\omega_0$.
We find the solution of (\ref{eq_four}) for the output field $a_{out,N}(\omega)=S(\omega)a_{in,1}(\omega)$ where TF is:
\begin{eqnarray}\label{sol_eq}
S(\omega)=e^{i\Phi(\omega)}\prod_{n}\frac{-\frac{k_n}{2}-i(\omega-\Delta_n')+\frac{N_ng_n^2}{{\delta}_n-i(\omega-\tilde{\Delta}_n)}}{\frac{k_n}{2}-i(\omega-\Delta_n')+\frac{N_n g_n^2}{{\delta}_n-i(\omega-\tilde{\Delta}_n)}},
\end{eqnarray}
where $\Phi(\omega)=(\omega+\omega_0)(z_N-z_1)/c$.
By using time-reversal property of photon echo protocols \cite{Moiseev2001,Moiseev2004,Riedmatten2008,Tittel2009,Moiseev2010} we can revise macroscopic spin coherence by rephasing spin states and initiating echo signal emission, respectively.
Following this way and taking into account weak spin decoherence $\gamma\ll\kappa_1$, we get from equation (\ref{eq}), the total QM efficiency $\eta_{tot}\approx\eta_{stor}^2$, where recording efficiency
\begin{eqnarray}\label{eff}
\eta_{stor}(\omega)= 1-|S(\omega)|^2-\sum_n\frac{\gamma_n\langle b_n^{\dagger}(\omega)b_n(\omega)\rangle}{\langle a_{in,1}^{\dagger}(\omega)a_{in,1}(\omega)\rangle}.
\end{eqnarray}
\noindent
where $\langle...\rangle$ are calculated by using initial field-atoms quantum state $|\Psi_{in}\rangle=|G_s\rangle|\psi_{in,1}\rangle$.
For calculation of optimal system parameters providing maximum quantum efficiency $\eta_{stor}(\omega)$, for simplicity we imply $\gamma_1=...=\gamma_N=\gamma=0$ (where $\eta_{stor}(\omega,\gamma=0)=\eta_{stor}^0(\omega)$).
In this case the upper bound of the efficiency is given by the expression $\eta_{stor}\approx\eta_{stor}^0-\xi\gamma/\kappa_1$ with $\xi\approx1\div10$, where
\begin{eqnarray}\label{eff0}
\eta_{stor}^0(\omega)=1-|S(\omega)|^2,
\end{eqnarray}
actually determines the upper limit of the qualitative capabilities of the studied cascade CFC QM for the storage of input signal field in the several long-lived spin ensembles via the ring high-Q resonators.
Further relaxation during storage time $T$ will be determined by the decoherent processes (with long decay time $T_2\gg 1/\kappa_1$) in the long-lived spin system ($\eta_{tot}^0(\omega,T)\approx e^{-2T/T_2}\eta_{stor}^0(\omega)$ for the input single-photon field).
\section{Optimization procedure}
In general, TF (\ref{sol_eq}) has very complex spectral properties \cite{Swanson2007,Koziel2011,Tamiazzo2017} and many mathematical aspects of our topic intersect with the fundamental problems of the theory of filters.
However, we show that under the certain conditions, the real system can acquire the spectral properties of TF corresponding to the highly effective sensor \cite{Cheng2014}, broadband filter \cite{Swanson2007,Rosenberg2013}, QM \cite{Perminov2017spectral} or QI.
An ideal broadband QI with an infinite working spectral bandwidth corresponds to $S(\omega)=0$.
Considering the last equality for $S(\omega)$ as an approximate condition for the problem of algebraic polynomial optimization, we have:
$S(0)\rightarrow0,\sum\limits_{m=1}^{N_{opt}}|\textrm{numer}(S(m\tilde{\Delta}_n/(2N_{opt})))|^2\rightarrow \textrm{min},N_{opt}=2N-1$
for $4N+1$ free parameters $\{\tilde{\delta}={\delta}_n,g_n,\Delta_n\}$ (where $\textrm{numer} (S(\omega))$ is the numerator of the rational function $S(\omega)$).
We solved the equality $S(\omega)\approx0$ accurately only for several spectral points.
We note that the first of the conditions of the form $S(0)=0$ is analogous to the standard condition for a impedance matched resonator QM \cite{Moiseev2010,Moiseev_2017_PRA}, which determines the spectral efficiency in the central part of the band for a broadband QM, and additional conditions allow improving the spectral efficiency at edges of the used frequency band.
Here we consider the particular case of four resonators ($N=2$) and after numerical calculations (for $\Delta=1$), we obtain a configuration which is a symmetric with respect to the indices
$\tilde{\delta}=1.8,k_{\mp1}=3.27,k_{\mp2}=2.03,N_1^{\frac{1}{2}}g_{\mp1}=1.78,N_2^{\frac{1}{2}}g_{\mp2}=1.49,\Delta_{\mp1}=0.48,\Delta_{\mp2}=1.13$ and the intensity for the reflection spectrum $|S(\omega)|^2$ depicted in figure \ref{int_spectra}.
\begin{figure}[t]
\includegraphics[width = 0.45\textwidth]{figure_2.pdf}
\caption{Intensity of the reflected spectrum $|S(\omega)|^2$ for the optimal CFC QI.}
\label{int_spectra}
\end{figure}
From where we found the optimal absorption coefficients of the resonator modes by spin ensembles $N_1g_1^2T_2^*=1.76$ and $N_2g_2^2T_2^*=1.23$.
The absorption coefficients are much smaller than the optimal absorption coefficient $\kappa$, which should be realized at the usual impedance matching conditions if the spin system will be placed in one common resonator with spectral width $\kappa\gg\Delta=1$ \cite{Moiseev2010}.
This observation shows that the total number of spins can be reduced in the studied cascade CFC QM scheme.
The spectral behavior of $|S(\omega)|^2$ in figure \ref{int_spectra} demonstrates the high quality of the CFC QI.
$|S(\omega)|^2$ is characterized by an almost rectangular spectral plateau of a fairly homogeneous form in the frequency range $\omega\in[-1.5;1.5]$, where the efficiency $\eta_{stor}^0$ of the absorption of the signal field by the CFC system reaches $99.99$ \%.
\section{Conclusion}
First experimental realization of QI on the system of ring microwave resonators \cite{Petrovnin2018} demonstrated a possibility of broadband storage for small number of resonators.
In this paper, we theoretically show that optimization of all the parameters in CFC QI with 4 ring resonators is possible for the signal storage on 4 spin ensembles covering a wide frequency band with a quite large efficiency 99.99\%.
The storage requires using sufficiently high-quality microresonators \cite{Megrant2012,Armani2003,Vahala2003} for which the total losses are limited by the effective decay constant $\sim\gamma/k_1+T/T_2$.
Rather simple optimization scheme used by us at several spectral points does not require large computational resources and it can be also analyzed analytically on the basis of applied methods of algebraic geometry \cite{Amari2006,Shakirov2007,Dolotin2007}.
The studied scheme can be also realized on the system of high-quality WGM (whispering gallery modes) microresonators \cite{Gorodetsky1994,Gorodetsky1999} coupled to the external optical waveguide by including resonant atomic ensembles in each microresonator.
Herein the coherent control of the optical atomic coherence can be carried out by an additional lasers tuned to other eigenfrequencies of WGM microresonators.
In addition, the proposed cascade CFC QI can be used to combine several different types of QM devices (for the systems with different types of inhomogeneous broadening) into a single broadband QM block with a higher quality spectral profile \cite{Perminov2017spectral}.
In the proposed way it is possible to create a superefficient hybrid QM devices \cite{Kurizki2015}, which can be integrated into the quantum computer circuit on the basis of already existing technologies \cite{Brecht_2016,Pierre2014,Du2016,Melloni2010,Romanenko2014,Huet2016,Gu2017}.
\ack
This work was financially supported by a grant of the Government of the Russian Federation, project No. 14.Z50.31.0040, February 17, 2017.
\section*{References}
\bibliographystyle{iopart-num}
|
\section{INTRODUCTION}
As we move to increasingly complex systems \cite{Malikopoulos2015} new decentralized control approaches are needed to optimize the impact on system behavior of the interaction between its entities \cite{Malikopoulos}. Centralized stochastic control has been the ubiquitous approach to control complex systems so far \cite{Malikopoulos2016c}. A key assumption in centralized stochastic control problems is that a singular decision maker perfectly recalls all previous control actions and observations. The information available to an agent when making a decision is called the \textit{information structure} of the system. The centralized information structure is classified as the \textit{classical information structure}.
While centralized systems have been extensively studied \cite{23}, the classical information structure does not apply to many applications involving multiple agents \cite{Malikopoulos2018}. In these applications, all agents simultaneously make a decision based only on their local information and information received through delayed, or costly communication with other agents \cite{Malikopoulos2018d}. Thus, a centralized knowledge of the complete information in the system is infeasible \cite{18}. These information structures are classified as \textit{non-classical information structures}. Ultimately, we end up with multi-stage optimization problems \cite{Malikopoulos2015b}, known as decentralized stochastic control problems.
Decentralized stochastic control has proven to be very challenging as the most common approach to derive centralized optimal control policies, dynamic programming (DP), is not directly applicable to non-classical information structures due to a lack of separation between estimation and control. There are three general approaches in the literature for these problems that use techniques from centralized stochastic control: (1) the person-by-person approach, (2) the designer's approach, and (3) the common information approach. Due to space limitations, it is very difficult to cite all the literature around these three approaches. For more details, the reader may refer to the tutorial by Mahajan et al. \cite{Mahajan2012} and the references therein.
1) In the \textit{person-by-person approach}, the control strategies of all agents except one are arbitrarily fixed. Only the control strategy of the chosen agent is then optimized for this new centralized problem. Repeating this process for all agents allows for the derivation of \textit{structural results} and DP for a person-by-person optimal strategy, that is not globally optimal in general. However, every globally optimal strategy must necessarily be person-by-person optimal. Some applications of this approach can be found in \cite{4,5,11,7,8,6,2,1}.
2) The \textit{designer's approach} takes the point of view of a designer with knowledge of the system model and statistics. The designer's task is the \text{blue}{computationally challenging selection of} the globally optimal control strategy for the system by transforming the problem into a centralized planning problem. Some applications of this approach can be found in \cite{3,19,10,lall_broadcast}.
3) A more recent development in this field is the \textit{common information approach} developed for problems with partial history sharing \cite{17}, and then formalized for general decentralized systems \cite{14}. The solution is derived by reformulating the system from the viewpoint of a fictitious \textit{coordinator} whose task is to prescribe control laws to every agent in the system. Some applications can be found in a variety of problems including symmetric delayed information sharing structures \cite{14}, control sharing information structures \cite{16}, stochastic games with asymmetric information \cite{20}, teams with mean-field sharing \cite{Mahajan2015} and systems with unreliable uplink channels \cite{asghari2018optimal}.There are some earlier papers that used similar ideas in analyzing specific information structures \cite{yuksel2009stochastic, yoshikawa1978decomposition, aicardi1987decentralized, casalino1984partially}.
In this paper, we introduce and study a decentralized system with multiple agents who communicate with \textit{word of mouth}. In a word-of-mouth communication, we have a network of agents, where each agent may directly communicate only with its neighbors in the network. Thus, information from each agent propagates in the network through its neighbors who share it with their neighbors, and so on. Each link in the network has a delay associated with it, which can be thought of as the time it takes for the information to transmit from an agent to its neighbor. This problem has a non-classical information structure because of the delays in communication.
We consider the common information approach to be the standard approach in solving a wide variety of decentralized stochastic control problems, including problems with a word-of-mouth information structure. However, we find that in problems with asymmetric communication, there may not be a lot of common information available to all agents in the system \cite{Aditya_2019}. This has motivated us to continue looking for structural results that can improve on the performance of the common information approach by taking into account the asymmetries in a system.
The contributions of this paper are:
1) We introduce and analyze a problem with a word-of-mouth information structure.
2) We present the prescription approach and its properties which lead to a reformulation of the problem from the point of view of every agent, with a state sufficient for input output mapping and information state for each reformulation.
3) We state some preliminary structural results with time-invariant domains that arise from the prescription approach.
The rest of the paper is organized as follows. In Section II, we present the problem and the information structure of the system. In Section III, we provide a reformulation of the problem and, in Section IV, we derive the preliminary results for optimal strategies. Finally, in Section V, we draw concluding remarks, and present some ideas for future work.
\subsection{Notation}
Random variables are denoted by upper case letters and their realization by the corresponding lower case letters. For integers $a<b$, $X_{a:b}$ is shorthand for the vector $(X_a,X_{a+1},\ldots,X_b)$ and $X^{a:b}$ is shorthand for the vector $(X^a,X^{a+1},\ldots,X^b)$. When $a>b$, the dimension of $X^{a:b}$ is 0. The combined notation with $c<d$ and $a<b$, we write $X_{a:b}^{c:d}$ to denote the vector $(X_i^j: i=a,\ldots,b; \, j=c,\ldots,d)$.
For sets $A$ and $B$, $\{A,B\}$ is the set $A \cup B$. For a singleton $\{a\}$ and set $B$, $\{a,B\}$ is the set $\{a\} \cup B$. The function $|\cdot|$ returns the cardinality of a set. The null set is represented by $\emptyset$. We have attempted to use notation consistent with \cite{17} as our work is closely related to it.
The probability and expectation measures that depend on a vector $\boldsymbol{g}$ are written as $\mathbb{P}^{\boldsymbol{g}}(\cdot)$ and $\mathbb{E}^{\boldsymbol{g}}(\cdot)$ respectively.
All equalities involving random variables hold with a probability of 1.
\section{PROBLEM FORMULATION}
\subsection{The Network of Agents}
Consider a network of $K \in \mathbb{N}$ agents represented by a strongly connected graph $\mathcal{G} =(\mathcal{K}, \mathcal{E})$, where $\mathcal{K} := \{1,\ldots,K\}$ is the set of agents and $\mathcal{E}$ is the set of links. A link from an agent $k \in \mathcal{K}$ to an agent $j \in \mathcal{K}$ is denoted by $(k,j) \in \mathcal{E}$. Every link $(k,j)$ represents a communication link from agent $k$ to $j$ which is characterized by a delay of $\delta^{[k,j]} \in \mathbb{N}$ time steps for transferring information from $k$ to $j$.
When agent $k$ sends out information to agent $j$ through link $(k,j)$, we call it \textit{transmission of information}. The information transmitted by agent $k$ at time $t$ is received by agent $j$ at time $t + \delta^{[k,j]}$. For any agent $k$, the acts of receiving and transmission of information occur at different instances within every time step as discussed in Section II-D.
\begin{definition}
\label{path}
Let $\mathcal{N} = \{1,\ldots,m : m \in \mathcal{K}\}$ be a set of indices. For any $k,j \in \mathcal{K}$, a \textit{path} $q^{[k,j]}_a$, $a \in \mathbb{N}$, from $k$ to $j$ is given by the sequence $\{k_n\}_{n\in\mathcal{N}}$ such that: (1) $k_1 = k$ and $k_m = j$, (2) $k_n \in \mathcal{K}$ for $n \in \mathcal{N}$, and (3) there exists a link $(k_{n-1},k_n) \in \mathcal{E}$ for $n \in \mathcal{N}\setminus\{1\}$.
\end{definition}
The set $\mathcal{Q}^{[k,j]} = \{q^{[k,j]}_a: a =1,\ldots,b; \; b \in \mathbb{N}\}$ includes all paths from agent $k$ to agent $j$.
\begin{definition}
\label{path_delay}
Let agents $k,j \in \mathcal{K}$ with a path $q^{[k,j]}_a$ from $k$ to $j$. The \textit{communication delay} $d^{[k,j]}_a \in \mathbb{N}$ for $q^{[k,j]}_a$ is defined as
\begin{equation*}
d^{[k,j]}_a = \delta^{[k,k_2]}+\cdots+\delta^{[k_{m-1},j]},
\end{equation*}
where $\delta^{[k_{n-1},k_n]}$ is the delay in information transfer through the link $(k_{n-1},k_n) \in \mathcal{E}$.
\end{definition}
The \textit{information path}, defined formally next, from agent $k$ to agent $j$ in the network is the path with the least possible delay.
\begin{definition}
The \textit{information path} from $k$ to $j$ denoted by $(k \rightarrow j)$ is given by a path $q_{{a}}^{[k,j]} \in \mathcal{Q}^{[k,j]}$ such that,
\begin{gather}
d^{[k,j]}_{{a}} = \min\left\{d_1^{[k,j]},\ldots,d_b^{[k,j]}\right\},
\end{gather}
where $b := |\mathcal{Q}^{[k,j]}|$.
\end{definition}
The strongly connected nature of the network ensures that there is always an information path $(k \rightarrow j)$ from every other agent $k \in \mathcal{K}$ to every agent $j \in \mathcal{K}$. We denote the associated delay simply by $d^{[k,j]}$ and, by convention, we set $d^{[k,k]} = 0$. Also note that because the links in the network are directed, the delay $d^{[k,j]}$ in communication from $k$ to $j$ is not equal to the delay $d^{[j,k]}$ in communication from $j$ to $k$.
\subsection{System Description}
The network of agents is considered a discrete time system that evolves up to a finite time horizon $T \in \mathbb{N}$. At time $t \in \mathcal{T}$, $\mathcal{T} = \{0,1,\ldots,T\}$, the state of the system $X_t$ takes values in a finite set $\mathcal{X}$ and the control variable $U_t^k$ associated with agent $k \in \mathcal{K}$, takes values in a finite set $\mathcal{U}^k$. Let ${U}_t^{1:K}$ denote the vector $(U_t^1,\ldots,U_t^K)$. Starting at the initial state $X_0$, the evolution of the system follows the state equation
\begin{equation}
X_{t+1}=f_t\left(X_t,U_t^{1:K},W_t\right), \label{st_eq}
\end{equation}
where $W_t$ is the uncontrolled disturbance to the system represented as a random variable taking values in a finite set $\mathcal{W}$. At time $t$ every agent $k$ makes an observation $Y_t^k$, given by
\begin{equation}
Y_t^k=h_t^k(X_t,V_t^k), \label{ob_eq}
\end{equation}
which takes values in a finite set $\mathcal{Y}^k$ through a noisy sensor, where $V_t^k$ takes values in the finite set $\mathcal{V}^k$ and represents the noise in measurement.
Agent $k$ selects a control action $U_t^k$ from the set of feasible control actions $\mathcal{U}_t^k$ as a function of its information structure. The information structure is different for each agent $k \in \mathcal{K}$ because of the means of communication and topology of the network. We discuss the information structure in Section II-D. After each agent $k$ generates a control action $U_t^k$, the system incurs a cost $c_t(X_t,U_t^{1:K}) \in \mathbb{R}$.
\subsection{Assumptions}
In our modeling framework above, we impose the following assumptions:
\begin{assumption}
\label{top_asu}
The network topology is arbitrary, known a priori, and does not change with time.
\end{assumption}
With a known and invariable network topology, every agent can keep track of what information is accessible to other agents in the network.
\begin{assumption}
\label{prim_asu}
The external disturbance $\{W_t: t \in \mathcal{T}\}$ and the noise in measurement $\{V_t^k: t \in \mathcal{T}, \, k \in \mathcal{K}\}$ are sequences of independent random variables that are also independent of each other and of the initial state $X_0$.
\end{assumption}
The external disturbance, noise in measurement, and initial state are referred to as the \textit{primitive} random variables, and they have known probability distributions.
\begin{assumption}
\label{func_asu}
The state functions $(f_{t}: t \in \mathcal{T})$, observation functions $(h_{t}^{k}: t \in \mathcal{T}, \, k \in \mathcal{K})$, the cost functions $(c_{t}: t \in \mathcal{T}),$ and the set of all feasible control policies $G$ are known to all agents.
\end{assumption}
These functions and the set of feasible control policies (explained in Section II-D) form the basis of the decision making problem.
\begin{assumption}
Each agent has perfect recall.
\end{assumption}
Perfect recall of the data from the memory of every agent is an essential assumption for the structural results derived in this paper.
\begin{figure}[ht!]
\centering
\captionsetup{justification=centering}
\includegraphics[width=0.9\linewidth, keepaspectratio]{seq.PNG}
\caption{Sequence of activities.}
\label{fig:update}
\end{figure}
We summarize below the sequence of activities taken by agent $k \in \mathcal{K}$ at time $t$ (Fig. \ref{fig:update}):
\begin{enumerate}
\item The state $X_t$ is updated based on (\ref{st_eq}).
\item Agent $k$ receives information from all agents in $\mathcal{K}$, collectively denoted by $\underline{I}_t^k$.
\item Agent $k$ makes an observation about the state $Y_t^k$ based on (\ref{ob_eq}).
\item Agent $k$ updates its memory, $M_t^k$, defined in Section II-D, on a given protocol.
\item Agent $k$ transmits information denoted by $\overline{I}_t^k$ to every agent $j \in \mathcal{K}$ through the shortest path $q_{{a}}^{[k,j]}$.
\item Agent $k$ generates a control action $U_t^k$.
\end{enumerate}
\subsection{Information Structure of the System}
The information structure of the system is characterized by the graph topology and delays along communication paths described in Section II-A. In the word-of-mouth information structure, every agent $j \in \mathcal{K}$ at time $t$ transmits the information $\overline{I}_t^j := \{Y_{t}^j,U_{t-1}^j\}$ to every other agent in the network through the relevant information paths. Agent $k \in \mathcal{K}$ receives information $\overline{I}_t^j$ at time $t + d^{[j,k]}$, where $d^{[j,k]}$ is the communication delay from $j$ to $k$.
Then, the information available to agent $k$ at time $t$ is the collection of information he received from every agent $j \in \mathcal{K}$ at time steps $0$ through $t$.
\begin{definition}
The \textit{memory} of agent $k \in \mathcal{K}$ is defined as the random variable $M_t^k$ that takes values in the finite set $\mathcal{M}_t^k$ and is given by
\begin{align}
M_t^k := &\left\{Y^j_{0:t-d^{[j,k]}},U_{0:t-d^{[j,k]}-1}^j: j \in \mathcal{K} \right\}, \label{eq_mem}
\end{align}
where $d^{[j,k]}$ is the delay in information transfer from every agent $j \in \mathcal{K}$ to agent $k$.
\end{definition}
At time $t$, agent $k$ accesses his memory $M_t^k$ to generate a control action, namely,
\begin{gather}
U_t^k := g_t^k(M_t^k), \label{u_basic}
\end{gather}
where $g_t^k$ is the control policy of agent $k$ at time $t$. We define the control policy for each agent as $\boldsymbol{g}^k := (g_0^k,\ldots,g_T^k)$ and the control policy of the system as $\boldsymbol{g} := (\boldsymbol{g}^1,\ldots,\boldsymbol{g}^K)$. The set of all feasible control policies is denoted by $G$.
The performance criterion for the system is given by the total expected cost:
\begin{equation}
\textbf{Problem 1:}~~~~~ \mathcal{J}(\boldsymbol{g}) = \mathbb{E}^{\boldsymbol{g}}\left[\sum_{t=0}^T{c_t(X_t,U_t^{1:K})}\right], \label{per_cri}
\end{equation}
where the expectation is with respect to the joint probability measure on the random variables $\{X_t, U_t^1,\ldots,U_t^K\}$.
The optimization problem is to select the optimal control policy $\boldsymbol{g}^* \in G$ that minimizes the performance criterion in \eqref{per_cri}, given the probability distributions of the primitive random variables $\{X_0,W_{0:T},V_{0:T}^1,\ldots,V_{0:T}^K\}$, and functions $\left\{c_t,f_t,h_t^{k}:t \in \mathcal{T}, \, k \in \mathcal{K} \right\}$.
\section{THE PRESCRIPTION APPROACH}
\subsection{Construction of Prescriptions}
For an agent $k \in \mathcal{K}$, we consider a scenario where the control action $U_t^k$ is generated in two stages:
(1) Agent $k$ generates a function based on information which is a subset of the information available in its memory $M_t^{{k}}$.
(2) This function takes as an input the compliment of the subset used to generate it, and yields the control action $U_t^{{k}}$.
We call these functions \textit{prescriptions}. They allow us to construct an optimization problem of selecting the optimal \textit{prescription strategy} that is equivalent to the problem of selecting the optimal control policy $\boldsymbol{g}^{*k}$ as we show next. In this section, we construct the subset of the memory $M_t^{{k}}$ and prescriptions for every agent ${{k}} \in \mathcal{K}$ without changing the information structure of the system. In order to simplify the notation, we first define the set of agents located beyond agent $k$ in the set of all agents.
\begin{definition}\label{def:setB}
For an agent ${{k}} \in \mathcal{K}$, the set of agents beyond ${{k}}$ is defined as $\mathcal{B}^{{{k}}} := \{{{j}} \in \mathcal{K}: {{j}} \geq {{k}}\}$.
\end{definition}
Now we can define the information used to generate prescriptions.
\begin{definition} \label{def_A}
Let $k \in \mathcal{K}$ and $M_t^{{k}}$ be the agent's memory at time $t$. The \textit{accessible information} of agent ${{k}}$ is defined as the set $A_t^{{k}}$ that takes values in the finite collection of sets $\mathcal{A}_t^{{k}}$ such that
\begin{gather}
A_t^{{k}} = \bigcap_{{{i}}=1}^{{k}}\left(M_t^{{i}}\right). \label{ainfo_def}
\end{gather}
\end{definition}
For example, we can write \eqref{ainfo_def} for agent $1$ as
\begin{gather}
A_t^{{1}} = M_t^{{1}},
\end{gather}
and for agent $2$ as
\begin{gather}
A_t^{{2}} = M_t^{{1}} \cap M_t^{{2}}.
\end{gather}
Based on Definition \ref{def_A}, the accessible information $A_t^{{k}}$ has the following properties:
\begin{align}
& A_{t-1}^{{k}} \subset A_t^{{k}}, \label{ainfo_prop_2} \\
& A_t^{{j}} \subset A_t^{{k}}, \; \forall {{j}} \in \mathcal{B}^{{k}}, \label{ainfo_prop_1}
\end{align}
where $\mathcal{B}^{{k}}$ is the set of agents beyond ${{k}}$.
Property \eqref{ainfo_prop_2} motivates the introduction of a new term to denote the new information added to accessible information $A_t^{{k}}$ at time $t$.
\begin{definition}
The \textit{new information} for agent $k$ at time $t$ is defined as the set $Z_t^{{k}}$ that takes values in a finite collection of sets $\mathcal{Z}_t^{{k}}$ such that
\begin{align}
Z_t^{{k}} := A_t^{{k}} \backslash A^{{k}}_{t-1}. \label{Z_def}
\end{align}
\end{definition}
We observe in \eqref{ainfo_prop_1} that the accessible information $A_t^{{j}}$ of any agent ${{j}} \in \mathcal{B}^{{k}}$ is a subset of the memory $M_t^{{k}}$. Thus, we can define the \textit{inaccessible information} of the agent ${{k}}$ with respect to the accessible information $A_t^{{j}}$ for every ${{j}} \in \mathcal{B}^{{k}}$.
\begin{definition}
The \textit{inaccessible information} of agent ${{k}}$ with respect to accessible information $A_t^{{j}}$, ${{j}} \in \mathcal{B}^{{k}}$, is defined as the set of random variables $L_t^{[{{k}},{{j}}]}$ that takes values in the finite collection of sets $\mathcal{L}_t^{[{{k}},{{j}}]}$ such that
\begin{gather}
L_t^{[{{k}},{{j}}]} := M_t^{{k}} \setminus A_t^{{j}}. \label{inacc_def}
\end{gather}
\end{definition}
The pair of sets $A_t^{{j}}$ and $L_t^{[{{k}},{{j}}]}$ forms a partition of the set $M_t^{{k}}$, such that
\begin{gather}
M_t^{{k}} = \{L_t^{[{{k}},{{j}}]},A_t^{{j}}\}, \; \forall {{j}} \in \mathcal{B}^{{k}}. \label{eq_partition}
\end{gather}
As an example, consider a system with three agents Fig. \ref{fig:inaccessible}. In this system, we have
\begin{align}
&A_t^{1} = M_t^{1}, \nonumber \\
&A_t^{3} \subset A_t^{2} \subset M_t^{1}, \nonumber \\
&M_t^{1} = \{A_t^{2},L_t^{[{{1}},{{2}}]}\} = \{A_t^{3},L_t^{[{{1}},{{3}}]}\}.
\end{align}
For agents $2$ and $3$, we can derive similar relationships as illustrated in Fig. \ref{fig:inaccessible}.
\begin{figure}[h!]
\centering
\captionsetup{justification=centering}
\includegraphics[height = 8cm, keepaspectratio]{inaccessible.PNG}
\caption{Memory partitions of three agents.}
\label{fig:inaccessible}
\end{figure}
Next, we use these partitions of the memory to define the \textit{prescription function}.
\begin{definition}
The \textit{prescription function} $\Gamma_t^{[{{k}},{{j}}]}$ of an agent ${{k}} \in \mathcal{K}$ for the agent ${{j}} \in \mathcal{K}$ is defined as follows
\begin{align} \label{pres_func_def}
\Gamma_t^{[{{k}},{{j}}]} :
\begin{cases}
\mathcal{L}_t^{[{{j}},{{k}}]} \longrightarrow \mathcal{U}_t^{{j}}, \text{ if } {{j}} \not\in \mathcal{B}^{{k}}, \\
\mathcal{L}_t^{[{{j}},{{j}}]} \longrightarrow \mathcal{U}_t^{{j}}, \text{ if } {{j}} \in \mathcal{B}^{{k}},
\end{cases}
\end{align}
and takes values in the set of feasible prescription functions $\mathscr{G}_t^{[{{k}},{{j}}]}$.
\end{definition}
\begin{remark}
In Definition 9, the inaccessible information of agent $k$ is defined with respect to the accessible information $A_t^{{j}}$ for ${{j}} \in \mathcal{B}^{{k}}$. Note that in the first part of \eqref{pres_func_def}, we have ${{k}} \in \mathcal{B}^{{j}}$, and thus \eqref{pres_func_def} holds.
\end{remark}
Every prescription function $\Gamma^{[{{k}},{{j}}]}$ is generated as follows
\begin{align} \label{eq_gen_pres}
\Gamma_t^{[{{k}},{{j}}]} :=
\begin{cases}
{\psi}_t^{[{{k}},{{j}}]}(A_t^{{k}}), \text{ if } {{j}} \not\in \mathcal{B}^{{k}}, \\
{\psi}_t^{[{{k}},{{j}}]}(A_t^{{j}}), \text{ if } {{j}} \in \mathcal{B}^{{k}},
\end{cases}
\end{align}
where we call ${\psi}_t^{[{{k}},{{j}}]}$ the \textit{prescription strategy} of the agent ${{k}}$ for the agent ${{j}}$ given by the mapping
\begin{align}
\psi_t^{[{{k}},{{j}}]} :
\begin{cases}
\mathcal{A}_t^{{{k}}} \longrightarrow \mathscr{G}_t^{[{{k}},{{j}}]}, \text{ if } {{j}} \not\in \mathcal{B}^{{k}}, \\
\mathcal{A}_t^{{{j}}} \longrightarrow \mathscr{G}_t^{[{{k}},{{j}}]}, \text{ if } {{j}} \in \mathcal{B}^{{k}}.
\end{cases}
\end{align}
We call $\boldsymbol{\psi}^{{k}} := (\boldsymbol{\psi}^{[{{k}},{{1}}]},\ldots,\boldsymbol{\psi}^{[{{k}},{{K}}]})$ the \textit{prescription strategy} of the agent ${{k}}$. The set of feasible prescription strategies for the agent ${{k}}$ is denoted by $\Psi^{{k}}$.
\begin{remark}
The prescription $\Gamma_t^{{[{{k}},{{j}}]}}$ of agent $k$ for agent $j$ is only available to agent $k$. The equivalent prescription available to agent $j$ is $\Gamma_t^{{[{{j}},{{j}}]}}$. The relationship between the two is given in Lemmas \ref{lem_psi_relation1} and \ref{lem_psi_relation2} in Section III-B.
\end{remark}
\begin{remark}
Every agent needs to generate prescriptions corresponding to every other agent in the system so that we can define the \textit{information state} in Section IV-B.
\end{remark}
Next, we define the \textit{complete prescription} of an agent ${{k}}$ below.
\begin{definition}
The \textit{complete prescription} for agent ${{k}}$ is given by the function
\begin{align}
\Theta_t^{{k}} : \;
\mathcal{L}_t^{[{{1}},{{k}}]} \times \cdots \times \mathcal{L}_t^{[{{k}},{{k}}]} &\times \mathcal{L}_t^{[{{k+1}},{{k+1}}]} \times \cdots \times \mathcal{L}_t^{[{{K}},{{K}}]} \nonumber \\
&\longrightarrow \mathcal{U}_t^{{1}} \times \cdots \times \mathcal{U}_t^{{k}},
\end{align}
which takes values in the set of functions $\mathscr{G}_t^{{k}}$.
\end{definition}
The complete prescription for agent ${{k}}$ is constructed as $\Theta_t^{{k}} = (\Gamma_t^{[{{k}},{{1}}]},\ldots,\Gamma_t^{[{{k}},{{K}}]})$.
\subsection{Relationships Between Prescriptions and Control Policies}
In this section we present the relationships between the different prescriptions and control policies.
The first result states that for an agent $k \in \mathcal{K}$ we can use the complete prescription $\Theta_t^{{k}}$ to generate control action $U_t^k$ instead of the control policy $g_t^k$.
\begin{lemma} \label{lem_psi_g_relation}
Let agent $k \in \mathcal{K}$ and let $\Theta_t^{{k}}$ be its complete prescription. For any given control policy $\boldsymbol{g} \in G$, there exists a prescription strategy $\boldsymbol{\psi}^{{k}} \in \Psi^{{k}}$ such that
\begin{gather}
U_t^{{k}} = \Gamma_t^{[{{k}},{{k}}]}\left(L_t^{[{{k}},{{k}}]}\right). \label{u_presc}
\end{gather}
\end{lemma}
\begin{proof}
Let $A_t^{{k}}$ and $L_t^{[{{k}},{{k}}]}$ be the accessible and inaccessible information, respectively, of agent $k$. For any control policy $\boldsymbol{g}$ that generates $U_t^k$ from \eqref{u_basic}, we can select a prescription strategy $\psi_t^{{k}} : \mathcal{A}_t^{{{k}}} \longrightarrow \mathscr{G}_t^{[{{k}},{{k}}]}$ such that
\begin{equation}
\Gamma_t^{[{{k}},{{k}}]}(\cdot) = \psi_t^{{k}}(A_t^{{{k}}})(\cdot) = g_t^{{k}}(A^{{k}}_t,\cdot).
\end{equation}
Then, the control action is
\begin{gather}
U_t^{{k}} = \Gamma_t^{[{{k}},{{k}}]}(L^{[{{k}},{{k}}]}_t)=g_t^{{k}}(A^{{k}}_t,L^{[{{k}},{{k}}]}_t) = g_t^{{k}}(M_t^{{k}}).
\end{gather}
\end{proof}
Similarly, for any prescription strategy $\boldsymbol{\psi}^{{k}}$, we can construct an appropriate control policy $\boldsymbol{g}$ that generates the same control actions $U_t^{{k}}$ for all agents in $\mathcal{K}$.
\begin{lemma}\label{lem_psi_g_relation_inv}
Let agent $k \in \mathcal{K}$ and let $\Theta_t^{{k}}$ be its complete prescription. For any given prescription strategy $\boldsymbol{\psi}^{{k}} \in \Psi^{{k}}$, there exists a control policy $\boldsymbol{g} \in G$ such that
\begin{gather}
U_t^{{k}} = \Gamma_t^{[{{k}},{{k}}]}(L_t^{[{{k}},{{k}}]}) = g_t^{{k}}(M_t^{{k}}). \label{u_presc_inv}
\end{gather}
\end{lemma}
\begin{proof}
For any prescription strategy $\boldsymbol{\psi}^{{k}},$ we can construct a control policy $\boldsymbol{g}^{{k}}$ such that
\begin{align}
U_t^{{k}} = g_t^{{k}}(M^{{k}}_t) = &g_t^{{k}}(A^{{k}}_t,L^{[{{k}},{{k}}]}_t) \nonumber \\
= &{\psi}_t^{[{{k}},{{k}}]}(A_t^{{k}})(L^{[{{k}},{{k}}]}_t).
\end{align}
\end{proof}
Lemmas \ref{lem_psi_g_relation} and \ref{lem_psi_g_relation_inv} imply that the control action $U_t^{{k}}$ of every agent $k \in \mathcal{K}$ generated through a prescription strategy $\boldsymbol{\psi}^{{k}}$, can also be generated through an appropriate policy $\boldsymbol{g}$ and vice versa.
\begin{definition}
Given two agents $k,j \in \mathcal{K}$, a \textit{positional relationship} from agent ${{k}}$ to agent ${{j}}$ is given by the function
\begin{gather}
e^{[{{j}},{{k}}]}: \Psi^{{k}} \longrightarrow \Psi^{{j}}.
\end{gather}
\end{definition}
Next we show the existence of a positional relationship $e^{[{{j}},{{k}}]}$ from any agent ${{k}} \in \mathcal{K}$ to every agent ${{j}} \in \mathcal{K}$ with desirable properties that allow us to construct optimal control policies of all agents from the optimal prescription strategy of just one agent. The following result establishes that using a positional relationship $e^{[{{j}},{{k}}]}=(e_1^{[{{j}},{{k}}]},\ldots,e_T^{[{{j}},{{k}}]})$, an agent ${{j}}$ can derive the prescription strategy for agent ${{i}} \in \mathcal{K}$, when given the prescription strategy of agent ${{k}}$ for agent ${{i}}$, namely
\begin{gather}
{\psi}_t^{[{{j}},{{i}}]} := e_t^{[{{j}},{{k}}]}\Big({\psi}_t^{[{{k}},{{i}}]}\Big), \; \forall {{i}} \in \mathcal{K}. \label{eq_e}
\end{gather}
\begin{lemma} \label{lem_psi_relation1}
Let agent $k \in \mathcal{K}$ and agent ${{j}} \in \mathcal{B}^{{k}}$. For any given prescription strategy $\boldsymbol{\psi}^{{k}}$ of agent ${{k}}$, there exists a positional relationship $e^{[{{j}},{{k}}]}$ such that a prescription strategy $\boldsymbol{\psi}^{{j}}$ of agent ${{j}}$ generated using \eqref{eq_e} yields:
\begin{align}
\text{\emph{1. }} &{\Gamma}_t^{[{{k}},{{i}}]}(L_t^{[{{i}},{{i}}]}) = {\Gamma}_t^{[{{j}},{{i}}]}(L_t^{[{{i}},{{i}}]}), \text{ \emph{if} } {{i}} \in \mathcal{B}^{{j}}, \nonumber \\
\text{\emph{2. }} &{\Gamma}_t^{[{{k}},{{i}}]}(L_t^{[{{i}},{{i}}]}) = {\Gamma}_t^{[{{j}},{{i}}]}(L_t^{[{{i}},{{j}}]}), \text{ \emph{if} } {{i}} \in \mathcal{B}^{{k}}, {{i}} \not\in \mathcal{B}^{{j}}, \nonumber \\
\text{\emph{3. }} &{\Gamma}_t^{[{{k}},{{i}}]}(L_t^{[{{i}},{{k}}]}) = {\Gamma}_t^{[{{j}},{{i}}]}(L_t^{[{{i}},{{j}}]}), \text{ \emph{if} } {{i}} \not\in \mathcal{B}^{{k}}.
\label{lem_3_condition}
\end{align}
\end{lemma}
\begin{proof}
Let $g_t^{{i}}$ denote the control policy of agent $i \in \mathcal{K}$ at time $t$. To prove the result, we construct $g_t^{{i}}$ and the prescription strategy ${\psi}_t^{{j}}$ for three cases, given a prescription strategy $\boldsymbol{\psi}^{{k}}$.
\begin{enumerate}
\item If ${{i}} \in \mathcal{B}^{{j}}$, the control policy $g_t^{{i}}$ can be constructed from the prescription strategy ${\psi}_t^{[{{k}},{{i}}]}$, namely,
\begin{align}
g_t^{{i}}(A_t^{{i}},{L}_t^{[{{i}},{{i}}]}) &= {\psi}_t^{[{{k}},{{i}}]}(A_t^{{i}})({L}_t^{[{{i}},{{i}}]}).
\end{align}
From \eqref{eq_gen_pres} we have
\begin{gather}
\Gamma_t^{[{{j}},{{i}}]} = \psi_t^{[{{j}},{{i}}]}(A_t^{{i}}), \; \forall {{i}} \in \mathcal{B}^{{j}},
\end{gather}
and thus,
\begin{gather}
\psi_t^{[{{j}},{{i}}]}(A_t^{{i}})(L_t^{[{{i}},{{i}}]}) = g_t^{{i}}(A_t^{{i}},{L}_t^{[{{i}},{{i}}]}).
\end{gather}
Hence,
\begin{gather}
{\psi}_t^{[{{k}},{{i}}]}(A_t^{{i}})({L}_t^{[{{i}},{{i}}]}) = \psi_t^{[{{j}},{{i}}]}(A_t^{{i}})(L_t^{[{{i}},{{i}}]}). \label{eq_e_1}
\end{gather}
\item If ${{i}}\in \mathcal{B}^{{k}}$ and ${{i}} \not \in \mathcal{B}^{{j}}$, the control policy $g_t^{{i}}$ can be constructed by the prescription strategy ${\psi}_t^{[{{k}},{{i}}]},$ namely,
\begin{align}
g_t^{{i}}(A_t^{{i}},{L}_t^{[{{i}},{{i}}]}) = {\psi}_t^{[{{k}},{{i}}]}(A_t^{{i}})({L}_t^{[{{i}},{{i}}]}).
\end{align}
From \eqref{eq_gen_pres} we have
\begin{gather}
\Gamma_t^{[{{j}},{{i}}]} = \psi_t^{[{{j}},{{i}}]}(A_t^{{j}}), \; \forall {{i}} \not\in \mathcal{B}^{{j}}.
\end{gather}
Thus,
\begin{gather}
{\psi}_t^{[{{j}},{{i}}]}(A_t^{{j}})({L}_t^{[{{i}},{{j}}]}) = g_t^{{i}}(A_t^{{j}},{L}_t^{[{{i}},{{j}}]}) = g_t^{{i}}(A_t^{{i}},{L}_t^{[{{i}},{{i}}]}).
\end{gather}
Hence,
\begin{align}
{\psi}_t^{[{{k}},{{i}}]}(A_t^{{i}})({L}_t^{[{{i}},{{i}}]})
={\psi}_t^{[{{j}},{{i}}]}(A_t^{{j}})({L}_t^{[{{i}},{{j}}]}). \label{eq_e_2}
\end{align}
\item If ${{i}}\not\in\mathcal{B}^{{k}}$, the control policy $g_t^{{i}}$ can be constructed by the prescription strategy ${\psi}_t^{[{{k}},{{i}}]},$ namely,
\begin{align}
g_t^{{i}}(A_t^{{k}},{L}_t^{[{{i}},{{k}}]}) =
{\psi}_t^{[{{k}},{{i}}]}(A_t^{{k}})({L}_t^{[{{i}},{{k}}]}).
\end{align}
From \eqref{eq_gen_pres} we have
\begin{gather}
\Gamma_t^{[{{j}},{{i}}]} = \psi_t^{[{{j}},{{i}}]}(A_t^{{j}}), \; \forall {{i}} \not\in \mathcal{B}^{{j}}.
\end{gather}
Thus,
\begin{align}
{\psi}_t^{[{{j}},{{i}}]}(A_t^{{j}})({L}_t^{[{{i}},{{j}}]}) &= g_t^{{i}}(A_t^{{j}},{L}_t^{[{{i}},{{j}}]}) \nonumber \\
&= g_t^{{i}}(A_t^{{k}},{L}_t^{[{{i}},{{k}}]}).
\end{align}
Hence,
\begin{align}
{\psi}_t^{[{{k}},{{i}}]}(A_t^{{k}})({L}_t^{[{{i}},{{k}}]})=
{\psi}_t^{[{{j}},{{i}}]}(A_t^{{j}})({L}_t^{[{{i}},{{j}}]}). \label{eq_e_3}
\end{align}
\end{enumerate}
To complete the proof, note that we can define a positional relationship $e^{[{{j}},{{k}}]} : \Psi^{{k}} \rightarrow \Psi^{{j}}$ with $e^{[{{j}},{{k}}]}=(e_1^{[{{j}},{{k}}]},\ldots,e_T^{[{{j}},{{k}}]})$ such that \eqref{eq_e} implies \eqref{eq_e_1}, \eqref{eq_e_2} and \eqref{eq_e_3} with ${{j}} \in \mathcal{B}^{{k}}$.
\end{proof}
\begin{lemma} \label{lem_psi_relation2}
Let agents $k, j\in \mathcal{K}$ with ${{j}} \not\in \mathcal{B}^{{k}}$. For any given prescription strategy $\boldsymbol{\psi}^{{k}}$ of agent ${{k}}$, there exists a positional relationship $e^{[{{j}},{{k}}]}$ such that a prescription strategy $\boldsymbol{\psi}^{{j}}$ of agent ${{j}}$ generated from \eqref{eq_e} yields:
\begin{align}
\text{\emph{1. }} &{\Gamma}_t^{[{{k}},{{i}}]}(L_t^{[{{i}},{{i}}]}) = {\Gamma}_t^{[{{j}},{{i}}]}(L_t^{[{{i}},{{i}}]}), \text{ \emph{if} } {{i}} \in \mathcal{B}^{{k}}, \nonumber \\
\text{\emph{2. }} &{\Gamma}_t^{[{{k}},{{i}}]}(L_t^{[{{i}},{{k}}]}) = {\Gamma}_t^{[{{j}},{{i}}]}(L_t^{[{{i}},{{i}}]}), \text{ \emph{if} } {{i}} \in \mathcal{B}^{{j}}, {{i}} \not\in \mathcal{B}^{{k}}, \nonumber \\
\text{\emph{3. }} &{\Gamma}_t^{[{{k}},{{i}}]}(L_t^{[{{i}},{{k}}]}) = {\Gamma}_t^{[{{j}},{{i}}]}(L_t^{[{{i}},{{j}}]}), \text{ \emph{if} } {{i}} \not\in \mathcal{B}^{{j}}.
\label{lem_4_condition}
\end{align}
\end{lemma}
\begin{proof}
The proof is very similar to the proof of Lemma \ref{lem_psi_relation1}. It is omitted due to space limitations.
\end{proof}
\begin{comment}
Let $i \in \mathcal{K}$ be an agent located at a position $q \in \mathcal{K}$ and $g_t^{{i}}$ denote the control policy of agent $i$ at time $t$. To prove the result, we construct $g_t^{{i}}$ and the prescription strategy ${\psi}_t^{{j}}$ for three cases, given a prescription strategy $\boldsymbol{\psi}^{{k}}$.
\begin{enumerate}
\item If ${{i}} \in \mathcal{B}^{{k}}$, the control policy $g_t^{{i}}$ can be constructed by the prescription strategy ${\psi}_t^{[{{k}},{{i}}]},$ namely
\begin{align}
g_t^{{i}}(A_t^{{i}},{L}_t^{[{{i}},{{i}}]}) &= {\psi}_t^{[{{k}},{{i}}]}(A_t^{{i}})({L}_t^{[{{i}},{{i}}]}).
\end{align}
$q \in \mathcal{B}^{{k}}$ and $r \not \in \mathcal{B}^{{k}}$ implies $q \in \mathcal{B}^{{j}}$.
From \eqref{eq_gen_pres}, we have
\begin{gather}
\Gamma_t^{[{{j}},{{i}}]} = \psi_t^{[{{j}},{{i}}]}(A_t^{{k}}), \; \forall {{i}} \in \mathcal{B}^{{j}}.
\end{gather}
Thus,
\begin{gather}
\psi_t^{[{{j}},{{i}}]}(A_t^{{i}})(L_t^{[{{i}},{{i}}]}) = g_t^{{i}}(A_t^{{i}},{L}_t^{[{{i}},{{i}}]}).
\end{gather}
Hence,
\begin{gather}
{\psi}_t^{[{{k}},{{i}}]}(A_t^{{i}})({L}_t^{[{{i}},{{i}}]})=\psi_t^{[{{j}},{{i}}]}(A_t^{{i}})(L_t^{[{{i}},{{i}}]}). \label{eq_e_21}
\end{gather}
\item If ${{i}}\in \mathcal{B}^{{j}}$, $q \not \in \mathcal{B}^{{k}}$, the control policy $g_t^{{i}}$ can be constructed by the prescription strategy ${\psi}_t^{[{{k}},{{i}}]},$ namely
\begin{align}
g_t^{{i}}(A_t^{{k}},{L}_t^{[{{i}},{{k}}]}) = {\psi}_t^{[{{k}},{{i}}]}(A_t^{{k}})({L}_t^{[{{i}},{{k}}]}).
\end{align}
From \eqref{eq_gen_pres}, we have
\begin{gather}
\Gamma_t^{[{{j}},{{i}}]} = \psi_t^{[{{j}},{{i}}]}(A_t^{{i}}), \; \forall {{i}} \in \mathcal{B}^{{j}}.
\end{gather}
Thus,
\begin{align}
{\psi}_t^{[{{j}},{{i}}]}(A_t^{{i}})({L}_t^{[{{i}},{{i}}]}) &= g_t^{{i}}(A_t^{{i}},{L}_t^{[{{i}},{{i}}]}) \nonumber \\
&= g_t^{{i}}(A_t^{{k}},{L}_t^{[{{i}},{{k}}]}).
\end{align}
Hence,
\begin{align}
{\psi}_t^{[{{j}},{{i}}]}(A_t^{{i}})({L}_t^{[{{i}},{{i}}]})
= {\psi}_t^{[{{k}},{{i}}]}(A_t^{{k}})({L}_t^{[{{i}},{{k}}]}). \label{eq_e_22}
\end{align}
\item If ${{i}}\not\in\mathcal{B}^{{k}}$, the control policy $g_t^{{i}}$ can be constructed by the prescription strategy ${\psi}_t^{[{{k}},{{i}}]},$ namely
\begin{align}
g_t^{{i}}(A_t^{{k}},{L}_t^{[{{i}},{{k}}]}) =
{\psi}_t^{[{{k}},{{i}}]}(A_t^{{k}})({L}_t^{[{{i}},{{k}}]}).
\end{align}
From \eqref{eq_gen_pres} we have
\begin{gather}
\Gamma_t^{[{{j}},{{i}}]} = \psi_t^{[{{j}},{{i}}]}(A_t^{{j}}), \; \forall {{i}} \not\in \mathcal{B}^{{j}}.
\end{gather}
Thus,
\begin{align}
{\psi}_t^{[{{j}},{{i}}]}(A_t^{{j}})({L}_t^{[{{i}},{{j}}]}) &= g_t^{{i}}(A_t^{{j}},{L}_t^{[{{i}},{{j}}]}) \nonumber \\
&= g_t^{{i}}(A_t^{{k}},{L}_t^{[{{i}},{{k}}]}).
\end{align}
Hence,
\begin{align}
{\psi}_t^{[{{j}},{{i}}]}(A_t^{{j}})({L}_t^{[{{i}},{{j}}]}) = {\psi}_t^{[{{k}},{{i}}]}(A_t^{{k}})({L}_t^{[{{i}},{{k}}]}). \label{eq_e_23}
\end{align}
\end{enumerate}
To complete the proof, note that we can define a positional relationship $e^{[{{j}},{{k}}]} : \Psi^{{k}} \rightarrow \Psi^{{j}}$ with $e^{[{{j}},{{k}}]}=(e_1^{[{{j}},{{k}}]},\ldots,e_T^{[{{j}},{{k}}]})$ such that \eqref{eq_e} implies \eqref{eq_e_21}, \eqref{eq_e_22} and \eqref{eq_e_23} when ${{j}} \not\in \mathcal{B}^{{k}}$.
\end{comment}
To this end, we consider a positional relationship function $e^{[{{j}},{{k}}]}$ from every ${{k}} \in \mathcal{K}$ to every position ${{j}} \in \mathcal{K}$ which satisfies the properties in Lemmas \ref{lem_psi_relation1} and \ref{lem_psi_relation2}. This implies that for any two agents ${{k}}$ and ${{j}}$, we have the relation,
\begin{gather}
U_t^{{j}} = {\Gamma}_t^{[{{j}},{{j}}]}(L_t^{[{{j}},{{j}}]}) =
\begin{cases}
{\Gamma}_t^{[{{k}},{{j}}]}(L_t^{[{{j}},{{k}}]}), \text{ if } {{j}} \not\in \mathcal{B}^{{k}}, \\
{\Gamma}_t^{[{{k}},{{j}}]}(L_t^{[{{j}},{{j}}]}), \text{ if } {{j}} \in \mathcal{B}^{{k}}.
\end{cases}
\label{eq_u_with_another_psi}
\end{gather}
\section{RESULTS}
\subsection{Equivalent Prescription Problems}
Lemmas \ref{lem_psi_g_relation} through \ref{lem_psi_relation2} lead to \eqref{eq_u_with_another_psi}. This implies that the control action $U_t^j$ for agent $j \in \mathcal{K}$ can be equivalently obtained through the prescription function $ {\Gamma}_t^{[{{k}},{{j}}]}$ of any other agent ${{k}} \in \mathcal{K}$, if the corresponding inaccessible information is available.
Using \eqref{eq_u_with_another_psi}, we can write the cost incurred by the system at time $t$ as
\begin{align}
c_t(X_t,U_t^1,\ldots,&U_t^K) \nonumber \\
=: c_t\big(&X_t,{\Gamma}_t^{[{{k}},1]}({L}_t^{[1,{{k}}]}),\ldots,{\Gamma}_t^{[{{k}},{{k}}]}({L}_t^{[{{k}},{{k}}]}), \nonumber \\
&{\Gamma}_t^{[{{k}},{{k+1}}]}({L}_t^{[{{k+1}},{{k+1}}]}),\ldots,{\Gamma}_t^{[{{k}},K]}({L}_t^{[{{K}},K]})\big). \label{o_cost}
\end{align}
We can then reformulate Problem 1 in terms of the prescription strategy of any agent ${{k}}$. The optimization problem is to select the optimal prescription strategy $\boldsymbol{\psi}^{{*k}} \in \Psi^{{k}}$ that minimizes the performance criterion given by the total expected cost:
\begin{multline}
\textbf{Problem 2:}~~~\mathcal{J}^{{k}}(\boldsymbol{\psi}^{{k}}) = \\
\mathbb{E}^{\boldsymbol{\psi}^{{k}}} \Big[\sum_{t=0}^T{c_t\big(X_t,{\Gamma}_t^{[{{k}},1]}({L}_t^{[1,{{k}}]}),\ldots,{\Gamma}_t^{[{{k}},{{k}}]}({L}_t^{[{{k}},{{k}}]})}, \\
{\Gamma}_t^{[{{k}},{{k+1}}]}({L}_t^{[{{k+1}},{{k+1}}]}),\ldots,{\Gamma}_t^{[{{k}},K]}({L}_t^{[{{K}},K]})\big)\Big]. \label{per_cri_2}
\end{multline}
The task of deriving optimal prescription strategy $\boldsymbol{\psi}^{{*k}},$ and subsequently, the complete prescription $\Theta_t^{{k}}$ for agent ${{k}}$ is assigned to a fictitious designer that can only access memory $M_t^{{k}}$. Note that this maintains the decentralized nature of the problem as the strategies are implemented by the agents in real time with asymmetric and incomplete information. Now, we show the equivalence between the two problems.
\begin{lemma} \label{lem_equivalence}
For any agent ${{k}} \in \mathcal{K}$, Problem 2 is equivalent to Problem 1.
\end{lemma}
\begin{proof}
Eq. \eqref{o_cost} implies that the performance criterion $\mathcal{J}^{{k}}(\boldsymbol{\psi}^{{k}})$ in \eqref{per_cri_2} is equal to the performance criterion $\mathcal{J}(\boldsymbol{g})$ in \eqref{per_cri}. Thus, given the optimal prescription strategy $\boldsymbol{\psi}^{{*k}}$, from Lemmas \ref{lem_psi_g_relation} and \ref{lem_psi_g_relation_inv}, we can derive the optimal policy $\boldsymbol{g}^*$ in Problem 1.
\end{proof}
Next, we present a state sufficient for input-output mapping in Problem 2 for agent ${{k}}$ following the exposition presented in \cite{mahajan2008sequential}.
\begin{lemma} \label{state_suff_k}
A state sufficient for input-output mapping for agent ${{k}} \in \mathcal{K}$ is
\begin{equation}
S^{{k}}_t := \left\{X_{t}, L_t^{[1,{{k}}]},\ldots,L_t^{[{{k-1}},{{k}}]},L_t^{[{{k}},{{k}}]},\ldots,L_t^{[{{K}},K]}\right\}. \label{S_k}
\end{equation}
\end{lemma}
\begin{proof}
The state $S^{{k}}_t$ satisfies the three properties stated by Witsenhausen \cite{witsenhausen1976some}:
1) There exist functions $\{\hat{f}^{{k}}_t$: $t \in \mathcal{T}\}$ such that
\begin{equation}
S^{{k}}_{t+1} = \hat{f}^{{k}}_{t}(S^{{k}}_t, W_t, V_{t+1}^{1:K}, \Theta_t^{{k}}). \label{eq_St_k1}
\end{equation}
2) There exist functions $\{\hat{h}^{{k}}_t$: $t \in \mathcal{T}\}$ such that
\begin{equation}
Z^{{k}}_{t+1} = \hat{h}^{{k}}_{t}(S^{{k}}_{t},\Theta_t^{{k}},V_{t+1}^{1:K}). \label{eq_St_k2}
\end{equation}
3) There exist functions $\{\hat{c}^{{k}}_t$: $t \in \mathcal{T}\}$ such that
\begin{align}
c_t(X_t,U_t^{1:K}) &= \hat{c}^{{k}}_t(S^{{k}}_t,\Theta^{{k}}_t). \label{eq_St_k3}
\end{align}
The three equations above can each be verified by substitution of variables on the LHS. The complete proof can be found in \cite{2019Aditya_arXiv}.
\end{proof}
\subsection{The Information States}
From the designer's point of view, the system behaves as a Partially Observed Markov Decision Process (POMDP) with state $S_t^{{k}}$, control input $\Theta_t^{{k}}$, output $Z_t^{{k}}$ (with $Z^{{k}}_{0:t} = A^{{k}}_t$) and cost $\hat{c}^{{k}}_t(S^{{k}}_t,\Theta^{{k}}_t)$ at time $t$.
The difference is that the prescription functions $\Gamma_t^{[{{k}},{{j}}]}$, ${{j}} \in \mathcal{B}^{{k}},$ are generated as functions of the accessible information $A_t^{{j}}$ instead of $A_t^{{k}}$. Thus, structural results for POMDPs cannot be directly applied to Problem 2. Before proceeding to structural results, we define the \textit{information state} for agent ${{k}}$.
\begin{definition}
Let $S_t^{{k}}$ be the state, $A_{t}^{{k}}$ the accessible information, and $\Theta^{{k}}_{0:t-1}$ the control inputs at time $t$ an agent ${{k}} \in \mathcal{K}$. The \textit{information state} is defined as a probability distribution $\Pi^{{k}}_t$ that takes values in the possible realizations $\mathscr{P}^{{k}}_t := \Delta(\mathcal{S}^{{k}}_t)$ such that,
\begin{equation}
\Pi^{{k}}_t(s^{{k}}_t) := \mathbb{P}^{\boldsymbol{\psi}^{{k}}}(S^{{k}}_t = s^{{k}}_t \big|A^{{k}}_t, \Theta^{{k}}_{0:t-1}).
\end{equation}
\end{definition}
Due to space limitation, the proofs of the following three properties are omitted but can be found in \cite{2019Aditya_arXiv}.
The first property establishes that the information state $\Pi_t^{{k}}$ is independent from the prescription strategy $\boldsymbol{\psi}^{{k}}$.
\begin{lemma} \label{pi_k_1}
At time $t$, there exists a function $F_t^{{k}}$ independent from the prescription strategy $\boldsymbol{\psi}^{{k}}$ such that
\begin{equation}
\Pi_{t+1}^{{k}} = F_{t+1}^{{k}}(\Pi_t^{{k}},\Theta_t^{{k}},Z_{t+1}^{{k}}).
\end{equation}
\end{lemma}
The second property of the information state $\Pi_t^{{k}}$ is that its evolution is Markovian.
\begin{lemma} \label{pi_k_2}
The evolution of the information state $\Pi_t$ is a controlled Markov Chain with $\Theta^{{k}}_{t}$ as the control action at time $t$
\begin{align}
\mathbb{P}^{\boldsymbol{\psi}^{{k}}}(\Pi_{t+1}^{{k}}|A_t^{{k}},\Pi^{{k}}_{0:t}, \Theta^{{k}}_{0:t}) = \mathbb{P}^{\boldsymbol{\psi}^{{k}}}(\Pi_{t+1}^{{k}}|\Pi^{{k}}_{t}, \Theta^{{k}}_{t}).
\end{align}
\end{lemma}
The third property of the information state $\Pi_t^{{k}}$ is that the expected cost incurred by the system at time $t$ can be written as a function of $\Pi_t^{{k}}$.
\begin{lemma} \label{pi_k_3}
There exists a function $C^{{k}}_t$, independent of the prescription strategy $\boldsymbol{\psi}^{{k}}$, such that
\begin{equation}
\mathbb{E}^{\boldsymbol{\psi}^{{k}}}\big[\hat{c}^{{k}}_t(S^{{k}}_t,\Theta^{{k}}_t)|A^{{k}}_t,\Theta^{{k}}_{0:t}\big] = C_t^{{k}}(\Pi_t^{{k}},\Theta_t^{{k}}).
\end{equation}
\end{lemma}
In Lemmas \ref{pi_k_1} through \ref{pi_k_3}, we established that the information state $\Pi_t^{{k}}$ evolves as a controlled Markov chain with control inputs $\Theta_t^{{k}}$.
\subsection{Structural Results}
We start by presenting a structural result for agent $K$. By definition, the set of agents beyond agent $K$ contains only agent $K$, i.e., $\mathcal{B}^{K} = \{K\}$. Using \eqref{eq_gen_pres}, this implies that for all agents ${{k}} \in \mathcal{K}$, the prescription component $\Gamma^{[{{K}},{{k}}]}_t$ is a function of the accessible information $A_t^{{[K]}}$. This leads to the following result derived in \cite{14} through the common information approach.
\begin{lemma}\label{lem_case_K}
Consider agent $K$. There exists an optimal prescription strategy $\boldsymbol{\psi}^{*K}$ of the form
\begin{equation}
\Gamma_t^{*[{{K}},{{k}}]} = \psi_t^{*[{{K}},{{k}}]}(\Pi_t^{K}), \label{eq_common_info}
\end{equation}
that optimizes the performance criterion \eqref{per_cri_2} in Problem 2.
\end{lemma}
We know that for any two agents ${{k}} \in \mathcal{K}$ and ${{j}} \in \mathcal{B}^{{k}}$, we have $A_t^{{j}} \subset A_t^{{k}}$. Given the accessible information $A_t^{{k}}$ and the optimal prescription strategy $\boldsymbol{\psi}^{*{{k}}}$, agent ${{k}}$ can derive the optimal complete prescriptions $\Theta_t^{*{{j}}}$ for every ${{j}} \in \mathcal{B}^{{k}}$. This leads to the following structural result, proved in \cite{2019Aditya_arXiv}.
\begin{theorem} \label{struct_result}
Consider agent ${{k}} \in \mathcal{K}$. There exists an optimal prescription strategy $\boldsymbol{\psi}^{*{{k}}}$ of the form
\begin{align}
\Gamma_t^{*[{{k}},{{j}}]}(\cdot) =
\begin{cases}
{\psi}_t^{*[{{k}},{{j}}]}(\Pi^{{k}}_t,\ldots,\Pi_t^{K}), \text{\emph{ if }} {{j}} \not\in\mathcal{B}^{{k}}, \\
{\psi}_t^{*[{{k}},{{j}}]}(\Pi^{{j}}_t,\ldots,\Pi_t^{K}), \text{\emph{ if }} {{j}} \in \mathcal{B}^{{k}},
\end{cases} \label{eq_struct_result}
\end{align}
that optimizes the performance criterion \eqref{per_cri_2} in Problem 2.
\end{theorem}
\subsection{A Comparison with Existing Approaches}
Among the existing approaches, the person-by-person approach and the designer's approach do not yield the kind of structural results presented in this paper. The graphical approach presented in \cite{mahajan2015algorithmic} has similarities with the prescription approach, but, it applies only to problems where agents have perfect observations. The common information approach in \cite{14} can be applied to the problem with the word-of-mouth communication structure to obtain the structural result presented in Lemma \ref{lem_case_K}, since by definition, the accessible information $A_t^{{K}}$ is the common information in the system. Thus, the control action for agent $k \in \mathcal{K}$ is given by
\begin{gather}
U_t^{{*k}} = \Gamma_t^{*[{{K}},{{k}}]}(L_t^{[{{k}},{{K}}]}).
\end{gather}
In contrast, we see that when we consider Problem 2 for agent ${{k}}$, the control action of agent ${{k}}$ is given by
\begin{gather}
U_t^{{*k}} = \Gamma_t^{*[{{k}},{{k}}]}(L_t^{[{{k}},{{k}}]}).
\end{gather}
Now, from \eqref{eq_partition}, we note that,
\begin{gather}
A_t^{{k}} \cup L_t^{[{{k}},{{k}}]} = A_t^{{K}} \cup L_t^{[{{k}},{{K}}]}, \label{eq_comparison_1}
\end{gather}
and from \eqref{ainfo_prop_1} we have the relation,
\begin{gather}
A_t^{{K}} \subset A_t^{{k}}, \label{eq_comparison_2}
\end{gather}
because $K \in \mathcal{B}^{{k}}$ for all ${{k}} \in \mathcal{K}$. Then, \eqref{eq_comparison_1} and \eqref{eq_comparison_2} imply,
\begin{gather}
L_t^{[{{k}},{{k}}]} \subset L_t^{[{{k}},{{K}}]}.
\end{gather}
Thus, the prescription functions generated through the prescription approach have an equal or smaller domain when compared with those generated through the common information approach.
\section{CONCLUSIONS}
In this paper, we introduce a network of agents with a word-of-mouth communication structure, and analyze it using the prescription approach, which yielded some desired properties. We showed that the structural result derived through the common information approach can be considered as the outcome of one reformulations using the prescription approach. Finally, we provided, without proof, a preliminary structural result arising from the prescription approach. A direction for future research should seek to extend these results for a broader class of decentralized systems.
\bibliographystyle{ieeetr}
|
\section{Appendix}
\label{lab:app}
Derivation of the Sparse Bayesian Learning algorithm using iterative reweigthed L1 minimalization.
For a fixed $\gamma$, we have an approximate prior as
\begin{eqnarray}
\hat{p}_{\gamma}(\theta) = \prod_{i=1}^{m}\mathcal{N}(\theta_i;0,\gamma_i)\zeta(\gamma_i).
\end{eqnarray}
Using the Gaussian prior and likelihood, we get an approximate posterior
\begin{eqnarray}
\hat{p}_{\gamma}(\theta|y) = \frac{p(y|\theta)\hat{p}_{\gamma}(\theta)}{\int p(y|\theta)\hat{p}_{\gamma}(\theta) d\theta} = \mathcal{N}(\theta;\mu_{\theta},\Sigma_{\theta})
\end{eqnarray}
where
\begin{eqnarray}
\label{eqn:cov_mtx}
\mu_{\theta} &=& \Gamma\Phi^\top (\lambda I + \Phi\Gamma\Phi^\top)^{-1}y \\
\Sigma_{\theta} &=& \Gamma - \Gamma\Phi^\top (\lambda I + \Phi\Gamma\Phi^\top)^{-1}\Phi\Gamma
\end{eqnarray}
where $\Gamma = \text{diag}[\gamma]$. The next step is estimating $\gamma$ in a way the is amenable to the above computation. According to \cite{Wipf2011}, an estimator for $\gamma$ can be constructed using variational representation which involves solving
\begin{eqnarray}
\gamma &=& \argmin_{\gamma} \int p(y|\theta)|p(\theta) - \hat{p}_{\gamma}(\theta)| d\theta \\
&=& \argmax_\gamma \int p(y|\theta) \prod_{i=1}^{m}\mathcal{N}(\theta_i;0,\gamma_i)\zeta(\gamma_i) d\theta_i.
\end{eqnarray}
The above expression has analytical solution, after applying $-2\log(\cdot)$, the cost function is
\begin{eqnarray}
\label{eq:gamma_cost_fcn}
C(\gamma) = y^\top \Sigma_y^{-1} y + \log|\Sigma_y| + \sum_{i=1}^{m}f(\gamma_i)
\end{eqnarray}
where $f(\gamma_i)=-2\log(\zeta(\gamma_i))$ and $\Sigma_{y}=\lambda I + \Phi\Gamma\Phi^{\top}$.
The above cost function in not convex, but an iterative optimization can be established mainly using results from convex analysis. First, the data dependent term in \eqref{eq:gamma_cost_fcn} can be reexpressed as
\begin{eqnarray}
y^\top \Sigma_y^{-1} y = \min_\theta \frac{1}{\lambda}||y-\Phi\theta||_2^2 + \sum_{i=1}^{n} \frac{\theta_i^2}{\gamma_i}.
\end{eqnarray}
Then, we can create a strict upper bounding function on $C(\gamma)$ by
\begin{eqnarray*}
C(\gamma,\theta) = \frac{1}{\lambda}||y-\Phi\theta||_2^2 + \sum_{i=1}^{n} \frac{\theta_i^2}{\gamma_i} + \log |\Sigma_y| + \sum_{i}f(\gamma_i).
\end{eqnarray*}
If we minimize over $\gamma$, then
\begin{eqnarray}
C(\theta) = \min_{\gamma \geq 0}C(\gamma,\theta) = ||y-\Phi\theta||_2^2 + \lambda g(\theta)
\end{eqnarray}
where
\begin{eqnarray}
\label{eq:g_fcn}
g(\theta) = \min_{\gamma \geq 0} \sum_{i=1}^{n} \frac{\theta_i^2}{\gamma_i} + \log |\Sigma_y| + \sum_{i}f(\gamma_i).
\end{eqnarray}
In the rest of the paper, we assume that $f(\gamma_i)=0$, for $i=1,\ldots,m$ because we are interested in maximal sparsity. For further details on $f(\gamma_i)$ and comparison of different regularization terms, see \cite{wipf2007performance}.
In the rest of the analysis, we focus on \eqref{eq:g_fcn}. The $\log |\cdot|$ is a concave function on semidefinite matrices \cite{Boyd2004} and $\Sigma_y$ is a positive semidefinite matrix and an affine function in $\gamma$. Thus $\log|\Sigma_y|$ is a concave, non-decreasing function in $\gamma$ and can be expressed as
\begin{eqnarray}
\log|\Sigma_y| = \min_{z \geq 0} z^\top \gamma - h^*(z)
\end{eqnarray}
where $h^*(z)$ is the concave conjugate of $\log|\Sigma_y|$ and given by
\begin{eqnarray}
h^*(z) = \min_{\gamma \geq 0 }z^\top\gamma - \log|\Sigma_y|.
\end{eqnarray}
Based on that, we rewrite \eqref{eq:g_fcn} as
\begin{eqnarray}
\label{eq:g_fcn_gamma_z}
g(\theta) = \min_{\gamma, z \geq 0} \theta^\top \Gamma^{-1} \theta + z^\top \gamma - h^*(z).
\end{eqnarray}
Fixing $\theta$ and $z$, then optimizing over $\gamma$, yields
\begin{eqnarray}
\gamma^{opt}_i = z_i^{-1/2}|\theta_i|,~~i=1,\ldots,m.
\end{eqnarray}
Substituting $\gamma^{opt}$ back to \eqref{eq:g_fcn_gamma_z}, we get
\begin{eqnarray}
g(\theta) = \min_{z \geq 0}\sum_{i=1}^{m}2z_i^{-1/2}|\theta_i| - h^*(z).
\end{eqnarray}
Now, we only need to calculate the optimal $z$. For a fixed $\gamma$, we get
\begin{eqnarray}
z^{opt} &=& \nabla_\gamma \log |\Sigma_y| \\
&=& \text{diag}[\Phi^\top\Sigma_y^{-1}\Phi] = \text{diag}[\Phi^\top(\lambda I + \Phi\Gamma\Phi^\top)\Phi]\nonumber.
\end{eqnarray}
Then, we can substitute $\Gamma=\text{diag}[\gamma^{opt}]$ to get a value for $z^{opt}$.
\section{BACKGROUND}
The concept of CRNs was introduced by Feinberg during his seminal lectures, and since then it has become a widely used modeling framework \cite{Feinberg:79}; first in process engineering, then later in Systems and Synthetic Biology.
The dynamics of CRNs can be described by the class of kinetic systems which offer certain algebraic proprieties that lead to several powerful results (see, e.g., \cite{Feinberg1987,Craciun2015,Shinar2010}). There has been a constant effort to characterize the systems theoretical properties of kinetic systems (see, \cite{Craciun2009,perez2012chemical,van2013mathematical} for more details). This paper follows an optimization-based approach which translates the dynamical and structural properties of kinetic systems into the constraint set, see \cite{Johnston2011conj,acs2016} for an overview.
This section defines the system class represented by polynomial Ordinary Differential Equations (ODEs) and defines Chemical Reaction Networks as well; then establishes a connection between the two. Building on these definitions, we can introduce the optimization problems to compute certain graph structures.
Let us define the following polynomial differential equation with state vector $x\in \mathbb{R}^n_{\geq 0}$, and
\begin{eqnarray}
\label{eqn:kinetic_sys}
\dot{x} = M\varphi(x), \quad x(0) \in \mathbb{R}^n_{\geq 0},
\end{eqnarray}
where the matrix $M\in\mathbb{R}^{n\times m}$ is the coefficient matrix and $\varphi(x): \mathbb{R}^n_{\geq 0} \rightarrow \mathbb{R}^m_{\geq 0}$ is a monomial-type vector mapping defined as $\varphi_j(x)= \prod_{i=1}^{n} x_i^{[B]_{ij}}, j=1,\ldots,m$ and $B\in \mathbb{N}_0^{n\times m}$. Note that the system in \eqref{eqn:kinetic_sys}, with certain sign constraints in matrix $M$, belongs to the class of nonnegative systems, i.e. $\mathbb{R}_{\geq 0}^n$ is forward invariant (see, e.g. Chapter 9 in \cite{Haddad2010}).
Next, we define chemical reaction networks, which can be characterized by three sets:
\begin{itemize}
\item a set of species: $\mathcal{S}= \{X_i\ |\ i=1,\ldots,n\}$,
\item a set of complexes: $\mathcal{C} = \{ C_j \ |\ j=1,\ldots,m\}$, where\\
$C_j = \sum \limits_{i=1}^{n} \alpha_{ji} X_i \qquad j=1,\ldots,m$ \text{and} \\
$\alpha_{ji} \in \mathbb{N}_0 \qquad \qquad ~~j=1,\ldots,m, \ i=1,\ldots,n$,\\
$\alpha_{ji}$ are called the stoichiometric coefficients,
\item and a set of reactions: $\mathcal{R} \subseteq \{(C_i,C_j)\ |\ C_i,C_j \in \mathcal{C}\}$, each ordered pair $(C_i,C_j)$ has a reaction rate coefficient $k_{ij} \in \mathbb{R}_{\geq 0}$ so that the corresponding reaction $C_i \rightarrow C_j$ takes place if and only if $k_{ij} >0$.
\end{itemize}
In the rest of the paper, we assume mass action kinetics, but the results summarized in this section have been extended to rational kinetics as well, see e.g. \cite{GaborA2016} for more details.
For computation purposes, we can characterize a CRN by two matrices: the complex composition matrix $Y\in \mathbb{N}_0^{n\times m}$ describes the complexes as follows
\[[Y]_{ij}=\alpha_{ji} \qquad i=1,\ldots,n, \ j=1,\ldots,m,\]
and the set of reactions is encoded by the Kirchhoff matrix $A_\kappa \in \mathbb{R}^{m\times m}$ as
\begin{eqnarray}
\label{eq:kirchhoff_mtx}
[A_\kappa]_{ij}= \begin{cases}
\hfil k_{ji} &\text{ if } i \neq j\\
-\sum \limits_{l=1, l\neq i}^{m} k_{il} & \text{ if } i=j.\\
\end{cases}
\end{eqnarray}
The dynamics of a CRN can be written as a nonnegative polynomial differential equation
\begin{eqnarray}
\label{eqn:crn_dynamics}
\dot{x} = Y A_\kappa \psi^Y(x), \quad x(0) \in \mathbb{R}^n_{\geq 0},
\end{eqnarray}
where $x$ represents the concentration vector of the species and the monomial vector mapping $\psi^Y(x)$ is defined as
\begin{eqnarray}
\psi_j^Y(x) = \prod_{i=1}^{n} x_i^{[Y]_{ij}} \quad j=1,\ldots,m.
\end{eqnarray}
At this point, we can make a connection between a nonnegative polynomial ODE and the dynamics of CRNs as follows:
a nonnegative polynomial ODE in \eqref{eqn:kinetic_sys} can be transformed into the form of \eqref{eqn:crn_dynamics}, i.e.
\begin{eqnarray}
\label{eqn:dyn_eq_full}
M\varphi(x) = Y A_\kappa \psi^Y(x),
\end{eqnarray}
if and only if the following condition is fulfilled
\begin{eqnarray}
\label{eqn:kinetic_logical}
\text{if } [M]_{ij}\leq 0,\text{then } [B]_{ij}>0, \\i=1,\ldots,n, j=1,\ldots,m. \nonumber
\end{eqnarray}
If the above condition is satisfied, we call \eqref{eqn:kinetic_sys} a kinetic system because it has at least one CRN realization. This condition also ensures that there are no negative cross-effects in the kinetic system, as it is explained in \cite{Erdi1989}. Furthermore, using this condition a so-called canonical realization can be computed from \eqref{eqn:kinetic_sys} (see \cite{Hars1981} for the details). However, it should be stressed that CRNs with different sets of complexes and reactions can generate the same dynamics \cite{Erdi1989,Craciun2008}.
In this paper, we assume that the set of used complexes is known, which defines the matrix $Y$ and consequently the monomial vector mapping $\psi^{Y}(x)$. Note that the left hand side of \eqref{eqn:dyn_eq_full} is a multivariate polynomial function, while the factorization on the right hand side defines a CRN structure. Therefore, the sets of monomials in $\varphi$ and $\psi^Y$ are not necessarily identical. The reason for this is that the monomials of pure product complexes do not appear in the kinetic equations, but $\psi^Y$ contains the monomials of each complex, even if some of them have zero coefficients in the equations. Naturally, $\psi^Y$ must contain all the monomials of $\varphi$. This means that without the loss of generality we can assume that $\varphi=\psi^Y$ and write the matrix $M$ accordingly. Using this assumption, dynamical equivalence can be simply represented as
\begin{eqnarray}
\label{eqn:dyn_eq}
M = Y A_\kappa.
\end{eqnarray}
The equation above shows that even with a fixed set of complexes, several different $A_\kappa$ matrices can lead to the same dynamics. Hence, these different $A_\kappa$ matrices are called dynamically equivalent realizations of a kinetic system. The total number of such realizations and their (structural and dynamical) properties are the main focus of the past and the current research presented in this paper.
It should be noted that a kinetic system---with a fixed set of complexes---is uniquely characterized by the matrix pair ($Y$, $M$), thus we can refer to it by this pair.
\subsection{Optimization-based computation of realizations}
\label{sec:opt_based_crn}
Optimization problems can be formulated to find realizations of the same kinetic system with different dynamical or structural properties \cite{acs2016}. Jonhston \textit{et al.} also proved that there exists a unique superstructure for each kinetic system, and all possible realizations are contained by the superstructure as subgraphs \cite{Johnston2012a}. This superstructure is also called the dense realization since it contains the most number of reactions for a given kinetic system.
Based on the superstructure property, an algorithm was developed to compute all possible realizations of a given kinetic system \cite{acs2016}. This algorithm effectively excludes different edge patterns from the dense realization, and by construction, it returns all possible realizations (see details of the proof in \cite{acs2016}). Additionally, this algorithm can be massively parallelized (see \cite{TuzaFosbe2016} for more details).
A kinetic system has not only a structurally unique dense realization but a structurally sparse realization as well. However, this sparse realization is not always structurally unique, meaning that multiple sparse realizations may exist and have the same minimum number of reactions (i.e. edges in the reactions graph).
Several graph properties can be translated as constraints in an optimization problem, a non-exhaustive list includes weak reversibility \cite{Szederkenyi2011b}, complex balance \cite{Szederkenyi2011a}, deficiency zero \cite{Liptak2015}, or deficiency one \cite{Johnston2016}. The resulting optimization problems not only give solutions with the given graph properties, but these may also give a certificate about the lack of such realizations. For example, if one is looking for a weakly reversible realization of a given kinetic system, and there is provably no feasible solution to the optimization problem complemented with the constraints of weak reversibility, then there exists no weakly reversible realization for a given dynamics. Thus, such optimization problems can be used to characterize some of the structural and dynamical properties of kinetic systems.
\subsubsection{Computation of the dense realization}
The computation of the dense reaction graph can be formulated as an optimization problem. A possible approach for that would be a mixed integer linear programming problem where the number of reactions in the network has to be maximized, see \cite{Szederkenyi2011} for the details. To make the computation tractable for large networks, an iterative algorithm to compute the dense realization was reported in \cite{Acs2015}. The main steps of this algorithm are summarized below.
First, by combining \eqref{eq:kirchhoff_mtx} and \eqref{eqn:dyn_eq}, a linear programming (LP) problem can be formulated with the following set of constrains
\begin{align}
\label{eq:kinetic_constraints}
\ & M = Y A_\kappa\nonumber \\
\ &[A_\kappa]_{ij} \geq 0 \quad i=1,\ldots,m,~~j=1,\ldots,m \ ~~i\neq j\\
\ & [A_\kappa]_{ii} =- \sum \limits_{\substack{j =1 \\ j \neq i}}^{m} [A_\kappa]_{ji} \quad i=1,\ldots, m,\nonumber
\end{align}
where dynamics of a kinetic system is given by ($Y$, $M$) and $A_\kappa\in\mathbb{R}^{m\times m}$ is now the decision variable of the optimization problem.
In many cases, the edge exclusion from the reaction graph is needed. Formally, a set $\mathcal{H}\subset \mathcal{R}$ of reactions has to be excluded from the network that can be written as a linear constraint:
\begin{equation}
\label{eq:edge_exclusion}
[A_\kappa]_{ji} = 0 \qquad (C_i,C_j) \in \mathcal{H}.
\end{equation}
Second, we formulate the linear cost function as
\begin{eqnarray}
\label{eq:cost_fcn}
\text{maximize}~\sum \limits_{i=1}^m \sum \limits_{j=1}^m [E]_{ij} [A_\kappa]_{ij},
\end{eqnarray}
where the binary matrix $E\in\{0,1\}^{m\times m}$ selects the elements of $A_\kappa$ into the cost function. Further details of the optimization problem and the algorithm itself is given in \cite{Acs2015}.
\subsection{Uncertain Kinetic Systems}
So far, we encoded dynamics of a kinetic system by ($Y$, $M$) and we assumed that the coefficients matrix $M$ is constant. We then computed certain graph structures. In this section, we define a family of kinetic systems where elements of the coefficient matrix belong to a set, denoted by $\mathcal{M}$ which contains all admissible parameter vectors of the kinetic system. Clearly, the properties of the kinetic system depend on the set $\mathcal{M}$. Therefore, we first characterize the type of uncertainty considered in this paper. Then, building on the results summarized in the previous section, we define a convex optimization problem to compute the dense realization and subsequently all realizations of the uncertain kinetic system.
\subsection{Optimization-based computation of uncertain realizations}
\label{sec:uncertain_opt}
We assume two properties of the parametric uncertainty. First, the nominal matrix $\bar{M}\in\mathcal{M}$ is given, thus we have one member of the family of kinetic systems. Second, the nominal $\bar{M}$ is perturbed by an unknown matrix $\Delta$, and we only know the upper bound of the uncertainty in some norm, e.g. in Frobenius norm: $||\Delta||_F \leq \rho$, where $\rho\in\mathbb{R}_{\geq 0}$. With these assumptions, we can characterize the possible parameters vectors as $\text{vec}(M) = \text{vec}(\bar{M}) + \rho u$ where $\text{vec}(M)\in\mathbb{R}^{nm}$ is the vectorization of the parameter matrix $M\in\mathbb{R}^{n\times m}$ and $u\in\mathbb{R}^{nm}$ is given as $u=\text{vec}(\Delta)$.
The uncertainty set around the nominal $\bar{M}$ is given by
\begin{eqnarray}
\label{eqn:uncertainty_set_ball}
\mathcal{M} =\{\text{vec}(\bar{M})+\rho u, ||u||_{2} \leq 1\}.
\end{eqnarray}
This type of uncertainty describes a sphere around $\text{vec}(\bar{M})$, which can be translated to an second order conic (SOC) constraint as follows
\begin{eqnarray}
\label{eq:socp_spherical}
||\text{vec}(M)-\text{vec}(\bar{M})||_2^2 \leq \rho
\end{eqnarray}
where $\text{vec}(M)$ is an optimization variable. We can add \eqref{eq:socp_spherical} as an SOC constraint to the optimization problem defined in \eqref{eq:kinetic_constraints}, \eqref{eq:edge_exclusion}, \eqref{eq:cost_fcn} and compute the dense realization with spherical uncertainty using the Algorithm 1 from \cite{Acs2015}.
In this paper, we are more interested in the case where the uncertainty not uniform in all directions. In this case, the ellipsoidal uncertainty is defined by $\Sigma\in\mathbb{R}^{nm\times nm}$ and $\Sigma \succeq 0$ and the corresponding uncertainty set is the following
\begin{eqnarray}
\mathcal{M} = \{\text{vec}(\bar{M})+ R u, ||u||_2 \leq 1\},\nonumber
\end{eqnarray}
where $R\in \mathbb{R}^{nm\times nm}$ is defined by the Cholesky decomposition, $\Sigma = R^\top R$. Then, using the same derivation as above, the modified $\mathcal{M}$ can be represented as a SOC constraint
\begin{eqnarray}
\label{eqn:SOC_with_cov}
||R^\top (\text{vec}(M)-\text{vec}(\bar{M}))||_2^2 \leq 1.
\end{eqnarray}
Again, just as in the spherical case, we can compute the uncertain dense realization. For the complete treatment of uncertain kinetic systems and the proofs, see \cite{acs2018}.
It should be mentioned that other types of uncertainty are possible for a kinetic system, as long as the uncertainty set can be translated as convex constraints, the dense realization exists, see \cite{acs2018}.
Besides the framework described in this paper, a useful application of this technique could be the design of dynamics, i.e. designing a kinetic system which operates inside the prescribed operational limits or design envelope, see \cite{liptak2016kinetic} for details on CRN controller design.
At this point, we have the tools to compute the uncertain dense realization. As it was shown in \cite{acs2018} that the all possible realizations can be computed in the uncertain case as well.
In summary, in order to calculate the dense and subsequently all realizations, we need to define $Y$, the nominal coefficient matrix $\bar{M}$ and a spherical or ellipsoidal uncertainty. Therefore, the next step is to compute $\bar{M}$ from time-series data with the assumption of that $Y$ is known.
\section{Parameter estimation}
\label{sec:paramest}
There are many possible ways to estimate the parameters of a kinetic system from time-series data, see e.g. \cite{August2009} or \cite{wei2017phd}. In this paper, we work with the following assumptions: all state variables can be measured, and the set of complexes (i.e. the matrix $Y$) is known \textit{a priori}. The former assumption can be relaxed, by using state estimation for the unmeasured states. However, $Y$ is usually assumed to be known, because it represents our knowledge about the participating chemical complexes.
For the purpose of parameter estimation, we need to discretize the kinetic system in \eqref{eqn:kinetic_sys}. Using sufficiently small sampling time, we apply the forward Euler method and get
\begin{eqnarray}
x_i(t_k) &=& x_i(t_{k-1}) + h[M]_{i,\cdot}\cdot\psi_i^Y(x(t_{k-1})),\\
x(0)&=&x_0,~~k=1,\ldots,N,~~i=1,\dots,n, \nonumber
\end{eqnarray}
where $t_k$ is the sampling time point, $x_i(t_k)$ is the $i$th state variable at time $t_k$, $x_0\in\mathbb{R}^n_{\geq 0}$ are the initial values, the $\psi^Y_i(x)$ is the $i$th element of vector mapping $\psi^Y(x)$, the vector $[M]_{i,\cdot}$ is the $i$th row of matrix $M$, the $h$ is the sampling time, and the $N$ is the last sampling time point.
For the framework later on, we need the time derivate of $x$. It can be estimated in many ways (see \cite{debrabanter2013} for details). In our case, it is given from the forward Euler method. Therefore, we assume that the measurement of $i$th state variable of the discrete kinetic system is available in this transformed form
\begin{eqnarray}
y^{(i)}(t_k) := \frac{x_i(t_k) - x_i(t_{k-1})}{h},~k=1,\ldots,N,~i=1,\dots,n. \nonumber
\end{eqnarray}
Then, we get a linear process for each state variable which is linear in parameters, and the $i$th state is given as
\begin{eqnarray}
\label{eq:lin_model}
\tilde{y}^{(i)}(t_k,\theta) = \Phi(t_{k-1})\theta^{(i)\top} +\nu_i(t_k),
\end{eqnarray}
where the parameter vector is defined as $\theta^{(i)}=[M]_{i,\cdot}$ and the regressor vector is given as
\begin{eqnarray}
&&\Phi(t_{k-1})= \\
&& \begin{bmatrix}
\psi_1(x(t_{k-1})), & \psi_2(x(t_{k-1})), & \ldots, & \psi_m(x(t_{k-1}))
\end{bmatrix}. \nonumber
\end{eqnarray}
and the measurement noise is $\nu_i \sim \mathcal{N}(0,\sigma^2)$. We assume that the distribution of measurement noise is the same for all the output channels.
It must be emphasized that kinetic system with mass action kinetics is always linear in parameters, therefore the standard algorithms and tools for analysis from the parameter estimation literature can be applied in this case \cite{ljung_sysid}. For example, we can use the well-known Least Squares method to calculate the parameters of \eqref{eq:lin_model} as the following
\begin{eqnarray}
\label{eqn:lsq_with_eq}
\hat{\theta}^{(i)} &=& \argmin_{\theta} \frac{1}{2}|| y^{(i)} - \Phi\theta ||_2^2
\end{eqnarray}
The problem with this path is that the Least Squares method does not promote sparsity. In fact, it rather tries to associate non-zero value to all parameters. However, the dynamics of a state variable is usually not driven by all the monomials in $\varphi(x)$, but only a subset of them.
Therefore, we need to have either a constrained parameter estimation method, which knowns \textit{a priori} the zero coefficients or a parameter estimation method that promotes sparsity. Among many candidates for the later one, Sparse Bayesian Learning gained popularity recently, mostly because of guarantees for convergence and sparsity. However, evolutionary computation \cite{Schmidt2009} or heuristic based \cite{Brunton2016} algorithms were also proposed recently to find parsimonious models from time-series data.
\subsection{Sparse Bayesian Learning}
\label{sec:SBL}
Sparse Bayesian Learning was proposed by Tipping and was applied to Relevance Vector Machines where the task is to find a sparse regression or classification \cite{Tipping2001}. Independently from the Bayesian framework, Candes \textit{et al.} developed a framework that uses iterative reweighting of the L1 norm penalty on the parameters \cite{Candes2008}. Candes \textit{et al.} makes the connection to MM algorithms, which is a fundamental way to iteratively solve non-convex optimizations problems. Recently, Wipf \textit{et al.}, building on the work of Candes and Tipping developed a framework which uses iterative reweighting of either L1 or L2 norm to find sparse solution of broad range of problems, e.g. sparse signal representation \cite{Wipf2010}, \cite{Wipf2011}, automatic relevance determination \cite{Wipf2008}, source localization on MRI measurements \cite{Wipf2009}.
Here, we only give a short outline of the Sparse Bayesian framework, and therefore readers are strongly encouraged to read \cite{Wipf2011} and \cite{Wipf2010} for a thorough treatment of the subject. The following introduction follows the notations from \cite{Wipf2010}.
Throughout the derivation, we assume that we have the following process model
\begin{eqnarray}
y = \Phi\theta + \nu
\end{eqnarray}
where $y\in\mathbb{R}^{N}$ is the measurement vector, $\Phi\in\mathbb{R}^{N\times m}$ is a dictionary of features, $\theta\in\mathbb{R}^{m}$ is the parameter vector, and $\nu \sim \mathcal{N}(0,\lambda I)$ is the measurement noise. Our goal is to estimate the sparsest $\theta$, which then select the sparest dictionary to describe the measurements.
As first step in the derivation, the Least Squares problem in \eqref{eqn:lsq_with_eq} can be transformed into a Gaussian likelihood as
\begin{eqnarray}
p(y|\theta) \propto \exp\left(-\frac{1}{2}||y-\Phi\theta||_2^2\right)
\end{eqnarray}
The prior information about the parameters is expressed in terms of non-negative latent variables $\gamma\in \mathbb{R}_{\geq 0}^{m}$ as follows
\begin{eqnarray}
p(\theta) \propto \prod_{i=1}^{m} p(\theta_i),\qquad p(\theta_i) = \max_{\gamma_i \geq 0} \mathcal{N}(\theta_i;0,\gamma_i)\zeta(\gamma_i)
\end{eqnarray}
where $\zeta(\gamma_i)$ is a non-negative function. This form of the prior distribution allows us to express wide range of penalty functions that are needed to promote sparsity in the parameter vector \cite{palmer2006variational}. This structure already hints that if we can set the variance of some of the parameters to zero, then the corresponding parameter value becomes zero. The derivation of the algorithm can be found in the Appendix.
To find the parameter values for $\theta$ and $\gamma$, we need to solve the following iterated optimization problem
\begin{itemize}
\item Step 1: initialize each $z_{i}=1, i=1,\ldots,m$
\item Step 2: $\hat{\theta} = \argmin_\theta ||y-\Phi \theta||_{2}^{2}+2\lambda \sum_{i}z_i^{-1/2}|\theta_i|$
\item Step 3: compute $\gamma^{opt}_i = z_i^{-1/2}|\hat{\theta}_i|, i=1,\ldots,m$
\item Step 4: compute $z^{opt} = \nabla_\gamma \log |\Sigma_y|$
\item Step 5: iterate Step 2, 3 and 4 until $\gamma$ is converged to some value.
\end{itemize}
It should be noted that the first iteration of the above algorithm is the LASSO optimization, and we try to improve on that, hence the name iterated reweighted L1 optimization.
\subsection{Examples}
We illustrate the framework shown in Figure \ref{fig:overall_scheme} on an example. This example was presented in \cite{August2009} as a benchmark problem for network inference. The network structure is shown in Figure \ref{fig:acc_network} and its CRN formulation is given as
\begin{small}
\begin{eqnarray}
Y = \begin{bmatrix}
1 & 0 & 1 & 0 & 0\\
0 & 2 & 0 & 0 & 1\\
0 & 0 & 1 & 0 & 0\\
0 & 0 & 0 & 1 & 0\\
0 & 0 & 0 & 0 & 1
\end{bmatrix}
\end{eqnarray}
\end{small}
and
\begin{small}
\begin{eqnarray}
\label{eq:example_ak_param_est}
A_k &=&
\begin{bmatrix}
-1.163 & 0 & 0 & 0 & 0.8492 \\
0.3386 & 0 & 0 & 0 & 0.4290 \\
0.8244 & 0 & -0.7364 & 0.5631 & 0 \\
0 & 0 & 0 & -0.5631 & 0 \\
0 & 0 & 0.7364 & 0 & -1.2782
\end{bmatrix},
\end{eqnarray}
\end{small}
the CRN has 5 complexes and 6 reactions, the coefficient matrix is given as $M=Y A_{\kappa}$. The dynamically equivalent (noiseless case) dense realization has also 6 reactions and the network has no other dynamically equivalent realization, i.e. in the noiseless case only one network structure exists.
\begin{figure}
\begin{center}
\input{ACC_model_CRN}
\input{ACC_model_CRN_dense_level_1_12}
\end{center}
\caption{Left: this example network is taken from \cite{August2009}. The parameters of the network are
$k_{1,2} = 0.3386$, $k_{1,3} = 0.8244$, $k_{5,1} = 0.8496$, $k_{5,2} = 0.4290$, $k_{3,5} = 0.7364$ and $k_{4,3} = 0.5630$. Right: the dense realization computed from data with $\sigma^2=10^{-4}$ and confidence level $\alpha=0.05$. The parameters of the network are $k_{1,2}=0.2920$, $k_{5,2}=0.2645$, $k_{5,1}=0.9208$, $k_{1,3}=1.2262$, $k_{4,3}=0.6108$, $k_{3,5}=0.6495$, $k_{3,1}=0.3227$, $k_{3,2}=0.1800$ and $k_{4,1}=0.5714$.}
\label{fig:acc_network}
\end{figure}
\subsubsection{Least Squares Method}
In the first part, we test the performance of the simple Least Squares method assuming that the zero elements of matrix $M$ are known, i.e. we have a constrained Least Squares with equality constraints (LSE). For this part, we generated 50 experiments with different initial values, sampled with Latin Hypercube sampling between $[0,1]$, and simulated the process for $T=10$ sec. For each state variable, we had the following measurement noise $\sigma^2=10^{-4}$, and the sampling time was $h=0.01$.
From the LSE, the covariance matrix is available in close form. We then set the confidence level at $\alpha=0.05$ and computed the dense realization of the uncertain kinetic system with Algorithm 1 from \cite{Acs2015}; the dense realization is shown in the right panel of Figure \ref{fig:acc_network}. This realization has 9 reactions, thus the extra three reactions open up the possibility for multiple realizations. After computing all possible realizations using the algorithm from \cite{TuzaFosbe2016}, we got 56 structurally different reaction networks.
Clearly, even in this generous scenario---all state variables are directly measured, limited measurement noise is added---gave us several different network structures within the given uncertainty bounds.
As a next step, we show for this example how the number of possible structures depend on the measurement noise. For that reason, we generated 100 different $\sigma^2$ values between $10^{-4}$ and $10^{1}$. Then, the overall procedure from Figure \ref{fig:overall_scheme} was done for all the different noise scenarios. The Figure \ref{fig:level_vs_real_num} shows the results.
\begin{figure}
\begin{tikzpicture}
\begin{axis}[xmode=log,
scale only axis=true,
width=0.4\textwidth,
height=0.22\textwidth,
extra y ticks={511},
extra tick style={grid=major},
ytick ={10,50,100,200,300,400},
xlabel=\textsc{$\sigma^2$ - $\log$ scale},
ylabel=\textsc{Number of realizations}]
\addplot[only marks,scatter,scatter src=explicit] table [x index=1,y index=3, meta index=0, col sep=comma] {noise_vs_realnum_paramest_ACC_model.csv};
\end{axis}
\end{tikzpicture}
\caption{Number of realizations depending on the variance of the measurement noise. LSE case.}
\label{fig:level_vs_real_num}
\end{figure}
As we can see in Figure \ref{fig:level_vs_real_num}, the number of realizations saturates above a given noise level. This is because we have reached the combinatorially possible number of realizations, which is given by
\begin{eqnarray}
\label{eqn:comb_max}
R_{max} = \sum_{i=1}^{R_d} \begin{pmatrix}
R_d\\
i
\end{pmatrix}
\end{eqnarray}
where $R_d$ is the number of edges in the dense realization. If we have some prior information on the minimum number edges in the network, then the index $i$ can start from that number. Clearly, then we have fewer number of possible network structures.
If we reach the maximum number of possible realizations for a kinetic system that means that the parameters estimation provided no restriction on the possible structures. Therefore, by computing $\frac{R_{\lambda}}{R_{max}}$, where $R_{\lambda}$ is the number of realizations for a given noise level, we have a simple measure which tells us how much information we gained from the parameter estimation about the possible network structures.
\subsubsection{Sparse Bayesian Learning}
In the second part, we test the performance of the SBL algorithm. The example is the same as before, but this case we did not assume we know the zero entries of $M$ by using equality constraints; it will be estimated by the SBL algorithm. We have generated only $10$ different initial values and simulated the system for $T=10$ sec with sampling time $h=0.1$. Thus, we have significantly less measurement data for all state variables, then in the previous case.
Again, we generated different measurement noise levels and executed the overall procedure from Figure \ref{fig:overall_scheme}. The results of this part is summarized in Figure \ref{fig:level_vs_real_num_SBL}. As we can see, we have significantly fewer possible realizations in the investigated range. In this range of noise level the position of the zero elements in the $M$ matrix was correctly estimated, outside this range the sparsity pattern was not estimated correctly. Additional measurements could potentially extend the range where the sparsity pattern can be restored, but this falls outside of the scope of the current paper.
At this stage, we completely characterized the possible network structures. By adding different \textit{a priori} knowledge, we can measure the effect of this knowledge on the number of structures and compare them with each other. From the dense realization, shown in right panel of Figure \ref{fig:acc_network}, we identify three reactions that are possible reactions given the data, but the are not part of the original network. By excluding these reaction one at the time, we can compare the resulting number of realizations. This is shown in Figure \ref{fig:level_vs_real_num_SBL}, shown with different symbols.
\begin{figure}
\begin{tikzpicture}
\begin{axis}
scale only axis=true,
width=0.4\textwidth,
height=0.15\textwidth,
xmode=log,
legend columns=2,
legend style={at={(0.5,-0.2)},anchor=north},
scatter/classes=
a={mark=square*,blue}
b={mark=triangle*,red}
c={mark=o,draw=blue},
d={mark=x,draw=black}},
ytick ={5,7,10,11},
ylabel absolute, ylabel style={xshift=-0.8cm,yshift=-0.3cm},
xlabel=\textsc{$\sigma^2$ - $\log$ scale},
ylabel=\textsc{Number of realizations}]
\addplot[scatter,only marks
scatter src=explicit symbolic
table[meta=label] {
x y label
0.000100 11 a
0.000100 5 b
0.000100 7 c
0.000100 7 d
0.000316 11 a
0.000316 5 b
0.000316 7 c
0.000316 7 d
0.001000 10 a
0.001000 5 b
0.001000 7 c
0.001000 7 d
0.001000 7 d
0.00316 11 a
0.00316 5 b
0.00316 7 c
0.00316 7 d
0.01000 10 a
0.01000 5 b
0.01000 7 c
0.01000 7 d
0.0316 11 a
0.0316 5 b
0.0316 7 c
0.0316 7 d
0.1000 11 a
0.1000 5 b
0.1000 7 c
0.1000 7 d
0.316 11 a
0.316 5 b
0.316 7 c
0.316 7 d
};
\addlegendentry{$\mathcal{H}=\emptyset$}
\addlegendentry{$\mathcal{H}=[C_4\rightarrow C_1]$}
\addlegendentry{$\mathcal{H}=[C_3\rightarrow C_1]$}
\addlegendentry{$\mathcal{H}=[C_3\rightarrow C_2]$}
\end{axis}
\end{tikzpicture}
\caption{Number of realizations depending on the variance of the measurement noise in the SBL case.}
\label{fig:level_vs_real_num_SBL}
\end{figure}
\section{INTRODUCTION}
One of the goals in the Systems and Synthetic Biology is to characterize possible network structures that can explain the observed data. This is usually done by incorporating the information content of noisy experimental data into parametrized process models \cite{ljung_sysid}. Then, these network structures give us a blueprint for the possible interactions between chemical species. Building on that, one can understand a complex biological process or even manipulated it by other chemical species. Synthetic Biology is particularly successful identifying small interaction networks in nature or even building artificial ones from well understood biological parts such as the repressilator. The need to grow this library of well-characterized biological parts to build more complex interaction networks drove the development of high-throughput data acquisition. This type of data collection soon generates a large volume of data which is incompatible with the current model building methodology. Mainly because of the current practice involves manually curating the data and the possible network structure. As the need for fast characterization of biological parts drove development of high-throughput methods, the abundant data from high-throughput methods drives the need for automatic model building methods.
Automatic model building is standard practice in, for example, machine learning, but usually, the building blocks of such models are general nonlinear functions, e.g. Gaussian kernels. In contrast to that, in Biochemistry we have a well-understood family of nonlinear functions that can capture the underlying chemical interactions. Even first principle models can be built by these nonlinearities. For example, Chemical Reactions Networks (CRNs) are often used to build such models.
In CRN modeling, it is usually assumed that there is a one-to-one correspondence between the dynamical model and the underlying CRN structure. However, this is only true if one builds the differential equations from the CRN structure. The other way may yield multiple structures that exhibit the same dynamics. Since the later direction is used in system identification, it needs to be carefully investigated.
The existence of multiple network structures for a given dynamics has been investigated extensively in \cite{Johnston2011conj,acs2016}. However, this investigation was done with the assumption of perfect measurements, and the uncertain case was investigated in \cite{acs2018}. As we show in this paper, this allows us to handle noisy time-series data and to compute network structures using the same computational tools developed for the noiseless case.
The goal of the paper is to develop a framework that characterizes all possible network structures from time-series data. All possible structure means that it can be proven the set of network structures, computed in our framework, is complete and no other other network structure exists for a given dynamical model. Complementing the underlying optimization problem with further constraints on the dynamics or on the structure helps to reduce the number of possible network structures, thus we can not only add \textit{a priori} knowledge to the identification process but measure its impact on the number of possible network structure. This allows us to compare the effect of different \textit{a priori} knowledge. Moreover, we can even characterize the set of assumptions needed for a unique the network structure from time-series data.
\paragraph*{Notations.} $\mathbb{R}_{\geq 0}$ denotes the nonnegative real numbers, $\mathbb{N}_0$ is the set of integers including zero. $[A]_{ij}$ is the entry in the $i$th row and $j$th column of the matrix $A$. Furthermore, $[A]_{i\cdot}$ is the $i$th row of matrix $A$, and diag$[a]$ is matrix which has the elements of $a$ in the diagonal and the rest is zero. Finally, $A\succeq 0$ is a positive semidefinite matrix and $\text{vec}(A) = [[A]_{1\cdot}, [A]_{2\cdot}, \ldots, [A]_{n\cdot}]^\top$ is transpose of the row expansion of matrix $A\in\mathbb{R}^{n\times m}$.
\section{Characterizing network structures}
At this point, we can merge the Sparse Bayesian Learning algorithm and the computational tools developed for CRNs into a framework which is shown in Figure \ref{fig:overall_scheme}. As a first step, time-series data are feed to the parameter estimation step (see, Section \ref{sec:SBL}), then using the optimization problems from Section \ref{sec:uncertain_opt}, the uncertain dense realization is computed. From this, all possible graph structures are computed. Thus, this framework transforms the available information in the time-series data into possible network structures relying only on few assumptions.
Both parts of the framework solve optimization problems. Therefore these optimization problems can be complemented with constraints enforcing structural or dynamical properties.
To visulize our the results, the so-called Feinberg-Horn-Jackson graph is used, which is a weighted directed graph. In this graph, the vertices are the complexes, the edges are the reactions, and the weights are the reaction rate coefficients ($k_{ij}$).
\begin{figure}
\centering
\begin{tikzpicture}[auto, node distance=3cm,>=latex']
\node [input, name=rinput] (rinput) {};
\node [block, below of=rinput, text width=1cm,node distance=1.2cm] (param_est) {\footnotesize Param estim. w/ SBL};
\node [block, right of=param_est, text width=1.95cm] (dense) {\footnotesize Comp. the dense realization};
\node [block, below of=dense, text width=1.85cm, , node distance=1.5cm] (core) {\footnotesize Struc. or dyn. constraints};
\node [block, right of=dense, text width=1.8cm,node distance=3.3cm] (all_real) {\footnotesize Comp. of all realizations};
\node [output, below of=all_real,node distance=1.2cm] (routput) {};
\draw [->] (rinput) -- node{$\mathcal{D}$} (param_est);
\draw [->] (param_est) -- node{$\bar{M},\Sigma$} (dense);
\draw [->,dashed,pos=0.1] (core) -| ++(1.5,1) |- node[xshift=0.4cm, yshift=-1.3cm]{} (all_real.200);
\draw [->,dashed] (core) -| node{} (param_est.south);
\draw [->,pos=0.41] (dense) -- node{$A_k^D$} (all_real);
\draw [->] (all_real) -- node[xshift=-1.195cm, yshift=-0.8cm]{$\overbrace{A_k^{(1)},\ldots,A_k^{(P)}}$} (routput);
\draw [->,dashed,pos=0.7,above] ([xshift=1.2cm, yshift=-0.5cm] routput) -| ++(0.15,1) |- node{further \textit{a priori} information} ([xshift=0.5cm, yshift=-0.32cm] rinput);
\end{tikzpicture}
\caption{The figure depicts the overall scheme of the framework developed in this paper. $\mathcal{D}$ denotes the time-series data collected from experiments. Then, this information is used to estimate the parameter of the model $\bar{M}$, along with the covariance matrix $\Sigma$. From these, the uncertain dense realization is computed by Algorithm 1 in \cite{Acs2015}. In the final step, all possible realizations are computed and this set can be further analyzed. The framework allows us to complement either the parameter estimation or the structure computation with constraints, thus the possible network structures can be further reduced.}
\label{fig:overall_scheme}
\end{figure}
\section{CONCLUSIONS AND FUTURE WORKS}
We have developed a framework that computes the possible network structures from time-series data. Assuming that all state variables are measured and the participating chemical complexes are known, we have used the Sparse Bayesian Learning to estimate the parameter of the dynamical model. From the statistics of the parameter estimation all possible network structures have been computed. We have given a simple metric to judge the information content of the time-series data as a ratio of the current number of network structures and combinatorially possible ones. We have also shown that we can control the number of network structures by adding additional {a priori} information. In fact, we can compare the effect of different {a priori} information on the number of possible structures, since without any constraints the framework---by construction---provides all possible network structures. As future work, we would like to further characterize data and noise dependence of the Sparse Bayesian Learning algorithm and add further constraints to limit the possible network structures or automatically generate the set of constraints needed for a unique network structure.
\section{ACKNOWLEDGMENTS}
This work was developed within the project called COSY-BIO (Control Engineering of Biological Systems for Reliable Synthetic Biology): H2020-FETOPEN-2016-2017, project number 766840. This work was also supported by the U.K. EPSRC Fellowship EP/M002187/1 to Dr Stan. The first author is thankful for Gabor Szederkenyi for his advises on the CRN definitions.
|
\section{Introduction}
\label{sect:intro}
Motivated by the exponentially increasing demand for higher information rate under limited wireless resources, and propelled by recent advances in \ac{RF} hardware, in-band \acf{FD} radio has emerged as a key technology for future wireless applications from \ac{5G} mobile communication systems to \ac{IoT} \cite{bharadia2013fullduplex ,Kim2015, Sabharwal2014SAiC}.
Practical communication in \ac{FD} mode requires dedicated solutions to mitigate the \acf{SI} caused by leakage of \ac{TX} signals into the \acf{RX} chain, due to the close proximity between \acf{TX} and \ac{RX} antennas \cite{Zhang2015CM,Choi2010MobiCom}.
Ironing out this fundamental issue of \ac{FD} technology is one of the main research topics in this field, motivating various authors to contribute with several \ac{SI} cancellation techniques for \ac{FD} systems \cite{bharadia2013fullduplex, Everett2011Asilomar, Wang2015WCSP, Altieri2014SAiC}.
Thanks to the added \ac{DoF} afforded by multiple antennas, bi-directional \ac{FD} radio systems with high spectral efficiency can be designed exploiting \ac{MIMO} technology \cite{vehkapera2013asymptotic, DaySP2012, jia2017signaling, Riihonen2013CISS}.
In particular, hybrid \ac{MIMO} \ac{SI} suppressing techniques combining analog and digital cancellers have been proposed for \ac{FD} radios \cite{GowdaTWC2018, MyListOfPapers:KorpiGlobecom2014, Sim2017CM, Cirik2013Asilomar} which proved very effective from a theoretical standpoint.
From a practical implementation standpoint, however, it has been recently demonstrated in real-world experiments that such \ac{MIMO} approaches are not devoid of its own technical challenges \cite{Jain2011MobiCom, korpi2016fullduplex, bharadia2013fullduplex}, one of which is the excessive cost incurred by the use of large numbers of antennas.
One approach to keep the hardware cost of hybrid \ac{MIMO} \ac{SI} suppressing techniques for \ac{FD} radios under control is to reduce the number of antennas while introducing temporal \acp{DoF} by means of \ac{TDL} processing in order to maintain the \ac{DoF} required to achieve the desired performance \cite{Alexandropoulos2017}.
In \cite{Alexandropoulos2017}, for instance, a joint hybrid \ac{TX}-\ac{RX} \ac{BF} design with limited hardware costs was proposed, in which the sum rate of a system with one \ac{MIMO} \ac{FD} radio communicating with two \ac{MIMO} \ac{HD} nodes was optimized.
In this paper, we contribute to the area of effective and feasible \ac{SI} canceller designs for \ac{MIMO} \ac{FD} radios as follows.
First, we combine the joint hybrid \ac{TX}-\ac{RX} approach of \cite{Alexandropoulos2017} with the analog cancellation technique referred to as \emph{multi-tap analog canceller} previously presented in \cite{Kolodziej2016TWC}.
The result is a new multi-tap \emph{hybrid} (analog and digital) \ac{TX}-\ac{RX} \ac{MIMO} \ac{FD} \ac{SI} cancellation scheme, in which the number of hardware components for analog cancellation becomes independent of the number of antennas.
Secondly, instead of maximizing the sum rate (which is of less practical interest), we formulate our problem to minimize the TX power while guaranteeing (when possible) prescribed \ac{QoS} targets defined in terms of maximum \acf{SINR}.
Thirdly and finally, we present a low-complexity solution to the latter problem in which
the \ac{TX} employs \ac{MRT} with powers optimized in closed-form via a \ac{PF} method, while the \ac{RX} maximizes the SINR by computing corresponding closed-form \ac{RX} \ac{BF} vectors from a \ac{RQ}, iteratively.
Our results show that our algorithm can outperform the similar methods previously proposed for 2-user MIMO systems in terms of both power efficiency and computational complexity.
\vspace{-0.6ex}
\section{System Model}
\label{System_Model}
\vspace{-0.2ex}
Consider the two-way \ac{FD} \ac{MIMO} communication system illustrated in Figure \ref{fig:System_model}.
This system consists of two node in which each equipped with $M$ \ac{TX} and $N$ receive antennas.
Both nodes are assumed to \ac{TX} and receive simultaneously to/from one another in the same resource unit.
\begin{figure}[H]
\center
\includegraphics[width=\columnwidth]{System_Model.eps}
\caption{System model of two-way full duplex MIMO with reduced hardware multi-tap analog cancellation.}
\label{fig:System_model}
\vspace{-1ex}
\end{figure}
A generic $k$-th node, with $k\in\{1,2\}$, is assumed to employ the digital \ac{TX} precoding vector $\bm{v}_{k}\in\mathbb{C}^{M\times 1}$ and the digital \ac{RX} \ac{BF} vector $\bm{u}_{k}\in\mathbb{C}^{1\times N}$, as well as the multi-tap analog cancellation matrix $\bm{C}_{k}\in\mathbb{C}^{N\times M}$ \cite{Alexandropoulos2017, Kolodziej2016TWC}.
It is capable of performing TX-RX digital \ac{BF} and analog \ac{SI} cancellation with the aim at suppressing \ac{SI} and maximizing rate simultaneously.
Finally, in order to model practical limitations, it is assumed that the \ac{TX}ted signal at the $k$-th node has a power upper bound, such that $\tr{\bm{v}_{k}\bm{v}^{\rm H}_{k}}= P_{k}\leq P_{\rm max}$.
Referring to Figure \ref{fig:System_model}, let $\bm{H}_{k\ell}\in\mathbb{C}^{N\times M}$ and $\bm{H}_{kk}\in\mathbb{C}^{N\times M}$ be the intended channel matrix between the two nodes and the \ac{SI} channel matrix at the $k$-th node, respectively, with $k\neq\ell\in\{1,2\}$.
It is also assumed throughout this paper that each node has full knowledge of the \ac{CSI} of both the communication links and their own \ac{SI} link.
Extension to imperfect \ac{CSI} knowledge is left for future work.
From all the above, the received signal at the $k$-th node after applying analog \ac{SI} cancellation can be written as
\begin{eqnarray}
\label{eq:received_k}
\bm{y}_{k} &=& \bm{H}_{\ell k}\bm{v}_{\ell}s_{\ell} + \left(\bm{H}_{kk} - \bm{C}_{k}\right)\bm{v}_{k}s_{k} + \bm{n}_{k}\\
&=& \underbrace{\bm{H}_{\ell k}\bm{v}_{\ell}s_{\ell}}_{\rm Intended} + \underbrace{\tilde{\bm{H}}_{kk}\bm{v}_{k}s_{k}}_{\rm SI} + \underbrace{\bm{n}_{k}}_{\rm Noise}\nonumber,
\end{eqnarray}
where the multi-tap analog cancellation matrix $\bm{C}_{k}$ consists of $N_{\rm tap}$ non-zero components and $MN - N_{\rm tap}$ zeros, $\bm{n}_{k}\sim \mathcal{CN}\left(0,\sigma^2\bm{I}_{N}\right)$ denotes the complex \ac{AWGN} vector under the assumption that $\bm{n}_{k}$ is independent from the \ac{TX}ted signal $s_{\ell}$, and $\tilde{\bm{H}}_{kk}\triangleq\bm{H}_{kk} - \bm{C}_{k}$ is the \ac{SI} channel matrix after performing the considered analog cancellation.
After digital down conversion and combining by the \ac{RX} \ac{BF} vector $\bm{u}_{k}$, the estimated signal $\tilde{s}_{\ell}$ corresponding to the intended signal $s_{\ell}$ at the $k$-th node can be expressed as
\begin{eqnarray}
\label{eq:estimated_k}
\tilde{s}_{\ell} &=& \bm{u}_{k}\bm{y}_{k}\\
&=&\bm{u}_{k}\bm{H}_{\ell k}\bm{v}_{\ell}s_{\ell} + \bm{u}_{k}\tilde{\bm{H}}_{kk}\bm{v}_{k}s_{k} + \bm{u}_{k}\bm{n}_{k}.\nonumber
\end{eqnarray}
Similarly, the received signal and symbol estimate at node $\ell\neq k$ after analog cancellation and \ac{RX} \ac{BF} are given, respectively, by
\begin{eqnarray}
\label{eq:received_l}
\bm{y}_{\ell} &=& \bm{H}_{k\ell}\bm{v}_{k}s_{k} + \tilde{\bm{H}}_{\ell\ell}\bm{v}_{\ell}s_{\ell} + \bm{n}_{\ell},\\
\label{eq:estimated_l}
\tilde{s}_{k} &=& \bm{u}_{\ell}\bm{y}_{\ell}\nonumber\\
&=&\bm{u}_{\ell}\bm{H}_{k\ell}\bm{v}_{k}s_{k} + \bm{u}_{\ell}\tilde{\bm{H}}_{\ell\ell}\bm{v}_{\ell}s_{\ell} + \bm{u}_{\ell}\bm{n}_{\ell},
\end{eqnarray}
where $\bm{n}_{\ell}\sim \mathcal{CN}\left(0,\sigma^2\bm{I}_{N}\right)$ is the AWGN vector that is assumed independent from the \ac{TX}ted symbol $s_{k}$.
Assuming that unit power information signals $s_{k}$ and $s_{\ell}$ are used, the average SINR estimates at the two nodes in Figure \ref{fig:System_model} can be, respectively, written as
\begin{equation}
\label{eq:SINR}
\gamma_{k} = \frac{\abs{\bm{u}_{k}\bm{H}_{\ell k}\bm{v}_{\ell}}^2}{\abs{\bm{u}_{k}\tilde{\bm{H}}_{kk}\bm{v}_{k}}^2 + \sigma^2} \quad \text{and}\quad
\gamma_{\ell} = \frac{\abs{\bm{u}_{\ell}\bm{H}_{k\ell}\bm{v}_{k}}^2}{\abs{\bm{u}_{\ell}\tilde{\bm{H}}_{\ell\ell}\bm{v}_{\ell}}^2 + \sigma^2},
\end{equation}
where we assume that the channel matrices in equation \eqref{eq:SINR} are constant for a number of signal transmissions and the \ac{RX} combining vector $\bm{u}_{k},\forall k$ has a unit norm, i.e., $\norm{\bm{u}_{k}}^2=1$.
\vspace{-1ex}
\section{QoS-Guaranteed Transmissions}
\label{sect:OptimalTS}
\vspace{-1ex}
Signal processing techniques for the joint TX-RX linear precoding/combining and adaptive \ac{TX} power allocation with the aim of maximizing data rate while suppressing the residual \ac{SI} power level have been proposed in the past \cite{Zheng2015TSP, IimoriSPAWC2018} demonstrating the feasibility of two-way \ac{FD} \ac{MIMO} systems.
Maximizing data rate is, however, not typically required by actual users, which instead tend to perceive the
quality of a communication system by comparing it to a given level of expectation dictated by the intended application.
We therefore consider instead the TX-RX beamformer optimization problem aiming at minimizing the individual \ac{TX} powers while satisfying individual target SINR requirements:
\begin{subequations}
\label{MMES_OP_Problem}
\begin{eqnarray}
\label{OP1}
\min_{\bm{v}_{k},\bm{v}_{\ell}}&& \sum^{2}_{k=1}\norm{\bm{v}_{k}}^2\\
\label{ConstraintP}
{\rm s.t.}&& \gamma_{k} \geq \Gamma_{k}\:\: \forall k,
\end{eqnarray}
\end{subequations}
where $\Gamma_{k}$ is the target SINR for the $k$-th node.
\vspace{-0.5ex}
\subsection{\ac{TX} Power Minimization with SINR Constraints}
\label{sect:minpower}
\vspace{-0.5ex}
Let us define the normalized precoding vector $\bar{\bm{v}}_{k}\triangleq\frac{\bm{v}_{k}}{\norm{\bm{v}_{k}}}$ and the \ac{TX} power $P_{k}=\norm{\bm{v}_{k}}^2$ such that the optimization problem in \eqref{MMES_OP_Problem} can be rewritten as
\begin{subequations}
\label{OP2}
\vspace{-0.5ex}
\begin{eqnarray}
\min_{P_{1},P_{2}}&& \sum^{2}_{k=1}P_{k}\\
\label{ConstraintP}
{\rm s.t.}&& \gamma_{k} \geq \Gamma_{k} \:\: \forall k.
\end{eqnarray}
\end{subequations}
The optimization problem described by equation \eqref{OP2} is well-known to be non convex due to the SINR constraints \cite{Malla2015CSIT}, although approximate solutions can be obtained for it with basis on convex optimization algorithms, such as interior point methods if the constraint can be convexified \cite{Boyd2004}.
In addition to the losses due to convex relaxation, such solutions tend also to be computationally demanding.
Therefore, we propose instead a low complexity alternating minimization method based on closed-form expressions of the optimal \ac{TX} powers $P_{1}$ and $P_{2}$.
In order to obtain the desired closed-form expressions for $P_{1}$ and $P_{2}$, notice that from equation \eqref{eq:SINR} and \eqref{OP2} we readily obtain
\begin{subequations}
\label{eqn:ineqPopt}
\begin{eqnarray}
&P_{2}\abs{\bm{u}_{1}\bm{H}_{21}\bar{\bm{v}}_{2}}^2 \geq \Gamma_{1}\Big({P_{1}\abs{\bm{u}_{1}\tilde{\bm{H}}_{11}\bar{\bm{v}}_{1}}^2 + \sigma^2}\Big),\\
&P_{1}\abs{\bm{u}_{2}\bm{H}_{12}\bar{\bm{v}}_{1}}^2 \geq \Gamma_{2}\left({P_{2}\abs{\bm{u}_{2}\tilde{\bm{H}}_{22}\bar{\bm{v}}_{2}}^2 + \sigma^2}\right).
\end{eqnarray}
\end{subequations}
The latter inequalities can be re-expressed in matrix form as
\begin{eqnarray}
\label{eqn:ineqPoptMat}
\left({\bf{I}}- \bm{\Gamma}\bm{M}\right)\bm{p}\geq\sigma^2 \bm{\Gamma} \bm{m},
\end{eqnarray}
where we define the \ac{TX} power vector $\bm{p} \triangleq \left[P_{1},P_{2}\right]^{\rm T}$ and the auxiliary matrices $\bm{\Gamma}$, $\bm{M}$ and $\bm{m}$ respectively by
\begin{subequations}
\begin{eqnarray}
\bm{\Gamma} &=&
\begin{bmatrix}
0 & \Gamma_{2} \\
\Gamma_{1} & 0
\end{bmatrix},\\
\bm{M} &=&
\begin{bmatrix}
\frac{|\bm{u}_{1}\tilde{\bm{H}}_{11}\bar{\bm{v}}_{1}|^2}{|\bm{u}_{1}\bm{H}_{21}\bar{\bm{v}}_{2}|^2} & 0\\
0 & \frac{|\bm{u}_{2}\tilde{\bm{H}}_{22}\bar{\bm{v}}_{2}|^2}{|\bm{u}_{2}\bm{H}_{12}\bar{\bm{v}}_{1}|^2}
\end{bmatrix},\\
\bm{m} &=&
\begin{bmatrix}
\abs{\bm{u}_{1}\bm{H}_{21}\bar{\bm{v}}_{2}}^{-2} &
\abs{\bm{u}_{2}\bm{H}_{12}\bar{\bm{v}}_{1}}^{-2}
\end{bmatrix}^\text{T}.
\end{eqnarray}
\end{subequations}
Taking advantage of the \ac{PF} theorem \cite{S-U-Pillai2005SPM} and the fact that $\bm{\Gamma}\bm{M}$ is a non negative matrix, the optimal \ac{TX} power vector can be computed in closed form as
\begin{eqnarray}
\label{eqn:popt}
\bm{p}^{*}=\sigma^2\left({\bf{I}}- \bm{\Gamma}\bm{M}\right)^{-1} \bm{\Gamma} \bm{m}.
\end{eqnarray}
\subsection{Optimal \ac{BF} Design for SINR Maximization}
\label{sect:beamdesign}
With possession of a closed-form optimal solution to the \ac{TX} power vector $\bm{p}$ as per equation \eqref{eqn:popt}, as well as a given analog cancellation matrix $\bm{C}_{k}$ obtained for example as discussed in \cite{Alexandropoulos2017, IimoriSPAWC2018}, we seek optimal \ac{BF} designs for $\bm{v}_{k}$ and $\bm{u}_{k}\: \forall k$, such that the average SINR at each node is maximized, while minimizing the effect of the \ac{SI}.
Taking into account the fact that the role of TX-RX beamformers is to minimize the effect of \ac{SI} while maximizing the downlink rate, we consider the \ac{MRT} TX beamformer with perfect \ac{CSI} known at the nodes, such that the instantaneous \ac{SINR} at each node is maximized under the assumption that the \ac{SI} power level can be significantly reduced after processing by the proposed optimal \ac{RX} combiner.
\vspace{1ex}
\subsubsection[]{Design of \ac{RX} Combiner $\bm{u}_{k}\forall k$}\quad\\[-2ex]
The role of the \ac{RX} combining vector $\bm{u}_{k}$ at the $k$-th node is to maximize the power of the signal from the $\ell$-th node, while supressing the interference-plus-noise signal.
In other words, the \ac{RX} \ac{BF} vector $\bm{u}_{k}$ must be designed so as to maximize the ratio between the power of the intended signal and that of interference-plus-noise term of equation \eqref{eq:estimated_k}, which can be mathematically expressed as
\begin{eqnarray}
\label{eq:gk_OP}
\max_{\bm{u}_{k} \atop \norm{\bm{u}_{k}}^2=1} \frac{\bm{u}_{k}\overbrace{\bm{H}_{\ell k}\bm{v}_{\ell}\bm{v}^{\rm H}_{\ell}\bm{H}^{\rm H}_{\ell k}}^{\triangleq \bm{Q}_{\bm{u}_{k}}}\bm{u}^{\rm H}_{k}}{\bm{u}_{k}\underbrace{\left(\tilde{\bm{H}}_{k k}\bm{v}_{k}\bm{v}^{\rm H}_{k}\tilde{\bm{H}}^{\rm H}_{k k} + \sigma^2{\bf{I}}\right)}_{^{\triangleq \bm{W}_{\bm{u}_{k}}}}\bm{u}^{\rm H}_{k}},
\end{eqnarray}
which holds a generalized \ac{RQ} structure, such that the optimal solution to $\bm{u}_{k}$ is obtained by \cite{Prieto2003ICASSP}
\begin{eqnarray}
\label{eq:ukopt}
\bm{u}^{*}_{k} = {\rm eigv}_{\rm max}\left(\bm{W}^{-1}_{\bm{u}_{k}}\bm{Q}_{\bm{u}_{k}}\right)^{\rm H}.
\end{eqnarray}
\begin{figure}[H]
\center
\vspace{-1ex}
\includegraphics[width=\columnwidth]{TransPowervsTargetRate.eps}
\vspace{-3.5ex}
\caption{Proposed and conventional preceding methods TX power comparison for different target rates.}
\label{fig:TransPower}
\vspace{-3ex}
\end{figure}
\begin{algorithm}[h]
\caption[]{Alt. \ac{TX} Power Min. with SINR Guarantees.}
\label{alg:main}
\begin{algorithmic}[1]
\State \textbf{Input}:$P_{k},\bm{H}_{kk},\bm{H}_{\ell k}$, \:$\bm{C}_{k} \forall k$ given by \cite{Alexandropoulos2017}.
\State Set $P_{k}=P_{\rm max} \forall k\in\{1,2\}$ and make arbitrary unit-norm vectors as initial \ac{RX} \ac{BF} vectors $\bm{u}_{k} \forall k$.
\Repeat
\State Compute $\bar{\bm{v}}_{k}\forall k$ from equation \eqref{eqn:MRTTX}.
\State Compute $\bm{u}^{*}_{k}\forall k$ from equation \eqref{eq:ukopt}.
\State Compute $\bm{p}^{*}$ from equation \eqref{eqn:popt}.
\Until{{\rm\bf convergence or reach maximum iterations.}}
\end{algorithmic}
\end{algorithm}
\setlength{\textfloatsep}{-2pt}
\vspace{-2ex}
\subsubsection{Design of TX Precoder $\bm{v}_{k}\forall k$}\quad\\[-2.3ex]
Assuming that the strong \ac{SI} caused by the leakage of own \ac{TX} signals due to the close proximity of \ac{TX} and \ac{RX} antennas can be sufficiently suppressed by the \ac{RX} combining vector $\bm{u}_{k}$, the role of the TX precoder $\bm{v}_{k}$ is only to direct the \ac{TX} beams so as to maximize the downlink rate.
For this purpose, it suffices to apply a simple \ac{MRT} TX precoder, namely
\vspace{-1ex}
\begin{eqnarray}
\label{eqn:MRTTX}
\vspace{-1ex}
\bar{\bm{v}}_{k} = \frac{\bm{H}^{\rm H}_{kl}\bm{u}^{\rm H}_{\ell}}{\left|\left|\bm{H}_{kl}\bm{u}_{\ell}\right|\right|}.
\vspace{-1ex}
\end{eqnarray}
Taking into account all the steps described in this section, the proposed algorithm for the optimization of \ac{TX} powers, as well as \ac{TX} and \ac{RX} \ac{BF} vectors can be compactly described by the pseudo-code offered in Algorithm \ref{alg:main}.
\vspace{-1ex}
\section{Simulation results}
\label{SimulationResults}
\vspace{-0.5ex}
In this section, we evaluate the proposed iterative algorithm in terms of consumed \ac{TX} power for different required SINR constraints via software simulation.
The downlink communication channels $\bm{H}_{12}$ and $\bm{H}_{21}$ are assumed to be block Rayleigh fading with $110$dB path loss, while the \ac{SI} channels $\bm{H}_{11},\bm{H}_{22}$ are assumed to be block Ricean with path loss of $40$dB and $K$-factor $35$dB \cite{Alexandropoulos2017,duarte2012experiment}.
Each node is assumed to be equipped with $4$ \ac{TX} and receive antennas, i$.$e$.$, $M=N=4$, with a noise floor of $-110$dBm, and the number of analog cancellation taps $N_{\rm tap}$ is set to $8$, which corresponds to $50$\% reduction in the number of elements in the analog cancellation matrix $\bm{C}_{k}$ compared to \cite{bharadia2013fullduplex, Everett2011Asilomar}.
\begin{figure}[H]
\center
\vspace{1ex}
\includegraphics[width=\columnwidth]{SIPowervsTargetRate.eps}
\vspace{-4ex}
\caption{Residual SI power comparisons of the proposed and conventional methods for different target rates after \ac{RX} \ac{BF}.}
\label{fig:ResidualSI}
\vspace{1ex}
\includegraphics[width=\columnwidth]{OutTransPowervsAvailPower.eps}
\vspace{-4ex}
\caption{\ac{TX} power outage probability for different available \ac{TX} powers with the fixed target rate $R_{k}=R_{\ell}=8$ [bps/Hz].}
\label{fig:OutPower}
\vspace{-2ex}
\end{figure}
\noindent For the modeling practical situations, the multi-tap analog canceller is assumed to be subjected to amplitude imperfection uniformly distributed between $-0.01$dB and $0.01$dB and phase noise uniformly distributed between $-0.065^{\circ}$ and $0.065^{\circ}$ \cite{Alexandropoulos2017, Kolodziej2016TWC}.
In all figures that follow, we compare the proposed \ac{TX} power minimization method in Algorithm \ref{alg:main} with $100$ maximum iterations against the conventional \ac{ZF} TX precoder, in which the proposed \ac{PF} power optimization is applied.
In addition, by noticing that the considered bi-directional FD MIMO corresponds to a special form of the MIMO X channel, we deploy relevant algorithms \cite{GeorgeSP2013, Alexandropoulos2016_CB, OmidTWC2018} targeting at TX-RX \ac{BF} design yielding sum rate maximization.
Particularly, our considered system is a MIMO X channel having $\bm{H}_{21}$ and $\bm{H}_{12}$ as the intended channels and $\tilde{\bm{H}}_{11}$ and $\tilde{\bm{H}}_{22}$ as the interference channels, having possibly larger powers than the intended ones.
\begin{table}[t!]
\caption{Run time comparisons for different methods.}
\vspace{-3ex}
\begin{center}
\begin{tabular}{|c|c|c|c|c|} \hline
Methods & ZF & RQ-RQ & Rec & Proposed \\ \hline
Average run time [s] & 0.0025 & 0.0028 & 0.1309 & 0.0022 \\\hline
\end{tabular}
\vspace{-1ex}
\label{Tab:Runtime}
\end{center}
\end{table}
First, average \ac{TX} power comparisons of the proposed algorithm for different target rates $\log_{2}(1+{\Gamma_{k}}) \forall k$ is shown in Figure \ref{fig:TransPower}, where \ac{ZF}-\ac{RQ}, \ac{RQ}-\ac{RQ} \cite{OmidTWC2018} and the Reconfigurable sum rate maximization algorithm \cite{GeorgeSP2013} are employed as a benchmark.
In order to fairly compare those algorithms, we adopt an alternating recalculation between \ac{TX}-\ac{RX} \ac{BF} for each algorithms until convergence or maximum number of iterations reached.
It is shown in Figure \ref{fig:TransPower} that the proposed method can decrease the \ac{TX} power by about $-4.5$dB compared to the conventional \ac{ZF}-\ac{RQ}, \ac{RQ}-\ac{RQ} methods and about $-0.8$dB compared against the Reconfigurable method.
Secondly, Figure \ref{fig:ResidualSI} outlines that the interference cancellation performance in terms of residual \ac{SI} power levels after processing by the \ac{RX} \ac{BF} are compared for the different \ac{TX}-\ac{RX} \ac{BF} schemes.
From Figure \ref{fig:TransPower} and \ref{fig:ResidualSI}, one can notice that although the \ac{ZF} method can perfectly suppress the effect of \ac{SI} at the \ac{RX} baseband, the proposed method can outperform the other schemes due to the fact that not only the residual \ac{SI} level of the proposed method is suppressed below the noise floor level but also it aims at maximizing the data rate performance.
In other words, the other methods devote too much available \ac{DoF}s to suppressing \ac{SI} power level at the \ac{RX} baseband.
Thirdly, the \ac{TX} power outage probability of the proposed method for different available \ac{TX} powers $P_{\rm max}$ with target data rate fixed at $R_{k}=R_{\ell}=8$ [bps/Hz] is compared with the outage performance of the other conventional methods in Figure \ref{fig:OutPower}, where we define the \ac{TX} power outage probability as ${\rm Pr}\left({\rm min}\left(P_{k}, P_{\ell}\right)>P_\mathrm{max}\right)$.
Lastly, the average run time comparisons until the convergence for each different algorithms are depicted in TABLE \ref{Tab:Runtime}, where we take an average from $500$ channel realizations.
From Figure \ref{fig:TransPower}, \ref{fig:ResidualSI} and \ref{fig:OutPower} and Table \ref{Tab:Runtime}, it can be observed that the proposed method can has much fast convergence rate compared with the Reconfigurable method and outperform the conventional \ac{ZF} and \ac{RQ} \ac{TX} \ac{BF} methods in terms of the \ac{TX} power outage probability performance.
\vspace{-0.4ex}
\section{Conclusion}
\label{Conclusion}
\vspace{-0.6ex}
In this paper, we considered bi-directional \ac{FD} \ac{MIMO} communications systems with limited number of analog canceller taps and designed TX-RX \ac{BF} vectors with the goal to minimize \ac{TX} power under \ac{SINR} constraints.
The proposed \ac{TX} power minimization \ac{BF} design was investigated in terms of system performance and complexity, and the \ac{PF} \ac{TX} power minimization approach was jointly offered with the proposed beamformers.
Simulation results demonstrate the capability of our proposed algorithm to suppress the \ac{SI} level to below $-110$dB which is the typical noise floor for wireless communications, while maximizing the downlink rate, and consequently, it minimizes the average \ac{TX} power for different target data rate.
\vspace{-0.4ex}
\section{Acknowledgement}
\vspace{-0.6ex}
Parts of this work were supported by JSPS KAKENHI, Grant Number JP17K06448, Denkitsushin Hukyu Foundation, and EU Project HIGHTS with grant number 636537.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The adiabatic limit of anti-self-dual connections on 4-manifolds has been extensively studied by many authors, with various interesting applications to problems in gauge theory, geometry, and physics. In \cite{DS,DS2}, Dostoglou and Salamon proved the Atiyah-Floer conjecture (see \cite{At2}) by showing that the adiabatic limits of self-dual connections on the product of $\mathbb{R}$ and the mapping cylinder of a principal $SO(3)$-bundle over a compact Riemann surface of higher genus (greater than one) produce holomorphic curves in the moduli space of flat connections on the $SO(3)$-bundle. Later, the behavior of anti-self-dual $SU(n)$-connections along the adiabatic degenerations of the product of two compact Riemann surfaces of higher genus was studied in \cite{Ch1} and \cite{Nis2} respectively, which gave mathematical rigorous proofs of the reduction from the 4-dimensional Yang-Mills theory to 2-dimensional sigma models discovered by physicists (cf. \cite{BJSV}).
Based on previous works of gauge theory on higher dimensional manifolds \cite{DT,T}, \cite{Ch2} generalized the 4-dimensional case to complex anti-self-dual connections on products of Calabi-Yau surfaces. The Atiyah-Floer conjecture was studied in
\cite{Du} for principal $PU(n)$-bundles.
Another motivation for the study of adiabatic limits of anti-self-dual connections arises in the context of the mirror symmetry.
In \cite{SYZ}, Strominger, Yau and Zaslow proposed a conjecture, called the SYZ conjecture, for constructing mirror Calabi-Yau manifolds via dual special Lagrangian fibrations. Gross, Wilson, Kontsevich, Soibelman and Todorov \cite{GW,KS,KS2} proposed an alternative version of the SYZ conjecture by using the collapsing of Ricci-flat K\"{a}hler metrics.
Motivated by the study of homological mirror symmetry,
a gauge theory analogue of the collapsing of Ricci-flat K\"{a}hler metrics was conjectured by Fukaya (Conjecture 5.5 in \cite{Fuk}), which relates the adiabatic limits of anti-self-dual connections on Calabi-Yau manifolds to special Lagrangian cycles on the mirror Calabi-Yau manifolds. This conjecture was studied in the preprints \cite{Fu,Nis1} for Hermitian-Yang-Mills connections on 2-dimensional complex torus, and in \cite{Ch3} for the case of Hermitian-Yang-Mills connections on higher dimensional semi-flat Calabi-Yau manifolds.
The present paper proves a version of Fukaya's conjecture for anti-self-dual connections on elliptically fibered K3 surfaces.
Let $M$ be a projective elliptically fibered $K3$ surface, $
f:M\rightarrow N\cong \mathbb{CP}^1,$ admitting a section $\sigma:N\rightarrow M$. Let $\alpha$ be an ample class on $M$, $\alpha_{t} =t \alpha + f^*c_1(\mathcal{O}_{\mathbb{CP}^1}(1))$, $t\in (0,1]$, and let $\omega_t \in \alpha_{t} $ be the unique Ricci-flat K\"{a}hler-Einstein metric in this class (from \cite{Ya}). We denote by $ \mathrm{g}_t$ the corresponding Riemannian metric of $\omega_t$, which is a HyperK\"{a}hler metric.
The limit behavior of $\omega_t$ as $t\rightarrow 0$ was studied by Gross and Wilson in \cite{GW}, for K3 surfaces with only type $I_1$ singular fibers. This was generalized to any elliptically fibered $K3$ surface in \cite{To1,GTZ,GTZ2}. More precisely, if $N_0\subseteq N$ denotes the complement of the discriminant locus of $f$, i.e. for any $w\in N_0$ the fiber $M_w=f^{-1}(w)$ is a smooth elliptic curve, then it is proved in \cite{GTZ} that $\omega_t$ converges to $f^*\omega$ in the locally $C^\infty$-sense on $M_{N_0}=f^{-1}(N_0)$, where $\omega$ is a K\"{a}hler metric on $N_0$ with Ricci curvature ${\rm Ric}(\omega)=\omega_{WP}$ (obtained previously by \cite{ST,To1}), and $\omega_{WP}$ denotes the Weil-Petersson metric of the fibers of $f$. Furthermore,
$(M, \omega_t)$ converges to a compact metric space $Y$ homeomorphic to $N$ in the Gromov-Hausdorff sense \cite{GTZ2}.
Assume that $f: M \rightarrow N$ has only singular fibers of Kodaira type $I_1$ and type $II$.
Let $P$ be a principal $SU(n)$-bundle on $M$, and
$(\mathcal{V}, H)$ be the smooth Hermitian vector bundle of rank $n$ obtained by the twisted product, i.e. $\mathcal{V}\cong P\times_{\rho}\mathbb{C}^n$ where $\rho$ is the standard $SU(n)$ representation on $\mathbb{C}^n$.
Assume that there is a family of anti-self-dual connections $\Xi_t$ on $P$
with respect to $\mathrm{g}_t$, for $t\in (0,1]$. This is equivalent to the curvature $F_{\Xi_t}$ satisfying $$F_{\Xi_t}\wedge \omega_t =0, \ \ \ {\rm and} \ \ F_{\Xi_t}\wedge \Omega =0, $$ where $\Omega$ is a holomorphic symplectic form on $M$. For each $t\in (0,1]$, $\Xi_t$ induces a holomorphic structure on $\mathcal{V}$, and we denote the resulting holomorphic bundle of rank $n$ as $V_t$.
Under some non-degeneracy assumptions on the behavior of $V_t$, the main result of this paper, Theorem \ref{thm-main},
asserts that for any sequence $t_k \rightarrow 0$, there exists a Zariski open subset $N^o \subset N_0$ such that $u_k(\Xi_{t_k}) $ converges subsequentially to $\Xi_0$ in the locally $C^{0,\alpha}$-sense on $M_{N^o}$, for some sequence of unitary gauge transformations $u_k$ on $P$. Furthermore, the restriction of the limit $\Xi_0 $ to any fiber is unitary gauge equivalent to a smooth flat $SU(n)$-connection induced by a holomorphic curve in $M$, which can be regarded as a multi-section of $f$. Furthermore, $\Xi_0 $ is the Fourier-Mukai transform of a certain flat $U(1)$-connection on the multi-section.
We refer the reader to Theorem \ref{thm-main} and Theorem \ref{thm-main2} for more precise statements. By performing the HyperK\"{a}hler rotation, we can use this result to show a version of Fukaya's conjecture, relating the connections $\Xi_{t_k}$ to a converging family of special Lagrangian multi-sections in the mirror HyperK\"ahler structure.
In comparison to previous results on the adiabatic limits of anti-self-dual connections, including, for example \cite{DS,Ch1,Nis2,Fuk2}, one essential difficulty we encounter is that the moduli space $\mathfrak{M}_E(n)$ of flat $SU(n)$-connections on a smooth elliptic curve is not smooth, and actually, the whole $\mathfrak{M}_E(n)$ is degenerated, i.e. there is no smooth point (cf. \cite{Nis11}). Specifically, since every flat connection is gauge equivalent to a reducible connection, Poincar\'e type inequalities may not follow, creating immense analytic difficulties. The same issue also appears for the case of $T^4=\mathbb{C}^2/\mathbb{Z}^4$ as in \cite{Fu,Nis1}. To overcome this, we take a totally different approach from \cite{Fu,Nis1}, which is inspired by the study of collapsing of Einstein 4-manifolds \cite{And,CT}. In addition we adapt some of the arguments from \cite{DS,DS2}, as suggested in \cite{Fuk}.
Fortunately, in the literature there is a very satisfactory theory about the moduli spaces of semi-stable holomorphic bundles of rank $n$ on elliptic curves in algebraic geometry. In the proof of Theorem \ref{thm-main}, we utilize the well understood results of holomorphic bundles on elliptic fibered surfaces in \cite{FMW,FMW0,Fr1}, as opposed to the pseudo-holomorphic curve theory in symplectic geometry used in \cite{DS,Nis1}. Additionally, in the course of our analysis, we obtain a Poincar\'{e} type inequality for the curvatures of $SU(n)$-connections on smooth elliptic curves, which relies on the earlier work of the first two named authors (cf. \cite{DJ}). This enables us to generalize certain arguments of \cite{DS} to the present case. Finally, the small energy estimates for sufficiently collapsed Einstein 4-manifolds developed in \cite{And} can be adapted to the case of Yang-Mills connections on collapsed 4-manifolds, which is used to finish the proof of the main theorem.
Here we outline the paper briefly. Section 2 reviews the background notions, and preliminary results, which are needed for the main theorem. We recall the standard background on gauge theory in Section 2.1, and the theory of holomorphic vectors bundles on elliptic curves in Section 2.2. Section 2.3 reviews the previous work about the gauge fixing on elliptic curves by the first two named authors, which is one essential ingredient in the proof of the main result of the present paper. Section 2.4 recalls the work of Friedman-Morgan-Witten \cite{FMW,FMW0}, where the relationship between holomorphic bundles and spectral covers on elliptic surfaces is established. This work is the algebro-geometric input needed to overcome the difficulty of non-smoothness of the moduli spaces of flat connections. In Section 2.5, we setup some notations for the collapsing of Ricci-flat K\"{a}hler Einstein metrics on K3 surfaces, and leave more detailed discussions to the Appendix. Section 2.6 reviews the notion of Fourier-Mukai transform. We adapt the small energy estimates for sufficiently collapsed Einstein 4-manifolds by Anderson \cite{And} to the present case in Section 2.7.
Section 3 is devoted to the main theorems of this paper. We state the main theorems, and in Section 3.1, we apply the main theorems to the SYZ mirror symmetry for K3 surfaces, which proves a version of Fukaya's conjecture in \cite{Fuk}. Section 4 contains the proof of Theorem \ref{thm-main} assuming some important a priori estimates, which are established in the sections that follow. Section \ref{Poincare} contains the key analytic result of the paper, namely the Poincar\'{e} type inequality mentioned above. In Section 6, we obtain a $C^0$-bound for curvature under the assumption of a certain decay rate of curvatures as the fibers collapse. Section 7 studies the relationship between the energy of curvature and the spectral covers. In Section 8, we use a blowup argument to prove the desired curvature decay rate, thereby completing the proof of Theorem \ref{thm-main}. Section 9 proves Theorem \ref{thm-main2}.
Finally, the appendix has some results of independent interest, where we study the collapsing rate of Ricci-flat K\"{a}hler-Einstein metrics on general Abelian fibered Calabi-Yau manifolds. Here we improve on the previous results of \cite{GTZ,GTZ2,TZ}.
\\
\noindent {\bf Acknowledgements:} We would like to thank Mark Haskins for introducing the authors to the question, and some valuable comments. The work was initiated when the second and the third named author attended the
First Annual Meeting 2017 of the Simons Collaboration on Special Holonomy in Geometry, Analysis and Physics. We thank the Simons Foundation and the organisers of the meeting for providing this opportunity.
We also thank Simon Donaldson, Mark Gross, Valentino Tosatti, Yuuji Tanaka, and Michael Singer for some discussions.
\section{Preliminaries}
In this section, we review the various notions, and preliminary results, which are needed for the main theorem. Although there is quite a bit of background to cover, we find it necessary to provide all the important details before we can state our results.
Let $M$ be a projective, elliptically fibered $K3$ surface. Denote the fibration by $
f:M\rightarrow N\cong \mathbb{CP}^1.$ Assume $f$ admits a section $\sigma:N\rightarrow M$, and furthermore assume $f$ has only singular fibers of Kodaira type $I_1$ and type $II$. Let $I$ denote the holomorphic structure on $M$ for which $f$ is holomorphic. We denote by $S_N$ the discriminant locus $f$, and $N_0=N\backslash S_N$ the regular locus. The preimage of the regular locus is denoted by $M_0:=f^{-1}(N_0)$. For any point $w\in N$, the fiber over this point is written $M_w:=f^{-1}(w)$. Additionally, for any subset $U\subset N$, we use the notation $M_U:=f^{-1}(U)$.
Let $P$ be a principal $SU(n)$-bundle on $M$, and $\mathcal{V} $ be the smooth vector bundle of rank $n$ equipped with an Hermitian metric $H$ induced by $P$, i.e. $\mathcal{V}=P\times_{\rho}\mathbb{C}^n$, where $\rho$ is the standard unitary representation of $SU(n)$ on $\mathbb{C}^n$. Note that first Chern class of $\mathcal{V}$ vanishes, i.e. $c_1(\mathcal{V})=0$.
For computing norms it is convenient to use a fixed K\"ahler form $\omega$ on $M$, which lies in a fixed K\"ahler class $\alpha$. Unless otherwise specified, all norms are computed with respect to $\omega$ and $H$. We let $\langle\cdot, \cdot\rangle_w$ denote the inner product of the space of forms induced by $\omega|_{M_w}$ on the fiber $M_w$, and $\|\cdot\|_w $ the respective $L^2$-norm on $M_w$.
Throughout the paper, we let $C$ denote constants, which only depend on fixed background data, whose value may change from line to line. The constants may depend on a compact or open sets contained in $N$, and this dependence is either explicitly stated, or clear from context.
\subsection{Anti-self-dual connections} We begin by recalling the standard background on anti-self-dual connections, and readers are referred to texts \cite{AtB,DK,FU,Kob} for details.
Given the definition of $P$ above, let $\Xi$ be a connection on $P$, or an $SU(n)$-connection of $\mathcal{V} $. If the curvature $F_{\Xi}$ satisfies
$$ F_{\Xi}^{0,2}=0, \ \ \ {\rm or \ \ equivalently} \ \ F_{\Xi}=F_{\Xi}^{1,1}, $$
then $\Xi$ induces a holomorphic structure on $\mathcal{V}$. We denote the resulting holomorphic bundle as $V_{\Xi}$, and $\bar{\partial}_{\Xi}$ the corresponding Cauchy-Riemann operator. Specifically, we can write the covariant derivative $d_{\Xi}:C^\infty (\wedge^q T^*M \otimes\mathcal{V})\rightarrow C^\infty (\wedge^{q+1} T^*M \otimes\mathcal{V}) $ as $d_{\Xi}= \partial_{\Xi}+\bar{\partial}_{\Xi}$, and the Cauchy-Riemann operator is the $(0,1)$-component.
By construction $\Xi$ is the unique Chern connection induced by $H$ and $\bar\partial_{\Xi}$.
Let $\mathcal{A}^{1,1}$ be the space of all unitary connections with vanishing $(0,2)$-component of curvatures on $P$, so for any $\Xi \in \mathcal{A}^{1,1}$, we have $F_{\Xi}^{0,2}=0$. If $\mathcal{G}$ denotes the unitary gauge group, i.e. the space of unitary automorphisms of $\mathcal{V}$ covering the identity on $M$, then $\mathcal{G}$ acts on $\mathcal{A}^{1,1} $ by $$u(\Xi)=\Xi+ u^{-1}(d_\Xi u),$$ for $u\in \mathcal{G}$ and $\Xi \in \mathcal{A}^{1,1}$. The $\mathcal{G}$-action extends to an action of
the complex gauge group $\mathcal{G}_{\mathbb{C}}$, which consists all automorphisms of $\mathcal{V}$ covering the identity on $M$, on $\mathcal{A}^{1,1}$ by $$ g(\Xi)= \Xi +g^{-1} (\bar{\partial}_{\Xi}g)-(g^{-1} (\bar{\partial}_{\Xi}g))^*, $$ for $g \in \mathcal{G}_{\mathbb{C}}$, where $(\cdot)^*$ denotes the conjugate transpose.
Any two connections $\Xi_1$ and $\Xi_2 \in \mathcal{A}^{1,1}$ induce isomorphic holomorphic structures on $\mathcal{V}$ if and only if $\Xi_1=g(\Xi_2)$ for a certain $g\in \mathcal{G}_{\mathbb{C}}$. Therefore the quotient space $\mathcal{A}^{1,1}/ \mathcal{G}_{\mathbb{C}}$ parameterizes the holomorphic structures on $\mathcal{V}$.
Note that if $g \in \mathcal{G}_{\mathbb{C}}$ is an Hermitian gauge, i.e. $g=g^*$, then for any $\Xi\in \mathcal{A}^{1,1}$, the curvature transforms via
\bea F_{ g(\Xi)}& = & F_{\Xi}+ \partial_{\Xi}(g^{-1}(\bar{\partial}_{\Xi}g)) - \bar{\partial}_{\Xi}( (\partial_{\Xi}g)g^{-1}) \nonumber\\ & & + \partial_{\Xi}g g^{-2} \bar{\partial}_{\Xi}g - g^{-1}\bar{\partial}_{\Xi}g \partial_{\Xi}g g^{-1}. \nonumber
\eea
The transformation of $\Xi$ to $g(\Xi)$ by a Hermitian gauge $g$ is equivalent to fixing the holomorphic structure on a bundle $V$, and then changing the Hermitian metric (see \cite{Don1} for details).
Given a K\"{a}hler class $\alpha$ on $M$, choose a K\"ahler form $\omega\in \alpha$, and let $\mathrm{g}$ be the corresponding Riemannian metric.
\begin{defn}
An $SU(n)$-connection $\Xi$ is called anti-self-dual with respect to the K\"{a}hler metric $\omega$ if $\Xi$ satisfies the equation \begin{equation}\label{asd} \star_{ \mathrm{g}} F_{\Xi}=- F_{\Xi} , \end{equation} where $\star_{\mathrm{g}} $ denotes the Hodge star operator of $\mathrm{g}$.
\end{defn}
For any anti-self-dual connection, Chern-Weil theory gives
\begin{equation}\label{cw0} \int_M |F_{\Xi}|_{\omega}^2 \omega^2=- \int_M {\rm tr}(F_{\Xi}\wedge F_{\Xi})= 8 \pi^2 c_2(\mathcal{V}). \end{equation} Furthermore, anti-self-dual connections are absolute minima of the Yang-Mills functional on $P$, and thus satisfy the Yang-Mills equations $$d_{\Xi}F_{\Xi}=0, \ \ \ {\rm and} \ \ d_{\Xi}^* F_{\Xi}=0.$$ This implies the following Weitzenb\"{o}ck formula for the curvature of $\Xi$
\begin{equation}\label{wei} 0=\Delta_{\Xi}F_{\Xi} =\nabla_\Xi^* \nabla_\Xi F_{\Xi}+R_\omega \# F_{\Xi}+ F_{\Xi}\# F_{\Xi}. \end{equation}
Here $R_\omega$ denotes the Riemannian curvature of $\omega$, and $S\#T$ denotes some algebraic bilinear expression involving the tensors $S$ and $T$, where the exact form is not important for the present paper.
In complex dimension 2, a connection $\Xi$ is anti-self-dual if and only if it is Hermitian-Yang-Mills \cite{DK},
which is given by the following set of equations
\begin{equation}\label{HYM0} F_{\Xi}^{1,1}\wedge \omega =0, \qquad \ \ {\rm and} \qquad \ \ F_{\Xi}^{0,2} =0. \end{equation} Thus an anti-self-dual connection $\Xi$ induces a holomorphic structure on $\mathcal{V} $, and we denote the resulting holomorphic vector bundle as $V_\Xi$.
For a given K\"{a}hler class $\alpha$ on $M$, a holomorphic vector bundle $V$ is called $\alpha$-stable (respectively $\alpha$-semi-stable), if for any proper torsion-free coherent subsheaf $\mathcal{F}$, the following inequality holds $$\frac{c_1(\mathcal{F})\cdot \alpha}{{\rm rank}(\mathcal{F}) }< \frac{c_1(V)\cdot \alpha}{{\rm rank}(V) } \ \ ({\rm respectively } \ \ \leq). $$
Fundamental work of Donaldson, Uhlenbeck, and Yau, asserts the equivalence between stability and the existence of Hermitian-Yang-Mills connections (cf. \cite{Don1,UY}). In partcular, we state the following Theorem, restricted to the $SU(n)$ case.
\begin{thm}[Donaldson \cite{Don1}, Uhlenbeck-Yau \cite{UY}]\label{DUY} Let $(\mathcal{V}, H)$ be the smooth Hermitian bundle induced by a principal $SU(n)$-bundle $P$, $\alpha$ be a K\"{a}hler class on $M$, and $\omega\in \alpha$ a K\"{a}hler metric. If the holomorphic bundle $V$ determined by a $\mathcal{G}_{\mathbb{C}}$-orbit $ O$ in $\mathcal{A}^{1,1}$ is $\alpha$-stable, then $O$ contains an anti-self-dual connection (equivalently a Hermitian-Yang-Mills connection). Furthermore, this connection is unique up to unitary gauge transformations. Conversely, if $\Xi$ is an anti-self-dual connection with respect to $\omega$, and the holomorphic bundle $V_\Xi$ induced by $\Xi$ is irreducible, then $V_\Xi$ is $\alpha$-stable.
\end{thm}
Note that if $\omega$ is a Ricci-flat K\"{a}hler-Einstein metric, then the corresponding Riemannian metric $\mathrm{g}$ is a HyperK\"{a}hler metric, and $(\omega, {\rm Re} (\Omega), {\rm Im} (\Omega))$ is a HyperK\"{a}hler triple (cf. \cite{GHJ}), where $\Omega$ is a holomorphic symplectic form such that $$\omega^2={\rm Re} ( \Omega)^2= {\rm Im} (\Omega)^2, \ \ \omega\wedge \Omega=0, \ \ {\rm and} \ \ {\rm Re} ( \Omega)\wedge {\rm Im} (\Omega)=0. $$ Complex structures making $\mathrm{g}$ HyperK\"{a}hler are parameterized by $S^2$, and any anti-self-dual connection $\Xi$ with respect to $ \mathrm{g}$ is also a Hermitian-Yang-Mills connection with respect to any such complex structure. In the HyperK\"{a}hler case, the anti-self-dual equation (\ref{asd}) and the Hermitian-Yang-Mills equation (\ref{HYM0}) are equivalent to the following system
\begin{equation}\label{hyper} F_{\Xi}\wedge \omega =0, \ \ \ {\rm and} \ \ F_{\Xi}\wedge \Omega =0. \end{equation} For the remainder of the paper, we mainly work with the above equations, as they are the most applicable to our setup.
The above equations \eqref{hyper} are given with respect to the complex structure $I$ making $f:M\rightarrow N$ holomorphic. By the HyperK\"{a}hler rotation, we have another complex structure $J$ such that the holomorphic symplectic form $\Omega_J= {\rm Im} (\Omega)+ i \omega $, and the K\"{a}hler form $\omega_J= {\rm Re}( \Omega)$. If $\Xi$ is an anti-self-dual connection with respect to $ \mathrm{g}$, then $\Xi$ also satisfies $ F_{\Xi}\wedge \omega_J =0, $ and $ F_{\Xi}\wedge \Omega_J =0$. Thus
$\Xi$ induces a holomorphic bundle structure on $\mathcal{V}$ with respect to the complex structure $J$, denoted as $V_{\Xi, J}$, and $\Xi$ is a Hermitian-Yang-Mills connection on $V_{\Xi, J}$.
We conclude this section by recalling Uhlenbeck's compactness theorems, which are divided into the cases of weak and strong compactness.
\begin{thm}[Uhlenbeck \cite{U2,Weh2}]\label{Ucompact} Let $K$ be a compact subset of $ M$.
\begin{itemize}
\item[i)]{\rm [Weak compactness]} If $\Xi_k$ is a sequence of unitary connections on $P|_K$ such that $\|F_{\Xi_k}\|_{L^p}\leq C$, for $p>2$, then there exists a sequence of unitary gauge transformations $u_k\in \mathcal{G}^{2,p}$ so that $u_k(\Xi_k)$ converges along a subsequence in $L^p_{1,loc}$ to a $L^p_1$-unitary connection $\Xi_\infty$ on $K$.
\item[ii)]{\rm [Strong compactness]} If we further assume that $\Xi_k$ is anti-self-dual with respect to a Riemannian metric $\mathrm{g}_k$, and $\mathrm{g}_k$ converges smoothly to a smooth Riemannian metric $\mathrm{g}_\infty$ locally on $K$, then $u_k(\Xi_k)$ converges to $\Xi_\infty$ in the locally $C^\infty$-sense, and $\Xi_\infty$ is anti-self-dual with respect to $\mathrm{g}_\infty$. \end{itemize} \end{thm}
\subsection{Gauge theory on elliptic curves}
While working with bundles over $M$, we need several preliminary results dealing with the restriction of a bundle to a fixed elliptic fiber, which we detail here.
Fix a point $w\in N_0$, and consider the fiber $M_w=E$, a smooth elliptic curve with period $\tau$, i.e. $E=\mathbb{C}/{\rm Span}_{\mathbb{Z}}\{1, \tau\}$. Equip $E$ with the flat metric $\omega^F_w:=i{\rm Im}(\tau)^{-1}\,dz\wedge d\bar z$. Let $V$ be a holomorphic vector bundle of rank $n$ with trivial determinant line bundle $\bigwedge^{n} V \cong \mathcal{O}_{E} $, let $\bar\partial$ be the Cauchy-Riemann operator, and fix a Hermitian metric $H$ on $V$. Let $A_{ch}$ be the unique Chern connection determined by the holomorphic structure and the Hermitian metric $H$, i.e. $A_{ch}=( \partial H )H^{-1}$ under a certain local holomorphic trivialization. Recall that $\|\cdot\|_w$ denotes the $L^2$ norm on $E$.
\begin{prop}\label{semistablelemma}
There exists a $\delta>0$, dependent only on $E$ and $V$, so that if $A$ is in the complexified gauge orbit of $A_{ch}$ and satisfies $\|F_A\|_{w} <\delta$, then the holomorphic bundle $V$ is semi-stable. \end{prop}
\begin{proof}
This proposition follows from the fact, proven by R$\mathring{\rm a}$de, that the critical values of the Yang-Mills functional (the $L^2$ norm of the curvature) are discrete, and that in real dimension $2$ and $3$ the Yang-Mills flow converges in $L^2_1$ \cite{Rade}. If $A$ satisfies $\|F_A\|_{w} <\delta$ for $\delta$ sufficiently small, then the Yang-Mills flow starting at $A$ must converge to a flat connection $A_0$, by discreteness of critical values. Thus $\|F_{A(t)}\|_{w}\rightarrow0$, where $A(t)$ denotes the flow of connections. Furthermore, the Yang-Mills flow preserves the complex gauge equivalence class of $A$, so $A(t)$ all define isomorphic holomorphic structures on $V$. As a result, $V$ admits an approximate Hermitian-Einstein structure, and is semi-stable \cite{Kob}.
\end{proof}
Although the Yang-Mills flow preserves the complex gauge equivalence class of $A$, it is not immediately clear whether the limiting flat connection $A_0$ is contained in the complexified gauge orbit, or only strictly in the closure. To better understand this, we turn to Atiyah's classification of semi-stable bundles on an elliptic curve.
Let $0\in E$ the identity of the group law. Denote the trivial line bundle by $\cO_E $, and given a point $q\in E$, let $\mathcal{O}_E(q-0)$ be the line bundle associated to the divisor $q-0$. Define $\cI_r$ inductively, with $\cI_1=\cO_E$ and $\cI_r$ the unique nontrivial extension of $\cI_{r-1}$ by $\cO_E$.
\begin{thm}[Atiyah \cite{At}]
Any semi-stable bundle $V$ over $E$ with trivial determinant bundle is isomorphic to a direct sum of bundles of the form $\mathcal{O}_E(q-0)\otimes\cI_r$, i.e. $$ V\cong \bigoplus_{j=1}^{\ell}\mathcal{O}_E(q_j-0)\otimes\cI_{r_j}. $$
\end{thm}
\begin{defn}
A semi-stable bundle $V$ is called regular if it is of the form $V\cong \bigoplus\limits_{j=1}^{\ell}\mathcal{O}_E(q_j-0)\otimes\cI_{r_j}$ with $q_j \neq q_i$ for any $j\neq i$.
\end{defn}
Now, in our setting one (and only one) of two things can happen. Either $V$ is isomorphic a direct sum of line bundles $V=\oplus \mathcal{O}_E(q-0)$, and the limiting flat connection $A_0$ is in the complex gauge orbit of $A$, or $V$ is isomorphic a direct sum of bundles of the form $\mathcal{O}_E(q-0)\otimes\cI_r$, with at least one $r>1$. In the latter case, $\mathcal{O}_E(q-0)\otimes\cI_r$ is strictly semi-stable, since $\mathcal{O}_E(q-0)\subset \mathcal{O}_E(q-0)\otimes\cI_r$ has degree zero but $\mathcal{O}_E(q-0)\otimes\cI_r$ does not split holomorphically. As a result $V$ does not admit a flat connection, and so $A$ is not complex gauge equivalent to $A_0$.
Note that if $V\cong \bigoplus\limits_{j=1}^{\ell}\mathcal{O}_E(q_j-0)\otimes\cI_{r_j}$, then $V$ is {\it S-equivalent} to the flat bundle $\bigoplus\limits_{j=1}^{\ell}\mathcal{O}_E(q_j-0)^{\oplus r_j}$ (see \cite{Fr1} for the precise definition of S-equivalence). Every S-equivalence class corresponds to a divisor $\sum\limits_{j=1}^{\ell} r_j q_j $ in the complete linear system $| n 0|$. Conversely, any divisor $\sum\limits_{j=1}^{\ell} r_j q_j \in | n 0|$ on $E$ induces an S-equivalence class of semi-stable bundles with trivial determinant, which contains $\bigoplus\limits_{j=1}^{\ell}\mathcal{O}_E(q_j-0)^{\oplus r_j}$. Therefore, the moduli space of S-equivalence classes of semi-stable bundles with trivial determinant is given by the complete linear system $| n 0| \cong \mathbb{CP}^{n-1}$.
Furthermore, the moduli space of flat line bundles on $E$ is the dual torus $\check{E}\cong H^{0,1}(E)/ H^1(E, \mathbb{Z})$, and we identify $E$ and $\check{E}$ by $q \mapsto \mathcal{O}_E(q-0)$. Another way to state this is that a point $q\in E$ corresponds to a flat connection $\pi ({\rm Im} \tau)^{-1} ( qd\bar{z}-\bar{q}dz)$ on the trivial Hermitian bundle $E\times \mathbb{C}$. Therefore the flat bundle structure of $\bigoplus\limits_{j=1}^{n}\mathcal{O}_E(q_j-0)$ is given by the flat connection
\be
\label{flatconnection}
A_0= \pi ({\rm Im} \tau)^{-1} ({\rm diag}\{q_1, \cdots, q_n\}d\bar{z}-{\rm diag}\{\bar{q}_1, \cdots, \bar{q}_n\}dz),
\ee
where $\sum\limits_{j=1}^{n}q_j \in | n 0|$. Note that the above connection has this form in a global unitary frame for $V$. Let $\mathfrak{M}_E(n) $ denote the moduli space of flat $SU(n)$ connections on $V$, which is naturally identified with $| n 0|$, the moduli space of S-equivalence classes of semi-stable bundles with trivial determinant.
We note that from the perspective of algebraic geometry, the linear system $| n 0|$ is a well behaved object. On the other hand, from the perspective of symplectic geometry, the moduli space $\mathfrak{M}_E(n) $ is quite complicated. In particular, any flat $SU(n)$-connection on $E$ is degenerate, the virtual dimension of $\mathfrak{M}_E(n) $ is zero, and the whole space $\mathfrak{M}_E(n) $ is regarded as singular, i.e. there is no smooth point (cf. \cite{Nis1,Nis2}). If we let $\mathcal{A}$ denote the space of all unitary connections on the trivial bundle on $E$, and $\mathcal{G}$ the unitary gauge group, then following Atiyah-Bott \cite{AtB}, one can construct $\mathfrak{M}_E(n) $ as the symplectic reduction $\mathfrak{M}_E(n)=\{A\in \mathcal{A}|F_A=0\}/\mathcal{G} $. Using this construction $\mathfrak{M}_E(n) $ is in the singular locus of $\mathcal{A}/\mathcal{G}$. Such ill behavior of $\mathfrak{M}_E(n)$ prevents us to generalize the arguments in \cite{Ch1,DS,Fuk2,Nis11} directly, where the moduli space of flat connections on Riemann surfaces of higher genus are considered. Instead we follow an algebro-geometric approach combined with estimates for the above non-linear partial differential equations.
\subsection{Gauge fixing} In this section we continue to work on a single elliptic curve $(E,\omega)$. Let $V$ be a regular, semi-stable, holomorphic vector bundle of rank $n$ which admits a flat connection $A_0$, equipped with a Hermitian metric $H$. Suppose $A$ is another connection in the complex gauge orbit of $A_0$, i.e. $A=g(A_0)$ for some $g\in\mathcal{G}_{\mathbb C}$. It will be important for us to know under what conditions we have control over the $C^0$ norm of $g$. Since the action of a fixed unitary gauge transformation will not affect this norm, without loss of generality we assume that $A=e^s(A_0)$ for a trace free Hermitian endomorphism $s$.
In general it is not reasonable to expect direct control of $s$. For example, if $e^s$ were a diagonal matrix of constants $c_1,..., c_n$ in the trivial frame, then $e^s( A_0)$ will also be a flat connection. However, one eigenvalue $c_i$ can be arbitrarily large while still preserving the condition that $s$ be trace free, so $s$ cannot be controlled. What does end up being true is that under a small curvature assumption, there exists a normalized endomorphism $\hat s$, which may be distinct from $s$, that nevertheless gives the same connection under the complexified gauge group action, and is uniformly controlled in $C^0$. The key result of the first two named authors is as follows.
\begin{thm}[Datar-Jacob \cite{DJ}]
\label{gaugef}
Let $e^s(A_0)$ be a connection on $V$ given by the action of a trace free Hermitian endomorphism $s.$ There exists constants $\epsilon_0>0$, and $C_0>0$, depending only on $\omega$, $A_0,$ and $H$, so that if
\be
\qquad \|F_{e^s(A_0)}\|^2_{C^0(E)}\leq\epsilon_0,\nonumber
\ee
then there exists another trace free Hermitian endomorphism $\hat s$ satisfying that $\hat s$ is perpendicular to the Kernel of $d_{A_0}$, in addition to
\be
e^s( A_0)=e^{\hat s}(A_0)\qquad{\rm and}\qquad \|\hat s\|_{C^0(E) }\leq C_0.\nonumber
\ee
\end{thm}
We remark that the assumptions that $V$ be regular and admit a flat connection are critical, as they imply that the holomorphic automorphism group of $V$ is precisely $n$ dimensional \cite{FMW}. The idea of the proof is that the linearization of the complex gauge group action of a Hermitian endomorphism on $A_0$ is $\star d_{A_0}s$. Restricting to endomorphisms perpendicular to the Kernel of $d_{A_0}$, a Poincar\'e inequality gives that the linearized map is invertible with bounded inverse. Thus, if $e^s(A_0)$ is sufficiently close to $A_0$, via the contraction mapping principle the results of the theorem hold. In order for the theorem to hold under the small curvature assumption, a connectedness argument is applied. We direct the reader to \cite{DJ} for further details.
\subsection{Spectral covers}
We now discuss holomorphic vector bundles over our elliptic fibration $M$, as opposed to a single elliptic curve.
We assume that $f: M \rightarrow N$ has only singular fibers of Kodaira type $I_1$ and type $II$. Then $ M $ coincides with the Weierstrass model $\check{f}:\check{M}\rightarrow N$, i.e. $M=\check{M}$ and $f=\check{f}$.
Let $V$ be a holomorphic vector bundle $V$ of rank $n$ on $M$ such that the determinant line bundle $ \bigwedge^{n} V$ is trivial, i.e. $\bigwedge^{n} V \cong \mathcal{O}_{M} $. If the restriction of $V$ on the generic fiber of $f$ is regular semi-stable, then a multi-valued section of $f$ is constructed in \cite{FMW}, which is called the spectral cover associated to $V$. More precisely, we have the following theorem.
\begin{thm}[\cite{FMW}]\label{special cover}
Assume that the restriction of $V$ on the generic fiber of $f$ is semi-stable and regular. Then there exists a divisor $$D_V \in | n \sigma (N)+m l|, $$ called the spectral cover associated to $V$, where $ l$ denotes effective divisor class of the fibers of $f$, $m\in \mathbb{Z}$ satisfies $0\leq m \leq c_2(V)$, and for a generic $w\in N_0$, $$ V|_{M_w}\cong \bigoplus_{j=1}^{\ell}\mathcal{O}_{M_w}(q_j-0)\otimes\cI_{r_j}, \ \ \ D_{V}\cap M_w =\sum_{j=1}^{\ell}r_j q_j\in | n \sigma (w)|. $$
\end{thm}
We recall the construction in \cite{FMW}.
Since $h^{0}(M_{w}, \mathcal{O}_{M_w}(n \sigma (w)))=n$ for any fiber $M_w$, the push forward $f_{*}\mathcal{O}_{M}(n \sigma)$ is a vector bundle of rank $n$ on $N$, and more precisely, $$f_{*}\mathcal{O}_{M}(n \sigma)=\mathcal{O}_{N}\oplus L^{-2}\oplus \cdots \oplus L^{-n},$$ where $L^{-1}=\sigma^* \mathcal{O}_M(\sigma)$ by Lemma 4.1 of \cite{FMW}.
We denote $p: \mathcal{P}_{n-1}\rightarrow N$ the projection bundle, so $\mathcal{P}_{n-1}=\mathbb{P}f_{*}\mathcal{O}_{M}(n \sigma)$ (cf. Section 4.1 of \cite{FMW}). For any $w\in N$, the fiber $p^{-1}(w)$ is the complete linear system $|n \sigma (w)| \cong \mathbb{CP}^{n-1}$, i.e. $p^{-1}(w)= |n \sigma (w)|$, and is identified as the coarse moduli space for semi-stable bundles of rank $n$ on $M_{w}$ (cf. Section 1 of \cite{FMW}). Since the restriction of $V$ to the generic fiber is semi-stable, there is a non-empty Zariski open subset $N'\subset N$ such that for any $w\in N'$, $V|_{M_w}$ is semi-stable, which defines a point $\varrho (V|_{M_w})\in |n \sigma (w))|$ by Theorem 1.2 in \cite{FMW}. Then Lemma 4.2 of \cite{FMW} defines a section $$\mathcal{A}_{V}: N' \rightarrow p^{-1}(N'), \ \ \ \ {\rm by} \ \ \ \mathcal{A}_{V}(w)=\varrho (V|_{M_w}),$$ and by Lemma 6.1 in \cite{FMW}, $\mathcal{A}_{V}$ extends to $N$ as a section of $\mathcal{P}_{n-1}$, denoted still by $\mathcal{A}_{V}: N \rightarrow \mathcal{P}_{n-1}$.
Section 4.3 in \cite{FMW} constructs an $n$-sheeted branched covering $ \varrho : \mathcal{T}\rightarrow \mathcal{P}_{n-1}$, which admits a $\mathbb{CP}^{n-2}$--fibration $r: \mathcal{T}\rightarrow M$. For any smooth fiber $M_w$, $\mathcal{T}_w= r^{-1}(M_w) \rightarrow M_w$ coincides with the construction in Section 2.1 of \cite{FMW} as follows. Let $\Pi_w \subset M_{w}^{\otimes n}$ be the subset such that $(q_1, \cdots, q_n)\in \Pi_w$ if and only if the divisor $q_1+ \cdots +q_n $ is linearly equivalent to $n \sigma (w)$. If $\mathbb{S}_n$ denotes the symmetric group, and $\mathbb{S}_{n-1} \subset \mathbb{S}_n$ is the subgroup fixing the last element, then $\mathbb{S}_n$ acts on $\Pi_w$, and the quotient $\Pi_w / \mathbb{S}_n = |n \sigma (w)| \cong \mathbb{CP}^{n-1}$. Also $\mathcal{T}_w=\Pi_w / \mathbb{S}_{n-1}$, $r|_{\mathcal{T}_w}:\mathcal{T}_w \rightarrow M_w$ is given by $(q_1, \cdots, q_{n-1}, q_n) \mapsto q_n$, and $\varrho|_{\mathcal{T}_w}: \mathcal{T}_w \rightarrow |n \sigma (w)|$ is a branched $n$-sheeted cover such that $\varrho|_{\mathcal{T}_w}$ is unbranched over $q_1+ \cdots +q_n \in | n \sigma (w)|$ if and only if $q_i\neq q_j$ for any $i\neq j$. Clearly, $r|_{\mathcal{T}_w}(\varrho|_{\mathcal{T}_w}^{-1}(q_1+ \cdots +q_n))=\{q_1, \cdots ,q_n\}\subset M_w$ for any $q_1+ \cdots +q_n \in | n \sigma (w)|$.
The spectral cover $D_{V}$ is defined as the scheme-theoretic inverse image of $\mathcal{A}_{V}(N)$, i.e. $D_{V}=\varrho^{-1}(\mathcal{A}_{V}(N)) $, which is a subscheme of $\mathcal{T}$, and $p\circ \varrho|_{D_{V}}:D_{V}\rightarrow N$ is finite and flat of degree $n$ (cf. Definition 5.3 in \cite{FMW}). By Lemma 5.4 of \cite{FMW}, $r|_{D_{V}}$ embeds $D_{V}$ in $M$ as an effective Cartier divisor, and $f\circ r|_{D_{V}}= p \circ \varrho|_{D_{V}}$. Therefore, we always regard $D_{V}$ as a divisor of $M$ in the present paper. Furthermore, Lemma 5.4 in \cite{FMW} shows that $\mathcal{O}_{M}(D_{V})\cong \mathcal{O}_{M}(n \sigma(N) )\otimes f^{*} \mathcal{L}_{V}$ where $\mathcal{L}_{V}=\mathcal{A}_{V}^{*}\mathcal{O}_{\mathcal{P}_{n-1}}(1)$. Thus $$D_{V} \in | n \sigma(N)+ m l |,$$ where $l$ denotes the effective divisor class of the fibers of $f$, and $m={\rm deg} \mathcal{L}_{V}\in \mathbb{Z}$.
The arguments in Section 6.1 of \cite{FMW} show that \begin{equation}\label{chern2} 0 \leq m={\rm deg} \mathcal{L}_{V} \leq c_{2}(V), \end{equation} which is sketched as follows. Since the restriction of $V$ to the generic fiber is regular semi-stable, there are only finite possible fibers such that the restrictions of $V$ are unstable. Lemma 6.2 of \cite{FMW} proves that by preforming finite allowable elementary modifications to $V$, one obtains a new bundle $V'$ such that the restriction of $V'$ to any fiber is semi-stable. Furthermore $c_2(V')\leq c_2(V)$, and equality holds if and only if $V'=V$, i.e. there is no elementary modification preformed.
The proof of Corollary 6.3 in \cite{FMW} shows that there is a coherent sheaf $V_0$, whose restriction on any fiber is regular semi-stable, and
a morphism $\psi:V_0 \rightarrow V'$, which is an isomorphism on $f^{-1}(U)$ for a nonempty Zariski open set $U\subset N$. The cokernel coherent sheaf $Q$ is a torsion sheaf supported on finite fibers, and admits a filtration by degree zero sheaves. Consequently, $c_2(V_0)=c_2(V')$. Note that $V_0$ is isomorphic to $V$ on $f^{-1}(U')$ for a nonempty Zariski open set $U'\subset N$, as the above two processes only change the restrictions of $V$ on finite fibers. Therefore we have $\mathcal{A}_{V_0}=\mathcal{A}_{V}$, $D_{V_0}=D_{V}$, and $\mathcal{L}_{V_0}=\mathcal{L}_{V} $. By Proposition 5.15 of \cite{FMW}, ${\rm deg} \mathcal{L}_{V_0} = c_{2}(V_0)$, and we obtain the inequality (\ref{chern2}).
The spectral cover $D_{V}$ gives a criterion of $V$ being stable.
\begin{thm}[Theorem 7.4 of \cite{FMW}]\label{stable criterion}
If $D_{V}$ is reduced and irreducible, then $V$ is stable with respect to $f^{*} c_{1}(\mathcal{O}_{\mathbb{CP}^{1}}(1))+t \alpha$, for all $0<t\leq (\frac{n^3}{4}c_{2}(V))^{-1}$, where $\alpha$ is an ample class on $M$.
\end{thm}
This theorem can be used to construct stable bundles on $M$ as follows. If $D\in | n \sigma (N)+m l|$, $m>2n$, is an effective reduced and irreducible divisor, then Lemma 5.4 in \cite{FMW} asserts that $D$ is the spectral cover of a unique section $\mathcal{A}$ of $\mathcal{P}_{n-1}$, which satisfies $m ={\rm deg}\mathcal{A}^* \mathcal{O}_{\mathcal{P}_{n-1}}(1)$. A holomorphic vector bundle $V$ is constructed from $\mathcal{A}$ (cf. Definition 5.2 in \cite{FMW}) such that the restriction of $V$ on every fiber is regular semi-stable with trivial determinant line bundle, and $D$ is the spectral cover of $V$, i.e. $D_{V}=D$.
We recall the construction in Section 5.1 of \cite{FMW} by assuming that $D$ is smooth, and does not intersect with any singular set of the singular fibers of $f$. If $\tilde{M}= D\times_{N} M$ denotes the base change, which is smooth, then there are morphisms $\tilde{f}: \tilde{M} \rightarrow D$ and $\nu_D: \tilde{M} \rightarrow M$ such that $f\circ \nu_D= f|_D \circ \tilde{f}$. We regard $\tilde{M}= D\times_{N} M\subset M\times_N M$ via the natural embedding $D\hookrightarrow M$. Then $\Sigma_D = \nu_D^* \sigma$ and $\Delta= \tilde{M} \cap \Delta_0$ are divisors, where $\Delta_0$ is the diagonal of $ M\times_N M$. For any $w\in N_0$, and $q_j(w)\in M_w\cap D$, we have $\tilde{M}_{(w, q_j(w))}=M_w$, $\Sigma_D \cap \tilde{M}_{(w, q_j(w))}= \{\sigma(w)\}$, and $\Delta \cap \tilde{M}_{(w, q_j(w))}= \{q_j(w)\}$.
Lemma 5.5 of \cite{FMW} asserts that the push forward $(\nu_D)_* \mathcal{O}_{\tilde{M}}(\Delta - \Sigma_D)$ satisfies that its restriction on every fiber is regular semi-stable with trivial determinant line bundle. Furthermore, for any line bundle $\tilde{L}$ on $D$, $(\nu_D)_* (\mathcal{O}_{\tilde{M}}(\Delta - \Sigma_D)\otimes \tilde{f}^*\tilde{L})$ also satisfies the required conditions.
Conversely, if $V$ is a holomorphic vector bundle whose restriction of $V$ on every fiber is regular semi-stable with trivial determinant line bundle, and $D$ is the spectral cover of $V$, then $$V=(\nu_D)_* (\mathcal{O}_{\tilde{M}}(\Delta - \Sigma_D)\otimes \tilde{f}^*\tilde{L})$$ for a certain line bundle $\tilde{L}$ on $D$ by Proposition 5.7 in \cite{FMW}.
Now, since ${\rm deg}L=-\sigma^2=2$, Proposition 5.12 of \cite{FMW} asserts that one can choose $V$ via a suitable $\tilde{L}$ on $D$ such that the first Chern class $c_1(V)=0$, and therefore, $V$ has trivial determinant line bundle on $M$.
Now Theorem 7.4 of \cite{FMW} shows that $V$ is stable with respect to $f^{*} c_{1}(\mathcal{O}_{\mathbb{CP}^{1}}(1))+t \alpha$ for $0<t \ll 1$.
In summary, we have
\begin{thm}\label{stable construction} If $D\in | n \sigma (N)+m l|$, $m>2n$, is an effective reduced and irreducible divisor, then there exists a holomorphic vector bundle $V$ of rank $n$ with $c_1(V)=0$ on $M$ such that the restriction of $V$ on every fiber is regular semi-stable, and $D$ is the spectral cover of $V$, i.e. $D_{V}=D$. Furthermore,
$V$ is stable with respect to $f^{*} c_{1}(\mathcal{O}_{\mathbb{CP}^{1}}(1))+t \alpha$, for all $0<t\leq (\frac{n^3}{4}c_{2}(V))^{-1}$, where $\alpha$ is an ample class on $M$.
\end{thm}
\subsection{Collapsing of Ricci-flat K\"{a}hler-Einstein metrics}
\label{basemetric}
We now introduce some preliminary results on our family of collapsing base metrics on $M$, and highlight a new decay estimate necessary for our main theorem. The reader is directed to Appendix A for a proof of this particular asymptotic decay.
Let $\alpha$ be an ample class on $M$, $\alpha_{t} =t \alpha + f^*c_1(\mathcal{O}_{\mathbb{CP}^1}(1))$, $t\in (0,1]$, and $\omega_t \in \alpha_{t} $ the unique Ricci-flat K\"{a}hler-Einstein metric, which satisfies the complex Monge-Amp\`ere equation $$ \omega_t^2=c_t t \Omega \wedge \overline{\Omega}. $$ Here $\Omega$ is a holomorphic symplectic form on $M$, and $c_t $ tends to a positive number $ c_0$ when $t\rightarrow 0$.
For any $t\in(0,1]$, there exists a family of K\"ahler metrics $\omega_t^{SF}$ on $M_0$, such that $\omega_t^{SF}|_{M_w}$ is the flat metric in the class $t\alpha|_{M_w}$. Such metrics are called {\it semi-flat}, and we recall their construction here.
Note that $M_0$ is obtained by the quotient of the holomorphic cotangent bundle $T^*N_0$ by a lattice subbundle $\Lambda$. More precisely, we have a covering map $p:T^*N_0\rightarrow M_0$, so that $p(\Lambda)=\sigma(N_0)$, and the pull-back $p^*\Omega$ is the canonical holomorphic symplectic form on $T^*N_0$.
If $U\subset N_0$ is a small open disk, we can choose a holomorphic coordinate $w$ on $U$ so that $\Lambda\cap T^*U={\rm Span}_{\mathbb Z}\{dw,\tau(w) dw\}$, where $\tau(w)$ is the period of the elliptic curve $M_w$. Under the trivialization $T^*U\cong U\times\mathbb C$ given by $zdw\mapsto(w,z)$, we see $p^*\Omega=dw\wedge dz$. Note that the $(1,1)$-form $$ i\partial\overline{\partial}{\rm Im}(\tau)^{-1}({\rm Im}(z))^2 =\frac i2W(dz+bdw)\wedge\overline{(dz+bdw)}$$ is invariant under the translation of any local constant section of $\Lambda$ (cf. Section 3 in \cite{GTZ}), where
\be
W={\rm Im}(\tau)^{-1}\qquad{\rm and}\qquad b= -\frac{{\rm Im}(z)}{{\rm Im}(\tau)}\frac{\partial\tau}{\partial w}.\nonumber
\ee
Thus the above $(1,1)$-form can be regarded as living on $f^{-1}(U)$. The semi-flat metric is defined as
\be
\label{SFmetric}
\omega_t^{SF} = \frac i2\left(tW(dz+bdw)\wedge\overline{(dz+bdw)}+W^{-1} dw\wedge d\bar w\right).
\ee For simplicity we denote $\omega^{SF}:=\omega_1^{SF}$, which we use as a fixed base metric, and $\theta= dz+bdw$.
We now state our decay result for $\omega_t$ as $t\rightarrow 0$, which is contained in Theorem \ref{ttm-decay} (see Appendix A below). Given $U\subset N_0$,
\cite{GTZ} asserts that
there exists a local section $\sigma_0$ such that for any $\ell\geq 0$, $$ \|T_{\sigma_0}^*\omega_t- \omega^{SF}_t\|_{C_{\rm loc}^\ell(M_U, \omega^{SF}_t)}\rightarrow 0,$$ when $t\rightarrow \infty$, where $T_{\sigma_0}$ denotes the fiberwise translation by $\sigma_0$ (cf. Lemma 4.7 in \cite{GTZ}). Theorem \ref{ttm-decay} shows that there is
a $(1,1)$-form $\chi_t$ satisfying $\chi_t \rightarrow 0$ in $C^\infty$ as $t\rightarrow 0$, so that $T_{\sigma_0}^*\omega_t$ approaches to $ \omega^{SF}_t+ f^*\chi_t$ faster than any polynomial rate, i.e. $$ T_{\sigma_0}^*\omega_t= \omega^{SF}_t+ f^*\chi_t+ o(t^{\frac{\nu}{2}}), $$ for any $\nu\gg 1$.
In the proof of the main theorem we need a slightly stronger statement. The difference between $T_{\sigma_0}^*\omega_t$ and $\omega^{SF}_t$ can be written out in components in the fiber and base directions:
$$T_{\sigma_0}^*\omega_t- \omega^{SF}_t=\varphi_{t, z \bar{z}}dz\wedge d\bar{z}+\varphi_{t, w \bar{w}}dw\wedge d\bar{w}+\varphi_{t, w \bar{z}}dw\wedge d\bar{z}+ \varphi_{t, z \bar{w}}dz\wedge d\bar{w}. $$ We need the following important lemma, which is a direct consequence of Lemma \ref{decay}.
\begin{lem}\label{lem-decay}
For any $\nu\gg 1$ and $\ell\geq 0$, there is a constant $C_{\ell,\nu}>0$ such that on $M_{U'}$, $U' \subset U$, $$\|\varphi_{t, w \bar{w}}-\chi_{t,w\bar{w}}\|_{C^0}\leq C_{0,\nu}t^{\frac{\nu}{2}}, $$ $$\|\frac{\partial}{\partial z}\varphi_{t, w \bar{w}}\|_{C^\ell}+\|\frac{\partial}{\partial \bar{z}}\varphi_{t, w \bar{w}}\|_{C^\ell}+ \|\varphi_{t, z \bar{z}}\|_{C^\ell}+ \|\varphi_{t, z \bar{w}}\|_{C^\ell}+ \|\varphi_{t, w \bar{z}}\|_{C^\ell}\leq C_{\ell,\nu} t^{\frac{\nu}{2}}, $$ and $\chi_{t, w \bar{w}} \rightarrow 0$ in the $C^\infty$-sense when $t\rightarrow 0$. Here $\chi_t=\chi_{t,w\bar{w}}dw\wedge d\bar{w} $, and the $C^\ell$-norms are calculated using the fixed K\"{a}hler metric $\omega^{SF}$ on $M_U$.
\end{lem}
In this section we also recall the blow-up limit of $t^{-1} \omega_t$, which shows up in the analysis to follow. Let $t_k \rightarrow 0$ and $w_k \rightarrow w_0$ in $U\subset N_0$.
By \cite{GTZ}, $$(M, t_k^{-1} \omega_{t_k}, p_k) \rightarrow (\mathbb{C} \times M_{w_0}, \omega_\infty =\omega_{w_0}^{F}+\frac{i}{2}W^{-1}(w_0)d\tilde{w}\wedge d\bar{\tilde{w}}, p_0), $$ in the $C^\infty$-Cheeger-Gromov sense, where $w_k=f(p_k)$, $p_k\rightarrow p_0\in M_{w_0} $, $\omega_{w_0}^{F}$ is the flat K\"{a}hler metric representing $\alpha|_{M_{w_0}}$, i.e. $\omega_{w_0}^{F}=\omega^{SF}|_{M_{w_0}} $, and $\tilde{w}$ denotes the coordinate of $ \mathbb{C}$. More precisely, if
$D_r=\{\tilde{w} \in \mathbb{C}| |\tilde{w}|<r\}$, we define smooth embeddings $\Phi_{k,r}: D_r \times M_{w_0} \rightarrow M_U$ by $$( \tilde{w}, a_1+a_2 \tau (w_0))\mapsto (w_k +\sqrt{t_k}\tilde{w}, a_1+a_2 \tau(w_k+\sqrt{t_k}\tilde{w})), \ \ a_1,a_2 \in\mathbb{R}/\mathbb{Z},$$ where we identify $M_U$ with $(U\times\mathbb{C})/{\rm Span}_{\mathbb{Z}}\{1, \tau\}$.
If $z=a_1+a_2 \tau (w_0)$, then $ a_1+a_2 \tau(w_k+\sqrt{t_k}\tilde{w})=z +h_k$, where $$h_k=i(2{\rm Im}\tau(w_0))^{-1}(\bar{z}-z)(\tau(w_k+\sqrt{t_k}\tilde{w})-\tau(w_0)), $$ which satisfies that
$\|h_k\|_{C^{\ell}}\rightarrow 0$ when $t_k \rightarrow 0$.
Therefore
$$\Phi_{k,r}^*(dz +bdw)= dz+dh_k +\sqrt{t_k }(b-{\rm Im}h_k({\rm Im}\tau)^{-1}\partial_w \tau )d\tilde{w}\rightarrow dz,$$ in the $C^\infty$-sense. Clearly, $d\Phi_{k,r}^{-1}I d\Phi_{k,r} \rightarrow I_\infty$, where $I$ is the complex structure of $M$ and $I_\infty$ denotes the complex structure of $ \mathbb{C}\times M_{w_0}$, and \begin{equation}\label{blowup1}\Phi_{k,r}^* t_k^{-1} \omega_{t_k}^{SF} \rightarrow \omega_\infty=\frac i2\left(W(w_0)dz \wedge d\bar{z} +W^{-1}(w_0) d\tilde{w}\wedge d\bar {\tilde{w}}\right), \end{equation} in the $C^\infty$-sense on $ D_r\times M_{w_0}$. Furthermore, \begin{equation}\label{blowup2}(T_{\sigma_0} \circ \Phi_{k,r})^* t_k^{-1} \omega_{t_k} =\Phi_{k,r}^* t_k^{-1}T_{\sigma_0}^* \omega_{t_k} \rightarrow \omega_\infty, \end{equation} in the $C^\infty$-sense, on $ D_r\times M_{w_0}$, when $t_k\rightarrow 0$ by \cite{GTZ}.
\subsection{Fourier-Mukai transform} In this section, we recall a notion, called the Fourier-Mukai transform (cf. \cite{AP,LYZ,Ch3,Ch4} etc.), and we present a little variant of the standard construction for the convenience of the proof of Theorem \ref{thm-main2}.
Let $N^o \subset N_0$ be a Zariski open subset, and $D^o \subset M_{N^o}$ be a smooth curve such that $f|_{D^o}: D^o \rightarrow N^o$ is a unbranched $n$-sheets cover.
Note that the moduli space of flat $U(1)$-connections on $D^o$ is the cohomology group $H^1(D^o, \mathcal{U}_c(1))\cong H^1(D^o,U(1))$, where $\mathcal{U}_c(1)$ is the $U(1)$-valued locally constant sheaf. For any $\Theta \in H^1(D^o, \mathcal{U}_c(1))$, the Fourier-Mukai transform takes the pair $(D^o, \Theta)$ to a unitary gauge equivalent class $ \mathcal{FM}(D^o, \Theta)$ of $U(n)$-connections on $M_{N^o}$. We review the construction as the following.
If $\tilde{M}^o= D^o\times_{N^o} M_{N^o}$ is the base change, then the projection $\tilde{f}: \tilde{M}^o\rightarrow D^o$ is a fibration with the fiber $\tilde{M}_p^o=M_{f(p)}$,
and $\nu_D: \tilde{M}^o \rightarrow M_{N^o}$ is a unbranched $n$-sheets cover satisfying
$f\circ \nu_D= f|_{D^o} \circ \tilde{f}$.
We embed $\tilde{M}^o= D^o\times_{N^o} M_{N^o} \hookrightarrow M_{N^o}\times_{N^o} M_{N^o}$ via the natural inclusion $D^o\hookrightarrow M_{N^o}$.
Let $\Sigma = \nu_D^* \sigma$ and $\Delta= \tilde{M}^o \cap \Delta_0$, where $\Delta_0$ denotes the diagonal of $ M_{N^o}\times_{N^o} M_{N^o}$. For any $x\in N^o$, and $q(x)\in M_x\cap D^o$, we have $\tilde{M}_{(x, q(x))}^o=M_x$, $\Sigma \cap \tilde{M}_{(x, q(x))}^o= \{\sigma(x)\}$, and $\Delta \cap \tilde{M}_{(x, q(x))}^o= \{q(x)\}$. We regard $\Sigma$ as the zero section of $\tilde{f}$, which is used to identify the fibers with elliptic curves, and view $\Delta$ as the pull back the multi-section $D^o$, which is a section of $\tilde{f}$.
There is a $U(1)$-connection $A^o$ on the smooth trivial line bundle $ \tilde{M}^o \times \mathbb{C}$, which is obtained by the
restriction of the Poincar\'{e} line bundle (cf. \cite{AP}) on $ M_{N^o}\times_{N^o} M_{N^o}$ by identifying $M_{N^o}$ with the Jacobian $\check{M}_{N^o}$. We exhibit $A^o$ explicitly.
If $U\subset D^o$ is an open disc such that $f|_U:U\rightarrow f(U)$ is biholomorphic, we choose the coordinate $w$ such that $\tilde{M}_{U}^o\cong T^*U/ {\rm Span}_{\mathbb{Z}}\{dw, \tau dw\}$, where $\tau(w)$ is the period of $\tilde{M}_w^o$. Here the section $\Sigma\equiv 0$ under this identification. If $z$ denotes the coordinate on the fiber, then the holomorphic symplectic form $\nu_D^*\Omega=dw\wedge dz$, and $\Delta \cap \tilde{M}_{U}^o$ is given by a holomorphic function $q=q(w)$ on $U$, i.e. $\Delta \cap \tilde{M}_{U}^o=\{(w, q(w))\}\subset U\times \mathbb{C}/ {\rm Span}_{\mathbb{Z}}\{1, \tau \} $. We have the $U(1)$-connection \begin{equation}\label{u(1)} A^o= \pi ( {\rm Im} (\tau))^{-1} (q\bar{\theta}-\bar{q}\theta), \end{equation} on $\tilde{M}_{U}^o $, where $\theta=dz+bdw$.
If $y_1$ and $y_2$ are real functions defined on $U\times \mathbb{C}$ by $z=y_1+\tau y_2$, then $dy_1$ and $dy_2$ are well-defined 1-forms on $\tilde{M}_U^o$. Note that $\tilde{M}_U^o$ is diffeomorphic to $U\times (\mathbb{R}/\mathbb{Z})^2$, and we can regard $y_1$ and $y_2$ as the angle coordinates of $\mathbb{R}/\mathbb{Z}$. We have the decomposition of the cotangent bundle $T^*\tilde{M}_U^o= {\rm Span}_{\mathbb{R}}\{dy_1, dy_2\}\oplus {\rm Span}_{\mathbb{R}}\{dx_1, dx_2\}$, where $w=x_1+ix_2$.
Since $dz=dy_1+\tau dy_2+y_2 d\tau$, $2i{\rm Im} (\tau )y_2=z-\bar{z}$, we have $\theta=dy_1+\tau dy_2$. If we write $q=q_1+\tau q_2$, then \begin{equation}\label{u(1)1} A^o= 2\pi i (q_2dy_1-q_1dy_2). \end{equation} If we choose another basis of the lattice ${\rm Span}_{\mathbb{Z}}\{1, \tau \}$, and let $y_1'$ and $y_2'$ be the corresponding angle coordinates, then $y_j'=\sum c_{ji}y_i$ and $q_j'=\sum c_{ji}q_i$ with $\det (c_{ji})=1 $ and $c_{ji}\in\mathbb{Z}$. A direct calculation shows that $A^o$ is independent of the choice of the basis, and therefore $A^o$ is a global defined $U(1)$-connection on the trivial line bundle $ \tilde{M}^o \times \mathbb{C}$.
Let $\check{y}_1$ and $\check{y}_2$ be the dual coordinates of $y_1$ and $y_2$ on the dual space $(\mathbb{R}^2)^*$, i.e. if we view $(\mathbb{R}^2)^*$ as the cotangent space, then $\check{y}_1$ and $\check{y}_2$ are coordinates with respect to the basis $dy_1$ and $dy_2$. We
identify $\mathbb{R}^2/\mathbb{Z}^2$ with the dual torus
$(\mathbb{R}^2)^*/(\mathbb{Z}^2)^*$ via the symplectic form $\omega=dy_2\wedge dy_1$, i.e. $v \mapsto \omega(v, \cdot)$. Then $q=(q_1,q_2)\in \mathbb{R}^2/\mathbb{Z}^2$ is mapped to $\check{q}=(q_2,-q_1)$ in $(\mathbb{R}^2)^*/(\mathbb{Z}^2)^*$. The Poincar\'{e} line bundle is a line bundle on $U\times \mathbb{R}^2/\mathbb{Z}^2 \times (\mathbb{R}^2)^*/(\mathbb{Z}^2)^*$ with the $U(1)$-connection $$A_P=2\pi i (\check{y}_1dy_1+\check{y}_2dy_2). $$
Thus $A^o= A_P|_{U\times \mathbb{R}^2/\mathbb{Z}^2\times \{\check{q}\}}$, which coincides with the constructions \cite{AP,LYZ}.
Let $\{U_\lambda| \lambda\in\Lambda\}$ be a locally finite open cover of $N^o$ such that any intersection $U_{\lambda_1}\cap \cdots \cap U_{\lambda_h}$ is contractible. On any $U_\lambda$, there are holomorphic functions $q_{1}, \cdots, q_{n}$, such that $D^o \cap M_{U_\lambda}=\{(w, q_{1}(w)), \cdots, (w, q_{n}(w))| w\in U_\lambda\}$. Furthermore, $D^o \cap M_{U_\lambda}=U_\lambda^1 \cup \cdots \cup U_\lambda^n$ is a disjoint union of open sets biholomorphic to $U_\lambda$ where $U_\lambda^j=\{(w, q_j(w))\}$, and $\{U_\lambda^j|\lambda\in\Lambda, j=1, \cdots, n \}$ is an open cover of $D^o $ such that any intersections are contractible.
If $\Theta\in H^1(D^o, \mathcal{U}_c(1))$, then we let $\{g_{\mu\lambda}^{ij_i}\}\in \mathcal{C}^1(\{U_\lambda^j\}, \mathcal{U}_c(1))$ be the cocycle representing $\Theta$, where $U_\mu^i \cap U_\lambda^{j_i} \neq\emptyset$, and $g_{\mu\lambda}^{ij_i}$ are $U(1)$-valued constant functions on $U_\mu^i \cap U_\lambda^{j_i} $. If $U_\mu^i \cap U_\lambda^j \cap U_\nu^k\neq\emptyset$, then $g_{\mu\lambda}^{ij}g_{\lambda\nu}^{jk}g_{\nu\lambda}^{ki}=1$. We identify $\tilde{f}^*g_{\mu\lambda}^{ij_i}=g_{\mu\lambda}^{ij_i}$, and regard $g_{\mu\lambda}^{ij_i}$ as $U(1)$-valued constant functions on $\tilde{M}^o_{U_\mu^i}\cap \tilde{M}^o_{U_\lambda^{j_i}}$.
Note that $(g_{\mu\lambda}^{ij_i})^{-1}A^o g_{\mu\lambda}^{ij_i}+(g_{\mu\lambda}^{ij_i})^{-1}dg_{\mu\lambda}^{ij_i}=A^o$. If $L_\Theta$ denotes the line bundle on $\tilde{M}^o $ given by the cocycle $\{(\tilde{M}^o_{U_\mu^i}\cap \tilde{M}^o_{U_\lambda^{j_i}}, g_{\mu\lambda}^{ij_i})\}$, then $A^o$ induces a $U(1)$-connection on $L_\Theta$ locally given by (\ref{u(1)}) denoted still by $A^o$.
The pushforward $(\nu_D)_* L_\Theta $ is a rank $n$ bundle on $ M_{N^o}$ given by the transitions $g_{\mu\lambda}={\rm diag}\{g_{\mu\lambda}^{1,j_1}, \cdots, g_{\mu\lambda}^{n,j_n}\}$ on $M_{U_\mu }\cap M_{U_\lambda}$, where $U_\mu^{i} \cap U_\lambda^{j_i}\neq\emptyset$. There is a natural $U(n)$-connection $\Xi$ on $(\nu_D)_* L_\Theta $ induced by $A^o$ given locally by \begin{eqnarray*} \Xi|_{M_{U_\lambda}}& =& {\rm diag}\{(\nu_D)_* A^o|_{\tilde{M}^o_{U_\lambda^{1}}}, \cdots, (\nu_D)_* A^o|_{\tilde{M}^o_{U_\lambda^{n}}}\} \\ & =& \pi ( {\rm Im} (\tau))^{-1} ({\rm diag}\{q_{1}, \cdots, q_{n}\}\bar{\theta}-{\rm diag}\{\bar{q}_{1}, \cdots, \bar{q}_{n}\}\theta), \end{eqnarray*} which satisfies $g_{\mu\lambda}^{-1}\Xi|_{M_{U_\mu}}g_{\mu\lambda}+g_{\mu\lambda}^{-1}dg_{\mu\lambda}=\Xi|_{M_{U_\lambda}}$.
If $\{g_{\mu\lambda}'^{ij_i}\}$ is an another cocycle representing $\Theta$, then there is a cycle $\{s_{\lambda}^{j}\}\in \mathcal{C}^0(\{U_\lambda^j\}, \mathcal{U}_c(1))$ such that $g_{\mu\lambda}'^{ij_i} s_{\lambda}^{j_i}=s_{\mu}^{i}g_{\mu\lambda}^{ij_i}$ when $U_\mu^{i} \cap U_\lambda^{j_i}\neq\emptyset$.
If we define $s_\lambda={\rm diag}\{s_{\lambda}^{1},\cdots, s_{\lambda}^{n} \}$ on $M_{U_\lambda}$, then $g_{\mu\lambda}' s_{\lambda}=s_{\mu}g_{\mu\lambda}$, and $\{s_\lambda| \lambda\in\Lambda\}$ induces a unitary gauge change of $(\nu_D)_* L_\Theta $.
\begin{defn} The Fourier-Mukai transform $\mathcal{FM} (D^o, \Theta)$ of $(D^o, \Theta)$ is defined as the unitary gauge equivalent class $[\Xi] $ of the $U(n)$-connection $\Xi$ on $(\nu_D)_* L_\Theta $, i.e. $$ \mathcal{FM} (D^o, \Theta)= [\Xi].$$
\end{defn}
For any $t\in (0,1]$, the semi-flat metric $\omega^{SF}_t$ is HyperK\"{a}hler, and by using the HyperK\"{a}hler rotation, we can find a new complex structure and a symplectic form such that $D^o$ is a special lagrangian submanifold.
In \cite{LYZ}, it is shown that the connection obtained by the Fourier-Mukai transform of a special lagrangian section satisfies the deformed Hermitian-Yang-Mills equation, and in the case of dimension 2, the standard Hermitian-Yang-Mills equation. In the present case, the bundle $(\nu_D)_* L_\Theta $ with the connection $\Xi$ splits locally, and therefore, it is a corollary of \cite{LYZ} that $\Xi$ is an anti-self-dual connection.
We give a direct calculation proof of this assertion.
\begin{prop}
\label{propFM} If $\Theta\in H^1(D^o, \mathcal{U}_c(1))$, then for any $\Xi\in \mathcal{FM} (D^o, \Theta)$, $\Xi$ is an anti-self-dual connection with respect to the semi-flat HyperK\"{a}hler structure $(\omega^{SF}_t, \Omega)$, $t\in (0,1]$, i.e. the curvature $F_\Xi$ satisfies that $$ F_\Xi\wedge \omega^{SF}_t=0, \ \ and \ \ F_\Xi\wedge \Omega =0.$$
\end{prop}
\begin{proof} Since the anti-self-dual equation is unitary gauge invariant, we only need to verify the split case, i.e. $\Xi|_{M_{U_\lambda}} = {\rm diag}\{ A^o|_{\tilde{M}^o_{U_\lambda^{1}}}, \cdots, A^o|_{\tilde{M}^o_{U_\lambda^{n}}}\}$, where we identify $M_{U_\lambda}$ and $\tilde{M}^o_{U_\lambda^{j}}$ via $\nu_D$. The curvature $$F_{\Xi}|_{M_{U_\lambda}} = {\rm diag}\{F_{ A^o}|_{\tilde{M}^o_{U_\lambda^{1}}}, \cdots, F_{A^o}|_{\tilde{M}^o_{U_\lambda^{n}}}\} , $$ and thus we need to prove that $F_{ A^o}$ satisfies the anti-self-dual equation.
By
$\overline{\partial}\tau=0$, we have $0=\overline{\partial}\tau_1+i\overline{\partial}\tau_2$, where
$\tau=\tau_1+i\tau_2$, and $\partial_{\bar{w}} \bar{\tau}=\partial_{\bar{w}} \tau_1 -i\partial_{\bar{w}} \tau_2=-2i\partial_{\bar{w}} \tau_2$. Thus $$F_{A^o}^{0,2}=\pi\overline{\partial} (\tau_2^{-1} q\bar{\theta})=\frac{\pi q}{2i\tau_2^2}\partial_{\bar{w}} \bar{\tau} d \bar{z}\wedge d\bar{w}+\frac{\pi q}{\tau_2^2}\partial_{\bar{w}}\tau_2d \bar{z}\wedge d\bar{w}=0, $$ which is equivalent to $F_{A^o}\wedge \Omega =0$.
By (\ref{u(1)1}), $$F_{A^o}=d A^o= 2\pi i \sum_{j=1,2}(\partial_{x_j} q_2 dx_j \wedge dy_1-\partial_{x_j} q_1 dx_j \wedge dy_2),$$ and by (\ref{SFmetric}),
$$\omega_t^{SF} = t dy_1 \wedge dy_2 +W^{-1} dx_1\wedge dx_2.$$
Thus $$F_{A^o} \wedge \omega_t^{SF}=0,$$ and we obtain the conclusion.
\end{proof}
Finally, we remark that the split $\Xi $ obtained by the Fourier-Mukai transform is $T^2$-invariant, and thus reduces to
solutions of Hitchin equations \cite{Hit} on the base $N^o$.
\subsection{Small energy estimates on collapsed K3 surfaces} Finally, we review small energy estimates for curvatures of anti-self-dual connections with respect to collapsed metrics.
As above, for $t\in (0,1]$, let $\omega_t \in \alpha_{t} =t \alpha + f^*c_1(\mathcal{O}_{\mathbb{CP}^1}(1))$ be the unique Ricci-flat K\"{a}hler-Einstein metric in $\alpha_t$, and consider $\Xi_t$, a family of anti-self-dual connections on $P$ with with respect to $\omega_t$.
For any $p\in M$, and $r>0$, we define the local energy of the curvature $F_{\Xi_t}$ as
\be
\label{energy}
\E_t(p,r)=\frac{r^4}{{\rm Vol}_{\omega_t}(B_{\omega_t}(p,r))}\int_{{B_{\omega_t}(p,r)}} |F_{\Xi_t}|^2_{\omega_t}\,\omega_t^2.
\ee
This energy is a continuous function of $p$ and $r$. By the Bishop-Gromov comparison Theorem, for $r_1\leq r_2$ it holds
\be
\E_t(p,r_1)\leq\E_t(p,r_2)\qquad{\rm and}\qquad \E_t (p,0)=0.\nonumber
\ee
We have the following small energy estimate for curvatures of anti-self-dual connections, which is essentially Theorem 4.4 in \cite{And}.
\begin{lem}
\label{energybound} There exists a universal constant $\varepsilon >0$, independent of $t$, such that if \be
\E_t (p,r)\leq \varepsilon, \nonumber
\ee for $p$ and $r$ satisfying $p \in M_K$ and $B_{\omega_t}(p,r)\subset M_{K'}$ (for fixed compact subsets $K\subset K'\subset N_0$), then \be
\sup_{{B_{\omega_t}(p,r/2)}} |F_{\Xi_t}|_{\omega_t}\leq \frac {C_{K'}\varepsilon^{\frac12}}{r^2} \nonumber
\ee for a constant $C_{K'}>0$.
\end{lem}
\begin{proof}
By Lemma 4.4 of \cite{GTZ}, the curvature $R_{\omega_t}$ is bounded by a uniform constant $c_{K'}$ on $M_{K'}$. The Weitzenb\"{o}ck formula (\ref{wei}) implies the Bochner formula
\be
\Delta_{\omega_t} |F_{\Xi_t}|_{\omega_t}\geq -|F_{\Xi_t}|^2_{\omega_t}-c_{K'} |F_{\Xi_t}|_{\omega_t}.\nonumber
\ee
One can now carry over the exact argument from \cite{And}, consisting of Moser iteration with the local Sobolev inequality $$ \frac{c_S}{3}\big(\frac{B_{\omega_t}(p,r)}{r^4}\big)^{\frac{1}{4}}\| \xi\|_{L^{4}(\omega_t)} \leq \| d \xi\|_{L^{2}(\omega_t)}$$ for any compactly supported function $\xi$ on $B_{\omega_t}(p,r)$, where $c_S$ is a universal constant (cf. (4.1) and Theorem 4.1 in \cite{And}).
If we
keep track of the extra $c_{K'}$ term, because this term is of lower order, it does not effect the choice of the uniform constant $\tau$, which is thus independent of $K$ and $K'$.\end{proof}
Choose $\varepsilon\ll 1$ such that $C_{K'}\varepsilon^{\frac12} \leq 4$.
This allows us to make the following definition.
\begin{defn}
For any $t\in(0,1]$, we define $R_t(p)>0$ be the minimal number such that
\be
\E_t (p, R_t(p))=\varepsilon. \nonumber
\ee
\end{defn}
In particular, for any compact set $K\subset N_0$, and $p\in M_K$, as long as $R_t(p)$ is small enough, it holds
\be
\label{curvradconclusion}
|F_{\Xi_t}|_{\omega_t}(p)\leq 4 R_t(p)^{-2},
\ee
and for any $r\geq R_t(p)$, $$\E_t(p, r)\geq \varepsilon.$$
\section{The main theorems}
In this section, we present the main theorems of this paper, and demonstrate its applications to SYZ mirror symmetry of K3 surfaces.
\begin{thm}\label{thm-main}
Let $M$ be a projective elliptically fibered $K3$ surface with fibration $
f:M\rightarrow N\cong \mathbb{CP}^1.$ Assume $f$ has a section $\sigma:N\rightarrow M$, and assume it has only singular fibers of Kodaira type $I_1$ and type $II$.
Let $\Omega$ be a holomorphic symplectic form on $M$, and let $\omega_t \in \alpha_{t} $ be the unique Ricci-flat K\"{a}hler-Einstein metric in $\alpha_{t} =t \alpha + f^*c_1(\mathcal{O}_{\mathbb{CP}^1}(1))$, $t\in (0,1]$, where $\alpha$ is an ample class on $M$. Let $P$ be a principal $SU(n)$-bundle on $M$, and let $\mathcal{V} $ be the smooth vector bundle of rank $n$ equipped with a Hermitian metric $H$ induced by $P$, i.e. $\mathcal{V}=P\times_{\rho}\mathbb{C}^n$.
Assume there exists a family of anti-self-dual $SU(n)$-connections $\Xi_t$ on $P$ with respect to $(\omega_t, \Omega)$, i.e. $$ F_{\Xi_t}\wedge \omega_t=0, \ \ and \ \ F_{\Xi_t}\wedge \Omega=0,$$ with $t\in (0,1]$. Let $V_t$ denote the holomorphic bundle of $\mathcal{V}$ equipped with the holomorphic structure induced by $\Xi_t$.
Furthermore, assume:
\begin{itemize}
\item[i)] The restriction of $V_t$ to a generic fiber of $f$ is semi-stable and regular.
\item[ii)] Let $D_{t} \in | n \sigma(N)+ m l |$ be the corresponding spectral cover of $V_t$, where $0<m \leq c_2(\mathcal{V})$. As $t\rightarrow 0$, $$ D_{t} \rightarrow D_{0} \ \ \ in \ \ | n \sigma(N)+ m l |.$$
\item[iii)] The limit $D_0$ can be written $$D_0= D_{0}^o+ D_0', $$ where $D_{0}^o \in | n \sigma(N)+ m' l |$ is reduced, for some $0\leq m' \leq m$, and $ D_0'\in |(m-m')l|$ consists of all irreducible components of $D_0$ supported on fibers.
\end{itemize}
Then the following holds:
\begin{itemize}
\item[i)]
For any sequence $t_k \rightarrow 0$, and any $p>2$, there exists a Zariski open subset $N^o \subset N_0$, a subsequence (still denoted $t_k$), a sequence of $L^{p}_2$ unitary gauge changes $u_k\in \mathcal{G}^{2,p}$ of $P|_{M_{N^o}}$, and a $L^{p}_1$ $SU(n)$-connection $\Xi_0$ on $P|_{M_{N^o}}$ so that on $M_{N^o}$
$$u_k(\Xi_{t_k}) \rightarrow \Xi_0 $$
in the locally $L^{p}_{1}$ sense. Here the norms are calculated using a fixed K\"{a}hler metric on $M$, and the Hermitian metric $H$ on $\mathcal{V}$.
\item[ii)] The curvature $F_{\Xi_{t_k}}$ of $\Xi_{t_k}$ is locally bounded,
i.e. for any compact subset $K\subset N^o$, there exists a constant $C_K$ so that $$ \|F_{\Xi_{t_k}}\|_{C^0(M_K)}\leq C_K.$$
\item[iii)] For any $w \in N^o$ and $0<\alpha <1$, there is a $C^{1,\alpha}$ unitary gauge $u_\infty$ on $M_w$ so that
$u_\infty(\Xi_0|_{M_w})$ is a smooth flat connection. This limiting connection satisfies that the bundle $\mathcal{V}|_{M_{w}}$ equipped with the holomorphic structure induced by $u_\infty(\Xi_0|_{M_w})$ is bi-holomorphic to $$\bigoplus_{q \in D_{0}^o \cap M_w}\mathcal{O}_{M_w}(q-\sigma(w)). $$\end{itemize}
\end{thm}
\begin{rk}\label{remark1}\rm
We remark that $D_0' \in |(m-m')l|$ is supported on fibers over a finite number of points, and we refer to these fibers as type III bubbles, which is the terminology used in the previous relevant works \cite{DS,Nis1,Nis2}.
\end{rk}
\begin{rk}\label{remark1+}\rm There is a topological constraint on $\mathcal{V}$ built into the above theorem, namely that $$c_2(\mathcal{V} )\geq 2n-2.$$
To see this, note that if $\sigma (N)$ is not an irreducible component of $D_{0}^o$, then
$D_{0}^o \cdot \sigma (N)=-2n+m' \geq 0$. Otherwise, $(D_{0}^o-\sigma (N))\cdot \sigma (N)=-2n+2+m' \geq 0$. In both cases, we have $m' \geq 2n-2$, which implies the inequality for the second Chern number.
\end{rk}
Let us demonstrate a case in which the hypotheses of Theorem \ref{thm-main} hold. For a given $m \in \mathbb{N}$ and $s\in (0,1]$, let $D_s$ be a family of effective reduced irreducible divisors in the complete linear system $|n \sigma (N)+m l|$ such that as $s\rightarrow 0,$ $$D_s \rightarrow D_0=D_{0}^o+\sum_j D_j \ \ \ {\rm in} \ \ |n \sigma (N)+m l|, $$ where $D_{0}^o$ is reduced and irreducible, $D_{0}^o\in |n \sigma (N)+m' l|$ for some $m' \leq m$, and $\sum_j D_j \in |(m-m')l|$. For example, we can take $D_s\equiv D$ for some fixed divisor.
By Theorem \ref{stable construction}, we can construct a family of holomorphic bundles $V_s$ of rank $n$ satisfying $c_1(V_s)=0$, the restriction of $V_s$ to any fiber $M_w$ is semi-stable and regular, and $D_s$ is the spectral cover of $V_s$. Furthermore, Proposition 5.15 of \cite{FMW} asserts that $c_2(V_s)=m$, and therefore, all of $V_s$ are smoothly isomorphic to the same smooth bundle, since $SU(n)$ is simply connected. Now,
Theorem 7.4 of \cite{FMW} shows that for any $s$ the bundle $V_s$ is stable with respect to $f^{*} c_{1}(\mathcal{O}_{\mathbb{CP}^{1}}(1))+t \alpha$ for $0<t \ll 1$ and $t\leq s$. As a result, by Theorem \ref{DUY} (and taking a diagonal sequence) we obtain a family of anti-self-dual connections $\Xi_t$, for which the hypotheses of Theorem \ref{thm-main} are verified.
\begin{thm}\label{thm-main2}
Under the setup of Theorem \ref{thm-main}, the unitary gauge equivalent class of the limit connection $\Xi_0$ is the Fourier-Mukai transform of a $\Theta\in H^1(D_{0}^o\cap M_{N^o}, \mathcal{U}_c(1))$, i.e. $$\Xi_0 \in\mathcal{FM}(D_{0}^o\cap M_{N^o}, \Theta),$$ where $\mathcal{U}_c(1)$ is the $U(1)$-valued locally constant sheaf.
\end{thm}
\subsection{Strominger-Yau-Zaslow mirror symmetry with anti-self-dual connections}
We now apply Theorem \ref{thm-main} to Fukaya's Conjecture 5.5 in \cite{Fuk}, which relates the adiabatic limits of anti-self-dual connections to
special Lagrangian cycles on the mirror Calabi-Yau manifolds. While describing the mirror symmetry background, we first consider the more general setup where $M$ is any projective elliptically fibered $K3$ surface admitting a section.
We normalize $\alpha_t$ by multiplying a constant, so that the normalized class $\tilde{\alpha}_t$ satisfies $ \tilde{\alpha}_t ^2=[{\rm Re}\Omega]^2=[{\rm Im}\Omega]^2$. Let $\tilde{\omega}_t\in \tilde{\alpha}_t$ be the Ricci-flat K\"{a}hler-Einstein metric in this class, and so $(\tilde{\omega}_t, {\rm Re}\Omega, {\rm Im}\Omega)$ is a HyperK\"{a}hler triple. Using the HyperK\"{a}hler rotation, we have a family of complex structures $J_t$ with corresponding K\"{a}hler form and the holomorphic symplectic from
$$\omega_{J_t}={\rm Im}\Omega \ \ {\rm and} \ \ \Omega_{J_t}=\tilde{\omega}_t+i {\rm Re}\Omega.$$
Using $\Omega|_{M_w}=0$ and $\Omega|_{\sigma (N)}=0$, under $J_t$ the fibration $f$ becomes a special Lagrangian fibration, and the section $\sigma$ is a special Lagrangian section with respect to $\omega_{J_t}$ and $\Omega_{J_t}$.
Mirror symmetry for K3 surfaces is well understood (cf. \cite{AM,Do,GW0,GroII,ABC}), and in particular the SYZ mirror symmetry of K3 surfaces was studied in Section 7 of Gross \cite{GroII} and in Gross-Wilson \cite{GW0}. For the reader's convenience we elaborate further on this setup. Let $[\sigma]$ denotes the class of the section $\sigma(N)$ in $ H^2(M, \mathbb{Z})$ and $l$ the fiber class. Then we have the following intersection pairings:
$$l^2=0, \qquad [\sigma]\cdot l=1, \qquad [\sigma]^2=-2, \qquad[\omega_{J_t}]\cdot [\sigma]=0,$$
$$[{\rm Im}\Omega_{J_t}]\cdot [\sigma]=0, \qquad [\omega_{J_t}]\cdot l=0,\qquad{\rm and}\qquad[{\rm Im}\Omega_{J_t}]\cdot l=0.$$
Now, the SYZ construction from Section 7 of \cite{GroII} uses the choice of a B-field $\mathbb{B}\in l^\bot /l \otimes \mathbb{R}/\mathbb{Z}$. However, Gross' assumptions are slightly different than those of the present paper. Namely, Gross assumes the K3 surface $M$ is generic, i.e. the picard group ${\rm Pic}(M)\cong \mathbb{Z}$, while in our case we have $\dim {\rm Pic}(M)\geq 2$. Nevertheless, the proof of Theorem 7.3 of \cite{GroII} shows that, in our case, if we further assume that $[\sigma]+(1+\frac{1}{2}[\omega_{J_t}]^2)l$ is an ample class on $M$, and the B-field $\mathbb{B}$ vanishes, then the SYZ mirror of $(M, \tilde{\omega}_t, \Omega_{J_t}) $ is $f: M \rightarrow N$ equipped with the HyperK\"{a}hler structure $(\check{\omega}_t, \check{\Omega}_t)$ and the B-field $\check{\mathbb{B}}_t$ satisfying $$ [\check{\Omega}_{t}]= (l\cdot [{\rm Re}\Omega_{J_t}])^{-1}([\sigma]+(1+\frac{1}{2}[\omega_{J_t}]^2)l-i[\omega_{J_t}]), \ \ \ [\check{\omega}_t]=(l\cdot [{\rm Re}\Omega_{J_t}])^{-1}[{\rm Im}\Omega_{J_t}],$$
$${\rm and} \ \ \ \check{\mathbb{B}}_t=(l\cdot [{\rm Re}\Omega_{J_t}] )^{-1} [{\rm Re}\Omega_{J_t}]-[\sigma]+{\rm mod} (l), $$ on the cohomological level.
We study the case that $[\sigma]+(1+\frac{1}{2}[\omega_{J_t}]^2)l$ is not necessarily ample. Recall that the Weierstrass model $\check{f}: \check{M}\rightarrow N$ of $f: M\rightarrow N$ is obtained by contracting the irreducible components of singular fibers of $f$, which do not intersect with the section $\sigma$ (cf. Chapter 7 in \cite{Fr1}). Denote by $\pi: M \rightarrow \check{M}$ the contraction morphism. Since $\pi$ contracts finitely many $(-2)$-curves, $ \check{M}$ has only orbifold A-D-E singularities.
\begin{prop}\label{lemma-mirror} Normalize $\Omega$ so that $[{\rm Im}\Omega]^2=4$. The SYZ mirror of $(M, \omega_{J_t}, \Omega_{J_t})$ with vanishing B-field is $(M, (l\cdot \tilde{\alpha}_t)^{-1} \check{\omega}, (l\cdot \tilde{\alpha}_t)^{-1}\check{\Omega})$ with the B-field $ \check{\mathbb{B}}_t$, where $$ \check{\Omega}= \pi^* \omega_{\check{M}}- i{\rm Im}\Omega, \ \ \ \check{\omega}={\rm Re}\Omega, \ \ {\rm and} $$ $$ \check{\mathbb{B}}_t=(l\cdot \tilde{\alpha}_t)^{-1} \tilde{\alpha}_t-[\sigma]+{\rm mod} (l). $$ Here $ \omega_{\check{M}}$ is the Ricci-flat K\"{a}hler-Einstein metric, possibly in the orbifold sense, such that $\pi^* \omega_{\check{M}} \in c_1(\mathcal{O}_{M}(\sigma (N)+3l))$.
\end{prop}
\begin{proof} Firstly, note that $([\sigma]+3l)^2= 4 >0$.
Now, let $D$ be an irreducible curve such that $( [\sigma]+3l)\cdot [D]\leq 0$. If $[D]\cdot l >0$, then $[\sigma]\cdot [D]<0$. Thus $D=\sigma$, and $( [\sigma]+3l)\cdot [D]=1>0$, which is a contradiction.
We obtain that $[D]\cdot l\leq 0$, and $D$ is an irreducible component of a fiber. Thus $[D]\cdot l=0$, and $[\sigma]\cdot [D]\leq 0$, which implies that $[\sigma]\cdot [D]=0$, and $D$ is an irreducible component of a singular fiber of $f$ which does not intersect with $\sigma$. Therefore $[\sigma]+3l$ is nef and big, and an irreducible curve $D$ satisfies $( [\sigma]+3l)\cdot [D]= 0$ if and only if $D$ is an irreducible component of a singular fiber of $f$ which does not intersect with $\sigma$. There is an ample class $\alpha_{\check{M}}$ on the Weierstrass model $\check{M}$ such that
$[\sigma]+3l=\pi^* \alpha_{\check{M}}$, and by \cite{KT}, there exists a unique Ricci-flat K\"{a}hler-Einstein metric $\omega_{\check{M}}\in \alpha_{\check{M}}$ on $\check{M}$ in the orbifold sense.
Since $[\pi^*\omega_{\check{M}}]^2=([\sigma]+3l)^2= [{\rm Im}\Omega]^2= [{\rm Re}\Omega]^2, $ $(\pi^*\omega_{\check{M}}, {\rm Re}\Omega, {\rm Im}\Omega)$ is a HyperK\"{a}hler triple on $\pi^{-1}(\check{M}_{reg})$.
By using the HyperK\"{a}hler rotation, we can find new complex structure $K$, and define a family of HyperK\"{a}hler structures
$$ \check{\Omega}_{t}= (l\cdot \tilde{\alpha}_t)^{-1}(\pi^* \omega_{\check{M}}-i{\rm Im}\Omega), \ \ \ \check{\omega}_t=(l\cdot \tilde{\alpha}_t)^{-1} {\rm Re}\Omega, $$
which satisfy
$$ [\check{\Omega}_{t}]= (l\cdot [{\rm Re}\Omega_{J_t}])^{-1}([\sigma]+3l-i[\omega_{J_t}]), \ \ {\rm and} \ \ [\check{\omega}_t]=(l\cdot [{\rm Re}\Omega_{J_t}])^{-1}[{\rm Im}\Omega_{J_t}].$$ By letting $$ \check{\mathbb{B}}_t=(l\cdot [{\rm Re}\Omega_{J_t}] )^{-1} [{\rm Re}\Omega_{J_t}]-[\sigma]+{\rm mod} (l), $$ the proof of Theorem 7.3 in \cite{GroII} shows that
$(M, \check{\omega}_{t}, \check{\Omega}_t)$ with $\check{\mathbb{B}}_t$ is the SYZ
mirror of $(M, \omega_{J_t}, \Omega_{J_t})$, i.e. $(f: M_{N_0}\rightarrow N_0, \check{\omega}_{t}, \check{\Omega}_t)$ is the dual special Lagrangian fibration of $(f: M_{N_0}\rightarrow N_0, \omega_{J_t}, \Omega_{J_t})$.
\end{proof}
We now assume that $M$ satisfies the hypotheses of Theorem \ref{thm-main}, which gives $M=\check{M}$ and $\pi$ is the identity. We can now see how Theorem \ref{thm-main} applies to Conjecture 5.5 in \cite{Fuk}. In our setup, the anti-self-dual connection $\Xi_t$ and the complex structure $J_t$ induce a holomorphic structure on $\mathcal{V}$ for any $t\in (0,1]$, and $\Xi_t$ satisfies the Hermitian-Yang-Mills equation $$ F_{\Xi_t}\wedge \omega_{J_t} =0, \ \ {\rm and} \ \ F_{\Xi_t}\wedge \Omega_{J_t} =0.$$ The spectral cover $D_t$ and the limit $D_0$ are special Lagrangian cycles with respect to the mirror HyperK\"{a}hler structure $( \check{\omega}, \check{\Omega})$. We now rephrase Theorem \ref{thm-main} and Theorem \ref{thm-main2} in the context of SYZ mirror symmetry.
\begin{thm}\label{conj} Under the assumptions of Theorem \ref{thm-main}, for any sequence $t_k \rightarrow 0$ and any $p>2$, there exists an open dense subset $N^o \subset N_0$, a subsequence (still denoted $t_k$), a sequence of $L^{p}_2$ unitary gauge changes $u_k$ of $P$, and a $L^{p}_1$ $SU(n)$-connection $\Xi_0$ on $P|_{M_{N^o}}$ so that
$$u_k(\Xi_{t_k}) \rightarrow \Xi_0 $$ in the locally $L^{p}_1$ sense on $M_{N^o}$. Here the norms are calculated by using a fixed metric on $M$.
For any $w \in N^o$, the restriction of $\Xi_0$ to the fiber $M_w$, denoted $\Xi_0|_{M_w}$, is $C^{1,\alpha}$ gauge equivalent to a smooth flat $SU(n)$-connection
$$u_\infty ( \Xi_0|_{M_w} )= \frac{\pi}{{\rm Im} (\tau)}({\rm diag}\{q_{1}(w), \cdots, q_{n}(w)\}d\bar{z}-{\rm diag}\{\bar{q}_{1}(w), \cdots, \bar{q}_{n}(w)\}dz), $$
where $u_\infty\in\mathcal{G}^{1,\alpha}(M_w)$,
$M_w \cong \mathbb{C}/\Lambda_{\tau}$, $\Lambda_{\tau}={\rm Span}_{\mathbb{Z}}\{1, \tau\}$, $\sigma(w)=0$,
and $z$ denotes the coordinate on $\mathbb{C}$.
As $w$ varies, $\{q_{1}(w), \cdots, q_{n}(w)\}\subset M_w$ forms a special Lagrangian multisection of $f^{-1}(N^o) \rightarrow N^o$ with respect to the SYZ mirror HyperK\"{a}hler structure $( \check{\omega}, \check{\Omega})$, and its closure $D_{0}^o$ is a special Lagrangian cycle, i.e. $$\check{\omega}|_{D_{0}^o}\equiv0, \ \ {\rm and} \ \ {\rm Im}\check{\Omega}|_{D_{0}^o}\equiv0. $$
The family of special Lagrangian submanifolds $D_t$ with respect to $( \check{\omega}, \check{\Omega})$ converges to $D_{0}^o$ on $f^{-1}(N^o)$ in the locally Hausdorff sense.
Furthermore,
the unitary gauge equivalent class of the limit connection $\Xi_0$ is the Fourier-Mukai transform of a flat $U(1)$-connection $\Theta$ on $D_{0}^o\cap M_{N^o}$, i.e. $$\Xi_0 \in\mathcal{FM}(D_{0}^o\cap M_{N^o}, \Theta).$$
\end{thm}
Conversely, if $D$ is a smooth special Lagrangian submanifold with respect to $( \check{\omega}, \check{\Omega})$ on $M$ such that $D$ represents $n [\sigma ]+m l \in H_2(M, \mathbb{Z})$ for some $m \in \mathbb{N}$, and $\Theta$ is a flat $U(1)$-connection on $D$, then $D$ is a smooth holomorphic curve in $M$.
The argument in Section 3.1 shows that there is a stable bundle $V$ of rank $n$ with respect to $f^{*} c_{1}(\mathcal{O}_{\mathbb{CP}^{1}}(1))+t \alpha$ for $0<t \ll 1$. The anti-self-dual connections $\Xi_t$ on $V$ are also Hermitian-Yang-Mills with respect to $(\omega_{J_t}, \Omega_{J_t})$.
In the context of mirror symmetry, a special Lagrangian submanifold with a flat $U(1)$-connection is called an A-cycle, and a Hermitian-Yang-Mills connection on a complex submanifold is called a B-cycle (cf. \cite{LYZ,JY,Va}). The correspondence between B-cycles and A-cycles is motivated by the study of homological mirror symmetry via the SYZ construction in \cite{AP,Fuk1,Fuk}, and the extended mirror symmetry with bundles \cite{LYZ,Va}.
Theorem \ref{conj} says that in the current case, the adiabatic limit of B-cycles is corresponding to an A-cycle on the mirror K3 surface.
\subsection{Remarks}
We conclude this section with a few more remarks.
\begin{rk}\label{remark2} \rm Note that the Levi-Civita connection of the Ricci-flat K\"{a}hler-Einstein metric $\omega_t$ is an anti-self-dual connection. However Theorem \ref{thm-main} does not apply to this case due to the following. If $M_w$ is a smooth fiber, then the restriction of the tangent bundle of $M$ satisfies a short exact sequence $$0 \rightarrow TM_w \rightarrow TM|_{M_w} \rightarrow f^*T_w N \rightarrow0,$$ and $TM|_{M_w}$ is S-equivalent to $\mathcal{O}_{M_w}\oplus \mathcal{O}_{M_w}$. Thus the special cover of $TM$ is $D_{TM}=2\sigma(N)$, and is not reduced. Consequently, the hypotheses of Theorem \ref{thm-main} are not satisfied.
The curvature $F_{\Xi_t}$ in Theorem \ref{thm-main} behaves very differently from the curvature of the Ricci-flat K\"{a}hler-Einstein metric $\omega_t$. In the metric case, the curvature $R_{\omega_t}$ of $\omega_t$ is bounded away from the singular fibers along the collapsing of $\omega_t$, i.e. $$\sup_{M_K}| R_{\omega_t}|_{\omega_t}\leq C_K,$$ for any compact subset $K\subset N_0$, by \cite{GW,GTZ}. Furthermore, there is a more general result in \cite{CT} that asserts the boundedness of curvatures of sufficiently collapsed Ricci-flat Riemannian Einstein metrics $\mathrm{g}$ on 4-manifolds away from finite metric balls. The readers are referred to \cite{CT} for details.
In Theorem \ref{thm-main},
it is shown that the curvature $F_{\Xi_t}$ is bounded with respect to any fixed metric on $M_U$.
However,
$F_{\Xi_t}$ can not be bounded with respect to the collapsed metric $\omega_t$ as the following demonstrates. If it were bounded, then Proposition \ref{prop2+0} of Section 7 shows that on any $U\subset N^o$, \bea \int_{U}\sum_{j=1,2}\|\partial_{x_j}A_{0,t}\|_w^2 dx_1dx_2 & \leq & C( \|F_{\Xi_t}\|_{L^2(M_{U}, \omega_t)}^2 +t)\nonumber\\ & \leq & C (\sup_{M_U}| F_{\Xi_t}|_{\omega_t}^2{\rm Vol}_{\omega_t}(M_U)+t) \nonumber \\ & \leq & C t \rightarrow 0, \nonumber\eea where $x_1$ and $x_2$ are coordinates on $U$, which implies $\partial_{x_j}A_{0}\equiv 0$, $j=1,2$. Thus $\partial_{x_j}({\rm Im} (\tau)^{-1} q_i(w))\equiv 0$, $j=1,2$, and $q_i(w)=c_i (\tau(w) -\bar{\tau}(w))$ for constants $c_i\in \mathbb{C}$, $i=1, \cdots, n$. Note that $q_i(w)$ is holomorphic, and $\tau(w)$ is not constant as the fibration $f$ is a Weierstrass fibration. We have $c_i=0$ and $q_i(w)\equiv 0$, $i=1, \cdots, n$. Hence $D_0^o\cap M_U=n \sigma (U)$, which contradicts to the assumption of $D_0^o$ being reduced.
\end{rk}
\begin{rk}\label{remark3-}\rm Theorem \ref{thm-main} is a compactness result, i.e. the convergence of $\Xi_t $ occurs along subsequences $t_k$. The convergence along the parameter $t$ may hold under certain stronger assumptions, for example the follows.
For any $t\ll 1$, we assume that $V_{t}|_{M_w}$ is regular semi-stable for any $w\in N$.
As in Section 2.4, Proposition 5.7 of \cite{FMW} shows that $$V_t= (\nu_{D_t})_* (\mathcal{O}_{\tilde{M}}(\Delta_t - \Sigma_{D_t})\otimes \tilde{f}^*\tilde{L}_t)$$ for a line bundle $\tilde{L}_t$ on $D_t$. If we assume further that $\tilde{L}_t$ converges to a $\tilde{L}_0$ on $D_0$ as divisors along the convergence of $D_t$ to $D_0$, then we expect that $\Xi_t $ converges away from finite fibers without passing to any subsequence, which would be left for the future study.
\end{rk}
\begin{rk}\label{remark3} \rm There are many more questions that the authors would like to investigate in the future. Firstly, we would like to understand what are the corresponding algebraic geometric descriptions of the type $I$ and type $II$ bubbles in the proof of Proposition \ref{estimate 1}. Secondly, we like to have an explicit formula for the second Chern number $c_2(\mathcal{V} )$ via the bubbles and the limit special cover $D_0$. Here a certain bubble tree convergence is expected.
Finally, we like to study the metric geometry of the moduli space of anti-self-dual Yang-Mills connections on collapsed K3 surfaces, inspired by the F-theory/heterotic string theory duality as in \cite{FMW0}. For any $0<t \leq (\frac{n^3}{4}c)^{-1}$, let $\mathfrak{M}_t(n, c)$ be the moduli space of anti-self-dual connections on $\mathcal{V}$ with respect to the HyperK\"{a}hler structure $(\omega_t, \Omega)$, where $c=c_2 (\mathcal{V})$, which is not empty (cf. Theorem \ref{stable criterion}). By Theorem 7.10 in \cite{Kob}, $(\omega_t, \Omega)$ induces a HyperK\"{a}hler structure $(\omega_{\mathfrak{M},t}, \Omega_{\mathfrak{M},t}) $ on the regular locus $\mathfrak{M}_t(n, c)^o$ of $\mathfrak{M}_t(n, c)$. Furthermore, it is expected that there is a holomorphic lagrangian fibration $\mathfrak{f}:\mathfrak{M}_t(n, c)^o\rightarrow \mathfrak{U}\subset |n \sigma (N)+ml|$ (cf. Section 2.4 of \cite{FMW0}). For example, if $D\in |n \sigma (N)+ml|$ is smooth, then the fiber $\mathfrak{f}^{-1}(D)$ is the Jacobian $\mathfrak{J}(D)$ of $D$, which parameterises the flat $U(1)$-connections on $D$. We would investigate the degeneration behavior of $(\omega_{\mathfrak{M},t}, \Omega_{\mathfrak{M},t}) $ when $t\rightarrow0$ in future study.
\end{rk}
\section{The proof of Theorem \ref{thm-main}}
\label{mainproof}
In this section we prove Theorem \ref{thm-main}, assuming some important estimates which will be proved in the subsequent sections. We begin with a bubbling result, which gives a decay estimate for curvature away from a finite set. This set may depend on the chosen sequence of times $t_k\rightarrow 0$.
Since we are interested in the behavior of the restriction of the connections $\Xi_{t_k}$ to a fiber $M_w$, we use the notation $A_{t_k}(w)=\Xi_{t_k}|_{M_w}$. In general we write this fiberwise connection as $A_{t_k}$, as the dependence on $w$ is clear from context.
\begin{prop}
\label{estimate 1} If $\Xi_t$
is a family of anti-self-dual connections on $P$ with respect to $(\omega_t, \Omega)$, then for any sequence $t_k \rightarrow 0$, there is a Zariski open subset $N_1 \subset N_0$, and a subsequence (still denoted $t_k$), so that the curvature $F_{\Xi_{t_k}}$ satisfies \be
\sup_{M_K}|F_{\Xi_{t_k}}|_{\omega_{t_k}}\leq \frac{\epsilon_k}{t_k} \nonumber
\ee
on any compact subset $K \subset N_1$. Here the constants $\epsilon_k$ may depend on $K$, and satisfy $\epsilon_k \rightarrow 0$ as $k \rightarrow \infty$. Consequently, for any $w\in K$ and $t_k \ll 1$,
$$\|F_{A_{t_k}}\|_{C^0(\omega^{SF}|_{M_w})}\rightarrow 0,$$ and
$V_{t_k}|_{M_w}$ is semi-stable.
\end{prop}
Note that the above assumptions are slightly weaker than those used in Theorem \ref{thm-main}. To prove the proposition, we follow a bubbling argument similar to arguments seen previously (for example \cite{DS}), however we present the details here for completeness.
\begin{proof} Suppose that there exists a sequence of points $p_k \in M$ so that $f(p_k) \rightarrow w$ in $N_0$, and furthermore
\begin{equation}\label{eq01}\liminf_{k \rightarrow\infty}t_k|F_{\Xi_{t_k}}|_{\omega_{t_k}}(p_k) >0. \end{equation} We claim that there is a universal constant $\varepsilon>0$ such that for any neighborhood $U_w$ of $w$, $$ \int_{M_{U_w}}|F_{\Xi_{t_k}}|_{\omega_{t_k}}^2\omega_{t_k}^2 \geq \varepsilon,$$ for $k \gg 1$. Once this is demonstrated, by \eqref{cw0} there can only be a finite number of such $w$.
By \cite{GTZ}, for some $p\in M_w$ we have
$$(M, t_k^{-1} \omega_{t_k}, p_k) \rightarrow (M_x \times \mathbb{C}, \omega_\infty =\omega_{w}^{F}+\frac{i}{2}W^{-1}(w)d\tilde{w}\wedge d\bar{\tilde{w}}, p)$$ in the pointed $C^\infty$-Cheeger-Gromov sense, where $\omega_{w}^{F}$ is the flat K\"{a}hler metric representing $\alpha|_{M_w}$, i.e. $\omega_{w}^{F}=\omega^{SF}|_{M_w} $, and $\tilde{w}$ denotes the scaled coordinate of $ \mathbb{C}$ (see Section 2.4). More precisely, if
$D_r=\{\tilde{w} \in \mathbb{C}| |\tilde{w}|<r\}$, there are smooth embeddings $\Phi_{t_k,r}: M_w \times D_r \rightarrow M_U$
such that $$\Phi_{t_k,r}^* t_k^{-1}\omega_{t_k} \rightarrow \omega_{\infty}, \ \ \ \Phi_{t_k,r}^* I\Phi_{t_k,r,*} \rightarrow I_{\infty},$$ in the $C^{\infty}$-sense on $M_w \times D_r$, where $I$ (resp. $I_\infty$) denotes the complex structure on $M$ (resp. $M_w \times \mathbb{C}$).
We have two cases. In the first case, for any compact subset $K\subset M_w \times \mathbb{C}$, there is a constant $C_K>0$ such that $$|F_{\Xi_{t_k}}|_{t_k^{-1}\omega_{t_k}}=t_k|F_{\Xi_{t_k}}|_{\omega_{t_k}}\leq C_K,$$ on $\Phi_{t_k,r}(K)$, $r\gg 1$. By passing a subsequence, Uhlenbeck's strong compactness theorem shows that there is a sequence of unitary gauge transformations $u_{K,k}$, and an anti-self-dual $SU(n)$-connection $\Xi_\infty$ on $M_w \times \mathbb{C}$ such that
$u_{K,k}(\Phi_{t_k,r}^*\Xi_{t_k})$ converges to $\Xi_\infty$ in the locally $C^\infty$-sense on $K$. Thus, in the $C^0$-sense on $K$,
$$ \Phi_{t_k,r}^*|F_{\Xi_{t_k}}|_{t_k^{-1}\omega_{t_k}} \rightarrow |F_{\Xi_\infty}|_{\omega_\infty}, \ \ {\rm and} \ \
|F_{\Xi_\infty}|_{\omega_\infty}(p)>0. $$
By \cite{Weh}, there is a constant $\mu=\mu(n)$ depending only on the group $SU(n)$, such that $$\int_{M_x \times \mathbb{C}} |F_{\Xi_\infty}|_{\omega_\infty}^2\omega_\infty^2 \geq \mu.$$ Furthermore if $n=2$, we know $\mu (2)=4\pi^2$. This is called the bubble of type $II$ in \cite{DS}. By choosing $K$ large enough, $$ \int_{M_{U_w}}|F_{\Xi_{t_k}}|_{\omega_{t_k}}^2\omega_{t_k}^2 \geq \int_{\Phi_{t_k,r}(K)}|F_{\Xi_{t_k}}|_{t_k^{-1}\omega_{t_k}}^2t_k^{-2}\omega_{t_k}^2\geq \frac{\mu}{2},$$ for $k\gg 1$.
The second case is that there are $p_k'\in M$ such that $$d_{t_k^{-1}\omega_{t_k}}(p_k,p_k')<C< \infty, \ \ {\rm and }, \ \ t_k|F_{\Xi_{t_k}}|_{\omega_{t_k}}(p_k')\rightarrow\infty,$$ when $k \rightarrow\infty$. In order to preform the bubbling argument, recall the following point choosing lemma.
\begin{lem}[Lemma 9.3 in \cite{DS}]
\label{pointpick}
Let $(Y,d_Y)$ be a complete metric space, and $\zeta$ be a continuous non-negative function. For any $y\in Y$, there exist $y' \in Y$ and $0<\rho \leq 1$ such that $$d_Y(y,y')\leq 1, \ \ \sup_{B_{d_Y}(y',\rho)}\zeta \leq 2 \zeta (y'), \ \ {\rm and} \ \ 2\rho \zeta(y')\geq \zeta(y).$$
\end{lem}
We apply this lemma to $\zeta=|F_{\Xi_{t_k}}|_{t_k^{-1}\omega_{t_k}}$, $y=p_k'$, and obtain $y'=p_k''$ and $0\leq \rho \leq 1$.
We further rescale the metric, and $(M, |F_{\Xi_{t_k}}|_{\omega_{t_k}}^{-1}(p_k'')\omega_{t_k}, p_k'')$ converges to the standard Euclidean space $(\mathbb{C}^2,\omega_E, 0)$ in the smooth Cheeger-Gromov sense by passing to a subsequence. The same argument as above shows that $\Xi_{t_k}$ smoothly converges to an non-trivial anti-self-dual $SU(n)$-connection $\Xi_\infty'$ on $\mathbb{C}^2$ by passing to certain unitary gauge changes and subsequences. We now have $$\int_{\mathbb{C}^2} |F_{\Xi_\infty'}|_{\omega_E}^2\omega_E^2 \geq \tau,$$ where $\tau$ is the constant in Lemma \ref{energybound}. This is called a bubble of type $I$, and is standard in the study of Yang-Mills fields on 4-manifolds (cf. \cite{DK,FU}). Just as above, $$ \int_{M_{U_w}}|F_{\Xi_{t_k}}|_{\omega_{t_k}}^2\omega_{t_k}^2 \geq \int_{\Phi_{K,k}(K)}|F_{\Xi_{t_k}}|_{t_k^{-1}\omega_{t_k}}^2t_k^{-2}\omega_{t_k}^2\geq \frac{\tau}{2},$$ for $k\gg 1$, where $K$ satisfies that $p_k'\in \Phi_{K,k}(K)$. We obtain the claim by letting $\varepsilon= \frac{1}{2}\min \{\mu, \tau\}$.
Let $S_1$ be the set of points $x\in N_0$ for which there is a sequence $p_k \in M$ such that $f(p_k) \rightarrow w$ in $N_0$, and (\ref{eq01}) is satisfied. By \eqref{cw0} $$ 8 \pi^2 c_2(\mathcal{V})= \lim_{k\rightarrow\infty}\int_{M}|F_{\Xi_{t_k}}|_{\omega_{t_k}}^2\omega_{t_k}^2\geq \sharp (S_1)\varepsilon,$$ and as a result $S_1$ is a finite set. Therefore $N_1=N_0\backslash S_1$ is a Zariski open subset, and for any compact subset $K\subset N_1$,
$$
\sup_{M_K}t_k|F_{\Xi_{t_k}}|_{\omega_{t_k}}\leq \epsilon_k \rightarrow 0,
$$ when $k \rightarrow \infty$.
Since $\Phi_{t_k,r}^*t_k^{-1}\omega_{t_k}$ converges smoothly to $\omega_\infty$ on $M_w\in \mathbb{C}$ for $w\in K$, we have $$ \|F_{A_{t_k}}\|_{C^0(\omega^F)}\leq 2 \|F_{A_{t_k}}\|_{C^0(t_k^{-1}\omega_{t_k}|_{M_w})}\leq 2 \sup_{M_K}|F_{\Xi_{t_k}}|_{t_k^{-1}\omega_{t_k}}\rightarrow 0. $$ By Proposition \ref{semistablelemma},
$V_{t_k}|_{M_w}$ is semi-stable, where as above $V_{t_k}$ denotes $\mathcal{V}$ equipped with the holomorphic structure induced by $\Xi_{t_k}$.
\end{proof}
Restricting to a fiber $M_w$, by the above proposition, weak Uhlenbeck compactness gives that for any $p>2$, there exists a sequence of unitary gauge $u_{w,k}$ such that along a subsequence of times, $u_{w,k}(A_{t_k})$ converges in $L^p_1$ to a flat $L^p_1$-connection $\Xi_{\infty,w}$ on $M_w$. In other words, we have fiberwise convergence of $\Xi_{t_k}$ up to gauge changes. However, it is not clear yet that $\Xi_{t_k}$ has any limit when $t_k \rightarrow 0$ on $M_K$. For this, we need the stronger assumptions in Theorem \ref{thm-main}, and further results and estimates.
We now work under the setup of Theorem \ref{thm-main}, and consider a sequence of connections $\Xi_{t_k}$ where $t_k\rightarrow 0$ as $k\rightarrow\infty$. Before we turn to the key estimates, we need to describe the explicit form of the holomorphic structure of the bundle $V_t$ in a local trivialization.
Note that $f|_{D_{0}^o}: D_{0}^o \rightarrow N$ is an $n$-sheeted branched covering. If $S_{D_{0}^o}$ denotes the subset of $D_{0}^o$ consisting all singular points of $D_{0}^o$ and all branch points of $f|_{D_{0}^o}$, then $f(S_{D_{0}^o})$ is a finite subset of $N$.
We define a Zariski open subset \begin{equation}\label{subset}N^o=N_1\backslash (f(D_0-D_{0}^o)\cup f(S_{D_{0}^o})). \end{equation}
On $N^o$, $f|_{D_{0}^o}: f|_{D_{0}^o}^{-1}(N^o) \rightarrow N^o$ is an $n$-sheeted unbranched covering, since $D_{0}^o$ is reduced. For any $w\in N^o$, $D_{0}^o \cap M_w$ consists $n$ distinct points in $M_w$, i.e. $D_{0}^o \cap M_w= \{q_1, \cdots, q_n\}$ where $q_i \neq q_j$ for any $i\neq j$. The trivial bundle $\mathcal{V}|_{M_w}$ equipped with the holomorphic structure induced by $D_{0}^o \cap M_w$ is isomorphic to the flat holomorphic bundle
$$ \mathcal{O}_{M_w}(q_1 - \sigma (w))\oplus \cdots \oplus \mathcal{O}_{M_x}(q_n - \sigma (w)).$$
Since $D_{t}$ converges to $D_0$ and $D_0-D_{0}^o \in |(m-m')l|$ is supported on fibers, for any compact subset $K\subset N^o$ we have that $f: D_{t} \cap M_{K} \rightarrow K$ is an $n$-sheeted unbranched covering for $t\ll 1$. For any $w\in K$, $D_{t} \cap M_{w}=\{q_{1,t}, \cdots, q_{n,t}\}$ such that $q_{i,t} \neq q_{j,t}$ for any $i\neq j$, and $q_{i,t} \rightarrow q_{i}$ when $t \rightarrow 0$. Furthermore, $V_{t}|_{M_w}$ is semi-stable, which implies that $V_{t}|_{M_w}$ is regular by Proposition 6.4 in \cite{FMW}, and $$ V_{t}|_{M_w} \cong \mathcal{O}_{M_w}(q_{1,t} - \sigma (w))\oplus \cdots \oplus \mathcal{O}_{M_w}(q_{n,t} - \sigma (w)).$$
For any $t\ll 1$, there is a Zariski open subset $N_t^o \supset K$ such that $V_{t}|_{M_w}$, $w\in N_t^o$, is regular semi-stable.
Proposition 5.7 of \cite{FMW} asserts that $$V_t|_{M_{N_t^o}}= (\nu_{D_t})_* (\mathcal{O}_{\tilde{M}_{N_t^o}}(\Delta_t - \Sigma_{D_t})\otimes \tilde{f}^*\tilde{L}_t)$$ for a certain line bundle $\tilde{L}_t$ on $D_t\cap M_{N_t^o}$. Here, as in Section 2.4, $$\nu_{D_t}:\tilde{M}_{N_t^o}=D_t\times_{N_t^o} M\rightarrow M_{N_t^o},$$
$\Sigma_{D_t} = \nu_{D_t}^* \sigma,$ and $\Delta_t= \tilde{M}_{N_t^o} \cap \Delta_0$ for the diagonal $\Delta_0$ of $ M\times_{N_t^o} M$ via the natural embedding $ \tilde{M}_{N_t^o} =D_t\times_{N_t^o} M \hookrightarrow M\times_{N_t^o} M$.
Let $U\subset K \subset N_t^o$ be an open subset biholomorphic to the unit disk, and $w$ be a coordinate on $U$. Then $M_U\cong (U\times \mathbb{C})/\Lambda$ for lattice subbundle $\Lambda={\rm Span}_{\mathbb{Z}}\{1, \tau\}$, where $\tau=\tau(w)$ varies holomorphically and is the period of the elliptic curve $M_w$. Furthermore under this identification the section $\sigma$ satisfies $\sigma \equiv 0$. If $z$ is the coordinate on $\mathbb{C}$, we define real functions $y_1$ and $y_2$ on $U\times \mathbb{C}$ by $z=y_1+\tau y_2$. Then $dy_1$ and $dy_2$ are well-defined 1-forms on $M_U$, and we have the decomposition of cotangent bundle $T^*M_U\cong {\rm Span}_{\mathbb{R}}\{dy_1,dy_2\}\oplus {\rm Span}_{\mathbb{R}}\{dx_1,dx_2\}$, where $w=x_1+ix_2$. Let $\theta=dy_1+\tau dy_2$, whose restriction $\theta|_{M_w}=dz$ on any fiber $M_w$. Note that
$\overline{\partial}\tau=0$, $d\tau=\partial \tau$ and $0=\overline{\partial}\tau_1+i\overline{\partial}\tau_2$, where
$\tau=\tau_1+i\tau_2$. Thus $dz=dy_1+\tau dy_2+y_2 d\tau$, $2i\tau_2y_2=z-\bar{z}$, and $\theta=dz-\frac{z-\bar{z}}{2i\tau_2}\partial_w \tau dw=dz+bdw.$
We fix the trivializations $P|_{M_U}\cong M_U \times SU(n)$ and $\mathcal{V}|_{M_U}\cong M_U \times \mathbb{C}^n$. The unitary gauge group consists of $SU(n)$ valued functions, in other words $\mathcal{G}=C^\infty(M_U, SU(n))$, and the complex gauge group is $\mathcal{G}_{\mathbb{C}}=C^\infty(M_U, SL(n, \mathbb{C}))$ under this trivialization. The respective Lie algebras are $\mathfrak{g}=C^\infty(M_U, \mathfrak{su}(n))$ and $\mathfrak{g}_{\mathbb{C}}=C^\infty(M_U, \mathfrak{sl}(n, \mathbb{C}))$. Note that there is the decomposition $\mathfrak{g}_{\mathbb{C}}=\mathfrak{g}\oplus i\mathfrak{g}$ induced by $\mathfrak{sl}(n, \mathbb{C})=\mathfrak{su}(n)+i\mathfrak{su}(n)$, and if $s\in \mathfrak{g}_{\mathbb{C}}$ is Hermitian (given by $s^*=s$), then $s\in i\mathfrak{g}$. Therefore any complex gauge $g$ can be written as $g=\exp(v+s)$, for a certain $v\in \mathfrak{g}$ and an $s \in i \mathfrak{g}$.
Note that $D_{0}^o \cap M_U$ (resp. $D_{t} \cap M_U$) is given by n distinct holomorphic functions $q_j(w)$ (resp. $q_{j,t}(w)$), and for any $j$, $q_{j,t}(w)\rightarrow q_{j}(w)$ in the $C^{\infty}$-sense as $t\rightarrow 0$. Thus $D_t \cap M_U$ consists of $ n $ distinct unit disks, and because $\tilde{L}_t|_{D_t \cap M_U}$ is holomorphically trivial, we obtain $$V_t|_{M_{U}}\cong \bigoplus_{j=1}^{n} \mathcal{O}_{M_{U}}(q_{j,t} (U)- \sigma (U)). $$
We define the background connections on the trivial bundle $\mathcal{V}|_{M_U}$ \begin{equation}\label{bconnections} A_{0,t}= \pi ({\rm Im} (\tau))^{-1} ({\rm diag}\{q_{1,t}, \cdots, q_{n,t}\}\bar{\theta}-{\rm diag}\{\bar{q}_{1,t}, \cdots, \bar{q}_{n,t}\}\theta), \end{equation} \begin{equation}\label{bconnections2} A_{0}= \pi ( {\rm Im} (\tau))^{-1} ({\rm diag}\{q_{1}, \cdots, q_{n}\}\bar{\theta}-{\rm diag}\{\bar{q}_{1}, \cdots, \bar{q}_{n}\}\theta). \end{equation} Thus $ A_{0,t} \rightarrow A_{0}$ in the $C^\infty$-sense when $t\rightarrow 0$, $V_t|_{M_w}$ is isomorphic to $\mathcal{V}|_{M_w}$ equipped with the holomorphic structure induced by the flat connection $A_{0,t}|_{M_w}$, and $A_{0}|_{M_w}$ induces the holomorphic bundle structure $\bigoplus \limits_{i=1}^{n} \mathcal{O}_{M_w}(q_{i}(w)-\sigma(w))$.
\begin{lem}
\label{background connection} The unitary connection $A_{0,t}$ on $\mathcal{V}|_{M_U}$ induces the holomorphic structure isomorphic to $V_t|_{M_U}$.
\end{lem}
\begin{proof}
In general, if $L$ is a holomorphic line bundle, and $h$ determines a Hermitian metric on $L$ in a local holomorphic trivialization, then the unique Chern connection is given by $A_h=\partial \log h$. If $\rho$ is a local unitary frame, i.e. $|\rho|_h^2=h|\rho|^2\equiv 1$, then we have smooth trivialization of $L$ via $\rho\mapsto 1$, and under such trivialization, $A_h$ is transformed to $A=\overline{\partial}\log \rho- \partial \log \bar{\rho}$. A different choice of $\rho$ gives a unitary gauge transformation of $A$.
Note that the holomorphic line bundle $ \mathcal{O}_{M_{U}}(q_{j,t} (U)- \sigma (U))$ is given by the multiplier $\{e_1\equiv 1, e_\tau =\exp (-2\pi i q_{j,t}(w))\}$, i.e. $ \mathcal{O}_{M_{U}}(q_{j,t} (U)- \sigma (U))$ is obtained by the quotient of $U\times \mathbb{C} \times \mathbb{C}$ via
$$ (w,z, \xi) \sim (w,z+1, e_1 \xi), \ \ \ (w,z, \xi) \sim (w,z+\tau, e_\tau \xi)$$
(cf. Section 6 in Chapter 2 of \cite{GH}). On $U\times \mathbb{C}$, if we let $$h=\exp \pi \big({\rm Im} (\tau)^{-1}(z-\bar{z})(q_{j,t}-\bar{q}_{j,t})\big), $$ then $h(w,z+1)=h(w,z)$ and $h(w,z+\tau)=|\exp (2\pi i q_{j,t}(w)) |^2h(w,z)$, and thus $h$ defines a Hermitian metric on $ \mathcal{O}_{M_{U}}(q_{j,t} (U)- \sigma (U))$. If $$\rho=\exp \big( -\pi {\rm Im} (\tau)^{-1} (z-\bar{z})q_{j,t}\big) , $$ then $h|\rho|^2=1$, $\rho(w,z+1)=\rho(w,z)$ and $\rho(w,z+\tau)=e_\tau \rho(w,z)$. Thus $\rho$ is a global
unitary frame, and under the trivialization induced by $\rho$, the Chern connection $\Xi_{0,t,j}=\Xi^{1,0}_j+\Xi^{0,1}_j$ is given by $\Xi^{1,0}_j=- \overline{\Xi^{0,1}_j} $ and $$\Xi^{0,1}_j=\overline{\partial}\log \rho=\pi {\rm Im} (\tau)^{-1}q_{j,t}d\bar{z}-\pi (z-\bar{z})q_{j,t}\overline{\partial}{\rm Im} (\tau)^{-1}=\pi {\rm Im} (\tau)^{-1}q_{j,t}\bar{\theta}, $$ by $$\bar{\theta}=d\bar{z}-\frac{z-\bar{z}}{2i{\rm Im} (\tau)}\partial_{\bar{w}} \bar{\tau} d\bar{w}=d\bar{z}+\frac{z-\bar{z}}{{\rm Im} (\tau)}\partial_{\bar{w}} {\rm Im} (\tau) d\bar{w}.$$ We obtain the desired conclusion.
\end{proof}
Since $\Xi_t$ and $A_{0,t}$ induce the same holomorphic structure on $\mathcal{V}|_{M_U}$ over $M_U$, there is a complex gauge $g\in \mathcal{G}_{\mathbb{C}}$ such that $g(\Xi_t)=A_{0,t}$. Note that $gg^*$ is Hermitian, and $gg^*=e^{2s_t}$ for some $s_t \in C^\infty (M_U, \mathfrak{sl}(n, \mathbb{C}))$ with $s_t^*=s_t$. If we let $u=e^{-s_t}g$, then $u^*=u^{-1}$, i.e. $u$ is a unitary gauge, and $g=e^{s_t} u$. Therefore, by a further unitary gauge change if necessary, we assume that
\be
\label{hermitiangauge}
e^{s_t}(\Xi_t)=A_{0,t}
\ee
for a Hermitian gauge $e^{s_t}$ on $M_U$.
In order to prove the main theorem, we need to improve the curvature estimates of Proposition \ref{estimate 1}.
\begin{prop}
\label{type3bubbling} For any compact set $K\subset N^o$, there exists a constant $C_K$ such that
\begin{equation}\label{eq02}
\sup_{M_K}|F_{\Xi_{t_k}}|_{\omega_{t_k}}\leq C_K {t_k}^{-\frac12}.\nonumber
\end{equation}
\end{prop}
The proof of this proposition can be found in Section 7. This implies the subsequence of connections $\Xi_{t_k}$ satisfies \eqref{hyp:curv}, which is the main assumption of
Proposition \ref{prop2} in Section \ref{C0boundsection}. Thus we can apply Proposition \ref{prop2} to $\Xi_{t_k}$ and achieve uniform $C^0$ control of the curvature, from which we conclude:
\begin{prop}
\label{thm01}
Along the sequence of connection $\Xi_{t_k}$, there exists a constant $C_1>0$ such that $$ \|F_{A_{t_k}}\|_{C^0(M_w)} \leq C_1t_k, \ \ and \ \ \|F_{\Xi_{t_k}}\|_{C^0(M_K)} \leq C_1, $$ for any $w\in K$. Consequently, for any $p>2$, by the weak Uhlenbeck compactness theorem \cite{U2} there exists a subsequence (still denoted $t_k$), a sequence of unitary gauge transformations $u_k\in \mathcal{G}^{2,p}$, and a limiting $L^p_1$ connection $\Xi_\infty$, so that $$ u_k (\Xi_{t_k} ) \rightarrow \Xi_\infty$$ in $L^p_1(M_K)$. Here all norms are calculated by using a fixed K\"{a}hler metric on $M$.
\end{prop}
In order to prove Theorem \ref{thm-main}, we also need a generalization of Theorem 1.1 in \cite{DJ}, which is a direct consequence of Lemma \ref{connectionC0lem}.
\begin{prop}
\label{lemma01} For any $w\in K$ and $0<\alpha <1$, there exists a constant $C_2>0$ so that $$\|A_{t_k}-A_{0, t_k}\|_{C^{0,\alpha}(M_w)} \leq C_2t_k.$$
\end{prop}
Granted these three propositions, we are ready to prove Theorem \ref{thm-main}.
\begin{proof}[Proof of Theorem \ref{thm-main}]
By Proposition \ref{thm01} and the Sobolev embedding theorem, there exists $u_k\in \mathcal{G}^{1,\alpha}$ and a limiting $C^{0,\alpha}$-connection $\Xi_0$, so that
$$ u_k (\Xi_{t_k} ) \rightarrow \Xi_0$$ in $C^{0,\alpha}(M_K)$. Thus, for any $w\in K$, the restriction $\Xi_0|_{M_w}$ of $\Xi_0$ is a $C^{0,\alpha}$-connection on $M_w$, and $ u_k (\Xi_{t_k} ) |_{M_w} $ converges to $\Xi_0|_{M_w}$ in the $C^{0,\alpha}$-sense. Proposition \ref{lemma01}, along with the fact that $ A_{0,t} \rightarrow A_{0}$ in the $C^\infty$-sense, gives
$$A_{t_k}\rightarrow A_0,$$ on $M_w$ in the $C^{0, \alpha}$-sense, where $A_0$ is given by (\ref{bconnections2}).
Since $$du_k= u_k \Xi_{t_k} |_{M_w}- u_k (\Xi_{t_k} ) |_{M_w} u_k,$$ and the $u_k$ are unitary, we have a $C^1$-bound for $u_k$, i.e. $\| u_k\|_{C^1(M_w)} \leq C$. As a result, the $C^{0, \alpha}$-convergence of $ u_k (\Xi_{t_k} ) |_{M_w} $ and $ \Xi_{t_k} |_{M_w} $ imply the $C^{1,\alpha}$-bound of $u_k$, i.e. $\| u_k\|_{C^{1,\alpha}(M_w)} \leq C'$. Thus by passing a subsequence, for $\alpha'<\alpha$ we have $u_k $ converges to a $C^{1,\alpha'}$-unitary gauge $u_\infty$ in the $C^{1,\alpha'}$-sense, which satisfies that $u_\infty( \Xi_0|_{M_w})=A_0$. This concludes the theorem.
\end{proof}
\section{A Poincar\'e inequality for $F_{A_t}$}\label{Poincare}
We continue to work under the setup of Theorem \ref{thm-main}, and choose a sequence of connections $\Xi_{t_k}$. We work on the fiber $M_w$ over a point $w\in N^o$, which is away from the discriminant locus of $f$, the bubbling points, and the ramification points and singularities of the spectral cover. As above we let $A_{t_k}$ denote the restriction of the anti-self-dual connection $\Xi_{t_k}$ to the smooth fiber $M_w$. The goal of this section is to derive a Poincar\'e type inequality for the curvature $F_{A_{t_k}}$, when $F_{A_{t_k}}$ is sufficiently small in the $C^0$-sense. The following proposition is the key analytic input to overcome the difficulty of the non-smoothness of the moduli spaces of flat connections on elliptic curves.
For notational simplicity we drop the subscript $k$, and denote our connections by $A_t$. We do this because, aside from being used to define $N^o$, the explicit sequence of times $t_k$ does not have any bearing on the results in this section.
\begin{prop}
\label{poincare} For any compact set $K\subset N^o$, there are constants $\epsilon_K>0$ and $C_K>0$ such that if
\be
\|F_{A_t}\|_{C^0{(M_w, \omega^{SF})}}\leq \epsilon_K\nonumber
\ee
for a certain $t\in (0,1]$ and $w\in K$, then
\be
\|F_{A_t}\|_w\leq C_K \|d_{A_t}^* F_{A_t}\|_{w}.\nonumber
\ee
\end{prop}
We begin by recalling part of our setup, as described in Theorem \ref{thm-main}. Fix an open subset $U\subset N^o$ biholomorphic to a disk in $\mathbb{C}$, satisfying $f^{-1}(U)\cong (U\times \mathbb{C})/{\rm Span}_{\mathbb{Z}}\{1, \tau\}$, where $\tau$ is a holomorphic function on $U$. Fix trivializations $P|_{M_U}\cong M_U \times SU(n)$ and $\mathcal{V}|_{M_U}\cong M_U \times \mathbb{C}^n$.
In Section \ref{mainproof} we define the connections $A_{0,t}={\rm diag}\{\alpha_{t,1}, \cdots, \alpha_{t,n}\}$ and $A_{0}={\rm diag}\{\alpha_{0,1}, \cdots, \alpha_{0,n}\}$ associated to the spectral covers, where
$$ \alpha_{t,j}= \pi {\rm Im} (\tau)^{-1} (q_{j,t}(w)\bar{\theta}-\bar{q}_{j,t}(w)\theta), \ \ \ \alpha_{0,j}= \pi {\rm Im} (\tau)^{-1} (q_{j}(w)\bar{\theta}-\bar{q}_{j}(w)\theta), $$ and $\theta|_{M_w}=dz$.
Here all points vary holomorphically in the base, and satisfy $$\sum\limits_{j=1}^n q_{j,t}(w)\equiv 0,\qquad \sum\limits_{j=1}^n q_{j}(w)\equiv 0.$$ We also have that $q_{j,t}$ converges to $q_{j}$ as $t\rightarrow 0$ as holomorphic functions. Furthermore, for any $w\in U$, $$q_{i,t}(w)\neq q_{j,t}(w) {\rm mod}(\mathbb{Z}+ \tau(w) \mathbb{Z}), \ \ \ q_{i}(w)\neq q_{j}(w){\rm mod}(\mathbb{Z}+ \tau(w) \mathbb{Z})$$ if $i\neq j$. The connections $d_{A_{0,t}}$ and $d_{A_{0}}$ act on $\eta\in C^\infty(M_w, \mathfrak{sl}(n, \mathbb{C}))$ via
$$d_{A_{0,t}} \eta=d\eta+[A_{0,t}, \eta], \ \ \ d_{A_{0}} \eta=d\eta+[A_{0}, \eta]. $$
Note that if $d_{A_{0,t}} \eta=0$, then $d\eta_{jj}=0$ and $d\eta_{ij}+(\alpha_{t,i}-\alpha_{t,j})\eta_{ij}=0$, which implies that $\eta_{ij}=0$ for $i\neq j$, and $\eta_{jj}$ are constants. Therefore $\ker d_{A_{0,t}}=\{{\rm diag} \{\eta_1, \cdots, \eta_n\} \in \mathfrak{sl}(n, \mathbb{C})\} $, and the same argument gives also $\ker d_{A_{0}}=\{{\rm diag} \{\eta_1, \cdots, \eta_n\} \in \mathfrak{sl}(n, \mathbb{C})\} $.
Since $A_{0,t}$ is flat ($F_{A_{0,t}}=d_{A_{0,t}}^2=0$), we have a de Rham complex $$ C^\infty(M_w, \mathfrak{sl}(n, \mathbb{C})) \stackrel{d_{A_{0,t}}}\longrightarrow C^\infty(T^* M_w \otimes \mathfrak{sl}(n, \mathbb{C})) \stackrel{d_{A_{0,t}}}\longrightarrow C^\infty(\wedge^2 T^* M_w\otimes \mathfrak{sl}(n, \mathbb{C})). $$
Furthermore, there is a well behaved Hodge theory (cf. \cite{AtB}). If $ \star_w$ denotes the Hodge star operator with respect to the flat metric $\omega^F_w:=\omega^{SF}|_{M_w}$, then $d^*_{A_{0,t}}=-\star_w d_{A_{0,t}} \star_w$
is the adjoint of $d_{A_{0,t}}$, and $d_{A_{0,t}}^*d_{A_{0,t}}+d_{A_{0,t}}d_{A_{0,t}}^*$ is the Hodge Laplacian. If we denote $\mathcal{H}_{A_{0,t}}^q(M_w, \mathfrak{sl}(n, \mathbb{C}))$ the space of $\mathfrak{sl}(n, \mathbb{C})$ valued harmonic $q$-forms, the Hodge theory asserts an orthogonal decomposition $$C^\infty(\wedge^q T^* M_w \otimes \mathfrak{sl}(n, \mathbb{C})) \cong \mathcal{H}_{A_{0,t}}^q(M_w, \mathfrak{sl}(n, \mathbb{C})) \oplus {\rm Im} d_{A_{0,t}} \oplus {\rm Im} d_{A_{0,t}}^*,$$ for $q=0,1,2$.
If we replace $\mathfrak{sl}(n, \mathbb{C})$ by the subalgebra $\mathfrak{su}(n)$, then we have the subcomplex $(C^\infty(\wedge^q T^* M_w \otimes \mathfrak{su}(n)), d_{A_{0,t}})$, the harmonic space of $\mathfrak{su}(n)$ valued $q$-forms $\mathcal{H}_{A_{0,t}}^q(M_w, \mathfrak{su}(n))$, and the respective Hodge decomposition. Note that we have the connection $A_t \in C^\infty(T^* M_w\otimes \mathfrak{su}(n))$ and the curvature $F_{A_t} \in C^\infty(\wedge^2 T^* M_w\otimes \mathfrak{su}(n))$. The virtual dimension of the moduli space $ \mathfrak{M}_{M_w}(n)$ of flat $SU(n)$-connections on $M_w$ is zero due to the Euler number of the complex $(C^\infty(\wedge^q T^* M_w \otimes \mathfrak{su}(n)), d_{A_{0,t}})$ vanishing, and thus the whole $ \mathfrak{M}_{M_w}(n)$ is regarded as degenerated, which causes many difficulties in the global analysis. However,
the flat connection $A_{0,t}$ belongs to the regular part of $ \mathfrak{M}_{M_w}(n)$, and $\mathcal{H}_{A_{0,t}}^1(M_w, \mathfrak{su}(n))$ is the tangent space at $A_{0,t}$. The infinitesimal deformation space under the action of the unitary gauge group is ${\rm Im} d_{A_{0,t}} \cap C^\infty(T^* M_w\otimes \mathfrak{su}(n))$, and by using the decomposition $\mathfrak{sl}(n, \mathbb{C})=\mathfrak{su}(n) \oplus i \mathfrak{su}(n)$, the space ${\rm Im} d_{A_{0,t}}^* \cap C^\infty(T^* M_w\otimes \mathfrak{su}(n))$ is identified with the infinitesimal deformation space induced by Hermitian gauges. The readers are referred to \cite{Nis11} for details of the above discussion.
We denote by $\Delta_{A_{0,t}}=-d_{A_{0,t}}^*d_{A_{0,t}}$ the Laplacian operator acting on $C^\infty(M_w, \mathfrak{sl}(n, \mathbb{C})) $, and have $\ker \Delta_{A_{0,t}}=\ker d_{A_{0,t}}$, ${\rm Im} \Delta_{A_{0,t}}={\rm Im} d_{A_{0,t}}^*,$ and $\ker \Delta_{A_{0,t}}\bot {\rm Im} d_{A_{0,t}}^*$ by the Hodge decomposition. We need a uniform estimate for the lower bounds of the first eigenvalue of $\Delta_{A_{0,t}}$.
\begin{lem}
\label{eigen} For any $w\in U$ and $t\in (0,1]$, if $\lambda_{w,t}$ is
the first eigenvalue of $-\Delta_{A_{0,t}}$ on the fiber $M_w$, then there is a constant $C_1>0$ independent of $t$ and $w$ such that $$\lambda_{w,t}\geq C_1.$$
\end{lem}
\begin{proof} If the above bound does not hold, there are sequences $w_k$ and $t_k$ such that $t_k \rightarrow t_0$ in $[0,1]$, $w_k \rightarrow w_0$ in $U$, and $$\lambda_{w_k,t_k}\rightarrow 0$$ when $k\rightarrow\infty$. Let $\psi_k \in C^\infty (M_{w_k}, \mathfrak{sl}(n, \mathbb{C}))$ be a normalized eigenvector of $\Delta_{A_{0,t_k}}$, i.e. $\Delta_{A_{0,t_k}} \psi_k =-\lambda_{w_k,t_k} \psi_k $ and $\|\psi_k\|_{w_k}=1$.
We regard $M_w$ as the 2-torus $T^2$ equipped with the complex structure $I_w$, and the K\"{a}hler metric $\omega^{F}_{w}$ as a metric on $T^2$ with respect to $I_w$. Since $\tau(w_k)\rightarrow \tau(w_0)$, we have that $I_{w_k}\rightarrow I_{w_0}$ and $\omega^{F}_{{w_k}} \rightarrow \omega^{F}_{{w_0}}$ in the $C^\infty$-sense. Note that $A_{0,t_k} \rightarrow A_{0,t_0}$ in the $C^\infty$-sense, and if $t_0=0$, then $A_{0,t_0}=A_{0}$. Standard elliptic estimates show that $\| \psi_k\|_{C^{\ell}}\leq C_{\ell}$ for constants $C_{\ell}>0$ independent of $k$, where the $C^{\ell}$-norms are calculated by using any fixed metric on $T^2$. By passing to a subsequence, we have that $ \psi_k \rightarrow \psi_\infty$ smoothly on $T^2$, $\|\psi_\infty\|_{w_0}=1$, and $\Delta_{A_{0,t_0}}\psi_\infty=0$. Thus $\psi_\infty \in \ker \Delta_{A_{0,t_0}}$ and can be represented as ${\rm diag} \{\eta_1, \cdots, \eta_n\} \in \mathfrak{sl}(n, \mathbb{C})$.
Since $\psi_k \bot \ker \Delta_{A_{0,t_k}}$, for any $\psi \in \ker \Delta_{A_{0,t_0}}= \ker \Delta_{A_{0,t_k}}$ we have $$0=\langle \psi_k, \psi\rangle_{w_k}\rightarrow \langle \psi_\infty, \psi\rangle_{w_0}.$$ So $\langle \psi_\infty, \psi\rangle_{w_0}=0$ yet $\|\psi_\infty\|_{w_0}=1$. This is a contradiction, and we obtain the conclusion.
\end{proof}
Again restricting our attention to a single fiber $M_w$ for $w\in U$, we can compute the norm of the fiberwise component of the curvature $F_{A_t}$ with respect to the semi-flat metric
\be
\|F_{A_t}\|^2_{C^0(M_w,\omega_t^{SF})}=\frac{1}{t^2}\|F_{A_t}\|^2_{C^0(M_w,\omega^{SF})}.\nonumber
\ee Because the error terms relating $\omega_t$ and $\omega_t^{SF}$ decay fast enough (see Theorem \ref{ttm-decay}), we have
\be
\|F_{A_t}\|^2_{C^0(M_w,\omega^{SF})}\leq Ct^2\|F_{A_t}\|^2_{C^0(M_w,\omega_t)}\leq Ct^2\|F_{\Xi_t}\|^2_{C^0(M_w,\omega_t)}.\nonumber
\ee
We assume that there is a constant $0< \epsilon \ll 1$, which is determined later, such that for a certain $t$ small enough it holds \be
\label{smallcurve}
\|F_{A_t}\|_{C^0(M_w, \omega^{SF})}\leq \epsilon,
\ee
for $w\in U$.
By Proposition \ref{estimate 1}, there is a sequence $t_k \rightarrow 0$ such that $$\|F_{A_{t_k}}\|^2_{C^0(M_w,\omega^{SF})}\leq Ct_k^2\|F_{\Xi_{t_k}}\|^2_{C^0(M_w,\omega_{t_k})}\leq \epsilon_k \rightarrow 0. $$
Here we used that $U$ is away from the bubbling set. Therefore, for any fixed $\epsilon >0$, if we take $t$ to be some time $t_k \ll 1$ such that $\epsilon_k <\epsilon$, then (\ref{smallcurve}) holds.
Recall by \eqref{hermitiangauge} that there exists a Hermitian gauge transformation $e^{-s_t}$ so that $e^{-s_t}(A_t)=A_{0,t}$. Although given above, we include the definition of this action here to emphasize that we are working exclusively on a fiber:
\be
\label{complexaction}
e^{-s_t}(A_{t})= A_{t}+e^{-s_t}\bar\partial_{ A_{t}} e^{ s_t}+\left( e^{- s_t}\bar\partial_{\ A_{t}} e^{s_t}\right)^*.
\ee
Given inequality \eqref{smallcurve}, the assumptions of Theorem 6.1 from \cite{DJ} are satisfied, which yields a new sequence of Hermitian gauge transformations $e^{\hat s_t}$ which are perpendicular to the kernel of $d_{A_{0,t}}$, bounded in $C^0$, and define the same connection $e^{-\hat s_t}_*A_t=A_{0,t}$.
For the remainder of this section we work on the fiber $M_w$, and so we may drop it from adorning norms when it is clear from context. Similarly, all norms in this section are computed with respect to $H$ and $\omega^{F}_w$.
\begin{lem}
\label{C0slem}
Given \eqref{smallcurve}, for every $w\in U$ the Hermitian endomorphism $\hat s_t$ satisfies
\be
\label{C0s}
\|\hat s_t\|_{C^0{(M_w, \omega^{SF})}}\leq C_2\epsilon
\ee
for a uniform constant $C_2$.
\end{lem}
\begin{proof}
To begin, we use that $\hat s_t$ is uniformly bounded in $C^0$. Following Appendix A of \cite{JW}, the fact that $A_{0,t}$ is flat, along with a standard formula for curvatures related by a complex gauge transformation, yields
\be
\label{thing4}
-\Delta_{w} |\hat s_t|^2\leq-|\partial_{A_{0,t}} \hat s_t|^2+{\rm Tr}\left( e^{ {\hat s_t}} \star_w F_{A_t}e^{- {\hat s_t}}\hat s_t\right),
\ee where $\Delta_{w}$ is the Laplacian with respect to the flat K\"{a}hler metric $\omega^{F}_{w}$.
Integrating the above equality over $M_w$, and using Lemma \ref{eigen} along with the fact that $\hat s_t$ is perpendicular to the kernel of $d_{A_{0,t}}$, gives
\be
\|\hat s_t\|_w^2\leq C\|d_{A_{0,t}} \hat s_t\|_w^2\leq C\epsilon \|\hat s_t\|_w.\nonumber
\ee
Therefore $\|\hat s_t\|_w\leq C\epsilon$.
Now we argue $\|\hat s_t\|_{C^0({M_w})}$ is also bounded by $C\epsilon$.
Note that \eqref{thing4} implies
\be
-\Delta_w |\hat s_t|^2\leq C\epsilon |\hat s_t|.\nonumber
\ee
Now, suppose the desired bound does not hold, so we can find a sequence of constants $C_t\rightarrow \infty$ so $\|\hat s_t\|_{C^0{}}\geq C_t \epsilon$. Set $\phi_t=|\hat s_t|^2/\|\hat s_t\|^2_{C^0{}}$. For $t$ small enough it holds
\be
-\Delta_w \phi_t\leq \frac {C \epsilon |\hat s_t|}{\|\hat s_t\|^2_{C^0{}}}\leq\frac{C}{C_t}\leq 1.\nonumber
\ee
If $y_t$ denotes the point in $M_w$ realizing $\sup|\hat s_t|^2$, in a fixed neighborhood of radius $a$ of $y_t$ we see $\phi_t$ is a $C^2$ function satisfying $-\Delta_w \phi_t\leq 1$, $0\leq \phi\leq 1,$ and $\phi_t(y_t)=1.$ Let $u_t$ be a $C^2$ function satisfying both $\Delta_w u_t=-1$ and $u_t(y_t)=1$. By making $a$ smaller if necessary we can guarantee that $u_t$ is strictly positive on $B_a(y_t)$, and this choice will only depend on $\omega^{F}_w$. Thus we have $-\Delta_w(\phi_t-u_t)\leq 0$ and $\phi_t(y_t)-u_t(y_t)=0$. Applying the mean value inequality to $\phi_t-u_t$ gives \be 0\leq \int_{B_a(y_t)}(\phi_t-u_t).\nonumber \ee By the positivity of $u_t$, there exists a constant $\delta>0$ independent of $t$ so that \be \delta\leq \int_{B_a(y_t)}u_t\leq \int_{B_a(y_t)}\phi_t.\nonumber \ee Rearranging terms gives \be \|\hat s_t\|^2_{C^0{}}\leq \frac1\delta \int_{B_a(y_t)}|\hat s_t|^2\leq \frac1\delta\|\hat s_t\|^2_w\leq {C\epsilon^2} ,\nonumber \ee which is our desired bound
\end{proof}
The above lemma has some strong consequences, which we now detail. First we need a few key formulas on $M_w$. The complex gauge action by a Hermitian endomorphism \eqref{complexaction} gives
\be
A_t=e^{\hat s_t}_*A_{0,t}= A_{0,t}+e^{\hat s_t}\bar\partial_{ A_{0,t}} e^{-\hat s_t}+\left( e^{\hat s_t}\bar\partial_{\ A_{0,t}} e^{-\hat s_t}\right)^*.\nonumber
\ee
For a given $s$ define ${\rm ad}_s:=[s,\cdot]$, and let $\Upsilon(s)\in{\rm End}({\rm End}(V_t))$ denote the power series
\be
\Upsilon(s)=\frac{e^{{\rm ad}_s}-1}{{\rm ad}_s}=\sum_{m=0}^{\infty}\frac{(-1)^m}{(m+1)!}({\rm ad}_s)^m.\nonumber
\ee
Note that the first term from the power series $\Upsilon(\hat s_t)$ is the identity, allowing us to write $\Upsilon(\hat s_t)=Id+\ti\Upsilon(\hat s_t)$. Now, recall the standard formula for the derivative of the exponential map
\be
e^{\hat s_t}\bar\partial_{ A_{0,t}} e^{-\hat s_t}=-\Upsilon(\hat s_t)\bar\partial_{A_{0,t}}\hat s_t.\nonumber
\ee
Following Appendix A in \cite{JW} we see
\bea
\label{connectionformula}
A_t&=&A_{0,t}-\bar\partial_{A_{0,t}}\hat s_t+\partial_{ A_{0,t}}\hat s_t-\ti\Upsilon(\hat s_t)\bar\partial_{A_{0,t}}\hat s_t+\ti\Upsilon(-\hat s_t)\partial_{ A_{0,t}}\hat s_t\nonumber\\
&=&A_{0,t}-i\star_w d_{A_{0,t}} \hat s_t+o(\hat s_t,\nabla_{A_{0,t}}\hat s_t),
\eea and \bea
\label{bigcurveexpression}
F_{ A_t} &=&F_{ A_{0,t}}+ \Upsilon(-\hat s_t ) \bar\partial_{A_{0,t}}\partial_{A_{0,t}} \hat s_t- \Upsilon(\hat s_t)\partial_{A_{0,t}}\bar\partial_{A_{0,t}}\hat s_t \nonumber\\
&& + \bar\partial_{A_{0,t}} \Upsilon(-\hat s_t) \wedge \partial_{A_{0,t}}\hat s_t - \partial_{A_{0,t}}\Upsilon(\hat s_t ) \wedge \bar\partial_{A_{0,t}}\hat s_t \\
&& - \Upsilon(-\hat s_t )\partial_{A_{0,t}}\hat s_t \wedge \Upsilon(\hat s_t )\bar\partial_{A_{0,t}}\hat s_t+ \Upsilon(\hat s_t)\bar\partial_{A_{0,t}}\hat s_t\wedge \Upsilon(-\hat s_t)\partial_{A_{0,t}}\hat s_t .\nonumber
\eea
This formula, along with the fact that $A_{0,t}$ is flat, leads to the following characterization of the curvature $F_{A_t}$
\be
\label{curvatureformula}
F_{A_t}=-i\, d_{A_{0,t}}\star_w d_{A_{0,t}} \hat s_t+T_1(\hat s_t,\nabla_{A_{0,t}}^2\hat s_t)+T_2(\partial_{A_{0,t}}\hat s_t,\bar\partial_{A_{0,t}}\hat s_t).
\ee
Thus we conclude
\be
\star_w F_{A_t}=-i\Delta_{A_{0,t}}\hat s_t+T_1+T_2\nonumber
\ee
where the tensors $T_1$ and $T_2$ satisfy
\be
\label{Tcontrol}
|T_1|\leq C\epsilon |\nabla^2_{A_{0,t}}\hat s_t|\qquad{\rm and}\qquad |T_2|\leq |\nabla_{A_{0,t}}\hat s_t|^2.
\ee
\begin{lem}
\label{connectionC0lem}
Given \eqref{smallcurve} and \eqref{C0s}, the following bound holds
\be
\label{connectionC0}
\|A_t-A_{0,t}\|_{C^{0,\alpha}{(M_w, \omega^{SF})}}\leq C_3\epsilon, \ \ \ \|\nabla_{A_{0,t}}\hat s_t\|_{C^{0,\alpha}{(M_w, \omega^{SF})}}\leq C_3\epsilon
\ee for any $0<\alpha <1$, by choosing $\epsilon$ small enough. Here the constant $C_3$ depends on $U\subset N^o$.
\end{lem}
\begin{proof}
We begin the proof with the standard elliptic a priori estimate (cf. \cite{GT,Bess})
\bea
\|\hat s_t\|_{L^p_2{}}&\leq& C\left(\|\Delta_{A_{0,t}}\hat s_t\|_{L^p{}}+\|\hat s_t\|_{L^p{}}\right)\nonumber\\
&\leq&C\left(\|F_{A_t}\|_{L^p{}}+\|T_1\|_{L^p{}}+\|T_2\|_{L^p{}}+\|\hat s_t\|_{L^p{}}\right)\nonumber\\
&\leq&C\left(\epsilon+\|T_1\|_{L^p{}}+\|T_2\|_{L^p{}}\right)\nonumber
\eea
where we have used \eqref{smallcurve} and \eqref{C0s} in the last inequality. We also use the assumption that $A_{0,t} \rightarrow A_0$ smoothly, and therefore all derivatives of $A_{0,t}$ are bounded independent of $t$. Thus all constants in the above inequality are independent of $t$.
The necessary bound for $T_1$ follows immediately $\|T_1\|_{L^p}\leq C\epsilon\|\hat s_t\|_{L^p_2{}}.$ For $T_2$ we use the interpolation inequality for tensors from \cite{Ham} (see also Section 7.6 in \cite{Au})
\be
\left(\int_{M_w}|\nabla_{A_{0,t}}\hat s_t|^{2p}\right)^{\frac1p}\leq (\sqrt 2+2p-2) \|\hat s_t\|_{C^0{}}\left(\int_{M_w}|\nabla_{A_{0,t}}^2\hat s_t|^p\right)^{\frac1p}.\nonumber
\ee
This implies $\|T_2\|_{L^p{}}\leq C\epsilon\|\hat s_t\|_{L^p_2{}}$. Thus
$$
\|\hat s_t\|_{L^p_2{}}\leq C\left(\epsilon+\epsilon\|\hat s_t\|_{L^p_2{}}\right)
$$ and
for $\epsilon$ small enough
\be
\label{lq2}
\|\hat s_t\|_{L^p_2{}}\leq C\epsilon.
\ee By Morrey's inequality, for large enough $p$ we can conclude
\be
\|\nabla_{A_{0,t}}\hat s_t\|_{C^{0,\alpha}{}}\leq C\epsilon,\nonumber
\ee
and the proof follows from \eqref{connectionformula}.
\end{proof}
Comparing this lemma to Theorem 3.11 of \cite{Nis11}, the bound of
(\ref{connectionC0}) is stronger, i.e. we have $\epsilon$ instead of $\epsilon^{\frac{1}{2}}$, due to our assumption that $A_{0,t}$ and $A_{0}$ are regular.
We now turn to the proof of the main proposition of this section
\begin{proof}[Proof of Proposition \ref{poincare}]
Once again we begin with the standard elliptic a priori estimate
\be
\|\hat s_t\|_{L^2_2{}}\leq C \left(\|\Delta_{A_{0,t}}\hat s_t\|_{w}+\|\hat s_t\|_{w}\right).\nonumber
\ee
Since $\hat s_t$ is perpendicular to the the kernel of $d_{A_{0,t}}$, we have
a stronger inequality
\be
\|\hat s_t\|_{L^2_2{}}\leq C \|\Delta_{A_{0,t}}\hat s_t\|_{w}\nonumber
\ee (cf. \cite{GT,Bess}). Again we use the fact that all derivatives of $A_{0,t}$ and $A_0$ are bounded independent of $t$.
Next, we recall \eqref{Tcontrol}. Applying the the interpolation inequality for tensors from the previous lemma for $p=2$, we have
\be
\|T_1+T_2\|_{w}\leq C \epsilon\|\hat s_t\|_{L^2_2{}}\leq C \epsilon\|\Delta_{A_{0,t}}\hat s_t\|_{w}.\nonumber
\ee
Let $F^o_t$ denote the projection of $\star_w F_{A_t}$ onto the kernel of $\Delta_{A_{0,t}}$, and set $F^\perp_t=\star_w F_{A_t}-F^o_t$. Because $\Delta_{A_{0,t}}\hat s_t$ is perpindicular to the kernel of $\Delta_{A_{0,t}}$, we can conclude
\bea
\|F^\perp_t\|_{w}&\geq& \|\Delta_{A_{0,t}}\hat s_t\|_{w}-\|F^\perp_t-\Delta_{A_{0,t}}\hat s_t\|_{w}\nonumber\\
&=&\|\Delta_{A_{0,t}}\hat s_t\|_{w}-\|(T_1+T_2)^\perp\|_{w}\nonumber\\
&\geq& (1-C\epsilon)\|\Delta_{A_{0,t}}\hat s_t\|_{w}\nonumber\\ & \geq & \frac12\|\Delta_{A_{0,t}}\hat s_t\|_{w}.\nonumber
\eea We take $\epsilon$ small enough such that $C \epsilon < \frac{1}{2}$.
Now, since $(\Delta_{A_{0,t}}\hat s_t)^o=0$, we also have
\be
\|F^o_t\|_{w}\leq \|(T_1+T_2)^o\|_{w}\leq C\epsilon \|\Delta_{A_{0,t}}\hat s_t\|_{w}\leq 2 C\epsilon \|F^\perp_t\|_{w},\nonumber
\ee
which implies
\bea
\|F_{A_t}\|_{w}&\leq& \|F^o_t\|_{w}+\|F^\perp_t\|_{w}\nonumber\\
&\leq& (1+2 C\epsilon)\|F^\perp_t\|_{w}\nonumber\\ & \leq & 2 \|F^\perp_t\|_w.\nonumber
\eea
Thus, applying the Poincar\'e inequality to $F^\perp_t$ and Lemma \ref{eigen}, we can conclude
\be
\|F_{A_t}\|_{w}\leq 2\|F_t^\perp\|_{w}\leq C\|d_{A_{0,t}}^* F_{A_t}\|_{w}.\nonumber
\ee
The proposition now follows from Lemma \ref{connectionC0lem}, which allows us to bound the difference between the connections $A_t$ and $A_{0,t}$
\bea
\|F_{A_t}\|_{w}&\leq&C \|d_{A_{0,t}}^*F_{A_t}\|_{w}\nonumber\\
&\leq&C \|d_{A_t}^*F_{A_t}\|_w+C \|A_t-A_{0,t}\|_{C^0{}}\|F_{A_t}\|_{w}\nonumber\\
&\leq&C \|d_{A_t}^* F_{A_t}\|_{w}+C \epsilon \|F_{A_t}\|_{w}.\nonumber
\eea
We choose further that $C \epsilon < \frac{1}{2}$, and obtain $$ \|F_{A_t}\|_{w} \leq 2 C \|d_{A_t}^* F_{A_t}\|_{w}.$$
For any $K\subset N^o$, we cover $K$ by finite open disks $U_\beta$, i.e. $K\subset \bigcup U_\beta \subset N^o$, and apply the above arguments to any $U_\beta$.
By letting $\epsilon_K=\min \{\epsilon\}$ over the covering, and $C_K$ the maximum constant over the covering, the proposition is proved.
\end{proof}
A corollary is the following Sobolev inequality.
\begin{cor}\label{sobolev}
For any $p\geq 2$, there exists a cosntant $C_p$ so that $$\|F_{A_t}\|_{L^p(M_w)} \leq C_p\|d_{A_t}^\star F_{A_t}\|_w.$$
\end{cor}
\begin{proof} In dimension two we have the Sobolev inequality $$\|\xi\|_{L^p} \leq C_p(\|\nabla_{A_{0,t}} \xi\|_{w} + \|\xi\|_{w})\leq C_p(\|\nabla_{A_{t}} \xi\|_{w} + \|(A_{t}-A_{0,t})\xi\|_{w}+\|\xi\|_{w}),$$ for any smooth section $\xi$ of ${\rm End}(\mathcal{V})$ and some constant $C_p$ independent of $w\in U$ and $t$.
Applying this to $ \xi=\star_w F_{A_t}$, we obtain $$\|F_{A_t}\|_{L^p} \leq C_p(\|d_{A_{t}}^\star F_{A_t}\|_{w} + (1+\epsilon) \|F_{A_t}\|_{w})\leq 2C_pC_K\|d_{A_t}^\star F_{A_t}\|_w,$$ by Proposition \ref{poincare}.
\end{proof}
\section{$C^0$ bounds on curvature}
\label{C0boundsection}
The main goal of this section is to prove Proposition \ref{prop2}, which establishes $C^0$ control for the curvature of a family of connections. It is a conditional result relying on assumption \eqref{hyp:curv}. To avoid confusion, we note that this result is applied twice. In Section \ref{lowerbounds}, in the proof of Proposition \ref{type3bubbling}, it is applied to a family of connections in scaled coordinates, for which \eqref{hyp:curv} can be verified directly. Once Proposition \ref{type3bubbling} is established, assumption \eqref{hyp:curv} holds for our main sequence of connections $\Xi_{t_k}$ from the statement of Theorem \ref{thm-main}, and so Proposition \ref{prop2} can be used to establish Proposition \ref{thm01}.
As above, let $U\subset\subset N^o$ be an open subset, compactly contained in $N_0$, and biholomorphic to a disk in $\mathbb{C}$. We have $f^{-1}(U)\cong (U\times \mathbb{C})/{\rm Span}_{\mathbb{Z}}\{1, \tau\}$, where the period $\tau$ is holomorphic on $U$. Let $w$ denote the complex coordinate on $U$, and $z$ the coordinate on $\mathbb{C}$. Furthermore, we fix a trivialization $P|_{M_U}\cong M_U \times SU(n)$ and $\mathcal{V}|_{M_U}\cong M_U \times \mathbb{C}^n$. Under such trivialization, the Hermitian metric $H$ is the absolute value $|\cdot|$, the connection $\Xi_t$ is a matrix valued 1-form, and the curvature $F_{\Xi_t}$ is a matrix valued 2-form, i.e. $\Xi_t\in C^\infty(T^*M_U, \mathfrak{su}(n))$ and $F_{\Xi_t} \in C^\infty(\wedge^2 T^*M_U, \mathfrak{su}(n))$.
Define real coordinates $(x_1, x_2)$ on $U$ satisfying $w=x_1+i x_2$, and recall that we have the decomposition $T^*M_U\cong {\rm Span}_{\mathbb{R}}\{dy_1,dy_2\}\oplus {\rm Span}_{\mathbb{R}}\{dx_1,dx_2\}$, where $z=y_1+\tau y_2$, and
$z$ is the coordinate on $\mathbb{C}$.
In these coordinates we write
\be\label{decomp}
\Xi_t=A_t+B_{t,1}dx_1+B_{t,2} dx_2,
\ee
where $A_t$ is a connection on the restriction to the fiber $\mathcal{V}|_{M_w}$, and $B_{t,i}$ is a section in $\Gamma(U, \Omega^0(M_w, \mathfrak{su}(n)))$ for $i=1,2$. Given this decomposition, the curvature can be written as
\be
\label{curve}
F_{\Xi_t}=F_{A_t}-\kappa_{t,1}\wedge dx_1-\kappa_{t,2}\wedge dx_2-F_{B,t} dx_1\wedge dx_2.
\ee
Here $F_{A_t}$ is the curvature of $A_t$, the mixed terms are given by
\be
\kappa_{t,i}=\frac{\partial}{\partial x_i} A_t-d_{A_t}B_{t,i}\qquad{\rm for}\qquad i=1,2,\nonumber
\ee
and the curvature in the base direction can be expressed as
\be
F_{B,t}=\frac{\partial}{\partial x_2} B_{t,1}-\frac{\partial}{\partial x_1} B_{t,2}-[B_{t,1},B_{t,2}].\nonumber
\ee
Because of the uniform equivalence $$ C_U^{-1}\omega^{SF}_t \leq \omega_t \leq C_U \omega^{SF}_t, \ \ \ {\rm and} \ \ \omega^{SF}_t|_{M_w}=t\omega^{SF}|_{M_w}, $$
the norms of the different curvature components satisfy $$ |F_{A_t}|_{\omega^{SF}}= t|F_{A_t}|_{\omega_t^{SF}}, \ \ \ |\kappa_{t,i}|_{\omega^{SF}} = \sqrt{t} |\kappa_{t,i}|_{\omega_t^{SF}}, \ \ \ |F_{B,t}|_{\omega^{SF}} =|F_{B,t}|_{\omega_t^{SF}}. $$
We now state the main assumption of this section. Assume that there is a constant $C_1>0$, so that for a $t\in (0,1]$ it holds
\be \label{hyp:curv}
\sup_{M_U} |F_{\Xi_t}|_{\omega_t} \leq C_1t^{-\frac{1}{2}}.
\ee
This implies
\be \label{hyp:curv+=} \sup_{M_U} |F_{A_t}|_{\omega^{SF}} \leq C_1t^{\frac{1}{2}}, \ \ \sup_{M_U} |\kappa_{t,i}|_{\omega^{SF}} \leq C_1, \ \ \sup_{M_U} |F_{B,t}|_{\omega^{SF}} \leq C_1t^{-\frac{1}{2}}.\nonumber
\ee
We assume that $t\ll 1$ small enough such that $C_1t^{\frac{1}{2}}<\epsilon_K$, where $\epsilon_K$ is the small constant controlling the curvature in Proposition \ref{poincare}, and $U\subset K$.
Thus by Proposition \ref{poincare}, we see that the curvature $F_{A_t}$ satisfies the Poincar\'{e} type inequality
\be \label{hyp:poincare}
\|F_{A_t}\|_{w}\leq C_2 \|d_{A_t}^* F_{A_t}\|_{w}.
\ee
This inequality, along with assumption \eqref{hyp:curv}, are instrumental in the following:
\begin{prop}\label{prop2} Let $\nabla_{x_i}=\partial_{x_i}+B_{t,i}$ for $ i=1,2$ denote covariant differentiation in the base direction. If (\ref{hyp:curv}) and (\ref{hyp:poincare}) hold for $t\ll 1$, for $U'\subset\subset U$ we have
the following inequalities: \begin{itemize} \item[i)] $$ \|F_{A_t}\|_{C^{0}(M_{U'},\omega^{SF})} \leq C_3 t, \ \ \ \|F_{B,t}\|_{C^{0}(M_{U'},\omega^{SF})} \leq C_3 , $$ \item[ii)] $$ \|\nabla_{x_i} F_{A_t}\|_{L^2(M_{U'}, \omega^{SF})} \leq C_3 t^{\frac{1}{2}} ,$$ \item[iii)] $$ \|F_{\Xi_t}\|_{C^0(M_{U'}, \omega^{SF})} \leq C_3, $$
\end{itemize} where the constant $C_3$ may depend on the distance from $U'$ to $\partial U$, but is independent of $t$.
\end{prop}
As above let $\star_w$ denote the Hodge star operator on the fiber $M_w$ with respect to the flat metric $\omega^F_w:=\omega^{SF}|_{M_w}=i{\rm Im}(\tau)^{-1}\,dz\wedge d\bar z$. Then $\star_w^2=-1$, $\star_wdz=-idz$ and $\star_w d\bar z=id\bar z$. We write the anti-self-dual equation under the decomposition (\ref{curve}).
\begin{lem}
The curvature of $\Xi_t$ satisfies
\be
\label{ASD1}
\star_w \kappa_{t,1}=\kappa_{t,2}
\ee
and
\be
\label{ASD2}
t^{-1}(1+G_0+G_1)\star_w F_{A_t}-(W+G_2) F_{B,t}=\sum_{j=1}^2\kappa_{t,j}\# G_3,
\ee
where $G_1$, $G_2$, $G_3$ are smooth functions depending on $t$ such that
\be
t^{-\frac{\nu}{2}}(\| G_1\|_{C^{0}(\omega^{SF})}+ \|\frac{\partial}{\partial z} G_1\|_{C^{\ell}(\omega^{SF})}+\|\frac{\partial}{\partial \bar{z}} G_1\|_{C^{\ell}(\omega^{SF})}+\sum_{j=2,3}\| G_j\|_{C^{\ell}(\omega^{SF})})\rightarrow 0, \nonumber
\ee
for any $\nu\in\mathbb{N}$, and $G_0$ is a function on $U$ such that $ \|G_0\|_{C^{\ell}(U)}\rightarrow 0$, when $t\rightarrow 0$.
\end{lem}
\begin{proof}
We first demonstrate that \eqref{ASD1} follows from $F_{\Xi_t}^{0,2}=F_{\Xi_t}^{2,0}=0$. Note that
\be
2(\kappa_{t,1}\wedge dx_1+\kappa_{t,2} \wedge dx_2)=(\kappa_{t,1} -i \kappa_{t,2}) \wedge dw+(\kappa_{t,1}+i\kappa_{t,2})\wedge d\bar w.\nonumber
\ee
This implies, using $\star_wdz=-idz$ and $\star_w d\bar z=id\bar z$, that
\be
\star_w(\kappa_{t,1} -i \kappa_{t,2})=i (\kappa_{t,1} -i \kappa_{t,2})=i\kappa_{t,1}+\kappa_{t,2}\nonumber
\ee
and
\be
\star_w(\kappa_{t,1}+i\kappa_{t,2})=-i (\kappa_{t,1}+i\kappa_{t,2})=-i\kappa_{t,1}+\kappa_{t,2}.\nonumber
\ee
Adding these two equations together proves \eqref{ASD1}.
We now concentrate on \eqref{ASD2}. Using $F_{\Xi_t}\wedge \omega_t=0$, along with the decompositions \eqref{SFmetric} and \eqref{curve}, we see
\bea
0=F_{\Xi_t}\wedge\omega_t&=&F_{\Xi_t}\wedge\omega_t^{SF}+F_{\Xi_t}\wedge i\partial\bar\partial\varphi_t\nonumber\\
&=& \frac i{2}(W^{-1}+2\varphi_{t,w \bar{w}})F_{A_t}\wedge dw \wedge d\bar{w}\nonumber\\
&&-\frac i2(tW+ 2\varphi_{t,z \bar{z}})F_{B,t}\,dx_1 \wedge dx_2 \wedge \theta \wedge \bar{\theta}\nonumber\\
&&+ (\kappa_{t,1}\wedge dx_1+\kappa_{t,2}\wedge dx_2)\wedge {\rm Im}\left(2\varphi_{t,w\bar{z}} dw \wedge d\bar{z}\right).\nonumber
\eea
Next, note that $\theta=dy_1+\tau dy_2=dz+bdw$,
\be
dx_1\wedge dx_2=\frac i2 dw\wedge d\bar w\qquad{\rm and}\qquad F_{A_t}=\frac i2(\star_w F_{A_t})W \theta\wedge \bar \theta.\nonumber
\ee
Thus, dividing out by the volume form $dz\wedge dw \wedge d\bar{z}\wedge d\bar{w}= \theta \wedge dw \wedge \bar{\theta}\wedge d\bar{w}$, the above equation can be rewritten as
\bea
0&=&(1+2\varphi_{t,w \bar{w}}W)\star_w F_{A_t}-(tW+ 2\varphi_{t,z \bar{z}})F_{B,t}\nonumber\\
&&+\sum_{i=1}^2 \kappa_{t,i}\, \#\left(\varphi_{t,z\bar w}+\varphi_{t,w\bar z}\right).\nonumber
\eea
We set $G_0=2\chi_{t,w\bar w}W$, $G_1=2(\varphi_{t,w\bar w}-\chi_{t,w\bar w})W$, $G_2=2t^{-1}\varphi_{t,z\bar z}$, and $G_3=t^{-1}(\varphi_{t,z\bar w}+\varphi_{t,w\bar z}).$ The proof now follows from Lemma \ref{lem-decay}.
\end{proof}
Next we turn to a Bochner type formula for $F_{A_t}$.
\begin{lem}\label{le2} If we denote $\Delta=\partial_{x_1}^2+\partial_{x_2}^2$, then \begin{eqnarray*}\Delta \| F_{A_t}\|_w^2 & \geq & \frac{1}{4} \sum_{i=1,2} \|\nabla_{x_i} F_{A_t}\|_w^2+ \frac{\delta}{t} \| d_{A_t}^* F_{A_t} \|_w^2\\ & & - C_4't(\sum_{j=1,2}\|[\kappa_{t,j}, \kappa_{t,j}]\|_w^2 +t^\nu)\\ & \geq & \frac{1}{4} \sum_{i=1,2} \|\nabla_{x_i} F_{A_t}\|_w^2+ \frac{\delta}{t} \| d_{A_t}^* F_{A_t} \|_w^2 - C_4t, \end{eqnarray*} for constants $\delta >0$, $C_4>0$ and $C_4'>0$. \end{lem}
\begin{proof} Note we can write the mixed and base curvature terms as $$ \nabla_{x_1}d_{A_t}-d_{A_t}\nabla_{x_1}=\kappa_{t,1}, \ \ \ \nabla_{x_2}d_{A_t}-d_{A_t}\nabla_{x_2}=\kappa_{t,2}, \ \ \ [ \nabla_{x_1}, \nabla_{x_2}] =F_{B,t}. $$ By the Bianchi identity $d_{\Xi_t}F_{\Xi_t}=0$, and so $$d_{A_t}F_{t,B}=\nabla_{x_1}\kappa_{t,2}-\nabla_{x_2}\kappa_{t,1}, \ \ \ \nabla_{x_1} F_{A_t}=d_{A_t} \kappa_{t,1}, \ \ {\rm and} \ \ \nabla_{x_2} F_{A_t}=d_{A_t} \kappa_{t,2}.$$
Recall that $\star_w dz=-i dz$, $\star_w d\bar{z}=id\bar{z}$ and $\star_w \frac{i}{2}Wdz\wedge d\bar{z}=1$. Also, $\star_w$ is independent of $w$ when acting on 1-forms, and $\partial_{x_i}\star_w= -W^{-1} (\partial_{x_i}W)\star_w $ in the other cases. By the above formulas, we derive
\bea
(\nabla_{x_1}^2+\nabla_{x_2}^2)F_{A_t}&=&\nabla_{x_1} d_{A_t} \kappa_{t,1}+ \nabla_{x_2} d_{A_t} \kappa_{t,2}\nonumber\\
&=&d_{A_t}(\nabla_{x_1} \kappa_{t,1} + \nabla_{x_2} \kappa_{t,2})+\sum_{j=1,2}[\kappa_{t,j}, \kappa_{t,j}].\nonumber
\eea
By \eqref{ASD1}, we also have
$$ \nabla_{x_1} \kappa_{t,1} =-\star_{w} \nabla_{x_1} \kappa_{t,2}, \ \ {\rm and} \ \ \nabla_{x_2} \kappa_{t,2} =\star_{w} \nabla_{x_2} \kappa_{t,1}. $$
Hence, using \eqref{ASD2}, we obtain a Weitzenb\"{o}ck type formula for $F_{A_t}$:
\bea
\label{thing20}
(\nabla_{x_1}^2+\nabla_{x_2}^2)F_{A_t} & = & d_{A_t}\star_{w}( \nabla_{x_2} \kappa_{t,1}- \nabla_{x_1} \kappa_{t,2})+\sum_{j=1,2}[\kappa_{t,j}, \kappa_{t,j}]\\ &=& - d_{A_t}\star_{w} d_{A_t}F_{B,t} +\sum_{j=1,2}[\kappa_{t,j}, \kappa_{t,j}] \nonumber\\ &=&- t^{-1} d_{A_t}\star_{w} d_{A_t}(G_4 \star_w F_{A_t}) +\sum_{j=1,2}[\kappa_{t,j}, \kappa_{t,j}] \nonumber\\ & & +d_{A_t}\star_{w} d_{A_t}( \sum_{i=1,2} \kappa_{t,i}\# G_5 ),\nonumber
\eea
where $$G_4= (W+ G_2)^{-1}(1+G_0+G_1), \ \ {\rm and} \ \ G_5= (W+ G_2)^{-1}G_3.$$ Note that for any differential form $\alpha$, $d_{A_t}\alpha =d^{f} \alpha$, where $d^{f}$ denotes the differential along the fiber direction, i.e. $d^f=\partial_{y_1}(\cdot)dy_1+\partial_{y_2}(\cdot)dy_2$, and $\nabla_{x_i}\alpha =\partial_{x_i} \alpha$.
Since $\| F_{A_t}\|_w^2=\int_{M_w}{\rm tr}F_{A_t}\wedge \star_w F_{A_t}$, a direct calculation shows
$$\partial_{x_i}^2\| F_{A_t}\|_w^2=\|\nabla_{x_i} F_{A_t}\|_w^2+2{\rm Re}\langle \nabla_{x_i}^2 F_{A_t} , F_{A_t} \rangle_w+T_i, $$
where the term $T_i$ arises from derivative on the fiber metric, and satisfies
\begin{eqnarray*} |T_i|& \leq & C(|\partial_{x_i}\star_w|\|\nabla_{x_i} F_{A_t}\|_w\| F_{A_t}\|_w+ |\partial_{x_i}^2 \star_w|\| F_{A_t}\|_w^2)\\ & \leq & \frac{1}{2} \|\nabla_{x_i} F_{A_t}\|_w^2+C\| F_{A_t}\|_w^2 . \end{eqnarray*}
Using the notation $\|\nabla_x F_{A_t}\|_w^2=\sum\limits_{i=1,2} \|\nabla_{x_i} F_{A_t}\|_w^2$, the above calculations give
$$ \Delta \| F_{A_t}\|_w^2=\|\nabla_x F_{A_t}\|_w^2+2{\rm Re}\langle (\nabla_{x_1}^2+\nabla_{x_2}^2) F_{A_t} , F_{A_t} \rangle_w+T_1+T_2. $$
To this equality, we can now apply \eqref{thing20}. Using $d_{A_t}^*=-\star_w d_{A_t} \star_w$, we see
\begin{eqnarray*}
{\rm Re}\langle (\nabla_{x_1}^2+\nabla_{x_2}^2)F_{A_t} , F_{A_t} \rangle_w &=& t^{-1} {\rm Re}\langle G_4 d_{A_t}^* F_{A_t} , d_{A_t}^* F_{A_t} \rangle_w \\
&&+ {\rm Re}\langle \sum_{j=1,2}[\kappa_{t,j}, \kappa_{t,j}] , F_{A_t} \rangle_w \\ & & - t^{-1} {\rm Re}\langle \star_{w} ( d^f G_4) \star_{w} F_{A_t} , d_{A_t}^* F_{A_t} \rangle_w\\
& & + {\rm Re}\langle \star_{w} d_{A_t}( \sum_{i=1,2} \kappa_{t,i}\# G_5 ), d_{A_t}^* F_{A_t} \rangle_w.
\end{eqnarray*}
Next, note that for a constant $\delta >0$, we have $${\rm Re}\langle G_4 d_{A_t}^* F_{A_t} , d_{A_t}^* F_{A_t} \rangle_w \geq 8 \delta \| d_{A_t}^* F_{A_t} \|_w^2. $$ Using \eqref{hyp:curv} to bound the mixed terms, and the Poincar\'e inequality \eqref{hyp:poincare}, we have
\begin{eqnarray*}|\langle \sum_{j=1,2}[\kappa_{t,j}, \kappa_{t,j}] , F_{A_t} \rangle_w| & \leq & C \sum_{j=1,2}\|[\kappa_{t,j}, \kappa_{t,j}]\|_w\| F_{A_t} \|_w\\ & \leq & C \sum_{j=1,2}\|[\kappa_{t,j}, \kappa_{t,j}]\|_w \| d_{A_t}^*F_{A_t} \|_w\\ & \leq & C t\sum_{j=1,2}\|[\kappa_{t,j}, \kappa_{t,j}]\|_w^2 +\frac{\delta}{t}\| d_{A_t}^*F_{A_t} \|_w^2. \end{eqnarray*}
Because $d^f W=0$, $ d^f G_0=0$, and $d^f G_4=o(t^{\nu})$ for $\nu\gg 1$, it follows that
\begin{eqnarray*} | t^{-1} {\rm Re}\langle \star_{w} ( d^f G_4) \star_{w} F_{A_t} , d_{A_t}^* F_{A_t} \rangle_w|& \leq & C \| F_{A_t}\|_w\|d_{A_t}^* F_{A_t}\|_w \\ &\leq & C t \| F_{A_t}\|_w^2+\frac{\delta}{t}\|d_{A_t}^* F_{A_t}\|_w^2. \end{eqnarray*}
Finally, $ |d^fG_5|_{\omega^{SF}}= o(t^{\nu})$ for any $\nu\gg 1$, and so
\begin{eqnarray*}|\langle \star_{w} d_{A_t}( \sum_{i=1,2} \kappa_{t,i}\# G_5 ), d_{A_t}^* F_{A_t} \rangle_w| & \leq & C\|d_{A_t}^* F_{A_t}\|_w (t^{\nu}+\sum_{i=1,2}\| d_{A_t} \kappa_{t,i}\|_w)\\ & = & C\|d_{A_t}^* F_{A_t}\|_w (t^{\nu}+\sum_{i=1,2}\| \nabla_{x_i} F_{A_t}\|_w) \\ & \leq & Ct(t^{\nu}+\| \nabla_x F_{A_t}\|_w^2)+ \frac{\delta}{t}\|d_{A_t}^* F_{A_t}\|_w^2. \end{eqnarray*}
Putting everything together \begin{eqnarray*} {\rm Re}\langle (\nabla_{x_1}^2+\nabla_{x_2}^2)F_{A_t} , F_{A_t} \rangle_w & \geq & \frac{4\delta}{t} \| d_{A_t}^* F_{A_t} \|_w^2 - Ct(t^{\nu}+ \| F_{A_t} \|_w ^2+\| \nabla_x F_{A_t}\|_w^2\\ & & +\sum_{j=1,2}\|[\kappa_{t,j}, \kappa_{t,j}]\|_w^2 ),\end{eqnarray*}
which implies
\begin{eqnarray*} \Delta \| F_{A_t}\|_w^2& \geq & \|\nabla_x F_{A_t}\|_w^2+ \frac{4\delta}{t} \| d_{A_t}^* F_{A_t} \|_w^2 -\frac{1}{2} \|\nabla_x F_{A_t}\|_w^2-2C\| F_{A_t}\|_w^2 \\ & &- Ct(t^{\nu}+ \| F_{A_t} \|_w ^2+\| \nabla_x F_{A_t}\|_w^2+\sum_{j=1,2}\|[\kappa_{t,j}, \kappa_{t,j}]\|_w^2 ). \end{eqnarray*} The Poincar\'e inequality (\ref{hyp:poincare}), along with Young's inequality, gives $$ \Delta \| F_{A_t}\|_w^2 \geq \frac{1}{4} \|\nabla_x F_{A_t}\|_w^2+ \frac{\delta}{t} \| d_{A_t}^* F_{A_t} \|_w^2- C t(\sum_{j=1,2}\|[\kappa_{t,j}, \kappa_{t,j}]\|_w^2 +t^\nu). $$ \end{proof}
We need the following elementary lemma, and we include the proof for the reader's convenience (cf. Sublemma 6.48 in \cite{Fuk2}). As in the pervious lemma, let $\Delta=\partial_{x_1}^2+\partial_{x_2}^2$ denote the coordinate Laplacian in the base.
\begin{lem}\label{le3}Let $\zeta$ be a non-negative real valued function satisfying $$ \Delta \zeta \geq \frac{\delta}{t}\zeta-t$$ on a disk $U\subset \mathbb{C}$. Then for an open subset $U'\subset\subset U$, there exists a constant $C_{5}$, which depends on the distance from $U'$ to $\partial U$, such that $$\sup_{U'}|\zeta|\leq C_{5}t^2.$$
\end{lem}
\begin{proof}
For any point $w_0\in U'$, let $d= \sup \{ |w-w_0\|w\in U\}$, and let $a$ be a positive number such that $4a^2d^2+4a <\delta$.
Consider the function $\xi=\zeta\exp(-\frac{a|w-w_0|^2}{\sqrt{t}})$. If $\xi$ achieves its maximum $w_1$ on $\partial U$, then $$\zeta(w_0)=\xi(w_0)\leq \xi(w_1)= \zeta(w_1) \exp(-\frac{a|w_1-w_0|^2}{\sqrt{t}})\leq C \exp(-\frac{ar^2}{\sqrt{t}}), $$ where $r$ is the distance from $w_0$ to $\partial U$. For $t$ small enough the right hand side is smaller than $Ct^2$.
Otherwise, at an interior maximum $w_1$, we see $$0=\partial_w \xi(w_1)=(-\frac{a(\bar{w}_1-\bar{w}_0)}{\sqrt{t}}\zeta (w_1) +\partial_w \zeta (w_1))\exp(-\frac{a|w_1-w_0|^2}{\sqrt{t}}), $$ and $\partial_{\bar{w}} \xi(w_1)=0 $. Furthermore, since $\Delta=2\partial_w \partial_{\bar{w}}$, at this maximum point
\bea
0 & \geq & \Delta \xi(w_1) \nonumber \\ &= & 2 \left(\partial_w \partial_{\bar{w}} \zeta (w_1)- \frac{a^2|w_1-w_0|^2+a \sqrt{t}}{t}\zeta (w_1) \right)\exp(-\frac{a|w_1-w_0|^2}{\sqrt{t}}) \nonumber \\
&\geq& \left(\frac{\delta}{t}\zeta (w_1)-2 \frac{a^2d^2+a }{t}\zeta (w_1)-t \right)\exp(-\frac{a|w_1-w_0|^2}{\sqrt{t}}) \nonumber \\ &\geq& \left(\frac{\delta}{2t}\zeta (w_1)-t \right)\exp(-\frac{a|w_1-w_0|^2}{\sqrt{t}}). \nonumber
\eea
Thus $$\xi(w_1)\leq \zeta (w_1) \leq 2 \delta^{-1} t^2,$$ and so $$\zeta(w_0)=\xi(w_0)\leq \xi(w_1) \leq 2 \delta^{-1} t^2. $$ \end{proof}
\begin{lem}\label{le-l2} For any $w\in U' \subset\subset U$, $$\| F_{A_t}\|_w \leq C_{6} t, \ \ \ and \ \ \|\nabla_{x_i} F_{A_t}\|_{L^2(U', \omega^{SF})} \leq C_{6} t^{\frac{1}{2}} , $$ for a constant $C_{6}>0$ independent of $t$ and $w$.
\end{lem}
\begin{proof} Lemma \ref{le2} and Lemma \ref{eigen} imply $$ \Delta \|F_{A_t}\|_w^{2} \geq \frac{1}{4} \|\nabla_x F_{A_t}\|_w^2+ \frac{\delta}{t} \| d_{A_t}^* F_{A_t} \|_w^2- Ct\geq \frac{\delta' }{t} \|F_{A_t}\|_w^{2} -Ct.$$ Thus by Lemma \ref{le3}, $$\|F_{A_t}\|_w^2 \leq C t^2.$$
Let $\vartheta$ be a smooth non-negative function on $U$ such that $\vartheta\equiv 1$ on $U'$, and $U'\subset {\rm supp}(\vartheta) \subset U$. By Lemma \ref{le2}, \bea
\int_{U'}\frac{1}{4} \|\nabla_x F_{A_t}\|_w^2dx_1dx_2 & \leq & \int_{U}\vartheta \Delta \|F_{A_t}\|_w^{2} dx_1dx_2 +C t \nonumber\\ & \leq& \int_{U}\max \{0, \Delta \vartheta\} \|F_{A_t}\|_w^{2} dx_1dx_2 +C_{22}t \nonumber \\ & \leq & C ( \int_{U} \|F_{A_t}\|_w^{2} dx_1dx_2 +t) \nonumber \\ & \leq & C t, \nonumber
\eea and we obtain the second estimate.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop2}] Firstly, we prove the $C^0$-estimate of $
F_{A_t}$. Assume that there is a sequence $t_k \rightarrow 0$ such that $$ t_k^{-1}\sup_{M_{w_k}}|F_{A_{t_k}}|_{\omega^{SF}}\rightarrow\infty, $$ where $w_k \rightarrow w_0$ in $U'$.
In Section 2.4, we saw that for
$D_r=\{\tilde{w} \in \mathbb{C}| |\tilde{w}|<r\}$, one can define smooth embeddings $\Phi_{k,r}: D_r \times M_{w_0} \rightarrow M_U$ by $$(\tilde{w}, a_1+a_2 \tau (w_0))\mapsto ( w_k+ \sqrt{t_k}\tilde{w}, a_1+a_2 \tau(w_k+\sqrt{t_k}\tilde{w})), \ \ \ a_1,a_2 \in\mathbb{R}/\mathbb{Z},$$ using the identification of $M_U$ with $(U\times\mathbb{C})/{\rm Span}_{\mathbb{Z}}\{1, \tau\}$.
We also demonstrated that $d\Phi_{k,r}^{-1}I d\Phi_{k,r} $ $\rightarrow I_\infty$, where $I$ is the complex structure of $M$, and $I_\infty$ denotes the complex structure of $\mathbb{C} \times M_{w_0}. $ Furthermore, as $t_k\rightarrow0$, we have both $$\Phi_{k,r}^* t_k^{-1} \omega_{t_k}^{SF} \rightarrow \omega_\infty\ \ \ {\rm and} \ \ \ (T_{\sigma_0} \circ \Phi_{k,r})^* t_k^{-1} \omega_{t_k} =\Phi_{k,r}^* t_k^{-1}T_{\sigma_0}^* \omega_{t_k} \rightarrow \omega_\infty $$ in the $C^\infty$-sense on $ D_r \times M_{w_0}$.
For any $t_k$, we identify $ D_r\times M_{w_0}$ with $ \Phi_{k,r}( D_r\times M_{w_0})$ by $\Phi_{k,r}$.
We have the curvature bound
$$ | F_{\Xi_{t_k}}|_{ t_k^{-1} \omega_{t_k}^{SF}}\leq C t_k^{\frac{1}{2}}, \ \ {\rm and} \ \ | F_{\Xi_{t_k}}|_{ \omega_\infty}\leq 2C t_k^{\frac{1}{2}},$$ by (\ref{hyp:curv}).
Since $\Xi_{t_k}$ is Yang-Mills, by the strong Uhlenbeck compactness theorem (cf. Theorem \ref{Ucompact}), there exists a subsequence and a family of unitary gauges $u_{t_k}$, such that $$\Xi_{t_k}'=u_{t_k}(\Xi_{t_k}) \rightarrow \Xi_\infty$$ in the locally $C^\infty$-sense on $D_r\times M_{w_0}$, where $\Xi_\infty$ is a flat $SU(n)$-connection. Note that $ F_{\Xi_{t_k}'}=u_{t_k}F_{\Xi_{t_k}}u_{t_k}^{-1}$, and so
$$| F_{\Xi_{t_k}'}|_{ t_k^{-1} \omega_{t_k}^{SF}}= | F_{\Xi_{t_k}}|_{ t_k^{-1} \omega_{t_k}^{SF}}\leq C t_k^{\frac{1}{2}}\ \ \ {\rm and} \ \ \ | F_{\Xi_{t_k}'}|_{ \omega_\infty}\leq 2C t_k^{\frac{1}{2}}.$$
Furthermore we have $ \| F_{\Xi_{t_k}'}\|_{ C^{\ell}( \omega_\infty)}\rightarrow 0 $ for any $\ell \geq 0$, when $t_k \rightarrow 0$.
Now, recall the Weitzenb\"{o}ck formula $$0= \Delta _{\Xi_{t_k}'}F_{\Xi_{t_k}'}=\nabla_{\Xi_{t_k}'}^\ast\nabla_{\Xi_{t_k}'}F_{\Xi_{t_k}'}+ R_{{t_k}^{-1}\omega_{t_k}}\#F_{\Xi_{t_k}'}+F_{\Xi_{t_k}'}\#F_{\Xi_{t_k}'},$$ which is an elliptic partial differential equation with smooth coefficients. The $L^p$-estimate for elliptic equations (cf. \cite{GT}, and the appendix of \cite{Bess}) gives $$ \|F_{\Xi_{t_k}'}\|_{L^p_2(\omega_\infty)}\leq C \|F_{\Xi_{t_k}'}\|_{L^p(\omega_\infty)} \leq C t_k^{\frac{1}{2}},$$ for any $p>2$.
We have $w-w_k=\sqrt{t_k}\tilde{w}$ through $\Phi_{k,r}$, and let $\tilde{w}=\tilde{x}_1+i\tilde{x}_2$. By (\ref{thing20}), \bea
\label{keyequality}
(\nabla_{x_1}^2+\nabla_{x_2}^2)F_{A_{t_k}'} &=&- t_k^{-1} d_{A_{t_k}'}\star_{w} d_{A_{t_k}'}(G_4 \star_w F_{A_{t_k}'}) +\sum_{ij}\kappa_{t_k,i}'\# \kappa_{t_k,j}' \nonumber\\ & & +d_{A_{t_k}'}\star_{w} d_{A_{t_k}'}( \sum_{i=1,2} \kappa_{t_k,i}'\# G_5 ),
\eea
where $\nabla_{x_j}=\partial_{x_j}+B_{t_k,j}'$, $G_4= (W+ G_2)^{-1}(1+G_0+G_1)$ and $G_5= (W+ G_2)^{-1}G_3$. Recall $$\|G_1\|_{C^{0}}+\|d^f G_1\|_{C^{\ell}}+\|G_j\|_{C^{\ell}}\leq C t_k^\nu $$ for $\nu \gg 1$. Let $z=\tilde{y}_1+i \tilde{y}_2$, and set $\nabla_{A_{t_k}', y_j}=\partial_{\tilde{y}_j}+A_{t_k,j}'$. By the Weitzenb\"{o}ck formula, $$d_{A_{t_k}'}d_{A_{t_k}'}^* F_{A_{t_k}'} =\nabla_{A_{t_k}'}^*\nabla_{A_{t_k}'} F_{A_{t_k}'}+ F_{A_{t_k}'}\#F_{A_{t_k}'}.$$ The connection Laplacian above is given by $$\nabla_{A_{t_k}'}^*\nabla_{A_{t_k}'}=-W^{-1}(\nabla_{A_{t_k}', \tilde{y}_1}^2+\nabla_{A_{t_k}',\tilde{y}_2}^2), $$ since $|\partial_{\tilde{y}_j}|^2_{\omega^{SF}}=W$.
We want to bound terms on the right hand side of \eqref{keyequality}. Scaling gives $B_{t_k,i}'dx_i=\sqrt{t_k}B_{t_k,i}'d\tilde{x}_i$ and $\kappa_{t_k,i}'dx_i=\sqrt{t_k}\kappa_{t_k,i}'d\tilde{x}_i$, in addition to $$ F_{B,t_k}dx_1\wedge dx_2=t_k F_{B,t_k}d\tilde{x}_1\wedge d\tilde{x}_2.$$
This leads to the following control of the mixed terms
$$ | \sqrt{t_k}\kappa_{t_k,i}' |_{\omega_\infty}\leq 2C t_k^{\frac{1}{2}},\ \ \ \ \| \sqrt{t_k}\kappa_{t_k,i}' \|_{C^{\ell}( \omega_\infty)} \rightarrow 0, $$
and $$ \| \sqrt{t_k}\kappa_{t_k,i}' \|_{L^p_2( \omega_\infty)}\leq \|F_{\Xi_{t_k}'}\|_{L^p_2(\omega_\infty)} \leq C t_k^{\frac{1}{2}}.$$
Additionally, writing $\nabla_{\tilde{x}_j}=\partial_{\tilde{x}_j}+\sqrt{t_k }B_{t_k,j}'$, we have $$ \nabla_{\tilde{x}_1}^2+\nabla_{\tilde{x}_2}^2=t_k(\nabla_{x_1}^2+\nabla_{x_2}^2).$$ The bound $|\partial^{\ell}_{y_j} G_5|\leq C $ gives $$\|t_k^{\frac{1}{2}}d_{A_{t_k}'}\star_{w} d_{A_{t_k}'}( \sum_{i=1,2} \kappa_{t_k,i}'\# G_5 )\|_{L^p(\omega_\infty)}\leq C t_k^{\frac{1}{2}}$$ for any $p>2$. Furthermore
$$\|\sum_{ij}\kappa_{t_k,i}'\# \kappa_{t_k,j}'\|_{ C^0(\omega_\infty)}\leq C.$$
Now, if we write $G_4= W^{-1}(1+G_0)+G_6$, then $$\frac{1}{2} W^{-1}(w_0)\leq G_4 \leq 2 W^{-1}(w_0), \ \ |\partial_{\tilde{y}_j}^{\ell}G_6|\leq C t_k^{\nu},$$
and
\bea d_{A_{t_k}'}d_{A_{t_k}'}^* G_4 F_{A_{t_k}'}& = &G_4 d_{A_{t_k}'}d_{A_{t_k}'}^* F_{A_{t_k}'}+d^f G_6\# \nabla_{A_{t_k}'} F_{A_{t_k}'}\nonumber \\ & & + \partial_{\tilde{y}_i \tilde{y}_j}^2 G_6\# F_{A_{t_k}'}. \nonumber \eea
We define the operator $$\mathcal{D}_k=\nabla_{\tilde{x}_1}^2+\nabla_{\tilde{x}_2}^2-G_4 \nabla_{A_{t_k}'}^*\nabla_{A_{t_k}'}=\nabla_{\tilde{x}_1}^2+\nabla_{\tilde{x}_2}^2+W^{-1}G_4(\nabla_{A_{t_k}', \tilde{y}_1}^2+\nabla_{A_{t_k}',\tilde{y}_2}^2), $$ which is a uniformly elliptic operator of order two. Then $ F_{A_{t_k}'}$ satisfies the following elliptic equation
\bea & &
\mathcal{D}_k F_{A_{t_k}'} - d^f G_6 \# \nabla_{A_{t_k}'} F_{A_{t_k}'}- \partial_{\tilde{y}_i \tilde{y}_j}^2 G_6 \# F_{A_{t_k}'} \\ &= & G_4 F_{A_{t_k}'}\#F_{A_{t_k}'} + t_k \sum_{ij}\kappa_{t_k,i}'\# \kappa_{t_k,j}' + t_k d_{A_{t_k}'}\star_{w} d_{A_{t_k}'}( \sum_{i=1,2} \kappa_{t_k,i}'\# G_5 ) \nonumber\\ & =& G_7. \nonumber
\eea
By the $L^p$-estimate for elliptic equations, for any $p>2$, $$ \| F_{A_{t_k}'}\|_{L_2^p(D_{r'}\times M_{w_0})}\leq C(\| F_{A_{t_k}'}\|_{L^2(D_{r}\times M_{w_0})}+\| G_7\|_{L^p(D_{r}\times M_{w_0})}),$$ for a $r'<r$. We obtain $$\| F_{A_{t_k}'}\|_{L_2^p(D_{r'}\times M_{w_0})}\leq C t_k,$$ since $$\| G_7\|_{L^p(D_{r}\times M_{w_0})}\leq C (\| F_{A_{t_k}'}\|_{C^0(D_{r}\times M_{w_0})}^2+t_k) \leq C t_k,$$
and $$ \| F_{A_{t_k}'}\|_{L^2(D_{r}\times M_{w_0})}^2=\int_{D_{r}}\|F_{A_{t_k}'}\|_w^2d\tilde{x}_1d\tilde{x}_2 \leq C t_k^2 $$ by Lemma \ref{le-l2}.
The Sobolev embedding theorem gives $$\|F_{A_{t_k}'}\|_{C^{1,\alpha}(D_{r'}\times M_{w_0})} \leq C t_k, $$ and thus $$\|F_{A_{t_k}}\|_{C^0(M_{w_k})}=\|F_{A_{t_k}'}\|_{C^0(M_{w_k})}\leq \|F_{A_{t_k}'}\|_{C^{1,\alpha}{(D_{r'}\times M_{w_0})}}\leq C t_k ,$$ which is a contradiction.
Therefore we obtain the $C^0$-estimate, i.e. $$ \|F_{A_t}\|_{C^{0}(M_{U'},\omega^{SF})} \leq C t,$$ for a constant $C >0$, and
$$ \|F_{B,t}\|_{C^{0}(M_{U'},\omega^{SF})} \leq C (t^{-1} \|F_{A_t}\|_{C^{0}(M_{U'},\omega^{SF})} + \|\kappa_{t,j}\|_{C^{0}(M_{U'},\omega^{SF})}) \leq C , $$ by (\ref{ASD2}).
\end{proof}
\section{Further estimates for small fiberwise curvature}
\label{furthreestimates}
We continue our discussion of the previous section, and prove further estimates under the exact same setup. Let $U\subset\subset N^o$ be an open subset, biholomorphic to a disk in $\mathbb{C}$, and $M_U\cong (U\times \mathbb{C})/{\rm Span}_{\mathbb{Z}}\{1, \tau\}$. Fix a trivialization $P|_{M_U}\cong M_U \times SU(n)$ and $\mathcal{V}|_{M_U}\cong M_U \times \mathbb{C}^n$. Under such trivialization, the Hermitian metric $H$ is the absolute value $|\cdot|$, the connection $\Xi_t$ is a matrix valued 1-form, and the curvature $F_{\Xi_t}$ is a matrix valued 2-form. Assume that for $t\ll 1$, (\ref{hyp:curv}) and (\ref{hyp:poincare}) hold, and thus all conclusions of Section \ref{C0boundsection} hold.
Recall that a fiberwise flat connection \begin{equation}\label{bconnections3} A_{0,t}= \pi ({\rm Im} (\tau))^{-1} ({\rm diag}\{q_{1,t}, \cdots, q_{n,t}\}\bar{\theta}-{\rm diag}\{\bar{q}_{1,t}, \cdots, \bar{q}_{n,t}\}\theta) \end{equation} is induced by $D_t \cap M_U$ (see Section 3.3), i.e. $D_t \cap M_w=\{q_{1,t}(w), \cdots, q_{n,t}(w)\}$. The goal of this section is the following proposition, which shows the relationship between the energy of curvature and the spectral covers. Here, as above, the coordinate derivative in the base is computed in our fixed frame.
\begin{prop}\label{prop2+0} If (\ref{hyp:curv}) and (\ref{hyp:poincare}) hold for $t\ll 1$,
we have the following inequalities. For $U'\subset\subset U$, $$ \|F_{\Xi_t}\|_{L^2(M_{U'}, \omega_t)}^2 \leq C_1(t+\int_{U'}\sum_{j=1,2}\|\partial_{x_j}A_{0,t}\|_w^2 dx_1dx_2), \ \ \ and$$ $$ \|F_{\Xi_t}\|_{L^2(M_{U'}, \omega_t)}^2 \geq C_1^{-1}(\int_{U'}\sum_{j=1,2}\|\partial_{x_j}A_{0,t}\|_w^2 dx_1dx_2-t), $$
where the constant $C_1$ may depend on the distance from $U'$ to $\partial U$, but is independent of $t$.
\end{prop}
The proof rests on several important lemmas.
\begin{lem}
There exists a constant $C_2$ such that for all $t\ll1$,
\be
\sup_{M_{U'}}|\nabla_{A_{0,t}}F_{A_t}|_{\omega^{SF}}\leq C_2t^{\frac{1}{2}}.\nonumber
\ee
\end{lem}
\begin{proof}
By \eqref{connectionC0}, it suffices to prove the above bound for $\nabla_{A_{t}}F_{A_t}$. We argue by contradiction. Let $t_k\rightarrow 0$ such that $$\lim_{k\rightarrow \infty}t_k^{-\frac{1}{2}}\sup_{M_{U'}}|\nabla_{A_{t_k}}F_{A_{t_k}}|_{\omega^{SF}} = \infty.$$
Let $p_k\in M_{U'}$ be the points where the supremum is attained, and in addition let $f(p_k) := w_k\rightarrow w_0 \in U$. As in Section 2.5, we consider the rescaled metrics $\hat \omega_{k} = t_k^{-1}\omega_{t_k}$ and the embeddings $\Phi_{k,r}: D_r \times M_{w_0} \rightarrow M_U$ defined by $$( \tilde{w}, a_1+a_2 \tau (w_0))\mapsto (w_k +\sqrt{t_k}\tilde{w}, a_1+a_2 \tau(w_k+\sqrt{t_k}\tilde{w})), \ \ a_1,a_2 \in\mathbb{R}/\mathbb{Z},$$ where $D_r = \{\tilde{w} \in \mathbb{C}| |\tilde{w}|<r\}$. We have seen that if $I$ is the complex structure of $M$, and $I_\infty$ the complex structure of $\mathbb{C} \times M_{w_0} $, then $d\Phi_{k,r}^{-1}I d\Phi_{k,r} \rightarrow I_\infty$, and in addition
$$\Phi_{k,r}^* t_k^{-1} \omega_{t_k}^{SF} \rightarrow \omega_\infty\ \ \ {\rm and} \ \ \ \Phi_{k,r}^* \hat{\omega}_{k} \rightarrow \omega_\infty$$
in the $C^\infty$-sense on $ D_r \times M_{w_0}$. Here $\omega_\infty$ is a flat K\"{a}hler metric on $ D_r \times M_{w_0}$.
Denote by $\hat \Xi_k$ the pull-back of $\Xi_{t_k}$ by $\Phi_{k,r}$, and identify $ D_r\times M_{w_0}$ with $ \Phi_{k,r}( D_r\times M_{w_0})$ via $\Phi_{k,r}$. By our hypothesis,
\begin{equation}\label{step1F}
\sup_{D_r\times M_{w_0}}t_k^{-\frac{1}{2}}|\nabla_{\hat \Xi_k}F_{\hat\Xi_k}|_{\omega_\infty} = \infty,
\end{equation} while by (\ref{hyp:curv}) we have the curvature bounds
$$ | F_{\hat{\Xi}_{k}}|_{ t_k^{-1} \omega_{t_k}^{SF}}\leq C t_k^{\frac{1}{2}} \ \ {\rm and} \ \ | F_{\hat{\Xi}_{k}}|_{ \hat{ \omega}_k}\leq 2C t_k^{\frac{1}{2}}.$$
Since $\hat\omega_k$ is equivalent to a fixed metric, standard Yang-Mills theory gives the first derivative bound $ |\nabla_{\hat\Xi_k}F_{\hat{\Xi}_{k}}|_{ \hat{ \omega}_k}\leq C $ (for instance see \cite{Wein}), but this is of course not enough to obtain a contradiction. So following \cite{Wein}, as in the proof of Lemma \ref{energybound}, we consider the the Bochner formula
\begin{equation}
0=\Delta_{\hat\omega_k} |F_{\hat\Xi_k}|_{\hat\omega_k}^2-2|\nabla_{\hat\Xi_k} F_{\hat\Xi_k}|_{\hat\omega_k}^2+F_{\hat\Xi_k}\#F_{\hat\Xi_k}\#F_{\hat\Xi_k}+R_{\hat\omega_k}\#F_{\hat\Xi_k}\#F_{\hat\Xi_k}.\nonumber
\end{equation}
We have seen that the curvature of the base metric satisfies $|R_{\omega_t}|_{\omega_t}^2\leq C$ on a compact subset of $N_0$, and scaling only improves this bound $|R_{\hat\omega_k}|_{\hat\omega_k}^2\leq C t_k^2.$ Rearranging terms, and multiplying by a positive function $\chi$ yields
\begin{equation}
2\chi|\nabla_{\hat\Xi_k} F_{\hat\Xi_k}|_{\hat\omega_k}^2\leq \chi\Delta_{\hat\omega_k} |F_{\hat\Xi_k}|_{\hat\omega_k\hat\Xi_k}^2+\chi |F_{\hat\Xi_k}|_{\hat\omega_k}^3+C \chi |F_{\hat\Xi_k}|_{\hat\omega_k}^2.\nonumber
\end{equation}
If $\eta$ is a positive bump function supported in $D_{r/2}$ and satisfying $\eta\equiv1$ in $D_{r/4}$, we specify $\chi=f^{-1}(\eta)$. Integrating the above inequality gives
\bea\label{L2controlderiv}
\int_{{D_{\frac r4}}\times M_{w_0}}|\nabla_{\hat\Xi_k} F_{\hat\Xi_k}|_{\hat\omega_k}^2\hat\omega_k^2 &\leq & \frac12\int_{D_{\frac r2}\times M_{w_0}}\Delta_{\hat\omega_k}\chi|F_{\hat{\Xi}_k}|_{\hat\omega_k}^2\hat\omega_k^2+C\int_{D_{\frac r2}\times M_{w_0}} t_k \nonumber \\ & \leq & Ct_k,
\eea
where the constant $ C$ depends on $r$, which again we take to be fixed.
We next turn to the higher order Bochner formula for Yang-Mills connections:
\begin{align}
0=&\Delta_{\hat\omega_k} |\nabla_{\hat\Xi_k} F_{\hat\Xi_k}|_{\hat\omega_k}^2-2|\nabla^2_{\hat\Xi_k} F_{\hat\Xi_k}|_{\hat\omega_k}^2+\nabla_{\hat \Xi_t} F_{\hat\Xi_k}\#\nabla_{\hat\Xi_k} F_{\hat\Xi_k}\#F_{\hat\Xi_k}\nonumber\\
&+R_{\hat\omega_k}\#\nabla_{\hat\Xi_k} F_{\hat\Xi_k}\#\nabla_{\hat\Xi_k} F_{\hat\Xi_k}+\nabla_{\hat\omega_k} R_{\hat\omega_k}\#F_{\hat\Xi_k}\#\nabla F_{\hat\Xi_k}.\nonumber
\end{align}
Since $|\nabla_{\hat\omega_k} R_{\hat\omega_k}|_{\hat\omega_k}\leq t_k|\nabla_{\omega_{t_k}} R_{\omega_{t_k}}|_{\omega_{t_k}}\leq C_U t_k$, we have
\begin{equation}
-\Delta_{\hat\omega_k} |\nabla_{\hat\Xi_k} F_{\hat\Xi_k}|_{\hat\omega_k}^2\leq C(t_k^{\frac{1}{2}} |\nabla_{\hat\Xi_k} F_{\hat\Xi_k}|_{\hat\omega_k}^2+t_k^{\frac{3}{2}} |\nabla_{\hat\Xi_k} F_{\hat\Xi_k}|_{\hat\omega_k}).\nonumber
\end{equation}
Set
$$\psi_k:= |\nabla_{\hat\Xi_k} F_{\hat\Xi_k}|_{\hat\omega_k}^2/\sup_{D_r\times M_{w_0}}|\nabla_{\hat\Xi_k} F_{\hat\Xi_k}|^2_{\hat\omega_k}.$$
The above Bochner formula, in addition to our hypothesis \eqref{step1F}, gives $$-\Delta_{\hat\omega_k} \psi_k\leq C( t_k^{\frac{1}{2}}+ t_k)\leq 1,$$ for $k\gg 1$. We now follow the argument used in Lemma \ref{C0s}. Let $\hat p_k$ be the pullbacks of the points $p_k$ via $\Phi_{k,r}$. These are the points realizing the supremum of $|\nabla_{\hat\Xi_k} F_{\hat\Xi_k}|^2_{\hat\omega_k}$, so that $\psi_k(\hat p_k) = 1$. Now construct a sequence of functions $u_k$ solving $\Delta_{\hat\omega_k} u_k=-1$ and $u_k(\hat p_k)=1$. Working on a small ball $B_{\hat\omega_k}(\hat p_k, r_0)$, we can assume that $u_k>\varepsilon_0$ for some $\varepsilon_0>0$ independent of $k$. Then since $-\Delta(\psi_k-u_k)\leq 0$, by the mean value inequality, there exists a $\delta>0$ depending only $\varepsilon_0$ and $r_0$ such that
\begin{equation}
\delta<\int_{B_{\hat\omega_k}(\hat p_k,r_0)}u_k\leq \int_{B_{\hat\omega_k}(\hat p_k,r_0)}\psi_k\leq \int_{D_{r/4}\times M_{w_0}}\psi_k\leq \frac{C_8t_k}{\sup\limits_{D_r\times M_{w_0}}|\nabla_{\hat\Xi_k} F_{\hat\Xi_k}|^2_{\hat\omega_k}}\nonumber
\end{equation}
where the final inequality follows from \eqref{L2controlderiv}. This contradicts \eqref{step1F}, completing the proof.
\end{proof}
Next, we have a $C^{1,\alpha}$-estimate for $A_t$.
\begin{lem}\label{prop2+00} For all $w\in U'$, and for all $t\ll 1$, $0<\alpha <1$, $$\|A_t-A_{0,t}\|_{C^{1,\alpha}(M_w)}\leq C_{3}t^{\frac{1}{2}} \ \ and \ \ \|\nabla^2_{A_{0,t}}\hat s_t\|_{C^{0,\alpha}(M_w)}\leq C_{4}t^{\frac{1}{2}},$$ for constants $C_{3}$ and $C_{4}$ independent of $w$ and $t$.
\end{lem}
\begin{proof}
We begin by recalling inequality \eqref{lq2}, which follows from Proposition \ref{prop2}, and properties of $\hat s_t$
\be
\|\hat s_t\|_{L^p_2(M_w)}\leq Ct.\nonumber
\ee
We would like to extend the above estimate to the case of $p=\infty$. To accomplish this, we turn to the higher order elliptic a priori estimate
\bea
\|\hat s_t\|_{L^p_3(M_w)}&\leq& C \left(\|\Delta_{A_{0,t}}\hat s_t\|_{L^p_1(M_w)}+\|\hat s_t\|_{L^p(M_w)}\right)\nonumber\\
&\leq&C \left(\|\Delta_{A_{0,t}}\hat s_t\|_{L^p_1(M_w)}+ t \right).\nonumber
\eea
Taking one fiber derivative of \eqref{curvatureformula}, and using the fact that $\|\hat s_t\|_{C^0(M_w)}$ and $\|\nabla_{A_{0,t}}\hat s_t\|_{C^0(M_w)}$ are controlled by $t$, we see that
\be
\|\Delta_{A_{0,t}}\hat s_t\|_{L^p_1(M_w)}\leq \|\nabla_{A_{0,t}}F_{A_t}\|_{L^p(M_w)}+t\|\hat s_t\|_{L^p_3(M_w)}+t\|\hat s_t\|_{L^p_2(M_w)}.\nonumber
\ee
Thus, for $t$ small enough
\be
\|\hat s_t\|_{L^p_3(M_w)}\leq C (t+ \|\nabla_{A_{0,t}}F_{A_t}\|_{L^p(M_w)})\leq C t^{\frac{1}{2}}.\nonumber
\ee
By Morrey's inequality we have
\be
\label{thing6}
\|\nabla^2_{A_{0,t}}\hat s_t\|_{C^{0,\alpha}(M_w)}\leq C t^{\frac{1}{2}}.
\ee
\end{proof}
If we let $\Xi_{t}^0=e^{-\hat{s}_t}(\Xi_t)$, then $\Xi_{t}^0|_{M_w}=A_{0,t}$, and we write $$\Xi_{t}^0=A_{0,t}+B_{t,1}^0dx_1+B_{t,2}^0dx_2, \ \ \ {\rm and} \ \ F_{\Xi_{t}^0}=-\kappa_{t,1}^0dx_1-\kappa_{t,2}^0dx_2-F_{B,t}^0dx_1\wedge dx_2,$$ where
$$\kappa_{t,j}^0=\partial_{x_j} A_{0,t}-d_{A_{0,t}}B_{t,j}^0.$$ Note that we still have $F_{\Xi_{t}^0}^{0,2}=0$, which implies
\begin{equation}\label{HYMB++}\star_w \kappa_{t,1}^0=\kappa_{t,2}^0,\end{equation} and thus $$ \star_w\partial_{x_1} A_{0,t}-\partial_{x_2} A_{0,t}=\star_w d_{A_{0,t}}B_{t,1}^0-d_{A_{0,t}}B_{t,2}^0. $$
Since $$\star_w\partial_{x_1} A_{0,t}-\partial_{x_2} A_{0,t} \in \ker \Delta_{A_{0,t}}, \ \ d_{A_{0,t}}B_{t,2}^0\in {\rm Im}d_{A_{0,t}}, \ \ {\rm and} \ \ \star_w d_{A_{0,t}}B_{t,1}^0 \in {\rm Im}d_{A_{0,t}}^*,$$ we have $ \star_w\partial_{x_1} A_{0,t}=\partial_{x_2} A_{0,t}$ and $d_{A_{0,t}}B_{t,j}^0=0$ by the Hodge decomposition. As a result we obtain
\begin{equation}\label{curv+++} \kappa_{t,j}^0=\partial_{x_j} A_{0,t}.\end{equation}
A direct calculation shows
\bea\label{for+==}
\kappa_{t,j}-\kappa_{t,j}^0 & =&\partial_{x_j} (A_t-A_{0,t})-d_{A_{t}}B_{t,j} \\
&= &\nabla_{x_j} (A_t-A_{0,t})-[B_{t,j}, A_t-A_{0,t}]\nonumber\\ & & -d_{A_{0,t}}B_{t,j} + [A_{0,t}-A_t,B_{t,j}]\nonumber \\ &=& \nabla_{x_j} (A_t-A_{0,t})-d_{A_{0,t}}B_{t,j}.
\nonumber \eea
Now, by (\ref{ASD1}), (\ref{HYMB++}) and (\ref{for+==}), $$ \star_w\nabla_{x_1} (A_t-A_{0,t})-\nabla_{x_2} (A_t-A_{0,t})=\star_w d_{A_{0,t}}B_{t,1}-d_{A_{0,t}}B_{t,2}, $$ and since $\star_w d_{A_{0,t}}B_{t,1} \bot d_{A_{0,t}}B_{t,2}$, i.e. $\langle\star_w d_{A_{0,t}}B_{t,1} , d_{A_{0,t}}B_{t,2}\rangle_w=0$, we have
$$ \|d_{A_{0,t}}B_{t,j}\|_w \leq \sum_{i=1,2} \|\nabla_{x_i} (A_t-A_{0,t}) \|_w,$$ for any $w\in U$. Consequently, for $j=1,2$ \begin{equation}\label{?+++}\|\kappa_{t,j}-\kappa_{t,j}^0\|_w \leq 2 \sum_{i=1,2} \|\nabla_{x_i} (A_t-A_{0,t}) \|_w. \end{equation}
Furthermore, if we decompose $B_{t,j}=B_{t,j}^o+B_{t,j}^\bot$, where $B_{t,j}^o\in\ker d_{A_{0,t}}$ and $B_{t,j}^\bot \bot\ker d_{A_{0,t}}$, then \begin{equation}\label{?+++} \|B_{t,j}^\bot\|_w \leq C \|d_{A_{0,t}}B_{t,j}\|_w \leq C \sum_{i=1,2} \|\nabla_{x_i} (A_t-A_{0,t}) \|_w,\end{equation} by Lemma \ref{eigen}. We need one more Lemma before we are ready to prove Proposition \ref{prop2+0}.
\begin{lem}
\label{prop1}
On $U'\subset\subset U$, we have
\be
\int_{U}\sum_{j=1,2}\|\nabla_{x_j}(A_t-A_{0,t})\|_w^2dx_1dx_2\leq C_{5}(t^2+\int_U\sum_{j=1,2}\|\nabla_{x_j} F_{ A_t} \|_{w}^2dx_1dx_2),\nonumber
\ee for a constant $C_{5}>0$. Consequently, by ii) of Proposition \ref{prop2}, $$ \int_{U}\sum_{j=1,2}\|\kappa_{t,j}-\kappa_{t,j}^0\|_w^2 dx_1dx_2\leq C_{6} t.$$ \end{lem}
\begin{proof}
We denote two important terms by $$\Lambda= \sum_{j=1,2}\|\nabla_{x_j}(A_t-A_{0,t})\|_w, \ \ \ \Theta=\sum_{j=1,2}\|\nabla_{x_j} F_{ A_t} \|_{w}. $$
First, for $j=1,2$, we decompose $\nabla_{x_j}\hat s_t=\nabla_{x_j}\hat s_t^o+\nabla_{x_j}\hat s_t^\bot$, where $\nabla_{x_j}\hat s_t^\bot$ is perpendicular to the kernel of $d_{A_{0,t}}$
and $\nabla_{x_j}\hat s_t^o\in \ker d_{A_{0,t}}$. Recall that $\ker d_{A_{0,t}}=\{{\rm diag}\{\eta_1, \cdots, \eta_n\}\in\mathfrak{sl}(n, \mathbb{C})\}$, and as a volume form $\omega^{SF}|_{M_w}=dv$ is independent of $w$ under the identification $M_w \cong T^2$. For any $\eta\in \ker d_{A_{0,t}}$, since $[B_{t,j}^o, \eta]=0$, $$\nabla_{x_j} \eta=\partial_{x_j}\eta+[B_{t,j},\eta]=[B_{t,j}^\bot,\eta].$$
Thus $$0=\partial_{x_j} \langle \hat s_t, \eta \rangle_w = \langle \nabla_{x_j} \hat s_t, \eta \rangle_w+\langle \hat s_t, \nabla_{x_j} \eta \rangle_w=\langle \nabla_{x_j} \hat s_t^o, \eta \rangle_w+\langle \hat s_t, [B_{t,j}^\bot, \eta] \rangle_w ,$$ and by (\ref{?+++})
$$ \|\nabla_{x_j}\hat s_t^o\|_w\leq C \|\hat s_t\|_{C^0} \|B_{t,j}^\bot \|_w\leq C t\Lambda.$$ Along with Lemma \ref{eigen}, this implies $$ \|\nabla_{x_j}\hat s_t\|_w\leq C (\|\nabla_{x_j}\hat s_t^\bot\|_w+ t\Lambda)\leq C (\|d_{A_{0,t}}\nabla_{x_j}\hat s_t\|_w+ t\Lambda). $$ Since $$d_{A_{t}}\nabla_{x_j}\hat s_t=d_{A_{0,t}}\nabla_{x_j}\hat s_t+[A_t-A_{0,t},\nabla_{x_j}\hat s_t], \ \ {\rm and} \ \ \| A_t-A_{0,t}\|_{C^0}\leq Ct, $$ we obtain $$ \|\nabla_{x_j}\hat s_t\|_w\leq C (\|d_{A_{t}}\nabla_{x_j}\hat s_t\|_w+ t\Lambda). $$
Next, take the derivative of \eqref{connectionformula} in the base direction to see
\bea
\|\nabla_{x_j}(A_t-A_{0,t})\|_{w}^2&\leq& 2\|\nabla_{x_j}(\Upsilon(\hat s_t))d_{A_{t}}\hat s_t\|_{w}^2+2\|\Upsilon(\hat s_t)\nabla_{x_j}(d_{A_{t}}\hat s_t)\|_{w}^2.\nonumber
\eea We concentrate on the two terms on the right hand side above separately. By Lemma \ref{connectionC0lem} and Proposition \ref{prop2}, $\hat s_t$, $\nabla_{A_{0,t}}\hat s_t$ and $A_t-A_{0,t}$ are bounded in $C^0$ by $t$, and so the first term satisfies
\be
\|\nabla_{x_j}(\Upsilon(\hat s_t))d_{A_{t}}\hat s_t\|_{w}^2\leq t^2 C \|\nabla_{x_j}\hat s_t\|_{w}^2\leq t^2 C (\|d_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}^2+t^2\Lambda^2).\nonumber
\ee
To bound the second of the two terms, note that $\kappa_{t,j}$ is bounded, and $ \nabla_{x_j} d_{A_{t}}-d_{A_{t}}\nabla_{x_j}=\kappa_{t,j} $. Thus
\be
\|\Upsilon(\hat s_t)\nabla_{x_j}(d_{A_{t}}\hat s_t)\|_{w}^2\leq C \|\hat s_t\|_{w}^2+2\|d_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}^2\leq C t^2+2\|d_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}^2,\nonumber
\ee
from which we conclude
\be
\Lambda^2\leq 2\sum_{j=1,2}\|\nabla_{x_j}(A_t-A_{0,t})\|_{w}^2\leq 6\sum_{j=1,2}\|d_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}^2+C t^2.\nonumber
\ee
Therefore it suffices to bound $\|d_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}^2$.
Integration by parts, along with Lemma \ref{eigen}, gives\bea
\int_{M_w}|d_{A_{t}}\nabla_{x_j}\hat s_t|^2\omega^{SF}&\leq& \int_{M_w}|\nabla_{x_j}\hat s_t\|\Delta_{A_{t}}\nabla_{x_j}\hat s_t|\omega^{SF} \nonumber\\
&\leq&\|\nabla_{x_j}\hat s_t\|_{w}\|\Delta_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}\nonumber\\
&\leq&C (\|d_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}+t\Lambda)\|\Delta_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}\nonumber
\eea
and so
\be
\label{thing7}
\|d_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}^2\leq C \|\Delta_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}^2+t^2\Lambda^2.
\ee
Thus we obtain $$\Lambda^2 \leq C (\sum_{j=1,2}\|\Delta_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}^2+t^2).$$
In order to bound $\Delta_{A_{t}}\nabla_{x_j}\hat s_t$, we turn to the equality (\ref{bigcurveexpression}) for the curvature of $A_t$, using the fact that $A_{0,t}$ is flat, \bea
F_{ A_t} &=&i\, d_{A_{t}}\star_w d_{A_{t}} \hat s_t - \tilde{\Upsilon}(\hat s_t ) \bar\partial_{A_{t}}\partial_{A_{t}} \hat s_t + \tilde{\Upsilon}(-\hat s_t)\partial_{A_{t}}\bar\partial_{A_{t}}\hat s_t \nonumber \\
&& - \bar\partial_{A_{t}} \tilde{\Upsilon}(\hat s_t) \wedge \partial_{A_{t}}\hat s_t + \partial_{A_{t}}\tilde{\Upsilon}(-\hat s_t ) \wedge \bar\partial_{A_{t}}\hat s_t \nonumber\\
&& - \Upsilon(\hat s_t )\partial_{A_{t}}\hat s_t \wedge \Upsilon(-\hat s_t)\bar\partial_{A_{t}}\hat s_t+ \Upsilon(-\hat s_t)\bar\partial_{A_{t}}\hat s_t\wedge \Upsilon(\hat s_t)\partial_{A_{t}}\hat s_t .\nonumber
\eea We take the derivative of this equation in the base direction, and calculate $\nabla_{x_j}F_{A_t}$.
Firstly,
\bea \nabla_{x_j} d_{A_{t}}\star_w d_{A_{t}} \hat s_t& = & d_{A_{t}}\nabla_{x_j}\star_w d_{A_{t}} \hat s_t+\kappa_{t,j}\# d_{A_{t}} \hat s_t
\nonumber \\ & = & d_{A_{t}}\star_w d_{A_{t}}\nabla_{x_j} \hat s_t+d_{A_{t}}[\star_w\kappa_{t,j} ,\hat s_t ]+\kappa_{t,j}\# d_{A_{t}} \hat s_t
\nonumber\\ & =& d_{A_{t}}\star_w d_{A_{t}}\nabla_{x_j} \hat s_t\pm [\nabla_{x_i}F_{A_t} ,\hat s_t ]+\kappa_{t,j}\# d_{A_{t}} \hat s_t \nonumber \eea by $\nabla_{x_i}F_{A_t}=d_{A_t}\kappa_{t,i}=\pm d_{A_t}\star_w\kappa_{t,j}$, which implies
\bea
& & | \nabla_{x_j} d_{A_{t}}\star_w d_{A_{t}} \hat s_t- d_{A_{t}}\star_w d_{A_{t}}\nabla_{x_j} \hat s_t|\nonumber \\
&& \qquad\qquad\qquad\qquad\leq C (|\nabla_{A_{0,t}}\hat s_t|+|A_t-A_{0,t}\| \hat s_t| +|\nabla_{x_i}F_{A_t}\|\hat s_t|),\nonumber \\
&& \qquad\qquad\qquad\qquad\leq C t(1+\sum_{i=1,2}|\nabla_{x_i}F_{A_t}|). \nonumber \eea
As a result, we have $$ \| \nabla_{x_j} d_{A_{t}}\star_w d_{A_{t}} \hat s_t- d_{A_{t}}\star_w d_{A_{t}}\nabla_{x_j} \hat s_t\|_w \leq C t(1+\Theta). $$
Secondly, note that $\nabla_{A_t}= \nabla_{A_{0,t}}+(A_t-A_{0,t})$, and $$\nabla_{A_t}^2=\nabla_{A_{0,t}}^2+(A_t-A_{0,t})\# \nabla_{A_{0,t}} +\nabla_{A_{0,t}}(A_t-A_{0,t}) +(A_t-A_{0,t})\# (A_t-A_{0,t}). $$
A direct calculation shows
\bea
&&\|\nabla_{x_j}( \tilde{\Upsilon}(\hat s_t) \bar\partial_{A_{t}}\partial_{A_{t}} \hat s_t)\|_w\nonumber\\
&& \qquad\qquad\leq C ( \|\nabla_{x_j}\hat s_t\|_w \|\nabla_{A_{t}}^2\hat s_t\|_{C^0}+\|\nabla_{x_j}\bar\partial_{A_{t}}\partial_{A_{t}} \hat s_t \|_w\| \hat s_t\|_{C^0}) \nonumber \\
&& \qquad\qquad \leq C \|\nabla_{x_j}\hat s_t\|_w (\|\nabla_{A_{0,t}}^2\hat s_t\|_{C^0}+\|A_t-A_{0,t}\|_{C^1}\|\hat s_t\|_{C^1}) \nonumber\\
& & \qquad\qquad \phantom{\leq}+C (\|\Delta_{A_t}\nabla_{x_j} \hat s_t \|_w + 1+t\Theta)\| \hat s_t\|_{C^0} \nonumber \\
&& \qquad\qquad\leq C (t\|\Delta_{A_t}\nabla_{x_j} \hat s_t \|_w+t^{\frac{1}{2} }\|\nabla_{x_j}\hat s_t\|_w +t+t^2\Theta) \nonumber \\
&& \qquad\qquad\leq C (t^{\frac{1}{2}}\|\Delta_{A_t}\nabla_{x_j} \hat s_t \|_w+ t+ t\Lambda +t^2\Theta), \nonumber \eea
where we used Lemma \ref{prop2+00}. For the later terms, we have
\bea & & \|\nabla_{x_j}(\bar\partial_{A_{t}} \tilde{\Upsilon}(\hat s_t) \wedge \partial_{A_{t}}\hat s_t)\|_w +\|\nabla_{x_j}(\Upsilon(\hat s_t)\partial_{A_{t}}\hat s_t \wedge \Upsilon(-\hat s_t)\bar\partial_{A_{t}}\hat s_t)\|_w \nonumber \\ && \qquad \leq C (\|\nabla_{A_{0,t}}\hat s_t\|_{C^0}+\|A_t-A_{0,t}\|_{C^0}\|\hat s_t\|_{C^0})(\|\nabla_{x_j}\hat s_t\|_w \nonumber\\
&&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad+\|d_{A_{t}}\nabla_{x_j}\hat s_t\|_w+\|\hat s_t\|_w) \nonumber\\
&& \qquad\leq C (t^2+t\|\Delta_{A_{t}}\nabla_{x_j} \hat s_t\|_w+t^2\Lambda). \nonumber
\eea
Returning to \eqref{thing7}, we put everything together to see $$\|\nabla_{x_j}F_{A_t}-i d_{A_{t}}\star_w d_{A_{t}}\nabla_{x_j} \hat s_t\|_w \leq C (t^{\frac{1}{2}}\|\Delta_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}+t\Lambda +t\Theta +t), $$
$$
\|\Delta_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}
\leq C (\Theta+t+t\Lambda),$$ and $$\|d_{A_{t}}\nabla_{x_j}\hat s_t\|_{w}\leq C (\Theta+t+t\Lambda). $$
Thus we conclude
$$\Lambda^2 \leq C (\Theta^2+t^2),$$
proving the lemma.
\end{proof}
Now, we are ready to prove Proposition \ref{prop2+0}.
\begin{proof}[Proof of Proposition \ref{prop2+0}]
Note that we have
\be
\|F_{ \Xi_t}\|^2_{L^2(M_{ U'},\omega_{ t})}\leq 2 \int_{M_{ U'}}( t^{-1}|F_{A_t}|^2_{\omega^{SF}}+\sum_{j=1,2}| \kappa_{ t,j}|^2_{\omega^{SF}}+ t|F_{ B,t}|^2_{\omega^{SF}})(\omega^{SF})^2.\nonumber
\ee
By (\ref{ASD2}), we have
\be
t|F_{ B,t}|^2_{\omega^{SF}}\leq C ( t^{-1}|F_{A_t}|^2_{\omega^{SF}}+t \sum_{j=1,2}| \kappa_{ t,j}|^2_{\omega^{SF}}),\nonumber
\ee
which in turn implies
\bea
\|F_{ \Xi_t}\|^2_{L^2(M_{ U'}, \omega_{ t})}&\leq& C \int_{ U'}(t^{-1}\|F_{A_t}\|^2_{w}+\sum_{j=1,2}\|\kappa_{ t,j}\|^2_{w})d x_1 d x_2\nonumber\\
&\leq& C ( t + \sum_{j=1,2}\int_{ U'} \|\kappa_{ t,j}-\kappa_{ t,j}^0\|^2_{w} d x_1 d x_2 \nonumber \\ & & + \sum_{j=1,2} \int_{U'} \|\partial_{x_j}A_{0,t}\|_w^2 d x_1 d x_2) \nonumber\\ &\leq& C ( t + \sum_{j=1,2} \int_{U'} \|\partial_{x_j}A_{0,t}\|_w^2 d x_1 d x_2)\nonumber
\eea
For the second inequality above we used $\|F_{A_t}\|^2_{w}\leq C t^2$ and $ \kappa_{ t,j}^0 =\partial_{x_j}A_{0,t} $.
Finally, \bea
\|F_{ \Xi_t}\|^2_{L^2(M_{ U'}, \omega_{ t})}&\geq & \frac{1}{2}\int_{ U'}\sum_{j=1,2}\|\kappa_{ t,j}\|^2_{w}d x_1 d x_2\nonumber\\
&\geq & \frac{1}{2} ( \sum_{j=1,2} \int_{U'} \|\partial_{x_j}A_{0,t}\|_w^2 d x_1 d x_2 \nonumber \\ & & - \sum_{j=1,2}\int_{ U'} \|\kappa_{ t,j}-\kappa_{ t,j}^0\|^2_{w} d x_1 d x_2 ) \nonumber\\ &\geq & C ( \sum_{j=1,2} \int_{U'} \|\partial_{x_j}A_{0,t}\|_w^2 d x_1 d x_2-t),\nonumber
\eea and we obtain the conclusion.
\end{proof}
We finish this section by a lemma that is needed in the proof of Theorem \ref{thm-main2}.
\begin{lem}\label{lem9.1} $$\sum_{j=1,2}\|[\kappa_{t,j}, \kappa_{t,j}]\|_{L^{2}(M_{U'},\omega^{SF})}^2 \leq C_7 t,$$ for a constant $C_7>0$.
\end{lem}
\begin{proof} Recall that
$ \kappa_{t,j}^0=\partial_{x_j} A_{0,t}$ by (\ref{curv+++}), and thus $$ [\kappa_{t,j}^0,\kappa_{t,j}^0]=0, \ \ \ j=1,2.$$
We have $$[\kappa_{t,j},\kappa_{t,j}]=2 [\kappa_{t,j}^0,\kappa_{t,j}-\kappa_{t,j}^0]+[\kappa_{t,j}-\kappa_{t,j}^0,\kappa_{t,j}-\kappa_{t,j}^0], $$ and by $|\kappa_{t,j}|\leq C$, $$|[\kappa_{t,j},\kappa_{t,j}]|\leq C | \kappa_{t,j}-\kappa_{t,j}^0|.$$
Lemma \ref{prop1} shows that
$$\int_{U}\sum_{j=1,2}\|[\kappa_{t,j},\kappa_{t,j}]\|_w^2 dx_1dx_2 \leq C \int_{U}\sum_{j=1,2}\|\kappa_{t,j}-\kappa_{t,j}^0\|_w^2 dx_1dx_2\leq C t.$$ We obtain the conclusion.
\end{proof}
\section{Proof of Proposition \ref{type3bubbling}}
\label{lowerbounds}
Now, we have the tools to verify assumption \eqref{hyp:curv} along our main subsequence of times $t_k$, which is chosen in Proposition \ref{estimate 1}.
\begin{proof}[Proof of Proposition \ref{type3bubbling}]
We work via contradiction, and assume the Proposition is false, in other words assumption \eqref{hyp:curv} fails for our sequence $\Xi_{t_k}$. By passing to a subsequence, there exists a sequence of points $p_k'\in M_K$ so that
\be
\label{contraassump}
t_k^{\frac{1}{2}}|F_{\Xi_{t_k}}|_{\omega_{t_k}}(p_k')\rightarrow\infty,
\ee
and $f(p_k')$ converges to a point $x\in K$, as $t_k\rightarrow 0$.
Applying Lemma \ref{pointpick}, we can pick new points near $p_k$ to carry out our argument. Specifically, if $r=\frac{1}{2}{\rm dist}_{\omega}(x, N\backslash K)$, there exists a sequence of real numbers $0<\rho_{k}<r$ and a sequence $p_k \in M$ so that $d_{\omega_{t_k}}(p_k,p_k')\leq r$,
\be
\sup_{B_{\omega_{t_k}}(p_k, \rho_{k})}
|F_{\Xi_{t_k}}|_{\omega_{t_k}}\leq 2|F_{\Xi_{t_k}}|_{\omega_{t_k}}(p_k),\nonumber
\ee
and
$$ 2\rho_{k}|F_{\Xi_{t_k}}|_{\omega_{t_k}}(p_k)\geq r |F_{\Xi_{t_k}}|_{\omega_{t_k}}(p_k').$$
If we set $\delta_{k}:=t_k^{-\frac{1}{2}}|F_{\Xi_{t_k}}|_{\omega_{t_k}}^{-1}(p_k)$, then \eqref{contraassump} and the above inequalities give $\delta_{k} \rightarrow 0$, and
$$\rho_{k} \delta_{k}^{-1}\geq r t_k^{\frac{1}{2}}|F_{\Xi_{t_k}}|_{\omega_{t_k}}(p_k') \rightarrow\infty.$$
Furthermore, define
\be
\ti t_k:= t_k \delta_{k}^{-2}={t_k}^2|F_{\Xi_{t_k}}|_{\omega_{t_k}}^{2}(p_k)\leq \epsilon_k^2\rightarrow 0,\nonumber
\ee
which goes to zero as $t_k\rightarrow0$ by Proposition \ref{estimate 1}.
We now consider the scaled matric $\tilde{\omega}_{\ti t_k}= \delta_{k}^{-2}\omega_{t_k}$, and claim that $\tilde{\omega}_{\ti t_k}$ satisfies the same collapsing properties of $\omega_{t_k}$. If $\tilde{w} = \delta_{k}^{-1} w$ denotes the scaled coordinate on $D_r=\{|w|<r\delta_{k}\}=\{|\tilde{w}|<r\}$, where $f(p_k)$ is given by $w=0$, then \be
\delta_{k}^{-2}\omega_{t_k}^{SF}=\frac {i}2\left(\tilde{t}_k W(dz+\tilde{b}d\ti w)\wedge\overline{(dz+ \tilde{b}d\ti w)}+W^{-1} d\ti w\wedge d\bar {\ti w}\right),\nonumber
\ee where $\ti b= -\frac{{\rm Im}(z)}{{\rm Im}(\tau)}\frac{\partial\tau}{\partial \ti w}$.
For a certain fiberwise translation $T_{\sigma_0}$, we write
\bea
T_{\sigma_0}^*\delta_{k}^{-2} \omega_{t_k}- \delta_{k}^{-2}\omega^{SF}_{t_k} & = & \delta_{k}^{-2}\varphi_{{t_k}, z \bar{z}}dz\wedge d\bar{z}+ \varphi_{{t_k}, w \bar{w}}d\tilde{w}\wedge d\bar{\tilde{w}}\nonumber \\ & & +\delta_{k}^{-1} \varphi_{{t_k}, w \bar{z}}d\tilde{w}\wedge d\bar{z}+ \delta_{k}^{-1} \varphi_{{t_k}, z \bar{w}}dz\wedge d\bar{\tilde{w}}.\nonumber
\eea
By Lemma \ref{lem-decay},
for $\nu\gg 1$, $$\|\delta_{k}^{-2}\varphi_{t_k, z \bar{z}}\|_{C_{\rm loc}^\ell} + \|\delta_{k}^{-1} \varphi_{t_k, z \bar{w}}\|_{C_{\rm loc}^\ell}+ \|\delta_{k}^{-1} \varphi_{t_k, w \bar{z}}\|_{C_{\rm loc}^\ell}\leq C_\ell \tilde{t}_k^\nu,$$
and
$$\|\frac{\partial}{\partial z}\varphi_{t_k, w \bar{w}}\|_{C_{\rm loc}^\ell}+\|\frac{\partial}{\partial \bar{z}}\varphi_{t_k, w \bar{w}}\|_{C_{\rm loc}^\ell}\leq C_{\ell} \tilde{t}_k^{\nu}, \ \ \,\,\, \ \ \|\varphi_{t_k, w \bar{w}}-\chi_{t_k,w\bar{w}}\|_{C_{\rm loc}^0}\leq C_0 \tilde{t}_k^{\nu}. $$
Here we used $t_k \leq \tilde{t}_k$, and that $\chi_{t_k, w \bar{w}}$ is a function on $D_r$ that satisfies $\chi_{t_k, w \bar{w}} \rightarrow 0$ in the $C^\infty$-sense as $t_k\rightarrow 0$. The $C_{\rm loc}^\ell$-norms are calculated in coordinates $z$ and $ \tilde{w}$.
Working in the scaled metrics, we have that $d_{\omega_{\ti t_k}}(p_k, p)\leq \rho_{k}\delta_{k}^{-1}$ for any $p\in B_{\omega_{t_k}}(p_k, \rho_{k})$, so the radius of the disk approaches infinity. In particular this implies that on $ B_{\ti \omega_{ \ti t_k}}(p_k,\rho_{k}\delta_{k}^{-1})$, we have the bound
\be
|F_{ \Xi_{t_k}}|_{\ti \omega_{\ti t_k}}= \delta_{k}^{2}|F_{\Xi_{t_k}}|_{\omega_{t_k}} \leq 2 \delta_{k}^{2} |F_{\Xi_{t_k}}|_{\omega_{t_k}}(p_k)= 2 t_k^{-1} |F_{\Xi_{t_k}}|_{\omega_{t_k}}^{-1}(p_k)= 2\ti t_k^{-\frac12}.\nonumber
\ee
Now, because the energy $\E_{t_k}(p, R_{t_k}(p_k))$ is scale invariant,
\bea
\varepsilon &= & \E_{t_k} (p_k, R_{t_k}(p_k))\nonumber\\ &=&
\frac{\delta_{k}^{-4}R_{t_k}(p_k)^{4}}{{\rm Vol}(B_{\ti \omega_{ \ti t_k}}(p_k,\delta^{-1}_{t_k}R_{t_k}(p_k)))}\int_{B_{\ti \omega_{\ti t_k}}(p_k,\delta^{-1}_{t_k}R_{t_k}(p_k))}|F_{ \Xi_{t_k}}|^2_{\ti \omega_{\ti t_k}}\ti\omega_{\ti t_k}^2. \nonumber
\eea
Additionally, note that
$$
\delta_{k}^{-1}R_{t_k}(p_k) = {t_k}^{\frac12}|F_{\Xi_{t_k}}|_{\omega_{t_k}}(p_k) R_{t_k}(p_k) \leq 4 t_k^{\frac12} |F_{\Xi_{t_k}}|_{\omega_{t_k}}^{\frac{1}{2}}(p_k) =4 \ti t_k^{\frac14},\nonumber
$$ since $ |F_{\Xi_{t_k}}|_{\omega_{t_k}}(p_k)\leq 4 R_{t_k}^{-2}(p_k)$ by (\ref{curvradconclusion}). Thus, on $B_{\ti \omega_{ \ti t_k}}(p_k,\rho_{k}\delta_{k}^{-1})$ we have \be
\label{thing10}
|F_{ \Xi_{t_k}}|_{\ti \omega_{\ti t_k}}\leq 2\ti t_k^{-\frac12}
\ee
and
\be
\label{thing9}
\varepsilon \leq\frac{4^4 \ti t_k}{{\rm Vol}(B_{\ti \omega_{ \ti t_k}}(p_k,4 \ti t_k^{\frac14}))}\int_{B_{\ti \omega_{\ti t_k}}(p_k,4 \ti t_k^{\frac14})}|F_{ \Xi_{t_k}}|^2_{\ti \omega_{\ti t_k}}\ti\omega_{\ti t_k}^2.
\ee
Inequality \eqref{thing10} gives assumption \eqref{hyp:curv} for our connections in scaled coordinates (with scaled parameter $\ti t$). Also \eqref{hyp:poincare} is also satisfied since the scaling does not effect the fiber direction. Thus Proposition \ref{prop2} holds in scaled coordinates, which in turn allows us to conclude Proposition \ref{prop2+0} as well.
To achieve our contradiction, we show these bounds force the energy on the right hand side of \eqref{thing9} to go to zero. We continue to use the notation $\|\cdot\|_{w}:=\|\cdot\|_{L^2(M_{w},\ti\omega^{SF})}$ since scaling does not affect the fiber direction.
Applying Proposition \ref{prop2+0}, on any $K\subset D_r$ we have
$$ \|F_{\Xi_{t_k}}\|_{L^2(M_{K}, \ti \omega_{\ti t_k})}^2 \leq C(\ti t_k +\int_{K}\sum_{j=1,2}\|\partial_{\tilde{x}_j}A_{0,t_k}\|_w^2 d\tilde{x}_1d\tilde{x}_2) $$
for a uniform constant $C$, where $\tilde{x}_1+i\tilde{x}_2=\tilde{w}$. Since $A_{0,t_k} \rightarrow A_{0}$ in the $C^\infty$-sense on $M_U$, we have
$$\|\partial_{\tilde{x}_j}A_{0,t_k}\|_w^2=\delta_k^2\|\partial_{x_j}A_{0,t_k}\|_w^2\leq C\delta_k^2, $$
and thus
$$ \|F_{\Xi_{t_k}}\|_{L^2(M_{K}, \ti \omega_{\ti t_k})}^2 \leq C(\ti t_k +\delta_k^2 \int_K d\tilde{x}_1d\tilde{x}_2 ). $$
Because the radius $\ti t_k^{\frac14}$ grows slower than the injectivity radius of the elliptic fibers in the metric $\ti \omega_{\ti t_k}$ (which is roughly $\ti t_k^{\frac12}$), we see that for $\ti t_k$ small enough
\be
\frac{\ti t_k}{{\rm Vol}(B_{\ti \omega_{ \ti t_k}}(p,4\ti t_k^{\frac14}))}\leq \frac{C \ti t_k }{\ti t_k \ti t_k^{\frac12}}=\frac{C }{ \ti t_k^{\frac12}}. \nonumber
\ee
Also $B_{\ti \omega_{ \ti t_k}}(p_k, 4\ti t_k^{\frac14})\subset M_{D_r}$. Thus, returning to \eqref{thing9}, we have
\bea
\varepsilon & \leq & \frac{4^4 \ti t_k}{{\rm Vol}(B_{\ti \omega_{ \ti t_k}}(p_k,4\ti t_k^{\frac{1}{4}}))}\int_{B_{\ti \omega_{\ti t_k}}(p_k,4\ti t_k^{\frac{1}{4}})}|F_{ \Xi_{t_k}}|^2_{\ti \omega_{\ti t_k}}\ti\omega_{\ti t_k}^2\nonumber \\
&\leq& \frac{C}{ \ti t_k^{\frac12}}(\ti t_k +\delta_k^2 \ti t_k^{\frac12} )\nonumber \\ &\leq & C(\ti t_k^{\frac12}+\delta_k^2).\nonumber
\eea
The right hand side above goes to zero, a contradiction.
\end{proof}
\section{The proof of Theorem \ref{thm-main2}}
At last, we prove Theorem \ref{thm-main2} in this section. Under the same setup as in Section 6, the first lemma shows that for any fixed $p\geq 2$, $$\| F_{B,t}\|_{L^p(M_U, \omega^{SF})}\rightarrow 0$$ when $t \rightarrow 0$.
\begin{lem}\label{prop9.2} If (\ref{hyp:curv}) and (\ref{hyp:poincare}) hold for $t\ll 1$, for any $p\geq 2$, we have
the following inequalities $$ \| F_{A_t}\|_{L^p(M_U, \omega^{SF})}^p\leq C_1t^{1+p}, \ \ and \ \ \| F_{B,t}\|_{L^p(M_U, \omega^{SF})}^p\leq C_1t^{1+\frac{1}{p}} , $$ where the constant $C_1$ is independent of $t$.
\end{lem}
\begin{proof} By Lemma \ref{le2},
$$ \Delta \| F_{A_t}\|_w^2 \geq \frac{\delta}{t} \| F_{A_t} \|_w^2 - CtZ_t, $$
where $$Z_t=\sum_{j=1,2}\|[\kappa_{t,j}, \kappa_{t,j}]\|_w^2 +t^\nu,$$ for $\nu\gg 1$ and a constant $C>0$. Lemma \ref{lem9.1} implies that $$\int_U Z_tdx_1dx_2 \leq Ct.$$
Let $\eta$ be a smooth function such that $0\leq \eta \leq 1$ and ${\rm supp} (\eta) \subset U$. Then
\begin{eqnarray*} \int_U \| F_{A_t} \|_w^2 dx_1dx_2 & \leq & t \delta^{-1} \int_U \| F_{A_t} \|_w^2 \Delta \eta dx_1dx_2 +t^2 C \int_U \eta Z_tdx_1dx_2\\ & \leq & t\tilde{C}\delta^{-1} \int_U \| F_{A_t} \|_w^2 dx_1dx_2+ t^2 C \int_U Z_tdx_1dx_2, \end{eqnarray*} for a constant $\tilde{C}\geq \sup\limits_{U} \Delta \eta$. Thus for $t\ll 1$, $$ \int_U \| F_{A_t} \|_w^2 dx_1dx_2 \leq Ct^{3}. $$
For any $p\geq 2$, $$ \| F_{A_t}\|_{L^p(M_U, \omega^{SF})}^p\leq C t^{p-2} \int_U \| F_{A_t} \|_w^2 dx_1dx_2 \leq Ct^{p+1}, $$ by Lemma \ref{le-l2}, and $$ \| F_{B,t}\|_{L^p(M_U, \omega^{SF})}^p\leq Ct^{1+\frac{1}{p}}, $$ by (\ref{ASD2}).
\end{proof}
Recall that for any sequence $t_k \rightarrow 0$, a subsequence of $\Xi_{t_k}$ $L^p_1\cap C^{0,\alpha}$-converges to a $L^p_1\cap C^{0,\alpha}$-connection $\Xi_0$ by preforming certain further unitary gauge changes if necessary on $M_K$
in Theorem \ref{thm-main}, where $K\subset N^o$. Thus the curvature $F_{\Xi_{t_k}}$ $L^p$-converges to $F_{\Xi_{0}}$ on $M_K$.
On any open disc $U\subset K$, we have the decompositions $$\Xi_{0}=\tilde{A}_0+\tilde{B}_{0,1}dx_1+\tilde{B}_{0,2}dx_2, \ \ \ {\rm and}$$ $$F_{\Xi_{0}} =F_{\tilde{A}_0}-\tilde{\kappa}_{0,1}\wedge dx_1-\tilde{\kappa}_{0,2}\wedge dx_2-F_{\tilde{B},0} dx_1\wedge dx_2,$$ where $\tilde{\kappa}_{0,j}=\frac{\partial}{\partial x_j} \tilde{A}_0-d_{\tilde{A}_0}\tilde{B}_{0,j} $.
By Lemma \ref{prop9.2} and the convergence, we obtain that $$F_{\tilde{A}_0}\equiv 0, \ \ F_{\tilde{B},0} \equiv 0, \ \ {\rm and} \ \ \star_w \tilde{\kappa}_{0,1}=\tilde{\kappa}_{0,2}.$$ Thus $\Xi_{0}$ is an anti-self-dual connection with respect to $(\omega^{SF}, \Omega)$, i.e. $$ F_{\Xi_{0}}\wedge \omega^{SF} =0, \ \ {\rm and} \ \ F_{\Xi_{0}}\wedge \Omega =0.$$ It is standard (cf. Theorem 9.4 of \cite{Weh2}) that by preforming a further unitary gauge change if necessary, we can have that $\Xi_{0} $ is smooth.
\begin{lem}\label{prop9.3} There is a unitary gauge $u$ such that $$u(\Xi_{0})=A_0$$ on $M_U$, where $A_0$ is given by (\ref{bconnections2}).
\end{lem}
\begin{proof} By Theorem \ref{thm-main}, for any $w\in U$, there is a unitary gauge $u_w$ on $M_w$ such that $u_w(\Xi_{0}|_{M_w})=A_0|_{M_w}$, and $u_w$ is smooth since both $\Xi_{0}|_{M_w}$ and $A_{0}|_{M_w}$ are smooth. We claim that one can choose $u_w$ depending on $w$ smoothly.
Note that $M_w \cong T^2$ and $P|_{M_w}\cong M_w\times SU(n)$. Let $\mathcal{A}^{\ell, p}$ be the space of $L^{ p}_{\ell}$ $SU(n)$-connections on the trivial bundle on $T^2$, $\ell \geq 1 $, and $\mathcal{G}^{\ell+1, p}$ be the $L^{ p}_{\ell+1}$ unitary gauge group. We have identifications $\mathcal{A}^{\ell, p}= L^{ p}_{\ell}(T^2, \mathfrak{sl}(n))$ and $\mathcal{G}^{\ell+1, p}=L^{ p}_{\ell+1}(T^2, SU(n))$ under the trivialization, and $\mathcal{G}^{\ell+1, p}$ acts on $\mathcal{A}^{\ell, p}$ by $u(A)=u^{-1}Au+u^{-1}du$. If we denote the orbit $O_w=\{u(\Xi_{0}|_{M_w})| u\in \mathcal{G}^{\ell+1, p}\}\subset \mathcal{A}^{\ell, p}$ for any $w\in U$, then $ A_{0}|_{M_w} \in O_w$. Define the orbit map $$\Psi: \mathcal{G}^{\ell+1, p}\times U \rightarrow \bigcup_{w\in U}O_w\subset \mathcal{A}^{\ell, p}, \ \ {\rm by} \ \ \Psi(u,w)=u(\Xi_{0}|_{M_w}). $$
For a fixed $w_0\in U$, let $\varrho_w: O_w\rightarrow O_{w_0}$ by $A \mapsto v(A_{0}|_{M_{w_0}})$, where $v(A_{0}|_{M_w})=A$ for a unitary gauge $v$. If $v'$ is an another unitary gauge such that $v'(A_{0}|_{M_w})=A$, then $v'v^{-1}(A_{0}|_{M_w})=A_{0}|_{M_w}$, and thus $v'v^{-1}\in T^{n-1}\subset SU(n)$, i.e. a diagonal matrix. Since $A_{0}|_{M_{w_0}}$ is a diagonal matrix valued 1-form, we have $v (A_{0}|_{M_{w_0}})=v' (A_{0}|_{M_{w_0}})$, and
$\varrho_w$ is well-defined.
Let $\Psi'=\varrho_w\circ \Psi: \mathcal{G}^{\ell+1, p}\times U \rightarrow O_{w_0}$ be the the composition.
Note that the tangent space $T_{A_{0}|_{M_{w_0}}}O_{w_0}={\rm Im}(d_{A_{0}|_{M_{w_0}}})$, and the first partial derivative of $\Psi'$ at $(u,w)$ such that $\Psi'(u,w)= A_{0}|_{M_{w_0}}$ is $D_1 \Psi'=-d_{A_{0}|_{M_{w_0}}}$. Thus $A_{0}|_{M_{w_0}}$ is a regular value of $\Psi'$, and $\Psi'^{-1}(A_{0}|_{M_{w_0}})$ is a smooth submanifold. Furthermore, the projection $\mathcal{G}^{\ell+1, p}\times U\rightarrow U$ induces a $T^{n-1}$-bundle structure on $\Psi'^{-1}(A_{0}|_{M_{w_0}})$ with fiber $T^{n-1}\subset SU(n)$.
If $\tilde{u}: U\rightarrow \Psi'^{-1}(A_{0}|_{M_{w_0}})$ is a smooth section, then $\tilde{u}(w)(\Xi_{0}|_{M_w})=A_{0}|_{M_w}$, and we can regard $\tilde{u}$ as a smooth unitary gauge change on $M_U$. Therefore we have
$$\tilde{u}(\Xi_{0})=A_{0}+ B_{0,1}dx_1+B_{0,2}dx_2, $$ which still satisfies $$\star_w \kappa_{0,1}=\kappa_{0,2}, \ \ {\rm with} \ \ \kappa_{0,j}=\frac{\partial}{\partial x_j} A_0-d_{A_0}B_{0,j}, \ \ j=1,2, \ \ {\rm and} $$ $$0=F_{B,0}=\frac{\partial}{\partial x_2} B_{0,1}-\frac{\partial}{\partial x_1} B_{0,2}-[B_{0,1},B_{0,2}].$$
Note that $\frac{\partial}{\partial x_j} A_0 \in \ker \Delta_{A_0}$, $j=1,2$, on any $M_w$, and $$\star_w \frac{\partial}{\partial x_1} A_0-\frac{\partial}{\partial x_2} A_0= \star_w d_{A_0}B_{0,1}-d_{A_0}B_{0,2} . $$ By the Hodge decomposition, $\ker \Delta_{A_0}$, ${\rm Im} ( d_{A_0}^*)$ and ${\rm Im} (d_{A_0})$ are orthogonal to each other. Thus $$d_{A_0}B_{0,j} \equiv 0, \ \ j=1,2,$$ on any $M_w$, and $B_{0,j}|_{M_w}$ is a diagonal matrix in $\mathfrak{sl}(n) $. If we write $B_{0,j}=i{\rm diag}\{b_{j,1}, \cdots, b_{j,n}\}$, then $\frac{\partial}{\partial x_2} B_{0,1}=\frac{\partial}{\partial x_1} B_{0,2}$ implies that there are real functions $ \vartheta_{\ell}$ on $U$ such that $b_{1,\ell}dx_1+b_{2,\ell}dx_2=- d \vartheta_{\ell}$, $\ell=1, \cdots,n$. If $\tilde{v}={\rm diag}\{\exp(i\vartheta_{1}), \cdots, \exp(i\vartheta_{n})\}$, and we regard $\tilde{v}$ as a unitary gauge change on $M_U$, then $$ \tilde{v}(\tilde{u}(\Xi_{0}))=A_{0}.$$ We obtain the conclusion by letting $u=\tilde{v}\cdot\tilde{u}$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm-main2}] Let $\{U_\lambda| \lambda\in\Lambda\}$ be an open cover of $N^o$ such that any intersection $U_{\lambda_1}\cap \cdots \cap U_{\lambda_h}$ is contractible. For any $U_\lambda$, $D_0^o \cap M_{U_\lambda}=U_\lambda^1 \cup \cdots \cup U_\lambda^n$ is a disjoint union of open sets biholomorphic to $U_\lambda$, and $\{U_\lambda^j|\lambda\in\Lambda, j=1, \cdots, n \}$ is an open cover of $D_0^o \cap M_{N^o}$ such that any intersections are contractible.
On any $M_{U_\lambda}$, there is a unitary gauge $u_\lambda$ such that $u_\lambda(\Xi_0)=A_0$ by Lemma \ref{prop9.3}. Recall that $$A_0={\rm diag}\{\alpha_1, \cdots, \alpha_n\}, \ \ \ \alpha_j=\pi({\rm Im}( \tau))^{-1}(q_j \bar{\theta}-\bar{q}_j\theta),$$ where $\{(w, q_j(w))\}=U_\lambda^j$ is one component of $D_0^o\cap M_{U_\lambda}$, and $\alpha_j$ is not unitary gauge equivalent to $\alpha_i$ if $j\neq i$.
On any intersection $M_{U_\lambda \cap U_\mu}$, $A_0=u_\mu \cdot u_\lambda^{-1}(A_0)$. Thus $ u_\mu \cdot u_\lambda^{-1} |_{M_w} \in T^{n-1}\subset SU(n)$ for any $w\in U_\lambda \cap U_\mu$. We can write $u_\mu \cdot u_\lambda^{-1}= {\rm diag}\{g_{\mu\lambda}^{1j_1}, \cdots, g_{\mu\lambda}^{nj_n}\}$, where $g_{\mu\lambda}^{ij_i}$ is a $U(1)$-valued function on $U_\lambda \cap U_\mu$, and is the unitary gauge change between $\alpha_i$ on $M_{U_\mu}$ and $\alpha_{j_i}$ on $M_{U_\lambda}$. Hence we have that $U_\mu^i \cap U_\lambda^{j_i} \neq\emptyset$, and $d \log g_{\mu\lambda}^{ij_i} =0$, which implies that $g_{\mu\lambda}^{ij_i}$, $i=1, \cdots, n$, are $U(1)$-valued constant functions on $U_\lambda \cap U_\mu$. By regarding $g_{\mu\lambda}^{ij_i}$ as a function on $U_\mu^i \cap U_\lambda^{j_i}$, we obtain a 1-chain $\{(U_\mu^i \cap U_\lambda^{j_i}, g_{\mu\lambda}^{ij_i})\} \in \mathcal{C}^1(\{U_\lambda^j\}, \mathcal{U}_c(1))$ for the $U(1)$-valued locally constant sheaf $\mathcal{U}_c(1)$ on $D_0^o \cap M_{N^o}$.
If $U_\mu^i \cap U_\lambda^{j}\cap U_\nu^{k}\neq\emptyset$, then $U_\mu \cap U_\lambda \cap U_\nu \neq\emptyset$, and by $u_\mu \cdot u_\lambda^{-1} \cdot u_\lambda \cdot u_\nu^{-1}\cdot u_\nu \cdot u_\mu^{-1}={\rm Id}$, we obtain that $g_{\mu\lambda}^{ij}g_{\lambda\nu}^{jk}g_{\nu\mu}^{ki}=1$. Therefore $\{(U_\mu^i \cap U_\lambda^{j_i}, g_{\mu\lambda}^{ij_i})\}$ satisfies the cocycle condition, and
defines a cohomological class $\Theta=[\{(U_\mu^i \cap U_\lambda^{j_i}, g_{\mu\lambda}^{ij_i})\}]\in H^1(D_0^o \cap M_{N^o}, \mathcal{U}_c(1))$, which is equivalent to a flat $U(1)$-connection on $D_0^o \cap M_{N^o}$. From the construction in Subsection 2.6, it is clear that $\Xi_0\in \mathcal{FM}(D_0^o \cap M_{N^o}, \Theta)$. \end{proof}
|
\section{Introduction}
In a broad sense, an object is called \emph{quasirandom} if, asymptotically, it has similar properties to random objects of the class it belongs to. The notion of quasirandomness has been studied for a variety of objects, including groups \cite{gowers07}, graphs \cite{chung89, lovasz08}, and $k$-uniform hypergraphs \cite{gowers06}. For permutations in particular \cite{cooper04, cooper08}, several different definitions of randomness are equivalent to a single concept of a \emph{quasirandom} permutation sequence. One such definition states that in the limit, as the lengths of the permutations in a sequence grow toward infinity, the densities of all pattern permutations of length $k$ approach $1/k!$.
The study of regularity and quasirandomness is an active area of research, and there have been fruitful prior results \cite{sliacan08}. In 2012, Kral and Pikhurko \cite{kral12} proved that if a permutation has uniform densities for all permutations of length 4, then it is quasirandom. With this result in mind, it is natural to consider permutation sequences that have some set of uniform densities but are not quasirandom.
Inflatable permutations are one example of a general construction for permutation sequences. We define the inflation of two permutations $\tau$ and $\gamma$ with lengths $n$ and $m$ to be a new permutation with length equal to the product $nm$. For each index $i \in [n]$ (the range of integers from $1$ to $n$) in the first permutation $\tau$, we replace it with $m$ numbers $n(\tau_i - 1) + \gamma_j$ for each $j \in [m]$. In a sense, we are ``inflating'' each number in the first permutation by substituting for it $m$ numbers in the order of the second permutation.
We can use this definition to build convergent permutation sequences. The main idea is to start from a base permutation and take its inflation with a sequence of random permutations with lengths tending to infinity. The $n$-th term of this sequence is the inflation of the base permutation with a permutation chosen uniformly at random from all elements of $S_n$.
A permutation is called $k$-inflatable when this construction results in a permutation sequence that has uniform densities of all length-$k$ subpermutations. Cooper and Petrarca \cite{cooper08} previously studied this topic in 2008, finding examples of 3-inflatable permutations. The result from Kral and Pikhurko \cite{kral12} implies that there do not exist any nontrivial $4$-inflatable permutations. In general, nontrivial $k$-inflatable permutations do not exist for $k \ge 4$.
We decided to write this paper when we discovered via random sampling that the example of a 3-inflatable permutation of length 9 in \cite{cooper08} is wrong. We found an explicit formula that calculates densities in permutation inflations and used this to find constraints on admissible lengths of such permutations. This proved that the shortest length of a 3-inflatable permutation is 17, and we found many examples of this length with an optimized computer search. We also found a way to multiply 3-inflatable permutations that provides an infinite number of examples.
Code for calculating the limit densities of patterns within inflations, searching for 3-inflatable permutations, and running randomized tests is available at \url{https://github.com/ekzhang/inflatable}.
We start with basic definitions in Section~\ref{sec:preliminaries}. This includes formal definitions for permutation density and the convergence of permutation sequences, as well as a definition of quasirandomness for permutations.
In Section~\ref{sec:inflation}, we prove a formula for the limit density of a permutation $\pi$ of length $k$ in the inflation of a permutation $\tau$ with permutations of a convergent sequence $\{ \gamma_j \}$, given the limit densities of patterns with lengths up to $k$ in that sequence. This allows the limit densities of induced patterns in inflations to be computed efficiently.
In Section~\ref{sec:random-inflation}, we examine the asymptotic properties of fixed permutations when inflated by uniformly chosen random permutations of increasing length.
In Section~\ref{sec:ksymmetry} we apply the formula to densities of permutations of length 3. We calculate necessary and sufficient conditions for a permutation to be $3$-inflatable. From this, we deduce that the lengths of $3$-inflatable permutations must have remainders 0, 1, 17, 64, 80, or 81 modulo 144. By computational search, we find examples of $3$-inflatable permutations for each of these lengths, including the shortest possible examples, with length $17$.
In Section~\ref{sec:structure3} we examine the structure of $3$-inflatable permutations. We show that if two permutations are both $3$-inflatable, then the inflation of one with the other is also $3$-inflatable. We then discuss how a rotational symmetry helps us find examples of $3$-inflatable permutations by reducing the number of needed equivalences.
\section{Preliminaries}
\label{sec:preliminaries}
A \textit{permutation} is an ordering of the elements of a set. We represent a permutation of length $n$ as an $n$-tuple of distinct positive integers, representing the image of $[n]$ under application of the permutation. For example, $(1, 2, 3)$ is the identity permutation of length 3, while $\tau = (3, 2, 1)$ is the permutation that swaps 1 and 3. For convenience, we can leave out the parentheses and write $\tau = 321$. We denote the length of the permutation $\tau$ as $|\tau|$.
We are interested in patterns that are formed by subsets in a given permutation. Suppose we have an ordered $k$-tuple $(a_1, \ldots, a_k)$ of distinct positive integers. We say that this tuple is \textit{order-isomorphic} to a permutation $\pi$ of size $k$, if and only if $a_i < a_j \iff \pi_{i} < \pi_{j}$ for all $i, j \in [k]$.
The \textit{density} of a pattern permutation $\pi$ of length $k$ in a permutation $\tau$ is defined as the probability that a randomly-selected $k$-point subset of $\tau$ is order-isomorphic to $\pi$. We denote this density by $t(\pi, \tau)$. For example, $t(12, 132) = 2/3$.
A sequence of permutations $\{ \tau_j \}$ is \textit{convergent} if and only if as $j$ increases, the length of the permutations $|\tau_j|$ approaches infinity, and for any permutation $\pi$, the sequence of densities $t(\pi, \tau_j)$ converges. The limit $\displaystyle \lim_{j \rightarrow\infty} t(\pi, \tau_j)$ is called the \textit{limit density} of $\pi$ in the sequence $\{\tau_j\}$.
This notion of density can be used to describe the property of ``randomness'' in a permutation. Notice that given any pattern permutation $\pi$, if one selects a permutation $\lambda$ uniformly at random from $S_n$ for $n \ge |\pi|$, we have by symmetry that $\E t(\pi, \lambda) = \frac{1}{|\pi|!}$. This motivates the following definition of what it means for a convergent sequence of permutations to be \textit{quasirandom}.
\begin{definition}
A convergent sequence of permutations $\{\tau_j\}$ is called \emph{quasirandom} if for every permutation $\pi$, \[
\lim_{j\rightarrow\infty} t(\pi, \tau_j) = \frac{1}{|\pi|!}.
\]
\end{definition}
\section{Inflation}
\label{sec:inflation}
We define an inflation of one permutation with respect to another permutation. Cooper and Petrarca call this the \textit{tensor product} of two permutations \cite{cooper08}.
\begin{definition}
The \textit{inflation} of $\tau \in S_n$ with respect to a permutation $\gamma \in S_m$ is defined as a permutation, denoted by $\inflate(\tau, \gamma)$, of length $mn$ that consists of $n$ blocks of length $m$, which are each order-isomorphic to $\gamma$, and such that any restriction of the permutation to one element in each block is order-isomorphic to $\tau$.
\end{definition}
For example, if $\tau = 12$ and $\gamma = 312$, then $\inflate(\tau, \gamma) = 312645$. The resulting permutation consists of two blocks 312 and 645, each of which is order-isomorphic to $\gamma$. Also, each number in the second block is greater than every number in the first block, as dictated by permutation $\tau$. As another example, $\inflate(\gamma, \tau) = 561234$.
Inflations of permutations are interesting objects of study, as it is possible to calculate densities in $\inflate(\tau, \gamma)$ through densities in $\tau$ and $\gamma$. This later allows for the construction of interesting examples of permutation sequences with some set densities.
There is a more general definition of inflation, which turns out to be useful for computing densities. For some permutation $\tau \in S_n$, rather than inflating $\tau$ with respect to a single permutation $\pi$, we can define an operation on $\tau$ with respect to a sequence of $n$ permutations $\gamma_1, \ldots, \gamma_n$ as follows.
\begin{definition}
The \textit{generalized inflation} of $\tau \in S_n$ with respect to a sequence of $n$ permutations $\gamma_1, \ldots, \gamma_n$ is defined as a permutation of length $|\gamma_1| + \cdots + |\gamma_n|$ that consists of $n$ blocks, such that the $i$-th block is order-isomorphic to $\gamma_i$, and any restriction of the permutation to one element in each block is order-isomorphic to $\tau$.
\end{definition}
Note that inflation is a special case of generalized inflation, where all the $\gamma_i$ for $i \in [n]$ are equal to some single permutation $\gamma$.
For example, the generalized inflation of the permutation $\tau = 231$ with the sequence $\{\gamma_j\} = (12, 231, 1)$ is the length-6 permutation $23|546|1$. Note how each element in the original permutation $\tau$ corresponds to a ``block'' of elements in the new permutation.
Given a permutation $\pi$, it is useful to see how $\pi$ can be represented as a generalized inflation. To this end, consider pairs $(b, \sigma)$, where $\sigma$ is a permutation and $b$ is a sequence of permutations $(\sigma_1, \ldots, \sigma_k)$ so that $\pi$ is the generalized inflation of $\sigma$ with respect to $\sigma_1, \ldots, \sigma_k$. We call the pair $(b, \sigma)$ a \emph{block-partition} of $\pi$, and define the function $B(\pi)$ to be the set of all \emph{block-partitions} of $\pi$. Each block-partition of a permutation is a way of representing it as a generalized inflation.
For example, if $\pi = 132$, then we have three block-partitions: first, $1|3|2$ with $b=(1, 1, 1)$ and $\sigma = 132$; second, $1|32$ with $b=(1,21)$ and $\sigma = 12$; and third, $132$ with $b=(132)$ and $\sigma = 1$.
This definition allows us to derive an explicit formula for the limit densities of patterns in the inflation of a permutation with a convergent sequence $\{ \gamma_j \}$.
\begin{theorem} \label{thm:inflate-density} The limit density of $\pi$ in $\inflate(\tau, \gamma_j)$ is
\[
\lim_{j\rightarrow\infty} t(\pi, \inflate(\tau, \gamma_j)) = \frac{|\pi|!}{|\tau|^{|\pi|}} \sum_{(b, \sigma) \in B(\pi)} \left(\binom{|\tau|}{|\sigma|} t(\sigma, \tau) \cdot \prod_{\alpha \in b} \frac{\lim_{j\rightarrow\infty} t(\alpha, \gamma_j)}{|\alpha|!}\right).
\]
\end{theorem}
\begin{proof}
There are $|\tau|^{|\pi|}$ ways of selecting (counting distinct orderings) $|\pi|$ blocks with replacement from the inflation of $\tau$.
For a given pair $(b, \sigma) \in B(\pi)$, let $k = |\sigma|$. There exist $\binom{|\tau|}{k} \cdot t(\sigma, \tau)$ order-respecting assignments of distinct blocks $a_1, a_2, \dots, a_{k}$ in the inflation of $\tau$ to the corresponding indices in $\sigma$. Furthermore, since we count distinct orderings, the number of ways to select these $k$ blocks with their respective multiplicities is given by the multinomial coefficient \[
\binom{|\pi|}{|\sigma_1|; |\sigma_2|; \dots; |\sigma_k|} = |\pi|! \cdot \prod_{\alpha \in b} \frac{1}{|\alpha|!}.
\]
The probability that a random $|\pi|$-point subset from the inflation of $\tau$ is order-isomorphic to $\pi$, and also blocked according to $a_1, a_2, \dots, a_k$, is thus equal to \[
\frac 1{|\tau|^{|\pi|}} \left( |\pi|! \cdot \prod_{\alpha \in b} \frac{1}{|\alpha|!} \right) \left( \prod_{\alpha \in b} \lim_{j\rightarrow\infty} t(\alpha, \gamma_j) \right) = \frac{|\pi|!}{|\tau|^{|\pi|}} \prod_{\alpha \in b} \frac{\lim_{j\rightarrow\infty} t(\alpha, \gamma_j)}{|\alpha|!}.
\]
Summing these probabilities for all assignments of $\{a_j\}$ mentioned above, as well as for each block-partition in $B(\pi)$, we arrive at the final formula.
\end{proof}
For example, let $\pi = 12$. Then $B(\pi)$ consists of two elements $((1, 1), 12)$ and $((12), 1)$. Plugging this into the formula above, we get\footnote{For clarity, here we abuse notation by taking $\inflate(\tau, \sigma)$ to mean, for $\{ \sigma_j \}$ being a convergent permutation sequence, the sequence of inflations $\{ \inflate(\tau, \sigma_j) \}$.}
\[
t(12, \inflate(\tau, \sigma)) = \frac{2!}{|\tau|^2} \left( \binom{|\tau|}2 t(12, \tau)/2 + \frac{|\tau|}4 \cdot t(12, \sigma)\right).
\]
In particular, if $t(12, \tau)= t(12, \sigma) = 1/2$, then $t(12, \inflate(\tau, \sigma)) = 1/2$. However, it is possible that $t(12, \inflate(\tau, \sigma)) = 1/2$, while $t(12, \tau) \neq t(12, \sigma)$.
\section{Uniform inflation}
\label{sec:random-inflation}
A special case of inflation is of particular interest to us, when we inflate a fixed permutation with a random second permutation. In particular, given a permutation $\pi$ and a uniform random permutation $\gamma$ of length not less than $|\pi|$, the expected value of the density $\E t(\pi, \gamma) = \frac{1}{|\pi|!}$ for symmetry reasons.
Using this idea, we can define the \emph{uniform inflation} of a single permutation as the convergent sequence of its inflations with random permutations of length tending to infinity.
\begin{definition}
The \emph{uniform inflation} of $\tau$, denoted by $\inflate(\tau)$, is a sequence of random permutations $\{ \inflate(\tau, \lambda_j) \}$, where $\lambda_j$ is selected uniformly at random from $S_j$. In other words, the $j$-th term of the sequence is the inflation of $\tau$ with a uniform random permutation of length $j$.
\end{definition}
It is not difficult to show that the densities of pattern permutations in the uniform inflation of $\tau$ converge for any permutation $\tau$, so the sequence is convergent. This convergence fact allows us to extend Theorem~\ref{thm:inflate-density} to the case of uniform inflation by plugging in $t(\alpha, \gamma) = \frac{1}{|\alpha|!}$, which yields the following result.
\begin{corollary}
\label{random-density}
If $\{ \lambda_j \}$ is a sequence of random permutations with lengths tending to infinity, then the limit density of $\pi$ in the sequence of permutations $\{\inflate(\tau, \lambda_j)\}$ is
\[
t(\pi, \inflate(\tau)) =
\frac{|\pi|!}{|\tau|^{|\pi|}} \sum_{(b, \sigma) \in B(\pi)} \left(\binom{|\tau|}{|\sigma|} t(\sigma, \tau) \cdot \prod_{\alpha \in b} \frac{1}{|\alpha|!^2}\right).
\]
\end{corollary}
\subsection{Example}
As we are primarily interested in the limit densities of length-3 pattern permutations, for our next example we choose $\pi = 132$. Consider some general permutation $\tau$ of length $n$.
The permutation 132 admits 3 block-partitions: $132$, $1|32$, and $1|3|2$. Our formula yields for the limit density of 132 in the inflation of $\tau$ with a random sequence:
\begin{align*}
&\frac{3!}{n^3} \left(
\binom{n}{1}t(1, \tau) \cdot \frac{1}{3!^2} +
\binom{n}{2}t(12, \tau) \cdot \frac{1}{1!^2 2!^2} +
\binom{n}{3}t(132, \tau) \cdot \frac{1}{1!^2 1!^2 1!^2}
\right) \\
&\qquad= \frac{3!}{n^3}
\left(
\frac{1}{36}\binom{n}{1} +
\frac{1}{4}\binom{n}{2}t(12, \tau) +
\binom{n}{3}t(132, \tau)
\right).
\end{align*}
Notice from this expression that there exists additional structure if we focus our attention on length-3 pattern permutations. Let us denote first $\frac{3!}{n^3} \binom{n}{3}$ by $a(n)$, second $\frac{3!}{n^3}\frac{1}{4}\binom{n}{2}$ by $b(n)$, and third $\frac{3!}{n^3}\frac 1{36}\binom n1$ by $c(n)$. Substituting yields \[
t(132,\inflate(\tau))= a(n)t(132,\tau) + b(n) t(12,\tau)+ c(n).
\]
Also, the limit densities of permutations 213, 312 and 231 in the inflation of $\tau$ are given by analogous formulas:
\[
t(213,\inflate(\tau))= a(n)t(213,\tau) + b(n) t(12,\tau)+ c(n),
\] \[
t(312,\inflate(\tau))= a(n)t(312,\tau) + b(n) t(21,\tau)+ c(n),
\] \[
t(231,\inflate(\tau))= a(n)t(231,\tau) + b(n) t(21,\tau)+ c(n).
\]
The reason for this similarity is that the block-partitions of $213$, $312$, and $132$ all follow the same pattern. Each has a partition into one block of size 3, two blocks of sizes 1 and 2, and three blocks of size 1.
Similarly, permutations 123 and 321 follow similar formulas, except the coefficient of the second term is doubled since there are two ways to partition each of $123$ and $321$ into two blocks. These limit densities are given by \[
t(123,\inflate(\tau))= a(n)t(123,\tau) + 2 b(n) t(12,\tau)+ c(n),
\] \[
t(321,\inflate(\tau))= a(n)t(321,\tau) + 2 b(n) t(12,\tau)+ c(n).
\]
To see an example of this symmetry, consider the length-9 permutation $\tau = 472951836$. The density of $132$ in $\tau$ is $17/84$, and the density of $12$ in $\tau$ is $1/2$. Therefore, \[
t(132, \inflate(\tau)) = \frac{29}{162}.
\]
Similarly, we then have \[
t(213, \inflate(\tau)) = t(231, \inflate(\tau)) = t(312, \inflate(\tau)) = \frac{29}{162}.
\] \[
t(123, \inflate(\tau)) = t(321, \inflate(\tau)) = \frac{23}{162}.
\]
\section[Inflation and k-symmetry]{Inflation and $k$-symmetry}
\label{sec:ksymmetry}
We now turn our attention to symmetric permutation sequences.
\begin{definition}
A convergent permutation sequence $\{\tau_j\}$ is called \textit{$k$-symmetric} (as defined in \cite{cooper08}) if for every permutation $\pi$ of length $k$, the limit density of $\pi$ in $\tau_j$ is $1/k!$.
\end{definition}
\begin{definition}
A permutation $\tau$ is called \textit{$k$-inflatable} if $\inflate(\tau)$ is $k$-symmetric.
\end{definition}
As a direct corollary of Theorem~\ref{thm:inflate-density} permutation $\tau$ is 2-inflatable if and only if $t(12, \tau) = 1/2$. We are interested in 3-inflatable permutations, that is that the $\inflate(\tau)$ sequence has 3-point densities all equal to $1/3!$.
It was claimed in Cooper and Petrarca \cite{cooper08} that the smallest $3$-inflatable permutations are of length $9$. They provide as examples $472951836$ and its inverse. However, by our calculations above, this is not the case. We have also verified this empirically by estimating the limit densities in the inflation using a Monte Carlo method, and the results were consistent with our formula.
\subsection[k-inflatable permutations]{$k$-inflatable permutations}
Theorem~\ref{thm:inflate-density} allows us to calculate the densities of the permutation we inflate, so that the result is 3-inflatable.
\begin{theorem} A permutation $\tau$ of length $n$ is $3$-inflatable if and only if $t(12, \tau) = 1/2,$ and
$$t(123, \tau) = t(321, \tau) = (2 n - 7)/(12 (n - 2)),$$
$$t(132, \tau) = t(213, \tau) = t(231, \tau) = t(312, \tau) = (4 n - 5)/(24 (n - 2)).$$
\end{theorem}
\begin{proof}
First, a 3-inflatable permutation must also be 2-inflatable, so the density of 12 in $\tau$ must be 1/2.
A specific application of Corollary~\ref{random-density} for three-point densities yields \[
t(123, \inflate(\tau)) = \frac{9 \cdot 8 \cdot 7}{9^3} t(123, \tau) + \binom 92 \cdot \frac{3}{9^3} \cdot 2 \cdot t(12, \tau) \cdot \frac 12 + \frac{9}{9^3} \frac{1}{3!}.
\]
Using this formula, we can now calculate the density of 123 in a 3-inflatable permutation. We have already showed that the density of inversions should be 1/2. Therefore, \[
\frac{1}{6} = \frac{n(n-1)(n-2)}{n^3} t(123, \tau) + \binom n2 \cdot \frac{3}{n^3} \cdot 2 \cdot
\frac 12 \cdot \frac 12 + \frac{n}{n^3} \frac{1}{3!}.
\]
Multiplying by $12n^2$ and rearranging yields \[
2n^2 - 9n + 7 = 12(n-1)(n-2) t(123, \tau).
\]
Then, we observe that the density is \[
t(123, \tau) = \frac{2n^2 - 9n + 7}{12(n-1)(n-2)} = \frac{(2n-7)(n-1)}{12(n-1)(n-2)} = \frac{2n-7}{12(n-2)}.
\]
Analogous applications of this formula for densities of the five other length-3 permutations yields the desired result.
\end{proof}
For a given integer $n$, these densities might not be possible for any permutation $\tau$ due to divisibility. The conditions are described in the following corollary.
\begin{corollary}
If a permutation of length $n$ is $3$-inflatable, then both:
\begin{enumerate}
\item $\binom n2$ is even.
\item $\binom n3$ is divisible by the reduced denominators of $\frac{2n-7}{12(n-2)}$ and $\frac{4 n - 5}{24 (n - 2)}$.
\end{enumerate}
\end{corollary}
The above divisibility criteria can be more explicitly described by determining the complete set of admissible residue classes modulo 144.
\begin{lemma}
For any 3-inflatable permutation $\tau$ with length $n$,
$$n \equiv 0, 1, 17, 64, 80, 81 \pmod{144}.$$
\end{lemma}
\begin{proof}
Let $k$ be the number of occurrences of $132$ in the permutation $\tau$. By the above criteria, we have that \[
\frac{k}{\binom{n}{3}} = \frac{4 n - 5}{24 (n - 2)},
\] \[
144 k = n(n-1)(4 n - 5),
\] \[
144 \mid n(n-1)(4n-5).
\]
Also, \[
\frac{k}{\binom{n}{3}} = \frac{2 n - 7}{12 (n - 2)},
\] \[
72k = n(n - 1)(2 n - 7),
\] \[
72 \mid n(n - 1)(2 n - 7).
\]
Together, these two divisibility requirements are equivalent to \[
n \equiv 0, 1, 17, 64, 80, 81 \pmod{144}.
\]
\end{proof}
The smallest possible nontrivial length for a 3-inflatable permutation is therefore 17. In this case, the binomials $\binom{17}{2} = 136$ and $\binom{17}{3} = 680$, while $(2n-7)/(12(n-2)) = 3/20$ and $(4 n - 5)/(24 (n - 2)) = 7/40$. Therefore, in a 3-inflatable permutation of length 17, the numbers of occurrences of the subpermutations 123 and 321 should both be 102, while the four other pattern permutations of length 3 should each occur 119 times.
\begin{figure}
\centering
\begin{asy}
unitsize(5cm);
drawperm("E534BGA9HC2D1687F");
label("E534BGA9HC2D1687F", (0.5, -0.03), S);
drawperm("G54ABC319HF678ED2", 1.2);
label("G54ABC319HF678ED2", (0.5 + 1.2, -0.03), S);
\end{asy}
\caption{Plots of selected minimal-length 3-inflatable permutations.}
\label{fig:plot-length17}
\end{figure}
Examples of these minimum-length 3-inflatable permutations were found through an optimized computer search. For example, the permutations \[\text{E534BGA9HC2D1687F and G54ABC319HF678ED2}\] are both 3-inflatable, where capital letters denote numbers greater than nine (A=10, B=11, C=12, etc.). Plots of these permutations are shown in Figure~\ref{fig:plot-length17}.
Although it was intractable to check all $17!$ of such permutations for the inflation criterion through our available computational resources, we could reduce the search space by checking only those permutations which were \emph{centrally symmetric} (further discussed in the next section). Of the $8 \cdot 2^8$ centrally symmetric permutations of length $17$, a computer search found that $750$ of them are $3$-inflatable.
Furthermore, additional computer searches found many more examples of centrally symmetric, 3-inflatable permutations of lengths $64$, $80$, $81$, $144$, and $145$. We have found 3-inflatable permutations of lengths belonging to each admissible residue class modulo 144. There also seems to be a large number of such permutations of each length. These empirical results suggest the following conjecture.
\begin{conjecture}
For any positive integer $x$ of length belonging to an admissible residue class modulo $144$, there exists a $3$-inflatable permutation of length $x$.
\end{conjecture}
\section{Structure of 3-inflatable permutations}
\label{sec:structure3}
We start by proving that the inflation of two $k$-inflatable permutations is itself $k$-inflatable. This gives us an explicit construction that shows there exists an infinite number of $3$-inflatable permutations, and also allows us to construct examples of arbitrarily large length.
\begin{theorem}
If $\tau_1$ and $\tau_2$ are two $k$-inflatable permutations, then $\inflate(\tau_1, \tau_2)$ is also $k$-inflatable.
\end{theorem}
\begin{proof}
Note that inflation is associative, so for a random permutation $\gamma$, \[
\inflate(\inflate(\tau_1, \tau_2), \gamma) = \inflate(\tau_1, \inflate(\tau_2, \gamma)).
\]
Then, by the definition of a $k$-inflatable permutation, for any length-$k$ permutation $\pi$, \[
\lim_{|\gamma| \rightarrow \infty} t(\pi, \inflate(\tau_2, \gamma)) = \frac{1}{k!}.
\]
Since our current discussion only considers densities of pattern permutations with length at most $k$, we can substitute $\inflate(\tau_2, \gamma)$ for a random permutation $\gamma_2$ of the same length without changing the expected value of the expression. Thus, \[
\lim_{|\gamma| \rightarrow \infty} t(\pi, \inflate(\tau_1, \inflate(\tau_2, \gamma))) = \lim_{|\gamma_2| \rightarrow \infty} t(\pi, \inflate(\tau_1, \gamma_2)) = \frac{1}{k!},
\]
So $\inflate(\tau_1, \tau_2)$ is $k$-inflatable.
\end{proof}
Given this result, it is natural to ask if the set of admissible lengths of 3-inflatable permutations modulo 144 are closed under multiplication --- in other words, if the set of numbers with remainders 0, 1, 17, 64, 80, and 81 modulo 144 is closed under multiplication. This is true, as shown in Figure~\ref{fig:mult_table}.
\begin{figure}
\centering
\begin{tabular}{|c|cccccc|}
\hline
$\times$ & 0 & 1 & 17 & 64 & 80 & 81 \\ \hline
0 & 0 & 0 & 0 & 0 & 0 & 0 \\
1 & 0 & 1 & 17 & 64 & 80 & 81 \\
17 & 0 & 17 & 1 & 80 & 64 & 81 \\
64 & 0 & 64 & 80 & 64 & 80 & 0 \\
80 & 0 & 80 & 64 & 80 & 64 & 0 \\
81 & 0 & 81 & 81 & 0 & 0 & 81 \\
\hline
\end{tabular}
\caption{Multiplication table of admissible lengths modulo 144.}
\label{fig:mult_table}
\end{figure}
For example, we can use the 3-inflatable permutations of size 17 that we found to build 3-inflatable permutations of sizes $17^n$.
Now we want to define a symmetry of permutations that is useful for increasing the likelihood of being symmetric. Let us define an operation $R$ on permutations which we call a rotation: $R(\pi)_i = n + 1 - \pi_{n + 1 - i}$. This is equivalent to drawing the permutation as a graph (on a square grid) then rotating it by 180 degrees about its center.
We call a permutation $\pi$ of length $n$ \emph{centrally symmetric} if it is equal to its rotation (for example, see the right plot in Figure~\ref{fig:plot-length17}). In other words, $\pi_i + \pi_j = n+1$ whenever $i+j = n + 1$. The importance of centrally symmetric permutations is explained by the following lemma.
\begin{lemma}
For permutations $\pi$ and $\gamma$,
\[t(\pi,\gamma) = t(R(\pi),R(\gamma)).\]
\end{lemma}
\begin{proof}
Let $k = |\pi|$, and let $S_1$ be the set of $k$-point subsets of $\gamma$ that are order-isomorphic to $\pi$. Similarly, let $S_2$ be the set of $k$-point subsets of $R(\gamma)$ that are order-isomorphic to $R(\pi)$. There is a one-to-one correspondence between $S_1$ and $S_2$ given by the rotation operation.
More formally, for any set of indices $\{ i_1, i_2, \dots, i_k \}$ in $\gamma$, there is a corresponding set of indices, $\{ n + 1 - i_1, n + 1 - i_2, \dots, n + 1 - i_k \}$ in $R(\gamma)$. If the former indices induce some permutation $\pi$ in $\gamma$, then the latter set induces $R(\pi)$ in $R(\gamma)$ from the definition of rotation. Thus, we have $|S_1| = |S_2|$, so the result follows.
\end{proof}
\begin{corollary}
If $\pi$ is centrally symmetric, then
\[t(\pi,\gamma) = t(\pi,R(\gamma)).\]
\end{corollary}
That means centrally symmetric permutations automatically give us some equalities among densities. For example, if $\pi$ is centrally symmetric, then
\[t(132,\pi) = t(231,\pi) \quad \text{and} \quad t(312,\pi) = t(213,\pi).\]
\section{Acknowledgements}
We are grateful to Prof.~Yufei Zhao for suggesting this project to us and for helpful discussions on the topic, as well as to the anonymous reviewer for their thorough review and suggestions. We would also like to thank the PRIMES program for giving us the opportunity to do this research.
\printbibliography
\end{document}
|
\section{Introduction
In this paper, we continue to develop the link between the theory of $Q$-polynomial distance-regular graphs and the theory of double affine Hecke algebras (DAHAs); cf. \cite{1984BanIto, 1973Delsarte, 1992Cherednik, 2013LeeLAA, 2017LeeJCTA, 2018LeeTanakaSIGMA}.
We briefly summarize our results concerning the link.
In \cite{2013LeeLAA}, we considered a $Q$-polynomial distance-regular graph that corresponds to $q$-Racah polynomials, at the top level (i.e. $_{4}\phi_{3}$) in the terminating branch of the $q$-Askey scheme \cite{2010KLS}.
Assuming that the graph contains a clique with maximal possible size (i.e. Delsarte clique), we introduced the generalized Terwilliger algebra $\mathbf{T}(x,C)$, which is a non-commutative semisimple matrix $\mathbb{C}$-algebra attached to every pair of a Delsarte clique $C$ and a vertex $x \in C$ of the graph.
We showed that each such pair $x,C$ gives rise to a vector space that has an irreducible module structure for both $\mathbf{T}(x,C)$ and a DAHA of type $(C^\vee_1, C_1)$, the most general DAHA of rank one \cite{2004OblomkovIMRN}.
In the following paper \cite{2017LeeJCTA}, we captured the non-symmetric $q$-Racah polynomials from that vector space, a discrete version of non-symmetric Askey-Wilson polynomials introduced by Sahi \cite{1999SahiAnnMath}, and gave a combinatorial interpretation for their orthogonality relations.
We note that, however, the results obtained in \cite{2013LeeLAA, 2017LeeJCTA} may remain at the purely algebraic level; because there is no known example of a (non-trivial) $Q$-polynomial distance-regular graph with large diameter (at least ten)\footnote{For small diameter, there are infinitely many examples
of bipartite $Q$-polynomial distance-regular graphs of $q$-Racah type, for which every edge is a Delsarte clique.} that corresponds to $q$-Racah polynomials and contains a Delsarte clique.
To complement this shortcoming, in the subsequent paper \cite{2018LeeTanakaSIGMA} we dealt with the dual polar graphs as a concrete combinatorial example in the context of the theory developed in \cite{2013LeeLAA, 2017LeeJCTA}.
The dual polar graphs are a classical family of $Q$-polynomial distance-regular graphs and correspond to the dual $q$-Krawtchouk polynomials.
Applying techniques of \cite{2013LeeLAA, 2017LeeJCTA} to a dual polar graph, we obtained an irreducible module for a nil-DAHA\footnote{This nil-DAHA is isomorphic to the confluent Cherednik algebra $\mathcal{H}_{\mathrm{III}}$ that corresponds to Al-Salam-Chihara polynomials; cf. \cite{2018LeeTanakaSIGMA, 2014Mazzocco}.} of type $(C^\vee_1, C_1)$ \cite{2015CheOrrMathZ}, which is a specialization of the DAHA of type $(C^\vee_1, C_1)$.
We then captured the non-symmetric dual $q$-Krawtchouk polynomials, a discrete version of non-symmetric Al-Salam-Chihara polynomials \cite{2014Mazzocco}, from a nil-DAHA module. We also described their recurrence and orthogonality relations from a combinatorial point of view.
In the present paper, as another specific combinatorial object with strong regularity, we discuss the Grassmann graphs in the context of our study to develop the theory of \cite{2013LeeLAA, 2017LeeJCTA} further.
The Grassmann graphs are a classical family of $Q$-polynomial distance-regular graphs and correspond to the dual $q$-Hahn polynomials which lie in between $q$-Racah and dual $q$-Krawtchouk polynomials in the $q$-Askey scheme; see Figure \ref{q-Askey scheme}.
\begin{figure}
\centering
\scalemath{0.9}{
\begin{tikzpicture}[>=stealth,thick, every node/.style={shape=rectangle,rounded corners}, align=center, text width=16em, minimum height=1em]
\node[draw, minimum height=2.7em] (c1) at (0,0) {\small $q$-Racah polynomials \\ ($Q$-polynomial distance-regular graphs)};
\node[draw, minimum height=2.7em] (c3) at (0, -2) {\small Dual $q$-Hahn polynomials \\ (Grassmann graphs)};
\node[draw, minimum height=2.7em] (c6) at (0, -4) {\small Dual $q$-Krawtchouk polynomials\\ (Dual polar graphs)};
\node[draw, minimum height=2.7em] (d1) at (8.5,0) {\small The DAHA $\mathcal{H}$ of type $(C^\vee_1, C_1)$};
\node[draw, minimum height=2.7em] (d2) at (8.5, -2) {\small The confluent Cherednik algebra $\mathcal{H}_\mathrm{V}$};
\node[draw, minimum height=2.7em] (d3) at (8.5, -4) {\small The confluent Cherednik algebra $\mathcal{H}_\mathrm{III}$};
\node (r1) [text width=1.5em] at (-4,0) {(${}_4\phi_3$)};
\node (r2) [text width=1.5em] at (-4,-2) {(${}_3\phi_2$)};
\node (r3) [text width=1.5em] at (-4, -4) {(${}_3\phi_2$)};
\draw[->] (c1) to (c3)
\draw[->] (c3) to (c6)
\draw[->] (d1) to (d2)
\draw[->] (d2) to (d3)
\draw[<->, dashed, thick] (c1) to (d1);
\draw[<->, dashed, thick] (c3) to (d2);
\draw[<->, dashed, thick] (c6) to (d3);
\end{tikzpicture}}
\caption[Caption Footnoot]{Part of the $q$-Askey scheme and the corresponding (degenerate) DAHAs\footnotemark}\label{q-Askey scheme}
\end{figure}
\footnotetext{Recently, the duality and its limit behavior of three families (Askey-Wilson, continuous dual $q$-Hahn, Al-Salam-Chihara) of the $q$-Askey scheme and the corresponding degenerate DAHAs were dealt with by Koornwinder and Mazzocco; cf. \cite{2018KoornMazzo}.}
The main results of this paper are as follows.
Let $J_q(N,D)$ denote a Grassmann graph, where $q$ is a prime power and $N$, $D$ are positive integers with $N \geq 2D$; see the definition in Section \ref{Section:Grassmann graphs}.
Fix a pair of a Delsarte clique $C$ in $J_q(N,D)$ and a vertex $x$ in $C$.
Applying the methods used in \cite{2013LeeLAA, 2017LeeJCTA, 2018LeeTanakaSIGMA} to $J_q(N,D)$, we construct a $2D$-dimensional irreducible $\mathbf{T}(x,C)$-module $\mathbf{W}$ and show that $\mathbf{W}$ has a module structure for the \emph{confluent Cherednik algebra} $\mathcal{H}_\mathrm{V}$ \cite{2016Mazzocco}; cf. Definition \ref{Def:CheAlgHv}.
We then demonstrate how the $\mathbf{T}(x,C)$-action on $\mathbf{W}$ is related to the $\mathcal{H}_\mathrm{V}$-action on $\mathbf{W}$; cf. Theorems \ref{thm:1stResult} and \ref{thm:2ndResult}.
From the $\mathcal{H}_\mathrm{V}$-module $\mathbf{W}$, we obtain \emph{non-symmetric dual $q$-Hahn polynomials}, a discrete version of non-symmetric continuous dual $q$-Hahn polynomials \cite[Section 2]{2014Mazzocco}, and describe their recurrence and orthogonality relations; cf. Theorems \ref{thm:3rdResult} and \ref{thm:4thResult}.
We should mention that all the formulas in the present paper are described in terms of the scalars $q$, $N$, and $D$, not depending on our fixed pair $x, C$.
This paper is organized as follows.
In Section \ref{Section:DRGs} we recall some preliminaries concerning $Q$-polynomial distance-regular graphs and the Terwilliger algebra.
In Section \ref{Section:Grassmann graphs} we discuss a Grassmann graph $\Gamma=J_q(N,D)$ with diameter $D\geq 3$ and its properties.
We also discuss the Terwilliger algebra of $\Gamma$ associated with a Delsarte clique.
In Section \ref{Section:algebraT} we fix a Delsarte clique $C$ and a vertex $x$ in $C$.
We discuss the generalized Terwilliger algebra $\mathbf{T}=\mathbf{T}(x,C)$ of $\Gamma$ and construct the so-called primary $\mathbf{T}$-module $\mathbf{W}$.
In Section \ref{Section:LS of dual q-Hahn} we discuss the theory of Leonard systems. In particular, we treat a family of Leonard systems that corresponds to dual $q$-Hahn polynomials.
In Section \ref{Section:primary T-module} we deal with four dual $q$-Hahn Leonard systems obtained from $\mathbf{W}$.
In Section \ref{Section:Hv} we discuss the confluent Cherednik algebra $\mathcal{H}_\mathrm{V}$ and construct a $\mathbb{C}$-algebra homomorphism from $\mathcal{H}_\mathrm{V}$ to $\mathrm{End}(\mathbf{W})$, which gives an $\mathcal{H}_\mathrm{V}$-module structure on $\mathbf{W}$.
We discuss a relationship between the action of $\mathcal{H}_\mathrm{V}$ and the action of $\mathbf{T}$ on $\mathbf{W}$.
We specialize the DAHA of type $(C^\vee_1, C_1)$ to get a nil-DAHA $\overline{\mathcal{H}}$ and discuss how $\overline{\mathcal{H}}$ is related to $\mathcal{H}_\mathrm{V}$.
In Section \ref{Section:nonsym dual q-Hahn poly} we introduce non-symmetric dual $q$-Hahn polynomials $\ell^\pm_i$ and give a combinational interpretation for $\ell^\pm_i$.
In Section \ref{Section:rec orth relations} we deal with recurrence relations and orthogonality relations for $\ell^\pm_i$.
Throughout this paper, we use the following notation.
For a non-empty finite set $X$, let $\mathrm{Mat}_X(\mathbb{C})$ denote the $\mathbb{C}$-algebra consisting of the complex square matrices indexed by $X$.
Let $\mathbb{C}^X$ denote the $\mathbb{C}$-vector space consisting of the complex column vectors indexed by $X$.
We endow $\mathbb{C}^X$ with the Hermitian inner product $\langle \cdot , \cdot \rangle = \langle \cdot , \cdot \rangle_{\mathbb{C}^X}$ which satisfies $\langle u, v \rangle = u^t\bar{v}$ for $u,v \in \mathbb{C}^X$, where $t$ denotes transpose and $\bar{ \ }$ denotes complex conjugate.
Abbreviate $\lVert u \rVert^2=\langle u, u \rangle$ for all $u \in \mathbb{C}^X$.
For $y \in X$, let $\hat{y}$ denote the vector in $\mathbb{C}^X$ with a $1$ in the $y$-coordinate and $0$ in all other coordinates.
For a subset $Y \subseteq X$, define $\hat{Y} = \sum_{y \in Y}\hat{y}$, called the \textit{characteristic vector} of $Y$.
Let $\mathbb{C}[\zeta, \zeta^{-1}]$ denote the space of Laurent polynomials in one variable $\zeta$.
A Laurent polynomial $f(\zeta)$ is said to be \emph{symmetric} if $f(\zeta)=f(\zeta^{-1})$, and \emph{non-symmetric} otherwise.
We view symmetric Laurent polynomials as ordinary polynomials in the variable $\lambda :=\zeta+\zeta^{-1}$.
Assume that $q \in \mathbb{C}^*$ is not a root of unity.
For $\alpha \in \mathbb{C}$,
\begin{equation}
(\alpha;q)_0 :=1 \quad \text{and} \quad (\alpha;q)_n := (1-\alpha)(1-\alpha q) \cdots (1-\alpha q^{n-1}), \quad n=1,2,3,\ldots.
\end{equation}
For $\alpha_1, \alpha_2, \ldots, \alpha_{s+1}, \beta_1, \beta_2, \ldots, \beta_{s} \in \mathbb{C}$,
\begin{equation}
{}_{s+1}\phi_s \left( \left. \begin{array}{c} \alpha_1, \alpha_2, \ldots, \alpha_{s+1} \\ \beta_1, \beta_2, \ldots, \beta_s \end{array}\right| q; \zeta \right) = \sum^\infty_{n=0}\frac{(\alpha_1;q)_n(\alpha_2;q)_n \cdots (\alpha_{s+1};q)_n}{(\beta_1;q)_n(\beta_2;q)_n\cdots (\beta_s;q)_n}\frac{\zeta^n}{(q;q)_n}.
\end{equation}
For integers $0 \leq m \leq n$, we denote the Gaussian binomial coefficient by
\begin{equation}
\gauss{n}{m} = \gauss{n}{m}_q = \frac{(q^n-1)(q^{n-1}-1)\cdots(q^{n-m+1}-1)}{(q^m-1)(q^{m-1}-1)\cdots(q-1)}.
\end{equation}
We remark that if $q$ is set to a prime power then $\gauss{n}{m}$ is equal to the number of $m$-dimensional subspaces of an $n$-dimensional vector space over a finite field $\mathbb{F}_q$.
In what follows, we assume that $q$ is a prime power unless otherwise stated.
\section{Preliminaries: Distance-regular graphs}\label{Section:DRGs
In this preliminary section, we recall some basic aspects of distance-regular graphs that we need later in the paper.
Let $\Gamma$ be a connected simple graph with finite vertex set $X$ and diameter $D\geq 3$.
For a vertex $x$ in $X$, define
\begin{equation}\label{dist.part(x)}
\Gamma_i(x) := \{ y \in X : \partial(x,y) = i \}, \quad 0 \leq i \leq D,
\end{equation}
where $\partial(x,y)$ is the shortest path-length distance function between $x$ and $y$.
We abbreviate $\Gamma(x)=\Gamma_1(x)$.
For an integer $k$, $\Gamma$ is said to be $k$-\emph{regular} (or \emph{regular with valency $k$}) whenever $|\Gamma(x)|=k$ for all $x$ in $X$.
We say that $\Gamma$ is \emph{distance-regular} whenever for every $i$, $0 \leq i \leq D$, and for every pair of vertices $x$, $y$ in $X$ with $\partial(x,y)=i$, there are constant numbers $a_i$, $b_i$, $c_i$ such that
\begin{equation}\label{IntNum.Gamma}
c_i = | \Gamma_{i-1}(x) \cap \Gamma(y) |, \qquad
a_i = | \Gamma_{i}(x) \cap \Gamma(y) |, \qquad
b_i = | \Gamma_{i+1}(x) \cap \Gamma(y) |,
\end{equation}
where $\Gamma_{-1}(x)$ and $\Gamma_{D+1}(x)$ are empty sets.
Observe that $c_0=b_D=0$, $b_{i-1}c_i \ne 0$, $1\leq i \leq D$, and $c_1=1$.
Observe also that $\Gamma$ is $b_0$-regular and $a_i+b_i+c_i=b_0$ for $0 \leq i \leq D$.
The constants $a_i$, $b_i$, $c_i$ are called the \emph{intersection numbers} of $\Gamma$.
Assume that $\Gamma$ is distance-regular.
For $0 \leq i \leq D$, define the matrix $A_i$ in $\mathrm{Mat}_X(\mathbb{C})$ such that $(x,y)$-entry of $A_i$ is $1$ if $\partial(x,y)=i$ and $0$ otherwise.
We call $A_i$ the \emph{$i$-th distance matrix} of $\Gamma$.
Observe that $A_0=I$, the identity matrix in $\mathrm{Mat}_X(\mathbb{C})$.
We abbreviate $A=A_1$ and call this the \emph{adjacency matrix} of $\Gamma$.
The \emph{Bose-Mesner algebra} of $\Gamma$ is the (commutative) semisimple subalgebra $M$ of $\mathrm{Mat}_X(\mathbb{C})$ generated by $I, A, A_2,\ldots,A_D$.
Observe that
\begin{equation*}
AA_i = b_{i-1}A_{i-1}+a_iA_i+c_{i+1}A_{i+1}, \qquad 0 \leq i \leq D,
\end{equation*}
where we set $b_{-1}A_{-1}=0$ and $c_{D+1}A_{D+1}=0$. From this recurrence, it follows that there is a polynomial $v_i \in \mathbb{C}[\lambda]$ of degree $i$ such that $v_i(A)=A_i$ for $0 \leq i \leq D$.
It follows that $A$ generates $M$, and that the matrices $A_i$, $0 \leq i \leq D$, form a basis for $M$.
Since $A$ is real symmetric and generates $M$, it has $D+1$ mutually distinct real eigenvalues $\theta_0, \theta_1, \ldots, \theta_D$.
We always set $\theta_0:=b_0$.
For $0 \leq i \leq D$, let $E_i \in \mathrm{Mat}_X(\mathbb{C})$ be the orthogonal projection onto the eigenspace of $\theta_i$.
Observe that $E_iE_j = \delta_{ij}E_i$, $0 \leq i, j \leq D$, and $\sum^D_{i=0}E_i=I$.
We have
\begin{equation*}
A = \sum^D_{i=0}\theta_i E_i,
\end{equation*}
so that the matrices $E_i$, $0 \leq i \leq D$, form another basis for $M$.
We recall the $Q$-polynomial property of $\Gamma$.
The Bose-Mesner algebra $M$ of $\Gamma$ is closed under entrywise multiplication, denoted by $\circ$, since $A_i\circ A_j = \delta_{ij}A_i$, $0 \leq i,j \leq D$.
We say that $\Gamma$ is \emph{$Q$-polynomial} with respect to the ordering $E_0, E_1, \ldots, E_D$ (or $\theta_0, \theta_1, \ldots, \theta_D$) if there are scalars $a^*_i$, $b^*_i$, $c^*_i$, $0 \leq i \leq D$, such that $b^*_D=c^*_0=0$, and $b^*_{i-1}c^*_i \ne 0$ for $1\leq i \leq D$, and
\begin{equation*}
|X| (E_1\circ E_i) = b^*_{i-1}E_{i-1} + a^*_iE_i+c^*_{i+1}E_{i+1}, \qquad 0 \leq i \leq D,
\end{equation*}
where we set $b^*_{-1}E_{-1}=0$ and $c^*_{D+1}E_{D+1}=0$.
From this recurrence, it follows that there is a polynomial $v^*_i \in \mathbb{C}[\lambda]$ of degree $i$ such that $v^*_i(E_1) = E_i$ for $0 \leq i \leq D$, where the multiplication is under $\circ$.
Write $E_1=|X|^{-1}\sum^D_{i=0}\theta^*_iA_i$. Then the scalars $\theta^*_i$, $0\leq i \leq D$, are real and mutually distinct.
We note that $\theta^*_0 = \mathrm{trace}(E_1) = \mathrm{rank}(E_1)$.
Assume that $\Gamma$ is $Q$-polynomial with respect to the ordering $E_0, E_1, \ldots, E_D$.
Fix a vertex $x$ in $X$.
For $0 \leq i \leq D$, let $E^*_i=E^*_i(x)$ denote the diagonal matrix in $\mathrm{Mat}_X(\mathbb{C})$ with $(y,y)$-entry $1$ if $\partial(x,y)=i$ and $0$ otherwise, i.e., $E^*_i = \mathrm{diag}(A_i\hat{x})$.
Observe that $E^*_iE^*_j=\delta_{ij}E^*_i$, $0\leq i,j \leq D$, and $\sum^D_{i=0}E^*_i=I$.
The \emph{dual Bose-Mesner algebra} of $\Gamma$ with respect to $x$ is the (commutative) semisimple subalgebra $M^*=M^*(x)$ of $\mathrm{Mat}_X(\mathbb{C})$ generated by $E^*_0, E^*_1, \ldots, E^*_D$.
Note that the matrices $E^*_i$, $0 \leq i \leq D$, form a basis for $M^*$.
Let $A^*=A^*(x)$ denote the diagonal matrix in $\mathrm{Mat}_X(\mathbb{C})$ with $(y,y)$-entry $(|X|E_1)_{xy}$ for $y\in X$, i.e., $A^* = \mathrm{diag}(|X|E_1\hat{x})$.
Then
\begin{equation}\label{A*;linComb(E*i)}
A^* = \sum^D_{i=0}\theta^*_iE^*_i,
\end{equation}
from which it follows that $A^*$ generates $M^*$. We call $A^*$ the \emph{dual adjacency matrix} of $\Gamma$ with respect to $x$.
Observe that the scalars $\theta^*_i$ are the eigenvalues of $A^*$, called the \emph{dual eigenvalues} of $\Gamma$.
The \emph{Terwilliger algebra} (or \emph{subconstituent algebra}) of $\Gamma$ with respect to $x$ is the subalgebra $T=T(x)$ of $\mathrm{Mat}_X(\mathbb{C})$ generated by $M$, $M^*$ \cite{1992TerJACO(1), 1993TerJACO(2), 1993TerJACO(3)}.
Note that the matrices $A$, $A^*$ generate $T$.
Note also that $T$ is (non-commutative) semisimple and any two non-isomorphic irreducible $T$-modules in $\mathbb{C}^X$ are orthogonal.
The following are relations in $T$:
\begin{equation*}
E^*_iAE^*_j = 0, \qquad E_iA^*E_j=0 \qquad \text{if} \quad |i-j|>1,
\end{equation*}
for $0 \leq i, j \leq D$; cf. \cite[Lemma 3.2]{1992TerJACO(1)}.
We observe that the subspace $M\hat{x}$ of $\mathbb{C}^X$ has bases $\{A_i\hat{x}\}^D_{i=0}$ and $\{E_i\hat{x}\}^D_{i=0}$, and that $A_i\hat{x}=E^*_i\hat{X}$, $0 \leq i \leq D$.
It follows that $M\hat{x}$ is same as the subspace $M^*\hat{X}$ of $\mathbb{C}^X$, and therefore $M\hat{x}$ is an irreducible $T$-module, called the \emph{primary} $T$-module.
The actions of $A$, $A^*$ on $M\hat{x}$ are given as follows: for $0 \leq i \leq D$,
\begin{equation*}
A.A_i\hat{x} = b_{i-1}A_{i-1}\hat{x}+a_iA_i\hat{x}+c_{i+1}A_{i+1}\hat{x}, \qquad A^*.A_i\hat{x} = \theta^*_iA_i\hat{x}.
\end{equation*}
For more information regarding distance-regular graphs, we refer to \cite{1984BanIto, 1989BCN, 2016DKT}.
\section{Grassmann graphs}\label{Section:Grassmann graphs
Recall $q$ a prime power. Let $N, D$ be positive integers with the restriction $N \geq 2D$.
Let $V$ be an $N$-dimensional vector space over a finite field $\mathbb{F}_q$, and let $X$ be the collection of $D$-dimensional subspaces of $V$.
The \emph{Grassmann graph} $J_q(N,D)$ has vertex set $X$, where two vertices are adjacent whenever their intersection has dimension $D-1$; cf. \cite[p. 268]{1989BCN}.
We readily see that the cardinality of $X$ is $\gauss{N}{D}$.
Observe that two vertices $x,y$ have distance $i$ if and only if $\dim(x\cap y)=D-i$.
Note that $J_q(N,D)$ is isomorphic to $J_q(N,N-D)$. By our restriction on $N$ and $D$, $J_q(N,D)$ has diameter $D$.
Throughout the rest of this paper, let $\Gamma$ denote the Grassmann graph $J_q(N,D)$ with diameter $D$; to avoid trivialities we assume $D\geq 3$.
We recall some basic results that we need; cf. \cite[Section 9.3]{1989BCN}.
The graph $\Gamma$ \emph{is} distance-regular with intersection numbers given by
\begin{equation}\label{Gr.IntNum;a,b,c}
a_i = \gauss{i}{1}\left(\gauss{i}{1}-q^{i+1}\gauss{D-i}{1}-q\gauss{N-D}{1}\right), \quad
b_i = q^{2i+1}\gauss{D-i}{1}\gauss{N-D-i}{1}, \quad
c_i = \gauss{i}{1}^2,
\end{equation}
for $0 \leq i \leq D$.
The eigenvalues of $\Gamma$ are given by
\begin{equation*}\label{Gr.Eigval}
\theta_i = q^{i+1}\gauss{D-i}{1}\gauss{N-D-i}{1}-\gauss{i}{1},
\end{equation*}
for $0 \leq i \leq D$.
The graph $\Gamma$ \emph{is} $Q$-polynomial with respect to the ordering $\{\theta_i\}^D_{i=0}$ with $\theta_0 > \theta_1> \cdots> \theta_D$.
The dual eigenvalues of $\Gamma$ are given by
\begin{equation}\label{Gr.dualEigval}
\theta^*_i = \frac{(q^N-q)(2-q^D-q^{N-D})}{(q-1)(q^D-1)(q^{N-D}-1)} + \frac{(q^N-q)(q^N-1)}{(q-1)(q^D-1)(q^{N-D}-1)}q^{-i},
\end{equation}
for $0 \leq i \leq D$; cf. \cite[Table 6.1, Theorem 8.4.1]{1989BCN}, \cite[Lemma 4.3]{2015GaGaHoLAA}.
Let $C$ be a collection of all $D$-dimensional subspaces of $V$ containing a fixed $(D-1)$-dimensional subspace.
Then $C$ is a maximal clique\footnote{There is the other type of maximal cliques in $\Gamma$, namely, the collection of all $D$-dimensional subspaces of $V$ contained in a fixed $(D+1)$-dimensional subspace. Note that these maximal cliques are not Delsarte unless $N=2D$.}
of $\Gamma$ and we have
\begin{equation}\label{size.C}
|C| = \gauss{N-D+1}{1}.
\end{equation}
From this, it follows that $C$ is a \emph{Delsarte} clique, i.e., $C$ attains the Hoffman bound $1-\theta_0/\theta_D$; cf. \cite[Proposition 4.4.6]{1989BCN}.
Take a Delsarte clique $C$ of $\Gamma$.
The \emph{covering radius} of $C$ is defined by $\max\{ \partial(y,C) : y \in X \}$, where $\partial(y,C) = \min\{\partial(y,z) : z \in C\}$.
Note that the covering radius of $C$ is given by $D-1$; cf \cite[Lemma 7.4]{1993GodsilAC}.
Define
\begin{equation}\label{dist.part(C)}
C_i := \{ y \in X :\partial(y,C) =i \}, \qquad 0 \leq i \leq D-1.
\end{equation}
For notational convenience, we set $C_{-1}:=\varnothing$ and $C_D:=\varnothing$.
We remark that $\{C_i\}^{D-1}_{i=0}$ is an \emph{equitable} partition, i.e., for all integers $i$ and $j$, $0 \leq i, j \leq D-1$, each vertex in $C_i$ has constant neighbors in $C_j$.
In particular, for each $z\in C_i$, $0 \leq i \leq D-1$, there exist constant numbers $\widetilde a_i$, $\widetilde b_i$, $\widetilde c_i$ such that
\begin{equation}\label{IntNum.C}
\widetilde c_i = |\Gamma(z) \cap C_{i-1}|, \qquad
\widetilde a_i = |\Gamma(z) \cap C_{i}|, \qquad
\widetilde b_i = |\Gamma(z) \cap C_{i+1}|,
\end{equation}
where $\widetilde c_0 = \widetilde b_{D-1}=0$ and $\widetilde b_{i-1}\widetilde c_i \ne 0$ for $1 \leq i \leq D-1$.
Observe that $\widetilde a_0 = |C|-1 = q\gauss{N-D}{1}$ and $\widetilde a_i + \widetilde b_i + \widetilde c_i = b_0$ for $0 \leq i \leq D-1$.
The constants $\widetilde a_i$, $\widetilde b_i$, $\widetilde c_i$ are called the \emph{intersection numbers} of $C$.
For $0 \leq i \leq D-1$ and $z \in C_i$, consider the subset $\{ y \in C \mid \partial(y,z)=i\}$ of $C$.
Then by the construction the cardinality of this set is given by
\begin{equation}\label{N(i)}
n_i(z):=|\{ y \in C \mid \partial(y,z)=i\}| = \gauss{i+1}{1},
\end{equation}
from which it follows that the cardinality $n_i(z)$ is independent of the choice of $z$ in $C_i$, and thus we write $n_i=n_i(z)$ for $0 \leq i \leq D-1$.
By definition, we have
\begin{equation}\label{AiC=lin.combo}
A_i\hat{C} = (|C|-n_{i-1})\hat{C}_{i-1} + n_i\hat{C}_i, \qquad 0 \leq i \leq D,
\end{equation}
where $(|C|-n_{-1})\hat{C}_{-1}=0$ and $n_D\hat{C}_D=0$.
\begin{lemma}\label{tilde;a,b,c(i)}
The intersection numbers of $C$ are given by
\begin{align}
& \widetilde{a}_i = \frac{1}{q-1}\left( (q^{D+1}-q^{i+1}+1)\gauss{i}{1} + q^{N-D+1}\gauss{i+1}{1} -q \gauss{2i+1}{1} \right), \label{IntNumC;a} \\
& \widetilde{b}_i = q^{2i+2}\gauss{D-i-1}{1}\gauss{N-D-i}{1}, \qquad \qquad
\widetilde{c}_i = \gauss{i+1}{1}\gauss{i}{1}, \label{IntNumC;bc}
\end{align}
for $0 \leq i \leq D-1$.
\end{lemma}
\begin{proof}
We recall the intersection numbers $b_i$, $c_i$ of $\Gamma$.
By \cite[Theorem 4.7]{2013LeeLAA} and \eqref{N(i)},
\begin{equation}\label{rels:tbc-bc}
\widetilde b_i = \frac{q^{D-N+i}-1}{q^{D-N+i+1}-1}b_{i+1}, \qquad \widetilde c_i = \frac{q^{i+1}-1}{q^i-1}c_i, \qquad 0\leq i \leq D-1.
\end{equation}
Evaluate \eqref{rels:tbc-bc} using \eqref{Gr.IntNum;a,b,c} to get \eqref{IntNumC;bc}.
To verify \eqref{IntNumC;a}, use $\widetilde a_i + \widetilde b_i + \widetilde c_i = b_0$ along with \eqref{Gr.IntNum;a,b,c}.
\end{proof}
We recall the Terwilliger algebra associated with $C$ in the sense of Suzuki \cite{2005SuzukiJACO}.
For $0\leq i \leq D-1$, let $\widetilde E^*_i=\widetilde E^*_i(C)$ denote the diagonal matrix in $\mathrm{Mat}_X(\mathbb{C})$ with $(y,y)$-entry $1$ if $y \in C_i$ and $0$ otherwise, i.e., $\widetilde E^*_i = \mathrm{diag}(\hat{C}_i)$.
Observe that $\widetilde E^*_i \widetilde E^*_j = \delta_{ij} \widetilde E^*_i$, $0 \leq i,j \leq D-1$ and $\sum^{D-1}_{i=0}\widetilde E^*_i = I$.
The \emph{dual Bose-Mesner algebra} of $\Gamma$ with respect to $C$ is the (commutative) semisimple subalgebra $\widetilde M^*=\widetilde M^*(C)$ of $\mathrm{Mat}_X(\mathbb{C})$ generated by $\widetilde E^*_0, \widetilde E^*_1, \ldots, \widetilde E^*_{D-1}$.
Note that the matrices $\widetilde E^*_i$, $0 \leq i \leq D-1$, form a basis for $\widetilde M^*$.
Define the diagonal matrix $\widetilde A^*=\widetilde A^*(C)$ in $\mathrm{Mat}_X(\mathbb{C})$ by
\begin{equation*}\label{dualadjmat(C)}
\widetilde A^* = \frac{1}{|C|}\sum_{y \in C} A^*(y) = \frac{|X|}{|C|}\mathrm{diag}(E_1\hat{C}).
\end{equation*}
Since $E_1=|X|^{-1}\sum^D_{i=0}\theta^*_iA_i$ and by \eqref{AiC=lin.combo}, we have
\begin{equation}\label{E1C=sum(tilE_i)}
\frac{|X|}{|C|}\mathrm{diag}(E_1\hat{C}) = \sum^{D-1}_{i=0} \left(\frac{n_i\theta^*_i+(|C|-n_{i})\theta^*_{i+1}}{|C|}\right)\widetilde E^*_i.
\end{equation}
For $0 \leq i \leq D-1$, let $\widetilde \theta^*_i$ denote the coefficient of each summand of $\widetilde E^*_i$ in \eqref{E1C=sum(tilE_i)}.
By \eqref{Gr.dualEigval}, \eqref{size.C} and \eqref{N(i)}, the $\widetilde\theta^*_i$ are given by
\begin{equation}\label{Gr.dualEigval.C}
\widetilde \theta^*_i = \frac{(q^{N-1}-1)(q+q^2-q^{D+1}-q^{N-D+2})}{(q-1)(q^D-1)(q^{N-D+1}-1)} + \frac{(q^N-q)(q^N-1)}{(q-1)(q^D-1)(q^{N-D+1}-1)}q^{-i},
\end{equation}
for $0 \leq i \leq D-1$.
Observe that the scalars $\widetilde\theta^*_i$ are real and mutually distinct.
We write
\begin{equation}\label{tilde(A);linComb(tildeE*i)}
\widetilde A^* = \sum^{D-1}_{i=0}\widetilde\theta^*_i \widetilde E^*_i,
\end{equation}
from which it follows that $\widetilde A^*$ generates $\widetilde M^*$.
We call $\widetilde A^*$ the \emph{dual adjacency matrix of $\Gamma$ with respect to $C$}.
Observe that the scalars $\widetilde\theta^*_i$ are the eigenvalues of $\widetilde A^*$, called the \emph{dual eigenvalues} of $\Gamma$ with respect to $C$.
The Terwilliger algebra of $\Gamma$ with respect to $C$ is the subalgebra $\widetilde T=\widetilde T(C)$ of $\mathrm{Mat}_X(\mathbb{C})$ generated by $M, \widetilde M^*$; cf. \cite{2005SuzukiJACO}.
Note that the matrices $A$, $\widetilde A^*$ generate $\widetilde T$.
Note also that $\widetilde T$ is (non-commutative) semisimple.
The following are relations in $\widetilde T$:
\begin{equation*}
\widetilde E^*_iA\widetilde E^*_j = 0, \qquad E_i\widetilde A^*E_j=0 \qquad \text{if} \quad |i-j|>1,
\end{equation*}
for $0 \leq i, j \leq D$, where we set $\widetilde E^*_D=0$; cf. \cite[Section 4]{2005SuzukiJACO}.
We note that the subspace $M\hat{C}$ of $\mathbb{C}^X$ has bases $\{A_i\hat{C}\}^{D-1}_{i=0}$, $\{\hat{C}_i\}^{D-1}_{i=0}$, and $\{E_i\hat{C}\}^{D-1}_{i=0}$.
By \eqref{AiC=lin.combo} and $\hat{C}_i = \widetilde E^*_i\hat{X}$, $0\leq i \leq D-1$,
the subspace $M\hat{C}$ is same as the subspace $\widetilde M^*\hat{X}$ of $\mathbb{C}^X$, and therefore $M\hat{C}$ is an irreducible $\widetilde T$-module, called the \emph{primary} $\widetilde T$-module.
The actions of $A, \widetilde A^*$ on $M\hat{C}$ are given as follows: for $0 \leq i \leq D-1$,
\begin{equation*}
A.\hat{C}_i = \widetilde b_{i-1}\hat{C}_{i-1} + \widetilde a_i\hat{C}_i+\widetilde c_{i+1}\hat{C}_{i+1}, \qquad \widetilde A^*.\hat{C}_i = \widetilde\theta^*_i\hat{C}_i.
\end{equation*}
\section{The generalized Terwilliger algebra of Grassmann graphs}\label{Section:algebraT
We continue to discuss the Grassmann graph $\Gamma=J_q(N,D)$.
Throughout the rest of the paper, we fix a Delsarte clique $C$ of $\Gamma$ and a vertex $x$ in $C$.
We recall the Terwilliger algebras $T=T(x)$ and $\widetilde T=\widetilde T(C)$ of $\Gamma$.
In this section, we treat the generalized Terwilliger algebra of $\Gamma$ associated with $x$ and $C$, and discuss its so-called primary module.
\begin{definition}[{\cite[Definition 5.20]{2013LeeLAA}}]\label{Algebra.T}
The \emph{generalized Terwilliger algebra} of $\Gamma$ with respect to $x$, $C$ is the subalgebra $\mathbf{T}=\mathbf{T}(x,C)$ of $\mathrm{Mat}_X(\mathbb{C})$ generated by $T$, $\widetilde T$.
Note that $A$, $A^*$, $\widetilde A^*$ generate $\mathbf{T}$, where $A^*\widetilde A^* = \widetilde A^* A^*$, and that $\mathbf{T}$ is (non-commutative) semisimple.
\end{definition}
Recall two partitions $\{\Gamma_i(x)\}^D_{i=0}$ and $\{C_i\}^{D-1}_{i=0}$ of $X$ from \eqref{dist.part(x)} and \eqref{dist.part(C)}, respectively.
Using these, we define a new partition $\{C^\pm_i\}^{D-1}_{i=0}$ of $X$ by
\begin{equation}\label{new.partition}
C^-_i := C_i \cap \Gamma_{i}(x), \qquad
C^+_i := C_i \cap \Gamma_{i+1}(x), \qquad 0 \leq i \leq D-1.
\end{equation}
See Figure \ref{2-dim'l partition}.
For notational convenience, we set $C^-_{-1}=C^+_{-1}=\varnothing$ and $C^-_{D}=C^+_{D}=\varnothing$.
Observe that $C_i = C^-_i \cup C^+_i$, $0 \leq i \leq D-1$, and $\Gamma_i(x) = C^+_{i-1} \cup C^-_i$, $0 \leq i \leq D$.
In particular, $x=C^-_0$ and $C=C^-_0 \cup C^+_0$.
From this and \eqref{size.C}, it easily follows that $|C^+_0| = \frac{q(q^{N-D}-1)}{q-1}$.
\begin{figure}
\centering
\scalemath{0.65}{
\begin{tikzpicture}
[scale=1,thick,auto=left,every node/.style={circle,draw}]
\node (n1) at (0,0) {$C^-_0$};
\node (n2) at (0,-2) {$C^+_0$};
\node (n3) at (2,-2) {$C^-_1$};
\node (n4) at (2,-4) {$C^+_1$};
\node (n5) at (4,-4) {$C^-_2$};
\node (n6) at (4,-6) {$C^+_2$};
\node (n7) at (6,-6) {$C^-_3$};
\node (n8) at (6,-8) {$C^+_3$};
\node[fill=black!10] (c0) at (0,2.5) {${~}C^{~}_0$};
\node[fill=black!10] (c1) at (2,2.5) {${~}C^{~}_1$};
\node[fill=black!10] (c2) at (4,2.5) {${~}C^{~}_2$};
\node[fill=black!10] (c3) at (6,2.5) {${~}C^{~}_3$};
\node[fill=blue!20] (r0) at (-2.5,0) {${~}\Gamma_0$};
\node[fill=blue!20] (r1) at (-2.5,-2) {${~}\Gamma_1$};
\node[fill=blue!20] (r2) at (-2.5,-4) {${~}\Gamma_2$};
\node[fill=blue!20] (r3) at (-2.5,-6) {${~}\Gamma_3$};
\node[fill=blue!20] (r4) at (-2.5,-8) {${~}\Gamma_4$};
\foreach \from/\to in {n1/n2,n2/n3,n3/n4,n4/n5,n5/n6,n6/n7,n7/n8,n1/n3,n3/n5,n5/n7,n2/n4,n4/n6,n6/n8, c0/c1,c1/c2,c2/c3, r0/r1, r1/r2, r2/r3, r3/r4}
\draw (\from) -- (\to)[line width=0.7mm] ;
\draw (c0) -- (n1) [dashed];
\draw (c1) -- (n3) [dashed];
\draw (c2) -- (n5) [dashed];
\draw (c3) -- (n7) [dashed];
\draw (r0) -- (n1) [dashed];
\draw (r1) -- (n2) [dashed];
\draw (r2) -- (n4) [dashed];
\draw (r3) -- (n6) [dashed];
\draw (r4) -- (n8) [dashed];
\end{tikzpicture}
\caption{The partition $\{C^{\pm}_i\}^{D-1}_{i=0}$ of $X$ when $D=4$}\label{2-dim'l partition}
\end{figure}
\begin{lemma}\label{cardinality:C+-}
For $0 \leq i \leq D-1$, the cardinality of each cell $C^{\pm}_i$ is given by
\begin{align*}
|C^-_i| & = q^{i(i+1)}\prod^i_{j=1} \frac{(q^{D-j}-1)(q^{N-D+1-j}-1)}{(q^j-1)^2}, \\
|C^+_i| & = \frac{q^{(i+1)^2}(q^{N-D}-1)}{q-1}\prod^i_{j=1} \frac{(q^{D-j}-1)(q^{N-D-j}-1)}{(q^j-1)(q^{j+1}-1)}.
\end{align*}
In particular, each of $C^\pm_i$ is non-empty.
\end{lemma}
\begin{proof}
Since $\Gamma$ is distance-regular and the partition $\{C_i\}^{D-1}_{i=0}$ is equitable, by \eqref{IntNum.Gamma}, \eqref{IntNum.C}, and \eqref{new.partition}, it follows
\begin{equation}\label{pf:|C-|,|C+|;eq}
\widetilde b_i |C^-_i| = c_{i+1}|C^-_{i+1}|, \qquad b_{i+1}|C^+_i| = \widetilde c_{i+1}|C^+_{i+1}|, \qquad 0 \leq i \leq D-2.
\end{equation}
Evaluate \eqref{pf:|C-|,|C+|;eq} using \eqref{Gr.IntNum;a,b,c}, \eqref{IntNumC;bc} and use induction on $i$ with $|C^-_0|=1$ and $|C^+_0| = \frac{q(q^{N-D}-1)}{q-1}$.
\end{proof}
We remark that from \eqref{new.partition} it turns out that the partition $\{C^\pm_{i}\}^{D-1}_{i=0}$ \emph{is} equitable; cf. \cite[Lemmas 5.1, 5.2]{2013LeeLAA}.
Let $\mathbf{W}$ be a subspace of $\mathbb{C}^X$ spanned by the set
\begin{equation}\label{Basis.W}
\mathcal{C}:=\{\hat{C}^-_0, \hat{C}^+_0,\hat{C}^-_1, \hat{C}^+_1, \ldots, \hat{C}^-_{D-1}, \hat{C}^+_{D-1}\}.
\end{equation}
Observe that $\mathcal{C}$ is an orthogonal ordered basis for $\mathbf{W}$.
Since $\{C^\pm_i\}^{D-1}_{i=0}$ is equitable, $\mathbf{W}$ is $A$-invariant.
Moreover, by the construction of \eqref{new.partition}, $\mathbf{W}$ is a module for both $M^*$ and $\widetilde M^*$.
Therefore, $\mathbf{W}$ is a $\mathbf{T}$-module.
Note that the $\mathbf{T}$-module $\mathbf{W}$ is generated by $\hat{x}$ since
\begin{equation*}
E^*_i\widetilde E^*_i\mathbf{J}\hat{x} = \hat{C}^-_i, \qquad
E^*_{i+1}\widetilde E^*_i\mathbf{J}\hat{x} = \hat{C}^+_i, \qquad
0 \leq i \leq D-1,
\end{equation*}
where $\mathbf{J}=\sum^D_{i=0}A_i$ and observe that $\mathbf{J}\hat{x}=\hat{X}$.
\begin{lemma}[{cf. \cite[Proposition 5.25]{2013LeeLAA}}]
The $\mathbf{T}$-module $\mathbf{W}$ is irreducible.
\end{lemma}
\begin{proof}
By semisimplicity of $\mathbf{T}$, $\mathbf{W}$ decomposes into an orthogonal direct sum of irreducible $\mathbf{T}$-modules.
Among such modules, take one, denoted by $\mathbf{W}_0$, which is not orthogonal to $\hat{x}$.
Then $E^*_0\mathbf{W}_0$ contains $\hat{x}$, from which it follows that the irreducible $\mathbf{T}$-module $\mathbf{W}_0$ contains $\hat{x}$.
Since the $\mathbf{T}$-module $\mathbf{W}$ is generated by $\hat{x}$, we have $\mathbf{W}=\mathbf{W}_0$. The result follows.
\end{proof}
We remark that the irreducible $\mathbf{T}$-module $\mathbf{W}$ is generated by $\hat{C}$ as well. We call $\mathbf{W}$ the \emph{primary} $\mathbf{T}$-module.
We describe the action of $\mathbf{T}$ on the basis $\mathcal{C}$ for $\mathbf{W}$.
Note that $\hat{C}^\pm_{-1}=0$ and $\hat{C}^\pm_{D}=0$.
\begin{lemma}\label{lem:actionAdjMatA}
The action of $A$ on $\hat{C}^\pm_i$, $0 \leq i \leq D-1$, is given by
\begin{align*}
A.\hat{C}^-_i
& = q^{2i}\gauss{D-i}{1}\gauss{N-D+1-i}{1}\hat{C}^-_{i-1} + q^{2i}\gauss{D-i}{1}\hat{C}^+_{i-1} \\
& \qquad + \left(q\gauss{D}{1}\gauss{N-D}{1} - q^{2i+1}\gauss{D-i}{1}\gauss{N-D-i}{1} - \gauss{i+1}{1}\gauss{i}{1} \right) \hat{C}^-_{i} \\
& \qquad + q^i\gauss{i+1}{1} \hat{C}^+_{i} + \gauss{i+1}{1}^2 \hat{C}^-_{i+1}, \\
A.\hat{C}^+_i
& = q^{2i+1}\gauss{D-i}{1}\gauss{N-D-i}{1}\hat{C}^+_{i-1} + q^{2i+1}\gauss{N-D-i}{1} \hat{C}^-_{i} \\
& \qquad + \left(q\gauss{D}{1}\gauss{N-D}{1} - q^{2i+2}\gauss{D-1-i}{1}\gauss{N-D-i}{1} - \gauss{i+1}{1}^2 \right)\hat{C}^+_{i} \\
& \qquad + q^{i+1}\gauss{i+1}{1} \hat{C}^-_{i+1} + \gauss{i+2}{1}\gauss{i+1}{1} \hat{C}^+_{i+1}.
\end{align*}
\end{lemma}
\begin{proof}
From the structure of \eqref{new.partition}, we routinely find both
\begin{align*}
A.\hat{C}^-_i & = \widetilde b_{i-1}\hat{C}^-_{i-1} + (\widetilde b_{i-1}-b_i)\hat{C}^+_{i-1} + (\widetilde a_i-b_i+\widetilde b_i) \hat{C}^-_i + (c_{i+1}-\widetilde c_i)\hat{C}^+_i + c_{i+1}\hat{C}^-_{i+1},\\
A.\hat{C}^+_i & = b_{i}\hat{C}^+_{i-1} + (b_{i}-\widetilde b_i)\hat{C}^-_{i} + (\widetilde a_i-c_{i+1}+\widetilde c_i) \hat{C}^+_i + (\widetilde c_{i+1}- c_{i+1})\hat{C}^-_{i+1} + \widetilde c_{i+1}\hat{C}^+_{i+1},
\end{align*}
for $0 \leq i \leq D-1$.
Evaluate these equations using \eqref{Gr.IntNum;a,b,c} and Lemma \ref{tilde;a,b,c(i)}. The result follows.
\end{proof}
\begin{lemma}\label{lem:actionMatA*,wtA*}
The actions of $A^*$, $\widetilde A^*$ on $\hat{C}^\pm_i$, $0 \leq i \leq D-1$, are given by
\begin{align*}
&&& A^*.\hat{C}^-_i = \theta^*_i \hat{C}^-_i, && A^*.\hat{C}^+_i = \theta^*_{i+1}\hat{C}^+_i, &&\\
&&& \widetilde A^*.\hat{C}^-_i = \widetilde \theta^*_i \hat{C}^-_i, && \widetilde A^*.\hat{C}^+_i = \widetilde \theta^*_{i}\hat{C}^+_i, &&
\end{align*}
where $\theta^*_i$ are from \eqref{Gr.dualEigval} and $\widetilde\theta^*_i$ are from \eqref{Gr.dualEigval.C}.
\end{lemma}
\begin{proof}
Immediate from \eqref{A*;linComb(E*i)} and \eqref{tilde(A);linComb(tildeE*i)}.
\end{proof}
\section{Leonard systems of dual $q$-Hahn type}\label{Section:LS of dual q-Hahn}
In this section, we discuss a family of Leonard systems said to have dual $q$-Hahn type and some properties we need in the paper.
We begin by recalling the notion of Leonard systems \cite{2001TerLAA}.
Let $\mathsf d$ be a non-negative integer and let $\mathsf V$ be a $\mathbb{C}$-vector space with dimension $\mathsf d+1$.
Assume that the element $\mathsf A \in \mathrm{End}(\mathsf V)$ is \emph{multiplicity-free}, i.e., $\mathsf A$ has $\mathsf d+1$ mutually distinct eigenvalues $\unslant\stheta_0, \unslant\stheta_1, \ldots, \unslant\stheta_\mathsf d$.
For $0\leq i \leq \mathsf d$, define $\mathsf E_i \in \mathrm{End}(\mathsf V)$ such that
$$
\mathsf E_i=\prod_{\substack{ 0 \le j \le \mathsf{d} \\ j\ne i} } \frac{\mathsf A-\unslant\stheta_j\mathsf I}{\unslant\stheta_i-\unslant\stheta_j},
$$
where $\mathsf I$ is the identity of $\mathrm{End}(\mathsf V)$.
Observe that (i) $\mathsf A\mathsf E_i = \unslant\stheta_i\mathsf E_i$, $0 \leq i \leq \mathsf d$, (ii) $\mathsf E_i\mathsf E_j = \delta_{ij}\mathsf E_i$, $0 \leq i,j \leq \mathsf d$, and (iii) $\sum^{\mathsf d}_{i=0} \mathsf E_i = \mathsf I$.
We call $\mathsf E_i$ the \emph{primitive idempotent} of $\mathsf A$ associated with $\unslant\stheta_i$.
\begin{definition}[{\cite[Definition 1.4]{2001TerLAA}}]\label{Def:LS}
By a \emph{Leonard system} on $\mathsf V$, we mean a sequence
\begin{equation}\label{def:LS}
\mathsf \Phi = \{ \mathsf A, \mathsf A^*, \{\mathsf E_i\}^{\mathsf d}_{i=0}, \{\mathsf E^*_i\}^{\mathsf d}_{i=0} \}
\end{equation}
of elements in $\mathrm{End}(\mathsf{V})$ that satisfy (i)--(iii) below.
\begin{itemize}
\item[(i)] Each of $\mathsf{A}, \mathsf{A}^*$ is multiplicity-free in $\mathrm{End}(\mathsf{V})$.
\item[(ii)] $\{\mathsf E_i\}^{\mathsf d}_{i=0}$ (resp. $\{\mathsf E^*_i\}^{\mathsf d}_{i=0}$) is an ordering of the primitive idempotents of $\mathsf A$ (resp. $\mathsf A^*$).
\item[(iii)] For $0 \leq i, j \leq \mathsf d$, both
\begin{equation}
\mathsf E_i \mathsf A^* \mathsf E_j = \begin{cases} 0 & \textrm{ if } |i-j|>1, \\ \ne 0 & \textrm{ if } |i-j|=1, \end{cases}
\quad \text{ and } \quad
\mathsf E^*_i \mathsf A \mathsf E^*_j = \begin{cases} 0 & \textrm{ if } |i-j|>1, \\ \ne 0 & \textrm{ if } |i-j|=1. \end{cases} \qquad
\end{equation}
\end{itemize}
We call $\mathsf d$ the \emph{diameter} of $\mathsf \Phi$.
\end{definition}
\begin{note}
In a common notational convention, $\mathsf A^*$ denotes the conjugate-transpose of $\mathsf A$.
We are not using this convention.
The elements $\mathsf A, \mathsf A^*$ in $\eqref{def:LS}$ are arbitrary subject to (i)--(iii) above.
\end{note}
Let $\mathsf \Phi=\{ \mathsf A, \mathsf A^*, \{\mathsf E_i\}^{\mathsf d}_{i=0}, \{\mathsf E^*_i\}^{\mathsf d}_{i=0} \}$ be a Leonard system on $\mathsf V$.
Let $\mathsf \Phi'$ be a Leonard system on a $(\mathsf d+1)$-dimensional $\mathbb{C}$-vector space $\mathsf V'$.
We say that $\mathsf \Phi'$ is \emph{isomorphic} to $\mathsf \Phi$ if there is a $\mathbb{C}$-algebra isomorphism $\sigma: \mathrm{End}(\mathsf{V}) \to \mathrm{End}(\mathsf{V'})$ such that $\mathsf \Phi'=\mathsf \Phi^\sigma = \{ \mathsf A^\sigma, \mathsf A^{*\sigma}, \{\mathsf E^\sigma_{i}\}^{\mathsf d}_{i=0}, \{\mathsf E^{*\sigma}_i\}^{\mathsf d}_{i=0} \}$.
Consider two sequences
\begin{equation}
\mathsf \Phi^*=\{ \mathsf A^*, \mathsf A, \{\mathsf E^*_{i}\}^{\mathsf d}_{i=0}, \{\mathsf E_i\}^{\mathsf d}_{i=0} \}, \qquad
\mathsf \Phi^{\Downarrow}=\{ \mathsf A, \mathsf A^*, \{\mathsf E_{\mathsf d-i}\}^{\mathsf d}_{i=0}, \{\mathsf E^*_i\}^{\mathsf d}_{i=0} \}
\end{equation}
Then both $\mathsf \Phi^*$ and $\mathsf \Phi^{\Downarrow}$ satisfy the conditions (i)--(iii) in Definition \ref{Def:LS}, and thus they are Leonard systems on $\mathsf V$.
For $0 \leq i \leq \mathsf d$, let $\unslant\stheta_i$ (resp. $\unslant\stheta^*_i$) be an eigenvalue of $\mathsf A$ (resp. $\mathsf A^*$).
Then there exists nonzero scalars $\unslant\svarphi_i$, $1 \leq i \leq \mathsf d$, and an isomorphism of $\mathbb{C}$-algebras $\natural$ from $\mathrm{End}(\mathsf{V})$ to the full matrix algebra $\mathbb{C}^{(\mathsf d+1)\times(\mathsf d+1)}$ such that (cf. \cite[Theorem 3.2]{2001TerLAA})
\begin{equation}
A^\natural = \begin{bmatrix}
\unslant\stheta_0 & & & & {\mathbf 0} \\[0.2em]
1 & \unslant\stheta_1 & \\[0.2em]
& 1 & \unslant\stheta_2 & \\
& & \ddots & \ddots \\[0.2em]
{\mathbf 0}& & & 1 & \unslant\stheta_\mathsf d
\end{bmatrix},
\qquad
A^{*\natural} = \begin{bmatrix}
\unslant\stheta^*_0 & \unslant\svarphi_1 & & & {\mathbf 0} \\[0.2em]
& \unslant\stheta^*_1 & \unslant\svarphi_2\\
& & \unslant\stheta^*_2 & \ddots \\
& & & \ddots & \unslant\svarphi_\mathsf d\\[0.2em]
{\mathbf 0}& & & & \unslant\stheta^*_\mathsf d
\end{bmatrix}.
\end{equation}
We call the sequence $\{\unslant\svarphi_i\}^{\mathsf d}_{i=1}$ the \emph{first split sequence} of $\mathsf \Phi$.
Let $\{\unslant\sphi_i\}^{\mathsf d}_{i=1}$ denote the first split sequence of $\mathsf \Phi^{\Downarrow}$ and call this the \emph{second split sequence} of $\mathsf \Phi$.
By the \emph{parameter array} of $\mathsf \Phi$, we mean the sequence
\begin{equation}\label{Def:PA}
( \{\unslant\stheta_i\}^{\mathsf d}_{i=0}, \{\unslant\stheta^*_i\}^{\mathsf d}_{i=0}, \{\unslant\svarphi_i\}^{\mathsf d}_{i=1}, \{\unslant\sphi_i\}^{\mathsf d}_{i=1} ).
\end{equation}
Take a non-zero vector $u$ in $\mathsf E_0 \mathsf V$.
Then the set $\{\mathsf E^*_iu\}^{\mathsf d}_{i=0}$ forms a \emph{$\mathsf \Phi$-standard basis}\footnote{Dually, we can consider a $\mathsf \Phi^*$-standard basis $\{\mathsf E_iu^*\}^{\mathsf d}_{i=0}$ for $\mathsf V$ with a non-zero $u^* \in \mathsf E^*_0\mathsf V$.} for $\mathsf V$, i.e., the set $\{\mathsf E^*_iu\}^{\mathsf d}_{i=0}$ satisfies both (i) $\mathsf E^*_iu \in \mathsf E^*_i\mathsf V$, $0\leq i \leq \mathsf d$; (ii) $\sum^\mathsf d_{i=0}E^*_iu \in \mathsf E_0\mathsf V$.
Applying $\mathsf A$ to $\mathsf E^*_iu$ and using Definition \ref{Def:LS}(iii), there exist the scalars $\mathsf a_i$, $\mathsf b_i$, $\mathsf c_i$, $0 \leq i \leq \mathsf d$, the so-called \emph{intersection numbers} of $\mathsf\Phi$, such that
$\mathsf b_\mathsf d=\mathsf c_0=0$, $\mathsf b_{i-1}\mathsf c_i \ne 0$, $1\leq i \leq \mathsf d$, and
\begin{equation}\label{eq;3-term.AE*i}
\mathsf A \mathsf E^*_i u = \mathsf b_{i-1}\mathsf E^*_{i-1}u + \mathsf a_i\mathsf E^*_i u + \mathsf c_{i+1}\mathsf E^*_{i+1}u,
\end{equation}
where $\mathsf b_{-1}\mathsf E^*_{-1}u=0$ and $\mathsf c_{\mathsf d+1}\mathsf E^*_{\mathsf d+1}u=0$.
Note that $\mathsf a_i+\mathsf b_i+\mathsf c_i = \unslant\stheta_0$ for $0 \leq i \leq \mathsf d$.
The intersection numbers $\mathsf b_i$ and $\mathsf c_i$ are given in terms of the parameter array \eqref{Def:PA} by (cf. \cite[Theorem 17.7]{2004TerLAA})
\begin{align}
\mathsf b_i &= \unslant\svarphi_{i+1}\frac{(\unslant\stheta^*_i-\unslant\stheta^*_0)(\unslant\stheta^*_i-\unslant\stheta^*_1)\cdots(\unslant\stheta^*_i-\unslant\stheta^*_{i-1})}{(\unslant\stheta^*_{i+1}-\unslant\stheta^*_0)(\unslant\stheta^*_{i+1}-\unslant\stheta^*_1)\cdots (\unslant\stheta^*_{i+1}-\unslant\stheta^*_i)}, \qquad 0 \leq i \leq \mathsf d-1, \label{i.n.Phi;b} \\
\mathsf c_i &= \unslant\sphi_{i}\frac{(\unslant\stheta^*_i-\unslant\stheta^*_{i+1})(\unslant\stheta^*_i-\unslant\stheta^*_{i+2})\cdots(\unslant\stheta^*_i-\unslant\stheta^*_{\mathsf d})}{(\unslant\stheta^*_{i-1}-\unslant\stheta^*_i)(\unslant\stheta^*_{i-1}-\unslant\stheta^*_{i+1})\cdots (\unslant\stheta^*_{i-1}-\unslant\stheta^*_\mathsf d)}, \qquad 1 \leq i \leq \mathsf d. \label{i.n.Phi;c}
\end{align}
Using the intersection numbers $\mathsf a_i$, $\mathsf b_i$, $\mathsf c_i$, define a sequence of polynomials $\{\mathsf v_i\}^{\mathsf d}_{i=0}$ in $\mathbb{C}[\lambda]$ as follows:
\begin{equation}\label{eq:poly.v(i)}
\mathsf v_0:=1, \qquad
\lambda \mathsf v_i = \mathsf b_{i-1}\mathsf v_{i-1} + \mathsf a_i\mathsf v_i + \mathsf c_{i+1}\mathsf v_{i+1}, \quad 0 \leq i \leq \mathsf d-1,
\end{equation}
where $\mathsf b_{-1}\mathsf v_{-1}=0$.
Observe that $\deg(\mathsf v_i)=i$ for $0 \leq i \leq \mathsf d$ since $\mathsf c_j \ne 0$, $1 \leq j \leq \mathsf d$.
We say that the polynomial $\mathsf v_i$ is \emph{associated with} $\mathsf\Phi$.
By \eqref{eq;3-term.AE*i}, it follows
\begin{equation}\label{vi(A)E*i}
\mathsf v_i(\mathsf A). \mathsf E^*_0 u=\mathsf E^*_iu, \qquad 0 \leq i \leq \mathsf d.
\end{equation}
We normalize the polynomial $\mathsf v_i$ by setting
\begin{equation}\label{poly:f}
\mathsf f_i:=\mathsf v_i/\mathsf k_i, \qquad 0 \leq i \leq \mathsf d,
\end{equation}
where $\mathsf k_i = \mathsf b_0\mathsf b_1\cdots \mathsf b_{i-1}/\mathsf c_1\mathsf c_2\cdots\mathsf c_i$.
Then it turns out that (cf. \cite[Theorem 17.4]{2004TerLAA})
\begin{equation}\label{poly;f}
\mathsf f_i(\lambda) = \sum^i_{n=0} \frac{(\unslant\stheta^*_i-\unslant\stheta^*_0)(\unslant\stheta^*_i-\unslant\stheta^*_1)\cdots (\unslant\stheta^*_i-\unslant\stheta^*_{n-1})(\lambda-\theta_0)\cdots (\lambda-\theta_{n-1})}{\unslant\svarphi_1\unslant\svarphi_2\cdots \unslant\svarphi_n}, \qquad 0 \leq i \leq \mathsf d.
\end{equation}
The Leonard system is uniquely determined up to isomorphism by the parameter array, cf. \cite[Theorem 1.9]{2001TerLAA}, and all families of the parameter arrays of Leonard systems are displayed in \cite{2005TerDCC} as parametric form.
We now recall the dual $q$-Hahn family of Leonard systems.
For the rest of this section, assume that $q$ is a nonzero scalar such that $q^i \ne 1$ for $1 \leq i \leq D$.
\begin{definition}[{\cite[Example 5.5]{2005TerDCC}}]\label{Def:LS.dualqHahn}
Let $\mathsf \Phi$ be a Leonard system on $\mathsf V$ with diameter $\mathsf d$.
Let the sequence \eqref{Def:PA} be the parameter array of $\mathsf\Phi$.
Then $\mathsf \Phi$ is said to have \emph{dual $q$-Hahn} type if there exist scalars $\mathsf a$, $\mathsf a^*$, $\mathsf b$, $\mathsf b^*$, $\mathsf c$, $\mathsf r$ such that
\begin{align*}
\unslant\stheta_i = \mathsf a+\mathsf b q^{-i} + \mathsf c q^i, \qquad \qquad
\unslant\stheta^*_i = \mathsf a^* + \mathsf b^*q^{-i},
\end{align*}
for $0 \leq i \leq \mathsf d$, and
\begin{align*}
\unslant\svarphi_i & = \mathsf b\sfb^*q^{1-2i}(1-q^i)(1-q^{i-\mathsf d-1})(1-\mathsf r q^i), \\
\unslant\sphi_i & = \mathsf c\mathsf b^* q^{\mathsf d+1-2i}(1-q^i)(1-q^{i-\mathsf d-1})(1-\mathsf b\mathsf r\mathsf c^{-1}q^{i-\mathsf d}),
\end{align*}
for $1 \leq i \leq \mathsf d$, where $\mathsf b$, $\mathsf b^*$, $\mathsf c$, $\mathsf r$ are nonzero\footnote{In the case $\mathsf r=0$, the Leonard system $\mathsf \Phi$ has dual $q$-Krawtchouk type; cf. \cite[Definition 5.2]{2018LeeTanakaSIGMA}} and neither of $\mathsf rq^i$, $\mathsf c \mathsf b^{-1} \mathsf r^{-1} q^{i-1}$ is equal to $1$ for $1\leq i \leq \mathsf d$.
We call $(\mathsf a, \mathsf a^*, \mathsf b, \mathsf b^*,\mathsf c, \mathsf r; q, \mathsf d)$ the \emph{parameter sequence of $\mathsf\Phi$}.
\end{definition}
From now on, let $\mathsf \Phi$ be a Leonard system of dual $q$-Hahn type as in Definition \ref{Def:LS.dualqHahn}.
From \eqref{i.n.Phi;b}, \eqref{i.n.Phi;c}, the intersection numbers of $\mathsf \Phi$ are given by
\begin{equation}\label{IntNum(dualqHahn)}
\mathsf b_i = \mathsf b(1-q^{i-\mathsf d})(1-\mathsf r q^{i+1}),
\qquad
\mathsf c_i = (1-q^{i})(\mathsf c-\mathsf b\mathsf r q^{i-\mathsf d}),
\end{equation}
for $0 \leq i \leq \mathsf d$.
Evaluate \eqref{poly;f} at $\lambda=\unslant\stheta_j$ using Definition \ref{Def:LS.dualqHahn}.
Then we get (cf. \cite[Example 5.5]{2005TerDCC})
\begin{equation}
\mathsf f_i(\unslant\stheta_j)
= {}_3\phi_2 \left( \left. \begin{array}{c} q^{-i}, \ q^{-j}, \ \mathsf t^2 q^j \\ \mathsf r q, \ q^{-\mathsf d} \end{array}\right| q, q\right), \qquad 0 \leq i, j \leq \mathsf d,
\end{equation}
where
\begin{equation}\label{scalar;t.b.c}
\mathsf t^2 = \mathsf b^{-1}\mathsf c.
\end{equation}
The polynomials $\mathsf f_i$ form the \emph{dual $q$-Hahn polynomials} \cite[Section 14.7]{2010KLS} in a variable $\lambda(x) = \mathsf a+\mathsf b q^{-x} + \mathsf c q^x$.
For notational convenience, fix a square root $\mathsf t$ of $\mathsf t^2$.
Set $x = \log_q(\mathsf t^{-1}\zeta)$ in $\lambda(x)$ so that
\begin{equation}
\lambda=\lambda(\log_q(\mathsf t^{-1}\zeta))=\mathsf a + \mathsf b\mathsf{t}\zeta^{-1} + \mathsf c\mathsf{t}^{-1}\zeta.
\end{equation}
We renormalize $\mathsf f_i(\lambda)$ by setting
\begin{equation}\label{eq;poly.h(i)}
\mathsf h_i(\zeta)
:= \frac{(\mathsf r q;q)_i(q^{-\mathsf d};q)_i}{\mathsf{t}^i} \mathsf f_i(\lambda)
= \frac{(\mathsf r q;q)_i(q^{-\mathsf d};q)_i}{\mathsf{t}^i} {}_3\phi_2 \left( \left. \begin{array}{c} q^{-i}, \ \mathsf{t}\zeta^{-1}, \ \mathsf{t}\zeta \\ \mathsf r q, \ q^{-\mathsf d} \end{array}\right| q, q\right),
\end{equation}
for $0 \leq i \leq \mathsf d$.
We note that $\mathsf h_i(\zeta)$ are \emph{monic} symmetric Laurent polynomials in a variable $\zeta$, i.e., the coefficient of their highest degree term in $\zeta$ is one, and note also that the $\mathsf h_i(\zeta)$ has the highest degree $i$ and the lowest degree $-i$.
Since $\mathsf h_i(\zeta)$ depends on the parameters $\mathsf b$, $\mathsf c$, $\mathsf r$, $\mathsf d$, and $q$, we write
\begin{equation}\label{LaurentPoly;h(i)}
\mathsf h_i = \mathsf h_i(\zeta) := \mathsf h_i(\zeta; \mathsf b, \mathsf c,\mathsf r, \mathsf d; q), \qquad 0 \leq i \leq \mathsf d,
\end{equation}
and say that $\mathsf{h}_i$ is \emph{associated with} $\mathsf\Phi$.
\begin{lemma}
Let $\widetilde{\mathsf V}$ be a $\mathbb{C}$-vector space containing $\mathsf V$ as a subspace.
Let $\mathsf X$ be an invertible element of $\mathrm{End}(\widetilde{\mathsf V})$ such that $\mathsf V$ is invariant under $\mathsf X+\mathsf X^{-1}$.
Suppose that the action of $\mathsf A$ on $\mathsf V$ is same as the action of $\mathsf a + \mathsf b\mathsf t(\mathsf X + \mathsf X^{-1})= \mathsf a + \mathsf b\mathsf{t}\mathsf X^{-1} + \mathsf c\mathsf{t}^{-1}\mathsf X$ on $\mathsf V$, where $\mathsf A$ is an element of $\mathsf\Phi$ as in Definition \ref{Def:LS.dualqHahn} and $\mathsf a$, $\mathsf b$, $\mathsf c$ are parameters of $\mathsf\Phi$ and $\mathsf t$ is from \eqref{scalar;t.b.c}. Then, on $\mathsf V$
\begin{equation}\label{hi(X)=vi(A)}
\mathsf h_i(\mathsf X) = \mathsf t^i(q;q)_i(\mathsf t^{-2}\mathsf r q^{1-\mathsf d};q)_i\mathsf v_i(\mathsf A), \qquad 0 \leq i \leq \mathsf d,
\end{equation}
where $\mathsf h_i$ and $\mathsf v_i$ are from \eqref{LaurentPoly;h(i)} and \eqref{eq:poly.v(i)}, respectively.
Moreover, for a non-zero vector $u \in \mathsf E_0\mathsf V$,
\begin{equation}\label{hi(X)=vi(A).E*iu}
\mathsf h_i(\mathsf X).\mathsf E^*_0u = \mathsf t^i(q;q)_i(\mathsf t^{-2}\mathsf r q^{1-\mathsf d};q)_i\mathsf E^*_iu, \qquad 0 \leq i \leq \mathsf d.
\end{equation}
\end{lemma}
\begin{proof}
From \eqref{eq;poly.h(i)}, we have
\begin{equation}\label{pf:eq.hi(X)=fi(A)}
\mathsf h_i(\mathsf X) = \frac{(\mathsf r q;q)_i(q^{-\mathsf d};q)_i}{\mathsf t^i}\mathsf f_i(\mathsf A),
\end{equation}
on $\mathsf V$.
Evaluate $\mathsf f_i(\mathsf A)$ in \eqref{pf:eq.hi(X)=fi(A)} using \eqref{poly:f}, \eqref{IntNum(dualqHahn)} and simplify the result to get \eqref{hi(X)=vi(A)}.
To obtain \eqref{hi(X)=vi(A).E*iu}, use \eqref{vi(A)E*i} and \eqref{hi(X)=vi(A)}.
\end{proof}
We finish this section with a comment.
With reference to $\mathsf \Phi$, we define the scalars
\begin{equation}\label{eq:scalar m(i)}
\mathsf m_i=\mathrm{trace}(\mathsf E_i \mathsf E^*_0), \qquad 0 \leq i \leq \mathsf d.
\end{equation}
By \cite[Theorem 17.12]{2004TerLAA}, the $\mathsf m_i$, $0 \leq i \leq \mathsf d$, are given in terms of the parameter array of $\mathsf \Phi$ by
\begin{equation}\label{formula.mi:PA}
\mathsf m_i = \frac{\unslant\svarphi_1\unslant\svarphi_2\cdots\unslant\svarphi_i \unslant\sphi_1\unslant\sphi_2\cdots \unslant\sphi_{\mathsf d-i}}{(\unslant\stheta^*_0-\unslant\stheta^*_1)\cdots(\unslant\stheta^*_0-\unslant\stheta^*_\mathsf d)(\unslant\stheta_i-\unslant\stheta_0)\cdots (\unslant\stheta_i-\unslant\stheta_{i-1})(\unslant\stheta_i-\unslant\stheta_{i+1})\cdots(\unslant\stheta_i-\unslant\stheta_\mathsf d)}.
\end{equation}
Applying the formulas in Definition \ref{Def:LS.dualqHahn} with \eqref{scalar;t.b.c} to \eqref{formula.mi:PA}, the $\mathsf m_i$ are given by
\begin{equation}\label{m_formula}
\mathsf m_i =
q^{\mathsf d-i}\mathsf r^{\mathsf d-i}
\frac{(q^{i+1};q)_{\mathsf d-i} (\mathsf r^{-1}\mathsf t^2q^i;q)_{\mathsf d-i}(\mathsf r q;q)_i(1-\mathsf t^2q^{2i})}{(q;q)_{\mathsf d-i}(\mathsf t^2q^i;q)_{\mathsf d+1}}.
\end{equation}
\section{The primary $\mathbf T$-module $\mathbf{W}$}\label{Section:primary T-module}
Recall the primary $\mathbf{T}$-module $\mathbf{W}$ of $\Gamma$ from Section \ref{Section:algebraT}.
In this section, we treat four dual $q$-Hahn Leonard systems that naturally arise from the structure of $\mathbf{W}$.
Since $\mathbf{W}$ is a module for both $T$ and $\widetilde T$, it contains both $M\hat{x}$ (as a $T$-module) and $M\hat{C}$ (as a $\widetilde T$-module).
Let $M\hat{x}^\perp$ (resp. $M\hat{C}^\perp$) denote the orthogonal complement of $M\hat{x}$ (resp. $M\hat{C}$) in $\mathbf{W}$.
Note that $M\hat{x}^\perp$ is an irreducible $T$-submodule of $\mathbf{W}$ with dimension $D-1$ and $M\hat{C}^\perp$ is an irreducible $\widetilde T$-submodule of $\mathbf{W}$ with dimension $D$; cf. \cite[Sections 6, 7]{2013LeeLAA}. Therefore, $\mathbf{W}$ decomposes in two ways:
\begin{align}
&&\mathbf{W} & = M\hat{x} \oplus M\hat{x}^\perp && \text{(orthogonal direct sum of irreducible $T$-modules)} && \label{ODS:Mx,Mxp}\\
&&& = M\hat{C} \oplus M\hat{C}^\perp && \text{(orthogonal direct sum of irreducible $\widetilde T$-modules)}\label{ODS:MC,MCp}.
\end{align}
For the rest of the paper, we set a non-zero scalar
\begin{equation}\label{scalar;tau}
\tau=-q^{(-N-1)/2}.
\end{equation}
Indeed, it turns out that $\tau^2=b^{-1}c$, where $b,c$ are from Proposition \ref{prop:LSonMx,Mxp}(I); cf. \eqref{scalar;t.b.c}.
\begin{proposition}\label{prop:LSonMx,Mxp}
Recall the matrices $A$, $A^*$, $\{E_i\}^D_{i=0}$, $\{E^*_i\}^D_{i=0}$ in $T$ and the irreducible $T$-submodules $M\hat{x}$, $M\hat{x}^\perp$ of $\mathbf{W}$ from \eqref{ODS:Mx,Mxp}.
Define the following sequences of matrices by
\begin{equation*}
\Phi:=(A, A^*,\{E_i\}^D_{i=0}, \{E^*_i\}^D_{i=0}) |_{M\hat{x}}, \quad
\Phi^\perp:=(A, A^*,\{E_i\}^{D-1}_{i=1}, \{E^*_i\}^{D-1}_{i=1}) |_{M\hat{x}^\perp},
\end{equation*}
where $|_{Z}$ means that each of the matrices in the sequence is restricted to the subspace $Z$ of $\mathbf{W}$.
The following {\rm (I)}, {\rm (II)} hold.
\begin{itemize}
\item[\rm{(I)}]
The sequence $\Phi$ is a Leonard system on $M\hat{x}$ that has dual $q$-Hahn type.
The parameter sequence of $\Phi$ is
$
(a, a^*, b, b^*,c,r; q, D),
$
where
\begin{align*}
& a = \frac{q-q^{N-D+1}-q^{D+1}-1}{(q-1)^2}, && a^* = \frac{(q^N-q)(2-q^D-q^{N-D})}{(q-1)(q^D-1)(q^{N-D}-1)},\\
& b=\frac{q^{N+1}}{(q-1)^2}, && b^*=\frac{(q^N-q)(q^N-1)}{(q-1)(q^D-1)(q^{N-D}-1)}, \\
& c=\frac{1}{(q-1)^2}, && r = q^{D-N-1}.
\end{align*}
Moreover, for $0 \leq i \leq D$, the folloinwg (i)--(iii) hold.
\begin{itemize}
\item[\rm (i)] The vectors $A_i\hat{x}$ $(=\hat{C}^+_{i-1}+\hat{C}^-_i)$ form a $\Phi$-standard basis for $M\hat{x}$.
\item[\rm (ii)] The intersection numbers $b_i$, $c_i$ of $\Phi$ are given by
\begin{equation*}
b_i = q^{2i+1}\gauss{D-i}{1}\gauss{N-D-i}{1}, \qquad c_i = \gauss{i}{1}^2. \qquad
\end{equation*}
\item[\rm (iii)] The monic dual $q$-Hahn polynomials $h_i$ associated with $\Phi$ (cf. \eqref{LaurentPoly;h(i)}) are given by
\begin{equation}\label{prop:eq:h(i)}
h_i(\zeta) = h_i(\zeta; b, c, r, D;q)=
\tau^i(q;q)^2_i v_i(a + b\tau \zeta^{-1} + c\tau^{-1}\zeta), \qquad
\end{equation}
where $v_i$ are the polynomials associated with $\Phi$ as in \eqref{eq:poly.v(i)}.
\end{itemize}
\item[\rm{(II)}]
The sequence $\Phi^\perp$ is a Leonard system on $M\hat{x}^\perp$ that has dual $q$-Hahn type.
The parameter sequence of $\Phi^\perp$ is
$
(a^\perp, a^{*\perp}, b^\perp, b^{*\perp},c^\perp,r^\perp; q, D-2),
$
where
\begin{equation*}
(a^\perp, a^{*\perp}, b^\perp, b^{*\perp},c^\perp,r^\perp)=(a, a^*, bq^{-1}, b^*q^{-1}, cq, rq).
\end{equation*}
Moreover, for $0 \leq i \leq D-2$, the folloinwg (i)--(iii) hold.
\begin{itemize}
\item[\rm (i)]
The vectors
\begin{equation}\label{prop:eq:u-perp(i)}
u^\perp_i = (q^{D-i-1}-1)\hat{C}^+_i + (q^{-i-1}-1)\hat{C}^-_{i+1} \qquad
\end{equation}
form a $\Phi^\perp$-standard basis for $M\hat{x}^\perp$.
\item[\rm (ii)]
The intersection numbers $b^\perp_i$, $c^\perp_i$ of $\Phi^\perp$ are given by
\begin{equation*}
b^\perp_i = q^{2i+3}\gauss{D-i-2}{1}\gauss{N-D-i-1}{1}, \qquad c^\perp_i = q\gauss{i}{1}\gauss{i+1}{1}. \qquad
\end{equation*}
\item[\rm (iii)] The monic dual $q$-Hahn polynomials $h^\perp_i$ associated with $\Phi^\perp$ are given by
\begin{equation}\label{prop:eq:h-perp(i)}
h^\perp_i(\zeta) = h^\perp_i(\zeta; bq^{-1}, cq, rq, D-2;q) =
\tau^iq^i(q;q)_i(q^2;q)_i v^\perp_i(a + b\tau \zeta^{-1} + c\tau^{-1}\zeta), \qquad
\end{equation}
where $v^\perp_i$ are the polynomials associated with $\Phi^\perp$ as in \eqref{eq:poly.v(i)}.
\end{itemize}
\end{itemize}
\end{proposition}
\begin{proof}
(I): Refer to \cite[Section 6]{2013LeeLAA} (or \cite[Section 6]{2018LeeTanakaSIGMA}).
Parts (i) and (ii) routinely follows.
For (iii), evaluate \eqref{eq;poly.h(i)} using \eqref{poly:f}, part (ii), and the parameter sequence of $\Phi$.\\
(II): Similar.
\end{proof}
\begin{remark}
We note that for each irreducible $T$-module $W$ the restrictions of $A$ and $A^*$ on $W$ induce a Leonard system of dual $q$-Hahn type; cf. \cite[Theorem 4.6]{2015GaGaHoLAA}.
\end{remark}
\begin{proposition}\label{prop:LSonMC,MCp}
Recall the matrices $A$, $\widetilde A^*$, $\{E_i\}^D_{i=0}$, $\{\widetilde E^*_i\}^{D-1}_{i=0}$ in $\widetilde T$ and the irreducible $\widetilde T$-submodules $M\hat{C}$, $M\hat{C}^\perp$ of $\mathbf{W}$ from \eqref{ODS:MC,MCp}.
Define the following sequences of matrices by
\begin{equation*}
\widetilde\Phi:=(A, \widetilde A^*,\{E_i\}^{D-1}_{i=0}, \{\widetilde E^*_i\}^{D-1}_{i=0})|_{M\hat{C}}, \quad
\widetilde\Phi^\perp:=(A, A^*,\{E_i\}^{D}_{i=1}, \{\widetilde E^*_i\}^{D-1}_{i=0})|_{M\hat{C}^\perp},
\end{equation*}
where $|_{Z}$ means that each of the matrices in the sequence is restricted to the subspace $Z$ of $\mathbf{W}$.
Recall the parameter sequence $(a, a^*, b, b^*,c,r; q, D)$ of $\Phi$ from Proposition \ref{prop:LSonMx,Mxp}.
The following {\rm (I)}, {\rm (II)} hold.
\begin{itemize}
\item[\rm{(I)}]
The sequence $\widetilde\Phi$ is a Leonard system on $M\hat{C}$ that has dual $q$-Hahn type.
The parameter sequence of $\widetilde\Phi$ is
$
(\widetilde a, \widetilde a^{*}, \widetilde b, \widetilde b^{*}, \widetilde c, \widetilde r; q, D-1),
$
where \footnote{See \cite[Proposition 4.6]{2011TanakaE-JC} for a general result.}
\begin{equation*}
(\widetilde a, \widetilde a^{*}, \widetilde b, \widetilde b^{*}, \widetilde c, \widetilde r) = \left( a, a^*+\frac{q-1}{q^{N-D+1}-1}b^*, b, \frac{q^{N-D}-1}{q^{N-D+1}-1}b^*, c, r \right).
\end{equation*}
Moreover, $0 \leq i \leq D-1$, the folloinwg (i)--(iii) hold.
\begin{itemize}
\item[\rm (i)] The vectors $\hat{C}_i$ $(=\hat{C}^-_i+\hat{C}^+_i)$ form a $\widetilde\Phi$-standard basis for $M\hat{C}$.
\item[\rm (ii)] The intersection numbers $\widetilde b_i$, $\widetilde c_i$ of $\widetilde \Phi$ are given by
\begin{equation*}
\widetilde{b}_i = q^{2i+2}\gauss{D-i-1}{1}\gauss{N-D-i}{1}, \qquad
\widetilde{c}_i = \gauss{i+1}{1}\gauss{i}{1}. \qquad
\end{equation*}
\item[\rm (iii)] The monic dual $q$-Hahn polynomials $\widetilde h_i$ associated with $\widetilde \Phi$ are given by
\begin{equation}\label{prop:eq:wt-h(i)}
\widetilde h_i(\zeta) = \widetilde h_i(\zeta; b, c, r, D-1;q)=
\tau^i(q;q)_i(q^2;q)_i\widetilde v_i(a + b\tau \zeta^{-1} + c\tau^{-1}\zeta), \qquad
\end{equation}
where $\widetilde v_i$ are the polynomials associated with $\widetilde \Phi$ as in \eqref{eq:poly.v(i)}.
\end{itemize}
\item[\rm{(II)}]
The matrices of $\widetilde \Phi^\perp$ act on $M\hat{C}^\perp$ as a Leonard system that has dual $q$-Hahn type.
The parameter sequence of $\widetilde \Phi^\perp$ is
$
(\widetilde a^\perp, \widetilde a^{*\perp}, \widetilde b^\perp, \widetilde b^{*\perp}, \widetilde c^\perp, \widetilde r^\perp; q, D-1),
$
where
\begin{equation*}
(\widetilde a^\perp, \widetilde a^{*\perp}, \widetilde b^\perp, \widetilde b^{*\perp}, \widetilde c^\perp, \widetilde r^\perp)=\left( a, a^*+\frac{q-1}{q^{N-D+1}-1}b^*, bq^{-1}, \frac{q^{N-D}-1}{q^{N-D+1}-1}b^*, cq, rq \right).
\end{equation*}
Moreover, $0 \leq i \leq D-1$, the folloinwg (i)--(iii) hold.
\begin{itemize}
\item[\rm (i)] The vectors
\begin{equation}\label{prop:eq:wt-u-perp(i)}
\widetilde u^\perp_i = (q^{N-D-i}-1)\hat{C}^-_i + (q^{-i-1}-1)\hat{C}^+_{i} \qquad
\end{equation}
form a $\widetilde\Phi^\perp$-standard basis for $M\hat{C}^\perp$.
\item[\rm (ii)] The intersection numbers $\widetilde b^\perp_i$, $\widetilde c^\perp_i$ of $\widetilde \Phi^\perp$ are given by
\begin{equation*}
\widetilde b^\perp_i = q^{2i+2}\gauss{D-i-1}{1}\gauss{N-D-i-1}{1}, \qquad c^\perp_i = q\gauss{i}{1}^2. \qquad
\end{equation*}
\item[\rm (iii)] The monic dual $q$-Hahn polynomials $\widetilde h^\perp_i$ associated with $\widetilde \Phi^\perp$ are given by
\begin{equation}\label{prop:eq:wt-h-perp(i)}
\widetilde h^\perp_i(\zeta) = \widetilde h^\perp_i(\zeta; bq^{-1},cq, rq, D-1;q)=
\tau^iq^i(q;q)^2_i\widetilde v^\perp_i(a+b\tau\zeta^{-1}+c\tau^{-1}\zeta),
\end{equation}
where $\widetilde v^\perp_i$ are the polynomials associated with $\widetilde \Phi^\perp$ as in \eqref{eq:poly.v(i)}.
\end{itemize}
\end{itemize}
\end{proposition}
\begin{proof}
Similar to Proposition \ref{prop:LSonMx,Mxp}.
\end{proof}
We comment on the decompositions \eqref{ODS:Mx,Mxp} and \eqref{ODS:MC,MCp} of $\mathbf{W}$.
We first consider the orthogonal direct sum of $\mathbf{W}$ from \eqref{ODS:Mx,Mxp}.
Let $\pi\in\mathrm{End}(\mathbf{W})$ be the orthogonal projection onto $M\hat{x}$, i.e., the element $\pi$ satisfies $(\pi-1)M\hat{x}=0$ and $\pi(M\hat{x}^\perp)=0$.
We give an action of $\pi$ on $\mathbf{W}$ as follows.
Consider a $\Phi$-standard basis $\{A_i\hat{x}\}^D_{i=0}$ for $M\hat{x}$ and a $\Phi^\perp$-standard basis $\{u^\perp_i\}^{D-2}_{i=0}$ for $M\hat{x}^\perp$.
From them, we find that
\begin{align}
& \hat{C}^+_{i-1} = \frac{q^i-1}{q^D-1}A_i\hat{x} + \frac{q^i}{q^D-1}u^\perp_{i-1}, \label{eq(1):C+-;Aix,upi}\\
& \hat{C}^-_i = \frac{q^D-q^i}{q^D-1}A_i\hat{x} + \frac{q^i}{1-q^D} u^\perp_{i-1}, \label{eq(2):C+-;Aix,upi}
\end{align}
for $1 \leq i \leq D-1$.
From \eqref{eq(1):C+-;Aix,upi} and \eqref{eq(2):C+-;Aix,upi}, the action of $\pi$ on $\hat{C}^\pm_j$ is given by
\begin{equation}\label{eq:pi.Cpm(i)}
\pi.\hat{C}^+_{i-1} = \frac{q^i-1}{q^D-1}(\hat{C}^+_{i-1}+\hat{C}^-_i),
\qquad
\pi.\hat{C}^-_i = \frac{q^D-q^i}{q^D-1}(\hat{C}^+_{i-1}+\hat{C}^-_i),
\end{equation}
for $1 \leq i \leq D-1$. Moreover, we have $\pi.\hat{C}^-_0 = \hat{C}^-_0$ and $\pi.\hat{C}^+_{D-1} = \hat{C}^+_{D-1}$.
Next, we consider the orthogonal direct sum of $\mathbf{W}$ from \eqref{ODS:MC,MCp}.
Let $\widetilde\pi\in\mathrm{End}(\mathbf{W})$ be the orthogonal projection onto $M\hat{C}$, i.e., the element $\widetilde\pi$ satisfies $(\widetilde\pi-1)M\hat{C}=0$ and $\widetilde\pi(M\hat{C}^\perp)=0$.
We give an action of $\widetilde\pi$ on $\mathbf{W}$ as follows.
Consider a $\widetilde\Phi$-standard basis $\{\hat{C}_i\}^{D-1}_{i=0}$ for $M\hat{C}$ and a $\widetilde\Phi^\perp$-standard basis $\{\widetilde u^\perp_i\}^{D-1}_{i=0}$ for $\widetilde M\hat{x}^\perp$.
From them, we find that
\begin{align}
& \hat{C}^-_i = \frac{q^{i+1}-1}{q^{N-D+1}-1}\hat{C}_i + \frac{q^{i+1}}{q^{N-D+1}-1}\widetilde u^\perp_i,\label{eq(1):C+-;Ci,tilde upi} \\
& \hat{C}^+_i = \frac{q^{N-D+1}-q^{i+1}}{q^{N-D+1}-1}\hat{C}_i - \frac{q^{i+1}}{q^{N-D+1}-1}\widetilde u^\perp_i, \label{eq(2):C+-;Ci,tilde upi}
\end{align}
for $0 \leq i \leq D-1$.
From \eqref{eq(1):C+-;Ci,tilde upi} and \eqref{eq(2):C+-;Ci,tilde upi}, the action of $\widetilde \pi$ on $\hat{C}^\pm_j$ is given by
\begin{equation}\label{eq:wtpi.Cpm}
\widetilde\pi(\hat{C}^-_i) = \frac{q^{i+1}-1}{q^{N-D+1}-1}(\hat{C}^-_i+\hat{C}^+_i),
\qquad
\widetilde\pi(\hat{C}^+_i) = \frac{q^{N-D+1}-q^{i+1}}{q^{N-D+1}-1}(\hat{C}^-_i+\hat{C}^+_i),
\end{equation}
for $0 \leq i \leq D-1$.
\section{The confluent Cherednik algebra $\mathcal{H}_{\mathrm V}$}\label{Section:Hv}
The double affine Hecke algebra (DAHA), or Cherednik algebra, for a reduced affine root system was defined by Cherednik \cite{1992Cherednik}, and the definition was extended to non-reduced affine root systems of type $(C^\vee_n, C_n)$ by Sahi \cite{1999SahiAnnMath}.
In \cite{2016Mazzocco} Mazzocco introduced seven new algebras as degenerations of the DAHAs of type $(C^\vee_1, C_1)$ and established a new relation between the theory of the Painlev\'e equations and the theory of the $q$-Askey scheme.
Among the seven algebras, the confluent Cherednik algebra $\mathcal{H}_{\mathrm{III}}$ \cite[(3.86)--(3.91)]{2016Mazzocco} has been shown to be recognized to a certain nil-DAHA of type $(C^\vee_1, C_1)$, which is associated with dual polar graphs; cf. \cite[Remark 8.4]{2018LeeTanakaSIGMA}.\footnote{In \cite[Definition 8.1]{2018LeeTanakaSIGMA} we overlooked the relation $\mathcal{U}'=q\mathcal{U}\mathcal{X}$ which is obtained by applying double-dot normalization to $\mathcal{T}_1=\mathcal{T}'^{-1}\mathcal{X}^{-1}$.
}
In the present paper, we shall focus our attention on the algebra $\mathcal{H}_\mathrm{V}$, another confluence Cherednik algebra among the seven, and discuss how $\mathcal{H}_\mathrm{V}$ is related to our Grassmann graph $\Gamma$.
\begin{definition}[{\cite[Theorem 3.2.(3.73)--(3.78)]{2016Mazzocco}}]\label{Def:CheAlgHv}
Let $k,k',u,q$ be non-zero scalars in $\mathbb{C}$.
The algebra $\mathcal{H}_{\mathrm V}=\mathcal{H}_{\mathrm V}(k,k',u;q)$ is the associative $\mathbb{C}$-algebra with generators $\mathcal{T}, \mathcal{T}'$, $\mathcal{U}$, $\mathcal{U}'$ and relations
\begin{align}
&&&(\mathcal{T}-k)(\mathcal{T}+k^{-1})=0, & & \mathcal{U}'(\mathcal{U}'+1)=0, && \label{Def:algHv.rel(1)} \\
&&&(\mathcal{T}'-k')(\mathcal{T}'+k'^{-1})=0, & & \mathcal{U}(\mathcal{U}+u^{-1})=0, && \label{Def:algHv.rel(2)}\\
&&& q^{1/2}\mathcal{T}'\mathcal{T}\mathcal{U}' = \mathcal{U} + u^{-1},
&& q^{1/2}\mathcal{U}\mathcal{T}'\mathcal{T} = \mathcal{U}' + 1. \label{Def:algHv.rel(3)}
\end{align}
We remark that $\mathcal{T}$, $\mathcal{T}'$ are invertible.
\end{definition}
We now construct an $\mathcal{H}_\mathrm{V}$-module structure.
Recall the prime power $q$.
In what follows, we set
\begin{equation}\label{scalars:k,k',u}
k=\sqrt{-1}q^{(D-N-1)/2}, \qquad k' = \sqrt{-1}q^{-D/2}, \qquad u=q^{D-\frac{N}{2}}.
\end{equation}
Using these scalars, define the matrices as follows: for $0 \leq i \leq D-1$
\begin{equation}\label{matrices:t(i),u'(i)}
t(i) := k\begin{bmatrix}
1-q^{i+1}+q^{N-D+1} & q^{N-D+1}(q^{D-N+i}-1) \\
1-q^{i+1} & q^{i+1}
\end{bmatrix},
\qquad
u'(i) := \begin{bmatrix}
-1 & 0 \\
q^{-i-1} & 0
\end{bmatrix},
\end{equation}
and for $1 \leq i \leq D-1$,
\begin{equation}\label{matrices:t'(i),u(i)}
t'(i) := k'\begin{bmatrix}
1-q^i+q^D & q^i-q^D \\
1-q^i & q^i
\end{bmatrix},
\qquad
u(i) := u^{-1}\begin{bmatrix}
-1 & 1-q^{D-i} \\
0 & 0
\end{bmatrix}.
\end{equation}
Moreover, define
\begin{equation}\label{matrices:t'(0,D),u(0,D)}
t'(0) := \begin{bmatrix} k' \end{bmatrix}, \quad
t'(D) := \begin{bmatrix} k' \end{bmatrix}, \quad
u(0) := \begin{bmatrix} 0 \end{bmatrix}, \quad
u(D) :=\begin{bmatrix} -u^{-1} \end{bmatrix}.
\end{equation}
\begin{lemma}\label{lem:Hv.rels}
The following (i), (ii) hold.
\begin{itemize}
\item[(i)] Let $0 \leq i \leq D-1$.
Then $\mathrm{trace}(t(i)) = k-k^{-1}$ and $\det(t(i)) = -1$. Moreover,
\begin{equation*}
(t(i)-k)(t(i)+k^{-1})=0, \qquad u'(i)(u'(i)+1) = 0.
\end{equation*}
\item[(ii)] Let $1 \leq i \leq D-1$.
Then $\mathrm{trace}(t'(i)) = k'-k'^{-1}$ and $\det(t'(i)) = -1$. Moreover,
\begin{equation*}
(t'(i)-k')(t'(i)+k'^{-1})=0, \qquad u(i)(u(i)+u^{-1}) = 0.
\end{equation*}
\end{itemize}
\end{lemma}
\begin{proof}
Follows from \eqref{matrices:t(i),u'(i)} and \eqref{matrices:t'(i),u(i)}.
\end{proof}
Using the matrices \eqref{matrices:t(i),u'(i)}--\eqref{matrices:t'(0,D),u(0,D)}, define the block diagonal matrices in $\mathbb{C}^{2D\times 2D}$:
\begin{align*}
&\pmb{T} := \mathrm{blockdiag}\Big[ t(0), t(1), \ldots , t(D-1) \Big],
&&\pmb{T}' := \mathrm{blockdiag}\Big[ t'(0), t'(1), \ldots , t'(D-1), t'(D) \Big], \\
& \pmb{U}' := \mathrm{blockdiag}\Big[ u'(0), u'(1), \ldots , u'(D-1) \Big],
&&{\pmb U} := \mathrm{blockdiag}\Big[ u(0), u(1), \ldots , u(D-1), u(D) \Big].
\end{align*}
\begin{lemma}\label{lem:mat.A*,wtA*}
We have both
\begin{align*}
qk (q^{-1/2}\pmb{U}\pmb{T}' + \pmb{T}\pmb{U}') & = \mathrm{diag}\Big( 1, q^{-1}, q^{-1}, q^{-2}, q^{-2}, \ldots, q^{-(D-1)}, q^{-(D-1)}, q^{-D} \Big), \\
qk (q^{1/2}\pmb{U}\pmb{T}' + \pmb{T}\pmb{U}') & = \mathrm{diag}\Big( 1, 1, q^{-1}, q^{-1}, q^{-2}, q^{-2}, \ldots, q^{-(D-1)}, q^{-(D-1)} \Big).
\end{align*}
\end{lemma}
\begin{proof}
Use \eqref{matrices:t(i),u'(i)}--\eqref{matrices:t'(0,D),u(0,D)}.
The result routinely follows.
\end{proof}
\begin{proposition}\label{prop:alg-homo:HtoC(2D)}
There exists a $\mathbb{C}$-algebra homomorphism from $\mathcal{H}_{\mathrm V}$ to the full matrix algebra $\mathbb{C}^{2D\times 2D}$ that sends
\begin{equation*}
\mathcal{T} \mapsto \pmb T, \qquad \mathcal{T}' \mapsto \pmb{T}', \qquad \mathcal{U} \mapsto \pmb{U}, \qquad \mathcal{U}' \mapsto \pmb{U}'.
\end{equation*}
\end{proposition}
\begin{proof}
The matrices $\pmb T$, $\pmb T'$, $\pmb U$, $\pmb U'$ satisfy the defining relations \eqref{Def:algHv.rel(1)}--\eqref{Def:algHv.rel(3)} by Lemma \ref{lem:Hv.rels}. The result follows.
\end{proof}
\noindent
Recall the $2D$-dimensional subspace $\mathbf{W}$ and its ordered basis $\mathcal{C}$ (cf. \eqref{Basis.W}) from Section \ref{Section:algebraT}.
\begin{corollary}\label{cor:Hv-moduleW}
There exists an $\mathcal{H}_{\mathrm{V}}$-module structure on $\mathbf{W}$ such that the matrices representing $\mathcal{T}$, $\mathcal{T}'$, $\mathcal{U}$, $\mathcal{U}'$ with respect to the ordered basis $\mathcal{C}$ are $\pmb{T}$, $\pmb{T}'$, $\pmb{U}$, $\pmb{U}'$, respectively.
\end{corollary}
\begin{proof}
Identifying $\mathrm{End}(\mathbf{W})$ with $\mathbb{C}^{2D\times 2D}$, we obtain a representation of $\mathcal{H}_\mathrm{V}$ on $\mathbf{W}$ by Proposition \ref{prop:alg-homo:HtoC(2D)}.
The result follows.
\end{proof}
By the comments below \eqref{Basis.W} and Corollary \ref{cor:Hv-moduleW}, the space $\mathbf{W}$ has a module structure for both $\mathbf{T}$ and $\mathcal{H}_{\mathrm V}$.
We shall discuss how the $\mathcal{H}_{\mathrm V}$-action on $\mathbf{W}$ is related to the $\mathbf{T}$-action on $\mathbf{W}$.
Recall the scalar $\tau$ from \eqref{scalar;tau}. Observe that $\tau=kk'$.
For notational convenience, we define
\begin{equation}\label{eq:X}
\mathcal{X} := \mathcal{T}'\mathcal{T}.
\end{equation}
Observe that $\mathcal{X}$ is invertible since $\mathcal{T}'$, $\mathcal{T}$ are invertible.
We also define
\begin{equation*}\label{eq:A,A*,wtA*}
\mathcal{A} := \mathcal{X}+\mathcal{X}^{-1}, \qquad \mathcal{A}^* := qk(q^{-1/2}\mathcal{U}\mathcal{T}'+\mathcal{T}\mathcal{U}'), \qquad \widetilde{\mathcal{A}}^* := qk(q^{1/2}\mathcal{U}\mathcal{T}'+\mathcal{T}\mathcal{U}').
\end{equation*}
We give the actions of the elements $\mathcal{A}$, $\mathcal{A}^*$, $\widetilde{\mathcal{A}}^*$ of $\mathcal{H}_\mathrm{V}$ on $\mathbf{W}$ with the basis $\mathcal{C}$.
\begin{lemma}\label{lem:actionX}
The following (i), (ii) hold.
\begin{itemize}
\item[(i)]
The actions of $\mathcal{X}$ and $\mathcal{X}^{-1}$ on $\hat{C}^-_i$, $0 \leq i \leq D-1$, are given as linear combination with the following terms and coefficients.
\begin{align*}
& \mathcal{X}.\hat{C}^-_i : && \mathcal{X}^{-1}.\hat{C}^-_i : \\
&\begin{tabular}{c|c}
term & coefficient \\
\hline \hline \\ [-0.8em]
$\hat C^+_{i-1}$ & $\tau(q^i-q^D)(q^{N-D+1}-q^{i+1}+1) $\\[0.5em]
$\hat C^-_{i}$ & $\tau q^i(q^{N-D+1}-q^{i+1}+1)$ \\[0.5em]
$\hat C^+_{i}$ & $\tau (1-q^{i+1})(q^D-q^{i+1}+1)$ \\[0.5em]
$\hat C^-_{i+1}$ & $\tau (q^{i+1}-1)^2$ \\[0.5em]
\end{tabular} \quad ,
&&
\begin{tabular}{c|c}
term & coefficient \\
\hline \hline \\ [-0.8em]
$\hat{C}^-_{i-1}$ & $\tau(1-q^{i-D})(q^{N+1}-q^{D+i}) $\\[0.5em]
$\hat C^+_{i-1}$ & $\tau (q^D-q^{i})(q^{N-D+1}-q^i+1)$ \\[0.5em]
$\hat C^-_{i}$ & $\tau q^{i+1}(q^D-q^i+1)$ \\[0.5em]
$\hat C^+_{i}$ & $\tau (q^{i+1}-1)(q^D-q^i+1)$ \\[0.5em]
\end{tabular}.
\end{align*}
\item[(ii)]
The actions of $\mathcal{X}$ and $\mathcal{X}^{-1}$ on $\hat{C}^+_i$, $0 \leq i \leq D-1$, are given as linear combination with the following terms and coefficients.
\begin{align*}
& \mathcal{X}.\hat{C}^+_i : && \mathcal{X}^{-1}.\hat{C}^+_i : \\
& \begin{tabular}{c|c}
term & coefficient \\
\hline \hline \\ [-0.8em]
$\hat C^+_{i-1}$ & $\tau q^{N+1}(q^{i-D}-1)(q^{D-N+i}-1)$\\[0.5em]
$\hat C^-_{i}$ & $\tau q^{N-D+1+i}(q^{D-N+i}-1)$ \\[0.5em]
$\hat C^+_{i}$ & $\tau q^{i+1}(q^D-q^{i+1}+1)$ \\[0.5em]
$\hat C^-_{i+1}$ & $\tau q^{i+1}(1-q^{i+1})$ \\[0.5em]
\end{tabular} \quad ,
&&\begin{tabular}{c|c}
term & coefficient \\
\hline \hline \\ [-0.8em]
$\hat C^-_{i}$ & $\tau q^{N-D+2+i}(1-q^{D-N+i})$\\[0.5em]
$\hat C^+_{i}$ & $\tau q^{i+1}(q^{N-D+1}-q^{i+1}+1)$ \\[0.5em]
$\hat C^-_{i+1}$ & $\tau q^{i+2}(q^{i+1}-1)$ \\[0.5em]
$\hat C^+_{i+1}$ & $\tau (q^{i+1}-1)(q^{i+2}-1)$ \\[0.5em]
\end{tabular}.
\end{align*}
\end{itemize}
\end{lemma}
\begin{proof}
Routine using \eqref{matrices:t(i),u'(i)}--\eqref{matrices:t'(0,D),u(0,D)} and Corollary \ref{cor:Hv-moduleW}.
\end{proof}
\begin{lemma}\label{lem:actionA}
The following (i), (ii) hold.
\begin{itemize}
\item[(i)] The action of $\mathcal{A}$ on $\hat{C}^{\pm}_i$, $0 \leq i \leq D-1$, is given as linear combination with the following terms and coefficients.
\begin{align*}
& \mathcal{A}.\hat{C}^-_i : && \mathcal{A}.\hat{C}^+_i : \\
& \begin{tabular}{c|c}
term & coefficient \\
\hline \hline \\ [-0.8em]
$\hat C^-_{i-1}$ & $\tau q^{2i}(q^{D-i}-1)(q^{N-D-i+1}-1)$\\[0.5em]
$\hat C^+_{i-1}$ & $\tau q^{2i}(q-1)(q^{D-i}-1)$\\[0.5em]
$\hat C^-_{i}$ & $\tau q^i(q^{N-D+1}+q^{D+1}-2q^{i+1}+q+1)$ \\[0.5em]
$\hat C^+_{i}$ & $\tau q^i(q-1)(q^{i+1}-1)$ \\[0.5em]
$\hat C^-_{i+1}$ & $\tau (q^{i+1}-1)^2$ \\[0.5em]
\end{tabular} \quad ,
&& \begin{tabular}{c|c}
term & coefficient \\
\hline \hline \\ [-0.8em]
$\hat C^+_{i-1}$ & $\tau q^{2i+1}(q^{D-i}-1)(q^{N-D-i}-1)$\\[0.5em]
$\hat C^-_{i}$ & $\tau q^{2i+1}(q-1)(q^{N-D-i}-1)$\\[0.5em]
$\hat C^+_{i}$ & $\tau q^{i+1}(q^{N-D+1}+q^D-2q^{i+1}+2)$ \\[0.5em]
$\hat C^-_{i+1}$ & $\tau q^{i+1}(q-1)(q^{i+1}-1)$ \\[0.5em]
$\hat C^+_{i+1}$ & $\tau (q^{i+1}-1)(q^{i+2}-1)$ \\[0.5em]
\end{tabular}.
\end{align*}
\item[(ii)] The actions of $\mathcal{A}^*$ and $\widetilde{\mathcal{A}}^{*}$ on $\hat{C}^\pm_i$, $0 \leq i \leq D-1$, are given by
\begin{align*}
&&& \mathcal{A}^*.\hat{C}^-_i = q^{-i}\hat{C}^-_i, && \mathcal{A}^*.\hat{C}^+_i = q^{-i-1}\hat{C}^+_i, &&\\
&&& \widetilde{\mathcal{A}}^*.\hat{C}^-_i = q^{-i}\hat{C}^-_i, && \widetilde{\mathcal{A}}^*.\hat{C}^+_i = q^{-i}\hat{C}^+_i.
\end{align*}
\end{itemize}
\end{lemma}
\begin{proof}
(i): Routine using Lemma \ref{lem:actionX}. \\
(ii): By Lemma \ref{lem:mat.A*,wtA*}.
\end{proof}
\begin{theorem}\label{thm:1stResult}
Recall the generators $A$, $A^*$, $\widetilde A^*$ of $\mathbf{T}$ and the elements $\mathcal{A}$, $\mathcal{A}^*$, $\widetilde{\mathcal{A}}^*$ of $\mathcal{H}_\mathrm{V}$.
On $\mathbf{W}$, we have
\begin{align}
&A = \tau b \mathcal{A} + a, \label{thm;eq(1);A,A*,wtA*} \\
&A^* = b^* \mathcal{A}^* + a^*, \label{thm;eq(2);A,A*,wtA*} \\
&\widetilde A^* = \widetilde b^* \widetilde{\mathcal{A}}^* + \widetilde a^*,\label{thm;eq(3);A,A*,wtA*}
\end{align}
where $\tau$ is from \eqref{scalar;tau} and $a$, $a^*$, $\widetilde a^*$, $b$, $b^*$, $\widetilde b^*$ are from Propositions \ref{prop:LSonMx,Mxp}(I) and \ref{prop:LSonMC,MCp}(I).
\end{theorem}
\begin{proof}
The identity \eqref{thm;eq(1);A,A*,wtA*} follows from Lemma \ref{lem:actionAdjMatA} and Lemma \ref{lem:actionA}(i).
The identities \eqref{thm;eq(2);A,A*,wtA*} and \eqref{thm;eq(3);A,A*,wtA*} follow from Lemma \ref{lem:actionMatA*,wtA*} and Lemma \ref{lem:actionA}(ii).
\end{proof}
\begin{remark}
(i) By Theorem \ref{thm:1stResult} and since $\mathbf{W}$ is irreducible as a $\mathbf{T}$-module, it follows that an $\mathcal{H}_\mathrm{V}$-module $\mathbf{W}$ is \emph{irreducible}.
\smallskip \noindent
(ii) On the $\mathcal{H}_\mathrm{V}$-module $\mathbf{W}$, the elements $\mathcal{T}$ and $\mathcal{T}'$ are both diagonalizable.
Moreover, the element $(\mathcal{T}+k^{-1})/(k+k^{-1})$ (resp. $(\mathcal{T}'+k'^{-1})/(k'+k'^{-1})$) acts as the projection from $\mathbf{W}$ onto the eigenspace of $\mathcal{T}$ (resp. $\mathcal{T}'$) corresponding to $k$ (resp. $k'$).
\end{remark}
\begin{theorem}\label{thm:2ndResult}
Recall the orthogonal projection $\pi$ (resp. $\widetilde \pi$) from $\mathbf{W}$ onto $M\hat{x}$ (resp. $M\hat{C}$). On $\mathbf{W}$, we have
\begin{equation}\label{thm;eq;pi=T'}
\pi = \frac{\mathcal{T}'+k'^{-1}}{k'+k'^{-1}}, \qquad \qquad
\widetilde\pi = \frac{\mathcal{T}+k^{-1}}{k+k^{-1}}.
\end{equation}
\end{theorem}
\begin{proof}
Use \eqref{eq:pi.Cpm(i)} and the matrix $\pmb T'$ to obtain the first identity in \eqref{thm;eq;pi=T'}.
Use \eqref{eq:wtpi.Cpm} and the matrix $\pmb T$ to obtain the second identity. The result follows.
\end{proof}
We should like to make a comment on a nil-DAHA of type $(C^\vee_1, C_1)$.
We first recall the definition of the (ordinary) DAHA of type $(C^\vee_1, C_1)$.
The DAHA $\mathcal{H}=\mathcal{H}(\kappa_0, \kappa_1,\kappa_0',\kappa_1';q)$ is the associative $\mathbb{C}$-algebra with generators $\mathbf{T}^{\pm1}_0$, $\mathbf{T}^{\pm1}_1$, and $\mathbf{X}^{\pm1}$ and relations (cf. \cite[Section 6.4]{2003Macdonald}, \cite[Section 3]{1999SahiAnnMath})
\begin{equation}\label{DAHArels(1)}
(\mathbf{T}_i-\kappa_i)(\mathbf{T}_i+\kappa_i^{-1})=0, \qquad
(\mathbf{T}'_i-\kappa'_i)(\mathbf{T}'_i+\kappa'^{-1}_i)=0, \qquad
i = 0,1,
\end{equation}
where,
\begin{equation*}\label{DAHArels(2)}
\mathbf{T}'_0 := q^{-1/2}\mathbf{X}\mathbf{T}^{-1}_0, \qquad
\mathbf{T}'_1 := \mathbf{X}^{-1}\mathbf{T}^{-1}_1.
\end{equation*}
In \cite[Remark 8.2]{2018LeeTanakaSIGMA}, we specialized some defining relations of $\mathcal{H}$ using the so-called ``double-dot normalization'' method; cf. \cite[Section 2.5]{2015CheOrrMathZ}.
We then obtained a certain nil-DAHA, which is isomorphic to the algebra $\mathcal{H}_{\mathrm{III}}$.
In the present paper, by employing the techniques used in \cite{2018LeeTanakaSIGMA}, we shall specialize the algebra $\mathcal{H}$ to obtain a new nil-DAHA, denoted by $\overline{\mathcal{H}}$, which is well-suited in the context of Grassmann graphs.
Set
\begin{equation}\label{ddot:T_1,T'_1}
\ddot{\mathbf{T}}_1 := \kappa_1\mathbf{T}_1, \qquad
\ddot{\mathbf{T}}'_1 :=\kappa_1\mathbf{T}'_1.
\end{equation}
Apply \eqref{ddot:T_1,T'_1} to the relations \eqref{DAHArels(1)} for $i=1$ to get
\begin{equation}\label{ddot-normal;eq(2)}
(\ddot{\mathbf{T}}_1-\kappa_1^2)(\ddot{\mathbf{T}}_1+1)=0, \qquad
(\ddot{\mathbf{T}}'_1-\kappa_1\kappa'_1)(\ddot{\mathbf{T}}'_1+\kappa_1\kappa'^{-1}_1)=0.
\end{equation}
Observe that $\mathbf{T}'_1 = \mathbf{X}^{-1}(\mathbf{T}_1-\kappa_1+\kappa_1^{-1})$ and $\mathbf{T}_1 = (\mathbf{T}'_1-\kappa'_1+\kappa'^{-1}_1)\mathbf{X}^{-1}$.
Use these and \eqref{ddot:T_1,T'_1} to get
\begin{equation}\label{ddot-normal;eq(1)}
\ddot{\mathbf{T}}'_1 = \mathbf{X}^{-1}(\ddot{\mathbf{T}}_1-\kappa^2_1+1), \qquad
\ddot{\mathbf{T}}_1 = (\ddot{\mathbf{T}}'_1-\kappa_1\kappa'_1+\kappa_1\kappa'^{-1}_1)\mathbf{X}^{-1}.
\end{equation}
Thus, $\mathcal{H}$ has a presentation with new generators $\mathbf{T}^{\pm1}_0$, $\ddot{\mathbf{T}}_1$, and $\mathbf{X}^{\pm1}$ and relations \eqref{DAHArels(1)} at $i=1$ and \eqref{ddot-normal;eq(2)} and \eqref{ddot-normal;eq(1)}.
We now specialize the parameters $\kappa_1, \kappa'_1$.
Let $u \in \mathbb{C}$ be a nonzero scalar.
Set $\kappa'_1=u^{-1}\kappa_1$ in \eqref{ddot-normal;eq(2)} and \eqref{ddot-normal;eq(1)}.
Then, letting $\kappa_1 \to 0$, the relations \eqref{ddot-normal;eq(2)} and \eqref{ddot-normal;eq(1)} become
$$
\ddot{\mathbf{T}}_1(\ddot{\mathbf{T}}_1+1)=0, \qquad
\ddot{\mathbf{T}}'_1(\ddot{\mathbf{T}}'_1+u)=0, \qquad
\ddot{\mathbf{T}}'_1=\mathbf{X}^{-1}(\ddot{\mathbf{T}}_1+1), \qquad
\ddot{\mathbf{T}}_1=(\ddot{\mathbf{T}}'_1+u)\mathbf{X}^{-1}.
$$
Define
$$
\mathcal{T}:=\mathbf{T}_0, \qquad
\mathcal{U}:=u^{-1}\ddot{\mathbf{T}}_1, \qquad
\mathcal{X}:=q^{-1/2}\mathbf{X}, \qquad
k:=\kappa_0, \qquad
k':=\kappa_0'
$$
Then the algebra $\overline{\mathcal{H}}=\overline{\mathcal{H}}(k,k',u;q)$ obtained from this specialization has a presentation with generators $\mathcal{T}^{\pm1}$, $\mathcal{U}$, $\mathcal{X}^{\pm1}$ and relations
\begin{align*}
(\mathcal{T}-k)(\mathcal{T}+k^{-1})&=0, \\
(\mathcal{T}'-k')(\mathcal{T}'+k'^{-1})&=0,\\
\mathcal{U}(\mathcal{U}+u^{-1})&=0, \\
\mathcal{U}'(\mathcal{U}'+1)&=0, \\
q^{1/2}\mathcal{U}\mathcal{X} &= \mathcal{U}'+1.
\end{align*}
where
$$
\mathcal{T}'=\mathcal{X}\mathcal{T}^{-1}, \qquad \qquad
\mathcal{U}'=q^{-1/2}\mathcal{X}^{-1}(\mathcal{U}+u^{-1}).
$$
We call $\overline{\mathcal{H}}$ a \emph{nil-DAHA} of type $(C^\vee_1, C_1)$.
We shall remark that the nil-DAHA $\overline{\mathcal{H}}$ is isomorphic to the algebra $\mathcal{H}_\mathrm{V}(k,k',u;q)$ from Definition \ref{Def:CheAlgHv}.
\section{Non-symmetric dual $q$-Hahn polynomials}\label{Section:nonsym dual q-Hahn poly}
In this section, we shall define non-symmetric dual $q$-Hahn polynomials and give them a combinatorial interpretation.
Recall the Leonard systems $\Phi$, $\Phi^\perp$, $\widetilde \Phi$, and $\widetilde \Phi^\perp$ from Propositions \ref{prop:LSonMx,Mxp} and \ref{prop:LSonMC,MCp}.
Recall the sequences of monic dual $q$-Hahn polynomials
\begin{equation*}
\{h_i\}^D_{i=0}, \qquad \{h^\perp_i\}^{D-2}_{i=0}, \qquad \{\widetilde h_i\}^{D-1}_{i=0}, \qquad \{\widetilde h^\perp_i\}^{D-1}_{i=0}
\end{equation*}
associated with $\Phi$, $\Phi^\perp$, $\widetilde \Phi$, $\widetilde \Phi^\perp$, respectively, from \eqref{prop:eq:h(i)}, \eqref{prop:eq:h-perp(i)}, \eqref{prop:eq:wt-h(i)}, \eqref{prop:eq:wt-h-perp(i)}.
Define the following monic Laurent polynomials in $\mathbb{C}[\zeta, \zeta^{-1}]$ by
\begin{equation}\label{eq:p-perp,p-tilde,p-tildeperp}
p^\perp = \zeta^{-1}(\zeta-\tau)(\zeta-\tau^{-1}q^{-D}), \qquad
\widetilde p = \zeta^{-1}(\zeta-\tau^{-1}q^{-D}), \qquad
\widetilde p^\perp = \zeta^{-1}(\zeta-\tau).
\end{equation}
\begin{lemma}\label{lem:p-perp,p-tilde,p-tildeperp}
On the $\mathcal{H}_\mathrm{V}$-module $\mathbf{W}$, we have
\begin{equation}\label{lem:eq:three-poly-action}
p^\perp(\mathcal{X}).\hat{x} = \tau q(1-q) u_0^\perp, \qquad
\widetilde p(\mathcal{X}).\hat{x} = (1-q) \hat{C}, \qquad
\widetilde p^\perp(\mathcal{X}).\hat{x} = q^{D-N}\widetilde u^\perp_0,
\end{equation}
where $u^\perp_0$ is from \eqref{prop:eq:u-perp(i)} and $\widetilde u^\perp_0$ is from \eqref{prop:eq:wt-u-perp(i)}.
\end{lemma}
\begin{proof}
Note that $\hat{x}=\hat{C}^-_0$.
Setting $i=0$ in Lemma \ref{lem:actionX}(i), the actions of $\mathcal{X}$ and $\mathcal{X}^{-1}$ on $\hat{C}^-_0$ are given by
\begin{align*}
& \mathcal{X}.\hat{x} = \tau (q^{N-D+1}-q+1)\hat C^-_{0} + \tau (1-q)(1-q+q^D)\hat C^+_{0} + \tau (1-q)^2\hat C^-_{1}, \\
& \mathcal{X}^{-1}.\hat{x} = \tau q^{D+1}\hat C^-_{0} + \tau q^D(q-1)\hat C^+_{0}.
\end{align*}
Evaluate $p^\perp(\mathcal{X}).\hat{x}$ using these equations and simplify the result using \eqref{prop:eq:u-perp(i)} at $i=0$ to get the first equation in \eqref{lem:eq:three-poly-action}.
The remaining two equations in \eqref{lem:eq:three-poly-action} are similarly obtained.
\end{proof}
We define the non-symmetric Laurent polynomials $\ell^\pm_i$ in $\mathbb{C}[\zeta, \zeta^{-1}]$ as follows. For $0 \leq i \leq D-1$,
\begin{align}
& \ell^-_i (\zeta) := \frac{q^D-q^i}{\tau^{i}(q^D-1)(q;q)^2_i}\left( h_i - \frac{1-q^i}{q^D-q^i}p^\perp h^\perp_{i-1} \right), \label{eq(1):NonSymDqHahn} \\
& \ell^+_i (\zeta) := \frac{q^{i+1}-1}{\tau^{i+1}(q^D-1)(q;q)^2_{i+1}}\left( h_{i+1} -p^\perp h^\perp_i \right),\label{eq(2):NonSymDqHahn}
\end{align}
where
\begin{equation}\label{hperp(D-1)}
h^\perp_{D-1}(\zeta) := \prod^{D-1}_{j=1} (\zeta+\zeta^{-1}-\tau q^j - \tau^{-1} q^{-j}) = \zeta^{1-D}\prod^{D-1}_{j=1} (\zeta - \tau q^j)(\zeta - \tau^{-1}q^{-j}).
\end{equation}
\begin{lemma}\label{lem:h-perp(D-1)vanish}
Recall the subspace $M\hat{x}^\perp$ of $\mathbf{W}$ from \eqref{ODS:Mx,Mxp}.
Then $h^\perp_{D-1}(\mathcal{X})$ vanishes on $M\hat{x}^\perp$.
\end{lemma}
\begin{proof}
On the $\mathcal{H}_\mathrm{V}$-module $\mathbf{W}$, using \eqref{thm;eq(1);A,A*,wtA*} we find
\begin{align*}
h^\perp_{D-1}(\mathcal{X}) = \prod^{D-1}_{j=1} (\mathcal{X}+\mathcal{X}^{-1}-\tau q^j - \tau^{-1} q^{-j}) = (\tau b)^{1-D}\prod^{D-1}_{j=1} (A-\theta_j),
\end{align*}
where we recall $\tau^2=b^{-1}c$ and $\theta_j = a + bq^{-j}+cq^{j}$.
Since $M\hat{x}^\perp=\sum^{D-1}_{j=1}E_j(M\hat{x}^\perp)$, the result follows.
\end{proof}
We define another non-symmetric Laurent polynomials $\widetilde \ell^\pm_i$ as follows. For $0 \leq i \leq D-1$,
\begin{align*}
& \widetilde \ell^-_i (\zeta) := \frac{1}{\tau^i(1-q^{N-D+1})(q;q)^2_i}\left( \widetilde{p} \widetilde{h}_i - q^{N-D+1} \widetilde{p}^\perp \widetilde{h}^\perp_i \right),\\
& \widetilde \ell^+_i (\zeta) := \frac{q^{N-D+1}}{\tau^i(q^{N-D+1}-1)(q;q)^2_i}\left( \frac{1-q^{D-N+i}}{1-q^{i+1}}\widetilde{p} \widetilde{h}_i - \widetilde{p}^\perp \widetilde{h}^\perp_i \right).
\end{align*}
We shall now give a description of the role of the Laurent polynomials $\ell^{\pm}_i$, $\widetilde \ell^\pm_i$ on $\mathbf{W}$.
Recall the basis $\mathcal{C}$ for $\mathbf{W}$ from \eqref{Basis.W}.
\begin{proposition}\label{prop:C_i=ell_i(X)x}
On the $\mathcal{H}_\mathrm{V}$-module $\mathbf{W}$,
\begin{equation}\label{lem:eq:C_i=ell_i(X)x}
\hat{C}^-_i = \ell^-_i(\mathcal{X}).\hat{x} = \widetilde\ell^-_i(\mathcal{X}).\hat{x}, \qquad \hat{C}^+_i = \ell^+_i(\mathcal{X}).\hat{x} = \widetilde \ell^+_i(\mathcal{X}).\hat{x},
\end{equation}
for $0 \leq i \leq D-1$.
\end{proposition}
\begin{proof}
We first show the equation $\hat{C}^+_i = \ell^+_i(\mathcal{X}).\hat{x}$, $0 \leq i \leq D-1$.
Recall the equation \eqref{eq(1):C+-;Aix,upi}.
We assume $0 \leq i \leq D-2$.
Applying \eqref{vi(A)E*i} and \eqref{thm;eq(1);A,A*,wtA*} to each summand of the right side of \eqref{eq(1):C+-;Aix,upi} and using \eqref{prop:eq:h(i)} and \eqref{prop:eq:h-perp(i)}, we get
\begin{equation*}\label{lem:pf:eq:C_i=ell_i(X)x}
\hat{C}^+_i = \frac{q^{i+1}-1}{q^D-1}\frac{h_{i+1}(\mathcal{X})}{\tau^{i+1}(q;q)^2_{i+1}}\hat{x} + \frac{q^{i+1}}{q^D-1}\frac{h^\perp_i(\mathcal{X})}{\tau^iq^i(q;q)_i(q^2;q)_i}u^\perp_0.
\end{equation*}
Simplify the right side of this equation using the first equation in \eqref{lem:eq:three-poly-action}.
We find $\hat{C}^+_i = \ell^+_i(\mathcal{X}).\hat{x}$, $0 \leq i \leq D-2$.
We now assume $i=D-1$.
By the first equation of line \eqref{lem:eq:three-poly-action} and Lemma \ref{lem:h-perp(D-1)vanish}, it follows
\begin{equation}\label{prop:eq:p-perp.h-perp=0}
p^\perp(\mathcal{X})h^\perp_{D-1}(\mathcal{X}).\hat{x} = \tau q(1-q) h^\perp_{D-1}(\mathcal{X}).u^\perp_0 =0.
\end{equation}
By this comment, we find $\hat{C}^{+}_{D-1}=\ell^+_{D-1}(\mathcal{X}).\hat{x}$.
The desired result follows.
The remaining equations in \eqref{lem:eq:C_i=ell_i(X)x} are obtained in a similar way using \eqref{eq(2):C+-;Aix,upi}, \eqref{eq(1):C+-;Ci,tilde upi} and \eqref{eq(2):C+-;Ci,tilde upi}.
\end{proof}
\begin{remark}\label{rmk:degree.ell}
By \eqref{lem:eq:C_i=ell_i(X)x}, we have $\ell^-_i = \widetilde \ell^-_i$, $\ell^+_i = \widetilde \ell^+_i$, $0 \leq i \leq D-1$. From this, it follows that
\begin{itemize}
\itemsep-0.1em
\item $\ell^-_i$ has the highest degree $i$ and the lowest degree is $-i$;
\item $\ell^+_i$ has the highest degree $i$ and the lowest degree $-i-1$.
\end{itemize}
\end{remark}
Consider the subspace $\mathcal{L}$ of $\mathbb{C}[\zeta, \zeta^{-1}]$ defined by
\begin{equation}\label{eq:space L}
\mathcal{L} = \sum^{D-1}_{i=-D} \mathbb{C}\zeta^i.
\end{equation}
Observe that $\dim{\mathcal{L}}=2D$ and by Remark \ref{rmk:degree.ell} $\{\ell^\pm_i\}^{D-1}_{i=0}$ forms a basis for $\mathcal{L}$.
We call $\ell^\pm_i=\widetilde \ell^\pm_i$, $0 \leq i \leq D-1$, the \emph{non-symmetric dual $q$-Hahn polynomials}; see Figure \ref{NonsymDualqHahn}.
\begin{figure}
\centering
\scalemath{0.7}{
\begin{tikzpicture}
[scale=.8,thick,auto=left, every node/.style={circle}]
\node[fill=black,label=right:{\Large$\ell^-_0, (0,0)$}] (n1) at (0,0) {};
\node[draw, label=left:{\Large$\ell^+_0, (0,-1)$}] (n2) at (0,-2) {};
\node[fill=black,label=right:{\Large$\ell^-_1, (1,-1)$}] (n3) at (2,-2) {};
\node[draw,label=left:{\Large$\ell^+_1, (1,-2)$}] (n4) at (2,-4) {};
\node[fill=black,label=right:{\Large$\ell^-_2, (2,-2)$}] (n5) at (4,-4) {};
\node[draw,label=left:{\Large$\ell^+_2, (2,-3)$}] (n6) at (4,-6) {};
\node[fill=black,label=right:{\Large$\ell^-_3, (3,-3)$}] (n7) at (6,-6) {};
\node[draw,label=left:{\Large$\ell^+_3, (3,-4)$}] (n8) at (6,-8) {};
\node[fill=black,label=right:{}] (n9) at (8,-8) {};
\node (n10) at (8,-10) {};
\foreach \from/\to in
{n1/n2,n2/n3,n3/n4,n4/n5,n5/n6,n6/n7,n7/n8}
\draw (\from) -- (\to);
\draw (n1) -- (n2) ;
\draw (n3) -- (n4) ;
\draw (n5) -- (n6) ;
\draw (n7) -- (n8) ;
\draw (n8) -- (n9) [dashed];
\draw (n9) -- (n10) [dashed];
\end{tikzpicture}
\caption{Non-symmetric dual $q$-Hahn polynomials $\ell^\pm_i$}\label{NonsymDualqHahn}
\end{figure}
In this figure, each black node (resp. white node) represents the non-symmetric dual $q$-Hahn polynomial $\ell^-_i$ (resp. $\ell^+_i$).
For each $\ell^\pm_i$, the corresponding ordered pair $(m,n)$ means that $m$ is the highest degree and $n$ is the lowest degree of $\ell^\pm_i$.
Figure \ref{NonsymDualqHahn} shows how $\ell^\pm_i$ are interpreted in a combinatorial sense; cf. Figure \ref{2-dim'l partition}.
We may regard the non-symmetric dual $q$-Hahn polynomials $\ell^\pm_i$ as a discretization of non-symmetric continuous dual $q$-Hahn polynomials.
We remark that non-symmetric continuous dual $q$-Hahn polynomials were used to prove the faithfulness of a so-called basic representation on the space of Laurent polynomials in one variable for the algebra $\mathcal{H}_\mathrm{V}$; cf. \cite[Section 2]{2014Mazzocco}.
\section{Recurrence and orthogonality relations}\label{Section:rec orth relations}
We continue to discuss non-symmetric dual $q$-Hahn polynomials $\ell^\pm_i$, $0 \leq i \leq D-1$.
In this section, we derive combinatorial recurrence and orthogonality relations for $\ell^\pm_i$ from the $\mathcal{H}_V$-module $\mathbf{W}$.
We begin with a lemma.
\begin{lemma}\label{lem:min.poly(X)}
Recall $p^\perp$ and $h^\perp_{D-1}$ from \eqref{eq:p-perp,p-tilde,p-tildeperp} and \eqref{hperp(D-1)}, respectively.
Recall $\mathcal{X}$ from \eqref{eq:X}.
The following (i)-(iii) hold.
\begin{itemize}
\item[(i)] The element $p^\perp(\mathcal{X})h^\perp_{D-1}(\mathcal{X})$ vanishes on $\mathbf{W}$.
\item[(ii)] Let $\mu$ be the polynomial in $\mathbb{C}[\zeta]$ defined by
\begin{equation}\label{eq:min.poly(X)}
\mu(\zeta) := \zeta^Dp^\perp h^\perp_{D-1} = (\zeta-\tau)(\zeta-\tau^{-1}q^{-D})\prod^{D-1}_{i=1} (\zeta - \tau q^i)(\zeta-\tau^{-1} q^{-i}).
\end{equation}
Then $\mu$ is the minimal polynomial of $\mathcal{X}$ on $\mathbf{W}$.
\end{itemize}
\end{lemma}
\begin{proof}
(i) Let $\varepsilon\in \{+,-\}$. By Lemma \ref{lem:h-perp(D-1)vanish}, Proposition \ref{prop:C_i=ell_i(X)x} and \eqref{prop:eq:p-perp.h-perp=0}, we have
\begin{equation*}
p^\perp(\mathcal{X})h^\perp_{D-1}(\mathcal{X}).\hat{C}^\varepsilon_i = \ell^\varepsilon_i(\mathcal{X})h^\perp_{D-1}(\mathcal{X})p^\perp(\mathcal{X}).\hat{x} = 0, \qquad 0 \leq i \leq D-1.
\end{equation*}
(ii) Observe that $\mu$ has degree $2D$ and $\mu(\mathcal{X})=0$ on $\mathbf{W}$ by part (i). From Proposition \ref{prop:C_i=ell_i(X)x}, the result follows.
\end{proof}
Define the Laurent polynomials $\ell^\pm_{-1}$ and $\ell^\pm_D$ in $\mathbb{C}[\zeta, \zeta^{-1}]$ by $\ell^\pm_{-1}:=0$ and
\begin{align*}
& \ell^-_D :=\frac{1}{\tau^D(q;q)^2_D}p^\perp h^\perp_{D-1} = \frac{1}{\tau^D(q;q)^2_D}(\zeta^D+\cdots +q^{-D}\zeta^{-D}), \\
& \ell^+_D :=\frac{\tau q^{D+1}-\zeta^{-1}}{\tau^{D+1}(q;q)_D(q;q)_{D+1}}p^\perp h^\perp_{D-1} = \frac{\tau q^{D+1}-\zeta^{-1}}{\tau^{D+1}(q;q)_D(q;q)_{D+1}}(\zeta^D+\cdots +q^{-D}\zeta^{-D}).
\end{align*}
Observe that
\begin{equation*}
\ell^-_D \equiv \frac{\zeta^D}{\tau^D(q;q)^2_D} \pmod{\mathcal{L}},
\qquad
\ell^+_D \equiv \frac{\tau q^{D+1}\zeta^D - q^{-D}\zeta^{-D-1}}{\tau^{D+1}(q;q)_D(q;q)_{D+1}} \pmod{\mathcal{L}}.
\end{equation*}
Moreover, by Lemma \ref{lem:min.poly(X)}(i) $\ell^\pm_D(\mathcal{X})$ vanish on $\mathbf{W}$.
By these comments and using \eqref{eq(1):NonSymDqHahn}, \eqref{eq(2):NonSymDqHahn} at $i=D-1$, we routinely find
\begin{align}
\zeta \ell^-_{D-1}& \equiv \tau(1-q^D)^2 \ell^-_D \qquad \pmod{\mathcal{L}} \label{eq:zeta.ell-(D-1)} \\
\zeta \ell^+_{D-1} &\equiv \tau q^D(1-q^D) \ell^-_D \qquad \pmod{\mathcal{L}} \label{eq:zeta.ell+(D-1)}\\
\zeta^{-1}\ell^+_{D-1} &\equiv \tau q^{D+1}(q^D-1) \ell^-_D + \tau(q^D-1)(q^{D+1}-1)\ell^+_D \qquad \pmod{\mathcal{L}}. \label{eq:zeta-1ell+(D-1)}
\end{align}
We shall now give the recurrence relations for $\ell^{\pm}_i$, $0 \leq i \leq D-1$.
\begin{theorem}\label{thm:3rdResult}
The following (i), (ii) hold.
\begin{itemize}
\item[(i)] For $0 \leq i \leq D-1$, $\zeta \ell^-_i$ and $\zeta^{-1} \ell^-_i$ are respectively given as linear combination with the following terms and coefficients.
\begin{align*}
&\zeta \ell^-_i: && \zeta^{-1} \ell^-_i: \\
& \begin{tabular}{c|c}
term & coefficient \\
\hline \hline \\ [-0.8em]
$\ell^+_{i-1}$ & $\tau(q^i-q^D)(q^{N-D+1}-q^{i+1}+1) $\\[0.5em]
$\ell^-_i$ & $\tau q^i(q^{N-D+1}-q^{i+1}+1)$ \\[0.5em]
$\ell^+_i$ & $\tau (1-q^{i+1})(q^D-q^{i+1}+1)$ \\[0.5em]
$\ell^-_{i+1}$ & $\tau (q^{i+1}-1)^2$ \\[0.5em]
\end{tabular}, \quad
&&
\begin{tabular}{c|c}
term & coefficient \\
\hline \hline \\ [-0.8em]
$\ell^-_{i-1}$ & $\tau(1-q^{i-D})(q^{N+1}-q^{D+i}) $\\[0.5em]
$\ell^+_{i-1}$ & $\tau (q^D-q^{i})(q^{N-D+1}-q^i+1)$ \\[0.5em]
$\ell^-_{i}$ & $\tau q^{i+1}(q^D-q^i+1)$ \\[0.5em]
$\ell^+_{i}$ & $\tau (q^{i+1}-1)(q^D-q^i+1)$ \\[0.5em]
\end{tabular}.
\end{align*}
\item[(ii)] For $0 \leq i \leq D-1$, $\zeta \ell^+_i$ and $\zeta^{-1} \ell^+_i$ are respectively given as linear combination with the following terms and coefficients.
\begin{align*}
& \zeta\ell^+_i : && \zeta^{-1}\ell^+_i : \\
&\begin{tabular}{c|c}
term & coefficient \\
\hline \hline \\ [-0.8em]
$\ell^+_{i-1}$ & $\tau q^{N+1}(q^{i-D}-1)(q^{D-N+i}-1)$\\[0.5em]
$\ell^-_{i}$ & $\tau q^{N-D+1+i}(q^{D-N+i}-1)$ \\[0.5em]
$\ell^+_{i}$ & $\tau q^{i+1}(q^D-q^{i+1}+1)$ \\[0.5em]
$\ell^-_{i+1}$ & $\tau q^{i+1}(1-q^{i+1})$ \\[0.5em]
\end{tabular}, \quad
&&
\begin{tabular}{c|c}
term & coefficient \\
\hline \hline \\ [-0.8em]
$\ell^-_{i}$ & $\tau q^{N-D+2+i}(1-q^{D-N+i})$\\[0.5em]
$\ell^+_{i}$ & $\tau q^{i+1}(q^{N-D+1}-q^{i+1}+1)$ \\[0.5em]
$\ell^-_{i+1}$ & $\tau q^{i+2}(q^{i+1}-1)$ \\[0.5em]
$\ell^+_{i+1}$ & $\tau (q^{i+1}-1)(q^{i+2}-1)$ \\[0.5em]
\end{tabular}.
\end{align*}
\end{itemize}
\end{theorem}
\begin{proof}
By Remark \ref{rmk:degree.ell}, the Laurent polynomials $\zeta\ell^\pm_i$, $\zeta^{-1}\ell^\pm_i$ belong to $\mathcal{L}$ except $\zeta \ell^-_{D-1}$, $\zeta \ell^+_{D-1}$, and $\zeta^{-1}\ell^+_{D-1}$.
By Lemma \ref{lem:actionX} and Proposition \ref{prop:C_i=ell_i(X)x}, the Laurent polynomials $\zeta\ell^\pm_i$, $\zeta^{-1}\ell^\pm_i$ belonging to $\mathcal{L}$ are given as linear combination as shown in the above tables.
For the remaining three cases, use \eqref{eq:zeta.ell-(D-1)}--\eqref{eq:zeta-1ell+(D-1)}.
Then again, by Lemma \ref{lem:actionX} and Proposition \ref{prop:C_i=ell_i(X)x}, the desired result follows.
\end{proof}
We now discuss orthogonality relations for $\ell^\pm_i$.
We first find the eigenvalues of $\mathcal{X}$ on $\mathbf{W}$.
From \eqref{eq:min.poly(X)}, $\mu$ has $2D$ mutually distinct zeros
\begin{equation}\label{eq:eigvaluesXonW}
\lambda_i :=
\begin{cases}
\tau q^i, & i=0,1,\ldots, D-1, \\
\tau^{-1} q^{i}, & i = -1, -2, \ldots, -D,
\end{cases}
\end{equation}
and hence $\mathcal{X}$ is multiplicity-free on $\mathbf{W}$.
Next, we find eigenvectors of $\mathcal{X}$ corresponding to $\lambda_i$, $-D \leq i \leq D-1$.
Recall a $\Phi^*$-standard basis $\{E_i\hat{x}\}^D_{i=0}$ for $M\hat{x}$ and a $\Phi^{\perp*}$-standard basis $\{E_iu^\perp_0\}^{D-1}_{i=1}$ for $M\hat{x}^\perp$.
We consider the following ordered basis $\mathcal{B}$ for $\mathbf{W}$:
\begin{equation*}\label{eq:dual-basis.W}
\mathcal{B} = \{E_0\hat{x}, E_1\hat{x}, E_1{u}^\perp_0, E_2\hat{x}, E_2{u}^\perp_0, \ldots, E_{D-1}\hat{x}, E_{D-1}{u}^\perp_0, E_D\hat{x} \}.
\end{equation*}
Observe that $\mathcal{B}$ is orthogonal.
Recall the projection $\pi$ (resp. $\widetilde \pi$) from $\mathbf{W}$ onto $M\hat{x}$ (resp. $M\hat{C}$).
\begin{lemma}\label{lem:Mat.wtpi_B}
The following (i), (ii) hold.
\begin{itemize}
\item[(i)] The matrix representing $\pi$ with respect to $\mathcal{B}$ is
\begin{equation*}
\mathrm{blockdiag}\Big[ \pi(0), \pi(1), \ldots, \pi(D-1), \pi(D) \Big],
\end{equation*}
where $\pi(0) = \pi(D)=[1]$ and $\pi(i)=\mathrm{diag}(1,0)$ for $1 \leq i \leq D-1$.
\item[(ii)] The matrix representing $\widetilde \pi$ with respect to $\mathcal{B}$ is
\begin{equation*}
\mathrm{blockdiag}\Big[\widetilde\pi(0), \widetilde\pi(1), \ldots, \widetilde\pi(D-1), \widetilde\pi(D)\Big],
\end{equation*}
where $\widetilde\pi(0) = [1]$, $\widetilde\pi(D)=[0]$, and $\widetilde\pi(i)$, $1 \leq i \leq D-1$, is a $2\times 2$ matrix
\begin{equation*}
\begin{bmatrix}
\dfrac{q^i(q^{D-i}-1)(q^{N-D-i+1}-1)}{(q^D-1)(q^{N-D+1}-1)} & \dfrac{q^{i-1}(q^i-1)(q^{D-i}-1)(q^{N-i+1}-1)(q^{N-D-i+1}-1)}{(q-1)(q^D-1)(q^{N-D+1}-1)} \\[1em]
\dfrac{q(q-1)}{(q^{N-D+1}-1)(q^D-1)} & \dfrac{(q^i-1)(q^{N-i+1}-1)}{(q^{N-D+1}-1)(q^D-1)}
\end{bmatrix}.
\end{equation*}
\end{itemize}
\end{lemma}
\begin{proof}
(i) Since $\pi.E_i\hat{x}=E_i\hat{x}$, $0 \leq i \leq D$, and $\pi(M\hat{x}^\perp)=0$, the result follows. \\
(ii) Since $E_0\hat{x} \in M\hat{C}$ and $E_D\hat{x} \in M\hat{C}^\perp$, it follows that $\widetilde \pi.E_0\hat{x} = E_0\hat{x}$ and $\widetilde \pi.E_D\hat{x} = 0$.
Assume $1 \leq i \leq D-1$.
We now compute $\widetilde \pi.E_i\hat{x}$ and $\widetilde \pi E_i.u^\perp_0$.
Recall $\hat{C}_0 = \hat{C}^-_0 + \hat{C}^+_0$.
Eliminate $\hat{C}^+_0$ using \eqref{eq(1):C+-;Aix,upi} at $i=1$ to obtain
\begin{equation}\label{eq(1):pf:mat.rep.wt pi_B}
\hat{C}_0 = \hat{C}^-_0 + \frac{q-1}{q^D-1}A\hat{x} + \frac{q}{q^D-1}u^\perp_0.
\end{equation}
In this equation, eliminate $\hat{C}_0$ using \eqref{eq(1):C+-;Ci,tilde upi} at $i=0$ and solve the result for $\widetilde u^\perp_0$ to obtain
\begin{equation}\label{eq(2):pf:mat.rep.wt pi_B}
\widetilde u^\perp_0 = (q^{N-D}-1)\hat{C}^-_0 + \frac{(q-1)^2}{q(1-q^D)}A\hat{x} + \frac{1-q}{q^D-1}u^\perp_0.
\end{equation}
Apply $E_i$ to both sides of each equation of \eqref{eq(1):pf:mat.rep.wt pi_B}, \eqref{eq(2):pf:mat.rep.wt pi_B} and simplify the result using $E_iA=\theta_iE_i$.
Then, by recalling $\hat{x}=\hat{C}^-_0$ and $\hat{C}=\hat{C}_0$, we have
\begin{align}
& E_i\hat{C} = \frac{q^i(q^{D-i}-1)(q^{N-D-i+1}-1)}{(q-1)(q^D-1)}E_i\hat{x} + \frac{q}{q^D-1}E_iu^\perp_0, \label{eq(3):pf:mat.rep.wt pi_B}\\
& E_i\widetilde u^\perp_0 = \frac{(q^i-1)(q^{N-i+1}-1)}{q(q^D-1)}E_i\hat{x} + \frac{1-q}{q^D-1}E_iu^\perp_0. \label{eq(4):pf:mat.rep.wt pi_B}
\end{align}
Solving the system of equations \eqref{eq(3):pf:mat.rep.wt pi_B}, \eqref{eq(4):pf:mat.rep.wt pi_B} for $E_i\hat{x}$ and $E_i u^\perp_0$, we find
\begin{align}
& E_i\hat{x} = \frac{q-1}{q^{N-D+1}-1}E_i\hat{C} + \frac{q}{q^{N-D+1}-1}E_i\widetilde u^\perp_0, \label{eq(5):pf:mat.rep.wt pi_B}\\
& E_i u^\perp_0 = \frac{(q^i-1)(q^{N-i+1}-1)}{q(q^{N-D+1}-1)}E_i\hat{C} + \frac{q^D(q^{i-D}-1)(q^{N-D-i+1}-1)}{(q^{N-D+1}-1)(q-1)}E_i\widetilde u^\perp_0.\label{eq(6):pf:mat.rep.wt pi_B}
\end{align}
Apply $\widetilde \pi$ to both sides of each of equations \eqref{eq(5):pf:mat.rep.wt pi_B}, \eqref{eq(6):pf:mat.rep.wt pi_B} and eliminate $E_i\hat{C}$ using \eqref{eq(3):pf:mat.rep.wt pi_B}.
Simplify the result to obtain $\widetilde \pi.E_i\hat{x}$ and $\widetilde \pi.E_iu^\perp_0$, which are given by linear combinations of $E_i\hat{x}$ and $E_iu^\perp_0$.
The desired result follows.
\end{proof}
\noindent
In Theorem \ref{thm:2ndResult} we have shown how the projection $\pi$ (resp. $\widetilde \pi$) is related to $\mathcal{T}'$ (resp. $\mathcal{T}$) on $\mathbf{W}$.
Using this result and Lemma \ref{lem:Mat.wtpi_B} we obtain the following lemma.
\begin{lemma}\label{lem:MatTT'_B}
The following (i), (ii) hold.
\begin{itemize}
\item[(i)] The matrix representing $\mathcal{T}$ with respect to $\mathcal{B}$ is
\begin{equation*}
\mathrm{blockdiag}\Big[\mathcal{T}(0), \mathcal{T}(1), \ldots, \mathcal{T}(D-1), \mathcal{T}(D)\Big],
\end{equation*}
where $\mathcal{T}(0) = [k]$, $\mathcal{T}(D)=[-k^{-1}]$, and $\mathcal{T}(i)$, $1 \leq i \leq D-1$, is a $2\times 2$ matrix
\begin{equation*}
\begin{bmatrix}
\dfrac{k(q^{N+1}+q^D-q^i-q^{N-i+1})}{q^D-1} & \dfrac{-kq^{i-1}(q^i-1)(q^{D-i}-1)(q^{N-i+1}-1)(q^{N-D-i+1}-1)}{(q-1)(q^D-1)} \\[1em]
\dfrac{-kq(q-1)}{q^D-1} & \dfrac{k(q^{N-i+1}+q^i-q^{N-D+1}-1)}{q^D-1}
\end{bmatrix}.
\end{equation*}
\item[(ii)] The matrix representing $\mathcal{T'}$ with respect to $\mathcal{B}$ is
\begin{equation*}
\mathrm{blockdiag}\Big[\mathcal{T'}(0), \mathcal{T'}(1), \ldots, \mathcal{T'}(D-1), \mathcal{T'}(D)\Big],
\end{equation*}
where $\mathcal{T'}(0) = \mathcal{T'}(D)=[k']$ and $\mathcal{T'}(i)=\mathrm{diag}(k',-k'^{-1})$ for $1 \leq i \leq D-1$.
\end{itemize}
\end{lemma}
\begin{proof}
Use Theorem \ref{thm:2ndResult} and Lemma \ref{lem:Mat.wtpi_B}.
\end{proof}
\begin{lemma}\label{lem:MatX_B}
The matrix representing $\mathcal{X}$ with respect to $\mathcal{B}$ is
\begin{equation*}
\mathrm{blockdiag}\Big[ \mathcal{X}(0), \mathcal{X}(1), \ldots, \mathcal{X}(D-1), \mathcal{X}(D) \Big],
\end{equation*}
where $\mathcal{X}(0) = [\tau]$, $\mathcal{X}(D) = [\tau^{-1} q^{-D}]$, and $\mathcal{X}(i)$, $1 \leq i \leq D-1$, is a $2\times 2$ matrix
\begin{equation}\label{lem:eq:MatX(i)}
\begin{bmatrix}
\dfrac{\tau(q^{N+1}-q^i+q^D-q^{N+1-i})}{q^D-1} & \dfrac{\tau(q^i-1)(q^{D-i}-1)(q^{N-i+1}-1)(q^{i-1}-q^{N-D})}{(q-1)(q^D-1)} \\[1em]
\dfrac{\tau q^{D+1}(1-q)}{q^D-1} & \dfrac{\tau(q^{N+D+1-i}+q^{D+i}-q^{N+1}-q^D)}{q^D-1}
\end{bmatrix}.
\end{equation}
\end{lemma}
\begin{proof}
Recall $\mathcal{X}=\mathcal{T}'\mathcal{T}$.
Use this and Lemma \ref{lem:MatTT'_B}. The result routinely follows.
\end{proof}
Note that for each $1\leq i \leq D-1$ the matrix $\mathcal{X}(i)$ of \eqref{lem:eq:MatX(i)} has the eigenvalues $\lambda_{i}$ and $\lambda_{-i}$; cf \eqref{eq:eigvaluesXonW}.
Now we find eigenvectors of $\mathcal{X}$ associated with $\lambda_i$, $-D \leq i \leq D-1$.
Define
\begin{align}
& \mathbf{y}_i := \frac{(q^{D-i}-1)(q^{N-i+1}-1)}{(q^D-1)(q^{N-2i+1}-1)} E_i\hat{x} + \frac{q^{D+1-i}(q-1)}{(q^D-1)(q^{N-2i+1}-1)}E_i u^\perp_0, \label{eq(1):eigvecX}\\
& \mathbf{y}_{-i} := \frac{q^{D-i}(q^i-1)(q^{N-D-i+1}-1)}{(q^D-1)(q^{N-2i+1}-1)} E_i\hat{x} - \frac{q^{D-i+1}(q-1)}{(q^D-1)(q^{N-2i+1}-1)}E_iu^\perp_0, \label{eq(2):eigvecX}
\end{align}
for $1 \leq i \leq D-1$. Moreover, define
\begin{equation}\label{eq(3):eigvecX}
\mathbf{y}_0 := E_0\hat{x}, \qquad \qquad \mathbf{y}_{-D} := E_D\hat{x}.
\end{equation}
Observe that (i) the vectors $\mathbf{y}_i$, $-D \leq i \leq D-1$, are real; (ii) $\mathbf{y}_i + \mathbf{y}_{-i} = E_i\hat{x}$, $1 \leq i \leq D-1$, so that
\begin{equation}\label{eq:sum(yi)=x}
\sum^{D-1}_{i=-D} \mathbf{y}_i = \mathbf{y}_0 + \sum^{D-1}_{i=1} (\mathbf{y}_i + \mathbf{y}_{-i}) + \mathbf{y}_{-D} = \sum^D_{i=0}E_i\hat{x} = \hat{x}.
\end{equation}
\begin{proposition}\label{prop:eigvectorX}
Let $-D \leq i \leq D-1$.
On $\mathbf{W}$, the vector $\mathbf{y}_i$ is an eigenvector of $\mathcal{X}$ associated with the eigenvalue $\lambda_i$.
Moreover, the vectors $\mathbf{y}_i$ form an eigenbasis of $\mathcal{X}$ for $\mathbf{W}$.
\end{proposition}
\begin{proof}
The first assertion routinely follows from \eqref{eq(1):eigvecX}, \eqref{eq(2):eigvecX}, \eqref{eq(3):eigvecX} and Lemma \ref{lem:MatX_B}.
The second assertion immediately follows from that $\mathcal{B}$ is an basis for $\mathbf{W}$.
\end{proof}
Recall the Hermitian inner product $\langle \cdot, \cdot \rangle$ defined on $\mathbb{C}^X$.
Since the basis $\mathcal{B}$ is orthogonal on $\mathbf{W}$, we observe that
\begin{equation}\label{iff:orthogonality}
\langle \mathbf{y}_i, \mathbf{y}_j \rangle \ne 0 \quad \text{ if and only if } \quad j \in \{ i, -i\}, \qquad -D\leq i,j \leq D-1.
\end{equation}
We compute explicitly the non-zero inner products.
We first compute $\lVert E_i\hat{x}\rVert^2$, $0 \leq i \leq D$, and $\lVert E_i u^\perp_0 \rVert^2$, $1 \leq i \leq D-1$.
For $0 \leq i \leq D$, let $m_i$ denote the scalar as in \eqref{eq:scalar m(i)} associated with $\Phi$. Using \eqref{m_formula} and the parameter sequence of $\Phi$ in Proposition \ref{prop:LSonMx,Mxp}(I), we routinely find
\begin{equation}\label{eq:m;q-formula}
m_i = \frac{q^i(1-q^{N-2i+1})(q^{i+1};q)_{D-i}}{(q^{N-D+1};q)_{D-i+1}}.
\end{equation}
For $0 \leq i \leq D-2$, let $m^\perp_i$ denote the scalar as in \eqref{eq:scalar m(i)} associated with $\Phi^\perp$.
Using \eqref{m_formula} and the parameter sequence of $\Phi$ in Proposition \ref{prop:LSonMx,Mxp}(II), we routinely find
\begin{equation}\label{eq:mperp;q-formula}
m^\perp_i = \frac{q^{N-1}(1-q^{N-D-i})(1-q^{D-i-1})(1-q^{2i-N+1})(q^{i+1};q)_{D-2-i}}{(q-1)(q^{N-D};q)_{D-i}}.
\end{equation}
\begin{lemma}\label{lem:norm.Eix} Both
\begin{itemize}
\item[(i)] $\lVert E_i\hat{x} \rVert^2 = \dfrac{q^i(1-q^{N-2i+1})(q^{i+1};q)_{D-i}}{(q^{N-D+1};q)_{D-i+1}}, \quad 0 \leq i \leq D,$
\item[(ii)] $\lVert E_iu^\perp_0 \rVert^2 = q^{2N-D-i}\dfrac{(1-q^{D-N+i-1})(1-q^{D-i})(1-q^{2i-N-1})(q^{i};q)_{D+1-i}}{(q-1)^2(q^{N-D+1};q)_{D-i}}, \quad 1 \leq i \leq D-1.$
\end{itemize}
\end{lemma}
\begin{proof}
(i) Since $E^*_0E_iE^*_0 = m_iE^*_0$, we have $\lVert E_i\hat{x} \rVert^2=\lVert E_iE^*_0\hat{x} \rVert^2= \langle \hat{x}, E^*_0E_iE^*_0\hat{x} \rangle=m_i$.
By this and \eqref{eq:m;q-formula}, the result follows.\\
(ii) Similarly to (i), we have $\lVert E_iu^\perp_0\rVert^2 = m^\perp_{i-1}\lVert u^\perp_0\rVert^2$, where $\lVert u^\perp_0\rVert^2 = \dfrac{(q^D-1)(q^{D-1}-1)(q^{n-D}-1)}{q-1}$ by \eqref{prop:eq:u-perp(i)} and Lemma \ref{cardinality:C+-}.
Using this and \eqref{eq:mperp;q-formula}, the result routinely follows.
\end{proof}
\begin{lemma}
Recall the eigenvectors $\mathbf{y}_i$, $-D \leq i \leq D-1$, for $\mathcal{X}$ on $\mathbf{W}$ from \eqref{eq(1):eigvecX}--\eqref{eq(3):eigvecX}.
For $1 \leq i \leq D-1$, we have
\begin{align}
& \lVert \mathbf{y}_i \rVert^2 = \frac{ (q^{D-i}-1)(q^{N+1}+q^D-q^{D+i}-q^i)(q^i;q)_{D-i} }{(q^i-1)(q^{N-2i+1}-1)(q^{N-D+1};q)_{D-i}}, \label{lem:eq(1):norm}\\
& \lVert \mathbf{y}_{-i} \rVert^2 = \frac{q^{D+N-2i+1}(q^{N-D-i+1}-1)(q^D-q^{D-N+i-1}-q^i+1)(q^i;q)_{D-i}}{(q^{N+1-i}-1)(q^{N+1-2i}-1)(q^{N-D+1};q)_{D-i}}, \label{lem:eq(2):norm}\\
& \langle \mathbf{y}_i, \mathbf{y}_{-i} \rangle = \frac{q^D(1-q^{D-i})(1-q^{N-D-i+1})(q^i;q)_{D-i}}{(1-q^{N-2i+1})(q^{N-D+1};q)_{D-i}}.\label{lem:eq(3):norm}
\end{align}
Moreover,
\begin{equation}\label{lem:eq(4):norm}
\lVert \mathbf{y}_0 \rVert^2 = \frac{(q;q)_D}{(q^{N-D+1};q)_D}, \qquad \qquad
\lVert \mathbf{y}_{-D} \rVert^2 = \frac{q^D(q^{N-2D+1}-1)}{q^{N-D+1}-1}.
\end{equation}
\end{lemma}
\begin{proof}
Evaluate $\lVert \mathbf{y}_i \rVert^2$ using \eqref{eq(1):eigvecX} and Lemma \ref{lem:norm.Eix}(i) and simplify the result to get \eqref{lem:eq(1):norm}.
Similarly, we obtain \eqref{lem:eq(2):norm}, \eqref{lem:eq(3):norm} using \eqref{eq(1):eigvecX}, \eqref{eq(2):eigvecX}.
Line \eqref{lem:eq(4):norm} follows from Lemma \ref{lem:norm.Eix}(i) at $i=0,D$.
\end{proof}
Recall the space $\mathcal{L}$ from \eqref{eq:space L} and let $f \in \mathcal{L}$.
By \eqref{eq:sum(yi)=x} and Proposition \ref{prop:eigvectorX} the action $f(\mathcal{X}).\hat{x}$ on $\mathbf{W}$ is given as
\begin{equation*}
f(\mathcal{X}).\hat{x} = f(\mathcal{X}).\sum^{D-1}_{i=-D}\mathbf{y}_i = \sum^{D-1}_{i=-D}f(\lambda_i)\mathbf{y}_i.
\end{equation*}
Using this and \eqref{iff:orthogonality}, we find that for $f, g \in \mathcal{L}$,
\begin{equation*}\label{eq:innerprod_W}
\langle f(\mathcal{X}).\hat{x}, g(\mathcal{X}).\hat{x} \rangle
= \sum^{D-1}_{i=-D} f(\lambda_i) \overline{g(\lambda_i)} \lVert \mathbf{y}_i \rVert^2 + \sum^{D-1}_{i=1} \Big( f(\lambda_i)\overline{g(\lambda_{-i})} + f(\lambda_{-i})\overline{g(\lambda_i)} \Big) \langle \mathbf{y}_i, \mathbf{y}_{-i} \rangle.
\end{equation*}
In particular, for basis elements $\ell^\pm_i, \ell^\pm_j$ of $\mathcal{L}$, by using Proposition \ref{prop:C_i=ell_i(X)x}
\begin{equation}\label{eq:innerprod,ell,W}
\langle \ell^\sigma_i (\mathcal{X}).\hat{x}, \ell^\nu_j(\mathcal{X}).\hat{x} \rangle
= \langle \hat{C}^\sigma_i, \hat{C}^\nu_j \rangle
= \delta_{i,j}\delta_{\sigma, \nu} |C^\sigma_i|,
\end{equation}
where $0 \leq i, j \leq D-1$ and $\sigma, \nu \in \{+,-\}$.
Motivated by these comments, we define the Hermitian form $\langle \cdot, \cdot \rangle_\mathcal{L}$ on $\mathcal{L}$ by
\begin{equation}\label{eq:HermitianFormL}
\langle f, g \rangle_{\mathcal{L}} := \sum^{D-1}_{i=-D} f(\lambda_i) \overline{g(\lambda_i)} \omega_i + \sum^{D-1}_{i=1} \Big( f(\lambda_i)\overline{g(\lambda_{-i})} + f(\lambda_{-i})\overline{g(\lambda_i)} \Big) \omega^\vee_i,
\end{equation}
where $f,g \in \mathcal{L}$ and the $\lambda_i$ are from \eqref{eq:eigvaluesXonW} and the non-zero (real) scalars $\omega_i$, $\omega^\vee_j$ are given by
\begin{equation*}
\omega_i := \lVert \mathbf{y}_i \rVert^2, \quad -D \leq i \leq D-1, \qquad \omega^\vee_i := \langle \mathbf{y}_i, \mathbf{y}_{-i} \rangle, \quad 1 \leq i \leq D-1.
\end{equation*}
We shall now give the orthogonality relations for $\ell^\pm_i$, $0 \leq i \leq D-1$.
\begin{theorem}\label{thm:4thResult}
Let $\langle \cdot , \cdot \rangle_{\mathcal{L}}$ be the Hermitian form as in \eqref{eq:HermitianFormL}.
For $0 \leq i, j \leq D-1$ and $\sigma, \nu \in \{+,-\}$, we have
\begin{align*}
\langle \ell^\sigma_i, \ell^\nu_j \rangle_{\mathcal{L}}
& = \delta_{i,j}\delta_{\sigma, \nu}|C^\sigma_i| \\
& =
\begin{cases}
\delta_{i,j}\delta_{\sigma, \nu} q^{i(i+1)}\displaystyle\prod^i_{h=1} \dfrac{(q^{D-h}-1)(q^{N-D+1-h}-1)}{(q^h-1)^2}, & \text{ if }\sigma=-,\\
\delta_{i,j}\delta_{\sigma, \nu} \dfrac{q^{(i+1)^2}(q^{N-D}-1)}{q-1}\displaystyle\prod^i_{h=1} \dfrac{(q^{D-h}-1)(q^{N-D-h}-1)}{(q^h-1)(q^{h+1}-1)}, & \text{ if } \sigma=+.
\end{cases}
\end{align*}
\end{theorem}
\begin{proof}
From \eqref{eq:innerprod,ell,W} and Lemma \ref{cardinality:C+-}, the result follows.
\end{proof}
\section*{Acknowledgement}
The author thanks to Hajime Tanaka and Paul Terwilliger for giving this paper a close reading and offering valuable suggestions.
The author also thanks the anonymous referee for careful reading and helpful comments.
|
\section{Introduction}
Named entity recognition(NER) is a subtask of sequence labeling. It is similar to other sequence labeling tasks considering its working procedure that is assigning
a certain label to each token of a sequence. But unlike part-of-speech(POS) tagging and other sequence labeling task evaluated on accuracy, the performance of
an NER system is evaluated on the whole named entity using precision, recall, and f1 score. Thus, the output of an NER system at each position is not independent
with each other and much related to its neighboring positions. The most common assumption of sequence labeling is the Markov property that the choice of label for a particular token is directly dependent only on the immediately adjacent labels; hence the labels of all the tokens in a sequence form a Markov chain.
Therefore, the widely-used statistical models for sequence labeling involve hidden Markov model (HMM) \cite{rabiner1986hmm}, maximum entropy Markov model (MEMM) \cite{DBLP:conf/icml/McCallumFP00} and conditional random field (CRF) \cite{lafferty2001conditional}.
In recent years, neural network architectures for NER \cite{huang2015bidirectional,P16-1101,D17-1283}
are proposed to reduce the efforts of feature engineering and the model complexity and have achieved great success. Currently, the dominative neural NER architecture consists of a bi-directional recurrent neural network (RNN) as the encoder and a conditional random field (CRF) as the decoder \cite{huang2015bidirectional}. The RNN encoder can effectively extract the context-aware features for each token, avoiding the cost of manually designing features.
Despite of their success, the network architecture of the encoder still need be manually designed for different tasks and lacks flexibility. Due to the inherent hierarchical structure of natural language, the crucial information for a token could appear in a changeable position. Taking the following NER instance for example, there are two entity mentions in the following sentence,
\begin{center}
\textit{He bought 30 shares of \underline{Acme} in \underline{2006}}.
\end{center}%
For the token ``\textit{2006}'', its self-information is enough to decide its entity type. But for the token ``\textit{Acme}'', the information from its neighbour ``\textit{shares}'' could be more important. Since the position of the crucial information for each token is different, the current rigid network architecture heavily depends on the ability of RNNs to capture the context information.
In this paper, we propose a deformable stacked structure to flexibly choose the most informative features as input. Unlike the vanilla stacked structure, the connections between two adjacent layers are dynamically constructed. The input of each position in the upper layer is dynamically chosen from the lower layer, instead of a fixed position. Specifically, we introduce a dynamical offset to indicate the input's position. The offsets are dynamically computed according to the current hidden states. To make the whole neural network end-to-end trainable, we further propose an approximate solution to use a continuous offset to softly select the inputs via a bilinear interpolation, instead of the exact discrete offset.
The contributions of the paper can be summarized as follows.
\begin{enumerate}
\item We propose a deformable stacked structure, whose stacking connections are dynamically determined, instead of in a pre-defined way. The deformable stacked way can effectively alleviate the pressure of RNNs for collecting the context information.
\item We also propose an approximate strategy to softly change the connections, which makes the whole neural network differentiable and end-to-end trainable.
\item Compared to the models with rigid network architecture, our model is more flexible and suitable for named entity recognition tasks and achieves the state-of-the-art performances for named entity recognition on OntoNotes dataset.
\end{enumerate}
\section{General Neural Architecture for Named Entity Recognition}
Given a sequence with $n$ tokens $X = \{x_1, \dots, x_n\}$, the aim of named entity recognition is to figure out the ground truth of labels $Y^* = \{y_1^*, \dots, y_n^*\}$:
\begin{equation}
Y^* = \argmax_{Y \in \mathcal{T}^n} p (Y | X), \label{eq:argmax}
\end{equation}
where $\mathcal{T}$ is the target set.
There are lots of prevalent methods to solve named entity recognition problem such as maximum entropy Markov model (MEMM), conditional random fields (CRF), etc.
Recently, neural models are widely applied to named entity recognition for their ability to minimize the effort in feature engineering
\cite{huang2015bidirectional,P16-1101}. Moreover, neural models also benefit from the distributed representations, which can enhance the generalization capabilities with the pre-trained word embeddings on the large-scale un-annotated corpus.
The general architecture of neural named entity recognition could be characterized by three components: (1) an embedding layer; (2) encoding layers consisting of several classical neural networks and (3) a decoding layer. The role of encoding layers is to extract features, which could be either convolution neural network or recurrent neural network. In this paper, we adopt the bidirectional long-short-term-memory (BiLSTM) neural networks followed by a CRF as decoding layer. Figure \ref{fig:arch} illustrates the general architecture.
\begin{figure}[t]
\centering
\hspace{2em}
\includegraphics[width=0.45\textwidth]{arch}
\caption{General neural architecture for named entity recognition. }\label{fig:arch}
\end{figure}
\subsection{Embedding Layer}
To represent discrete tokens as distributed vectors, the first step is usually to map them to distributed embedding vectors. Formally, we lookup embedding vector from embedding matrix for each token $x_i$ as $\mathbf{e}_{x_i} \in \mathbb{R}^{d_e}$, where $d_e$ is a hyper-parameter indicating the size of embedding.
\subsection{Encoding Layers}
To incorporate information from both sides of sequence, we use bi-directional LSTM with forward and backward directions.
Notably, the parameters of two LSTMs with different orientations are independent.
The update of each hidden state can be written precisely as follows:
\begin{align}
\textbf{h}_i &= \overrightarrow{\textbf{h}}_i \oplus {\overleftarrow{\textbf{h}}_i},\\
&= \text{BiLSTM}(\mathbf{e}_{x_{1:n}}, i, \theta),
\end{align}
where $\overrightarrow{\textbf{h}}_i$ and $\overleftarrow{\textbf{h}}_i$ are the hidden states at position $i$ of the forward and backward LSTMs respectively; $\oplus$ is concatenation operation; $\theta$ denotes all the parameters in BiLSTM model.
\subsection{Decoding Layer}
After extracting features, we employ a conditional random fields (CRF) layer to inference tags. In CRF layer, $p (Y | X)$ in Eq (\ref{eq:argmax}) could be formalized as:
\begin{equation}
p (Y | X) = \frac{\Psi (Y | X)}{\sum_{Y^\prime \in \mathcal{T}^n} \Psi (Y^\prime | X)}.
\end{equation}
Here, $\Psi (Y | X)$ is the potential function, and we only consider interactions between two successive labels (first order linear chain CRFs):
\begin{gather}
\Psi (Y | X) = \prod_{i = 2}^n \psi (X, i, y_{i-1}, y_i),\\
\psi (\mathbf{x}, i, y^\prime, y) = \exp(s(X, i)_{y} + b_{y^\prime y}),
\end{gather}
where $b_{y^\prime y} \in \mathbb{R}$ is transition parameter, indicating how possible a label $y^\prime$ will transfer to another label $y$. Score function $s(X, i) \in \mathbb{R}^{|\mathcal{T}|}$ assigns score for each label on tagging the $i$-th character:
\begin{equation}
s(X, i) = \mathbf{W}_s^\top \mathbf{h}_i + \mathbf{b}_s,
\end{equation}
where $\mathbf{h}_i$ is the hidden state of BiLSTM at position $i$; $\mathbf{W}_s \in \mathbb{R}^{d_h \times |\mathcal{T}|}$ and $\mathbf{b}_s \in \mathbb{R}^{|\mathcal{T}|}$ are trainable parameters.
At test phase, the Viterbi algorithm is employed to decode the best target sequence in polynomial time complexity.
\begin{figure}[t!]
\centering
\subfloat[Vanilla stacked structure]{
\includegraphics[width=0.5\textwidth]{rnn-stacked-unfolded} \label{fig:stacked-rnn}
}
\\
\subfloat[Deformable stacked structure]{
\includegraphics[width=0.5\textwidth]{rnn-stacked-unfolded-deform} \label{fig:Model-III}
}
\caption{Two ways to stack layers. The dashed lines denote the connections are dynamically changed.}\label{fig:arch-comparison}
\end{figure}
\section{Deformable Stacked Structure}
The critical factor of neural named entity recognition models is the encoding layer, whose role is to extract useful features to judge the label of each token. To better model the complex compositional features, we could increase the depth of neural network by stacking the recurrent encoding layers.
\paragraph{Vanilla Stacked Structure}
A conventional way to stack layers is to take the output of the lower layer as the input of upper layer at each position \cite{pascanu2013construct}.
For each position $i$ at $(l+1)$-th layer, its input is taken from position $i$ at $l$-th layer.
\begin{align}
\textbf{h}^{(l+1)}_i &= f(\textbf{h}^{(l)}_i, \theta^{(l+1)})
\end{align}
where $f$ is a non-linear function.
Despite being successful for NER, the vanilla stacked structure has a limitation of the fixed geometric structures. The stacking structures are manually designed and lack flexibility.
\paragraph{Deformable Stacked Structure}
Due to the inherently hierarchical structure of natural language, the crucial information for a token could appear in a changeable position. As usually seen in NER, while in some case the information of a token itself is enough to decide its entity type, in another case the information of its neighbors could be more important.
To flexibly capture the most informative features, we propose a deformable stacked structure to choose the input's position from the lower layers dynamically.
For each position $i$ at $(l+1)$-th layer, its input is taken from position $i+o$ at $(l)$-th layer.
\begin{align}
\textbf{h}^{(l+1)}_i &= f(\textbf{h}^{(l)}_{i+o}, \theta^{(l+1)})
\end{align}
where $o$ is a offset, $o \in \mathbb{Z}$.
The offsets $o$ is predicted by an extra module based on the hidden states of the lower layers.
We can adapt deformable stacked structure between different layers, which we will further explain in the experiment section.
Figure \ref{fig:arch-comparison} gives the comparison of two different stacked structure.
\section{Differentiable Deformable Stacked Structure}
In our proposed deformable stacked structure, the index of the lower layer is discrete, which results in a non-differentiable network. Although we can use reinforcement learning to learn the parameters, in this paper, we propose a differentiable variant to utilize the strengths of back-prorogation of the neural network.
Inspired by \cite{dai2017deformable}, we use bilinear interpolation to replace the exact discrete index.
Figure \ref{fig:rnn-stacked-unfolded-deform-imp} shows the
architecture of deformable stack.
\begin{figure}[t]
\centering
\hspace{2em}
\includegraphics[width=0.48\textwidth]{rnn-stacked-unfolded-deform-imp}
\caption{A differentiable implementation of deformable stacked structure. $f(\cdot)$ is defined by Eq (\ref{eq:offset}), $\oslash$ denotes the bilinear interpolation. }\label{fig:rnn-stacked-unfolded-deform-imp}
\end{figure}
\paragraph{Offset Learning}
Instead of a discrete offset, we use a continuous offset to make the whole network to be differentiable.
The offset is calculated by a simple function. The offset at the position $i$ can be given as:
\begin{equation}
o_i^{(l)} = \mathbf{v} ^\mathrm{\scriptscriptstyle T} \textbf{h}^{(l)}_i, \label{eq:offset}
\end{equation}
where $\mathbf{v} \in \mathbb{R} ^{d}$ is the parameter vector, $\textbf{h}^{(l)}_i \in \mathbb{R}^d$ is the hidden states at position $i$ of the $(l)$-th layer.
\paragraph{Deformable Input}
To bridge the continuous offset and discrete position, we apply bilinear interpolation to select the inputs from the lower layer softly.
For each position $i$ at $(l+1)$-th layer, its input $\mathbf{z}^{(l+1)}_i$ can be given as:
\begin{equation}
\mathbf{z}^{(l+1)}_i = \sum_{j = 1}^{n} g ( i + o_i^{(l)} , j ) \cdot \textbf{h}^{(l)}_j
\end{equation}
where $g( i + o_i^{(l)} , j )$ is a bilinear interpolation kernel and can be given as:
\begin{equation}
g ( i + o_i^{(l)} , j ) = \max ( 0, 1 - |
i + o_i^{(l)} - j | ),
\end{equation}
thus $\{g(i + o_i^{(l)} , j)\}_{j=1}^n$ can be regarded as a mask vector which has two non-zero elements at most.
For example, assuming that the length of input sequence is 5, and the current position is 2, a continuous offset $o=1.2$ gives the mask vector
\begin{align}
\mathbf{g} = [0,0,0.8,0.2,0].
\end{align}
By the above strategy, we can make the whole neural network to be differentiable, which can be end-to-end trained efficiently.
\subsection{Multi-Offset Extension}
To make the input of each recurrent layer more flexible, we can allow it to choose information from the multiple positions in the lower layer.
Therefore, a simple extension of our model is the multi-offset deformable structure, which allows the model to utilize the information at different positions jointly.
In multi-offset extension, we can predict $k$ offsets at position $i$,
\begin{equation}
\mathbf{o}^{(l)}_i = V \textbf{h}^{(l)}_i,
\end{equation}
where $\mathbf{o}^{(l)}_i \in \mathbb{R}^k$ is a vector consisting of $k$ offsets, and $V \in \mathbb{R} ^{k \times d}$ is the parameter matrix.
With $k$ offsets, we can obtain $k$ deformable inputs $\mathbf{z}^{(l+1)}_{i1}, \mathbf{z}^{(l+1)}_{i2}, \cdots ,\mathbf{z}^{(l+1)}_{ik}$ via bilinear interpolation. Then we concatenate these inputs
to get the final input of position $i$ to the next layer.
\begin{equation}
\mathbf{z}^{(l+1)}_i = \mathbf{z}^{(l+1)}_{i1} \oplus \mathbf{z}^{(l+1)}_{i2} \oplus \cdots \oplus \mathbf{z}^{(l+1)}_{ik},
\end{equation}
where $\oplus$ indicates the concatenation operation.
\subsection{Wide-Window Extension}
Another extension of our model is to involve neighbor hidden states when calculating the offsets.
In wide-window extension, we can predict $k$ offsets at position $i$ with window size $w$ by a convolutional neural network,
\begin{equation}
\mathbf{o}^{(l)}_i = W \textbf{h}^{(l)}_{i:i+d}
\end{equation}
\begin{equation}
\mathbf{O}^{(l)} = \mathrm{Conv}(\mathbf{H}, W, w),
\end{equation}
where $\mathbf{O} \in \mathbb{R} ^{n \times k}$ is a matrix consisting of $k$ offsets in $n$ positions. $W \in \mathbb{R} ^{w \times k \times d}$ is the parameter matrix of the convolutional neural network.
\section{Training}
Given a trainset $(X^{(n)}, Y^{(n)})_{n = 1}^N$, the objective is to minimize the cross entropy loss $\mathcal{L}(\theta)$:
\begin{equation}
\mathcal{L}(\theta) = \frac{1}{N}\sum_n \log p(Y^{(n)}|X^{(n)}) + \lambda \| \theta \| ^2,
\end{equation}
where $\theta$ represents all the parameters, $\lambda$ represents the regularizer factor.
We use stochastic gradient descent with a momentum of 0.9.
The initial learning rate is set according to the dataset and task.
To avoid overfitting, dropout is applied after each recurrent or convolutional layer.
\paragraph{Initialization}
We take advantage of pre-trained word embeddings such as Glove \cite{pennington2014glove} to transfer more knowledge from large unlabeled data.
For the words that don't appear in Glove, we randomly initialize their embeddings from a normal distribution with mean 0 and standard deviation $\sqrt{\frac{1}{dim}}$ following \cite{P16-1101}, $dim$ is the dimension of word embedding.
The network weights are initialized with Xavier normalization \cite{glorot2010understanding} to maintain the variance of activations throughout the forward and backward passes. Biases are uniformly set to zero when the network is constructed.
\paragraph{Character Embedding}
Following \cite{P16-1101}, we also apply convolutional neural network (CNN) to extract character-level features of words.
The character-level feature of words helps the model better handle the OOV (out of vocabulary)
problems. For each character, we randomly initialize its embedding from a normal distribution
with mean 0 and standard deviation 1.
\section{Experiment}
We consider three different kinds of deformable stacked structure in our paper.
\begin{enumerate}
\item Deformable stacked structure between BiLSTM layers.
\item Deformable stacked structure between the encoder layer (BiLSTM in our paper) and decoder layer (CRF in our paper).
\item Both 1 and 2.
\end{enumerate}
\begin{table}[t]
\setlength{\tabcolsep}{2pt}
\centering
\begin{tabular}{lccc}
\toprule
\textbf{Dataset} & \textbf{Train} & \textbf{Dev} & \textbf{Test} \\
\midrule
{CoNLL-2003} & 204,567 & 51,578 & 46,666 \\
\midrule
{OntoNotes 5.0} & \multirow{2}*{1,088,503} & \multirow{2}*{147,724} & \multirow{2}*{152,728} \\
{(CoNLL-2012)} \\
\bottomrule
\end{tabular}
\caption{Number of tokens in different dataset.}
\label{tab:dataset-detail}
\end{table}
\subsection{Datasets}
We evaluate our model on two datasets: CoNLL-2003 NER dataset \cite{tjongkimsang2003conll} dataset and
OntoNotes 5.0 \cite{W13-3516} dataset. We adapt structure 2 on CoNLL-2003 dataset and structure 1, 2, 3 on OntoNotes 5.0 dataset.
For both datasets, we perform the experiments on the English
portion.
For the OntoNotes 5.0 dataset, we split the
data according to the CoNLL-2012 shared task following \cite{Q16-1026}.
The Pivot Text portion is excluded because it lacks gold annotations for named entities.
The original tags are converted to tags of ``BIOES'' (begin, inside, outside, end, singleton) tagging.
The CoNLL-2003 dataset contains only 4 types of named entities: PERSON, LOCATION, ORGANIZATION, AND MISCELLANEOUS, while the OntoNotes 5.0
dataset contains 18 types of named entities, including works of art,
dates, cardinal numbers, languages, and events.
With BIOES tagging, we have 18 labels for
CoNLL-2003 dataset and 74 labels for OntoNotes 5.0 dataset(including padding label).
For the digits in both of the datasets, we replace them with digit 0.
The details of the two datasets are shown in Table \ref{tab:dataset-detail}.
\subsection{Hyper-parameters}
Hyper-parameters of our models are shown in Table \ref{tab:hyper-parameters-ner}.
For the word embedding, we use Glove pre-trained embedding of 100 dimensions on 6 billion words.
\subsection{Result}
\begin{table}[t!]
\centering
\setlength{\tabcolsep}{4pt}
\begin{tabular}{lcc}
\toprule
& \textbf{CoNLL} & \textbf{OntoNotes} \\
& \textbf{2003} & \textbf{5.0} \\
\midrule
Offset CNN window size & 3 & 3\\
Number of offsets $k$ & 3 & 3\\
Char-CNN filters & 30 & 30 \\
Char-CNN window size & 3 & 3 \\
Size of LSTM state & 256 & 200 \\
LSTM layers & 1 & 2 \\
Learning rate & 0.008 & 0.005 \\
Dropout & 0.5 & 0.5 \\
Batch size & 10 & 8 \\
\bottomrule
\end{tabular}
\caption{Hyper-parameters.}
\label{tab:hyper-parameters-ner}
\end{table}
\begin{table}[t!]\setlength{\tabcolsep}{3pt}
\centering
\begin{tabular}{lccc}
\toprule
\textbf{Model} & \textbf{P} & \textbf{R} & \textbf{F} \\
\midrule
\cite{N09-1037}$^{\star}$ & 84.04 & 80.86 & 82.42 \\
\cite{W09-1119} & 82.00 & 84.95 & 83.45 \\
\cite{Q14-1037} & 85.22 & 82.89 & 84.04 \\
\cite{chiu2016sequential} & 86.16 & 86.65 & 86.40 \\
\cite{D17-1283} & - & - & 86.84 \\
\cite{D17-1283}$^{\star\star}$ & - & - & 86.99 \\
\midrule
BiLSTM-CNN-CRF & 87.03 & 86.97 & 87.00\\
\midrule
Deformable stacked structure & \textbf{88.02}& \textbf{88.01} & \textbf{88.01} \\
\bottomrule
\end{tabular}
\caption{Performances on OntoNotes 5.0 dataset. $^{\star}$ denotes the result from \cite{W13-3516}. $^{\star\star}$ denotes the baseline from their paper. \cite{N09-1037}: joint parsing
and NER model. \cite{W09-1119}: using many resources, such as Wikipedia, non-local features. \cite{Q14-1037}: combining coreference resolution, entity linking, and NER into a
single CRF model with cross-task interaction factors. \cite{chiu2016sequential}: BiLSTM-CRF network with with many composite features. \cite{D17-1283}: iterated dilated CNN and CRF.}
\label{tab:compare-ner-ontonotes}
\end{table}
\begin{table}[th]\setlength{\tabcolsep}{3pt}
\centering
\begin{tabular}{lccc}
\toprule
\textbf{Model} & \textbf{P} & \textbf{R} & \textbf{F} \\
\midrule
\cite{collobert2011natural} & - &- &86.96\\
\cite{D15-1104} & - & - & 91.20 \\
\cite{N16-1030} & - &- & 90.33 \\
\cite{P16-1101} & 91.35 & 91.06 & 91.21 \\
\cite{D17-1283} & - & - & 90.54 \\
\midrule
Bi-LSTM-CNN-CRF & {91.03} & {91.11} & {91.07} \\
\midrule
Deformable stacked structure & 91.01 & 91.24 & 91.12 \\
\bottomrule
\end{tabular}
\caption{Performances on CoNLL-2003 dataset.
\cite{collobert2011natural}: a earlier neural model. \cite{D15-1104}:
joint NER/entity linking model. \cite{N16-1030}: BiLSTM-CRF with RNN-based char information. \cite{P16-1101}: BiLSTM-CRF with CNN-based char information.
\cite{D17-1283}: iterated dialated CNN and CRF.}
\label{tab:compare-ner-conll}
\end{table}
\begin{table*}[t]
\centering
\setlength{\tabcolsep}{4pt}
\begin{tabular}{l|ccc|ccc}
\toprule
\multirow{2}*{\textbf{Model}} & \multicolumn{3}{c|}{\textbf{CoNLL-2003}} & \multicolumn{3}{c}{\textbf{OntoNotes 5.0}} \\
& \textbf{P} & \textbf{R} & \textbf{F} & \textbf{P} & \textbf{R} & \textbf{F} \\
\midrule
BiLSTM-CRF & 87.97 & 86.49 & 87.22 & 85.26 & 85.87 & 85.56 \\
\midrule
Deformable stacked structure & 87.91 & 86.79 & 87.33 & 85.79 & 86.14 & 85.96\\
\bottomrule
\end{tabular}
\caption{Performances of models without character embedding on test set.}
\label{tab:compare-no-char}
\end{table*}
\begin{table}[t]
\centering
\setlength{\tabcolsep}{4pt}
\begin{tabular}{ccccc}
\toprule
\textbf{Structure} & \textbf{\# of offsets} & \textbf{P} & \textbf{R} & \textbf{F} \\
\midrule
baseline & - &87.03 & 86.97 & 87.00 \\
\midrule
1 & 1 & 87.14 & 88.66 & 87.81 \\
1 & 2 & 86.96 & 88.68 & 87.81 \\
1 & 3 & 87.34 & 88.51 & 87.83 \\
\midrule
2 & 1 & 87.20 & 87.86 & 87.53 \\
2 & 2 & 86.92 & 88.18 & 87.55 \\
2 & 3 & 86.99 & 88.27 & 87.63 \\
\midrule
3 & 1 & 87.24 & 88.56 & 87.90 \\
3 & 2 & 87.15 & \textbf{88.81} & 87.97 \\
3 & 3 & \textbf{87.64} & 88.38 & \textbf{88.01} \\
\bottomrule
\end{tabular}
\caption{Effects of different structure setting on OntoNotes 5.0 datset. Structure 1: deformable stacked structure between LSTM layers. Structure 2:
deformable stacked structure between LSTM layers and CRF. Structure 3: both structure 1 and structure 2.}
\label{tab:res-different-structure}
\end{table}
The results of our model on OntoNotes 5.0 and CoNLL-2003 dataset datasets are shown in Table \ref{tab:compare-ner-ontonotes} and Table \ref{tab:compare-ner-conll}.
On both of the two datasets, our model outperforms the vanilla stacked BiLSTM-CNN-CRF baseline. We also compare our model with existing models
on the two datasets. On OntoNotes dataset, our model also achieves the state-of-the-art result.
Deformable stacked structure utilizes the feature of neighbor words. Thus, it should make better predictions when OOV occurs.
We conduct experiments on models without character embedding. The results are shown in Table \ref{tab:compare-no-char}.
\begin{figure*}[t]
\centering\small
\subfloat[1 offset, structure 1]{
\includegraphics[width=0.33\textwidth]{bot_k1} \label{fig:k1-bot}
}
\subfloat[2 offsets, structure 1]{
\includegraphics[width=0.33\textwidth]{bot_k2}\label{fig:k2-bot}
}
\subfloat[3 offsets, structure 1]{
\includegraphics[width=0.33\textwidth]{bot_k3} \label{fig:k3-bot}
}
\\\vspace{-1em}
\subfloat[1 offset, structure 2]{
\includegraphics[width=0.33\textwidth]{up_k1} \label{fig:k1-up}
}
\subfloat[2 offsets, structure 2]{
\includegraphics[width=0.33\textwidth]{up_k2} \label{fig:k2-up}
}
\subfloat[3 offsets, structure 2]{
\includegraphics[width=0.33\textwidth]{up_k3} \label{fig:k3-up}
}
\caption{Kernel density estimation of offsets on test set. Y-axis represents the kernel density, and X-axis represents different values of offset.
(a)(d): offset number $k=1$. (b)(e): offset number $k=2$. (c)(f): offset number $k=3$. Structure 1: deformable stacked structure between LSTM layers. Structure 2:
deformable stacked structure between LSTM layers and CRF. Structure 3: both structure 1 and structure 2.}
\label{fig:vis-offset}
\end{figure*}
\paragraph{Comparison of different deformable stacked structure}
On OntoNotes dataset, we adopt three different deformable stacked structures. We evaluate the performance of the three different structure with varying
numbers of offsets.
As we can see from Table \ref{tab:res-different-structure}, the deformable stacked structure between LSTM layers has greater improvement
compared with the deformable stacked structure between the encoder layer and the decoder layer. These two structures differ on the function of their
next layers, and the difference in improvement also comes from that. For structure 1, the deformable stacked structure reconstructs the input of the next LSTM layer. Thus, it plays a role of feature augmentation
for the next LSTM layer to better extract related information. And better extraction of feature also helps with the decoding procedure. For structure 2, the deformable
stacked structure reconstructs the input of the CRF layer and plays a role of feature selection for decoding.
Taking information from multiple positions helps as increasing the number of offsets shows improvement.
We can also draw a significant character of the deformable stacked structure from Table \ref{tab:res-different-structure}. Compared with the baseline model,
the deformable stacked structure has great improvement in recall. It indicates that the deformable stacked structure can discover more named entity in the
sentence.
\subsection{Offsets Visualization}
To give an intuitive impression of how the offset dynamically changing, we show
the kernel density estimation of offset values of different deformable stacked structures on the OntoNotes test set in Figure \ref{fig:vis-offset}.
For the multiple offsets setting, we give the kernel density estimation of each offset. From Figure \ref{fig:vis-offset}, we can see that these offsets have the normal distribution in general.
For multiple offsets of deformable stacked structure between LSTM layers, we observe that at least one of the offsets has a distribution similar to
a normal distribution with mean 0 and the means of the rest distribution are slightly shifted from 0. For multiple offsets of deformable stacked structure
between the encoder layer and the decoder layer, the distributions shift from 0 significantly.
\subsection{Case Study}
We also visualize the offsets of a real case of structure 3 with offset number $k=3$ in Figure \ref{fig:vis}.
For simplicity, we only display the part of sentences that have different
predictions from the vanilla stacked model.
With deformable stacking, our model can better recognize the borders of named entities.
We can see that the positions that our model predicts correctly while the vanilla stacked structure predicts wrongly
have similar offsets between the CRF layer and the second LSTM layer. These offsets provide features from the
following positions and let the model know that the named entity doesn't break at the word "Faith".
Correctly predicting the borders of named entities shows great performance improvement when using BIOES tagging.
\begin{figure*}[t]
\centering
\includegraphics[width=0.95\textwidth]{visualize-sample2}
\caption{A real case on OntoNotes. The dashed lines indicate the dynamic offsets in our model. The red tags denote wrong predictions by the vanilla stacked structure, and the green tags denote the correct predictions by our
model. The black tags denote the correct predictions by both models.}
\label{fig:vis}
\end{figure*}
\section{Related Work}
There are mainly two lines of work related to ours.
One is the neural architecture for named entity recognition. Recently, several different neural network architectures have been proposed and successfully applied to NER. Among these neural architectures,
BiLSTM+CRF \cite{huang2015bidirectional} has become a fundamental architecture, which consists of a bi-directional LSTM as the encoder and a conditional random field (CRF) as the decoder. Some work \cite{Q16-1026,P16-1101,chen2017feature} also introduced a CNN layer before BiLSTM layer to model character-level information and achieved better performances. Besides BiLSTM, there is also some work to adopt CNN as encoder to capture the context information. \cite{D17-1283} use a dilated convolutional neural networks to efficiently aggregate broad context information.
Compared to these models, our model can effectively increase the input width of stacked layers and
help aggregate more broad context.
Another is neural architecture search \cite{pham2018efficient,zoph2016neural,AAAI1816537}.
Neural architecture search aims to automatically design the architecture of neural networks for a specific task. The current methods mainly adopt reinforcement learning to maximize the expected accuracy of the generated architectures on a validation set.
Our model can be regarded as a ``lightweight'' architecture search model, and changes the connections between the adjacent stacked layers. Moreover, we use an approximate strategy to change the connections softly.
\section{Conclusion}
We present deformable stacked structure, in which connections between two adjacent layers are dynamically generated. Three different deformable stacked structures
are designed and evaluated. Moreover, we also propose an approximate strategy to softly change the connections, which makes the whole neural network differentiable and end-to-end trainable.
Our model achieves the state-of-the-art performances on the OntoNotes dataset.
There are several potential directions for future
work. First, we hope to extend this work to build more flexible neural architecture. We have already established deformable connections between the most of layers of the encoder,
but some layers are still vanilla stacked such as the embedding layer.
Another exciting direction is to apply our model to other NLP tasks, such as parsing. Since our model does not require any task-specific knowledge, it might be effortless to apply it to these tasks.
\bibliographystyle{aaai}
|
\section{Introduction}
One of today's most relevant cryptographic protocols is Transport Layer Security (TLS). It provides communication security by symmetrically encrypting the data between two parties that have access to a shared secret which is generated at the beginning of a conversation. TLS is widely used in applications that provide secure communication over the internet. For example, web browsers, email, smartphone applications and the more recent class of smart home devices (e.g., Amazon Alexa) rely on the security of correct TLS configurations.
However, recent studies on TLS \cite{durumeric2013analysis, holz2015tls} show, that many endpoints are vulnerable to publicly known attacks like ``Heartbleed''\footnote{\url{https://heartbleed.com}} or ``DROWN'' \cite{aviram2016drown}. In addition, especially web servers are often poorly configured and allow for Man-in-the-Middle attacks~\cite{fahl2014eve}.
\begin{marginfigure}[0pc]
\begin{minipage}{\marginparwidth}
\centering
\begin{tikzpicture}
\begin{axis}[
height=9cm,width=\marginparwidth,
xbar, enlargelimits=0.15, xmin=0, xmax=60,
xlabel={Percentage},
symbolic y coords={.br,.uk,.jp,.de,.net,.org,.ru,.com},
ytick=data,
nodes near coords, nodes near coords align={horizontal},
legend columns=2,
legend style={at={(0.5,-0.2)},anchor=north},
]
\addplot[pattern=crosshatch dots] coordinates {(43.3,.com) (3.6,.org) (2.1,.ru) (4.7,.net) (4.6,.de) (2.8,.uk) (1.4,.br) (1.1,.jp)};
\addplot[pattern=grid] coordinates {(46.4,.com) (5.1,.org) (4.7,.ru) (4.0,.net) (3.6,.de) (2.2,.uk) (2.0,.br) (1.7,.jp)};
\legend{Our Data, W3Techs}
\end{axis}
\end{tikzpicture}
\caption{The distribution of the most used top level domains in our sample}~\label{fig:tldpercentagew3}
\end{minipage}
\end{marginfigure}
Fahl et al. \cite{fahl2014eve} found that webmasters often lack a detailed understanding of TLS which results in insecure and vulnerable communication channels. A recent lab experiment related to the configuration process of TLS (i.e., enabling HTTPS) \cite{krombholz2017have} confirmed, that even experienced users have problems to execute the configuration task in the most secure way. Krombholz et al. \cite{krombholz2017have} reported poor usability as the main reason for weak configuration results. While this is a important finding as it indicates that administrators, and thus the security of a TLS configuration, would benefit from more usable processes, the researchers excluded important real-world aspects. Most importantly, the evaluation was based on a short-term lab experiment and focused on the initial deployment process of a HTTPS-enabled web server configuration and thus, excluded long-term maintenance tasks, such as certificate renewal and the administrators' reactions to newly discovered vulnerabilities.
In contrast to this previous work, we analyzed real TLS security test results provided by the web service \textit{Qualys}\footnote{\url{https://www.ssllabs.com/ssltest/}} and propose to use this data for a complementary research approach. \textit{Qualys} is a web service which assesses the vulnerability of a TLS server configuration through grades from A to F. We argue that analyzing such data can provide important insights into real-world problems concerning TLS configurations and can be used to confirm or question previous knowledge gained from lab studies and self-reports.
Over a time span of 2.5 months, we collected 445,187 test results from 144,292 unique domains.
We identified two usage patterns: the \textit{Initial Configuration} and \textit{Reoccurring Checks}. Regarding the number of tests, we observed that the grades are increasing when more tests are done. In summary, this paper provides preliminary real-world insights into the difficult task of configuring TLS and indicates that more usable concepts are needed. In addition, we confirm previous findings which found that expert users are often unable to decide on the appropriate level of security. Our results highlight the need for more usable configuration processes and feasible support in decision making.
\section{Approach}
To collect real-world data, we used one of the most popular online service for SSL testing from \textit{Qualys SSL Labs}. We specifically focused on the \textit{Recently Seen}, the \textit{Recent best} and the \textit{Recent worst} grades which were shown on the services' front page. We wrote a python script that visited the site every five seconds\footnote{During the development phase, five seconds was found to be optimal for our purpose.}, extracted the domains with their related grades and stored them in a database. We collected data from September, 29th 2017 to December 14th, 2017, a time period of about 2.5 months (77 days). Overall, we analyzed 269,097 tests from 143,660 unique domains.
\subsection{Qualys Rating Scale}
\textit{Qualys} scoring is a mix of three categories which influence the security of a servers' TLS configuration. According to \textit{Qualys}, the final grade is calculated based on the following aspects:
\begin{itemize}
\item \textit{Protocol support (30\%)}: Depending on the supported protocols (SSL 2.0, SSL 3.0, TLS 1.0, TLS 1.1, TLS 1.2) the grade is defined by the weakest one available.
\item \textit{Key exchange (30\%)}: Based on the parameters at key exchange (e.g. key length).
\item \textit{Cipher strength (40\%)}: Qualys scores the length of the used cipher.
\end{itemize}
\begin{margintable}[-8pc]
\begin{minipage}{\marginparwidth}
\centering
\begin{tabular}{l r}
\small\textit{Sessions} & \small \textit{Occurrences} \\
\midrule
1 & 132,396 \\
2 & 7,399 \\
3 & 1,950 \\
4 & 826 \\
5 & 386 \\
$\geq$6 & 669 \\
\midrule
\textbf{Total} & \textbf{143,626}
\end{tabular}
\caption{Number of clustered testing sessions per domain}~\label{tab:sessiondistribution}
\end{minipage}
\end{margintable}
\begin{margintable}[0pc]
\begin{minipage}{\marginparwidth}
\centering
\begin{tabular}{l r r}
{\small\textit{Grade}} & {\small \textit{Total}} & {\small \textit{Percentage}}\\
\midrule
A+ (6) & 17,427 & 14.5\% \\
A (5) & 50,630 & 42.0\% \\
A- (4) & 11,976 & 10.0\% \\
B (3) & 14,759 & 12.2\% \\
C (2) & 9,152 & 7.6\% \\
F (1) & 10,047 & 8.3\% \\
T (0) & 6,550 & 5.4\% \\
\midrule
\textbf{Total} & \textbf{120.541} &
\end{tabular}
\caption{The distribution of the last grade we observed for each domain}~\label{tab:lastgradedist}
\end{minipage}
\end{margintable}
The combination of these three categories results in a final score from 0 to 100, that gets translated in a grade from A+ to F.
More details about the used scoring rules (e.g., the differences of A+, A and A-) are available online\footnote{\url{https://github.com/ssllabs/research/wiki/SSL-Server-Rating-Guide} -- accessed: 05/24/2018}.
\section{Preliminary Results}
Even though our observation was limited to 77 days and thus represents a snap shot of the real world, we are confident that the data is suited to inform our research question. Fahl et al. \cite{fahl2014eve} observed that 13.6\% of all servers triggered a warning in the users' browser. Our data confirms this finding as 13.7\% of all servers were graded with F or T, representing a very weak configuration (e.g., use a broken cipher like RC4, a key length of 512 and SSL 2 enabled) and a non-trusted certificate. Such configurations usually trigger warning messages in modern web browsers.
In addition, the Top-Level-Domain (TLD) distribution of our dataset matches the overall distribution as compared to publicly available W3Techs data \footnote{\url{https://w3techs.com/technologies/overview/top_level_domain/all} -- accessed: 05/24/2018} (see figure \ref{fig:tldpercentagew3}).
\subsection{Usage}
In 77 days, we observed 269,097 tests from 143,660 domains. The average number of tests per domain was $\mu=1.87$ ($\sigma=32.86, min=1, max=6,857$). Since a small number of domains was responsible for a large number of tests, we opted to sanitize the data by focusing on the domains whose test count was in the interval of $[0, \mu+3\sigma = 160]$. This way, 34 domains were identified as outliers and excluded from the analyses. The final data set comprised 143,626 domains and 219,877 tests ($\mu=1.53, \sigma=2.53$ tests per domain).
We grouped multiple consecutive tests into the same \emph{Test Session}. As previous lab experiments reported configuration sessions of four hours \cite{krombholz2017have}, a session was defined as closed whenever we did not observe a test within four hours after the last test had been logged. Table \ref{tab:sessiondistribution} shows the number of sessions we observed and their corresponding occurrences. It shows that the vast majority of users performed just one test session within the given time frame of 77 days.
\textbf{Initial Configuration}\\
\textit{Initial Configurations} are defined as the first occurrence of a domain in the data set having at least two checks. We set a minimum of two checks as we were interested in evaluating iterative configuration processes. We assume that this data represents the process where an administrator tests the server multiple times until she is satisfied with the resulting grade and the properties of the configuration.
The vast majority - 132,396 (92.18\%) - just tested once in their first appearing four-hour-time-interval, indicating no iterative process. In the remaining 11,225 \textit{initial phases}, we observed 2.97 ($\sigma = 1.72$) tests on average, meaning that most people checked their server configuration three times before seeing their task finished or giving up. The maximum of performed checks within a test session was 25.
\textbf{Reoccurring Checks}\\
In addition, we found that some administrators use the Qualys service more frequently over longer periods of time. For example, one domain checked its configuration approximately every 2-5 minutes resulting in 6,857 checks within 2.5 months (\textasciitilde90 tests/day). Table \ref{tab:sessiondistribution} gives more details.
\begin{marginfigure}[0pc]
\begin{minipage}{\marginparwidth}
\centering
\begin{tikzpicture}
\begin{axis}[
width=1.15*\marginparwidth, height=10cm,
stack plots=y,
area style,
enlarge x limits=false,
transpose legend,
legend columns=2,
legend style={at={(0.5,-0.2)},anchor=north},
xlabel={Number of tests},
symbolic x coords={2, 3, 4, 5, 6, 7, 8, 9, $\geq$10},
xtick=data,
]
\addplot[pattern=horizontal lines, pattern color=green] coordinates {(2,12.95) (3,16.98) (4,17.62) (5,23.71) (6,24.51) (7,27.51) (8,25.11) (9,25.60) ($\geq$10,25.55)} \closedcycle;
\addplot[pattern=vertical lines, pattern color=green] coordinates {(2,40.22) (3,39.53) (4,39.51) (5,39.07) (6,39.77) (7,37.23) (8,40.06) (9,37.68) ($\geq$10,43.77)} \closedcycle;
\addplot[pattern=grid, pattern color=green] coordinates {(2,10.87) (3,10.22) (4,10.96) (5,8.74) (6,9.71) (7,9.60) (8,9.45) (9,10.39) ($\geq$10,9.63)} \closedcycle;
\addplot[pattern=crosshatch, pattern color=yellow] coordinates {(2,12.63) (3,10.95) (4,11.28) (5,11.38) (6,9.93) (7,10.87) (8,10.44) (9,12.32) ($\geq$10,8.86)} \closedcycle;
\addplot[pattern=north east lines, pattern color=orange] coordinates {(2,8.39) (3,6.50) (4,7.51) (5,5.54) (6,6.66) (7,4.51) (8,5.78) (9,6.04) ($\geq$10,5.20)} \closedcycle;
\addplot[pattern=dots, pattern color=red] coordinates {(2,9.07) (3,9.10) (4,9.26) (5,6.66) (6,5.93) (7,5.09) (8,6.49) (9,4.83) ($\geq$10,3.98)} \closedcycle;
\addplot[pattern=fivepointed stars, pattern color=red] coordinates {(2,5.87) (3,6.72) (4,3.87) (5,4.89) (6,3.49) (7,5.20) (8,2.68) (9,3.14) ($\geq$10,3.02)} \closedcycle;
\legend{A+,A,A-,B,C,F,T}
\end{axis}
\end{tikzpicture}
\caption{Grade Distribution compared to the number of checks done with Qualys SSL Test}
\label{fig:gradecountdistribution}
\end{minipage}
\end{marginfigure}
\subsection{Grade Distribution}
To get insights into the overall distribution of the final grades, we focused on the last observed grade domains that were successfully tested at least twice\footnote{Qualys' test sometimes fails (e.g. due to ``Certificate name mismatch'') and requires a manual restart of the test process.}.
Overall, we logged 120.541 successfully performed tests. Table \ref{tab:lastgradedist} shows the distribution of the grades. The average of all ``final'' ratings is close to A- (3.97) with a $\sigma$ of 1.75.
\subsection{Grade changes}
Figure \ref{fig:gradecountdistribution} shows that with an increasing number of tests the proportion of all A grades rises from 64\% concerning two tests to nearly 75\% with six tests. This indicates that administrators tend to test and reconfigure their servers multiple times to achieve good security and tools like Qualys can support them in doing so.
\section{Discussion}
In the following, we discuss the implications of our preliminary findings.
\subsection{Securing TLS Configurations Takes Time}
Hardening TLS Configurations is an iterative process. Our analysis shows that the resulting TLS grade and thus, the servers' security is improving with the number of tests. This indicates that short-term lab studies might oversee some aspects when testing the ability of administrators to configure and harden a service that uses TLS as administrators often follow a trial and error approach and secure configurations may take several hours.
\subsection{Not All Services Achieve Best Grades}
Despite the fact that Qualys suggests improvements that can be done to harden the security of a configuration, we found that more than 21\% of all logged domains finished their last test with a grade of C or worse. To get more insights into the reasons for this real-world behavior, future studies should collect qualitative feedback to analyze why some administrators seem satisfied with bad grades.
\subsection{Only a Small Subset of Webservices Make Use of Qualys}
The number of all publicly known active websites\footnote{\url{https://www.netcraft.com/active-sites/} -- accessed: 05/24/2018} in combination with \textit{Let's Encrypt}'s statistics\footnote{\url{https://letsencrypt.org/stats/} -- accessed: 05/24/2018} indicates that there are nearly 115 million domains using TLS. We observed tests by only 0.12\% of these websites. Even though we assume that the number of administrators who use services like Qualys is much higher than observed, we still conclude that a minority of HTTPS-enabled web servers is actively tested and that the security of such services is seldom monitored.
\section{Conclusion \& Future Work}
In this paper we proposed to analyze \textit{Qualys SSL Server Test} results as a new information vector to understand the problems of the TLS configuration process.
In the future we plan to combine our quantitative data with qualitative feedback by contacting some of the observed web services directly.
Finally, it would be interesting to observe and analyze testing behaviour whenever new TLS vulnerabilities like DROWN are found and publicly discussed.
\section{Acknowledgments}
The authors would like to thank Qualys Inc. for granting access to their services.
\balance{}
\bibliographystyle{SIGCHI-Reference-Format}
|
\section{Introduction}
Wikipedia, the multilingual encyclopedia project, which is supported by the Wikimedia Foundation, and based on a model of openly editable content, is available in 291
languages. Even though there is no central authority who dictates which topics should be covered, one of the challenges for
Wikipedia, in the last decade, has been to balance the coverage of content across its different languages \cite{West2016,omnipedia}. For this purpose, a recommendation system is
applied by the Wikimedia Foundation in order to encourage Wikipedians to fill that gap \cite{Manske}.
\\
\\
However, contributing to Wikipedia means more than writing encyclopedic contents. Indeed, it allows communities to store cultural memories of events, to show the reality by their
own lens and to document their prominent people and places \cite{samolienko}. In this sense, even though we understand the inconvenience of the imbalance between the information
among several languages, we hypothesize that this collectively genuine gap has some important implications. It represents legitimate preferences among individuals sharing the
same language, which is a footprint of the groups’ collective identity.
\\
\\
Our goal, in this communication, is to analyze the broad preferences of the population editing Wikipedia (WP), depicted by categories over several worldwide languages. In addition,
we are also investigating cultural language-based footprints, since they have a tendency to disappear due to nowadays globalization. Our analysis is, hence, limited to the first
10 years of the edits in each language, when no intervention to cover the gap between languages had yet been done. Our study covers twelve Wikipedias: the ones written in
English (EN-WP), Spanish (ES-WP), French (FR-WP), Portuguese (PT-WP), Italian (IT-WP), Hungarian (HU-WP), German (DE-WP), Russian (RU-WP), Arabic (AR-WP), Japanese (JA-WP),
Chinese (ZH-WP) and Vietnamese (VI-WP). Our selection has been done based on the interplay between a worldwide view and the WP sizes. Some limitations are present in our study,
as the fact that to some extent, some WP languages have more global than local character, as for example the English one (EN-WP), which is worldwide edited. This language is only used
for comparative purposes.
\section{Methodology}
In order to have a broad view regarding preferences among individuals sharing the same language, we have chosen the categories already classified as in
the main branch of the tree structure, defined by the Wikimedia Foundation itself, and found in \cite{tree}. For the sake of avoiding a strong overlap between categories,
after taking a look at the pages within each category, our only intervention has been the decision of leaving out of the study the categories: Culture,
Humanities, Law, Life, Matter, People, Reference Works, Science and Technology, Society, Universe, and World. Given that the pages within such categories appeared in several of them.
Then, over the original 22 categories, our
study is restricted to the 13 following ones: Arts, Sports, Right, Events, Philosophy, Geography, History, Games, Mathematics, Nature, Politics, Religion,
and Health. We, then, have used the petscan API \cite{API} in order to download the names of all the pages within each category. This procedure has been done
for the 12 languages under study.
In order to ensure the same category between languages, we started from the category in English (EN) and link
to the WP page for the same category in the desired language. Once the name of the category in the new language was obtained, then, all the pages within that category were collected, by
fixing the category and the new language in the petscan API \cite{API}, without any link to the EN-WP. In this sense, different languages have different number of pages in each category.
All the activity of edits on the openly available Wikipedia was downloaded, also available in \cite{konect}. The starting point has been to set the starting date for each language, which ranges from
$11/10/2001$ to $28/03/2010$ for all of them. Because some information was not related to edits, but to other processes of Wikipedia's routines, only the pages fitting with the pages in one of
the studied categories were taken into account. We have removed the activity of the bots, by removing all the edits done by users whose name contains the word bot in any combination
of both uppercase and lowercase. In Table \ref{tab:twitternets} we show the number of pages studied and the number of edits for each language. This will constitute
the data-set of our study.
\begin{table}[h!]
\centering
\begin{tabular}{|c|c|r|r|}
\hline
WP & \# pages & \# edits \\
\hline
$EN$ & 1999263 & 222417366 \\
$ES$ & 1144177 & 47728243 \\
$FR$ & 2936383 & 58325545 \\
$PT$ & 894521 & 19937771 \\
$IT$ & 1084333 & 22200807 \\
$HU$ & 248808 & 5758998 \\
$DE$ & 1111265 & 39689676 \\
$RU$ & 1134752 & 14199590 \\
$AR$ & 624118 & 7674946 \\
$JA$ & 690795 & 24584471 \\
$ZH$ & 495855 & 3657770 \\
$VI$ & 238859 & 12618296 \\
\hline
\end{tabular}
\vspace{0.5cm}
\caption{Characteristics of the data-set.}
\label{tab:twitternets}
\end{table}
\section{Results}
\subsection{By languages}
We start by showing the number of edits in Fig.~\ref{fig1}. In the edits, we include any action done by WP-users, including correction, addition, and deletion. The motivation for
including all these actions is the hypothesis stating that the higher the number of edits, the greater the interest on that subject \cite{gandica}. The Y-axis represents, in different
colors, the proportion of edits in each category. The x-axis shows the language
at the bottom and the total number of edits at the top. Naturally, the EN-WP surpasses the other languages. FR-WP and ES-WP follow as the next more edited languages. The most
edited categories are `Arts' and `History', followed by `Nature' and `Politics'. All the languages follow roughly similar patterns, with some interesting particularities. For
example, the category `Arts' is predominantly edited in the DE-WP. While `History' dominates HU-WP and JA-WP. `Nature' appears more important for RU-WP, and `Politics' is the most
edited category in VI-WP.
\begin{figure}[tbp]
\scalebox{0.5}{\includegraphics{new_edits2}}
\caption{Distribution of edits for each language. A colour is associated to each category.
The colouring of each column gives the proportion of each category with respect to the total number of edits for the given language.}
\label{fig1}
\end{figure}
It is important to notice that the number of edits, apart from being the result of the collaborative process of editing in order to complement the information, can be highly influenced
by the disagreement between editors, which can lead to edit wars, characterized by a large number of edits \cite{yg}. In this sense, another interesting observable is the total number of
pages for each language, which is shown in Fig.~\ref{fig2}. Unexpectedly, the highest number of pages is not
in the EN-WP but in the FR-WP. This is an example of the fact that even though the EN-WP is the largest edition, several articles about local places and events, are mostly written only
in the local languages of those locations \cite{understanding}.
\\
The categories with more WP pages are `Arts', `History' and `Nature'. `Arts' continues to be dominated by DE-WP. But now `History' seems predominant for HU-WP and RU-WP. Then, we can
hypothesize that the predominant editing behavior in JA-WP could be highly influenced by edit wars. This hypothesis is an interesting point
to be studied in a future work. JA-WP seems to be more interested in developing subjects regarding `Nature' (also taking into account the results for the number of editors). VI-WP
continues focused on `Politics', this last pattern seeming robust.
\begin{figure}[tb]
\hspace{-0.5cm}\scalebox{0.48}{\includegraphics{new_pages2}}
\caption{Distribution of pages for each language. A colour is associated to each category.
The colouring of each column gives the proportion of each category with respect to the total number of pages for the given language.}
\label{fig2}
\end{figure}
We now show in Fig.~\ref{fig3} the number of editors for each language. As in the number of edits, the biggest categories are `Arts' and `History', followed by `Nature' and `Politics'.
The effect of colonialism starts to be visible. ES-WP and PT-WP appear as the most populated. This last calculation confirms the dominant interest on art by DE-WP, while the population
editing in HU-WP is more interested in `History'.
\begin{figure}[tbp]
\hspace{-0.5cm}\scalebox{0.48}{\includegraphics{new_users2}}
\caption{Distribution of editors for each language. A colour is associated to each category.
The colouring of each column gives the proportion of each category with respect to the total number of editors for the given language.}
\label{fig3}
\end{figure}
\subsection{Homogeneity of the preferences}
The degree of homogeneity among the categories, in terms of the number of pages and editors was calculated by means of the $\chi^2$-test with respect to the
uniform distribution. In this sense, the lower the value, the more homogeneous the category distribution will be for that language. Results are reported in Fig.~\ref{fig4}. Our limit case,
the EN-WP, has low values for both, for the number of pages and for the number of editors, probably due to the expected diversity associated with worldwide editing. Surprisingly, those low values are
comparable with the ones by IT-WP and FR-WP. Since the first hypothesis for the homogeneity of preferences is a heterogeneous distribution of people from several nationalities
--- background---, it is not easy to give an explanation for the last result. In the other extreme, we have DE-WP and HU-WP with the highest variability, as can be noticed in figures
\ref{fig2} and \ref{fig3}, following local patterns of pronounced preferences.
\begin{figure}[htbp]
\scalebox{0.35}{\includegraphics{var}}
\caption{Degree of homogeneity on the number of pages (in green) and the number of users (in black) among the categories. Calculated by the $\chi^2$-test.}
\label{fig4}
\end{figure}
\subsection{By categories}
Now let us explore the results in terms of the categories. How is the proportion for the number of edits, pages and editors in each WP language? The aggregation of all the values
will not be equal to one. However, looking at each value separately gives information about the real proportion in that language and can be compared with the proportion in the other
ones. As we will see the values are clearer in this representation. Results are depicted in Fig.~\ref{fig5},\ref{fig6} and \ref{fig7}, for the number of edits, pages and editors, respectively. Looking at the first icon in the three figures makes it clear
that DE-WP dominates in the category `Arts', followed by ZH-WP. The representation by languages helps to uncover second and third places, which are relevant as population's
preferences. The category `Sports' is mostly dominated by DE-WP and IT-WP. `Right' and `Philosophy' are singularly dominant for ES-WP.
The category `Geography' and `History' seem signally important for HU-WP. Taking into account the number of pages and editors, the role of ZH-WP and WI-WP looks important for the
development of category `Mathematics' and JA-WP for `Nature'. For `Politics' it is remarkable the low activity by IT-WP and JA-WP, while a high activity by VI-WP. Finally, the category
`health' seems to be more important for IT-WP and FR-WP, in comparison with the rest.
\begin{figure*}[htbp]
\scalebox{0.6}{\includegraphics{edits2.pdf}}
\caption{Each circle represents the proportion of edits among all the WP languages in each category. The colors in each circle represent a category, as indicated in the legend.}
\label{fig5}
\end{figure*}
\begin{figure*}[htbp]
\scalebox{0.6}{\includegraphics{pages2.pdf}}
\caption{Each circle represents the proportion of pages among all the WP languages in each category. The colors in each circle represent a category, as indicated in the legend.}
\label{fig6}
\end{figure*}
\begin{figure*}[htbp]
\scalebox{0.6}{\includegraphics{users2.pdf}}
\caption{Each circle represents the proportion of editors among all the WP languages in each category. The colors in each circle represent a category, as indicated in the legend.}
\label{fig7}
\end{figure*}
\section{Limitations}
The main limitation of working with Wikipedia data is the selection-bias, as all the works assume that the proportion of people editing WP is representative of the whole
group under study. The second limitation is the language-based comparisons for cultures, being language only one of the dimensions of the rich and complex elements composing culture
in society. Both limitations are present in all current-state-of-the-art studies. However, it is still possible to gather relevant information from
this data source despite such limitations \cite{mining,cul}.
\\
\\
Further limitations, specific to our work, include the fact that the categorization could be differently defined among the WP languages. Another limitation is that the same WP page
could belong to several categories. We have tried to reduce the last limitation by leaving out from the study the categories: Culture, Humanities, Law, Life, Matter,
People, Reference Works, Science and Technology, Society, Universe, and World. Such categories should be included whenever the overlap is not an undesirable situation in the
study.
\section{Discussion}
Collective interests of a language-speaking community have been studied by means of the voluntary process of editing. The $12$ languages that have been studied here have
been selected based on the interplay between a worldwide view and their Wikipedia editing sizes. The categories were taken from the main
branches of the Wikipedia's tree structure, defined by the same Wikimedia Fundation. We have shown the number of edits, pages and editors for each language, separately
by categories.
\\
\\
We have found stable patterns of preferences by languages, through the number of edits, pages and editors. Some results were unexpected. We were not able to find differences regarding languages spoken in
one country with respect to the ones spoken in several countries. A contrasting example was the case of the Italian language, which shows characteristics of diversity comparable to the
EN-WP. However, that could be due to the high number of Italian citizens living around the word.
\\
\\
To know the human preferences by languages could help to match data from different languages. The weight of preferences by categories could give some insights about how to perform a
better match when doing three kinds of analysis among different languages: Machine Translation, Speech Recognition and Sentiment Analysis; All those fields belong to Natural Language Processing. The present work
is a starting point for that purpose. We hope that our contribution can inspire future attempts at using the differences of human preferences by categories when performing
Multilingual NLP.
\\
\\
Interesting extensions of our work include discarding the effect of edit wars between wikipedians, in order to have more precise results. Also, the possibility of obtaining
the results for each country, instead of for each language, would be a great contribution. The study of the cultural importance of languages along the lines of the
study of the global influence that each language produces over the rest, presented in \cite{hidalgo_pnas}, would be an important development.
\\
\section{Acknowledgment}
Computational resources have been provided by the Consortium des équipements de Calcul Intensif (CBI), funded by the Fonds de la Recherche Scientifique de Belgique
(F.R.S.-FNRS) under Grant No.2.5020.11. YG thanks to J\'{e}r\^{o}me Kunegis and Silvia Chiachiera for valuable discussions and technical advises. YG thanks Julieta Barba by technical
and grafical support. YG thanks to http://www.opensym.org/os2018/ for a previous round of reviewing. YG thanks Fernando Sampaio Dos Aidos for proofreading the article.
|
\section{Acknowledgement}
\subsection{Image Sampling Strategies}
\label{sec:im_sampling_strat}
Using the representation for the images in our database from Section~\ref{sec:embedding}, our task is to select the $K$ most informative images to present to the user for feedback in order to quickly locate target image $I_t$. We begin by obtaining a short list of candidate images in Section~\ref{sec:candidates}, before refining this list using more powerful, but computationally expensive methods in Section~\ref{sec:sampling}.
\subsubsection{Candidate Selection Methods}
\label{sec:candidates}
Most selection strategies focus on trying to reduce uncertainty in the current model, or exploit the information obtained in order to make fine-grained distinctions. In practice, however, many search engines provide means to filter results based on meta-data labels. For example, when searching for clothing a search engine may allow you to filter results based on its category (\eg pants), subcategory (\eg jeans), and color, amongst others. Coupled with the initial query, this provides a strong signal to initialize an active learning algorithm. Thus, the criteria that follow focus on the exploitation of existing knowledge.
\smallskip
\noindent{\bf Nearest Neighbors.} As a baseline, we perform an iterative nearest neighbors query to obtain candidate images. Given query image $I_q$, this method returns the $K$-nearest neighbors to $I_q$ that have not been previously selected. Which ever image is selected as most relevant to the target image is used as the query in the next iteration.
\smallskip
\noindent{\bf Feedback Constraint Satisfaction.} Inspired by~\cite{KovashkaIJCV2015}, we find the samples which satisfy the maximum number of feedback constraints provided by the user. For each iteration that a new candidate query $I^*_{q_{i+1}}$ is accepted by the user, then we know that $I^*_{q_{i+1}}$ is closer to the target image than $I_{q_{i}}$. Analogously, if the candidate is not accepted, then we know $I^*_{q_{i+1}}$ is farther away from the target image than $I_{q_{i}}$. These become constraints $F$ where each element is a tuple $(I_x, I_y)$ where $I_x$ is closer to the target image than $I_y$. We define $l$ as a binary variable which indicates that we want to count the number of unsatisfied constraints (\ie for this criterion $l = 0$ so we count satisfied constraints). Then we can calculate the portion of constraints in $F$ a candidate image $I_o$ satisfies, \ie,
\begin{equation}
S(I_o | l, F) = \frac{1}{|F|} \sum_{\forall I_{x_n}, I_{y_n} \in F} \mathbbm{1}_{fcs}(I_o, I_{x_n}, I_{y_n}) \oplus l,
\label{eq:constraint_satisfaction}
\end{equation}
\noindent where $\mathbbm{1}_{fcs}$ is an indicator function which returns one if $D(I_o, I_{x_i}) < D(I_o, I_{y_i})$ under some distance function $D$. Thus, our criteria for the next proposed query from the set of candidates $\mathcal{O}$ is:
\begin{equation}
I^*_{q_{i+1}} = \argmax_{I_o \in \mathcal{O}} S(I_o | l = 0, F).
\end{equation}
\noindent Ties are broken using nearest neighbors sampling between the candidates and the query image.
\subsubsection{Candidate Re-ranking Methods}
\label{sec:sampling}
Many methods that measure how informative a sample is are computationally expensive, making it infeasible to run over a large database. Therefore, we begin by obtaining a short list of candidates using the methods from Section~\ref{sec:candidates}, then re-rank them based on how informative they are. Below we discuss two such re-ranking methods.
\smallskip
\noindent{\bf Expected Error Reduction.} Initially proposed in~\cite{royICML2001}, this refinement strategy focuses on reducing generalization error of the current model of the desired target image. Ergo, it can be seen as inherently balancing exploration and exploitation criteria. We measure the entropy of the current model by calculating the portion of constraints an image satisfies as done in Eq.\ (\ref{eq:constraint_satisfaction}), \ie,
\begin{equation}
H(F) = -\sum_{I_o \in \mathcal{O}}\sum_l \hspace{2mm} S(I_o | l, F)\log(S(I_o | l, F)).
\end{equation}
\noindent We use the highest ranked item which hasn't been presented to the user, which we denote as $I_{t^*}$, as a proxy for the target image when predicting the user's response $r$. The simulated response either accepts or rejects some image $I_c$ from our short list of candidate images $\mathcal{C}$ to create a new constraint. For example, $(I_q, I_c, r=0)$ would indicate that a constraint should be added to $F$ that says $I_c$ is farther away from the target image than the current query $I_q$. We decide if a new constraint would be satisfied by measuring the likelihood that the candidate image shares the same attributes with the target image. The candidate images are then selected according to the following:
\begin{equation}
I^*_{q_{i+1}} = \argmin_{I_c \in \mathcal{C}} \sum_r \sigma(r | D(I_c, I_{t^*}), \phi) H(F \cup (I_q, I_c, r)),
\end{equation}
\noindent where $\sigma$ converts the distances in the attribute's embedding space to probabilities using Platt scaling~\cite{platt1999} whose parameters $\phi$ we estimate using the training set. Effectively, we select the $I_c$ which we are the most uncertain about that is also similar to our best guess at the target image.
\smallskip
\noindent{\bf Learned Re-ranking Criteria.} So far only hand-crafted strategies have been discussed. Learned criteria can easily adapt to the exact task and dataset, making it an attractive option. To this end, we train a Deep Q-Network (DQN)~\cite{mnih-atari-2013} with experience replay to learn how to select informative images. In this paradigm, we learn a function $Q$ that estimates the reward $\rho$ we would get by taking some action given the current state of the system $\Psi$. We define $\rho$ as the change in the percentile rank of the target image under the current model after obtaining feedback from the user. We represent each image in the list of candidates obtained from the methods in Section~\ref{sec:candidates} as the difference between its visual embedding and the query image. This is fed into our DQN as the current state $\Psi$, which then selects which image to present to the user (\ie the set of actions asks which image to choose). At test time, the selection criteria simply need to maximize the expected reward if we were to select image $I_c$ to present to the user:
\begin{equation}
I^*_{q_{i+1}} = \argmax_{I_c \in \mathcal{C}} Q(I_c, \Psi).
\end{equation}
\noindent Our model is trained using a Huber loss on top of the temporal difference error between expected and observed rewards.
\section{Image Search with Active Feedback}
Our objective is to quickly locate a target image $I_t$ in a database given a query $q$. While the initial query can take multiple forms (\eg keywords, images~\cite{chumICCV2007,gongCVPR2011}, or sketches~\cite{yu2016sketch}), we will assume it is provided as an image $I_{q_0}$ which shares some desirable attribute with the target image. In order to locate the target image by obtaining feedback from the user, we need a representation where we can measure similarity between images in the database as well as a selection strategy which uses this representation to find informative images to present to the user. In this paper we provide enhancements for both learning the image representation, which we shall discuss in Section~\ref{sec:embedding}, and sampling strategies, which we will present in Section~\ref{sec:im_sampling_strat}.
\subsection{Globally-Consistent Attribute Embeddings}
\label{sec:embedding}
To compare two images, we train a set of embeddings, each representing a different attribute we wish to capture. This provides multiple senses of each image which we can use to select informative images to the user. Due to its state-of-the-art performance and efficiency, we chose to build upon the CSN model~\cite{veitCVPR2017} which we shall briefly review before describing our modifications.
\subsubsection{Conditional Similarity Network}
The CSN model was designed to learn a disentangled embedding for different attributes in a single model. A general image representation is learned through the image encoding layers of the model. Then a trained mask is applied to the representation to isolate the features important to that specific attribute. This enables each embedding to share some common parameters across concepts, while the mask is tasked with transforming the features into a discriminative representation. After obtaining the general embedding features between two images $G_i, G_j$, they are compared using a masked distance function,
\begin{equation}
D_{m}(G_i, G_j; m_a) = ||G_i \star m_a - G_j \star m_a ||_2,
\label{eq:mask_distance}
\end{equation}
\noindent where $m_a$ is the mask for some attribute and $\star$ denotes an element-wise multiplication. Then, given a triplet of embedding features $(G_x, G_y, G_z)$ where the pair $(G_x, G_y)$ share the same attribute label which is also not shared by $G_z$, the CSN model is trained using the margin based loss function given by
\begin{eqnarray}
\label{eq:triplet_loss}
\lefteqn{L_T(G_x, G_y, G_z; m_a) = } \\
& & \max\{0, h + D_{m}(G_x, G_y; m_a) - D_{m}(G_x, G_z; m_a)\} \,,
\nonumber
\end{eqnarray}
\noindent where $h$ controls the minimum margin between positive and negative pairs. The general unmasked embedding representation $G$ is L2 regularized to encourage regularly in the latent embedding space. The masks $m$ are $L1$ regularized to encourage a sparse feature selection. Thus, the complete loss function is
\begin{eqnarray}
\lefteqn{L_{CSN}(G_x, G_y, G_z; m_a) = } \\ \label{eq:csn_loss}
& & L_T(G_x, G_y, G_z; m_a) + \lambda_1||G||^2_2 + \lambda_2||m_a||_1 \,,
\nonumber
\end{eqnarray}
\noindent where $\lambda_{1-2}$ are scalar parameters.
We modify the original CSN model by L2 normalizing the final attribute representation (\ie $G \star m$ in Eq.~(\ref{eq:mask_distance})) as this tends to make training more stable~\cite{SchroffCVPR2015}. In addition, since the masks can be viewed as an attention over the general embedding features we force them to sum to 1 as typically done for attention models (\eg ~\cite{xu2015show}).
\subsubsection{Incorporating Global Compatibility}
\label{sec:global_sim}
Since our goal is to traverse our embeddings in order to locate some target image, it is desirable that they provide natural transitions from image to image. For example, if we were to transition from the anchor image to the rightmost image in Figure~\ref{fig:global_sim_motivation} it would be considered a significant divergence. The center image, while still different, seems like a more logical transition even though all three images belong to the boot category. Therefore, to make our embedding spaces more intuitive, we also take into account the overall similarity between two images beyond the attribute being encoded. Given the set of attributes $A_x, A_y, A_z$ for each of the images in our training triplet, we compute the difference in shared attributes between the negative and positive pairs:
\begin{equation}
w(A_x, A_y, A_z) = \max\{0, \frac{1}{\mathcal{E}}(|A_x \cap A_y| - |A_x \cap A_z|)\}
\end{equation}
\noindent where $\mathcal{E}$ represents the number of embeddings being trained. We prevent negative values of $w$ to maintain a minimum margin between negative and positive pairs of the triplet. We define our new margin $h'$ for Eq. (\ref{eq:triplet_loss}) as
\begin{equation}
h'(A_x, A_y, A_z) = \zeta + \eta w(A_x, A_y, A_z),
\label{eq:global_similarity}
\end{equation}
\noindent where $\eta, \zeta$ are a scalar parameters.
\begin{figure}
\centering
\begin{tabular}{c|cc}
Anchor Image & & \\
\includegraphics[width=0.06\textwidth]{images/anchor_boot} & \includegraphics[width=0.06\textwidth]{images/similar_boot} & \includegraphics[width=0.06\textwidth]{images/different_boot}\\
\hline
Shared Attributes with Anchor: & boot & boot\\
& mid-calf & \\
\hline
\end{tabular}
\caption{During training, we take into account the overall similarity between images based on the number of shared attributes. This encourages the model to maintain the left-to-right ordering of the images above in a category embedding space even though they all belong to the boot category}
\label{fig:global_sim_motivation}
\end{figure}
\section{Conclusion}
In this paper, we addressed the problem of active image search, but without expensive annotations or user requirements used in prior work. Instead, we introduced enhancements to the Conditional Similarity Network which improved its ability to make relative attribute comparisons. We used this representation in our experiments on active image search where we demonstrated the effectiveness of our DQN selection criterion and showed it was competitive with prior work which used expensive relative attribute annotations. In future work, we would like to build upon our current system by taking advantage of a hierarchical clustering method to organize our data which has proven useful in prior work~\cite{KovashkaICCV2013,MacAodhaCVPR2014}. Our model could also benefit from taking into account the diversity of selected images on each iteration by incorporating elements used in batch mode active learning approaches (\eg~\cite{Elhamifar_2013_ICCV,guoNIPS2008,kulesza2012determinantal}).
\section{Experiments}
\label{sec:experiments}
We begin by validating our image representation's ability to identify if two images share the same attributes in Section~\ref{sec:attributes}. Then we analyze the ability of our approach to perform our active image search task in Section~\ref{sec:active_search}.
\smallskip
\noindent{\bf Dataset.} Experiments were performed on the UT Zappos50K (UT Zap-50K) dataset~\cite{yuCVPR2014}. This dataset consists of just over 50K images taken from the Zappos website of shoes in a canonical view and homogeneous backgrounds. Each image has eight meta-data attributes associated with it: category, closure, gender, heel, insole, material, subcategory, and toestyle. Only the category and subcategory labels are required, resulting in a sparse labeling of the remaining attributes. We split images in the dataset by their productID, keeping 5000 products for testing, 1000 for validation, and used the remaining for the training.
\subsection{Attribute Embedding Experiments}
\label{sec:attributes}
\noindent{\bf Implementation Details.} We generally follow the training procedure described in~\cite{veitCVPR2017}. For each attribute in the dataset, we randomly sampled 200K triplets for training, 40K for testing, and 20K for the validation set from their respective images. We did not use the same triplets as~\cite{veitCVPR2017}, however, since they split their images randomly which could result in same product appearing in both the training and testing splits. Although we tried semi-hard negative sampling of triplets~\cite{SchroffCVPR2015}, it did not provide performance benefits in our experiments. The models were trained for 200 epochs with a batch size of 256. The best model is selected using the validation set. We set our parameters as the following: ($h = 0.3$, $\lambda_1 = 5e^{-4}$, and $\lambda = 5e^{-6}$ in Eq. (\ref{eq:triplet_loss}), and $\eta = h$ in Eq. (\ref{eq:global_similarity}). We initialize our model using an 18 layer Deep Residual Network~\cite{He_2016_CVPR} that was trained on ImageNet~\cite{deng2009imagenet}. All images are resized to $112 \times 112$ before being fed into the network.
\noindent{\bf Evaluation Metric.} Following~\cite{veitCVPR2017} we report the triplet satisfaction rate of our model (\ie the percentage of valid triplets for the 320K samples in the test set).
\smallskip
\noindent{\bf Results.} Table~\ref{tab:csn_results} reports our triplet satisfaction rate on the test set and compares our approach to the state-of-the-art. The first two lines of Table~\ref{tab:csn_results}(b) show that doubling the number of training triplets for the baseline model results in a very small improvement to performance. However, the third line of Table~\ref{tab:csn_results}(b) demonstrates that by including the $L2$ normalization on the embedding outputs and forcing each mask to sum to 1 (referred to in the table as ``constraints''), we can better leverage the additional training data, improving our performance by almost 2\%. Our full model, which includes these constraints as well as our attention to the global similarity between images (described in Section~\ref{sec:global_sim}) results in a 3\% improvement over the baseline.
\begin{table}[t]
\setlength{\tabcolsep}{3pt}
\centering
\begin{tabular}{|l|c|c|}
\hline
\multirow{ 2}{*}{Method} & triplets/ & \multirow{ 2}{*}{Accuracy}\\
& concept & \\
\hline
\hline
CSN & 50K & 79.29\\
CSN & 100K & 79.35\\
CSN + constraints & 100K & 81.27\\
CSN + constraints + global similarity & 100K & \textbf{82.28}\\
\hline
\end{tabular}
\caption{Triplet satisfaction rate and number of training triplets used for the UT-Zap50K dataset~\cite{yuCVPR2014}. Note: we used all eight meta-data labels rather than just the four reported in Veit~\etal~\cite{veitCVPR2017}.}
\label{tab:csn_results}
\end{table}
\begin{table*}[t]
\setlength{\tabcolsep}{4.8pt}
\centering
\begin{tabular}{|l|c|c|c|c|c|c|c|c|}
\hline
Method & Category & Closure & Gender & Heel & Insole & Material & Subcategory & Toestyle\\
\hline
\hline
CSN & 93.69 & 77.17 & 77.27 & 88.49 & 58.64 & 71.53 & 90.21 & 77.31\\
CSN + constraints & 93.07 & 79.80 & 80.15 & 89.40 & 60.35 & 75.00 & 92.90 & 79.46\\
CSN + constraints + global similarity & \textbf{94.48} & \textbf{81.63} & \textbf{81.37} & \textbf{89.62} & \textbf{61.68} & \textbf{75.75} & \textbf{93.98} & \textbf{79.83}\\
\hline
\end{tabular}
\caption{Triplet satisfaction rate using models trained with 100K triplets/concept on the UT-Zap50K dataset~\cite{yuCVPR2014} separated by attribute.}
\label{tab:csn_concept_results}
\end{table*}
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth,trim=.9cm 0cm 1.6cm 0cm,clip]{images/closure_tsne}
\caption{{\bf Visualization of the learned embedding.} t-SNE~\cite{vandermatten2008JMLR} of the closure embedding space using our improved CSN Model. Starting from the top left and moving down and to the right, in the first two pair of boxes we see that the embedding has learned to separate sandals based on heel-height despite both being slip-ons. The next three show that the embedding has learned subcategories of shoes despite all having lace-up closure mechanisms, with the last pair showing how different closing mechanisms for boots have also been separated.}
\label{fig:tsne}
\end{figure}
We break down the performance of our model by the attribute being learned in Table~\ref{tab:csn_concept_results}. The material and gender attributes reported the largest performance improvement at 4\% over the baseline CSN model. The subcategory attribute brought up its performance by just over 3.5\%, putting its performance more in line with the category attribute. While our model did improve the category attribute by almost 1\%, including just the constraints did lower performance slightly. However, this loss was more than made up by the improvements in the other attributes, and the model which included global similarity did best across all attributes.
To provide insight into the structure of the learned embedding spaces, we provide a t-SNE visualization~\cite{vandermatten2008JMLR} of the closure attribute in Figure~\ref{fig:tsne}\footnote{The closure attribute embedding is also provided in Figure 5(a) of Veit~\etal~\cite{veitCVPR2017} which mixes sandals, heels, and slippers in the same local space when not encouraging globally consistent embeddings.}. The highlighted boxes show how our embedding has learned to separate shoes with heels from those without, or athletic and dress shoes, despite having the same closing mechanism. This demonstrates how our representation can make more intuitive transitions while navigating the learned embedding space.
\subsection{Active Image Search Experiments}
\label{sec:active_search}
\begin{figure*}[t]
\centering
\includegraphics[width=0.73\textwidth,trim=0cm 0.5cm 5.5cm 0cm,clip]{images/graph_traversal_example}
\caption{{\bf Active Image Search Examples.} Each row is an example of the images selected by our system as we refine our search results going from our initial query to our target image which is contained by a red box}
\label{fig:traversal_example}
\end{figure*}
\begin{figure*}[t]
\centering
\topinset{\bfseries(a)}{\topinset{\bfseries(b)}{\includegraphics[width=0.95\textwidth,trim=.2cm 0.8cm .2cm 0.8cm,clip]{images/rank_per_iteration2}}{-.0in}{.25in}}{0.0in}{-3.05in}
\caption{Comparison of our DQN Refinement strategy over an embedding trained using binary attribute labels and the Attribute Pivots method~\cite{KovashkaICCV2013} which utilizes relative attribute annotations.}
\label{fig:target_ranking}
\end{figure*}
\noindent{\bf Implementation Details.} On each iteration, we select one image per attribute type (8 total) to present to the user. For our refinement experiments, we select 4 candidates per attribute using our Feedback Constraint Satisfaction criterion (32 total) and then re-rank them to select the top 8 images. In our DQN experiments, we use a replay memory size of 20,000, a discount rate of 0.999, and a batch size of 2048. Our DQN models are trained using simulated user feedback which we describe below. During training, we begin by performing a random action 90\% of the time and decay this randomized action rate exponentially until we reach 5\%.
\smallskip
\noindent{\bf Query-Target Pairs.} For each attribute in the dataset, we sample 2,000 pairs for training, 500 for validation, and 1,000 for testing resulting in 16,000, 4,000, and 8,000 pairs, respectively.
Each sample was randomly selected from the set of image pairs which share at least one attribute without restrictions. This means pairs can be semantically distant from each other (\eg a boot and a sandal without a heel can be sampled for pairs sharing that attribute), adding to the challenges faced by our model.
\smallskip
\noindent{\bf User Feedback.} User feedback is simulated by averaging the Euclidean distance between the candidate image and the target image between all embedding spaces and selecting the closest one. We evaluated the appropriateness of our user feedback mechanism by presenting triplets of images to 13 human annotators. Each triplet contained a target image and a pair of images to evaluate. Annotators were asked to select the image from the pair which most resembles the target image. We obtained 100 triplets selected at random from our active learning experiments (\ie they exactly reflected the decision process made by our algorithm). After removing triplets where at least 5 annotators disagreed with the majority, our simulated input agreed with the human annotators 79\% of the time over the remaining 86 samples. Human performance was similar, as individual annotators agreed with the majority 74-88\% of the time (84\% average). We also experimented with two other embeddings to use for our simulated feedback: features from our fine-tuned ResNet-18 model initially trained on ImageNet and the unmasked embedding representation output by our model. Relative performance remained consistent across feedback types in our active learning experiments.
\smallskip
\noindent{\bf Comparison to Relative Attribute Approaches.} In addition to our own baselines, we also adapt our embedding approach to produce relative attribute scores for four common concepts using the annotations provided in~\cite{yuCVPR2014}. We encode a pair of images $I_i, I_j$ using the Conditional Similarity Network to obtain an embedding representation for an attribute. We concatenate together the embedding from each image which is fed into a fully connected layer followed by a softmax with an output dimension of 3 and is trained jointly with the embeddings. The output of this model indicates if image $I_i$ has more, less, or the same amount of an attribute as image $I_j$. Using this model, we reproduce the binary tree EER-based approach of~\cite{KovashkaICCV2013}. We initialize the activate image search model using the query image provided at test time. The remaining implementation details follow~\cite{KovashkaICCV2013}. We shall refer to this reproduction as Attribute Pivots henceforth.
\smallskip
\noindent{\bf Evaluation Metric.} We evaluate performance based on how many iterations were required to go from the initial query to the target image.
\smallskip
\noindent{\bf Results.} We report our active image search performance in Table~\ref{tab:query_refinement_performance_multispace}\footnote{We don't include a comparison to Attribute Pivots~\cite{KovashkaICCV2013} in Table~\ref{tab:query_refinement_performance_multispace} as it regularly satisfied its stopping criterion (\ie~it was no longer able to improve its model) before finding the target image. Altering the stopping criterion so that it would only stop when it located the target image resulted in poor performance on this task.}. As seen in the top two lines of the table, using the feedback constraint satisfaction criterion reduces the number of iterations required to find a target image by 2 over the nearest neighbors baseline. Refining the top 4 candidates using expected error reduction reduces the number of steps by 1.5, and our DQN refinement reduces this further, making the total reduction approach 5 iterations fewer than the baseline. It is important to note that this would be considered the toughest settings for this task. In practice, a user could remove a lot of images from consideration by filtering by the meta-data labels.
\begin{table}[t]
\centering
\begin{tabular}{|l|c|}
\hline
Selection Strategy & \#Steps\\
\hline
\hline
Nearest Neighbor & 26.40\\
Feedback Constraint Satisfaction & 24.62\\
Expected Error Reduction & 23.07\\
DQN Refinement & 21.79\\
\hline
\end{tabular}
\caption{Active image search performance on UT Zap-50K.}
\label{tab:query_refinement_performance_multispace}
\end{table}
We provide examples of the images selected by our system as we refine image search results in Figure~\ref{fig:traversal_example}. In the first row, we see how the boots change in style on each iteration, deciding on the heel first before refining the style of the boot. The second row demonstrates how our system is capable of even changing the category of the shoe, traversing from a boot to a sneaker. The third row shows how the system switched from changing the style of shoe to the type of closing mechanism before locating the target image, with the last two rows demonstrating how our system can handle even relatively fine-grained differences between the initial query and the target image.
A good search refinement algorithm need not produce the exact target image in the refinement stage, but simply obtain enough information that the image ranks sufficiently high in the search results. To this end, we provide the rank of the target image per iteration in Figure~\ref{fig:target_ranking}(a) with a comparison to our implementation of the Attribute Pivots approach~\cite{KovashkaICCV2013} which takes advantage of relative attribute annotations. Here we see the two methods perform comparably despite our method using only binary attribute labels. It is important to note that Attribute Pivots produced a more consistent algorithm as exemplified with the lower per-iteration rank standard deviation seen in Figure~\ref{fig:target_ranking}(b), even if the average performance was slightly lower than ours. This may be due to the limited number of relative attributes available in the UT Zap-50K dataset, which suggests it would beneficial to further explore the trade-off between annotation cost vs.\ performance in future work.
\section{Introduction}
In image search applications the user often has the mental picture of their desired content. The ultimate goal of image search is to convey this mental picture to the system and overcome the difference between the lower-level image representation and the higher-level conceptual content. Describing the desired image may be time consuming, however, and an image search system may not find the image even with an accurate description. To remedy this issue, interactive search techniques (\eg ~\cite{caoACMM2010,coxTIP2000,ferecatuICCV2007,FogartyCHI2008,liICME2001,macarthur2000,rasiwaslaTM2007,tongACMMM2001,ZhangACMM2012,Zhao_2017_CVPR,zhouACMMS2003}) obtain user feedback to iteratively refine system returned results, often by asking questions in visual form.
In particular, there has been a recent focus of this type of iterative refinement using relative attribute feedback~\cite{KovashkaICCV2013,KovashkaIJCV2015,ladECCV2014,modiBMVC2017}. As seen in Figure~\ref{fig:feedback}(a), this enables a system to provide targeted feedback, but require relative attribute annotations not typically found in image datasets.
\begin{figure}[t]
\centering
\topinset{\bfseries Examples of User Feedback}{\topinset{\bfseries(a)}{\topinset{\bfseries(b)}{\includegraphics[width=0.48\textwidth,trim=0cm 3.5cm .0cm 0cm, clip]{images/user_feedback}}{.23in}{.55in}}{.23in}{-.53in}}{-.0in}{.21in}
\caption{{\bf Comparison of active feedback types for image search.} Recent work in iterative image search leverages relative attribute annotations enabling the collection of targeted feedback as seen in {\bf (a)}. However, these relative attributes are not typically collected natively in many datasets. Instead, we use the meta-data or categorical labels already present in most datasets to build a representation for our active image search approach. While this relies on more ambiguous feedback since users can define their own similarity criterion, as shown in {\bf (b)}, our experiments show it is sufficient most of the time.}
\label{fig:feedback}
\end{figure}
In this paper, we propose an interactive image search system which doesn't use relative attribute annotations. Instead, we learn an image embedding trained on meta-data labels which are collected natively in e-commerce datasets. These labels identify attributes with clear definitions (\eg does the shirt have long sleeves?), and are also be useful for other tasks such as organizing and filtering a dataset. Relative attributes, by comparison, can be subjective in nature, and annotators may prefer to label as many as 40\% of image pairs as having equal amounts of an attribute~\cite{yuCVPR2014}, making their usefulness beyond active image search unclear. This suggests a trade-off between the low annotation cost of our approach (due to using existing annotations) and expected performance gains from targeted feedback using relative attributes. This paper takes a step towards characterizing the nature of this trade-off.
We begin an interactive image search session by presenting a user with an initial set of candidate images after receiving an initial query. A user simply selects the image which is the most visually similar to their target image (see Figure~\ref{fig:feedback}(b) for an example of the expected feedback). We incorporate the new information provided by the user into our model and then select the next set of images. Thus, the goal is to select the most informative images to present to the user on each iteration. While this style of feedback provides less information than many relative attribute approaches, since each user determines their own similarity criteria resulting in different responses, our experiments show it is ``good enough'' in many cases.
A popular selection criterion is Expected Error Reduction (EER)~\cite{bransonECCV2010,KovashkaICCV2013,KovashkaICCV2011,MacAodhaCVPR2014,MensinkCVPR2011,royICML2001}. This strategy chooses images that provide the largest reduction in the generalization error of the current model, but its high computational cost is disqualifying for many tasks. As such, EER is typically computed on a short list of candidates (\eg exemplars from hierarchical clusters~\cite{KovashkaICCV2013,MacAodhaCVPR2014}). We experiment with two low-cost sampling strategies to obtain a candidate list in this work: a nearest neighbor baseline, which largely ignores user feedback, and a criterion that greedily selects images reflecting the feedback from prior iterations. Figure~\ref{fig:query_refinement_process} contains an overview of our active image search process.
A limitation of hand-crafted criteria like EER is the inability take advantage of the interplay between attributes. For example, knowing the target shirt has a collar also provides some information about the type of shirt being searched for. A good sampling criteria should be able to take advantage of such information as well as adapt to user behavior. Thus, we employ reinforcement learning using a Deep Q-Network~\cite{mnih-atari-2013} to learn to select informative images which we use to refine our list of candidates.
\begin{figure}[t]
\centering
\includegraphics[width=0.35\textwidth,trim=0cm 6.6cm 0cm 0cm, clip]{images/query_refinement_process}
\caption{{\bf Search refinement process.} At test time we are given an initial query as input to our system. On each iteration, we search our database using our ``candidate selector'' strategies to obtain an initial set of candidates. We use a ``candidate re-ranker'' on this set of images using informative, but computationally expensive selection criteria. During the ``user feedback'' step the user indicates if the new refined candidates are more representative of their desired image or not. If they accept a new image, it becomes the query for the next iteration.}
\label{fig:query_refinement_process}
\end{figure}
To learn a good feature representation for our task, we introduce several enhancements to the state-of-the-art Conditional Similarity Network (CSN)~\cite{veitCVPR2017}. This model trains a single network to learn an embedding for multiple attributes. It accomplishes this by learning a masking function for each attribute which selects features in a general representation for an image that is important to separating images in that concept space. This provides multiple views of the images in our database which has proven useful on similar tasks (\eg\cite{heICDM2016,KovashkaICCV2013,KovashkaIJCV2015,modiBMVC2017}) and tends to perform better than training separate embedding models for each concept. Whereas the authors of the CSN model considered the label for each embedding in isolation (\ie an embedding trained for colors would only care if both images were blue), we also factor the overall similarity between two images when training our representation.
The resulting model encourages samples to separate into homogeneous subgroups in each embedding space. Therefore, as we traverse an attribute embedding, \eg heel height, a transition from boot to a stiletto in a single step would be unlikely even if they both have the same sized heel. Combined with constraints which enable us to better exploit our training data, we show significant performance improvements in measuring the similarity between two images with regards to a specific concept. We provide an overview of our approach's parts in Figure~\ref{fig:dqn_structure}.
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth,trim=0 2.5cm 13cm 0,clip]{images/new_figure_idea}
\caption{\textbf{Model Overview.} Our model consists of three major components. First, we train a feature extractor which computes an embedding representation for each image in our database that can be projected into an attribute specific space using a learned mask. These are fed into our Candidate Selectors, which obtains a list of likely candidates. Finally, from these candidates we select the most informative image according to each attribute using a DQN consisting of three fully connected layers followed by a ReLU.}
\label{fig:dqn_structure}
\end{figure}
Our primary contributions are summarized below:
\begin{itemize}
\item We build a system which refines image search results without using the relative attribute annotations or attribute inputs required in prior work.
\item We introduce enhancements to the Conditional Similarity Network which encourages smooth transitions as we traverse the learned embedding space (Section~\ref{sec:embedding}).
\item We propose a Deep-Q Network-based selection criteria instead of hand-crafted methods (Section~\ref{sec:im_sampling_strat}).
\end{itemize}
Our experiments in Section~\ref{sec:experiments} show our image representation reduces attribute matching errors by 2.5-3\% on the UT Zap-50K~\cite{yuCVPR2014} and OSR~\cite{oliviaIJCV2001} datasets while also finding a specific image in a database faster than hand-crafted sampling strategies.
\section{Related Work}
\label{sec:related}
\noindent{\bf Attribute-based interactive image search.} A key difference between this paper and prior work (\eg~\cite{KovashkaIJCV2015}) is that we train our models using annotations which already exist in many datasets. Much of the recent work on this task has focused on how to best utilize models trained using expensive relative attribute annotations (\eg\cite{KovashkaICCV2013,KovashkaIJCV2015,modiBMVC2017}) or requires a user to specify which attributes their target image has (\eg\cite{ZhangACMM2012,Zhao_2017_CVPR}). These assumptions provide attribute specific feedback so the model knows exactly how to alter the current image to make it more like the target image, but are more costly than our approach in terms of annotations, user requirements, or both.
\smallskip
\noindent{\bf Reinforcement learning and active learning.} Recently there has been an ever-growing trend of abandoning hand-crafted approaches in favor of learned models. Training models for selecting informative examples in active learning, however, has primarily focused on how best to combine hand-crafted sampling strategies (\eg\cite{BaramJMLR2004,OsugiICDM2005}). In~\cite{ebert12cvpr}, the authors used reinforcement learning to select which hand-crafted strategy to use on each iteration. This idea was extended in~\cite{Long_2015_ICCV} to select which annotator to use as well as finding informative samples. In contrast, our approach creates an entirely new criterion rather than combining hand-crafted strategies, sharing a similar spirit to some early work in relevance feedback (\eg~\cite{yinICCV2003,Yin:2005:IRF:1083822.1083983}).
\smallskip
\noindent{\bf Relative Attributes.} Prior work in predicting relative attributes include using pairwise supervision to learn linear rankers~\cite{parikhICCV2011}, multi-task learning~\cite{Chen_2014_CVPR}, and fusing binary and relative attribute labels in a model that would make predictions for both types of labels~\cite{Wang_2016_CVPR}. Some works found training local rankers could lead to improved performance~\cite{yuCVPR2014,Yu_2015_ICCV}. Deep networks have also been used to rank attributes~\cite{souriACCV2016} as well as localize them~\cite{singhECCV2016}. However, most of these approaches rely solely on expensive pairwise supervision in order to train their models. An exception is Yu~\etal~\cite{Yu_2017_ICCV} which augmented their annotated pairs with synthetically generated images. In our work, we use labels which are typically found natively in many datasets rather than rely on relative attribute annotations.
\subsection{OSR Experiments}
\label{sec:osr}
To demonstrate our approach's ability to generalize we provide experiments on the Outdoor Scene Recognition (OSR) dataset~\cite{oliviaIJCV2001}. This dataset consists of 2688 images with six attributes annotated for the eight scene categories. We randomly sampled 400 images for the test set (50/category), 160 images for the validation set (20/category), and used the rest for training. To train our representation, we randomly sampled 100K triplets for training, 40K for testing, and 20K for validation. In our active image search experiments we sampled 16,000 pairs for training, 4,000 pairs for testing, and 800 pairs for validation. All other settings are the same that were used for the UT Zap-50K dataset.
\smallskip
\noindent{\bf Results.} As seen in our attribute experiments in Table~\ref{tab:scene_csn_results}, our additional constraints and global similarity enhancements provide a 1.5\% and 2.5\% improvement over the baseline, respectively. Our results on the active image search task in Table~\ref{tab:scene_query_refinement_performance_multispace} also follow the results on UT Zap-50K, where our DQN refinement strategy outperforms the EER alternative as well as the feedback constraint satisfaction and nearest neighbor baselines. Despite the OSR dataset being from a very different domain from UT Zap-50K, our model still provides a performance improvement over prior work.
\begin{table}[t]
\centering
\begin{tabular}{|l|c|}
\hline
Method & Accuracy\\
\hline
CSN & 96.84\\
CSN + constraints & 98.58\\
CSN + constraints + global similarity & \textbf{99.42}\\
\hline
\end{tabular}
\caption{Triplet satisfaction rate on the OSR dataset.}
\label{tab:scene_csn_results}
\end{table}
\begin{table}[t]
\centering
\begin{tabular}{|l|c|}
\hline
Selection Strategy & \#Steps\\
\hline
\hline
Nearest Neighbor & 6.21\\
Feedback Constraint Satisfaction & 5.57\\
Expected Error Reduction & 4.92\\
DQN Refinement & \textbf{4.54}\\
\hline
\end{tabular}
\caption{Active image search performance on the OSR dataset}
\label{tab:scene_query_refinement_performance_multispace}
\end{table} |
\section{Background}
\label{sec:back}
\subsection{Preliminaries - Dialog and Generating Response with Learning and Reasoning}
A {\em dialog} is made up of a series of {\em turns}, where each turn is a series of {\em utterances} by one or more participants playing one or more {\em roles}. We will assume a customer support setting where the roles are {\em customer} and {\em support agent}.
[Add related work on learning dialog policies]
[Add related work on reasoning/ slot filling dialogs]
\subsection{Example - Using Dialogs to Find Information Arranged in a Taxonomy}
One popular category of user dialogs in business setting is inquiry of information that is organized in a taxonomy. For instance, a person wants to ask about a product on a company's site (e.g., {\em pliers} - product, job openings) but does not know which particular type. An example of taxonomy is
United Nations Standard Products and Services Code (UNSPSC\footnote{https://www.unspsc.org/}), which {\em "is an open, global, multi-sector standard for efficient, accurate classification of products and services"}. Table~\ref{tab:unspsc} gives a summary of the levels in the taxonomy and its data. There are 4,302 items arranged into class, family, segment and commodity (lowest level). The {\em characteristic} column conveys the nature of abstraction across levels by reporting number of distinct (dissimilar) values, unique (count=1) values and their percentage, followed by highest frequency at a given level.
So, if the user inquires about a product, the chatbot can refer to the taxonomy to confirm what the user wants. As example, a query on {\em pliers} may refer to a surgical tool (Code: 42292303) or a vehicle servicing equipment (Code: 25191716) or a hand tool (Code: 27112106), and many variations therein. Thus, there can be a varieties strategies for a chatbot to help the user. It can go from highest level to lowest, or reverse, or ask randomly from the tree. Furthermore, if there is a restriction on how many turns the chatbot can take on this query, we want to explore if there are querying strategies that are optimal.
\begin{table}[t!]
\scriptsize
\centering
\begin{tabular}{|l|l|l|}
\hline
\textbf{Level} & \textbf{Name} & \textbf{Characteristic} \\ \hline
\textbf{4 - Highest} & Segment & 36; 3 (0\%); 825 \\ \hline
\textbf{3} & Family & 173; 23 (1\%); 443 \\ \hline
\textbf{2} & Class & 826; 216 (5\%); 49 \\ \hline
\textbf{1 - Lowest} & Commodity & 4302; 4302 (100\%); 1 \\ \hline
\end{tabular}
\caption{Levels in UNSPSC Taxonomy.}
\label{tab:unspsc}
\end{table}
\subsection{Related Work}
Conversational interface is considered especially suitable for feature or preference elicitation for at least two reasons~\cite{brennan1990conversation}. First, it is naturally interactive and thus users are more willing to provide additional information. Moreover, it can pro-actively and dynamically channel the information querying process in desired directions.
Suppose there are $m$ items described using
$n$ features, $F_1$, .., $F_n$. To narrow down users' queries, previous work
has explored asking to confirm (Yes/No) clarification questions such as ``\textit{are you asking about topic X}''~\cite{horvitz1999computational}, and specify values for a feature by direct inquiry (e.g., ``\textit{What VALUE would you like for FEATURE X"}'')
or multi-choice questions (e.g.``\textit{choices for FEATURE X are VALUE 1... VALUE N''})~\cite{chu2000mimic}.
We will assume that the features are binary.
The expected number of features that the user would report in answering a question is $\lambda_{NF}$, the recall rate of the NLP system is $R_{nlp}$, and each unknown feature's information gain is $IG$.
Such a setting can be used to model an online store selling shirts and trousers (i.e., items) to users. The features can be color, fabric and price category. A person searching the store may have preferences about fabric and constraints about color when they are buying items together.
\section{Backup Text - To Remove Later}
In this paper we present the details of the prototype which addresses above-mentioned drawbacks. Below points outline our contributions:
\begin{enumerate}
\item We present a web accessible application which provides delay insights of trains in India (having one of the largest train networks in world) with weeks or months ahead of journey. The predicted delays at in-line stations can quench all of the above-mentioned queries of a consumer.
\item Conversation is a hallmark of intelligence, and consequently, businesses wanting to build AI-based systems to increase productivity and improve customer experience, are interested in automated dialog systems. Therefore along with REST APIs (which enable a machine to communicate with our application), we consider a general chat-bot that can work with learned models and interact with a human.
\item We discuss applicable AI methods like RNNs and LSTMs which can help in predicting more accurate delays and capturing the dynamics of a rail network.
\end{enumerate}
\section{Discussion}
We demonstrated a novel conversation agent using AI methods to help common man gain insights about train delays and work around them. We used a slot-based dialog approach for front-end and a learning-based predictor trained on 2 years of training data. The system maintains user context and supports analytical queries.
The system can be easily extended by enhancing the scope of train data (number, history), the learning models and languages for conversation.
\vspace{-10pt}
\section{Example Illustration}
A sample example is shown in Figure~\ref{fig:example}. The user starts by asking about a train without specifying the location and date for which she is interested in. The agent assumes the intended location is
destination station on its journey and date as current, and responds in natural language with the predicted delay. It also inquires if the user's request is satisfied. The user specifies another location which the agent now satisfies. The user wants to know the first location for delay and the system checks the stations on the journey to know first occurrence of delay.
\begin{figure}
\center{
\fbox{
\begin{minipage}[t]{4in}
{\tiny
\begin{tabbing}
{\bf User:}\= // Turn 1 \\
\> Is train 12307 on time? \\
{\bf Agent:}\= // Answering assuming destination \\
\> Train Number 12307 will be delayed by 159 minutes at \\
\> JU station on 2018-09-21. \\
{\bf User:}\= // Turn 2 \\
\> How about for Varanasi? \\
{\bf Agent:}\= // Handling error condition \\
\> Train 12307 does not stop at Varanasi. \\
\> Here is the list of stations on this route. ...\\
{\bf User:}\= // Turn 3 \\
\> No, I meant for Allahabad. \\
{\bf Agent:}\= // User-specified station \\
\> Train 12307 will be delayed further after station ALD\\
\> on 2018-09-21 by 110.8 minutes \\
{\bf User:}\= // Turn 4 \\
\> What is the average train delay? \\
{\bf Agent:}\= ... // Route analytics \\
\end{tabbing}
}
\end{minipage}
}}
\caption{\label{fig:example} A sample conversation.; video at https://youtu.be/a-ABv29H6XU}
\vspace{-10pt}
\end{figure}
\section{Introduction}
Trains provide an inexpensive form of inter-city and long-distance movement for people and goods in many countries around the world like India. To keep them well maintained and provide good service to travelers, AI is seen as a promising new technology \cite{train-ai}.
However, in developing countries, trains often suffer from endemic delays which can be credited to obsolete technology, large network size and external factors like weather.
Consider train numbers 12305 and 13050 on Indian Railways\footnote{https://etrain.info/in?TRAIN=12305, https://etrain.info/in?TRAIN=13050. Trains are often late by tens of hours or even days in extreme weather.}. Train 12305, Kolkata Rajdhani, is a premium train which runs for nearly 18 hours to cover 1529 km (in 5 states that speak Hindi, Bengali and English), stops at
10 stations and was late over 1 hour at destination for 19\% of the days between 1-Oct-2017 and 1-Sep-2018. Train 13050, Amritsar Howrah Express, runs for nearly 46 hours over 3 days to cover 1922 km (in 7 states that speak Hindi, Bengali, Punjabi and English), stops at 112 stations and was late over 1 hour at destination for 63\% of the days during the same period.
In this context, travelers often want to know when a train may arrive at a station, whether it is delayed and what alternative travel options (like another lesser delayed train) they may have and in their own languages. Currently, people are given delay information about trains from over-crowded public booths by human agents or search-interfaces on websites
in English (which is spoken by a small segment of train-riding population), systems are not aware of a user's context and they
provide delay estimates incrementally which usually increases as time elapses; making re-planning hard.
There has been some recent efforts to understand these delays and predict them after learning from past delay data in India \cite{train-estimation}, Serbia \cite{train-estimation-serbia} and USA \cite{train-estimation-us}. But can these analytic model help travelers and other stakeholders (e.g., businesses serving passengers, people affected by travelers, train managers) who need guidance for their activities?
In response, we propose a conversation interface for common man to inquire about train delay which may further be personalized to user's need \cite{personal-chatbot}. This gives some unique benefits:
(1) {\bf New Temporal insights}: Currently, methods have delay information of on-going journeys but no estimate on the probable delays for journeys in upcoming weeks or months, or past performance. This does not enable travelers to schedule their journey plans efficiently with sufficient time in hand and hence, cause frustration.
(2) {\bf New Journey insights}:
Apart from delays at a certain station, people want to know (a) if a train will be delayed further in journey, (b) the bottleneck station which becomes the root cause of the delay,
and (c) if the delay gets mitigated or not during travel,
(d) the delay characteristics of trains and routes, (e) similarity of trains based on delays they encounter.
(3) {\bf New Interaction Mode}: The conversation interface allows people to access information without knowledge of computer systems which can be personalized to their need and language, and is aware of their traveling context.
\section{System Details}
\begin{figure}
\centering
\includegraphics[width=0.35\textwidth]{figs/chatbot-arch.jpg}
\caption{The architecture of train assistant.}
\label{fig:chatbot-arch}
\vspace{-20pt}
\end{figure}
\noindent {\bf System Architecture:} Figure~\ref{fig:chatbot-arch} shows the working of system at conceptual level.
Here, the user's utterance is analyzed to detect her intent and a
policy for response is selected. This policy may call for predicting a given train's delay at a given station and time, and result is returned which is used by response generator to create a response using templates. The system can select defaults for missing values, get prediction from previously trained models, and decide if it has a reasonable estimate (timely value and enough confidence in an estimate's correctness) to respond.
In principle, the chatbot could have been implemented in any framework since the call to predict delays is a REST call to cloud-based service hosting the learned predictive models. However, in practice, the duration of backend call varies a lot depending on the nature of query and train, and the behavior of different chatbot platforms is not the same. We experimented with Google's DialogFlow and Microsoft's Bot Framework, and eventually chose the latter.
\noindent {\bf Learned Models:} The details of data collection and model training are given in \cite{train-estimation}. We collected delay information of 135 Indian trains over two years from March 2016 to February 2018, and categorized the 135 trains into two groups based on the amount of data collected: 52 \textit{Known Trains} (52KT) for large and 83 \textit{Unknown Trains} (83UT) for small sets. We trained Random Forest Regressor (RFR) and Ridge Regressor (RR) models for each station en-route in 52KT's journey \textit{Known Station}, and learned a late minutes prediction framework which could
generalize to 83UT too.
We conducted exhaustive experiments in multiple settings and chose Confidence Intervals (CI) to articulate aggregated accuracy results. Adopting the popular CI classes of 68\%, 95\% and 99\%, we achieved 28\%, 56\% and 67\% accuracy in the respective classes for 52KT's test data. The trained model used in the demo corresponds to RFR at 99\% CI which correspond to highest accuracy obtained.
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{figs/Figure_1.png}
\caption{The time - accuracy (in RMSE) trade-off for train 13050 with largest number of stations in its journey}
\label{fig:time_acc_toff}
\vspace{-10pt}
\end{figure}
Since the learned models are used by a chatbot which needs to converse in real-time, apart from accuracy, we also consider the response time of models to predict delays for new trips. The performance of RR depends only on the number of stations of the train and took 20-22 seconds to predict late minutes for the longest train we considered - Train 13050 with 112 stations.
RFR was more accurate; for it,
we experimented with the number of stations a train has, the number of trees in the RFR, the time to predict and Root Mean Square Error (RMSE). In Figure~\ref{fig:time_acc_toff}, we show the result for Train 13050. The prediction time increases linearly with size of forest while error reduces sharply.
We find that 5-50 trees RFR give reasonable trade-off and use it in the system.
\vspace{-12pt}
|
\section{Introduction}
Let $\Sigma_0\subset\mathbb{R}^2$ be an embedded, smooth curve, parametrised by the embedding $X_0:\cI\to\mathbb{R}^2$, where $\cI\in\{\mathbb{S}^1,\mathbb{R}\}$.
We seek a one-parameter family of maps $X:\cI\times[0,T)\to\mathbb{R}^2$ with $X(\,\cdot\,,0)=X_0$ satisfying the evolution equation
\begin{align}\label{eq:ccf}
\fracp{X}{t}(p,t)=\big(h(t)-\kappa(p,t)\big)\boldsymbol{\nu}(p,t)
\end{align}
for $(p,t)\in\cI\times(0,T)$, where the vector $\boldsymbol{\nu}$ is the outward pointing unit normal to the curve $\Si_t:=X(\cI,t)$ and $\kappa$ is the curvature function.
The global term $h$ is smooth and smoothly bounded whenever the curvature is bounded.
For the curve shortening flow (CSF), $h\equiv0$.
For closed curves, the enclosed area preserving curve shortening flow (APCSF) has the global term
\begin{align}\label{eq:h_ap}
h(t)=\frac{2\pi}{L_t}\,,
\end{align}
where $L_t=L(\Si_t)$ is the length of the curve.
The length preserving curve flow (LPCF) has the global term
\begin{align}\label{eq:h_lp}
h(t)=\frac1{2\pi}\int_{\Si_t}\kappa^2\,ds_t\,.
\end{align}
The total curvature of a curve $\Si_t=X(\cI,t)$ is given by
\begin{align}\label{eq:def_alpha}
\alpha(t):=\int_{\Si_t}\kappa\,ds_t\,,
\end{align}
where $\alpha=2\pi$ if the curve $\Sigma=X(\mathbb{S}^1)$ is embedded, closed and positively oriented.
For $\cI=\mathbb{R}$, we assume that $\Sigma_0=X_0(\mathbb{R})$ is smoothly asymptotic to two distinct lines $\mathcal{L}_1$ and $\mathcal{L}_2$ for $p\to-\infty$ and $p\to\infty$.
We will need that
\begin{align}\label{eq:alpha}
\alpha(t)\equiv\alpha\in(-\pi,\pi)
\end{align}
and, for a fixed vector $\mathbf{v}\in\mathbb{R}^2$,
\begin{align}\label{eq:asymp}
\lim_{p\to-\infty}\measuredangle(\boldsymbol{\tau}(p,t),\mathbf{v})\equiv\frac{\pi-\alpha}2
\quad\text{ and }\quad
\lim_{q\to\infty}\measuredangle(\boldsymbol{\tau}(q,t),\mathbf{v})\equiv\frac{\pi+\alpha}2
\end{align}
is independent of time, where $\boldsymbol{\tau}$ is the tangent vector in direction of the parametrisation, as well as
\begin{align}\label{eq:intabskappa}
\int_{\Sigma_0}|\kappa|\,ds<\infty\,.
\end{align} \\
The APCSF was first studied by Gage~\cite{Gage86}.
He proved that initially embedded, closed, convex curves stay embedded, smooth and convex, and converge smoothly to a circle of radius $\sqrt{A_0/\pi}$, where $A_0=A(\Sigma_0)$ is the enclosed area of the initial curve.
In~\cite{Maeder15}, Maeder-Baumdicker studied APCSF for convex curves with Neumann boundary on a convex support curve and showed smooth convergence to an arc for sufficiently short, convex, embedded initial curves.
She proved a monotonicity formula and excluded type-I singularities for embedded, convex curves under the APCSF.
For the LPCF, Pihan~\cite{Pihan98} showed that initially embedded, closed, convex curves stay embedded, smooth and convex, and converge smoothly and exponentially to a circle of radius $L_0/2\pi$. \\
In this paper, we will adapt theory from CSF.
For CSF in the plane, Gage--Hamilton and Grayson~\cite{GageHamilton86,Grayson87} showed that all embedded, closed initial curves stay embedded until they smoothly and exponentially shrink to a round point.
In~\cite{Huisken95}, Huisken gave a different proof for this result by bounding the ratio of the extrinsic distance
$$d(p,q,t):=\Vert X(q,t)-X(p,t)\Vert_{\mathbb{R}^2}$$
and the intrinsic distance
$$l(p,q,t):=\int_p^qds_t$$
for curves $\Si_t=X(\mathbb{R},t)$ with asymptotic ends, respectively the extrinsic distance and the function
$$\psi(p,q,t):=\frac{L_t}\pi\sin\!\left(\frac{\pi\,l(p,q,t)}{L_t}\right)$$
for curves $\Si_t=X(\mathbb{S}^1,t)$, below away from zero, and by applying singularity theory for CSF.
In~\cite{AndrewsBryan11}, Andrews and Bryan found an explicit function to proof curvature bounds via the distance comparison principle.
To analyse curvature blow-ups, one distinguishes between type-I and type-II singularities and rescales the curve near a point of highest curvature.
Using his famous monotonicity formula in~\cite{Huisken90}, Huisken showed that if an immersed curve develops a type-I singularity under CSF, the curves $\Si_t$ have to be asymptotic to a homothetically shrinking solution around the singular point.
Abresch and Langer~\cite{AbreschLanger86} had previously classified all embedded, homothetically shrinking solutions of CSF as circles.
One concludes, in case of a type-I singularity, that the curves shrink to a round point.
For the type-II singularities, Hamilton~\cite{Hamilton89} and Altschuler~\cite{Altschuler91} showed that each rescaling sequence converges to a translating solution.
For curves in the plane, the only solution of this kind is the so-called grim reaper which is, for all $\tau\in\mathbb{R}$, given by the graph of the function $u(\sigma,\tau)=\tau-\log\cos(\sigma)$, where $\sigma\in(-\pi/2,\pi/2)$.
On the grim reaper $\inf(d/l)=0$, so that type-II singularities can be excluded.
Since $T<\infty$ and a singularity has to form, it has to be of type~I. \\
This paper is structured as follows.
In Section~\ref{sec:evo}, we state evolution equations for the geometric quantities under~\eqref{eq:ccf} and draw first conclusions.
In Section~\ref{sec:theta2}, we consider angles of tangent vectors and derive a strong maximum principle for the local total curvature
\begin{align}\label{eq:intkappa}
\theta(p,q,t):=\int_p^q\kappa\,ds_t\,.
\end{align}
In the subsequent sections, we study the flow~\eqref{eq:ccf} for embedded, positively oriented, smooth initial curves $\Sigma_0=X_0(\cI)$ with
\begin{align}\label{eq:intkappageqminuspi}
\theta_0(p,q)=\int_p^q\kappa\,ds\geq-\pi
\end{align}
for all $p,q\in\cI$.
Note that for convex curves $\theta_0\geq0$.
Figure~\ref{fig:example} is an example for condition~\eqref{eq:intkappageqminuspi}, where all the angles lay between $-\pi$ and $3\pi$, e.\,g. $\theta(p,q)=-\pi$, $\theta(q,p)=3\pi$, $\theta(q,r)=2\pi$, $\theta(r,q)=0$, $\theta(r,p)=\pi$.
\begin{figure}
\centering
\begin{minipage}{.35\textwidth}
\centering
\includegraphics[width=\textwidth]{fig_example}
\captionsetup{width=.9\textwidth}
\caption{Example for condition~\eqref{eq:intkappageqminuspi}}
\label{fig:example}
\end{minipage}
\hfill
\begin{minipage}{.60\textwidth}
\centering
\includegraphics[width=\textwidth]{fig_cexample}
\caption{Counterexample~\ref{ex:cexample}}
\label{fig:cexample}
\end{minipage}
\end{figure}\\
In Sections~\ref{sec:noncompact} and~\ref{sec:noncollapsing}, we modify the distance comparison principles from~\cite{Huisken95} and prove that, for
\begin{align}\label{eq:h_1}
h(t)\in
\begin{dcases}
[0,\infty)&\qquad\text{ for }\cI=\mathbb{R} \\
\left[0,\frac1{2\pi}\int_{\Si_t}\kappa^2\,ds_t+\frac{2\pi}{L_t}\right]&
\qquad\text{ for }\cI=\mathbb{S}^1
\end{dcases}
\end{align}
and if the initial embedding $\Sigma_0$ satisfies~\eqref{eq:intkappageqminuspi}, the ratio $d/l$ for $\cI=\mathbb{R}$ and $d/\psi$ for $\cI=\mathbb{S}^1$ is bounded from below away from zero uniformly in time.
We conclude that the curves $\Si_t$ stay embedded for all $t\in[0,T)$.
We also show that the condition~\eqref{eq:intkappageqminuspi} is sharp, that is, one can construct initial curves which violate~\eqref{eq:intkappageqminuspi} arbitrarily mildly and for which the resulting flow self-intersects in finite time.
An example is the initial curve in Figure~\ref{fig:cexample} with length sufficiently large compared to the $C^{3,\alpha}$-norm of its embedding and for which $\min_{\cI\times\cI}\theta_0<-\pi$, e.\,g. $\theta(p_1,p_2)<-\pi$. \\
In Section~\ref{sec:singana}, we assume that $T<\infty$ and there exist constants $0<c,C<\infty$ so that
\begin{align}\label{eq:h_2}
L_t\geq c
\qquad\text{ and }\qquad
0\leq h(t)\leq C
\end{align}
for $t\in[0,T)$ and study curvature blow ups via parabolic rescaling.
We use the distance comparison principles from Sections~\ref{sec:noncompact} and~\ref{sec:noncollapsing} in the same fashion as for CSF in~\cite{Huisken95} to exclude type-II singularities and conclude that the flow exists for all positive times. \\
In Section~\ref{sec:convexity}, we assume that a solution is immortal, that is, it exists for all positive times and the global term satisfies the following.
Let $\delta\in(0,\infty)$ be given so that $\delta A_0$ is the desired limit area and
\begin{align}\label{eq:h_3a}
h(t)=(1-\gamma)\frac{2\pi}{L_t}+\gamma\frac1{2\pi}\int_{\Si_t}\kappa^2\,ds_t\,,
\quad\text{ where }\quad
\gamma=\frac{(\delta-1)A_0}{L_0^2/4\pi-A_0}\,.
\end{align}
We prove that the curves become convex in finite time.
The global term above ensures that the enclosed area is bounded away from zero and the length is bounded away from infinity throughout the flow.
In Section~\ref{sec:longtimebehaviour}, we assume that an immortal solution of~\eqref{eq:ccf} with $h$ satisfying~\eqref{eq:h_3a} is convex.
We expand Gage's and Pihan's results and show smooth and exponential convergence to a round circle. \\
Note that the global term~\eqref{eq:h_ap} of the APCSF satisfies conditions~\eqref{eq:h_1},~\eqref{eq:h_2} and~\eqref{eq:h_3a}.
The global term~\eqref{eq:h_lp} of the LPCF satisfies a priori~\eqref{eq:h_1} and~\eqref{eq:h_3a}. \\
This paper extends results from the author's PhD thesis~\cite{Dittberner18}.\\
\textbf{Acknowledgements.}
The author wants to thank Klaus Ecker, Theodora Bourni, Mat Langford and Steven Lynch for interest in the work and helpful discussions.
\section{Evolution equations and first consequences}\label{sec:evo}
Let $\cI\in\{\mathbb{S}^1\!,\mathbb{R}\}$ and $X:\cI\to\mathbb{R}^2$ be a smooth, embedded curve with length element $v:\cI\to\mathbb{R}$ by $v(p):=\big\Vert\fracd{}{p}X(p)\big\Vert$.
For a fixed point $p_0\in\cI$, the arc length parameter $s:\cI\to[0,L]$ is given by $s(p):=\int_{p_0}^pv(r)\,dr$, so that $ds=vdp$ and $\fracd{}{s}=\frac1{v}\fracd{}{p}$.
For $\Sigma=X(\mathbb{S}^1)$, the arc length parameter is given by $s:\mathbb{S}^1\to\mathbb{S}^1_{L/2\pi}$ and $\tilde X:=X\circ s^{-1}:\mathbb{S}^1_{L/2\pi}\to\mathbb{R}^2$ parametrises $\Sigma$ by arc length.
For $\Sigma=X(\mathbb{R})$, the arc length parameter is given by $s:\mathbb{R}\to\mathbb{R}$.
The unit tangent vector field $\boldsymbol{\tau}$ to $\Sigma$ in direction of the arc length parametrisation is given by $\boldsymbol{\tau}:=\fracd{}{s}\tilde X$.
The outward unit normal is given by $\boldsymbol{\nu}:=(\boldsymbol{\tau}_2,-\boldsymbol{\tau}_1)$.
We define the curvature by
$$\kappa:=-\left\langle\fracd{\boldsymbol{\tau}}{s},\boldsymbol{\nu}\right\rangle=\left\langle\boldsymbol{\tau},\fracd{\boldsymbol{\nu}}{s}\right\rangle$$
and the curvature vector by $\boldsymbol\kappa:=-\kappa\boldsymbol{\nu}$.
The Frenet--Serret equations read as
$$\fracd{\boldsymbol{\tau}}{s}=-\kappa\boldsymbol{\nu}\qquad\text{ and }\qquad
\fracd{\boldsymbol{\nu}}{s}=\kappa\boldsymbol{\tau}\,.$$
Let $X:\cI\!\times[0,T)\to\mathbb{R}^2$ be a one parameter family of maps.
For fixed $t\in[0,T)$, we can parametrise $\Si_t=X(\cI,t)$ by arc length via the arc length parameter $s(\,\cdot\,,t)$, where $s(\cI,t)\in\{\mathbb{S}^1_{L_t/2\pi},\mathbb{R}\}$ and the arc length parametrisation is given by $\tilde X(\,\cdot\,,t)=X(\,\cdot\,,t)\circ s^{-1}(\,\cdot\,,t):s(\cI,t)\to\mathbb{R}^2$.
The evolution equation~\eqref{eq:ccf} applied to the arc length parametrisation reads
$$\fracp{\tilde X}{t}(s,t)=\fracp{^2\tilde X}{s^2}(s,t)+h(t)\tilde\boldsymbol{\nu}(s,t)$$
for $s\in s(\cI,t)$, where $\tilde\boldsymbol{\nu}(s,t)=\tilde\boldsymbol{\nu}(s(p,t),t)=\boldsymbol{\nu}(p,t)$ and we used the identity $\Delta_{\Sigma}\tilde X=\fracd2{s^2}\tilde X=\boldsymbol\kappa$ for the curvature vector.
Whenever we will calculate via the arc length parametrisation, we will do so at a fixed time.
Since the images $X(\cI,t)=\tilde X(s(\cI,t),t)$ are the same and $X$ and $\tilde X$ only differ by a tangential diffeomorphism, we will omit the ``$\sim$'' in the following above geometric quantities related to $\tilde X$ if these depend on $s$ rather than $p$.
\begin{Thm}[Short time existence, Huisken~{\cite[p.~36]{Huisken87}} for the APCSF, Pihan~{\cite[Thms.~4.3 and Corollary~4.4]{Pihan98}} for the LPCF]\label{thm:ste}
For $\alpha\in(0,1)$ and $k\geq3$, let $\Sigma_0$ be an embedded, closed $C^{k,\alpha}$-curve in $\mathbb{R}^2$, parametrised by a $C^{k,\alpha}$-embedding $X_0:\mathbb{S}^1\to\mathbb{R}^2$.
Then there exists a time $T=T(\Vert X_0\Vert_{C^{3,\alpha}})>0$ such that the initial value problem~\eqref{eq:ccf} with $h(0)>0$ has a unique solution $X\in C^{k,\alpha;\lfloor k/2\rfloor,\alpha/2}(\mathbb{S}^1\times(0,T))$.
In particular, if $\Sigma_0$ is smooth so is $\Si_t$ for all $t\in(0,T)$.
\end{Thm}
\begin{Rem}
There is yet no result for short time existence for $\cI=\mathbb{R}$ or $h(0)<0$.
In~\cite{Pihan98}, Pihan considers the rescaled PDE,
$$\fracp{X}{t}=\left(1-\frac\kappa h\right)\boldsymbol{\nu}
=\frac1h\fracp{^2X}{s^2}+\boldsymbol{\nu}$$
which is why the positivity of $h$ is needed.
In this paper, we will assume short time existence for $\cI=\mathbb{R}$ or $h(0)<0$.
\end{Rem}
\begin{Lemma}[\hspace{-.005em}{\cite{Gage86}}]\label{lem:evolutionequations}
Let $X:\cI\times(0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf}.
Then, for $t\in(0,T)$,
\begin{alignat*}{2}
\fracp{v}{t}&=\kappa(h-\kappa)v\,,\qquad
\fracp{}{t}\fracp{}{s}&&=\fracp{}{s}\fracp{}{t}-\kappa(h-\kappa)\fracp{}{s}\,, \\
\fracp{\boldsymbol{\tau}}{t}&=-\fracp{\kappa}{s}\boldsymbol{\nu}\,,\qquad\qquad
\fracp{\boldsymbol{\nu}}{t}&&=\fracp{\kappa}{s}\boldsymbol{\tau}\,,\qquad\qquad
\fracp{\kappa}{t}=\fracp{^2\kappa}{s^2}-(h-\kappa)\kappa^2\,.
\end{alignat*}
\end{Lemma}
\begin{Cor}[\hspace{-.005em}{\cite[Thm.~1.3]{Huisken87}}]\label{cor:strongmaxpkappa}
Let $X:\mathbb{S}^1\!\times[0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} and let $\kappa\geq0$ on $\Sigma_0$.
Then $\kappa>0$ on $\Si_t$ for all $t\in(0,T)$.
\end{Cor}
\begin{Lemma}[\hspace{-.005em}{\cite{Gage86}}]\label{lem:dtL}
Let $X:\mathbb{S}^1\!\times(0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf}.
Then, for $t\in(0,T)$,
$$\fracd{A}{t}=hL-2\pi
\qquad\text{ and }\qquad
\fracd{L}{t}=2\pi h-\int_{\Si_t}\kappa^2\,ds_t\,.$$
\end{Lemma}
\begin{Prop}\label{prop:T<infty}
Let $X:\cI\times[0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$.
If $T<\infty$, then $\max_{p\in\cI}|\kappa(p,t)|\to\infty$ for $t\to T$.
\end{Prop}
\begin{proof}
Like in~\cite[Section~6.3]{Pihan98} (see also~\cite[Chapter~4]{Dittberner18}), we can bound the derivatives of the curvature in terms of the curvature as long as the curvature is bounded.
The proposition then follows like in~\cite[Thm.~8.1]{Huisken84}.
\end{proof}
\section{Angles and local total curvature}\label{sec:theta2}
We want to exploit the relationship between angles of tangent vectors and local total curvatures and prove a strong maximum principle for the latter. \\
Define $\vartheta:\cI\times[0,T)\to[0,2\pi)$ to be the angle between the $x_1$-axis and the tangent vector, so that
$$\vartheta(p,t)=\begin{cases}
\arccos(\langle\mathbf{e}_1,\boldsymbol{\tau}(p,t)\rangle)&\qquad\text{ if }\langle\mathbf{e}_2,\boldsymbol{\tau}(p,t)\rangle\geq0 \\
2\pi-\arccos(\langle\mathbf{e}_1,\boldsymbol{\tau}(p,t)\rangle)&\qquad\text{ if }\langle\mathbf{e}_2,\boldsymbol{\tau}(p,t)\rangle\leq0\,.
\end{cases}$$
Since $\boldsymbol{\nu}=(\boldsymbol{\tau}_2,-\boldsymbol{\tau}_1)$,
\begin{align}\label{eq:defvartheta}
\cos(\vartheta)=\langle\mathbf{e}_1,\boldsymbol{\tau}\rangle=-\langle\mathbf{e}_2,\boldsymbol{\nu}\rangle
\quad\text{ and }\quad
\sin(\vartheta)=\langle\mathbf{e}_2,\boldsymbol{\tau}\rangle=\langle\mathbf{e}_1,\boldsymbol{\nu}\rangle\,.
\end{align}
For a fixed time $t\in[0,T)$, we can define the angle $\tilde\vartheta$ via the arc length parameter by $\tilde\vartheta:s(\cI,t)\to[0,2\pi)$.
As explained earlier, we can omit the ``$\sim$'' for simplicity.
\begin{Lemma}[\hspace{-.005em}{\cite[Lem.~3.1.5]{GageHamilton86}}]\label{lem:dsvartheta}
Let $X:\cI\times[0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$.
Then
$$\fracp{\vartheta}{s}=\kappa
\qquad\text{ and }\qquad
\fracp{\vartheta}{t}=\fracp{\kappa}{s}
\quad\text{ for }\;
t\in(0,T).$$
\end{Lemma}
Like in~\eqref{eq:intkappa}, we define the total local curvature $\theta:\cI\times\cI\times[0,T)\to\mathbb{R}$ by
\begin{align}\label{eq:deftheta}
\theta(p,q.t):=\int_p^q\kappa(r,t)\,ds_t\,,
\end{align}
where we integrate in direction of the parametrisation.
The total curvature $\alpha(t)$ is given by the full integral over the curvature as stated in~\eqref{eq:def_alpha}.
For $\cI=\mathbb{S}^1$ and $p,q\in[0,2\pi)$, we set
\begin{align*}
\theta(p,q.t)=
\begin{dcases}
\int_p^q\kappa\,ds_t&\qquad\text{ if }p\leq q\\
\int_p^{2\pi}\kappa\,ds_t+\int_0^q\kappa\,ds_t&\qquad\text{ if }q<p\,.
\end{dcases}
\end{align*}
Then
\begin{align}\label{eq:2pithetatheta}
2\pi=\theta(p,q,t)+\theta(q,p,t)
\end{align}
for all $p,q\in\mathbb{S}^1$.
For $\cI=\mathbb{R}$ and $p<q$, we set $\theta(q,p,t)=-\theta(p,q,t)$.
By Lemma~\ref{lem:dsvartheta},
\begin{align}\label{eq:thetavartheta}
\theta(p,q,t)
=\int_p^q\kappa\,ds_t
=\int_p^q\frac1v\fracp{\vartheta}{r}\,vdr
=\vartheta(q,t)-\vartheta(p,t)+2\pi\omega\,,
\end{align}
where $\omega(p,q,t)\in\mathbb{Z}$ is the local winding number.
Hence, $\theta$ is the angle between the tangent vectors at two points $X(p,t)$ and $X(q,t)$ modulo the local winding number.
If a curve $\Sigma=X(\cI)$ is embedded and convex, then
\begin{align}\label{eq:0theta2pi}
0\leq\theta(p,q)<\alpha
\end{align}
for all $p,q\in\cI$.
\begin{Lemma}\label{lem:minmaxtheta}
Let $\Sigma=X(\mathbb{S}^1)$ be an embedded, closed curve.
Then
$$\sup_{\mathbb{S}^1\!\times\mathbb{S}^1}\theta=2\pi-\min_{\mathbb{S}^1\!\times\mathbb{S}^1}\theta\,.$$
\end{Lemma}
\begin{proof}
If $\Sigma$ is convex, the claim follows directly from~\eqref{eq:0theta2pi}.
For non-convex curves, we have $\min_{\mathbb{S}^1\!\times\mathbb{S}^1}\theta<0$ and $\max_{\mathbb{S}^1\!\times\mathbb{S}^1}\theta>2\pi$.
Let the maximum of $\theta$ be attained at $p_0,q_0\in\mathbb{S}^1$, that is, by~\eqref{eq:2pithetatheta},
\begin{align}\label{eq:2pithetatheta_2}
\max_{\mathbb{S}^1\!\times\mathbb{S}^1}\theta=\theta(p_0,q_0)=2\pi-\theta(q_0,p_0)\,.
\end{align}
Then, for all $p,q\in\mathbb{S}^1$, $p\ne q$, by~\eqref{eq:2pithetatheta} and~\eqref{eq:2pithetatheta_2},
$$2\pi-\theta(q_0,p_0)
=\theta(p_0,q_0)
\geq\theta(q,p)
=2\pi-\theta(p,q)\,.$$
Consequently, $\theta(q_0,p_0)\leq\theta(p,q)$ for all $p,q\in\mathbb{S}^1$, $p\ne q$, which implies with~\eqref{eq:2pithetatheta_2},
\[\min_{\mathbb{S}^1\!\times\mathbb{S}^1}\theta=\theta(q_0,p_0)
=2\pi-\max_{\mathbb{S}^1\!\times\mathbb{S}^1}\theta\,.\qedhere\]
\end{proof}
\begin{Lemma}\label{lem:minmaxthetaentire}
Let $\Sigma=X(\mathbb{R})$.
Then
$$\sup_{\mathbb{R}\times\mathbb{R}}\theta\leq\alpha-2\inf_{\mathbb{R}\times\mathbb{R}}\theta\,.$$
\end{Lemma}
\begin{proof}
For $p,q\in\mathbb{R}$, $p<q$,
\[\alpha
=\int_{-\infty}^p\kappa\,ds+\int_p^q\kappa\,ds+\int_q^\infty\kappa\,ds
\geq2\inf_{\mathbb{R}\times\mathbb{R}}\theta+\theta(p,q)\,.\qedhere\]
\end{proof}
For $t\in[0,T)$, we define
$$\theta_{\min}(t):=\min_{(p,q)\in\mathbb{S}^1\!\times\mathbb{S}^1}\theta(p,q,t)\leq0$$
and
$$\theta_{\inf}(t):=\inf_{(p,q)\in\mathbb{R}\times\mathbb{R}}\theta(p,q,t)\leq\alpha\,.$$
\begin{Thm}\label{thm:dtdstheta}
Let $X:\cI\times(0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf}.
Then
$$\left(\fracp{}{t}-\Delta_{\Si_t}\right)\theta(p,q,t)=0$$
for all $p,q\in\cI$ and $t\in(0,T)$.
Moreover, let $t_0\in(0,T)$.
\begin{enumerate}[(i)]
\item For $\cI=\mathbb{S}^1$, suppose $\theta_{\min}(t_0)<0$, then $\theta_{\min}(t_0)<\theta_{\min}(t)$ for all $t\in(t_0,T)$.
\item For $\cI=\mathbb{R}$, let~\eqref{eq:alpha} and~\eqref{eq:asymp} be satisfied.
Suppose $\theta_{\inf}(t_0)<\min\{0,\alpha\}$, then $\theta_{\inf}(t_0)<\theta_{\inf}(t)$ for all $t\in(t_0,T)$.
\end{enumerate}
\end{Thm}
\begin{proof}
We differentiate $\theta$ at $p,q\in\cI$ and use~\eqref{eq:thetavartheta} and Lemma~\ref{lem:dsvartheta} to obtain
$$\boldsymbol{\tau}_p(\theta)
=-\boldsymbol{\tau}_p(\vartheta_p)
=-\kappa_p
\qquad\text{ and }\qquad
\boldsymbol{\tau}_q(\theta)
=\boldsymbol{\tau}_q(\vartheta_q)
=\kappa_q$$
as well as
\begin{align}\label{eq:ds2theta}
\Delta_{\Si_t}\theta
=\boldsymbol{\tau}_p^2(\theta)+\boldsymbol{\tau}_q^2(\theta)
=\boldsymbol{\tau}_q(\kappa_q)-\boldsymbol{\tau}_p(\kappa_p)
=\fracp{\vartheta_q}{t}-\fracp{\vartheta_p}{t}
=\fracp{\theta}{t}\,.
\end{align}
Since $X$ is smooth, all spatial and time derivatives of $\theta$ are smooth in $\cI\times\cI\times[0,T)$.
Let $\cI=\mathbb{S}^1$.
For fixed $p\in\mathbb{S}^1$ and $t\in[0,T)$,
$$\lim_{q\searrow p}\theta(p,q,t)
=\lim_{q\searrow p}\int_p^q\kappa\,vdr
=0$$
and
$$\lim_{q\nearrow p}\theta(p,q,t)
=\lim_{q\nearrow p}\int_0^q\kappa\,vdr+\int_p^{2\pi}\kappa\,vdr
=\int_{\mathbb{S}^1}\kappa\,vdr
=2\pi\,.$$
Hence, $\theta$ is discontinuous along the diagonal $\{p=q\}\subset\mathbb{S}^1\!\times\mathbb{S}^1$.
The set $S:=\mathbb{S}^1\!\times\mathbb{S}^1\setminus\{p=q\}$ is an oriented cylinder.
The closure $\bar S$ has two boundaries
$$(\partial S)_-=\left\{(p,p)\,\big|\, p\in\mathbb{S}^1\right\}\quad\text{ and }\quad
(\partial S)_+=\left\{\left(\lim_{r\nearrow p}r,p\right)\,\bigg|\, p\in\mathbb{S}^1\right\}\,,$$
where $\theta\equiv0$ on $(\partial S)_-\times[0,T)$ and $\theta\equiv2\pi$ on $(\partial S)_+\times[0,T)$.
The claim~(i) now follows from the strong maximum principle with boundary conditions.
Let $\cI=\mathbb{R}$.
For $p\in\mathbb{R}$, define
$$\theta^-(p,t):=\lim_{q\to-\infty}\theta(q,p,t)
\quad\text{ and }\quad
\theta^+(p,t):=\lim_{q\to\infty}\theta(p,q,t)\,.$$
If $\theta_{\inf}<\min\{0,\alpha\}$, $\theta_{\inf}$ it either equals $\theta_{\inf}^+$ or $\theta_{\inf}^-$, or is attained at a local minimum.
Suppose $\theta_{\inf}<0$ is attained at a local minimum.
We have that $\theta\equiv0$ on $\{p=q\}\times[0,T)=\partial\{p\ne q\}\times[0,T)$.
Let $t\in[0,T)$ and $\theta$ attain a local minimum at points $p,q\in\mathbb{R}$, $p<q$, with $\theta(p,q,t)<0$.
Let $p$, $q$ be so that there exists $\varepsilon>0$ with
$$\theta(r,s,t)>\theta(p,q,t)=\min_{[p-\varepsilon,q+\varepsilon]}\theta(\,\cdot\,,\,\cdot\,,t)$$
for all $r\in[p-\varepsilon,p)$ and $s\in(q,q+\varepsilon]$.
By smoothness of the flow, there exists $\delta>0$ so that
$$\theta(p-\varepsilon,q+\varepsilon,\tau)>\theta(p,q,t)$$
for all $\tau\in[t,t+\delta]$.
The strong maximum principle yields that
$$\theta(r,s,\tau)>\theta(p,q,t)$$
for all $r,s\in[p-\varepsilon,q+\varepsilon]$, $r<s$, and $\tau\in(t,t+\delta]$.
Hence, every negative local minimum of $\theta$ is strictly decreasing.
Suppose $\theta_{\inf}(t)<\min\{0,\alpha\}$ equals $\theta_{\inf}^+$ or $\theta_{\inf}^-$.
By~\eqref{eq:alpha}, $\alpha$ is constant.
By~\eqref{eq:asymp}, $\lim_{q\to\pm\infty}\vartheta_q$ is constant in time.
Hence, by~\eqref{eq:ds2theta},
$$\left(\fracp{}{t}-\Delta_{\Si_t}\right)\theta^{\pm}=0\,.$$
If $\theta_{\inf}^{\pm}<\min\{0,\alpha\}$, then $\theta_{\inf}^{\pm}$ is attained at a point $p\in\mathbb{R}$.
The strong maximum principle yields that $\theta_{\inf}^{\pm}$ is strictly increasing as long as $\theta_{\inf}^{\pm}<\min\{0,\alpha\}$.
\end{proof}
We define the extrinsic distance $d:\cI\times\cI\times[0,T)\to\mathbb{R}$ by
$$d(p,q,t):=\Vert X(q,t)-X(p,t)\Vert_{\mathbb{R}^2}$$
and the vector $\mathbf{w}:\big(\cI\times\cI\times[0,T)\big)\setminus\{d=0\}\to\mathbb{R}^2$ by
$$\mathbf{w}(p,q,t):=\frac{X(q,t)-X(p,t)}{d(p,q,t)}\,.$$
\begin{Lemma}\label{lem:thetabeta}
Let $\Sigma=X(\cI)$ be an embedded curve and $p,q\in\cI$ with $d(p,q)\ne0$.
Let $\langle\mathbf{w},\boldsymbol{\tau}_p\rangle=\langle\mathbf{w},\boldsymbol{\tau}_q\rangle=\cos(\beta/2)$ for $\beta\in[0,\pi]$.
Then either
\begin{enumerate}[(i)]
\item $\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=-\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=-\sin(\beta/2)$ and $\theta(p,q)=2\pi k+\beta$,
\item $\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=-\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=\sin(\beta/2)$ and $\theta(p,q)=2\pi k-\beta$, or
\item $\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=\pm\sin(\beta/2)$ and $\theta(p,q)=2\pi k$
\end{enumerate}
for $k\in\mathbb{Z}$.
\end{Lemma}
\begin{proof}
The angles are invariant under rotations in the plane, thus we may assume $\mathbf{w}=\mathbf{e}_1$.
Since $\beta\in[0,\pi]$, the definition~\eqref{eq:defvartheta} of $\vartheta\in[0,2\pi)$ yields
$$\cos(\vartheta_p)=\langle\mathbf{e}_1,\boldsymbol{\tau}_p\rangle=\cos\!\left(\frac\beta2\right)\geq0
\quad\text{ and }\quad
\sin(\vartheta_p)=\langle\mathbf{e}_1,\boldsymbol{\nu}_p\rangle=\pm\sin\!\left(\frac\beta2\right)$$
as well as
$$\cos(\vartheta_q)=\langle\mathbf{e}_1,\boldsymbol{\tau}_q\rangle=\cos\!\left(\frac\beta2\right)\geq0
\quad\text{ and }\quad
\sin(\vartheta_q)=\langle\mathbf{e}_1,\boldsymbol{\nu}_q\rangle=\pm\sin\!\left(\frac\beta2\right)\,.$$
Hence, by~\eqref{eq:thetavartheta},
\begin{align}\label{eq:theta_2}
\vartheta_p,\vartheta_q\in\left\{\frac\beta2,2\pi-\frac\beta2\right\}\quad\text{ and }\quad
\theta=2\pi\omega+\vartheta_q-\vartheta_p\,,
\end{align}
where $\omega\in\mathbb{Z}$. \\
(i) Assume that $\sin(\vartheta_p)=-\sin(\beta/2)\leq0$ and $\sin(\vartheta_q)=\sin(\beta/2)\geq0$.
From~\eqref{eq:theta_2} it follows that $\vartheta_p=2\pi-\beta/2$, $\vartheta_q=\beta/2$ and $\theta=2\pi(\omega+1)+\beta$.\\
(ii) Assume that $\sin(\vartheta_p)=\sin(\beta/2)\geq0$ and $\sin(\vartheta_q)=-\sin(\beta/2)\leq0$.
From~\eqref{eq:theta_2} it follows that $\vartheta_p=\beta/2$, $\vartheta_q=2\pi-\beta/2$ and $\theta=2\pi(\omega+1)-\beta$.\\
(iii) Assume that $\sin(\vartheta_p)=\pm\sin(\beta/2)$ and $\sin(\vartheta_q)=\pm\sin(\beta/2)$.
From~\eqref{eq:theta_2} it follows that either $\vartheta_p=\vartheta_q=\beta/2$ or $\vartheta_p=\vartheta_q=2\pi-\beta/2$ and thus $\theta=2\pi\omega$.
\end{proof}
\section{Distance comparison principle for non\-compact curves}\label{sec:noncompact}
We adapt the methods from Huisken~\cite{Huisken95} to obtain estimates that imply a certain non-collapsing behaviour of the evolving curves.\\
The intrinsic distance $l:\cI\times\cI\times[0,T)\to\mathbb{R}$ is given by
$$l(p,q,t):=\int_p^qv(r,t)\,dr\,.$$
We set $d/l\equiv1$ on $\{p=q\}\times[0,T)$, then $d/l\in C^0(\mathbb{R}\times\mathbb{R}\times[0,T))$.
\begin{Rem}\label{rem:dl}
Embedded curves satisfy $(d/l)(p,q)>0$ for all $p,q\in\mathbb{R}$.
If a curve is not a line, then there exist $p,q\in\mathbb{R}$ so that $d(p,q)<l(p,q)$ and thus $\inf_{\mathbb{R}\times\mathbb{R}}(d/l)<1$.
\end{Rem}
\begin{Lemma}
\label{lem:mindltheta2pik}
Let $\Sigma=X(\mathbb{R})$ be an embedded curve.
Let $p,q\in\mathbb{R}$, $p\ne q$, such that $\Sigma$ crosses the connecting line between $X(p)$ and $X(q)$ at $X(r)$ with $r\notin[p,q]$.
Then $(d/l)(p,q)$ cannot be the infimum.
\end{Lemma}
\begin{proof}
Let $X(\mathbb{R})$ cross the connecting line between $X(p)$ and $X(q)$ at $X(r)$.
Then $X(r)=X(p)+\mathbf{w}(p,q)\Vert X(r)-X(p)\Vert$.
Set $d:=d(p,q)$, $d_1:=d(p,r)$ and $d_2:=d(r,q)$.
Then $d=d_1+d_2$.
Furthermore, set $l:=l(p,q)$, $l_1:=l(p,r)$ and $l_2:=l(r,q)$.
If $p<q<r$, then $l<l_1$ and $d/l>d_1/l_1$.
If $r<p<q$, then $l<l_2$ and $d/l>d_2/l_2$.
Thus, $d/l$ cannot be a global minimum.
\end{proof}
Now we can prove a similar result to~\cite[Thm.~2.1]{Huisken95}.
\begin{Thm}
\label{thm:minimumdl}
Let $\Sigma_0=X_0(\mathbb{R})$ be a smooth, embedded curve satisfying~\eqref{eq:intkappageqminuspi}.
Let $X:\mathbb{R}\times[0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} satisfying~\eqref{eq:alpha},~\eqref{eq:asymp} and~\eqref{eq:h_1}, and with initial curve $\Sigma_0$.
Then there exists a constant $c(\Sigma_0)>0$ such that
$$\inf_{(p,q,t)\in\mathbb{R}\times\mathbb{R}\times[0,T)}\frac dl(p,q,t)\geq c\,.$$
\end{Thm}
\begin{proof}
Let $\Sigma_0=X_0(\mathbb{R})$ be an embedded curve satisfying~\eqref{eq:alpha},~\eqref{eq:asymp} and~\eqref{eq:intkappageqminuspi}.
Lemma~\ref{lem:minmaxthetaentire} implies that $\theta_0\in[-\pi,\alpha+2\pi]$.
From the maximum principle for $\theta$, Theorem~\ref{thm:dtdstheta}, it follows that
\begin{align}\label{eq:thetainteral2dl}
\theta(p,q,t)\in(-\pi,\alpha+2\pi)
\end{align}
for all $p,q\in\mathbb{S}^1$ and $t\in(0,T)$.
By~\eqref{eq:alpha},
\begin{align}\label{eq:limdl}
\inf_{t\in[0,T)}\lim_{p\to\infty}\frac dl(p,-p,t)=:c_1(\Sigma_0)\in(0,1]\,.
\end{align}
Since $d/l$ is continuous and initially positive, there exists a time $T'\in(0,T]$ so that $d/l>0$ on $[0,T')$.
Fix $t_0\in(0,T')$.
If $\Sigma_{t_0}$ is a line, then Remark~\ref{rem:dl} yields that $d/l\equiv1$ on $\mathbb{R}\times\mathbb{R}$.
Assume that $\Sigma_{t_0}$ is not a line so that $\inf_{\mathbb{R}\times\mathbb{R}}(d/l)<1$ at $t_0$.
Let $p,q\in\mathbb{R}$, $p\ne q$, be points where a local spatial minimum of $d/l$ at $t_0$ is attained and assume w.l.o.g.\ that $s(p,t_0)<s(q,t_0)$.
We have for all $\mathbf{\xi}\in T_{X(p,t_0)}\Sigma_{t_0}\bigoplus T_{X(q,t_0)}\Sigma_{t_0}$,
$$0<\frac dl(p,q,t_0)<1\,,\quad
\mathbf{\xi}\!\left(\frac dl\right)\!(p,q,t_0)=0
\quad\text{ and }\quad
\mathbf{\xi}^2\!\left(\frac dl\right)\!(p,q,t_0)\geq0\,.$$
In the following, we always calculate at the point $(p,q,t_0)$.
The spatial derivatives of $d$ and $l$ are all given in~\cite{Huisken95} (for detailed calculations, see~\cite[Lems.~6.2 and~7.4]{Dittberner18}).
The first spatial derivative of $d/l$ at $(p,t_0)$ in direction of the vector $\mathbf{\xi}=\boldsymbol{\tau}_p\oplus0$ is given by
\begin{align}\label{eq:wtaupdl}
0=(\boldsymbol{\tau}_p\oplus0)\!\left(\frac dl\right)
=-\frac1l\langle\mathbf{w},\boldsymbol{\tau}_p\rangle+\frac d{l^2}\,.
\end{align}
At $(q,t_0)$ and for the vector $\mathbf{\xi}=0\oplus\boldsymbol{\tau}_q$, we have
\begin{align}\label{eq:wtauqdl}
0=(0\oplus\boldsymbol{\tau}_q)\!\left(\frac dl\right)
=\frac1l\langle\mathbf{w},\boldsymbol{\tau}_q\rangle-\frac d{l^2}\,.
\end{align}
Since $d/l\in(0,1)$, and by~\eqref{eq:wtaupdl} and~\eqref{eq:wtauqdl}, there exists $\beta\in(0,\pi)$ with
\begin{align}\label{eq:wtaucosthetadl}
\langle\mathbf{w},\boldsymbol{\tau}_p\rangle=\langle\mathbf{w},\boldsymbol{\tau}_q\rangle
=\frac dl
=\cos\!\left(\frac\beta2\right)
\in(0,1)\,.
\end{align}
By Lemma~\ref{lem:thetabeta},~\eqref{eq:thetainteral2dl} and~\eqref{eq:wtaucosthetadl}, either
\begin{align*}
\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=-\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=-\sin\!\left(\frac\beta2\right)&\quad\text{ and }\quad
2\pi k+\beta=\theta\in(0,\pi)\cup(2\pi,\alpha+2\pi)\,, \\
\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=-\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=\sin\!\left(\frac\beta2\right)&\quad\text{ and }\quad
2\pi k-\beta=\theta\in(-\pi,0)\cup(\pi,2\pi)\,,\quad\text{ or } \\
\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=\pm\sin\!\left(\frac\beta2\right)&\quad\text{ and }\quad
\theta\in\{0,2\pi\}
\end{align*}
for $k\in\mathbb{Z}$.
We use the evolution equation~\eqref{eq:ccf} and Lemma~\ref{lem:evolutionequations} to differentiate the ratio in time,
\begin{align}\label{eq:dtdl}
\fracp{}{t}\left(\frac dl\right)
&=\frac1l\big(h\langle\mathbf{w},\boldsymbol{\nu}_q-\boldsymbol{\nu}_p \rangle
+\langle\mathbf{w},\boldsymbol\kappa_q-\boldsymbol\kappa_p\rangle\big) \notag\\
&\quad+\frac d{l^2}\left(\int_p^q\kappa^2\,ds_t-h\int_p^q\kappa\,ds_t\right)\,.
\end{align}
We are now considering four different cases.\\
(i) Assume that
\begin{align}\label{eq:wnusinthetadl}
\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=-\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=-\sin\!\left(\frac\beta2\right)\quad\text{ and }\quad
\beta=\theta\in(0,\pi)\,.
\end{align}
Subtracting~\eqref{eq:wtaupdl} from~\eqref{eq:wtauqdl} yields $\langle\mathbf{w},\boldsymbol{\tau}_q-\boldsymbol{\tau}_p\rangle=0$.
For unit tangent vectors, we have
$$\langle\boldsymbol{\tau}_p+\boldsymbol{\tau}_q,\boldsymbol{\tau}_q-\boldsymbol{\tau}_p\rangle=\Vert\boldsymbol{\tau}_p\Vert^2-\Vert\boldsymbol{\tau}_q\Vert^2=0\,.$$
Thus, $\mathbf{w}$ and $\boldsymbol{\tau}_p+\boldsymbol{\tau}_q$ are both perpendicular to $\boldsymbol{\tau}_q-\boldsymbol{\tau}_p$ and are therefore parallel, that is, $\measuredangle(\mathbf{w},\boldsymbol{\tau}_q+\boldsymbol{\tau}_p)=0$.
Using $\Vert\mathbf{w}\Vert=1$, we calculate
\begin{align}\label{eq:wtautaupplustauqdl}
0\leq2\frac dl
=\langle\mathbf{w},\boldsymbol{\tau}_q+\boldsymbol{\tau}_p\rangle
=\Vert\boldsymbol{\tau}_q+\boldsymbol{\tau}_p\Vert\,.
\end{align}
By~\eqref{eq:wnusinthetadl},
\begin{align}\label{eq:wkappasinthetadl}
\langle\mathbf{w},\boldsymbol\kappa_q-\boldsymbol\kappa_p\rangle
=-\kappa_q\langle\mathbf{w},\boldsymbol{\nu}_q\rangle+\kappa_p\langle\mathbf{w},\boldsymbol{\nu}_p\rangle
=-(\kappa_p+\kappa_q)\sin\!\left(\frac\theta2\right)\,.
\end{align}
We differentiate $d/l$ at $(p,q,t_0)$ twice with respect to the vector $\mathbf{\xi}=\boldsymbol{\tau}_p\ominus\boldsymbol{\tau}_q$ and calculate with~\eqref{eq:wtautaupplustauqdl} and~\eqref{eq:wkappasinthetadl},
\begin{align*}
0&\leq(\boldsymbol{\tau}_p\ominus \boldsymbol{\tau}_q)^2\left(\frac dl\right)
=\frac1l\langle\mathbf{w},\boldsymbol\kappa_q-\boldsymbol\kappa_p\rangle
=-\frac1l(\kappa_p+\kappa_q)\sin\!\left(\frac\theta2\right)\,.
\end{align*}
We abbreviate $\kappa:=(\kappa_p+\kappa_q)/2$ and obtain
\begin{align*}
2\kappa\sin\!\left(\frac\theta2\right)\leq0\,.
\end{align*}
Since $\sin(\theta/2)>0$ for $\theta\in(0,\pi)$, we conclude $\kappa\leq0$ and
$$h-\kappa\geq h>h-\frac\theta l\,.$$
Furthermore, the inequality
$$\sin\!\left(\frac\theta2\right)>\frac\theta2\cos\!\left(\frac\theta2\right)>0$$
holds for $\theta\in(0,\pi)$.
Hence,
\begin{align}\label{eq:proofdtdlcase2a2}
(h-\kappa)\sin\!\left(\frac\theta2\right)>\frac\theta2\cos\!\left(\frac\theta2\right)\left(h-\frac\theta l\right)
\end{align}
for $\theta\in(0,\pi)$.
Cauchy--Schwarz and the definition~\eqref{eq:deftheta} of $\theta$ imply
\begin{align}\label{eq:CSthetadl}
\int_p^q\kappa^2\,ds_t
\geq\frac1l\left(\int_p^q\kappa\,ds_t\right)^{2}
=\frac{\theta^2}l\,.
\end{align}
Then~\eqref{eq:dtdl},~\eqref{eq:wtaucosthetadl},~\eqref{eq:wnusinthetadl},~\eqref{eq:wkappasinthetadl},~\eqref{eq:proofdtdlcase2a2} and~\eqref{eq:CSthetadl} yield
\begin{align*}
\fracp{}{t}\left(\frac dl\right)
&\geq\frac2l\left((h-\kappa)\sin\!\left(\frac\theta2\right)
-\frac\theta2\cos\!\left(\frac\theta2\right)\left(h-\frac\theta l\right)\right)
>0
\end{align*}
at $(p,q,t_0)$. \\
(ii) Assume that $\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=\pm\sin(\beta/2)$ and $\theta=0$.
By~\eqref{eq:wtaucosthetadl},
\begin{align}\label{eq:wnusinthetadl_0}
\boldsymbol{\tau}_p=\boldsymbol{\tau}_q\qquad\text{ and }\qquad\boldsymbol{\nu}_p=\boldsymbol{\nu}_q\,.
\end{align}
We differentiate $d/l$ at $(p,q,t_0)$ twice with respect to the vector $\mathbf{\xi}=\boldsymbol{\tau}_p\oplus\boldsymbol{\tau}_q$ and calculate with~\eqref{eq:wnusinthetadl_0},
\begin{align}\label{eq:wtautaupplustauqdl_0}
0&\leq(\boldsymbol{\tau}_p\oplus\boldsymbol{\tau}_q)^2\left(\frac dl\right)
=\frac1l\langle\mathbf{w},\boldsymbol\kappa_q-\boldsymbol\kappa_p\rangle\,.
\end{align}
We conclude with~\eqref{eq:deftheta},~\eqref{eq:dtdl},~\eqref{eq:wnusinthetadl_0},~\eqref{eq:wtautaupplustauqdl_0},
\begin{align*}
\fracp{}{t}\left(\frac dl\right)
\geq\frac d{l^2}\int_p^q\kappa^2\,ds_t
>0
\end{align*}
at $(p,q,t_0)$. \\
(iii) Assume that $\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=-\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=\sin(\beta/2)$ and $-\beta=\theta\in(-\pi,0)$.
Again by continuity of $d/l$, there exists $t_1(\Sigma_0)\in(0,T')$ so that
\begin{align}\label{eq:Cstar1dl}
\inf_{\mathbb{R}\times\mathbb{R}}\frac dl(\,\cdot\,,\,\cdot\,,t)\geq\frac12\inf_{\mathbb{R}\times\mathbb{R}}\frac dl(\,\cdot\,,\,\cdot\,,0)
\end{align}
for all $t\in[0,t_1]$.
If $t_0\in(t_1,T)$, Theorem~\ref{thm:dtdstheta} and~\eqref{eq:thetainteral2dl} yield $-\pi<\theta_{\inf}(t_1)\leq\theta_{\inf}(t_0)<0$ so that~\eqref{eq:wtaucosthetadl} and the monotone behaviour of the cosine on $(-\pi,0)$ imply
\begin{align}\label{eq:Cstar2dl}
\frac dl=\cos\!\left(\frac\theta2\right)
\geq\cos\!\left(\frac{\theta_{\inf}(t_0)}2\right)
\geq\cos\!\left(\frac{\theta_{\inf}(t_1)}2\right)>0\,.
\end{align}
We deduce with~\eqref{eq:Cstar1dl} and~\eqref{eq:Cstar2dl} that
$$\frac dl\geq
\min\left\{\frac12\inf_{\mathbb{R}\times\mathbb{R}}\frac dl(\,\cdot\,,\,\cdot\,,0),\cos\!\left(\frac{\theta_{\inf}(t_1)}2\right)\right\}=:c_2(\Sigma_0)>0$$
at $(p,q,t_0)$.\\
(iv) Assume that $\theta\in(\pi,\alpha+2\pi)$.
By~\eqref{eq:wtaucosthetadl}, $\langle\mathbf{w},\boldsymbol{\tau}_q\rangle=\langle\mathbf{w},\boldsymbol{\tau}_p\rangle=d/l\in(0,1)$.
Since $\Sigma_{t_0}$ is embedded and has asymptotic ends and $X(\,\cdot\,,t_0)$ is continuous, the curve has to cross the line segment between $X(p,t_0)$ and $X(q,t_0)$ at least once at $X(r,t_0)$ with $r\notin[p,q]$.
Lemma~\ref{lem:mindltheta2pik} implies that $d/l$ cannot attain the infimum at $(p,q,t_0)$ (it could still, however, attain a local minimum a this point).
Hence,
$$\frac dl>\inf_{\mathbb{R}\times\mathbb{R}}\frac dl(\,\cdot\,,\,\cdot\,,t_0)\,,$$
where $\inf_{\mathbb{R}\times\mathbb{R}}d/l(\,\cdot\,,\,\cdot\,,t_0)$ is either the infimum from~\eqref{eq:limdl} or a local minimum as discussed in cases~(i),~(ii) and~(iii). \\
Assume that $d/l$ falls below $c:=\min\{c_1,c_2\}$ and attains $\Lambda\in(0,c)$ for the first time at time $t_2\in(0,T)$ and points $p,q\in\mathbb{S}^1$, $p\ne q$, so that
\begin{align}\label{eq:c1mindl}
c>\Lambda=\frac dl(p,q,t_2)=\inf_{\mathbb{R}\times\mathbb{R}}\frac dl(\,\cdot\,,\,\cdot\,,t_2)
\end{align}
is the infimum and
\begin{align}\label{eq:dtdlleq0}
\fracp{}{t}_{|_{t=t_2}}\!\left(\frac dl\right)(p,q,t)\leq0\,.
\end{align}
Cases~(i) and~(ii) contradict~\eqref{eq:dtdlleq0}, and cases~(iii) and~(iv) contradict~\eqref{eq:c1mindl}.
\end{proof}
\begin{Cor}
\label{cor:embeddednessdl}
Let $\Sigma_0=X_0(\mathbb{R})$ be a smooth, embedded curve satisfying~\eqref{eq:intkappageqminuspi}.
Let $X:\mathbb{R}\times[0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} satisfying~\eqref{eq:alpha},~\eqref{eq:asymp} and~\eqref{eq:h_1}, and with initial curve $\Sigma_0$.
Then $\Si_t=X(\mathbb{R},t)$ is embedded for all $t\in(0,T)$.
\end{Cor}
\begin{Rem}
Counterexample~\ref{ex:cexample} shows that in order for embeddedness to be preserved it is crucial to assume that the initial local total curvature lies above $-\pi$.
\end{Rem}
\section{Distance comparison principle for closed curves}\label{sec:noncollapsing}
We continue to adapt the methods from Huisken~\cite{Huisken95}.
Let $X(\mathbb{S}^1)$ be a circle of radius $R$.
Then
$$\frac{d(p,q)}2
=\frac{L}{2\pi}\sin\!\left(\frac{\pi l(p,q)}{L}\right)$$
for all $p,q\in\mathbb{S}^1$.
This motivates the definition of the function $\psi:\mathbb{S}^1\!\times\mathbb{S}^1\!\times[0,T)\to\mathbb{R}$ with
\begin{align}\label{eq:defpsi}
\psi(p,q,t):=\frac{L_t}\pi\sin\!\left(\frac{\pi l(p,q,t)}{L_t}\right)\,,
\end{align}
where $L_t<\infty$.
We set $d/\psi\equiv1$ on $\{p=q\}\times[0,T)$, then $d/\psi\in C^0(\mathbb{S}^1\!\times\mathbb{S}^1\!\times[0,T))$.
\begin{Rem}\label{rem:dpsi}
Since $\sin(\pi-\alpha)=\sin(\alpha)$,
we have $\psi(p,q,t)=\psi(q,p,t)$.
Hence, we will later assume that $l\leq L/2$.
Embedded curves satisfy $d/\psi>0$.
If a closed curve $\Si_t$ is not a circle, then there exist $p,q\in\mathbb{S}^1$ so that $d(p,q,t)<\psi(p,q,t)$ and thus $\min_{\mathbb{S}^1\!\times\mathbb{S}^1}(d/\psi)<1$.
\end{Rem}
\begin{Lemma}
\label{lem:mindpsitheta2pik}
Let $\Sigma=X(\mathbb{S}^1)$ be an embedded, closed curve.
Let $p,q\in\mathbb{S}^1$, $p\ne q$, such that $\Sigma$ crosses the connecting line between $X(p)$ and $X(q)$.
Then $(d/\psi)(p,q)$ cannot be a global minimum.
\end{Lemma}
\begin{proof}
Let $\Sigma=X(\mathbb{S}^1)$ be an embedded, closed curve that crosses the connecting line between $X(p)$ and $X(q)$.
That is, there exists an $r\in\mathbb{S}^1$, $r\ne p,q$, with $X(r)=X(p)+\mathbf{w}(p,q)\Vert X(r)-X(p)\Vert$.
Set $d:=d(p,q)$, $d_1:=d(p,r)$ and $d_2:=d(r,q)$.
Then
\begin{align}\label{eq:d1plusd2}
d=d_1+d_2\,.
\end{align}
Furthermore, set $l:=l(p,q)$, $l_1:=l(p,r)$ and $l_2:=l(r,q)$.
Since $l_1,l_2\in(0,L)$,
\begin{align*}
\sin\!\left(\frac{\pi(l_1+l_2)}{L}\right)
&<\sin\!\left(\frac{\pi l_1}{L}\right)+\sin\!\left(\frac{\pi l_2}{L}\right)\,.
\end{align*}
Set $\psi:=\psi(p,q)$, $\psi_1:=\psi(p,r)$ and $\psi_2:=\psi(r,q)$ and assume that $d/\psi$ attains its global minimum at $(p,q)$.
We parametrise $\Sigma$ by arc length, so that $s(p)=0$.
Then either $0=s(p)<s(r)<s(q)$ or $0=s(p)<s(q)<s(r)$.
By the properties of the sine function, in both cases, $\psi<\psi_1+\psi_2$.
Since $d/\psi$ is a global minimum, we can estimate with~\eqref{eq:d1plusd2},
$$\frac{d_1}{\psi_1}\geq\frac d\psi>\frac{d_1+d_2}{\psi_1+\psi_2}
\qquad\text{ and }\qquad
\frac{d_2}{\psi_2}\geq\frac d\psi>\frac{d_1+d_2}{\psi_1+\psi_2}$$
so that
$$d_1(\psi_1+\psi_2)>(d_1+d_2)\psi_1
\qquad\text{ and }\qquad
d_2(\psi_1+\psi_2)>(d_1+d_2)\psi_2\,.$$
Adding both inequalities yields a contradiction.
Thus, $d/\psi$ cannot be a global minimum.
\end{proof}
Now we can prove a similar result to~\cite[Thm.~2.3]{Huisken95}.
\begin{Thm}
\label{thm:minimumdpsi}
Let $\Sigma_0=X_0(\mathbb{S}^1)$ be a smooth, embedded curve satisfying~\eqref{eq:intkappageqminuspi}.
Let $X:\mathbb{S}^1\!\times[0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} satisfying~\eqref{eq:h_1} and with initial curve $\Sigma_0$.
Then there exists a constant $c(\Sigma_0)>0$ such that
$$\inf_{(p,q,t)\in\mathbb{S}^1\!\times\mathbb{S}^1\!\times[0,T)}\frac d\psi(p,q,t)\geq c\,.$$
\end{Thm}
\begin{proof}
Let $\Sigma_0=X_0(\mathbb{S}^1)$ be an embedded closed curve satisfying~\eqref{eq:intkappageqminuspi}.
Lemma~\ref{lem:minmaxtheta} implies that $\theta_0\in[-\pi,3\pi]$.
From the maximum principle for $\theta$, Theorem~\ref{thm:dtdstheta}, it follows that
\begin{align}\label{eq:thetainteral2}
\theta(p,q,t)\in(-\pi,3\pi)
\end{align}
for all $p,q\in\mathbb{S}^1$ and $t\in(0,T)$.
Since $d/\psi$ is continuous and initially positive, there exists a time $T'\in(0,T]$ so that $d/\psi>0$ on $[0,T')$.
Fix $t_0\in(0,T')$.
If $\Sigma_{t_0}$ is a circle, then Remark~\ref{rem:dpsi} yields that $d/\psi\equiv1$ on $\mathbb{S}^1\!\times\mathbb{S}^1$.
Assume that $\Sigma_{t_0}$ is not a circle so that $\min_{\mathbb{S}^1\!\times\mathbb{S}^1}(d/\psi)<1$ at $t_0$.
Let $p,q\in\mathbb{S}^1$, $p\ne q$, be points where a local spatial minimum of $d/\psi$ at $t_0$ is attained and assume w.l.o.g.\ that $s(p,t_0)<s(q,t_0)$.
Again by Remark~\ref{rem:dpsi}, we can assume that $l(p,q,t_0)\leq L_{t_0}/2$.
We have for all $\mathbf{\xi}\in T_{X(p,t_0)}\Sigma_{t_0}\bigoplus T_{X(q,t_0)}\Sigma_{t_0}$,
$$0<\frac d\psi(p,q,t_0)<1\,,\quad
\mathbf{\xi}\!\left(\frac d\psi\right)\!(p,q,t_0)=0
\quad\text{ and }\quad
\mathbf{\xi}^2\!\left(\frac d\psi\right)\!(p,q,t_0)\geq0\,.$$
In the following, we always calculate at the point $(p,q,t_0)$.
The spatial derivatives of $d$ and $\psi$ are all given in~\cite{Huisken95} (for detailed calculations, see~\cite[Cor.~7.12 and Thm.~7.21]{Dittberner18}).
The first spatial derivative of $d/\psi$ at $(p,t_0)$ in direction of the vector $\mathbf{\xi}=\boldsymbol{\tau}_p\oplus0$ is given by
\begin{align}\label{eq:wtaup}
0=(\boldsymbol{\tau}_p\oplus0)\!\left(\frac d\psi\right)
=-\frac1\psi\langle\mathbf{w},\boldsymbol{\tau}_p\rangle+\frac d{\psi^2}\cos\!\left(\frac{\pi l}{L}\right)\,.
\end{align}
At $(q,t_0)$ and for the vector $\mathbf{\xi}=0\oplus\boldsymbol{\tau}_q$, we have
\begin{align}\label{eq:wtauq}
0=(0\oplus\boldsymbol{\tau}_q)\!\left(\frac d\psi\right)
=\frac1\psi\langle\mathbf{w},\boldsymbol{\tau}_q\rangle-\frac d{\psi^2}\cos\!\left(\frac{\pi l}{L}\right)\,.
\end{align}
Since $l\in(0,L/2]$ and $d/\psi\in(0,1)$, and by~\eqref{eq:wtaup} and~\eqref{eq:wtauq}, there exists $\beta\in(0,\pi]$ with
\begin{align}\label{eq:wtaucostheta}
\langle\mathbf{w},\boldsymbol{\tau}_p\rangle=\langle\mathbf{w},\boldsymbol{\tau}_q\rangle
=\frac d\psi\cos\!\left(\frac{\pi l}{L}\right)
=\cos\!\left(\frac\beta2\right)
\in[0,1)\,.
\end{align}
By Lemma~\ref{lem:thetabeta} and~\eqref{eq:thetainteral2}, either
\begin{align*}
\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=-\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=-\sin\!\left(\frac\beta2\right)&\quad\text{ and }\quad
2\pi k+\beta=\theta\in(0,\pi]\cup(2\pi,3\pi)\,, \\
\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=-\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=\sin\!\left(\frac\beta2\right)&\quad\text{ and }\quad
2\pi k-\beta=\theta\in(-\pi,0)\cup[\pi,2\pi)\,,\quad\text{ or } \\
\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=\pm\sin\!\left(\frac\beta2\right)&\quad\text{ and }\quad
\theta\in\{0,2\pi\}
\end{align*}
for $k\in\mathbb{Z}$.
We are now considering three different cases.\\
(i) Assume that
\begin{align}\label{eq:wnusintheta}
\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=-\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=-\sin\!\left(\frac\beta2\right)\quad\text{ and }\quad
\beta=\theta\in(0,\pi]\,.
\end{align}
By~\eqref{eq:wtaucostheta}, also
$$\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=-\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=-\frac d\psi\sin\!\left(\frac{\pi l}{L}\right)$$
so that
\begin{align}\label{eq:sinthetasinpilL}
2\sin\!\left(\frac\theta2\right)
=\langle\mathbf{w},\boldsymbol{\nu}_q-\boldsymbol{\nu}_p\rangle
=2\frac d\psi\sin\!\left(\frac{\pi l}{L}\right)\,.
\end{align}
Like in the proof of Theorem~\ref{thm:minimumdl},
$\measuredangle(\mathbf{w},\boldsymbol{\tau}_q+\boldsymbol{\tau}_p)=0$.
Using $\Vert\mathbf{w}\Vert=1$ and~\eqref{eq:wtaucostheta}, we calculate
\begin{align}\label{eq:wtautaupplustauq}
0\leq2\frac d\psi\cos\!\left(\frac{\pi l}{L}\right)
=\langle\mathbf{w},\boldsymbol{\tau}_q+\boldsymbol{\tau}_p\rangle
=\Vert\boldsymbol{\tau}_q+\boldsymbol{\tau}_p\Vert\,.
\end{align}
Since $d/\psi<1$, and again by~\eqref{eq:wtaucostheta},
\begin{align}\label{eq:costhetacosdpsi}
\cos\!\left(\frac\theta2\right)
=\frac d\psi\cos\!\left(\frac{\pi l}{L}\right)
<\cos\!\left(\frac{\pi l}{L}\right)\,.
\end{align}
As $\pi l/\!L\in(0,\pi/2]$ and the cosine function is axially symmetric and monotonically decreasing on $(0,\pi/2]$,~\eqref{eq:costhetacosdpsi} implies
\begin{align}\label{eq:thetalR}
\theta>\frac{2\pi l}{L}\,.
\end{align}
By~\eqref{eq:wnusintheta},
\begin{align}\label{eq:wkappasintheta}
\langle\mathbf{w},\boldsymbol\kappa_q-\boldsymbol\kappa_p\rangle
=-\kappa_q\langle\mathbf{w},\boldsymbol{\nu}_q\rangle+\kappa_p\langle\mathbf{w},\boldsymbol{\nu}_p\rangle
=-(\kappa_p+\kappa_q)\sin\!\left(\frac\theta2\right)\,.
\end{align}
We differentiate $d/\psi$ at $(p,q,t_0)$ twice with respect to the vector $\mathbf{\xi}=\boldsymbol{\tau}_p\ominus\boldsymbol{\tau}_q$ and calculate, using the definition~\eqref{eq:defpsi} of $\psi$,~\eqref{eq:wtautaupplustauq} and~\eqref{eq:wkappasintheta},
\begin{align*}
0&\leq (\boldsymbol{\tau}_p\ominus\boldsymbol{\tau}_q)^2\!\left(\frac d\psi\right)
=-\frac1\psi(\kappa_p+\kappa_q)\sin\!\left(\frac\theta2\right)
+\frac{4\pi^2d}{L^2\psi}\,.
\end{align*}
We abbreviate $\kappa:=(\kappa_p+\kappa_q)/2$ and obtain
\begin{align}\label{eq:kappasinthetadpsi}
2\kappa\sin\!\left(\frac\theta2\right)\leq\frac{4\pi^2}{L^2}d\,.
\end{align}
Since the sine function is positive and monotonically increasing on $(0,\pi/2]$, we conclude with $d/\psi<1$,~\eqref{eq:defpsi},~\eqref{eq:thetalR} and~\eqref{eq:kappasinthetadpsi} that
$$2\kappa\sin\!\left(\frac\theta2\right)
\leq\frac{4\pi^2}{L^2}d
<\frac{4\pi^2}{L^2}\psi
=\frac{4\pi^2}{L^2}\frac{L}\pi\sin\!\left(\frac{\pi l}{L}\right)
<\frac{4\pi}{L}\sin\!\left(\frac\theta2\right)\,.$$
Since $\sin(\theta/2)>0$ for $\theta\in(0,\pi]$, we can divide by it to obtain with Cauchy--Schwarz,
\begin{align}\label{eq:kappaeq1R}
\kappa<\frac{2\pi}{L}\leq\frac1{2\pi}\int_{\Si_t}\kappa^2\,ds_t\,.
\end{align}
The definition~\eqref{eq:deftheta} of $\theta$ and~\eqref{eq:thetalR} imply
\begin{align}\label{eq:thetah}
-h\int_p^q\kappa\,ds_t+\int_p^q\kappa^2\,ds_t
\geq-h\theta+\frac{\theta^2}l
>-\theta\left(h-\frac{2\pi}{L}\right).
\end{align}
We use the evolution equation~\eqref{eq:ccf} and Lemma~\ref{lem:evolutionequations} to differentiate the ratio in time and obtain by~\eqref{eq:sinthetasinpilL},~\eqref{eq:costhetacosdpsi},~\eqref{eq:wkappasintheta} and~\eqref{eq:thetah},
\begin{align}\label{eq:dtdpsi}
&\fracp{}{t}\!\left(\frac d\psi\right) \notag\\
&\;=\frac1\psi\big(h\langle\mathbf{w},\boldsymbol{\nu}_q-\boldsymbol{\nu}_p\rangle+\langle\mathbf{w},\boldsymbol\kappa_q-\boldsymbol\kappa_p\rangle\big)
-\frac d{\psi^2}\cos\!\left(\frac{\pi l}{L}\right)
\left(h\int_p^q\kappa\,ds_t-\int_p^q\kappa^2\,ds_t\right) \notag\\
&\;\quad+\frac d{\pi\psi^2}\left(\int_{\Si_t}\kappa^2\,ds_t-2\pi h\right)
\left(\sin\!\left(\frac{\pi l}{L}\right)-\frac{\pi l}{L}\cos\!\left(\frac{\pi l}{L}\right)\right)\notag\\
&\;\geq\frac2\psi(h-\kappa)\sin\!\left(\frac\theta2\right)
-\frac\theta\psi\cos\!\left(\frac\theta2\right)\left(h-\frac{\theta}{l}\right)\notag\\
&\;\quad+\frac1\psi\left(\frac1{\pi}\int_{\Si_t}\kappa^2\,ds_t-2h\right)
\left(\sin\!\left(\frac\theta2\right)-\frac{\pi l}L\cos\!\left(\frac\theta2\right)\right) \notag\\
&\;=\frac2\psi\left(\frac1{2\pi}\int_{\Si_t}\kappa^2\,ds_t-\kappa\right)\left(\sin\!\left(\frac\theta2\right)
-\frac\theta2\cos\!\left(\frac\theta2\right)\right) \notag\\
&\;\quad+\frac\theta\psi\cos\!\left(\frac\theta2\right)
\left(\frac1{2\pi}\int_{\Si_t}\kappa^2\,ds_t-\kappa
-h+\frac{\theta}{l}-\frac l{L\theta}\int_{\Si_t}\kappa^2\,ds_t+\frac{h2\pi l}{L\theta}\right)\,,
\end{align}
where we just added a zero in the last step.
Since $h$ satisfies~\eqref{eq:h_1}, that is, $h\leq\int_{\Si_t}\kappa^2\,ds_t/2\pi+2\pi/L$, we estimate with~\eqref{eq:thetalR} and~\eqref{eq:kappaeq1R},
\begin{align}\label{eq:dpsiA}
&\frac1{2\pi}\int_{\Si_t}\kappa^2\,ds_t-\kappa
-h+\frac{\theta}{l}-\frac l{L\theta}\int_{\Si_t}\kappa^2\,ds_t+\frac{h2\pi l}{L\theta} \notag\\
&\quad>\left(\frac1{2\pi}\int_{\Si_t}\kappa^2\,ds_t-h\right)\left(1-\frac{2\pi l}{L\theta}\right)
+\frac\theta l\left(1-\frac{2\pi l}{L\theta}\right) \notag\\
&\quad>\left(\frac1{2\pi}\int_{\Si_t}\kappa^2\,ds_t+\frac{2\pi}L-h\right)\left(1-\frac{2\pi l}{L\theta}\right)\geq0\,.
\end{align}
Furthermore, the inequality
\begin{align}\label{eq:proofdtdpsicase2asintheta}
\sin\!\left(\frac\theta2\right)>\frac\theta2\cos\!\left(\frac\theta2\right)
\end{align}
holds for all $\theta\in(0,\pi]$.
Thus, we conclude with,~\eqref{eq:kappaeq1R},~\eqref{eq:dtdpsi},~\eqref{eq:dpsiA} and~\eqref{eq:proofdtdpsicase2asintheta} that
$$\fracp{}{t}\!\left(\frac d\psi\right)>0$$
at $(p,q,t_0)$.\\
(ii) Assume that $\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=-\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=\sin(\beta/2)$ and $-\beta=\theta\in(-\pi,0)$.
Again by continuity of $d/\psi$, there exists $t_1(\Sigma_0)\in(0,T')$ so that
\begin{align}\label{eq:Cstar1}
\min_{\mathbb{S}^1\!\times\mathbb{S}^1}\frac d\psi(\,\cdot\,,\,\cdot\,,t)\geq\frac12\min_{\mathbb{S}^1\!\times\mathbb{S}^1}\frac d\psi(\,\cdot\,,\,\cdot\,,0)
\end{align}
for all $t\in[0,t_1]$.
For $t_0\in(t_1,T)$,
Theorem~\ref{thm:dtdstheta} applied to the initial time $t_1$ and~\eqref{eq:thetainteral2} yield $-\pi<\theta_{\min}(t_1)<\theta_{\min}(t_0)<0$ so that the monotone behaviour of the cosine on $(-\pi,0)$ implies
\begin{align}\label{eq:costhetatthetat0}
\cos\!\left(\frac{\theta_{\min}(t_0)}2\right)>\cos\!\left(\frac{\theta_{\min}(t_1)}2\right)>0\,.
\end{align}
From $0<l\leq L/2$ it follows that $1>\cos(\pi l/\!L)\geq0$ so that, by~\eqref{eq:wtaucostheta},~\eqref{eq:costhetatthetat0} and again the monotone behaviour of the cosine on $(-\pi,0)$,
\begin{align}\label{eq:Cstar2}
\frac d\psi>\frac d\psi\cos\!\left(\frac{\pi l}{L}\right)=\cos\!\left(\frac\theta2\right)
\geq\cos\!\left(\frac{\theta_{\min}(t_0)}2\right)
>\cos\!\left(\frac{\theta_{\min}(t_1)}2\right)\,.
\end{align}
With~\eqref{eq:Cstar1} and~\eqref{eq:Cstar2}, we deduce that
$$\frac d\psi
\geq\min\left\{\frac12\min_{\mathbb{S}^1\!\times\mathbb{S}^1}\frac d\psi(\,\cdot\,,\,\cdot\,,0),\cos\!\left(\frac{\theta_{\min}(t_1)}2\right)\right\}=:c(\Sigma_0)>0$$
at $(p,q,t_0)$. \\
(iii) Assume that $\theta\in\{0\}\cup(\pi,3\pi)$ or $\theta=\pi$ and $\langle\mathbf{w},\boldsymbol{\nu}_p\rangle=-\langle\mathbf{w},\boldsymbol{\nu}_q\rangle=1$.
By~\eqref{eq:wtaucostheta}, $\langle\mathbf{w},\boldsymbol{\tau}_q\rangle=\langle\mathbf{w},\boldsymbol{\tau}_p\rangle\in[0,1)$.
Since $\Sigma_{t_0}$ is closed and $X(\,\cdot\,,t_0)$ is continuous, $\Sigma_{t_0}$ has to cross the straight line segment between $X(p,t_0)$ and $X(q,t_0)$ at least once.
Lemma~\ref{lem:mindpsitheta2pik} implies that the ratio $d/\psi$ cannot have a global minimum at $(p,q,t_0)$ (it could still, however, attain a local minimum a this point).
Hence,
$$\frac d\psi>\min_{\mathbb{S}^1\!\times\mathbb{S}^1}\frac d\psi(\,\cdot\,,\,\cdot\,,t_0)\,,$$
where $\min_{\mathbb{S}^1\!\times\mathbb{S}^1}(d/\psi)(\,\cdot\,,\,\cdot\,,t_0)$ is attained at a point which was treated in cases~(i) and~(ii). \\
Assume that $d/\psi$ falls below $c$ and attains $\Lambda\in(0,c)$ for the first time at time $t_2\in(0,T)$ and points $p,q\in\mathbb{S}^1$, $p\ne q$, so that
\begin{align}\label{eq:c1mindpsi}
c>\Lambda=\frac d\psi(p,q,t_2)=\min_{\mathbb{S}^1\!\times\mathbb{S}^1}\frac d\psi(\,\cdot\,,\,\cdot\,,t_2)
\end{align}
is a global minimum and
\begin{align}\label{eq:dtdpsileq0}
\fracp{}{t}_{|_{t=t_2}}\!\left(\frac d\psi\right)(p,q,t)\leq0\,.
\end{align}
Case~(i) contradicts~\eqref{eq:dtdpsileq0}, and cases~(ii) and~(iii) contradict~\eqref{eq:c1mindpsi}.
\end{proof}
\begin{Cor}
\label{cor:embeddedness}
Let $\Sigma_0=X_0(\mathbb{S}^1)$ be a smooth, embedded curve satisfying~\eqref{eq:intkappageqminuspi}.
Let $X:\mathbb{S}^1\!\times[0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} satisfying~\eqref{eq:h_1} and with initial curve $\Sigma_0$.
Then $\Si_t=X(\mathbb{S}^1\!,t)$ is embedded for all $t\in(0,T)$.
\end{Cor}
The next example shows, why the condition $\min\theta_0\geq-\pi$ is sharp.
\begin{Coex}\label{ex:cexample}
Gage~\cite[p.~53]{Gage86} suggested the following counterexample.
Pihan~\cite[Section~5.4]{Pihan98} gave an incomplete proof for its validity which we will fix here.
If we allow local total curvature smaller than $-\pi$, then there exist counterexamples for any given minimum $\min\theta_0<-\pi$.
For the curve in Figure~\ref{fig:cexample},
$\theta_{\min}=\theta(p_1,p_2)<-\pi$.
We will construct a solution of~\eqref{eq:ccf} with embedded initial curve $\Sigma_0$ that intersects itself in finite time.
Fix $K_0>0$.
Let $\mathcal{S}$ be the set of all smooth, embedded curves in $\mathbb{R}^2$ that satisfy
\begin{align}\label{eq:FC3alpha0}
\min\theta_0<-\pi\,,\quad
\Vert X_0\Vert_{C^{3,\alpha}(\mathbb{S}^1)}\leq K_0
\quad\text{ and }\quad
L(\Sigma)=L_0\geq8\pi K_0\,,
\end{align}
where $L_0$ is chosen big enough so that curves like in Figure~\ref{fig:cexample} are in $\mathcal{S}$.
By the short time existence, Theorem~\ref{thm:ste}, there exists a time $T=T(K_0)$ so that
$$\Vert X\Vert_{C^{3,\alpha;1\lfloor\alpha/2\rfloor}(\mathbb{S}^1\!\times[0,T/2])}\leq K_1(K_0)\,.$$
In particular,
\begin{align}\label{eq:F1K1alpha}
\left|\fracp{X^1}{t}(p,t)-\fracp{X^1}{t}(p,0)\right|\leq K_1t^{\alpha/2}\,,
\end{align}
where $X^1:=\langle X,\mathbf{e}_1\rangle$, and, by~\eqref{eq:ccf} and~\eqref{eq:F1K1alpha},
\begin{align}\label{eq:F1K1}
-K_1t^{\alpha/2}
\leq(h(t)-\kappa(p,t))\boldsymbol{\nu}^1(p,t)-(h(0)-\kappa(p,0))\boldsymbol{\nu}^1(p,0)
\leq K_1t^{\alpha/2}
\end{align}
for all $p\in\mathbb{S}^1$ and for all $t\in[0,T/2]$, where $\boldsymbol{\nu}^1:=\langle\boldsymbol{\nu},\mathbf{e}_1\rangle$.
Assume $h(0)>0$ and set
\begin{align}\label{eq:t1}
t_1=t_1(K_0):=\min\left\{\frac T2,\left(\frac{h(0)}{2K_1}\right)^{-\alpha/2}\right\}\,.
\end{align}
Then~\eqref{eq:F1K1} holds for $t\in[0,t_1]$.
Let $\Sigma\in\mathcal{S}$ be a curve like in Figure~\ref{fig:cexample}, which is symmetric about the $x_2$-axis.
Let $p,q\in\mathbb{S}^1$ be located as in the picture so that
\begin{align}\label{eq:nukappa0}
\boldsymbol{\nu}(p,0)=-\boldsymbol{\nu}(q,0)=-\mathbf{e}_1\,\qquad\text{ and }\qquad
\kappa(p,0)=\kappa(q,0)=0\,.
\end{align}
We estimate with~\eqref{eq:ccf},~\eqref{eq:F1K1},~\eqref{eq:t1} and~\eqref{eq:nukappa0},
\begin{align}\label{eq:dtF1leq_1}
\fracp{X^1}{t}(p,t)
&=(h(t)-\kappa(p,t))\boldsymbol{\nu}^1(p,t)
\leq(h(0)-\kappa(p,0))\boldsymbol{\nu}^1(p,0)+K_1t_1^{\alpha/2} \notag\\
&\leq-h(0)+\frac{h(0)}2
=-\frac{h(0)}2
\end{align}
and likewise
\begin{align}\label{eq:dtF1leq_2}
\fracp{X^1}{t}(q,t)
&=(h(t)-\kappa(q,t))\boldsymbol{\nu}^1(q,t)
\geq(h(0)-\kappa(q,0))\boldsymbol{\nu}^1(q,0)-K_1t_1^{\alpha/2} \notag\\
&\geq h(0)-\frac{h(0)}2
=\frac{h(0)}2
\end{align}
for $t\in[0,t_1]$.
Since $\min\theta_0<-\pi$, we can smoothly deform a curve like in Figure~\ref{fig:cexample} to achieve arbitrarily small distance between $X(p,0)$ and $X(q,0)$ without exceeding the upper bound $K_0$ in~\eqref{eq:FC3alpha0} or changing the length or enclosed area.
Hence, we can choose an embedded initial curve $\Sigma_0$ with
\begin{align}\label{eq:F10}
X^1(p,0)=-X^1(q,0)\leq\frac{h(0)t_1}4\,.
\end{align}
Then, by~\eqref{eq:dtF1leq_1} and~\eqref{eq:F10}
$$X^1(p,t_1)=X^1(p,0)+\int_0^{t_1}\fracp{X^1}{t}(p,t)\,dt
\leq\frac{h(0)t_1}4-\frac{h(0)t_1}2<0$$
and by~\eqref{eq:dtF1leq_2} and~\eqref{eq:F10}
$$X^1(q,t_1)=X^1(q,0)+\int_0^{t_1}\fracp{X^1}{t}(q,t)\,dt
\geq-\frac{h(0)t_1}4+\frac{h(0)t_1}2>0$$
so that the curve has crossed itself by the time $t_1$.
\end{Coex}
\section{Singularity analysis}\label{sec:singana}
Proposition~\ref{prop:T<infty} states that the curvature blows up if $T<\infty$.
In this section, we assume $T<\infty$ and investigate curvature blow-ups for embedded flows~\eqref{eq:ccf} that satisfy~\eqref{eq:intkappageqminuspi} and~\eqref{eq:h_2}.
We adapt techniques from the theory of CSF to show that the curvature does not blow up in finite time and conclude $T=\infty$. \\
Proposition~\ref{prop:T<infty} motivates the following definition.
We say that a solution $X:\cI\times[0,T)\to\mathbb{R}^2$ of~\eqref{eq:ccf} develops a singularity at $T\leq\infty$ if $\max_{p\in\mathbb{S}^1}|\kappa(p,t)|\to\infty$ for $t\nearrow T$.
\begin{Lemma}
\label{lem:lowerblowuprate}
Let $X:\cI\times[0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} satisfying~\eqref{eq:h_2} and with
$T<\infty$.
Then, for all $t\in(0,T)$,
$$\max_{p\in\mathbb{S}^1}|\kappa(p,t)|\geq\frac1{2\sqrt{T-t}}\,.$$
\end{Lemma}
\begin{proof}
The proof is as in~\cite[Lem.~1.2]{Huisken90}, see also~\cite[Prop.~4.1]{Maeder15} or~\cite[Lem.~9.5]{Dittberner18}.
\end{proof}
Like for CSF, we distinguish between two kinds of singularities according to the blow-up rate from Lemma~\ref{lem:lowerblowuprate}.
Let $X:\cI\times(0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} with $T<\infty$.
We say that a singularity is of type I, if there exists a constant $C_0>0$ so that
$$\max_{p\in\mathbb{S}^1}|\kappa(p,t)|\leq\frac{C_0}{\sqrt{T-t}}$$
for all $t\in(0,T)$.
A singularity is said to be of type II, if such a constant does not exist, that is,
$$\limsup_{t\to T}\max_{p\in\mathbb{S}^1}|\kappa(p,t)|\sqrt{T-t}=\infty\,.$$
Type-I singularities have already been exploited in~\cite[Section~4]{Maeder15}.
We refer also to~\cite[Section~11]{White97} for a characterisation of singularities for almost Brakke flows with bounded global terms, using a monotonicity formula and a result of~\cite{Ilmanen95}.
\begin{Thm}[\hspace{-.005em}{\cite[Prop.~4.12]{Maeder15}}]
\label{thm:notypeIsing}
Let $X:\cI\times(0,T)\to\mathbb{R}^2$ be a smooth, embedded solution of~\eqref{eq:ccf} satisfying~\eqref{eq:h_2} and with $T<\infty$.
Then a type-I singularity cannot form at $T$.
\end{Thm}
\begin{proof}
In~\cite[Prop.~4.12]{Maeder15} the theorem is only stated for convex curves.
But the proof does not use the convexity, see also~\cite[Section~9.4]{Dittberner18}.
By Corollaries~\ref{cor:embeddednessdl} and~\ref{cor:embeddedness}, initially embedded curves stay embedded.
Since the global term is bounded, it will vanish in any limit flow of a type-I rescaling where we rescale by the maximal curvature.
Also, since the length of the curves are bounded away from zero, the curves of any limit flow will be of infinite length.
Like in the analysis in~\cite{Huisken90} of type-I singularities of mean curvature flow, a monotonicity formula yields that any limit flow of a type-I rescaling is an embedded homothetically shrinking solution of CSF with non-vanishing curvature.
By~\cite{AbreschLanger86}, this is an embedded shrinking circle.
This contradicts the unbounded length.
\end{proof}
To investigate type-II singularities, we want to rescale the curves $\Si_t$ near a singular point as $t\to T<\infty$.
The following rescaling technique for type-II singularities was introduced in~\cite[Proof of Thm.~16.4]{Hamilton95b} for Ricci flow, and applied to type-II singularities of MCF in~\cite[p.~11]{HuiskenSinestrari99}.
Let $(p_k,t_k)_{k\in\mathbb{N}}$ be a sequence in $\cI\times[0,T-1/k]$ with
$$T_k:=\kappa^2(p_k,t_k)\left(T-\frac1k-t_k\right)=\max_{(p,t)\in\mathbb{S}^1\!\times[0,T-1/k]}\left(\kappa^2(p,t)\left(T-\frac1k-t\right)\right)$$
for each $k\in\mathbb{N}$.
We set $\lambda_k^2:=\kappa^2(p_k,t_k)$, $\alpha_k:=-\lambda_k^2t_k$ and define the rescaled embeddings $X_k:\cI\times[\alpha_k,T_k]\to\mathbb{R}^2$ by
\begin{align}\label{eq:FkII}
X_k(p,\tau):=\lambda_k\left(X\!\left(p,t_k+\frac\tau{\lambda_k^2}\right)-X(p_k,t_k)\right)\,.
\end{align}
\begin{Thm}
\label{thm:Finftycsf}
Let $X:\cI\times(0,T)\to\mathbb{R}^2$ be a smooth, embedded solution of~\eqref{eq:ccf} with $T<\infty$ and satisfying~\eqref{eq:intabskappa} for $\cI=\mathbb{R}$ and~\eqref{eq:h_2} for $\cI\in\{\mathbb{S}^1,\mathbb{R}\}$.
Then there exists a sequence of intervals $0\in I_k\subset\mathbb{R}$ and rescaled embeddings
$$\left(\bar X_{k}:I_k\times[\alpha_k,T_k]\to\mathbb{R}^2\right)_{k\in\mathbb{N}}$$
that converges for $k\to\infty$ along a subsequence, uniformly and smoothly on compact subsets $I\times J\subset\mathbb{R}\times\mathbb{R}$ with $0\in I$ and compact subsets in $\mathbb{R}^2$ to a maximal, smooth, strictly convex or strictly concave limit solution $X_{\infty}:\mathbb{R}\times\mathbb{R}\to\mathbb{R}^2$ which satisfies
$$\fracp{X_{\infty}}{\tau}(p,\tau)=-\kappa_{\infty}(p,\tau)\boldsymbol{\nu}_{\infty}(p,\tau)\,.$$
Moreover, $L(\Sigma_\tau^{\infty})=\infty$ for all $\tau\in\mathbb{R}$, $X_{\infty}(0,0)=0$, $\sup_{\mathbb{R}\times\mathbb{R}}|\kappa_{\infty}|=|\kappa_{\infty}(0,0)|=1$.
\end{Thm}
\begin{proof}
The convergence follows similar lines to those of~\cite[Rem.~4.22(2)]{Ecker04} and~\cite[Prop.~4.7]{Maeder15}.
For details, see also~\cite[Thm.~9.13]{Dittberner18}.
The strict convexity/concavity is proofed like in~\cite[Thms.~5.14 and~7.7]{Altschuler91}, where we use that, away from $T$, the coefficients in the evolution equation for the curvature are bounded and that $\int_{\Si_t}|\kappa|\,ds_t<\infty$ (see property~\eqref{eq:intabskappa} for $\cI=\mathbb{R}$).
A more detailed proof can be found in~\cite[Prop.~4.3.2]{Mantegazza11} or in~\cite[Prop.~9.16]{Dittberner18}.
\end{proof}
We now can proceed as in~\cite[Thm.~2.4]{Huisken95}.
\begin{Thm}
\label{thm:notypeIIsing}
Let $\Sigma_0=X_0(\cI)$ be a smooth, embedded curve satisfying~\eqref{eq:intkappageqminuspi}.
Let $X:\cI\times[0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} with $T<\infty$, initial curve $\Sigma_0$ and satisfying~\eqref{eq:alpha},~\eqref{eq:asymp} and~\eqref{eq:intabskappa} for $\cI=\mathbb{R}$ and~\eqref{eq:h_2} for $\cI\in\{\mathbb{S}^1,\mathbb{R}\}$.
Then a type-II singularity cannot form at~$T$.
\end{Thm}
\begin{proof}
Theorem~\ref{thm:Finftycsf} yields that the limit flow consists of strictly convex or concave curves $\Sigma^{\infty}_\tau$ for $\tau\in\mathbb{R}$ satisfying $\sup_{\mathbb{R}\times\mathbb{R}}|\kappa_{\infty}|=|\kappa_{\infty}(0,0)|=1$.
If $\kappa_{\infty}<0$, we change the direction of parametrisation so that $\kappa_{\infty}>0$.
Since the curvature attains its maximum at the point $(0,0)\in\mathbb{R}\times\mathbb{R}$,~\cite[Main Theorem~B]{Hamilton95a} yields that $X_{\infty}$ is a translating solution of CSF.
~\cite[Thm.~8.16]{Altschuler91} implies that $\Sigma^{\infty}_\tau$ is the grim reaper for every $\tau\in\mathbb{R}$.
The grim reaper is asymptotic to two parallel lines of distance $\pi$ from inside.
Let $\tau\in\mathbb{R}$.
We can find a sequence of points $(p_j,q_j)_{j\in\mathbb{N}}$ in $\mathbb{R}\times\mathbb{R}$ with $d_{\infty}(p_j,q_j,\tau)\leq\pi$ for all $j\in\mathbb{N}$ and $l_{\infty}(p_j,q_j,\tau)\to\infty$ for $j\to\infty$.
Hence,
$$\inf_{\mathbb{R}\times\mathbb{R}}\frac{d_{\infty}}{l_{\infty}}(\,\cdot\,,\,\cdot\,,\tau)=0\,.$$
However, like in~\cite[Thms.~2.4 and~2.5]{Huisken95} (for details, see~\cite[Thm.~9.21]{Dittberner18}), the lower bound $\inf_{\mathbb{R}\times\mathbb{R}\times[0,T)}(d/l)\geq c$ from Theorem~\ref{thm:minimumdl}, respectively the lower bound $\inf_{\mathbb{S}^1\!\times\mathbb{S}^1\!\times[0,T)}(d/\psi)\geq c$ from Theorem~\ref{thm:minimumdpsi} imply that
$$\inf_{\mathbb{R}\times\mathbb{R}}\frac{d_{\infty}}{l_{\infty}}(\,\cdot\,,\,\cdot\,,\tau)\geq c$$
for every limit flow of rescalings according to~\eqref{eq:FkII}.
\end{proof}
\begin{Cor}\label{cor:T=infty}
Let $\Sigma_0=X_0(\cI)$ be a smooth, embedded curve satisfying~\eqref{eq:intkappageqminuspi}.
Let $X:\cI\times[0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$ and satisfying~\eqref{eq:alpha},~\eqref{eq:asymp} and~\eqref{eq:intabskappa} for $\cI=\mathbb{R}$ and~\eqref{eq:h_2} for $\cI\in\{\mathbb{S}^1,\mathbb{R}\}$.
Then $T=\infty$.
\end{Cor}
\begin{proof}
By Theorems~\ref{thm:notypeIsing} and~\ref{thm:notypeIIsing} neither a type-I nor a type-II singularity can form at $T$ so that curvature stays bounded on $[0,T]$ by a constant $C(\Sigma_0,T)$.
We can extend the flow beyond $T$ and repeat the above argument.
Hence, for every time $T'<\infty$, there exists a constant $C(\Sigma_0,T')<\infty$ so that $\max_{p\in\mathbb{S}^1}|\kappa(p,t)|\leq C$ for all $t\in[0,T')$.
Applying Proposition~\ref{prop:T<infty} yields that the short time solution can be extended to a smooth solution on $(0,\infty)$.
\end{proof}
\section{Convexity in finite time}\label{sec:convexity}
In this section, we show that a smooth, embedded solution $X:\mathbb{S}^1\!\times(0,\infty)\to\mathbb{R}^2$ of~\eqref{eq:ccf} with a global term $h$ satisfying~\eqref{eq:h_3a} becomes convex in finite time.
\begin{Rem}\label{rem:h}
We observe that, by Lemma~\ref{lem:dtL},
\begin{align}\label{eq:hdtAdtL}
h=\frac1L\left(2\pi+\fracd{A}{t}\right)
=\frac1{2\pi}\left(\int_{\Si_t}\kappa^2\,ds_t+\fracd{L}{t}\right)
\end{align}
and
\begin{align}\label{eq:L2piintkappa}
-\fracd{}{t}\left(\frac{L^2}{4\pi}-A\right)
=\frac L{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi\,.
\end{align}
In respect of~\eqref{eq:L2piintkappa}, choose
\begin{align}\label{eq:dtAa}
\fracd{A}{t}=\gamma\left(\frac L{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi\right)
\end{align}
and
\begin{align}\label{eq:dtLa}
\frac L{2\pi}\fracd{L}{t}
=\frac1{4\pi}\fracd{L^2}{t}
=-(1-\gamma)\left(\frac L{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi\right)\,,
\end{align}
where $\gamma\in\mathbb{R}$.
Then,~\eqref{eq:hdtAdtL} yields
$$h=(1-\gamma)\frac{2\pi}{L}+\frac{\gamma}{2\pi}\int_{\Si_t}\kappa^2\,ds_t\,.$$
For arbitrary $\gamma<0$, however, the positivity of $h$ is not guaranteed.
\end{Rem}
\begin{Lemma}\label{lem:AL}
Let $X:\mathbb{S}^1\!\times[0,T)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$ and $h$ satisfying~\eqref{eq:h_3a}.
Then, $A$ and $L$ are monotone and there exist constants $0<c<C<\infty$ such that $c\leq A,L\leq C$ on $[0,T)$ and
$$\frac{L}{2\pi}\int_{\Sigma_\tau}\kappa^2\,ds_\tau-2\pi\in L^1([0,T))\,.$$
\end{Lemma}
\begin{proof}
By~\eqref{eq:dtAa} and~\eqref{eq:dtLa},
\begin{align}\label{eq:dtAdtL2gamma}
(1-\gamma)\fracd{A}{t}=-\frac\gamma{4\pi}\fracd{L^2}{t}
\end{align}
so that, with $\delta\in(0,\infty)$ and
$$\gamma=\frac{(\delta-1)A_0}{L_0^2/4\pi-A_0}\in\left(-\frac{A_0}{L_0^2/4\pi-A_0},\infty\right)\,,$$
integrating~\eqref{eq:dtAdtL2gamma} yields
\begin{align}\label{eq:AL2gamma}
(1-\gamma)A_t+\gamma\frac{L^2_t}{4\pi}
=(1-\gamma)A_0+\gamma\frac{L^2_0}{4\pi}
=\delta A_0
\end{align}
for all $t\in(0,T)$.
For $\delta\in(0,1)$, we have $\gamma<0$ and $-(1-\gamma)<0$, so that by~\eqref{eq:dtAa} and~\eqref{eq:dtLa},
$$\fracd{A}{t}<0
\qquad\text{ and }\qquad
\fracd{L}{t}<0\,.$$
Hence, $A$ and $L$ are uniformly bounded away from infinity.
By~\eqref{eq:AL2gamma},
$$(1-\gamma)A_t
>(1-\gamma)A_t+\gamma\frac{L^2_t}{4\pi}
=\delta A_0$$
and so that, by the isoperimetric inequality, $A$ and $L$ are uniformly bounded away from zero.
For $\delta\in[1,L_0^2/4\pi A_0]$, we have $\gamma\in[0,1]$ and by~\eqref{eq:dtAa} and~\eqref{eq:dtLa},
$$\fracd{A}{t}\geq0
\qquad\text{ and }\qquad
\fracd{L}{t}\leq0\,.$$
Hence, $A$ and $L$ are uniformly bounded away from zero and infinity.
For $\delta>L_0^2/4\pi A_0$, we have $\gamma>1$ and by~\eqref{eq:dtAa} and~\eqref{eq:dtLa},
$$\fracd{A}{t}>0
\qquad\text{ and }\qquad
\fracd{L}{t}>0\,.$$
Hence, $A$ and $L$ are uniformly bounded away from zero.
By~\eqref{eq:AL2gamma},
$$\gamma\frac{L^2_t}{4\pi}
<(1-\gamma)A_t+\gamma\frac{L^2_t}{4\pi}
=\delta A_0$$
and so that $A$ and $L$ are uniformly bounded away from infinity.
The uniform bounds on the area and length from above and~\eqref{eq:L2piintkappa} yield
\[\frac{L}{2\pi}\int_{\Sigma_\tau}\kappa^2\,ds_\tau-2\pi\,d\tau\in L^1([0,T))\,.\qedhere\]
\end{proof}
Like in~\cite[Section~7]{Maeder15}, we use the following Gagliardo--Nirenberg interpolation inequality.
\begin{Thm}[Gagliardo--Nirenberg interpolation inequality,~{\cite[pp.~125]{Nirenberg59}}, see also~{\cite[Thm.~3.70]{Aubin98}}]\label{thm:GN}
Let $f\in C^\infty(\mathbb{S}^1)$.
Let $p>2$ and $\sigma\in[0,1)$ with $\sigma=1/2-1/p$.
Then there exist constants $C_1=c_1(p,\sigma)$ and $C_2=c_2(p,\sigma)$ such that
\begin{align*}
\left(\int_{\mathbb{S}^1}|f|^p\,dx\right)^{1/p}
&\leq C_1\left(\int_{\mathbb{S}^1}\left(\fracd{f}{x}\right)^{2}dx\right)^{\sigma/2}
\left(\int_{\mathbb{S}^1} f^2\,dx\right)^{(1-\sigma)/2} \notag\\
&\quad+C_2\left(\int_{\mathbb{S}^1} f^2\,dx\right)^{1/2}\,.
\end{align*}
\end{Thm}
\begin{Lemma}[\hspace{-.005em}{\cite[see proof of Cor.~7.5]{Maeder15}}]\label{lem:dtfintffto0}
Let $f\in C^1((0,\infty))\cap L^1((0,\infty))$ with $f\geq0$ and $\fracd{}{t}f\leq C(C+f)^3$ for $C\geq0$.
Then $f(t)\to0$ for $t\to\infty$.
\end{Lemma}
\begin{Lemma}\label{lem:dthlp}
Let $X:\mathbb{S}^1\!\times(0,\infty)\to\mathbb{R}^2$ be a smooth, embedded solution of~\eqref{eq:ccf} and $h$ satisfying~\eqref{eq:h_3a}.
Then there exists a constant $c>0$ such that, for all $t\in(0,\infty)$,
$$\fracd{}{t}\left(\frac{L}{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi\right)
\leq C\left(C+\frac{L}{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi\right)^3\,.$$
\end{Lemma}
\begin{proof}
By the isoperimetric inequality we can estimate
\begin{align}\label{eq:dtintkappa_33}
h\leq C\left(1+\int_{\Si_t}\kappa^2\,ds_t\right)\,,
\end{align}
where $C=C(\gamma)$.
We deduce with Theorem~\ref{thm:GN} for $p=4$ and $\sigma=1/4$ and Young's inequality for $p=4/3$ and $q=4$ as well as for $p=q=2$,
\begin{align}\label{eq:dtintkappa_4}
\int_{\Si_t}\kappa^4\,ds_t
&\leq\left(C\left(\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}\,ds_t\right)^{1/8}
\left(\int_{\Si_t}\kappa^2\,ds_t\right)^{3/8}
+C\left(\int_{\Si_t}\kappa^2\,ds_t\right)^{1/2}\right)^4 \notag\\
&\leq\delta\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}\,ds_t
+C(\delta)\left(\int_{\Si_t}\kappa^2\,ds_t\right)^3
+C\left(\int_{\Si_t}\kappa^2\,ds_t\right)^{2}
\end{align}
for a constants $C>0$.
Again, by Theorem~\ref{thm:GN} for $p=3$ and $\sigma=1/6$ and Young's inequality for $p=4$ and $q=3/4$,
\begin{align}
&\int_{\Si_t}\kappa^3\,ds_t \notag\\
&\quad\leq\left(C\left(\int_{\Si_t}\left(\fracp{\kappa}{p}\right)^{2}\,ds_t\right)^{1/12}
\left(\int_{\Si_t}\kappa^2\,ds_t\right)^{5/12}
+C\left(\int_{\Si_t}\kappa^2\,ds_t\right)^{1/2}\right)^3\label{eq:dtintkappa_3}\\
&\quad\leq\delta\int_{\Si_t}\left(\fracp{\kappa}{p}\right)^{2}\,ds_t
+C(\delta)\left(\int_{\Si_t}\kappa^2\,ds_t\right)^{5/3}
+C\left(\int_{\Si_t}\kappa^2\,ds_t\right)^{3/2}\label{eq:dtintkappa_3b}
\end{align}
so that with Young's inequality for $p=3/2$ and $q=3$ as well as for $p=4$ and $q=4/3$, as well as~\eqref{eq:dtintkappa_3},
\begin{align}\label{eq:dtintkappa_32}
&\int_{\Si_t}\kappa^2\,ds_t\int_{\Si_t}\kappa^3\,ds_t \notag\\
&\quad\leq C\left(\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}\,ds_t\right)^{1/4}
\left(\int_{\Si_t}\kappa^2\,ds_t\right)^{9/4}
+C\left(\int_{\Si_t}\kappa^2\,ds_t\right)^{5/2} \notag\\
&\quad\leq\delta\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}\,ds_t
+C(\delta)\left(\int_{\Si_t}\kappa^2\,ds_t\right)^3
+C\left(\int_{\Si_t}\kappa^2\,ds_t\right)^{5/2}.
\end{align}
We use Lemma~\ref{lem:evolutionequations},~\eqref{eq:dtintkappa_33},~\eqref{eq:dtintkappa_4},~\eqref{eq:dtintkappa_3b},~\eqref{eq:dtintkappa_32} and integration by parts to calculate with $\delta=1/3$,
\begin{align}\label{eq:dtintkappa2}
\fracd{}{t}\int_{\Si_t}\kappa^2\,ds_t
&=2\int_{\Si_t}\left(\kappa\fracp{^2\kappa}{s^2}-(h-\kappa)\kappa^3\right)ds_t
+\int_{\Si_t}\kappa^3(h-\kappa)\,ds_t \notag\\
&=-2\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}ds_t
-h\int_{\Si_t}\kappa^3\,ds_t
+\int_{\Si_t}\kappa^4\,ds_t \notag\\
&\leq C\sum_{i=1}^5\left(\int_{\Si_t}\kappa^2\,ds_t\right)^{p_i}
\leq C\left(1+\int_{\Si_t}\kappa^2\,ds_t\right)^3
\end{align}
for all $t\in(0,\infty)$, where $p_i\in\{2,3/2,5/3,5/2,3\}$.
By Lemmata~\ref{lem:dtL} and~\ref{lem:AL},~\eqref{eq:dtintkappa_33} and~\eqref{eq:dtintkappa2},
\begin{align*}
&\fracd{}{t}\left(\frac{L}{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi\right)
=\frac1{2\pi}\fracd{L}{t}\int_{\Si_t}\kappa^2\,ds_t
+\frac L{2\pi}\fracd{}{t}\int_{\Si_t}\kappa^2\,ds_t \notag\\
&\quad\leq \frac1{2\pi}\left(2\pi h-\int_{\Si_t}\kappa^2\,ds_t\right)\int_{\Si_t}\kappa^2\,ds_t
+C\frac{L}{2\pi}\left(1+\int_{\Si_t}\kappa^2\,ds_t\right)^3 \notag\\
&\quad\leq C\left(1+\int_{\Si_t}\kappa^2\,ds_t\right)^2
+C\left(1+\int_{\Si_t}\kappa^2\,ds_t\right)^3 \notag\\
&\quad\leq C\left(C+\frac{L}{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi\right)^3\,.\qedhere
\end{align*}
\end{proof}
\begin{Lemma}\label{lem:inthkappa}
Let $X:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}^2$ be a smooth, embedded solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$ and $h$ satisfying~\eqref{eq:h_3a}.
Then,
$$Lh\to2\pi
\qquad\text{ and }\qquad
\frac{L}{2\pi}\int_{\Si_t}\kappa^2\,ds_t\to2\pi$$
for $t\to\infty$, and there exist a time $t_0\geq0$ and constants $0<c<C<\infty$ such that $\inf_{[t_0,\infty)}h\geq c$ and
$$\sup_{[0,\infty)}h
+\sup_{[0,\infty)}\left|\fracd{h}{t}\right|
+\sup_{[0,\infty)}\int_{\Si_t}\kappa^2\,ds_t
\leq C\,.$$
\end{Lemma}
\begin{proof}
Lemmata~\ref{lem:AL},~\ref{lem:dtfintffto0} and~\ref{lem:dthlp} yield that
\begin{align}\label{eq:L2piintkappa2to2pi}
\frac{L}{2\pi}\int_{\Si_t}\kappa^2\,ds_t\to2\pi
\end{align}
for $t\to\infty$.
By Lemma~\ref{lem:AL}, $L$ is bounded away from zero and infinity.
By~\eqref{eq:dtAa}
and~\eqref{eq:L2piintkappa2to2pi},
$$\fracd{A}{t}=\int_{\Si_t}(h-\kappa)\,ds_t=Lh-2\pi\to0$$
for $t\to\infty$.
Hence, there exist a time $t_0\in[0,\infty)$ and constants $0<c<C<\infty$ so that $c\leq h\leq C$ on $[t_0,\infty)$.
By~\eqref{eq:dtLa}
and~\eqref{eq:L2piintkappa2to2pi},
$$\fracd{L}{t}=\int_{\Si_t}\kappa^2\,ds_t-2\pi h\to0$$
for $t\to\infty$ so that there exist $0<C<\infty$ with $\big|\fracd{}{t}L\big|+\int_{\Si_t}\kappa^2\,ds_t\leq C$ on $[0,\infty)$.
This yields
\[\left|\fracd{h}{t}\right|
=C\left(\left|\fracd{L}{t}\right|+\int_{\Si_t}\kappa^2\,ds_t+\left|\fracd{g}{t}\right|\right)
\leq C\,.\qedhere\]
\end{proof}
\begin{Lemma}\label{lem:dtinthkappa2}
Let $X:\mathbb{S}^1\!\times(0,\infty)\to\mathbb{R}^2$ be a smooth, embedded solution of~\eqref{eq:ccf}.
Then there exists a constant $C>0$ such that
\begin{align*}
\fracd{}{t}\int_{\Si_t}(h-\kappa)^2\,ds_t
&\leq-\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}\,ds_t
+C\fracd{h}{t}\int_{\Si_t}(h-\kappa)\,ds_t \\
&\quad+C\sum_{i=1}^5\left(\int_{\Si_t}(h-\kappa)^2\,ds_t\right)^{p_i}h^{q_i}
\end{align*}
for all $t\in(0,\infty)$, where $p_i\in[1,3]$ and $q_i\in[0,2]$.
\end{Lemma}
\begin{proof}
We follow the lines of~\cite[Lems.~7.3 and~7.4]{Maeder15}.
Write $\kappa=h-(h-\kappa)$.
Then
$$(h-\kappa)^3\kappa=h(h-\kappa)^3-(h-\kappa)^4$$
and
$$(h-\kappa)^2\kappa^2
=h^2(h-\kappa)^2-2h(h-\kappa)^3+(h-\kappa)^4\,.$$
Lemma~\ref{lem:evolutionequations} and integration by parts yields
\begin{align}\label{eq:dtinthminuskappa_1}
&\fracd{}{t}\int_{\Si_t}(h-\kappa)^2\,ds_t \notag\\
&\,=\int_{\Si_t}(h-\kappa)^3\kappa\,ds_t
+2\fracd{h}{t}\int_{\Si_t}(h-\kappa)\,ds_t
+2\int_{\Si_t}(h-\kappa)
\left(-\fracp{^2\kappa}{s^2}+(h-\kappa)\kappa^2\right)\,ds_t \notag\\
&\,=-2\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}ds_t
+2\fracd{h}{t}\int_{\Si_t}(h-\kappa)\,ds_t
+\int_{\Si_t}(h-\kappa)^4\,ds_t \notag\\
&\qquad-3h\int_{\Si_t}(h-\kappa)^3\,ds_t
+2h^2\int_{\Si_t}(h-\kappa)^2\,ds_t\,.
\end{align}
Like in~\cite[Cor.~7.4]{Maeder15}, we use Theorem~\ref{thm:GN} with $p=4$ and $\sigma=1/4$ and Young's inequality with $p=4/3$ and $q=4$ as well as for $p=q=2$, to estimate
\begin{align}\label{eq:dtinthminuskappa_2}
&\int_{\Si_t}(h-\kappa)^4\,ds_t \notag\\
&\,\leq\left(C\left(\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}\,ds_t\right)^{1/8}
\left(\int_{\Si_t}(h-\kappa)^2\,ds_t\right)^{3/8}
+C\left(\int_{\Si_t}(h-\kappa)^2\,ds_t\right)^{1/2}\right)^4 \notag\\
&\,\leq\frac12\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}\,ds_t
+C\left(\int_{\Si_t}(h-\kappa)^2\,ds_t\right)^3
+C\left(\int_{\Si_t}(h-\kappa)^2\,ds_t\right)^{2}.
\end{align}
Again by Theorem~\ref{thm:GN} with $p=3$ and $\sigma=1/6$ and Young's inequality for $p=3/2$ and $q=3$ as well as for $p=4$ and $q=4/3$ we obtain
\begin{align}\label{eq:dtinthminuskappa_3}
3h\int_{\Si_t}(h-\kappa)^3\,ds_t
&\leq 3h\left(C\left(\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}\,ds_t\right)^{1/12}
\left(\int_{\Si_t}(h-\kappa)^2\,ds_t\right)^{5/12}\right. \notag\\
&\hspace{4em}+\left.{\vphantom{C\left(\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}\,ds_t\right)^{1/12}
\left(\int_{\Si_t}(h-\kappa)^2\,ds_t\right)^{5/12}}}{} C\left(\int_{\Si_t}(h-\kappa)^2\,ds_t\right)^{1/2}\right)^3 \notag\\
&\leq\frac12\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}\,ds_t
+Ch^{4/3}\left(\int_{\Si_t}(h-\kappa)^2\,ds_t\right)^{5/3} \notag\\
&\quad+Ch\left(\int_{\Si_t}(h-\kappa)^2\,ds_t\right)^{3/2}.
\end{align}
Altogether,~\eqref{eq:dtinthminuskappa_1},~\eqref{eq:dtinthminuskappa_2},~\eqref{eq:dtinthminuskappa_3} yield the claim.
\end{proof}
\begin{Lemma}\label{lem:dtinthkappa}
Let $X:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}^2$ be a smooth, embedded solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$ and $h$ satisfying~\eqref{eq:h_3a}.
Then
$$\int_{\Si_t}(h-\kappa)^2\,ds_t\to0$$
for $t\to\infty$ and
$$\int_0^\infty\int_{\Si_t}(h-\kappa)^2\,ds_t\,dt+\int_0^\infty\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}\,ds_t\,dt<\infty\,.$$
\end{Lemma}
\begin{proof}
Similar to~\cite[p.~47]{Huisken87}, Lemma~\ref{lem:dtL} yields
$$\int_{\Si_t}(h-\kappa)^2\,ds_t
=Lh^2-4\pi h+\int_{\Si_t}\kappa^2\,ds_t
=h\fracd{A}{t}-\fracd{L}{t}\,.$$
Lemmata~\ref{lem:AL} and~\ref{lem:inthkappa} imply for $0<\varepsilon<\tau<\infty$,
$$\int_\varepsilon^\tau\int_{\Si_t}(h-\kappa)^2\,ds_t\,dt
=\sup_{t\in[\varepsilon,\tau]}h(t)(A_\tau-A_\varepsilon)+(L_\varepsilon-L_\tau)\leq C\,.$$
We let $\varepsilon\to0$ and $\tau\to\infty$ to obtain
\begin{align}\label{eq:intinthkappa2}
\int_0^\infty\int_{\Si_t}(h-\kappa)^2\,ds_t\,dt<\infty\,.
\end{align}
By Lemma~\ref{lem:inthkappa},
$$\fracd{h}{t}\int_{\Si_t}(h-\kappa)\,ds_t\,dt
\leq\sup_{[0,\infty)}\left|\fracd{h}{t}\right||Lh-2\pi|
\leq C\,,$$
so that Lemma~\ref{lem:dtinthkappa2} implies
$$\fracd{}{t}\int_{\Si_t}(h-\kappa)^2\,ds_t
\leq C\left(1+\int_{\Si_t}(h-\kappa)^2\,ds_t\right)^3\,.$$
Like in~\cite[Cor.~7.5]{Maeder15}, Lemma~\ref{lem:dtfintffto0} yields
$$\int_{\Si_t}(h-\kappa)^2\,ds_t\to0$$
for $t\to\infty$.
Consequently, there exists a time $t_0\geq0$ so that
$$\int_{\Si_t}(h-\kappa)^2\,ds_t<1$$
for all $t>t_0$, and thus
\begin{align}\label{eq:inthkappa3}
\left(\int_{\Si_t}(h-\kappa)^2\,ds_t\right)^{p}\leq\int_{\Si_t}(h-\kappa)^2\,ds_t
\end{align}
for all $p\geq1$ and $t>t_0$.
By Lemma~\ref{lem:AL}, $\fracd{}{t}A$ has a sign so that
$$\int_\varepsilon^\tau\left|\int_{\Si_t}(h-\kappa)\,ds_t\right|\,dt
=|A_\tau-A_\varepsilon|
\leq C\,,$$
where $C>0$ is independent of time.
Sending $\varepsilon\to0$ and $\tau\to\infty$ yields with Lemma~\ref{lem:inthkappa},
$$\int_0^\infty\fracd{h}{t}\int_{\Si_t}(h-\kappa)\,ds_t\,dt
\leq\sup_{[0,\infty)}\left|\fracd{h}{t}\right|\int_0^\infty\left|\int_{\Si_t}(h-\kappa)\,ds_t\right|\,dt
\leq C.$$
Thus, with Lemma~\ref{lem:dtinthkappa2},~\eqref{eq:intinthkappa2} and~\eqref{eq:inthkappa3} we obtain
\begin{align*}
\int_{t_0}^\infty\int_{\Si_t}\left(\fracp{\kappa}{s}\right)^{2}\,ds_t\,dt
\leq\int_{\Sigma_{t_0}}(h-\kappa)^2\,ds_t\,dt
+C+C\int_{t_0}^\infty\int_{\Si_t}(h-\kappa)^2\,ds_t\,dt<\infty\,.
\end{align*}
Since $\Si_t$ is smooth for $t\in[0,t_0]$, the claim follows.
\end{proof}
\begin{Thm}
\label{thm:convexity}
Let $X:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}^2$ be a smooth, embedded solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$ and $h$ satisfying~\eqref{eq:h_3a}.
Then there exists a time $T_0\geq0$ such that $\Si_t$ is strictly convex for $t>T_0$.
\end{Thm}
\begin{proof}
By Lemma~\ref{lem:inthkappa},
\begin{align}\label{eq:hleqpiL}
h\geq c_h>0
\end{align}
on $[t_0,\infty)$ for $t_0\geq0$ and $c_h>0$.
Lemma~\ref{lem:dtinthkappa} implies that there exists a sequence $(t_k)_{k\in\mathbb{N}}$ with $t_k\to\infty$ for $k\to\infty$ so that
\begin{align}\label{eq:dpkappato0}
\int_{\Sigma_{t_k}}\left(\fracp{\kappa}{s}\right)^{2}ds_{t_k}\to0
\end{align}
for $k\to\infty$.
Hence, there exists $k_0\in\mathbb{N}$ so that for all $k\geq k_0$
\begin{align}\label{eq:dpkappaC1}
\int_{\Sigma_{t_k}}\left(\fracp{\kappa}{s}\right)^{2}ds_{t_k}<1\,.
\end{align}
We employ~{\cite[Thm.~7.26(ii)]{GilbargTrudinger83}} to obtain that $W^{1,2}(\mathbb{S}^1)$ is compactly embedded in $C^0(\mathbb{S}^1)$.
Furthermore, $C^0(\mathbb{S}^1)\subset L^2(\mathbb{S}^1)$, and
$\Vert f\Vert_{L^2(\mathbb{S}^1)}
\leq\sqrt{2\pi}\Vert f\Vert_{C^0(\mathbb{S}^1)}$
for every $f\in C^0(\mathbb{S}^1)$.
Hence, $C^0(\mathbb{S}^1)$ is continuously embedded in $L^2(\mathbb{S}^1)$.
Let $f\in W^{1,2}(\mathbb{S}^1)$.
By Ehrling's lemma, for all $\varepsilon>0$ there exists a constant $C(\varepsilon)>0$ so that
\begin{align}\label{eq:Ehrling_W12}
\Vert f\Vert_{C^0(\mathbb{S}^1)}\leq\varepsilon\Vert f\Vert_{W^{1,2}(\mathbb{S}^1)}+C(\varepsilon)\Vert f\Vert_{L^2(\mathbb{S}^1)}\,.
\end{align}
Lemma~\ref{lem:dtinthkappa} and~\eqref{eq:dpkappato0} yield $h(t_k)-\kappa(\,\cdot\,,t_k)\in W^{1,2}(\mathbb{S}^1)$ for each $k\in\mathbb{N}$.
Hence, we can use~\eqref{eq:dpkappaC1} and~\eqref{eq:Ehrling_W12} to estimate
\begin{align}\label{eq:maxhlpkappa_0}
\max_{p\in\mathbb{S}^1}|h(t_k)-\kappa(p,t_k)|
&\leq\varepsilon\left(\int_{\Sigma_{t_k}}\left(\fracp{\kappa}{s}\right)^{2}\,ds_{t_k}\right)^{1/2}
+\varepsilon\left(\int_{\Sigma_{t_k}}(h-\kappa)^2\,ds_{t_k}\right)^{1/2} \notag\\
&\quad+C(\varepsilon)\left(\int_{\Sigma_{t_k}}(h-\kappa)^2\,ds_{t_k}\right)^{1/2} \notag\\
&\leq\varepsilon+C(\varepsilon)\left(\int_{\Sigma_{t_k}}(h-\kappa)^2\,ds_{t_k}\right)^{1/2}
\end{align}
for all $k\geq k_0$.
Choose $\varepsilon=c_h/4$ to deduce with~\eqref{eq:maxhlpkappa_0}
\begin{align}\label{eq:maxhlpkappa}
\max_{p\in\mathbb{S}^1}|h(t_k)-\kappa(p,t_k)|
\leq\frac{c_h}4+C\left(\int_{\Sigma_{t_k}}(h-\kappa)^2\,ds_{t_k}\right)^{1/2}\,.
\end{align}
Lemma~\ref{lem:dtinthkappa} implies that there exists $k_1\geq k_0$ so that for all $k\geq k_1$
$$\int_{\Sigma_{t_k}}(h-\kappa)^2\,ds_{t_k}<\left(\frac{c_h}{4C}\right)^2\,.$$
By~\eqref{eq:maxhlpkappa},
$$\max_{p\in\mathbb{S}^1}|h(t_{k_1})-\kappa(p,t_{k_1})|\leq\frac{c_h}2\,.$$
With~\eqref{eq:hleqpiL}, we conclude that $\kappa>0$ at $t_{k_1}$.
From Corollary~\ref{cor:strongmaxpkappa} it follows that $\kappa>0$ for all $t>t_{k_1}$.
Hence, the claim holds for $T_0=t_{k_1}$.
\end{proof}
\section{Longtime behaviour}\label{sec:longtimebehaviour}
In this section we show that convex solutions of~\eqref{eq:ccf} that exist for all positive times converge exponentially and smoothly to a round circle.
This was already shown in~\cite{Gage86} for the APCSF and in~\cite{Pihan98} for the LPCF.
We repeat and extend the arguments here for $h$ satisfying~\eqref{eq:h_3a}
for the sake of completeness.
We mostly follow the lines of~\cite[Section~5]{GageHamilton86} for rescaled convex CSF,~\cite{Gage86} for convex APCSF, and~\cite[Chapter~7]{Pihan98} for convex LPCF.
For further details, see~\cite[Chapter~11]{Dittberner18}.
\begin{Lemma}[Isoperimetric inequality,~\cite{Gage83}]\label{lem:intkappaLA}
For a closed, convex $C^2$-curve in the plane,
$$\int_{\Sigma}\kappa^2\,ds\geq\frac{\pi L}{A}$$
with equality if and only if the curve is a circle.
\end{Lemma}
\begin{Lemma}
\label{lem:isoleqexp}
Let $\Sigma_0=X_0(\mathbb{S}^1)$ be a smooth, embedded, convex curve.
Let $X:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$.
Then there exists a constant $C=C(\Sigma_0)>0$, such that, for all $t>0$,
$$\left(\frac{L^2}{A}-4\pi\right)\leq C\exp\!\left(-\int_0^t\frac{2\pi}{A}\,d\tau-\log\frac{A_t}{A_0}\right)\,.$$
\end{Lemma}
\begin{proof}
We follow the lines of~{\cite[Cor.~2.4]{Gage86}} and~{\cite[Lem.~7.7]{Pihan98}} and use Lemma~\ref{lem:intkappaLA} to estimate for $t>0$
\begin{align*}
\fracd{}{t}\!\left(\frac{L^2}{A}-4\pi\right)
&=-\frac{2L}{A}\left(\int_{\Si_t}\kappa^2\,ds_t-2\pi h\right)
-\frac{L^2}{A^2}\left(Lh-2\pi\right) \\
&\leq-\frac{L}{A}\left(\frac{2\pi L}A-4\pi h+\frac{L^2}{A}h-\frac{2\pi L}{A}\right)
=-\frac{hL}{A}\left(\frac{L^2}{A}-4\pi\right)\,.
\end{align*}
By~\eqref{eq:hdtAdtL},
\[\int_0^t\frac{hL}{A}\,d\tau
=\int_0^t\frac{2\pi}{A}+\fracd{}{t}\log A\,d\tau
=\int_0^t\frac{2\pi}{A}\,d\tau+\log\frac{A_t}{A_0}\,.\qedhere\]
\end{proof}
\begin{Prop}[Bonnesen isoperimetric inequality,~{\cite[Thm.~4\,(21)]{Osserman79}}]\label{prop:Osserman}
For an embedded, closed curve $\Sigma$ in the plane,
$$\frac{L^2}{A}-4\pi\geq\frac{\pi^2}{A}(r_{\cir}-r_{\inner})^2\geq0\,,$$
where $r_{\cir}$ and $r_{\inner}$ are the circumscribed and inscribed radius of $\Sigma$.
\end{Prop}
\begin{Prop}
\label{prop:C0convergence}
Let $\Sigma_0=X_0(\mathbb{S}^1)$ be a smooth, embedded, convex curve.
Let $X:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$ and $h$ satisfying~\eqref{eq:h_3a}.
Then
$$r_{\cir}(t)-r_{\inner}(t)\to0$$
for $t\to\infty$ and $\Si_t=X(\mathbb{S}^1\!,t)$ converges in $C^0$ to a circle of radius
$$R:=\lim_{t\to\infty}\frac{L_t}{2\pi}=\lim_{t\to\infty}\sqrt{\frac{A_t}{\pi}}\in(0,\infty)\,.$$
Moreover, for all $\beta\in(0,1)$ there exist a time $t_0>0$ and a constant $C>0$ such that, for all $t\geq t_0$,
$$\left(\frac{L^2}{4\pi}-A\right)
\leq C\exp\!\left(-\frac{2\beta t}{R^2}\right)
\quad\text{ and }\quad
\frac{L}{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi
\leq C\exp\!\left(-\frac{\beta t}{R^2}\right)\,.$$
\end{Prop}
\begin{proof}
By Lemma~\ref{lem:AL},
$$\int_0^t\frac{2\pi}{A}\,d\tau+\log\frac{A_t}{A_0}
\geq\frac{2\pi t}{C}+\log\frac{c}{A_0}
\to\infty$$
for $t\to\infty$.
Thus, Lemma~\ref{lem:isoleqexp} and Proposition~\ref{prop:Osserman} yield $r_{\cir}(t)-r_{\inner}(t)\to0$ for $t\to\infty$.
For embedded, closed, convex curves, we can estimate
$$r_{\inner}=\sqrt{\frac{A_{\inner}}\pi}\leq\sqrt{\frac{A}\pi}
\leq\frac{L}{2\pi}\leq\frac{L_{\cir}}{2\pi}=r_{\cir}\,,$$
where $A_{\inner}$ is the area of the inscribed circle, and $L_{\cir}$ is the length of the circumscribed circle.
Hence,
$$\frac{L}{2\pi}-\sqrt{\frac{A}\pi}\to0$$
for $t\to\infty$.
This and the bounds from Lemma~\ref{lem:AL} imply the second claim,
since $L/2\pi=\sqrt{A/\pi}$ only holds on a circle.
Let $\beta\in(0,1)$ and $\varepsilon(\beta,R)>0$ so that
$$\left(1-\varepsilon R^2\right)\geq\beta\,.$$
We can choose $t_0(\beta)>0$ so that for all $t\geq t_0$,
$$\left(\frac1{R^2}-\varepsilon\right)\leq\frac\pi A\,.$$
Hence,
$$-\int_0^t\frac{2\pi}{A}\,d\tau
\leq-2\left(1-\varepsilon R^2\right)\frac{t}{R^2}
\leq-\frac{2\beta t}{R^2}$$
and again by the bounds on $A$ from Lemma~\ref{lem:AL},
\begin{align}\label{eq:L4piAexp}
\left(\frac{L^2}{4\pi}-A\right)
=\frac{A}{4\pi}\left(\frac{L^2}{A}-4\pi\right)
\leq C\exp\!\left(-\frac{2\beta t}{R^2}\right)
\end{align}
for all $t\geq t_0$.
Let $f\in C^2([0,\infty))$.
Since $C^2([0,\infty))$ is compactly embedded in $C^1([0,\infty))$ and $C^1([0,\infty))$ is continuously embedded in $C^0([0,\infty))$, Ehrling's Lemma yields that for every $\delta>0$ there exists $C(\delta)>0$ so that
$$\Vert f\Vert_{C^1([0,\infty))}\leq\delta\Vert f\Vert_{C^2([0,\infty))}+C(\delta)\Vert f\Vert_{C^0([0,\infty))}\,.$$
We set $\delta=1/2$ and conclude
\begin{align}\label{eq:dtfEhrlich}
\sup_{[0,\infty)}\left|\fracd{f}{t}\right|
\leq\sup_{[0,\infty)}\left|\fracd{^2f}{t^2}\right|+C\sup_{[0,\infty)}|f|\,.
\end{align}
Let $\eta>0$ and define $f_\eta:[0,\infty)\to\mathbb{R}$ by $f_\eta(t):=f(\eta t)$.
Then
$$\fracd{f_\eta}{t}(t)=\eta\fracd{f}{t}(\eta t)
\qquad\text{ and }\qquad
\fracd{^2f_\eta}{t^2}(t)=\eta^2\fracd{^2f}{t^2}(\eta t)$$
as well as with~\eqref{eq:dtfEhrlich},
\begin{align}\label{eq:dtfeta}
\sup_{[0,\infty)}\left|\fracd{f}{t}\right|
&=\frac1\eta\sup_{[0,\infty)}\left|\fracd{f_\eta}{t}\right|
\leq\frac1\eta\sup_{[0,\infty)}\left|\fracd{^2f_\eta}{t^2}\right|+\frac C\eta\sup_{[0,\infty)}|f_\eta| \notag\\
&=\eta\sup_{[0,\infty)}\left|\fracd{^2f}{t^2}\right|+\frac C\eta\sup_{[0,\infty)}|f|\,.
\end{align}
By Lemmata~\ref{lem:dthlp} and~\ref{lem:inthkappa}, there exists a time $t_1\geq t_0$ so that for all $t\geq t_1$,
\begin{align}\label{eq:L2piintkappaleqc}
\fracd{}{t}\left(\frac{L}{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi\right)
\leq C\left(C+\frac{L}{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi\right)^3
\leq C(C+1)^3\,.
\end{align}
We choose
$$\eta=\exp\!\left(-\frac{\beta t}{R^2}\right)$$
to obtain by~\eqref{eq:L2piintkappa},~\eqref{eq:L4piAexp},~\eqref{eq:dtfeta} and~\eqref{eq:L2piintkappaleqc}, for all $t\geq t_1$,
\begin{align*}
\frac{L}{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi
&\leq\sup_{[t,\infty)}\left(\frac{L}{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi\right) \\
&\leq\eta c\sup_{[t,\infty)}\left(1+\frac{L}{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi\right)^3
+\frac C\eta\sup_{[t,\infty)}\left|\frac{L^2}{4\pi}-A\right| \\
&\leq C\exp\!\left(-\frac{\beta t}{R^2}\right)\,.\qedhere
\end{align*}
\end{proof}
By Corollary~\ref{cor:strongmaxpkappa}, $\Si_t$ is strictly convex for all $t>0$.
Like introduced in Section~\ref{sec:theta2} let $\vartheta:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}$ be the angle between the $x_1$-axis and the tangent vector at the point $X(p,t)$.
Since $\Si_t$ is strictly convex on $(0,\infty)$, $\vartheta(\,\cdot\,,t)$ is injective for each $t\in(0,\infty)$.
We want to use $\vartheta$ as spatial coordinate and define $\tau$ to be a new time variable so that $\tau=t$ as well as
\begin{align}\label{eq:taut}
\fracd{\tau}{t}=1\qquad\text{ and }\qquad\fracp{\vartheta}{\tau}=0\,.
\end{align}
The spatial derivatives transforms according to
$\frac1v\fracp{}{p}=\fracp{}{s}=\kappa\fracp{}{\vartheta}$.
In the following, we use the coordinates $(\vartheta,\tau)$ on $\mathbb{S}^1\!\times(0,\infty)$.
\begin{Lemma}[\hspace{-.005em}{\cite[Lem.~4.1.3]{GageHamilton86}} and~{\cite[Lem.~6.12]{Pihan98}}]\label{lem:dtaukappa}
Let $X:\mathbb{S}^1\!\times(0,\infty)\to\mathbb{R}^2$ be a smooth, strictly convex solution of~\eqref{eq:ccf}.
Then, for $\tau\in(0,\infty)$,
$$\fracp{\kappa}{\tau}=\kappa^2\fracp{^2\kappa}{\vartheta^2}-(h-\kappa)\kappa^2\,.$$
\end{Lemma}
For $\tau>0$, we define
\begin{align}\label{eq:defmtau}
m(\tau):=\max_{\bar\tau\in[0,\tau]}\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\bar\tau)\,.
\end{align}
\begin{Lemma}
\label{lem:intdthetakappa}
Let $\Sigma_0=X_0(\mathbb{S}^1)$ be a smooth, embedded, convex curve.
Let $X:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$ and $h$ satisfying~\eqref{eq:h_3a}.
Then there exists a constant $C(\Sigma_0)>0$ such that, for all $\tau>0$,
$$\int_{\mathbb{S}^1}\left(\fracp{\kappa}{\vartheta}\right)^{2}d\vartheta
\leq\int_{\mathbb{S}^1}\kappa^2\,d\vartheta+C(m(\tau)+1)\,.$$
\end{Lemma}
\begin{proof}
We follow similar lines to~\cite[Lem.~3.4 and Cor.~3.5]{Gage86} and~\cite[Lem.~6.9]{Pihan98}.
We observe that
\begin{align}\label{eq:intkappakappamax}
0<\int_{\Si_t}\kappa^2\,ds_t
=\int_{\mathbb{S}^1}\kappa\,d\vartheta
\leq2\pi\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)
\end{align}
and use the time-independency~\eqref{eq:taut} of $\vartheta$, Lemma~\ref{lem:dtaukappa}, integration by parts to estimate
\begin{align}\label{eq:dtauintkappadthetaap_1}
\fracd{}{\tau}\int_{\mathbb{S}^1}\left(\kappa^2-\left(\fracp{\kappa}{\vartheta}\right)^{2}-2h\kappa\right)d\vartheta
&=\intsp2\left(\kappa+\fracp{^2\kappa}{^2\vartheta}-h\right)\fracp{\kappa}{\tau}\,d\vartheta
-2\fracd{h}{\tau}\int_{\mathbb{S}^1}\kappa\,d\vartheta \notag\\
&=2\int_{\mathbb{S}^1}\kappa^2\left(\kappa+\fracp{^2\kappa}{^2\vartheta}-h\right)^{2}d\vartheta
-2\fracd{h}{\tau}\int_{\mathbb{S}^1}\kappa\,d\vartheta \notag\\
&\geq-2\fracd{h}{\tau}\int_{\mathbb{S}^1}\kappa\,d\vartheta
\end{align}
for all $\tau>0$.
By~\eqref{eq:h_3a},
\begin{align}\label{eq:dtauhintkappa}
\fracd{h}{\tau}\int_{\mathbb{S}^1}\kappa\,d\vartheta
=-(1-\gamma)\frac{2\pi}{L^2}\fracd{L}{\tau}\int_{\mathbb{S}^1}\kappa\,d\vartheta
+\frac{\gamma}{4\pi}\fracd{}{\tau}\left(\int_{\mathbb{S}^1}\kappa\,d\vartheta\right)^2\,.
\end{align}
By Lemma~\ref{lem:AL}, $\fracd{}{\tau}L$ has a sign so that
$$\int_\varepsilon^\tau\left|\fracd{L}{\bar\tau}\right|\,d\bar\tau\leq|L_\tau-L_0|\leq C$$
for all $0<\varepsilon<\tau<\infty$.
We integrate~\eqref{eq:dtauhintkappa} from $\varepsilon$ to $\tau$ and conclude with $\varepsilon\to0$, the upper bound from Lemma~\ref{lem:inthkappa}, the definition~\eqref{eq:defmtau} of $m$ and~\eqref{eq:intkappakappamax},
$$\int_0^\tau\fracd{h}{\bar\tau}\int_{\mathbb{S}^1}\kappa\,d\vartheta\,d\bar\tau
\leq C\max_{\bar\tau\in[0,\tau]}\int_{\mathbb{S}^1}\kappa\,d\vartheta
+C\left(\int_{\mathbb{S}^1}\kappa(\vartheta,\tau)\,d\vartheta\right)^2
\leq C(m(\tau)+1)$$
for all $\tau\in(0,\infty)$.
Hence, integrating~\eqref{eq:dtauintkappadthetaap_1} and the bounds from Lemma~\ref{lem:inthkappa} yield the claim.
\end{proof}
For $\tau>0$, define
\begin{align}\label{eq:defmstartau}
m^*(\tau):=1+\frac{\sqrt{m(\tau)+1}}{\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)}\,.
\end{align}
\begin{Lemma}\label{lem:varekappamax}
Let $\Sigma_0=X_0(\mathbb{S}^1)$ be a smooth, embedded, convex curve.
Let $X:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}^2$ be a solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$ and $h$ satisfying~\eqref{eq:h_3a}.
Let $\tau>0$, $\vartheta_1,\vartheta_2\in\mathbb{S}^1$ and $\delta\in(0,\pi/2]$.
If $|\vartheta_1-\vartheta_2|<\delta$, then there exists $C\geq\sqrt{2\pi}$ with
$$|\kappa(\vartheta_1,\tau)-\kappa(\vartheta_2,\tau)|<Cm^*(\tau)\sqrt\delta\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)\,.$$
\end{Lemma}
\begin{proof}
We follow similar lines to~\cite[Paragraph~4.3.6]{GageHamilton86} and~\cite[Lem.~7.1]{Pihan98}.
Lemma~\ref{lem:AL} provides
$$\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)\geq\frac{L}{2\pi}\geq c>0\,.$$
Let $\delta\in(0,\pi/2]$.
For $|\vartheta_1-\vartheta_2|<\delta$, Cauchy--Schwarz and Lemma~\ref{lem:intdthetakappa} imply
\begin{align*}
|\kappa(\vartheta_1,\tau)-\kappa(\vartheta_2,\tau)|
&\leq|\vartheta_1-\vartheta_2|^{1/2}\left(\int_{\vartheta_1}^{\vartheta_2}
\left(\fracp{\kappa}{\vartheta}(\vartheta,\tau)\right)^{2}d\vartheta\right)^{1/2} \notag\\
&\leq\sqrt{\delta}\left(\int_{\mathbb{S}^1}\kappa^2(\vartheta,\tau)\,d\vartheta+C(m(\tau)+1)\right)^{1/2} \notag\\
&\leq\sqrt{\delta}\left(\sqrt{2\pi}\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)+\sqrt{C(m(\tau)+1)}\right) \notag\\
&\leq\sqrt{\delta}C\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)\left(1+\frac{\sqrt{m(\tau)+1}}{\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)}\right)\,,
\end{align*}
where we used $\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)>0$ for $\tau>0$.
\end{proof}
\begin{Lemma}\label{lem:kappamaxrin}
Let $\Sigma_0=X_0(\mathbb{S}^1)$ be a smooth, embedded, convex curve.
Let $X:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}^2$ be a smooth, embedded solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$ and $h$ satisfying~\eqref{eq:h_3a}.
Let $\varepsilon\in(0,1)$ and $\tau>0$.
Then
$$\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)r_{\inner}(\tau)
\leq\left\{(1-\varepsilon)\left[1-K\!\left(\left(\frac\varepsilon{Cm^*(\tau)}\right)^2\right)\left(\frac{r_{\cir}(\tau)}{r_{\inner}(\tau)}-1\right)\right]\right\}^{-1}\,,$$
where $K:(0,\pi]\to[0,\infty)$ is a positive decreasing function with $K(\omega)\to\infty$ for $\omega\searrow0$ and $K(\pi)=0$.
\end{Lemma}
\begin{proof}
The proof follows with the help of Lemma~\ref{lem:varekappamax} and can be found in~\cite[Cor.~5.2]{GageHamilton86} and~\cite[Lem.~7.11]{Pihan98}.
For details, see also~\cite[Lem.~11.11]{Dittberner18}.
\end{proof}
\begin{Cor}\label{cor:kappamaxrinvare}
Let $\Sigma_0=X_0(\mathbb{S}^1)$ be a smooth, embedded, convex curve.
Let $X:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}^2$ be a smooth, embedded solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$ and $h$ satisfying~\eqref{eq:h_3a}.
For every $\varepsilon\in(0,1)$, there exists a time $\tau_0>0$ such that, for all $\tau\geq\tau_0$,
$$\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)r_{\inner}(\tau)\leq\frac1{(1-\varepsilon)^2}\,.$$
\end{Cor}
\begin{proof}
We extend the proof of~\cite[Prop.~5.3]{GageHamilton86} and~\cite[Cor.~7.12]{Pihan98}.
Proposition~\ref{prop:C0convergence} implies that, for every $\delta>0$, there exists a time $\tau_0(\delta)>0$ so that $r_{\cir}(\tau)-r_{\inner}(\tau)\leq\delta$ for all $\tau\geq\tau_0$, and thus
\begin{align}\label{eq:routrindelta}
\frac{r_{\cir}(\tau)}{r_{\inner}(\tau)}-1\leq\frac\delta{r_{\inner}(\tau)}\,.
\end{align}
Recall the definitions~\eqref{eq:defmtau} and~\eqref{eq:defmstartau} of $m$ and $m^*$.
We define
$$I_1:=\{\tau\geq\tau_0\;|\;m(\tau)=\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)\}$$
and
$$I_2:=\{\tau\geq\tau_0\;|\;m(\tau)>\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)\}\,.$$
Then, $m$ is monotonically increasing on $I_1$ and constant on every connected subinterval of $I_2$.
By Lemma~\ref{lem:AL}, $L$ is uniformly bounded from above.
Hence there exits a constant $c>0$ so that $\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)\geq c$ for all $\tau\in[\tau_0,\infty)$ and
$$m^*(\tau)\leq2+\frac1c$$
for $\tau\in I_1$.
We distinguish between three cases.
\begin{enumerate}[(i)]
\item Assume that $\sup_{[\tau_0,\infty)}m<\infty$.
Then $\sup_{[\tau_0,\infty)}m^*<\infty$.
\item Assume that $\sup_{[\tau_0,\infty)}m=\infty$ and $\sup\{\tau\in I_2\}=:\tau_1<\infty$.
Then $[\tau_1,\infty)\subset I_1$ and
$$\sup_{[\tau_0,\infty)}m^*=\sup_{I_1}m^*<2+\frac1c\,.$$
\item Assume that $\sup_{[\tau_0,\infty)}m=\infty$ and $\sup\{\tau\in I_2\}=\infty$.
Assume there exists $\tau_2\in[\tau_0,\infty)$ so that $(\tau_2,\infty)\subset I_2$, then $m(\tau)=m(\tau_2)<\infty$ for all $\tau\in(\tau_2,\infty)$.
This contradicts $\sup_{[\tau_0,\infty)}m=\infty$.
Hence, $I_2$ consists of infinitely many disjoint open intervals $I_{2,k}$, $k\in\mathbb{N}$ and $\sup_{I_1}m^*\leq2+1/c$.
Define the sequence
$$\big(\tau_k:=\sup\{\tau\in I_{2,k}\}\in I_1\big)_{k\in\mathbb{N}}\,.$$
Then $\tau_k\to\infty$ for $k\to\infty$ and for all $k\in\mathbb{N}$, and since $\tau_k\in I_1$,
$$m(\tau)=m(\tau_k)=\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau_k)$$
as well as
$$m^*(\tau)\leq1+\frac{m(\tau)+1}{\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)}
=1+\frac{m(\tau_k)+1}{\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau_k)}
\leq2+\frac1c$$
for all $\tau\in I_{2,k}$.
Hence,
$$\sup_{\tau\in[\tau_0,\infty)}m(\tau)
=\sup_{\tau\in I_1\cup I_2}m(\tau)
\leq2+\frac1c\,.$$
\end{enumerate}
Thus, for any $\tau\geq\tau_0$, $m^*$ is independent of time.
Recall that $K$, as defined in Lemma~\ref{lem:kappamaxrin}, is a positive decreasing function that satisfies $K(\omega)\to\infty$ for $\omega\searrow0$ and $K(\pi)=0$.
By Proposition~\ref{prop:C0convergence}, $r_{\inner}(\tau)\geq c>0$ for all $\tau\geq0$.
Hence, for given $\varepsilon\in(0,1)$, we can choose $\delta>0$ and $\tau_0(\delta)>0$ so that
\begin{align}\label{eq:rinKvare}
\frac\delta{r_{\inner}(\tau)}\leq\frac\varepsilon{K\big((\varepsilon/Cm^*)^2\big)}
\end{align}
for all $\tau\geq\tau_0$.
Combining~\eqref{eq:routrindelta} and~\eqref{eq:rinKvare} yields
$$\frac{r_{\cir}(\tau)}{r_{\inner}(\tau)}-1\leq\frac\varepsilon{K\big((\varepsilon/Cm^*)^2\big)}$$
so that
$$1-\varepsilon\leq1-K\!\left(\left(\frac\varepsilon{Cm^*}\right)^2\right)\left(\frac{r_{\cir}(\tau)}{r_{\inner}(\tau)}-1\right)$$
for all $\tau\geq\tau_0$.
This and Lemma~\ref{lem:kappamaxrin} imply
$$\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)r_{\inner}(\tau)
\leq\left\{(1-\varepsilon)\left[1-K\!\left(\left(\frac\varepsilon{Cm^*}\right)^2\right)\left(\frac{r_{\cir}(\tau)}{r_{\inner}(\tau)}-1\right)\right]\right\}^{-1}
\leq\frac1{(1-\varepsilon)^2}$$
for any $\tau\geq\tau_0$.
\end{proof}
\begin{Cor}
\label{cor:C2congergences}
Let $\Sigma_0=X_0(\mathbb{S}^1)$ be a smooth, embedded, convex curve.
Let $X:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}^2$ be a smooth, embedded solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$ and $h$ satisfying~\eqref{eq:h_3a}.
Then
$$\frac{\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)}{\min_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)}\to1\,,\qquad
\kappa(\vartheta,\tau)\to\frac1{R}\qquad\text{ and }\qquad
h(\tau)\to\frac1{R}$$
for every $\vartheta\in\mathbb{S}^1$ and for $\tau\to\infty$, where $R$ is given in Proposition~\ref{prop:C0convergence}.
\end{Cor}
\begin{proof}
We follow the lines of~{\cite[Cor.~7.14]{Pihan98}}.
By Proposition~\ref{prop:C0convergence}, $\Sigma_\tau$ is strictly convex for $\tau\in(0,\infty)$.
Like in~\cite[Thm.~5.4]{GageHamilton86},~\cite[Prop.~7.13]{Pihan98} or~\cite[Prop.~11.13]{Dittberner18}, we first conclude with the help of Corollary~\ref{cor:kappamaxrinvare} that $\kappa(\vartheta,\tau)r_{\inner}(\tau)\to1$ for all $\vartheta\in\mathbb{S}^1$ and for $\tau\to\infty$.
Hence, it also holds that $\max_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)r_{\inner}(\tau)\to1$ and $\min_{\vartheta\in\mathbb{S}^1}\kappa(\vartheta,\tau)r_{\inner}(\tau)\to1$ for $\tau\to\infty$ and the first claim follows.
By Proposition~\ref{prop:C0convergence}, the curve converges to a circle of radius $R$.
This yields the second claim.
The third claim follows from Lemma~\ref{lem:inthkappa} and $L\to2\pi R$.
\end{proof}
\begin{Thm}
\label{thm:kappanconvergence}
Let $\Sigma_0=X_0(\mathbb{S}^1)$ be a smooth, embedded, convex curve.
Let $X:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}^2$ be a smooth, embedded solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$ and $h$ satisfying~\eqref{eq:h_3a}.
Then, for all $n\in\mathbb{N}$, $\fracp{^n}{\vartheta^n}\kappa\to0$ uniformly for $\tau\to\infty$.
Hence, the curves converge uniformly in $C^\infty$ to a circle of radius $R$.
\end{Thm}
\begin{proof}
The proof uses Corollary~\ref{cor:C2congergences} and can be found in~\cite[Prop.~7.17]{Pihan98} or~\cite[Thm.~11.17]{Dittberner18}.
\end{proof}
We summarise our results in the following and two theorems.
\begin{Thm}
\label{thm:main1}
Let $\Sigma_0=X_0(\mathbb{S}^1)$ be a smooth, embedded curve.
Let $X:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}^2$ be a smooth, embedded solution of~\eqref{eq:ccf} with initial curve $\Sigma_0$ and $h$ satisfying~\eqref{eq:h_3a}.
Then the evolving surfaces $\Si_t=X(\mathbb{S}^1\!,t)$ are contained in a uniformly bounded region of the plane for all times.
And, for all $\beta\in(0,1)$, there exists a time-independent constant $C>0$ such that, for all $t\geq0$,
\begin{enumerate}[(i)]
\item $|\max_{p\in\mathbb{S}^1}\kappa(p,t)-\min_{p\in\mathbb{S}^1}\kappa(p,t)|\leq C\exp\!\left(-\frac{\beta}{R^2}t\right)$,
\item $|\kappa(p,t)-1/R|\leq C\exp\!\left(-\frac{\beta}{R^2}t\right)$ for all $p\in\mathbb{S}^1$,
\item $|h(t)-1/R|\leq C\exp\!\left(-\frac{\beta}{R^2}t\right)$, and
\item $\left|\fracp{^n}{t^m}\fracp{^n}{p^n}\kappa(p,t)\right|\leq C\exp\!\left(-\frac{\beta}{(n+2m+1)R^2}t\right)$ for all $p\in\mathbb{S}^1$ and all $n,m\in\mathbb{N}$.
\end{enumerate}
Hence, the solution converges smoothly and exponentially to a circle of radius $R$.
\end{Thm}
\begin{proof}
By Theorem~\ref{thm:convexity}, there exists a time $T_0>0$ so that the curves are strictly convex on $(T_0,\infty)$.
Like in~\cite{GageHamilton86},~\cite[Section~7.5]{Pihan98} and~\cite[Section~11.4]{Dittberner18}, we can show for convex curves with the help of Wirtinger's inequality and the smooth convergence of Theorem~\ref{thm:kappanconvergence} exponential decay of the $L^2$-norm of the derivative of the curvature.
The proof is independent of the particular form of $h$, which is why we do not repeat it here.
Interpolation inequalities then yield that for $\beta\in(0,1)$ and $m,n\in\mathbb{N}\cup\{0\}$, $m+n>0$, there exist constants $C_{n,m}>0$ such that
\begin{align}\label{eq:dtaundthetamkappa}
\max_{\vartheta\in\mathbb{S}^1}\left|\fracp{^m}{\tau^m}\fracp{^n\kappa}{\vartheta^n}(\vartheta,\tau)\right|
\leq C_{n,m}\exp\!\left(-\frac{\beta\tau}{(n+2m+1)R^2}\right)
\end{align}
for $\tau$ large enough.
To prove~(i), we follow the lines of~\cite[Prop.~7.27]{Pihan98}.
For $t\geq0$, let $p_1,p_2\in\mathbb{S}^1$ be the points where the curvature attains its maximum and minimum.
By Lemma~\ref{lem:AL} and~\eqref{eq:dtaundthetamkappa}, there exists a time-independent constant $C>0$ so that
$$\left|\max_{p\in\mathbb{S}^1}\kappa(p,t)-\min_{p\in\mathbb{S}^1}\kappa(p,t)\right|
=|\kappa(p_2,t)-\kappa(p_1,t)|
\leq\int_{\Si_t}\left|\fracp{\kappa}{s}\right|ds_t
\leq C\exp\!\left(-\frac{\beta t}{R^2}\right)\,.$$
for $t$ large enough.
To show claim (ii), we observe that for embedded, closed, convex curves,
\begin{align}\label{eq:kappaminALkappamax}
\min_{p\in\mathbb{S}^1}\kappa(p)
\leq\frac1{r_{\cir}}
=\sqrt{\frac\pi{A_{\cir}}}
\leq\sqrt{\frac\pi{A}}
\leq\frac{2\pi}{L}
\leq\frac1{2\pi}\int_{\Si_t}\kappa^2\,ds_t
\leq\max_{p\in\mathbb{S}^1}\kappa(p)\,.
\end{align}
By the intermediate value theorem and~\eqref{eq:kappaminALkappamax} there exist points $p_0,p_1,p_2\in\mathbb{S}^1$ with $\kappa(p_0,t)=\sqrt{\pi/A}$, $\kappa(p_1,t)=2\pi/L$ and $\kappa(p_2,t)=\int_{\Si_t}\kappa^2\,ds_t/2\pi$, so that for $p\in\mathbb{S}^1$ with~\eqref{eq:dtaundthetamkappa},
\begin{align}\label{eq:kappapiA}
\left|\kappa(p,t)-\sqrt{\frac\pi{A}}\right|
=|\kappa(p,t)-\kappa(p_0,t)|
\leq\int_{\Si_t}\left|\fracp{\kappa}{s}\right|ds_t
\leq C\exp\!\left(-\frac{\beta t}{R^2}\right)
\end{align}
and likewise
\begin{align}\label{eq:kappa2piL}
\left|\kappa(p,t)-\frac{2\pi}{L}\right|
+\left|\kappa(p,t)-\frac1{2\pi}\int_{\Si_t}\kappa^2\,ds_t\right|
\leq C\exp\!\left(-\frac{\beta t}{R^2}\right)
\end{align}
for $t$ large enough.
Furthermore, Proposition~\ref{prop:C0convergence} and yields
\begin{align}\label{eq:AR}
\left|\sqrt{\frac\pi{A}}-\frac1R\right|
\leq\frac{\sqrt{|A-\pi R^2|}}{R\sqrt{A}}
\leq C\sqrt{\int_t^\infty\left|\fracd{A}{\tau}\right|\,d\tau}
\leq C\int_t^\infty\left|\fracd{A}{\tau}\right|\,d\tau
\end{align}
and
\begin{align}\label{eq:LR}
\left|\frac{2\pi}{L}-\frac1R\right|
=\frac{|2\pi R-L|}{LR}
\leq C\int_t^\infty\left|\fracd{L}{\tau}\right|\,d\tau\,.
\end{align}
By~\eqref{eq:L2piintkappa},~\eqref{eq:dtAa} and Proposition~\ref{prop:C0convergence}, there exists a constant $C>0$ so that
\begin{align}\label{eq:piAR_a}
\sqrt{\int_t^\infty\left|\fracd{A}{\tau}\right|\,d\tau}
&=\sqrt{\gamma\left(\frac{L^2}{4\pi}-A\right)}
\leq C\exp\!\left(-\frac{\beta t}{R^2}\right)
\end{align}
for $t$ large enough.
By~\eqref{eq:kappapiA},~\eqref{eq:kappa2piL},~\eqref{eq:AR},~\eqref{eq:LR} and~\eqref{eq:piAR_a}, for $p\in\mathbb{S}^1$,
\begin{align}\label{eq:kappa1Rleq}
\left|\kappa(p,t)-\frac1{R}\right|
\leq\left|\kappa(p,t)-\sqrt{\frac\pi{A}}\right|+\left|\sqrt{\frac\pi{A}}-\frac1{R}\right|
\leq C\exp\!\left(-\frac{\beta t}{R^2}\right)\,.
\end{align}
Likewise with Proposition~\ref{prop:C0convergence},~\eqref{eq:kappapiA} and~\eqref{eq:kappa2piL},
\begin{align}\label{eq:kappahleq}
|\kappa(p,t)-h(t)|
&\leq\left|\kappa(p,t)-\frac{2\pi}L\right|+\frac{|\gamma|}{L}\left|\frac L{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi\right| \notag\\
&\leq C\exp\!\left(-\frac{\beta t}{R^2}\right)
\end{align}
for $t$ large enough.
The boundedness of the curvature on $[0,T_0]$ yields the claim for all $t\geq0$.
For claim (iii), we estimate with~\eqref{eq:kappa1Rleq} and~\eqref{eq:kappahleq},
$$\left|h-\frac1{R}\right|
\leq C\exp\!\left(-\frac{\beta t}{R^2}\right)$$
for all $t\geq0$.
For claim~(iv), we use Lemma~\ref{lem:evolutionequations} and~\eqref{eq:kappahleq} to estimate
$$\fracp{v}{t}
=\kappa(h-\kappa)v
\leq C\left(\exp\!\left(-\frac{\beta t}{R^2}\right)\right)v$$
for all $t\geq0$.
Hence, $v\geq C$ on $\mathbb{S}^1\!\times[0,\infty)$
and with~\eqref{eq:dtaundthetamkappa}, for every $m,n\in\mathbb{N}\cup\{0\}$, $m+n>0$,
\begin{align*}
\left|\fracp{^m}{t^m}\fracp{^n\kappa}{p^n}\right|
=\left|\fracp{^m}{t^m}\!\left(v\fracp{}{s}\right)^{\!n}\!\kappa\,\right|
\leq C\exp\!\left(-\frac{\beta t}{(n+2m+1)R^2}\right)\,.
\end{align*}
To show that the curves stay in a bounded region, we observe that with~\eqref{eq:kappahleq},
$$\Vert X(p,t)-X(p,0)\Vert_{\mathbb{R}^2}
\leq\int_0^t|\kappa(p,\tau)-h(\tau)|\,d\tau
\leq C\int_0^t\exp\!\left(-\frac{\beta\tau}{R^2_0}\right)\,d\tau\leq C$$
for all $p\in\mathbb{S}^1$ and $t\in(0,\infty)$, where $C$ is independent of time.
\end{proof}
\begin{Rem}\label{rem:dtAg}
All the proofs leading up to Theorem~\ref{thm:main1} also work, if we prescribe the derivative of the area or the length by a function $g\in C^\infty([0,\infty))\cap L^1([0,\infty))$.
If we prescribe the derivative of the area, Lemma~\ref{lem:dtL} and~\eqref{eq:hdtAdtL} yield
\begin{align}\label{eq:dtAdtLb}
\fracd{A}{t}=g\,,\quad
h=\frac{2\pi+g}L
\quad\text{ and }\quad
\fracd{L}{t}
=-\frac{2\pi}L\left(\frac L{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi-g\right)\,,
\end{align}
where either
\begin{align*}
-2\pi<g\leq0\,,&\quad\fracd{g}{t}\geq0\quad\text{ and }\quad\int_0^\infty g\,dt>-A_0\,,\qquad\text{ or } \\
0\leq g<\frac{L_t}{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi\,,&\quad\fracd{g}{t}\leq0\quad\text{ and }\quad\int_0^\infty g\,dt\leq\frac{L_0^2}{4\pi}-A_0\,,
\end{align*}
since need $A$ and $L$ to be monotone and bounded and we will need $h$ to be positive in Remark~\ref{rem:main2}.
If we prescribe the derivative of the length, Lemma~\ref{lem:dtL} and~\eqref{eq:hdtAdtL} yield
\begin{align*}
\fracd{L}{t}=g\,,\quad
h=\frac1{2\pi}\left(\int_{\Si_t}\kappa^2\,ds_t+g\right)
\quad\text{ and }\quad
\fracd{A}{t}
=\frac L{2\pi}\int_{\Si_t}\kappa^2\,ds_t-2\pi+\frac{Lg}{2\pi}\,,
\end{align*}
where either
\begin{align*}
-\int_{\Si_t}\kappa^2\,ds_t+\frac{4\pi}{L_t}<g\leq0\,,&\quad\fracd{g}{t}\geq0\quad\text{ and }\;\int_0^\infty g\,dt>-L_0\,,\quad\text{ or } \\
0\leq g\,,&\quad\fracd{g}{t}\leq0\quad\text{ and }\quad\int_0^\infty g\,dt<\infty\,,
\end{align*}
since again need $A$ and $L$ to be monotone and bounded.
Then Theorem~\ref{thm:main1} holds with the addition in the cases
\begin{enumerate}
\item[(ii)] $|\kappa(p,t)-1/R|\leq C\exp\!\left(-\frac{\beta}{R^2}t\right)+C\int_t^\infty g\,d\tau$ for all $p\in\mathbb{S}^1$, and
\item[(iii)] $|h(t)-1/R|\leq C\exp\!\left(-\frac{\beta}{R^2}t\right)+C\int_t^\infty g\,d\tau+Cg(t)$
\end{enumerate}
for all $\beta\in(0,1)$ and $t\geq0$, where $C>0$ is time-independent.
\end{Rem}
\begin{Thm}\label{thm:main2}
Let $\Sigma_0=X_0(\mathbb{S}^1)$ be a smooth, embedded curve satisfying~\eqref{eq:intkappageqminuspi}.
Then there exists a unique, smooth, embedded solution $X:\mathbb{S}^1\!\times[0,\infty)\to\mathbb{R}^2$ to~\eqref{eq:ccf} with initial curve $\Sigma_0$ and $h$ satisfying~\eqref{eq:h_ap}.
The evolving curves $\Si_t=X(\mathbb{S}^1\!,t)$ are contained in a uniformly bounded region and converge smoothly and exponentially to a circle of radius $R$.
\end{Thm}
\begin{proof}
By the short time existence, Theorem~\ref{thm:ste}, there exists a unique solution $X\in C^\infty(\mathbb{S}^1\!\times[0,T))$
By Lemma~\ref{lem:AL}, $c\leq L\leq C$ so that $h$ is uniformly bounded from above and below away from zero.
By Corollary~\ref{cor:embeddedness} the curves remain embedded on $(0,T)$.
Corollary~\ref{cor:T=infty} yields that $T=\infty$.
Hence, we can apply Theorem~\ref{thm:main1}.
\end{proof}
\begin{Rem}\label{rem:main2}
Theorem~\ref{thm:main2} also holds for $h$ satisfying~\eqref{eq:dtAdtLb}.
\end{Rem}
|
\section{Introduction}\label{sec:intro}
This paper deals with inverse problems: one aims at recovering an unknown signal from its corrupted measurements.
To be more specific, the motivation of this work is the reconstruction of an unknown \emph{continuous-domain} and \emph{periodic} signal $f$ from its $M$ noisy measurements $y_m \approx \langle \nu_m, f \rangle = \int_0^1 \nu_m(t)f(t) \mathrm{d} t $ for $m=1\ldots M$,
where the $\nu_m$ are measurement functions.
The goal is then to build an output signal ${f}_{\mathrm{opt}}$ that is as close as possible to $f$.
\subsection{Inverse Problems in the Continuous Domain}
Inverse problems are often formulated in the discrete domain~\cite{Banham1997digital,Karayiannis1990regularization,Figueiredo2003algorithm,Afonso2011augmented,Bertero1998introduction}.
This is motivated by the need of manipulating digital data on computers.
Nevertheless, many naturally occurring signals depend on continuous variables (\emph{e.g.}, time or position). This leads us to attempt recovering a signal $f_{\mathrm{opt}}(t)$ that depends on the continuous variable $t \in [0,1]$. In contrast with the classical discrete setting, our search space for this reconstructed signal is thus infinite-dimensional~\cite{Adcock2015generalized}. Moreover, we choose a regularization based on true derivatives (as opposed to finite differences) to impose some smoothness on the reconstructed signal, a concept that is absent in the discrete setting.
When considering continuous-domain reconstruction methods, a majority of works, typically in machine learning, deal with sampling measurements. The goal is then to recover $f$ from its (possibly noisy) values $y_m\approx f(t_m)$ at fixed location $t_m$.
In order to investigate a more general version of inverse problems, we shall consider generalized measurements~\cite{Papoulis1977generalized,Eldar2006minimum}. They largely exceed the sampling case and include Fourier sampling or convolution (\emph{e.g.}, MRI, x-ray tomography~\cite{Piccolomini2002regularization, Bostan2013Sparse}).
Our only requirement is that the measurements $y_m$ depend linearly on, and evolve continuously with, the unknown signal $f$ up to some additive noise, so that $y_m \approx \langle \nu_m, f \rangle$.
\subsection{Variational vs. Statistical Methods}
In the discrete domain, two standard strategies are used to reconstruct
an input signal $\bf{x}$ from its noisy measurements $\bf{y} \approx \mathbf{H} x$, where $\mathbf{H}$ models the acquisition process~\cite{Bertero1998introduction}.
The first approach is deterministic and can be tracked back to the '60s with Tikhonov's seminal work~\cite{Tikhonov1963solution}.
The ill-posedness of the problem usually imposes the addition of a regularizer.
By contrast, Wiener filtering is based on the stochastic modelization of the signals of interest and the optimal estimation of the targeted signal $\bf{x}$.
This paper generalizes these ideas for the reconstruction of \emph{continuous} signals from their \emph{discrete} measurements.
In the variational setting, the reconstructed signal is a solution to an optimization problem that imposes some smoothness conditions~\cite{cassel2013variational}. More precisely, the optimization problem may take the form
\begin{equation}\label{eq: intro}
{f}_{\mathrm{opt}} = \argmin_f \bigg( \sum_{m=1}^M \big(y_m -\langle \nu_m, f \rangle \big)^2 + \lambda \| \mathrm{L}f\| ^2_{L_2} \bigg),
\end{equation}
where $\mathrm{L}$ is a linear operator.
The first term in~\eqref{eq: intro} controls the data fidelity. The regularization term $\| \mathrm{L}f\| ^2_{L_2} $ constrains the function to satisfy certain smoothness properties (for this reason, the variational approach is sometimes called a smoothing approach).
The parameter $\lambda$ in~\eqref{eq: intro} quantifies the tradeoff between the fidelity to the data and the regularization constraint.
In the statistical setting, the signal is modeled as a random process and is optimally reconstructed using estimation theory~\cite{Moon2000mathematical}.
More precisely, one assumes that the continuous-domain signal is the realization of a stochastic process $s$ and that the samples are given by $y_m= \langle \nu_m , s \rangle + \epsilon_m$, where $\epsilon_m$ is a random perturbation and $\nu_m$ a linear measurement function.
In this case, one specifies the reconstructed signal as the optimal statistical estimator in the mean-square sense
\begin{equation}
{f}_{\mathrm{opt}} = \argmin_{\tilde{s}} \mathbb{E} \left[ \lVert s - \tilde{s}(\cdot | \mathbf{y}) \rVert_{L_2}^2 \right],
\end{equation}
where the estimators $t \mapsto \tilde{s}(t | \mathbf{y})$ are computed from the generalized samples $y_m$. The solution depends on the measurement function $\nu_m$ and the stochastic models specified for $s$ and $\epsilon_m$. In our case, the random process $s$ is characterized by a linear operator $\mathrm{L}$ that is assumed to have a whitening effect (it transforms $s$ into a periodic Gaussian white noise), while the perturbation is i.i.d. Gaussian.
\subsection{Periodic and General Setting}
The variational and statistical approaches have been extensively studied for continuous-domain signals defined on the infinitely supported real line.
However, it is often assumed in practice that the input signals are periodic.
In fact, a standard computational approach to signal processing is to extend by periodization the signals of otherwise bounded support.
Periodic signals arise also naturally in applications such as the parametric representation of closed curves~\cite{Cohen1994part, Delgado2012ellipse,Badoual2017subdivision}.
This has motivated the development of signal-processing tools and techniques specialized to periodic signals in sampling theory, error analysis, wavelets, stochastic modelization, or curve representation~\cite{Vetterli2002FRI,Maravic2005sampling,blu2008sparse,Jacob2002sampling,Triebel2008function, Fageot2017besov,Badoual2016inner}.
In this paper, we develop the theory of the variational and statistical approaches for periodic continuous-domain signals in a very general context, including the following aspects:
\begin{itemize}[leftmargin=10 pt]
\item We consider a broad class of measurement functions, with the only assumptions that they are linear and continuous.
\item Both methods refer to an underlying linear operator $\mathrm{L}$ that affects the smoothness properties of the reconstruction. We deal with a very broad class of linear operators acting on periodic functions.
\item We consider possibly non-quadratic data fidelity terms in the smoothing approach.
\end{itemize}
\subsection{Related Works}
The topics investigated in this paper have already received some attention in the literature, mostly in the non-periodic setting.
\paragraph{Reconstruction over the Real Line}
Optimization problems of the form~\eqref{eq: intro} appear in many fields and receive different names, including inverse problems in image processing~\cite{Bertero1998introduction}, representer theorems in machine learning~\cite{Scholkopf2001generalized}, or sometimes interpolation elsewhere.
Schoenberg was the first to show the connection between~\eqref{eq: intro} and spline theory~\cite{Schoenberg1964spline}.
Since then, this has been extended to other operators~\cite{Unser2007self}, or to the interpolation of the derivative of the signal~\cite{Uhlmann2016hermite,Condat2011quantitative}.
Many recent methods are dealing with non-quadratic regularization, especially the ones interested in the reconstruction of sparse discrete~\cite{Candes2006sparse,Donoho2006} or continuous signals~\cite{Adcock2015generalized,Denoyelle2015support,Unser2016splines,gupta2018continuous}.
We discuss this aspect more extensively in Section~\ref{sec:TV}.
A statistical framework requires the specification of the noise and of the signal stochastic model. The signal is then estimated from its measurements. A classical measure of the quality of an estimator is the mean-square error. This criterion is minimized by the minimum mean-square error (MMSE) estimator~\cite{Moon2000mathematical,Tarantola2005inverse}.
The theory has been developed mostly for Gaussian processes and in the context of sampling measurements~\cite{Berlinet2011reproducing}. We are especially interested in innovation models, for which one assumes that the signal can be whitened (\emph{i.e.}, transformed into a white noise) by the application of a linear operator~\cite{Kailath1968innovationsA,Unser2014unifiedContinuous}. Non-periodic models have been studied in many situations, including the random processes associated with differential~\cite{Kimeldorf1970spline, Uhlmann2015SampTA} or fractional operators~\cite{Blu2007self}. Extensions to non-Gaussian models are extensively studied by Unser and Tafti~\cite{Unser2014sparse}.
The statistical and variational frameworks are deeply connected.
It is remarkable that the solution of either problem can be expressed as spline functions in relation with the linear operator $\mathrm{L}$ involved in regularization (variational approach) or whitening (statistical approach).
Wahba has shown that the two approaches are strictly equivalent in the case of stationary Gaussian models~\cite{Wahba1990spline}. This equivalence has also been recognized by several authors since then, as shown by Berlinet and Thomas-Agnan~\cite{Berlinet2011reproducing}, and Unser and Blu~\cite{Unser2005generalized}.
In the non-stationary case, this equivalence is not valid any more and the existence of connections has received less attention.
\paragraph{Reconstruction of Periodic Signals}
Some strong practical concerns have motivated the need for an adaptation of the theory to the periodic setting.
Important contributions in that direction have been proposed.
Periodic splines are constructed and applied to sampling problems by Schoenberg~\cite{Schoenberg1964trigonometric} and Golomb~\cite{Golomb1968approximation}. The smoothing spline approach is studied in the periodic setting by Wahba~\cite{Wahba1990spline} for derivative operators of any order. Although the periodic extension of the classical theory is briefly mentioned by several authors~\cite{Berlinet2011reproducing,Wahba1990spline,de1978practical}, we are not aware of a global treatment. Providing a general analysis in the periodic setting is precisely what we propose in this paper.
\subsection{Outline and Main Contributions}
Section~\ref{sec:mathematical_background} contains the main notations and tools for periodic functions and operators.
In Section~\ref{sec:representer_th}, we state the periodic representer theorem (Theorem~\ref{th: periodic representer thm}). It fully specifies the form of the solution in the variational approach in a very general setting. For the specific case of sampling measurements, we show that this solution is a periodic spline (Proposition~\ref{prop: Spline}).
Section~\ref{sec:processes_MMSE} is dedicated to the statistical approach. We introduce a class of periodic stationary processes (the \emph{Gaussian bridges}) for which we specify the MMSE estimator in the case of generalized linear measurements (Theorem~\ref{th: MMSE}). We also provide a theoretical comparison between the variational and statistical approaches by reformulating the MMSE estimation as the solution of a new optimization problem (Proposition~\ref{prop: opt pb}). This highlights the strict equivalence of the two approaches for invertible operators and extends known results from sampling to generalized linear measurements.
For non-invertible operators, we complete our analysis with simulations in Section~\ref{sec:simulations}. In particular, we give empirical evidence of the practical relevance of the variational approach for the reconstruction of periodic stationary signals.
We provide in Section~\ref{sec:periodicvsline} a comparison between our results in the periodic setting and the known results over the real line.
Finally, we conclude in Section~\ref{sec:conslusions}.
All the proofs have been postponed to the Appendix sections.
\section{Mathematical Background for Periodic Signals}\label{sec:mathematical_background}
Throughout the paper, we consider periodic functions and random processes. Without loss of generality, the period can always be normalized to one. Moreover, we identify a periodic function over $\mathbb{R}$ with its restriction to a single period, chosen to be $\mathbb{T} = [0,1)$.
We use the symbols $f$, $s$, and $\tilde{s}$ to specify a function, a random process, and an estimator of $s$, respectively.
We call $\mathcal{S}(\mathbb{T})$ the space of $1$-periodic functions that are infinitely differentiable, $\mathcal{S}^{\prime}(\mathbb{T})$ the space of $1$-periodic generalized functions (dual of $\mathcal{S}(\mathbb{T})$), and $L_2(\mathbb{T})$ the Hilbert space of square integrable $1$-periodic functions associated with the norm ${\lVert f \rVert_{L_2} = (\int_0^1 \lvert f(t) \rvert^2 \mathrm{d} t )^{1/2}}$. Working with $\mathcal{S}'(\mathbb{T})$ allows us to deal with functions with no pointwise interpretation, such as the Dirac comb defined by
\begin{equation}
\Sha=\sum\limits_{k \in \mathbb{Z}}\delta(\cdot -k),
\end{equation}
where $\delta$ is the Dirac impulse.
The duality product between an element $f \in \mathcal{S}'(\mathbb{T})$ and a smooth function $g \in \mathcal{S}(\mathbb{T})$ is denoted by $\langle f, g \rangle$. For instance, $\langle \Sha , g \rangle = g(0)$ for every $g$.
When the two real functions are in $L_2(\mathbb{T})$, we simply have the usual scalar product $\langle f , g \rangle = \int_{0}^1 f(t) {g(t)} \mathrm{d} t$.
All these concepts are extended to complex-valued functions in the usual manner with the convention that $\langle f , g \rangle = \int_{0}^1 f(t) \overline{g(t)} \mathrm{d}t$ for square-integrable functions.
The complex sinusoids are denoted by $e_k(t)=\mathrm{e}^{\mathrm{j} 2\pi k t}$ for any $k \in \mathbb{Z}$ and $t \in \mathbb{T}$.
Any periodic generalized function $f \in \mathcal{S}'(\mathbb{T})$ can be expanded as
\begin{equation}\label{eq: Fourier serie}
f(t)=\sum_{k\in \mathbb{Z}}\widehat{f}[k] \mathrm{e}^{\mathrm{j} 2\pi k t}=\sum_{k\in \mathbb{Z}}\widehat{f}[k] e_k(t),
\end{equation}
where the $\widehat{f}[k]$ are the Fourier coefficients of $f$, given by
$\widehat{f}[k] = \langle f , e_k \rangle$. Finally, the convolution between two periodic functions $f$ and $g$ is given by
\begin{equation}\label{eq: periodic convolution}
(f \ast g)(t)=\langle f, g(t-\cdot)\rangle.
\end{equation}
If $f, g \in L_2(\mathbb{T})$, we have that $(f \ast g)(t)=\int_0^1f(\tau)g(t-\tau)\mathrm{d}\tau$.
\subsection{Linear and Shift-Invariant Operators} \label{subsec:LSIop}
Let $\mathrm{L}$ be a linear, shift-invariant (LSI), and continuous operator from $\mathcal{S}(\mathbb{T})$ to $\mathcal{S}'(\mathbb{T})$.
The shift invariance implies the existence of $\widehat{L}[k] \in \mathbb{C}$ such that
\begin{equation}\label{eq: LSI}
\mathrm{L}e_k=\widehat{L}[k]e_k,
\end{equation}
for any $k \in \mathbb{Z}$. We call $\widehat{L}[k]$ the frequency response of the operator $\mathrm{L}$; it is also given by
\begin{equation}
\widehat{L}[k]= \langle \mathrm{L} \{\Sha\}, e_k \rangle = \int_0^1 \mathrm{L}\{\Sha\}(t)\mathrm{e}^{-\mathrm{j} 2\pi k t} \mathrm{d}t.
\end{equation}
The sequence $(\widehat{L}[k])$ is the Fourier series of the periodic generalized function $\mathrm{L}\{\Sha\}$, and is therefore of slow growth~\cite[Chapter VII]{Schwartz1966distributions}.
This implies that $\mathrm{L}$, a priori from $\mathcal{S}(\mathbb{T})$ to $\mathcal{S}'(\mathbb{T})$, actually continuously maps $\mathcal{S}(\mathbb{T})$ into itself. This is a significant difference with the non-periodic setting --- we discuss this point in the conclusion in Section~\ref{sec:conslusions}. Therefore, one can extend it by duality from $\mathcal{S}^{\prime}(\mathbb{T})$ to $\mathcal{S}^{\prime}(\mathbb{T})$. Then, for every $f \in \mathcal{S}^{\prime}(\mathbb{T})$, we easily obtain from~\eqref{eq: LSI} that
\begin{equation}
\mathrm{L}f(t)=\sum_{k\in \mathbb{Z}}\widehat{(\mathrm{L}f)}[k]e_k(t), \ \mbox{where } \widehat{(\mathrm{L}f)}[k]=\widehat{f}[k]\widehat{L}[k].
\end{equation}
The null space of $\mathrm{L}$ is $\mathcal{N}_{\mathrm{L}} = \{ f \in \mathcal{S}'(\mathbb{T}) \ | \ \mathrm{L} f = 0\}$.
We shall only consider operators whose null space is finite-dimensional, in which case $\mathcal{N}_{\mathrm{L}}$ can only be made of linear combinations of sinusoids at frequencies that are annihilated by $\mathrm{L}$.
We state this fact in Proposition~\ref{prop: null space} and prove it in Appendix~\ref{app: null space}.
\begin{prop}\label{prop: null space}
Let $\mathrm{L}$ be a continuous LSI operator. If $\mathrm{L}$ has a finite-dimensional null space $\mathcal{N}_{\mathrm{L}}$ of dimension $N_0$, then the null space is of the form
\begin{equation}\label{eq: null space}
\mathcal{N}_{\mathrm{L}}=\mathrm{span}\{e_{k_n}\}_{n=1}^{N_0},
\end{equation}
where the $k_n \in \mathbb{Z}$ are distinct.
\end{prop}
\smallskip
\noindent From~\eqref{eq: LSI} and~\eqref{eq: null space}, we deduce that $\widehat{L}[k]=0$ if and only if $k=k_n$ for some $n\in [1 \ldots N_0]$.
In the following, we consider real-valued operators. In that case, we have the Hermitian symmetry
$ \overline{\widehat{L}[-k]}={\widehat{L}}[k]$. Moreover, $e_{k_n} \in \mathcal{N}_{\mathrm{L}}$ if and only if $e_{-k_n} \in \mathcal{N}_{\mathrm{L}}$.
The orthogonal projection of $f$ on the null space $\mathcal{N}_{\mathrm{L}}$ is given by
\begin{equation}\label{eq: Proj null space}
\mathrm{Proj}_{\mathcal{N}_{\mathrm{L}}} \{f\}=\sum_{n=1}^{N_0}\widehat{f}[k_n]e_{k_n}.
\end{equation}
Let ${\mathcal{K}_{\mathrm{L}}=\mathbb{Z}\backslash \{k_n\}_{ n \in \{1 \ldots N_0\}}}$. Then,~\eqref{eq: Fourier serie} can be re-expressed as
${f=\mathrm{Proj}_{\mathcal{N}_{\mathrm{L}}} \{f\}+ \sum_{k \in \mathcal{K}_{\mathrm{L}}} \widehat{f}[k]e_k}$ and we have that
$ \mathrm{L} f (t)= \sum_{k \in \mathcal{K}_{\mathrm{L}}} \widehat{f}[k] \widehat{L}[k] e_k(t),
$ which yields the Parseval relation
\begin{equation}\label{eq: Parseval}
\int_0^1| \mathrm{L}f (t)|^2 \mathrm{d}t=\sum_{k \in \mathcal{K}_{\mathrm{L}}} \big | \widehat{f}[k] \big |^2 \big |\widehat{L}[k] \big |^2 .
\end{equation}
\subsection{Periodic $\mathrm{L}$-Splines} \label{sec : periodic splines}
Historically, splines are functions defined to be piecewise polynomials~\cite{Schoenberg1973cardinal}. A spline is hence naturally associated to the derivative operator of a given order~\cite{Unser1999splines} in the sense that, for a fixed $N\geq 1$, a spline function $f : \mathbb{R}\rightarrow \mathbb{R}$ satisfies ${\mathrm{L} f (t) = \sum a_m \delta(t-t_m)}$ with $\mathrm{L} = \mathrm{D}^N$ the $N$th derivative.
Splines have been extended to differential~\cite{schumaker2007spline,Schultz1967Lsplines,Unser2005cardinal,Unser2005think}, fractional~\cite{Unser2007self,Panda2006fractional} or, more generally, spline-admissible operators~\cite{Unser2014sparse}.
We adapt here this notion to the periodic setting, where the Dirac impulse $\delta$ is replaced by the Dirac comb $\Sha$.
\begin{definition} \label{def:Lspline}
Consider an LSI operator $\mathrm{L}$ with finite-dimensional null space.
We say that a function $f$ is a \emph{periodic $\mathrm{L}$-spline} if
\begin{equation}
\mathrm{L} f (t) = \sum_{m=1}^M a_m \Sha( t - t_m)
\end{equation}
for some integer $M \geq 1$, weights $a_m \in \mathbb{R}$, and knot locations $t_m \in \mathbb{T}$.
\end{definition}
Periodic $\mathrm{L}$-splines play a crucial role in the variational and statistical approaches for the resolution of inverse problems in the periodic setting. We represent some periodic splines associated to different operators in Figure~\ref{fig: spline}.
%
\begin{figure}
\centering
\hspace*{-0.2cm}
\subfigure[$\mathrm{L}=\mathrm{D}+\mathrm{I}$ and $M=4$]{\includegraphics[scale=0.32]{spline.pdf}}
\hspace*{+0.2cm}
\subfigure[$\mathrm{L}=\mathrm{D}^2$ and $M=5$]{\includegraphics[scale=0.32]{spline2.pdf}}
\caption{Illustrations of periodic $\mathrm{L}$-splines. Dots: nodes $\big( t_m, f(t_m) \big)$. The spline in (a) corresponds to the periodization of an exponential B-spline (see Figure 1 in~\cite{Unser2005cardinal}). }
\label{fig: spline}
\end{figure}
\section{Periodic Representer Theorem}\label{sec:representer_th}
We now consider a continuous LSI operator $\mathrm{L}$ with finite-dimensional null space $\mathcal{N}_{\mathrm{L}}$. Let $\boldsymbol{\nu}$ be the vector of the linear measurement functions $\nu_1, \ldots, \nu_M$.
They usually are of the form $\nu_m = \delta(\cdot - t_m)$ for time-domain sampling problems.
Here, we consider general linear measurements to include any kind of inverse problems.
In this section, our goal is to recover a function $f$ from observed data $\mathbf{y}=(y_1, \ldots ,y_M)$ such that $y_m \simeq \langle \nu_m, f\rangle$.
To do so, we consider the variational problem
\begin{equation}\label{eq: opt pb}
\underset{f}{\min}\bigg( F(\mathbf{y},\boldsymbol\nu (f))+\lambda \| \mathrm{L}f\| _{L_2}^2\bigg),
\end{equation}
where $F: \mathbb{R}^M \times \mathbb{R}^M \rightarrow \mathbb{R}^+$ is a strictly convex and continuous function called the \textit{cost function}. This function controls the fidelity to data. A special attention will be given to the quadratic data fidelity of the form
\begin{equation}\label{eq: quadratic F}
F(\mathbf{y},\boldsymbol\nu (f))=\sum\limits_{m=1}^M (y_m-\langle \nu_m, f\rangle )^2.
\end{equation}
We give the solution of~\eqref{eq: opt pb} for the space of $1$-periodic functions in Theorem~\ref{th: periodic representer thm}. To derive this solution, we first introduce and characterize the space of functions on which~\eqref{eq: opt pb} is well-defined.
\subsection{Search Space} \label{sec:searchspace}
The optimization problem~\eqref{eq: opt pb} deals with functions such that $\mathrm{L} f$ is square-integrable, which leads us to introduce ${\mathcal{H}_{\mathrm{L}}=\{f \in \mathcal{S}^{\prime} (\mathbb{T}) \ | \ \mathrm{L}f \in L_2(\mathbb{T})\}}$. Due to~\eqref{eq: Parseval}, we have that
\begin{equation}\label{eq: search space}
\mathcal{H}_{\mathrm{L}}=\{f \in \mathcal{S}^{\prime} (\mathbb{T}) \ | \ \sum_{k\in \mathcal{K}_{\mathrm{L}}}|\widehat{f}[k]|^2|\widehat{L}[k]|^2 < + \infty\}.
\end{equation}
Similar constructions have been developed for functions over $\mathbb{R}$ or for sequences by Unser \textit{et al.}~\cite{Unser2016splines,Unser2016representer}.
We now identify a natural Hilbertian structure on $\mathcal{H}_{\mathrm{L}}$. If $\mathrm{L}: \mathcal{H}_{\mathrm{L}} \rightarrow L_2(\mathbb{T})$ is invertible, then $\mathcal{H}_{\mathrm{L}}$ inherits the Hilbert-space structure of $L_2$ via the norm $\|\mathrm{L}f\|_{L_2}$. However, when $\mathrm{L}$ has a nontrivial null space, $\|\mathrm{L}f \|_{L_2}$ is only a semi-norm, in which case there exists $f \neq 0$ (any element of the null space of $\mathrm{L}$) such that $\| \mathrm{L}f \|_{L_2}=0$. To obtain a \textit{bona fide} norm, we complete the semi-norm with a special treatment for the null-space components in Proposition~\ref{prop: inner product}.
\begin{prop}\label{prop: inner product}
Let $\mathrm{L}$ be a continuous LSI operator whose finite-dimensional null space is defined by ${\mathcal{N}_{\mathrm{L}}=\mathrm{span}\{e_{k_n}\}_{n=1}^{N_0}}$.
We fix $\gamma^2 > 0$. Then, $\mathcal{H}_{\mathrm{L}}$ is a Hilbert space for the inner product
\begin{equation}\label{eq: inner product}
\langle f,g\rangle _{\mathcal{H}_{\mathrm{L}}}=\langle \mathrm{L} f, \mathrm{L} g\rangle +\gamma^2 \sum_{n=1}^{N_0}\widehat{f}[k_n]\overline{\widehat{g}[k_n]}.
\end{equation}
\end{prop}
\noindent The proof is given in Appendix~\ref{app: inner-product}. We have that
${\| f \|_{\mathcal{H}_{\mathrm{L}}}^2=\| \mathrm{L}f\|_{L_2}^2+ \gamma^2 \| \mathrm{Proj}_{\mathcal{N}_{\mathrm{L}}} \{f\}\|_{L_2}^2}$, where $\mathrm{Proj}_{\mathcal{N}_{\mathrm{L}}} \{f\}$ is given by~\eqref{eq: Proj null space}. The coefficient $\gamma^2$ balances the contribution of both terms.
\subsection{Periodic Reproducing-Kernel Hilbert Space} \label{sec:conditionRKHS}
Reproducing-kernel Hilbert spaces (RKHS) are Hilbert spaces on which the evaluation maps $f \mapsto f(t)$ are well-defined, linear, and continuous.
In this section, we answer the question of when the Hilbert space $\mathcal{H}_{\mathrm{L}}$ associated to an LSI operator $\mathrm{L}$ with finite-dimensional null space is a RKHS. This property is relevant to us because periodic function spaces that are RKHS are precisely the ones for which one can use measurement functions of the form $\nu_m = \Sha(\cdot - t_m)$ in~\eqref{eq: opt pb}.
\begin{definition}\label{def: RKHS}
Let $\mathcal{H}\subseteq \mathcal{S}^{\prime} (\mathbb{T})$ be a Hilbert space of $1$-periodic functions and $\mathcal{H}^{\prime}$ be its dual. Then, we say that $\mathcal{H}$ is a \emph{RKHS} if the shifted Dirac comb $\Sha (\cdot - t_0) \in \mathcal{H}^{\prime}$ for any $t_0 \in \mathbb{T}$.
\end{definition}
This implies that any element $f$ of a RKHS has a pointwise interpretation as a function $t \rightarrow f(t)$.
As is well known, for any RKHS there exists a unique function $h : \mathbb{T} \times \mathbb{T} \rightarrow \mathbb{R}$ such that $h(\cdot, t_0) \in \mathcal{H}^{\prime}$ and ${\langle f, h(\cdot, t_0)\rangle=f(t_0),}$ for every $ t_0 \in \mathbb{T}$ and $ f \in \mathcal{H}$. We call $h$ the \emph{reproducing kernel} of $\mathcal{H}$.
\begin{prop}\label{prop: RKHS}
Let $\mathrm{L}$ be a continuous LSI operator with finite-dimensional null space. The Hilbert space $\mathcal{H}_{\mathrm{L}}$ (see~\eqref{eq: search space}) is a RKHS if and only if
\begin{equation} \label{eq:condition RKHS}
\sum\limits_{k \in \mathcal{K}_{\mathrm{L}}}\frac{1}{|\widehat{L}[k]|^2}< + \infty.
\end{equation}
Then, the reproducing kernel for the scalar product~\eqref{eq: inner product} is given by $h(t,\tau)=h_{\gamma}(t-\tau)$, where $h_{\gamma} \in \mathcal{S}'(\mathbb{T})$ is
\begin{equation}\label{eq: reproducing kernel}
h_{\gamma}(t)=\sum_{n=1}^{N_0}\frac{e_{k_n}(t)}{\gamma^2}+\sum_{k \in \mathcal{K}_{\mathrm{L}}}\frac{e_k(t)}{|\widehat{L}[k]|^2}.
\end{equation}
\end{prop}
\smallskip
\noindent The proof is given in Appendix~\ref{app: RKHS}. Note that the reproducing kernel only depends on the difference $(t-\tau)$.
\subsection{Periodic Representer Theorem}
Now that we have defined the search space of the optimization problem~\eqref{eq: opt pb}, we derive the representer theorem that gives the explicit form of its unique periodic solution.
\begin{thm}\label{th: periodic representer thm}
We consider the optimization problem
\begin{equation}\label{eq: periodic representer thm}
\underset{f \in \mathcal{H}_{\mathrm{L}}}{\min}\bigg( F(\mathbf{y},\boldsymbol\nu (f))+\lambda \| \mathrm{L}f\| _{L_2}^2\bigg),
\end{equation}
where
\begin{itemize}
\item $F: \mathbb{R}^M \times \mathbb{R}^M \rightarrow \mathbb{R}^+$ is strictly convex and continuous;
\item $\mathrm{L}$ is an LSI operator with finite-dimensional null space;
\item $\boldsymbol \nu = (\nu_1, \dots ,\nu_M) \in (\mathcal{H}'_{\mathrm{L}})^M$ such that $\mathcal{N}_{\mathrm{L}}\cap \mathcal{N}_{\boldsymbol \nu}=\{0\}$;
\item $\mathbf{y}=(y_1, \ldots ,y_M) \in \mathbb{R}^M$ are the observed data; and
\item $\lambda > 0$ is a tuning parameter.
\end{itemize}
\noindent Then,~\eqref{eq: periodic representer thm}
has a unique solution of the form
\begin{equation}\label{eq:formsolutionRT}
f_{\mathrm{RT}}(t)=\sum\limits_{m=1}^M a_m \varphi_m(t)+\sum_{n=1}^{N_0}b_n e_{k_n}(t),
\end{equation}
where $a_m, b_n \in \mathbb{R}$, $\varphi_m=h_{\gamma} \ast \nu_m$, and $h_{\gamma}$ is given by~\eqref{eq: reproducing kernel}. Moreover, the vector $\mathbf{a} = (a_1, \ldots ,a_{M})$ satisfies the relation $\mathbf{P}^{\mathsf{T}} \mathbf{a} = \mathbf{0}$, with $\mathbf{P}$ the $(M\times N_0)$ matrix with entries ${[\mathbf{P}]_{m,n} = \langle e_{k_n}, \nu_m \rangle}$.
\end{thm}
\noindent The proof of Theorem~\ref{th: periodic representer thm} is given in Appendix~\ref{app: per rep th}.
The optimal solution depends on $(M+N_0)$ coefficients, but the condition $\mathbf{P}^\mathsf{T} \mathbf{a} = \mathbf{0}$ implies that there are only $(M+N_0-N_0)= M$ degrees of freedom.
In the case when $F$ is quadratic of the form~\eqref{eq: quadratic F}, the solution is made explicit in Proposition~\ref{prop: interpolation pb}.
\begin{prop}\label{prop: interpolation pb}
Under the conditions of Theorem~\ref{th: periodic representer thm}, if $F$ is given by~\eqref{eq: quadratic F}, then the vectors $\mathbf{a}$ and $\mathbf{b}$ satisfy the linear system
\begin{equation} \label{eq:matrixform}
\begin{pmatrix}
\mathbf{a} \\
\mathbf{b}
\end{pmatrix}=
\begin{pmatrix}
\mathbf{G}+ \lambda \mathbf{I} & \mathbf{P} \\
\mathbf{P}^{\mathsf{T}} & \mathbf{0}
\end{pmatrix}^{-1}
\begin{pmatrix}
\mathbf{y} \\
\mathbf{0}
\end{pmatrix},
\end{equation}
where $\mathbf{P} \in \mathbb{C}^{M\times N_0}$ is defined by $[\mathbf{P}]_{m,n}=\langle e_{k_n}, \nu_m \rangle$ and $\mathbf{G} \in \mathbb{R}^{M \times M}$ is a Gram matrix such that
\begin{equation}
\label{eq:Chgamma}
[\mathbf{G}]_{m_1,m_2} = \int_0^1\int_0^1 \nu_{m_1}(t)h_{\gamma}(t-\tau)\nu_{m_2}(\tau)\mathrm{d}t \mathrm{d}\tau.\end{equation}
\end{prop}
\noindent The proof is given in Appendix~\ref{app: interpolation pb}.
{In the case of sampling measurements, we show moreover in Proposition~\ref{prop: Spline} that the optimal solution is a periodic spline in the sense of Definition~\ref{def:Lspline}.
We recall that such measurements are valid as soon as the search space $\mathcal{H}_{\mathrm{L}}$ is a RKHS, a situation that has been fully characterized in Proposition~\ref{prop: RKHS}.}
\begin{prop}\label{prop: Spline}
{Under the conditions of Proposition~\ref{prop: interpolation pb}, if $\mathrm{L}$ satisfies~\eqref{eq:condition RKHS} and if the measurements are of the form ${\nu_m=\Sha(\cdot-t_m)}$, $t_m \in \mathbb{T}$, then the unique solution of~\eqref{eq: periodic representer thm} is a periodic $(\mathrm{L}^*\mathrm{L})$-spline with weights $a_m$ and knots $t_m$. }
\end{prop}
\noindent The proof is given in Appendix~\ref{app: Spline}.
\section{Periodic Processes and MMSE}\label{sec:processes_MMSE}
In this section, we change perspective and consider the following statistical problem: given noisy measurements of a zero-mean and real periodic Gaussian process, we are looking for the optimal estimator (for the mean-square error) of the complete process over $\mathbb{T}$.
\subsection{Non-Periodic Setting} \label{sec:Non-Per Setting}
In a non-periodic setting, it is usual to consider stochastic models where the random process $s$ is a solution to the stochastic differential equation~\cite{Unser2014sparse}
\begin{equation} \label{eq:Ls=w}
\mathrm{L} s = w,
\end{equation}
where $\mathrm{L}$ is a linear differential operator and $w$ a continuous domain (non-periodic) Gaussian white noise. When the null space of the operator is nontrivial, it is necessary to add boundary conditions such that the law of the process $s$ is uniquely defined.
\subsection{Gaussian Bridges} \label{sec:GGB}
In the periodic setting, the construction of periodic Gaussian processes has to be adapted. We first introduce the notion of periodic Gaussian white noise, exploiting the fact that the law of a zero-mean periodic Gaussian process $s$ is fully characterized by its covariance function $r_s(t,\tau)$ such that
\begin{equation}\label{eq: esp Gaussian process}
\mathbb{E} [\langle s , f \rangle \langle s , g \rangle ] = \int_0^1 \int_0^1 f(t) r_s(t,\tau) \overline{g(\tau)} \mathrm{d} t \mathrm{d} \tau.
\end{equation}
\begin{definition}\label{def: Gaussian white noise}
A \emph{periodic Gaussian white noise}\footnote{Without loss of generality, we only consider Gaussian white noise with zero-mean and variance $1$.} is a Gaussian random process $w$ whose covariance is ${r_w(t,\tau) = \Sha(t-\tau)}$.
\end{definition}
For any periodic real function $f$, the random variable $\langle w, f \rangle$ is therefore Gaussian with mean $0$ and variance $\lVert f \rVert_{L_2}^2$. Moreover, $\langle w, f \rangle$ and $\langle w, g \rangle$ are independent if and only if $\langle f , g \rangle = 0$. Hence, the Fourier coefficients $\widehat{w}[k]=\langle w,\mathrm{e}_k \rangle$ of the periodic Gaussian white noise satisfy the following properties:
\begin{itemize}
\item $\widehat{w}[k]=\Re(\widehat{w}[k])+ \mathrm{j} \ \Im(\widehat{w}[k])$;
\item $\overline{\widehat{w}[-k]}=\widehat{w}[k]$;
\item $\Re(\widehat{w}[k]), \ \Im(\widehat{w}[k]) \sim \mathcal{N}(0, \frac{1}{2})$, $\forall k>0$;
\item $\widehat{w}[0] \in \mathbb{R}$ and $\widehat{w}[0] \sim \mathcal{N}(0, 1)$;
\item $\Re(\widehat{w}[k]), \ \Im(\widehat{w}[k])$, and $\widehat{w}[0]$ are independent.
\end{itemize}
\smallskip
\noindent Put differently, for any nonzero frequency $k$, $\mathbb{E}[\widehat{w}[k]^2]=0$ and ${\mathbb{E}[\widehat{w}[k] \overline{\widehat{w}[k]}]=1}$. This means that $\widehat{w}[k]$, $k\neq 0$, follows a complex normal distribution with mean 0, covariance $1$, and pseudo-covariance $0$~\cite{Goodman1963}.
\begin{table*}
\caption{Gaussian bridges for several operators.}\label{Table: Gaussian Bridges}
\vspace*{-0.3cm}
\begin{center}
\begin{tabularx}{\textwidth}{M{1.2cm} | Y | Y | Y | Y }
\hline
\hline
\multicolumn{1}{M{1cm} |}{} & $\mathrm{D}+ \mathrm{I}$ & $\mathrm{D}$ & $\mathrm{D}^2+ 4 \pi^2 \mathrm{I}$ & $\mathrm{D}^2$\Tstrut\Bstrut\\
\hline
$\widehat{L}[k]$ & $\mathrm{j}2\pi k +1$ & $\mathrm{j}2\pi k$ & $4\pi ^2(1-k^2)$ & $-4\pi^2k^2$\Tstrut\\
$\mathcal{N}_{\mathrm{L}}$ & $\mathrm{span}\{0\}$ & $\mathrm{span}\{e_0\}$ & $\mathrm{span}\{e_1, e_{-1}\}$ & $\mathrm{span}\{e_0\}$\Tstrut\Bstrut\\
Gaussian bridges \tiny{$\gamma_0^2=1$} &
\begin{minipage}{0.5\textwidth}
\includegraphics[width=0.4\columnwidth]{D+I.pdf}
\end{minipage}&
\begin{minipage}{0.5\textwidth}
\includegraphics[width=0.4\columnwidth]{derivee.pdf}
\end{minipage}&
\begin{minipage}{0.5\textwidth}
\includegraphics[width=0.4\columnwidth]{ellipse.pdf}
\end{minipage}&
\begin{minipage}{0.5\textwidth}
\includegraphics[width=0.4\columnwidth]{D2.pdf}
\end{minipage}\Tstrut\Bstrut\\
\hline
\hline
\end{tabularx}
\end{center}\vspace*{-0.2cm}
\end{table*}
When $\mathrm{L}$ has a nontrivial null space, there is no hope to construct a periodic process $s$ solution of~\eqref{eq:Ls=w} with $w$ a periodic Gaussian white noise. Indeed, the operator $\mathrm{L}$ kills the null-space frequencies, which contradicts that $\widehat{w}[k_n] \neq 0$ almost surely for $n=1\ldots N_0$. One should adapt~\eqref{eq:Ls=w} accordingly by giving special treatment to the null-space frequencies.
We propose here to consider a new class of periodic Gaussian processes: the \emph{Gaussian bridges}. Given some operator $\mathrm{L}$ and $\gamma_0 >0$, we set
\begin{equation}\label{eq:Lgamma}
\mathrm{L}_{\gamma_0} = \mathrm{L} + \gamma_0 \mathrm{Proj}_{\mathcal{N}_{\mathrm{L}}},
\end{equation}
where $\mathrm{Proj}_{\mathcal{N}_{\mathrm{L}}}$ is given by~\eqref{eq: Proj null space}. Note that $\mathrm{L}_{\gamma_0} = \mathrm{L}$ for any $\gamma_0$ when the null space of $\mathrm{L}$ is trivial.
Moreover, we remark that
\begin{equation}
\lVert \mathrm{L}_{\gamma_0} f \rVert_{L_2}^2 = \lVert \mathrm{L} f \rVert_{L_2}^2 + \gamma_0^2 \lVert \mathrm{Proj}_{\mathcal{N}_{\mathrm{L}}} \{f\} \rVert_{L_2}^2 = \lVert f \rVert_{\mathcal{H}_{\mathrm{L}}}^2,
\end{equation}
where $\lVert f \rVert_{\mathcal{H}_{\mathrm{L}}}^2 = \langle f, f \rangle_{\mathcal{H}_{\mathrm{L}}}$ is given in~\eqref{eq: inner product} (with $\gamma =\gamma_0$).
\begin{definition}\label{def: model process}
A \emph{Gaussian bridge} is a periodic Gaussian process $s$, solution to the stochastic differential equation
\begin{equation} \label{eq:GB}
\mathrm{L}_{\gamma_0} s = w,
\end{equation}
with $w$ a periodic Gaussian white noise and $\mathrm{L}_{\gamma_0}$ given by~\eqref{eq:Lgamma} for some LSI operator $L$ with finite-dimensional null space and ${\gamma_0} > 0$. We summarize this situation with the notation $s\sim \mathcal{GB}(\mathrm{L}, \gamma_0^2)$. When the null space is trivial, in which case the parameter $\gamma_0^2$ is immaterial, we write $s \sim\mathcal{GB}(\mathrm{L})$.
\end{definition}
\noindent The Gaussian-bridge terminology is inspired by the Brownian bridge, the periodic version of the Brownian motion\footnote{Our definition differs from the classical one, in which the Brownian bridge is zero at the origin instead of being zero-mean~\cite{Revuz2013continuous}.}.
Several realizations of our Gaussian bridges for various operators are shown in Table~\ref{Table: Gaussian Bridges} for $\gamma_0^2=1$.
The influence of the parameter $\gamma_0^2$ is illustrated in Figure~\ref{fig: influence of gamma0}.
\begin{prop} \label{prop:covariancebridge}
The covariance function of the Gaussian bridge $s \sim \mathcal{GB}(\mathrm{L}, \gamma_0^2)$ is
\begin{equation}\label{eq: covariance GB}
r_s(t,\tau) = h_{\gamma_0} (t-\tau),
\end{equation}
where $h_{\gamma_0}$ is defined in~\eqref{eq: reproducing kernel}. It implies that
\begin{equation}\label{eq: esperance GB}
\mathbb{E} [\langle s , f \rangle \langle s , g \rangle ] =\langle h_{\gamma_0} \ast f,g\rangle.
\end{equation}
In particular, we have that
\begin{equation} \label{eq:variancesk}
\mathbb{E} [\lvert \widehat{s} [k]\rvert^2 ] = \widehat{h}_{\gamma_0} [k].
\end{equation}
\end{prop}
\noindent The proof of Proposition~\ref{prop:covariancebridge} is given in Appendix~\ref{app:covariance}. An important consequence is that a Gaussian bridge is stationary since its covariance function only depends on the difference $(t-\tau)$.
\subsection{Measurement Model and MMSE Estimator}
For this section, we restrict ourselves to operators $\mathrm{L}$ for which the native space $\mathcal{H}_{\mathrm{L}}$ is a RKHS.
In that case, using~\eqref{eq:variancesk} and~\eqref{eq: reproducing kernel}, the Gaussian bridge $s$ satisfies
\begin{equation}
\mathbb{E} [ \lVert s \rVert_{L_2}^2]
= \sum_{k\in \mathbb{Z}} \mathbb{E} [\lvert \widehat{s} [k]\rvert^2 ] = \sum_{k \in \mathcal{K}_{\mathrm{L}}} \frac{1}{\lvert \widehat{L} [k] \rvert^2} + \sum_{n=1}^{N_0} \frac{1}{\gamma_0^2},
\end{equation}
which is finite according to~\eqref{eq:condition RKHS}. Therefore, the Gaussian bridge $s$ is (almost surely) square-integrable.
The observed data $\mathbf{y}$ are assumed to be generated as
\begin{equation}\label{eq: noisy measurement}
\mathbf{y}= \langle \boldsymbol\nu , s \rangle + \boldsymbol{\epsilon},
\end{equation}
where $s \sim \mathcal{GB}(\mathrm{L}, \gamma_0^2)$ is a Gaussian bridge (see Definition~\ref{def: model process}), $\boldsymbol{\nu} = ( \nu_1, \ldots , \nu_M) $ is a vector of $M$ linear measurement functions, and $\boldsymbol{\epsilon}$ are independent random perturbations such that $\boldsymbol{\epsilon}\sim \mathcal{N}(\mathbf{0}, \sigma_0^2 \mathbf{I})$.
Given $\mathbf{y}$ in~\eqref{eq: noisy measurement}, we want to find the estimator $\tilde{s}$ of the Gaussian bridge $s$, imposing that it minimizes the quantity $\mathbb{E} [ \lVert s - \tilde{s} \rVert_2^2]$.
\begin{thm}\label{th: MMSE}
Let $\mathbf{y}=(y_1, \ldots ,y_M)$ be the noisy measurement vector~\eqref{eq: noisy measurement} of the Gaussian bridge $s \sim \mathcal{GB}(\mathrm{L}, \gamma_0^2)$, with measurement functions $\nu_m \in \mathcal{H}_{\mathrm{L}}'$, $m=1\ldots M$. Then,
the MMSE estimator of $s$ given the samples $\{y_m\}_{m \in [1 \ldots M]}$ is
\begin{equation}
\tilde{s}_{\mathrm{MMSE}}(t)=\sum_{m=1}^M d_m \varphi_m(t),
\end{equation}
\noindent where $\varphi_m=h_{\gamma_0} \ast \nu_m$ with $\nu_m \in \mathcal{H}_{\mathrm{L}}'$, $\mathbf{d}=(d_1, \ldots ,d_M)=(\mathbf{G}+ \sigma_0^2 \mathbf{I})^{-1}\mathbf{y}$, and $\mathbf{G}$ is the Gram matrix defined in~\eqref{eq:matrixform}.
\end{thm}
\noindent The proof is given in Appendix~\ref{app: MMSE}.
{Theorem~\ref{th: MMSE} can be seen as a generalization of the classical Wiener filtering, designed for discrete signals, to the hybrid case where the input signal is in a (periodic) continuous-domain and the (finite-dimensional) measurements are discrete.}
A leading theme of this paper is that the form of the MMSE estimator $\tilde{s}_{\mathrm{MMSE}}$ is very close to the one of the solution of the representer theorem $f_{\mathrm{RT}}$ with $\lambda = \sigma_0^2$ and for a quadratic cost function. This connection is exploited in Section~\ref{sec:discussion}.
\subsection{MMSE Estimation as a Representer Theorem}\label{sec:discussion}
The MMSE estimator given in Theorem~\ref{th: MMSE} can be interpreted as the solution of the optimization problem described in Proposition~\ref{prop: opt pb}.
\begin{prop}\label{prop: opt pb}
Consider an LSI operator $\mathrm{L}$ with finite-dimensional null space, $\gamma >0$, and $\nu_m \in \mathcal{H}_{\mathrm{L}}'$ for ${m=1\ldots M}$. We set $\mathrm{L}_{\gamma}$ as in~\eqref{eq:Lgamma}.
Then, the solution of the optimization problem
\begin{equation} \label{eq:L2reguWithGamma}
\underset{f \in \mathcal{H}_{\mathrm{L}}}{\min} \bigg ( \sum\limits_{m=1}^M (y_m-\langle f,\nu_m \rangle )^2 +\lambda \| \mathrm{L}_\gamma f \|_{L_2}^2 \bigg )
\end{equation}
exists, is unique, and given by
\begin{equation} \label{eq:optigamma}
f_\mathrm{opt}(t)= \sum\limits_{m=1}^M d_m \varphi_m(t),
\end{equation}
where $\varphi_m=h_\gamma \ast \nu_m$ and $\mathbf{d}=(d_1, \ldots ,d_M)=(\mathbf{G}+ \lambda \mathbf{I})^{-1}\mathbf{y}$.
In particular, the unique minimizer of~\eqref{eq:L2reguWithGamma} is the MMSE estimator given in Theorem~\ref{th: MMSE} for $\lambda = \sigma^2_0$ and $\gamma = \gamma_0$.
\end{prop}
\smallskip
\noindent The proof of Proposition~\ref{prop: opt pb} follows the same steps as the ones of Theorem~\ref{th: periodic representer thm} (form of the minimizer for the periodic representer theorem) and Proposition~\ref{prop: interpolation pb} (explicit formulas in terms of system matrix for the vectors $\mathbf{a}$ and $\mathbf{b}$), with significant simplifications that are detailed in Appendix~\ref{sec:proofRTgamma}. Proposition~\ref{prop: opt pb} has obvious similarities with Theorem~\ref{th: periodic representer thm}, but it also adds new elements.
\smallskip
\begin{itemize}[leftmargin=*]
\item Proposition~\ref{prop: opt pb} gives an interpretation of the MMSE estimator of a Gaussian bridge given its measurements as the solution to an optimization problem. This problem is very close to the periodic representer theorem (Theorem~\ref{th: periodic representer thm}) for a quadratic cost function. However,~\eqref{eq:L2reguWithGamma} differs from~\eqref{eq: periodic representer thm} because the regularization also penalizes null-space frequencies.
\smallskip
\item If the null space $\mathcal{N}_{\mathrm{L}}$ is trivial, then
\begin{equation}\label{eq:equivalence}
f_{\mathrm{RT}} = \tilde{s}_{\mathrm{MMSE}}
\end{equation}
for ${\lambda=\sigma_0^2}$. This means that Theorem~\ref{th: periodic representer thm} (smoothing approach) and~\ref{th: MMSE} (statistical approach) correspond to the same reconstruction method. This equivalence is well-known for stationary processes on $\mathbb{R}$ in the case of time-domain sampling measurements~\cite{Wahba1990spline}.
Our results extend this to the periodic setting and to the case of generalized linear measurements.
\smallskip
\item If the null space is nontrivial, then Theorem~\ref{th: periodic representer thm} and Proposition~\ref{prop: opt pb} yield different reconstructions. In particular, this implies that one cannot interpret the optimizer $f_{\mathrm{RT}}$ in Theorem~\ref{th: periodic representer thm} as the MMSE estimator of a Gaussian bridge. Yet, the solutions get closer and closer as $\gamma_0 \rightarrow 0$. In Section~\ref{sec:simulations}, we investigate more deeply this situation.
\end{itemize}
\section{Quality of the Estimators on Simulations}\label{sec:simulations}
We consider
$\tilde{s}_{\gamma, \lambda}(t|\mathbf{y})=\sum_{m=1}^M d_m \varphi_m(t)$ as the linear estimator of $s$ given $\mathbf{y}$, where $\varphi_m=h_{\gamma} \ast \nu_m$, $\mathbf{d}=(\mathbf{G}+ \lambda \mathbf{I})^{-1}\mathbf{y}$, and $\mathbf{G}$ is defined in Proposition~\ref{prop: interpolation pb}.
To simplify notations, we shall omit $\mathbf{y}$ when considering $\tilde{s}_{\gamma,\lambda}( \cdot | \mathbf{y}) = \tilde{s}_{\gamma,\lambda}$.
Each pair $(\lambda,\gamma)$ gives an estimator. In particular, if $s$ is a Gaussian bridge, then ${\tilde{s}_{\mathrm{MMSE}}=\tilde{s}_{\gamma_0, \sigma_0^2}}$, according to Theorem~\ref{th: MMSE}. The mean-square error (MSE) of $\tilde{s}_{\gamma, \lambda}$ over $N$ experiments is computed as $\mbox{MSE}=\frac{1}{N}\sum_{n=1}^N \| s_n -\big(\tilde{s}_{\gamma,\lambda}\big)_n \|_{L_2}^2$, where the $s_n$ are independent realizations of $s$ that yield a new noisy measurement $\mathbf{y}_n$ and $\big(\tilde{s}_{\gamma, \lambda}\big)_n = \tilde{s}_{\gamma,\lambda}(\cdot |\mathbf{y}_n)$ is the estimator based on $\mathbf{y}_n$. We define the normalized mean-square error (NMSE) by
\begin{equation}\label{eq: error lambda}
\mbox{NMSE}=\frac{\mbox{MSE}}{\frac{1}{N}\sum_{n=1}^N \| s_n \|_{L_2}^2}\approx \frac{\mathbb{E}[\| s-\tilde{s}_{\gamma,\lambda} \|_{L_2}^2]}{\mathbb{E}[\| s \|_{L_2}^2]}.
\end{equation}
In this section, we first detail the generation of Gaussian bridges (Section~\ref{sec: generation gaussian bridge}). We then investigate the role of the parameters $\lambda$ (Section~\ref{sec: lambda}) and $\gamma^2$ (Section~\ref{sec: gamma}) on the quality of the estimator $\tilde{s}_{\gamma, \lambda}$. We primarily focus on time-domain sampling measurements with $\langle \boldsymbol{\nu} , s \rangle =(s(t_1), \ldots, s(t_M) )^{\mathsf{T}}$, where the $t_m$ are in $\mathbb{T}$.
\subsection{Generation of Gaussian Bridges}\label{sec: generation gaussian bridge}
We first fix the operator $\mathrm{L}$ with null space $\mathcal{N}_{\mathrm{L}}$ of dimension $N_0$ and $\gamma_0>0$. Then, we generate $(2N_{\mathrm{coef}}+1)$ Fourier coefficients $\{\widehat{w}[k]\}_{k\in [-N_{\mathrm{coef}} \ldots N_{\mathrm{coef}}]}$ of a Gaussian white noise according to Definition~\ref{def: Gaussian white noise}. Finally, we compute the Gaussian bridge $s$ as
\begin{align}\label{eq: Gaussian bridge}
s(t)
&=\sum\limits_{\substack{k \in \mathcal{K}_{\mathrm{L}} \\ |k|\leq N_{\mathrm{coef}}}} \frac{\widehat{w}[k]}{\widehat{L}[k] }e_k(t) + \sum\limits_{n=1}^{N_0} \frac{\widehat{w}[k_n]}{\gamma_0}e_{k_n}(t).
\end{align}
Since $N_0 < \infty$,~\eqref{eq: Gaussian bridge} provides a mere approximation of the Gaussian bridge. However, the approximation error can be made arbitrarily small by taking $N_{\mathrm{coef}}$ large enough.
In Figure~\ref{fig: influence of gamma0}, we generate $s \sim \mathcal{GB}(\mathrm{D}^2 + 4\pi^2 \mathrm{I}, \gamma_0^2 )$ for four values of $\gamma_0^2$.
For small values of $\gamma_0^2$, the null-space component dominates, which corresponds in this case to the frequency $|k|= 1$. When $\gamma_0^2$ increases, the null-space component has a weaker influence.
%
\begin{figure}
\center
\subfigure[$\gamma_0^2=10^0$.]{
\includegraphics[width=0.4\columnwidth]{MMSE_gamma01}
}
\subfigure[$\gamma_0^2=10^2$.]{
\includegraphics[width=0.4\columnwidth]{MMSE_gamma02}
}
\subfigure[$\gamma_0^2=10^3$.]{
\includegraphics[width=0.4\columnwidth]{MMSE_gamma03}
}
\subfigure[$\gamma_0^2=10^6$.]{
\includegraphics[width=0.4\columnwidth]{MMSE_gamma04}
}
\caption{Illustration of $s \sim \mathcal{GB}(\mathrm{D}^2+4\pi^2 \mathrm{I}, \gamma_0^2)$ for different values of $\gamma_0^2$.}
\label{fig: influence of gamma0}
\end{figure}
\subsection{Influence of $\lambda$}\label{sec: lambda}
We evaluate the influence of the parameter $\lambda$ for the case of the invertible operator $\mathrm{L}= \mathrm{D} + \mathrm{I}$.
In this case we have that $\mathrm{Proj}_{\mathcal{N}_\mathrm{L}} = 0$ (since $\mathcal{N}_{\mathrm{L}} = \{0\}$), which simplifies~\eqref{eq:Lgamma}. Hence, the parameter $\gamma_0^2$ is immaterial and we denote by $\tilde{s}_\lambda$ the estimator associated to $\lambda > 0$.
We consider $s \sim \mathcal{GB}(\mathrm{D} + \mathrm{I})$ and $\sigma_0^2 = 10^{-2}$.
\smallskip
\textit{Time-Domain Sampling Measurements.}
We generated ${N=500}$ realizations of $s$. From each one, we extracted ${M=30}$ noisy measurements.
We then computed $30$ estimators $\{\big(\tilde{s}_{\lambda}\big)_n\}_{\lambda \in \mathcal{L}_1}$, where $\mathcal{L}_1$ is the set of values obtained by uniform sampling of the interval $[0.001, 0.03]$.
The plot of the NMSE (approximated according to~\eqref{eq: error lambda}) as a function of $\lambda$ is given in Figure~\ref{fig: plot lambda} (a).
The minimum error is obtained for $\lambda\simeq 0.01$, which corresponds to $\sigma_0^2$. This result validates the theory presented in Theorem~\ref{th: MMSE}.
Actually, when $\lambda$ is small, the estimator interpolates the noisy measurements while, for a large $\lambda$, the estimator tends to oversmooth the curve. The MMSE estimator makes an optimal tradeoff between fitting the data and smoothing the curve.
These observations about $\lambda$ retain their validity for other operators, including noninvertible ones.
\smallskip
\textit{Fourier-Domain Sampling Measurements.}
We consider complex exponential measurement functionals, inducing $\langle \bm{\nu}, s \rangle = (\widehat{s}[k_1], \ldots , \widehat{s}[k_M])^{\mathsf{T}}$, where the $k_m$ are in $\mathbb{Z}$. We define $\mathcal{N}_{\bm{\nu}} = \{k_m\}_{m=1\ldots M}$, such that $(-k_m) \in \mathcal{N}_{\bm{\nu}}$ for every $k_m \in \mathcal{N}_{\bm{\nu}}$.
We consider the measurements $\bm{\nu} = ( e_{k_1}, \ldots , e_{k_M})$. Note that these measurement functionals are complex, which calls for a slight adaptation of the framework presented so far\footnote{One could equivalently consider cosine and sine measurements, to the cost of heavier formulas.}. The noise $\bm{\epsilon} = (\epsilon_1, \ldots, \epsilon_M)$ is then also complex and satisfies the properties:
\begin{itemize}
\item $\epsilon_m=\Re(\epsilon_m)+ \mathrm{j} \ \Im(\epsilon_m)$;
\item $\epsilon_{m_1}=\overline{\epsilon_{m_2}}$, $k_{m_1}=-k_{m_2}$;
\item $\Re(\epsilon_m), \ \Im(\epsilon_m) \sim \mathcal{N}(0,\frac{\sigma_0^2}{2})$, $\forall k_m \neq 0$;
\item $\epsilon_m \in \mathbb{R}$ and $\epsilon_m \sim \mathcal{N}(0,\sigma_0^2)$, $k_m=0$;
\item $\Re(\epsilon_m), \ \Im(\epsilon_m)$ and $\epsilon_{m_1}$, $k_{m_1}=0$, are independent.
\end{itemize}
This means that $\mathbb{E}[\lvert \epsilon_m \rvert^2]=\sigma_0^2$ for every $m$.
We repeated the experiment done with the time-domain sampling using exactly the same procedure and parameters, and $\mathcal{N}_{\bm{\nu}}=\{-2,-1,0,1,2\}$. The experimental curve of the evolution of the NMSE with $\lambda$ is given in Figure~\ref{fig: plot lambda} (b). Again, the minimum is obtained for $\lambda\simeq 0.01=\sigma_0^2$. We now want to compare this curve to the theoretical one.
For the Fourier-sampling case, we were also able to derive the corresponding closed-form formulas for the NMSE~\eqref{eq: error lambda}.
\begin{prop}\label{prop:fouriersampling}
Let $s$ be a Gaussian bridge associated with an invertible operator $\mathrm{L}$, and $y_m = \widehat{s}[k_m] + \epsilon_m$, $m=1\ldots M$, with $k_m \in \mathcal{N}_{\bm{\nu}}$ the sampled frequencies and $\epsilon$ a complex Gaussian noise with variance $\sigma_0^2$ as above.
Then, the MSE of the estimator $\tilde{s}_{\lambda} = \tilde{s}_\lambda(\cdot | \mathbf{y})$ is given by
\begin{equation}\label{eq:FouriersamplingNMSE}
\mathbb{E}\left[ \lVert s - \tilde{s}_\lambda \rVert_{L_2}^2 \right] = \sum_{m=1}^M \frac{\widehat{h}[k_m] (\lambda^2 + \widehat{h}[k_m] \sigma_0^2)}{(\widehat{h}[k_m]+\lambda)^2} + \sum_{k\notin \mathcal{N}_{\bm{\nu}}} \widehat{h}[k],
\end{equation}
where $h$ is the reproducing kernel of $\mathcal{H}_{\mathrm{L}}$.
\end{prop}
\noindent The proof is given in Appendix~\ref{app:Fouriersampling}. Note that ${\widehat{h}[k]= 1 /\rvert \widehat{L}[k] \lvert^2}$ is real-valued and strictly positive for every $k$. From~\eqref{eq:FouriersamplingNMSE}, we also recover the property that the optimum is reached for $\lambda = \sigma_0^2$ since each of the $M$ terms that appear in the first sum is minimized for this value of $\lambda$.
The theoretical curve for $\mathcal{N}_{\bm{\nu}}=\{-2,-1,0,1,2\}$ is given in Figure~\ref{fig: plot lambda} (b) and is in good agreement with the experimental curve. We explain the slight variation ($0.15 \%$ for the $L_2$-norm over $\lambda\in [0.001,0.03]$) by the fact that~\eqref{eq: error lambda} is only an estimation of the theoretical NMSE.
\begin{figure}
\center\hspace*{-0.2cm}
\subfigure[{Time-domain sampling}.]{
\includegraphics[width=0.48\columnwidth]{Plots_lambda_Interpolation.pdf}
}\hspace*{+0.05cm}
\subfigure[Fourier-domain sampling.]{
\includegraphics[width=0.48\columnwidth]{Plots_lambda_Fourier.pdf}
}
\caption{
Evolution of the NMSE in terms of $\lambda$ for $s \sim \mathcal{GB}(\mathrm{D+I})$ for time and Fourier-domain sampling measurements.
}
\label{fig: plot lambda}
\end{figure}
\subsection{Influence of $\gamma^2$}\label{sec: gamma}
In this section, we only consider noninvertible operators since invertibility has already been addressed in Section~\ref{sec:discussion} (see~\eqref{eq:equivalence}).
In order to evaluate the specific influence of $\gamma$, we set $\lambda = \sigma_0^2$. Hence, $\tilde{s}_{\gamma, \sigma_0^2} = \tilde{s}_\gamma$. We generated $N=500$ realizations of a Gaussian bridge $s$, and from each one, we extracted $M=30$ noisy measurements. We repeated this for several operators $\mathrm{L}$ and values of $\gamma_0^2$ and $\sigma_0^2$. For each case, we compared $\tilde{s}_{\mathrm{MMSE}}$ to $\tilde{s}_{\gamma \rightarrow 0}$, $\tilde{s}_{\gamma \rightarrow \infty}$, and $f_{\mathrm{RT}}$ in~\eqref{eq:formsolutionRT}, seen here as an additional estimator. The corresponding NMSEs (see~\eqref{eq: error lambda}) are given in Table~\ref{Table: Comparison NMSE}. We make four observations.
\smallskip
1) In each case, the best result is obtained with $\tilde{s}_{\mathrm{MMSE}}$, as expected.
We see, moreover, that
${\lim_{\gamma \rightarrow 0}\mathbb{E}[\| s -\tilde{s}_{\gamma} \|_{L_2}^2] \simeq \mathbb{E}[\| s -f_\mathrm{RT} \|_{L_2}^2]}$.
This is in line with the fact that the functional~\eqref{eq: periodic representer thm} to minimize in Theorem~\ref{th: periodic representer thm} corresponds to~\eqref{eq:L2reguWithGamma} with $\gamma = 0$.
\smallskip
2) For small values of $\gamma_0^2$ (\textit{i.e.}, $10^{-3}$ or $10^0$), we see that $\mathbb{E}[\| s -f_\mathrm{RT} \|_{L_2}^2] \simeq \mathbb{E}[\| s -\tilde{s}_{\mathrm{MMSE}} \|_{L_2}^2]$. This means that the performances of $\tilde{s}_{\mathrm{MMSE}}$ and $f_{\mathrm{RT}}$ are very similar. This is illustrated in Figure~\ref{fig: plot Table} (a), where $\tilde{s}_{\mathrm{MMSE}}$ and $f_{\mathrm{RT}}$ do coincide. Meanwhile, we see that ${\lim_{\gamma \rightarrow \infty}\mathbb{E}[\| s -\tilde{s}_{\gamma} \|_{L_2}^2] \gg \mathbb{E}[\| s -\tilde{s}_{\mathrm{MMSE}} \|_{L_2}^2]}.$ This is also illustrated in Figure~\ref{fig: plot Table} (a) for $\mathrm{L=D}$. The reconstruction for $\gamma \rightarrow + \infty $ significantly fails to recover the original signal $s$, as the corresponding estimator tends to have zero-mean.
\smallskip
3) For intermediate values of $\gamma_0^2$ (\textit{i.e.}, $\gamma_0^2=10^3$ or $10^6$ according to $\sigma_0$ and the order of the operator), the minimal NMSE is obtained for $\tilde{s}_{\mathrm{MMSE}}$ only. We also observe that ${\mathbb{E}[\| s -f_\mathrm{RT} \|_{L_2}^2] < \lim_{\gamma \rightarrow \infty}\mathbb{E}[\| s -\tilde{s}_{\gamma} \|_{L_2}^2]}.$ This is illustrated in Figure~\ref{fig: plot Table} (b) for $\mathrm{L=D^2+4\pi^2 I}$, $\gamma_0^2=10^6$ and $\sigma_0^2=10^{-4}$, where we can distinguish $\tilde{s}_{\mathrm{MMSE}}$, $\tilde{s}_{\gamma \rightarrow \infty}$, and $f_{\mathrm{RT}}$.
\smallskip
4) For large values of $\gamma_0^2$ (\textit{i.e.}, $\gamma_0^2 =10^9$), we observe that ${\lim_{\gamma \rightarrow \infty}\mathbb{E}[\| s -\tilde{s}_{\gamma} \|_{L_2}^2] \simeq \mathbb{E}[\| s -\tilde{s}_{\mathrm{MMSE}} \|_{L_2}^2]}$ and ${\mathbb{E}[\| s -f_{\mathrm{RT}} \|_{L_2}^2] > \mathbb{E}[\| s -\tilde{s}_{\mathrm{MMSE}} \|_{L_2}^2]}.$ In fact, for large $\gamma_0^2$, the Gaussian bridge tends to have vanishing null-space frequencies (with~\eqref{eq: Gaussian bridge}, we have that $\widehat{s}[k_n] = \widehat{w}[k_n] / \gamma_0$ for $n=1 \ldots N_0$).
Meanwhile, the reconstructed signal $f_{\mathrm{RT}}$ is not constrained to attenuate null-space frequencies.
The null-space part in~\eqref{eq:formsolutionRT} is mainly responsible for a higher error compared to $\tilde{s}_{\mathrm{MMSE}}$. This is highlighted in Figure~\ref{fig: plot Table} (c).
\smallskip
Observations 2), 3), and 4) suggest the existence of three regimes.
For further investigation, we present in Figure~\ref{fig: plot gamma} the evolution of NMSE as a function of $\log{\gamma^2}$ for $\mathrm{L=D}$ and $\gamma_0^2=10^0, 10^3$, and $10^6$.
The minimal error is always obtained for $\gamma^2 \simeq \gamma^2_0$, as predicted by the theory.
For the three cases, we observe two plateaus: one for $\gamma^2 \in (0,v_1)$ and the other for $\gamma^2 \in (v_2, \infty)$, where $v_1$, $v_2 > 0$. It means that, for each value of $\gamma_0^2$, the estimators $\tilde{s}_{\gamma}$ with $\gamma^2 \in (0,v_1)$ ($(v_2, \infty)$, respectively) are very similar and the reconstruction algorithms are practically indistinguishable. The values of $v_1$ and $v_2$ depend on $\gamma_0^2$. When $\gamma_0^2=10^0$ ($10^6$, respectively), we have that $\gamma_0^2 \in (0,v_1)$ ($(v_2, \infty)$, respectively). However, $\gamma_0^2=10^3 \in [v_1, v_2]$ belongs to none of the plateaus.
Two main conclusions can be drawn from our experiments.
First, we have strong empirical evidence that
\begin{equation}
\tilde{s}_\gamma \underset{\gamma\rightarrow 0}{\longrightarrow} f_{\mathrm{RT}},
\end{equation}
which we conjecture to be true for any Gaussian-bridge model.
This is remarkable because it presents the reconstruction based on the periodic representer theorem as a limit case of the statistical approach.
Second, we empirically see that, for reasonably small values of $\gamma_0^2$, the estimators corresponding to $\gamma^2 \leq \gamma_0^2$ are practically indistinguishable from the MMSE estimator.
This is in particular valid for the representer-theorem reconstruction, for which we then have that
\begin{equation}
f_{\mathrm{RT}} \approx \tilde{s}_{\mathrm{MMSE}}.
\end{equation}
\noindent The variational method is \emph{theoretically} suboptimal to reconstruct Gaussian bridges. However, based on our experiments, it is reasonable to consider this method as \emph{practically} optimal for small values of $\gamma_0^2$ and $\lambda=\sigma_0^2$.
\begin{table*}
\caption{Comparison of NMSE for $\tilde{s}_{\gamma \rightarrow 0}$, $f_{\mathrm{RT}}$, $\tilde{s}_{\mathrm{MMSE}}$, and $\tilde{s}_{\gamma \rightarrow \infty}$ over $N=500$ iterations. \textbf{Bold}: optimal result.}\label{Table: Comparison NMSE}
\vspace*{-0.3cm}
\begin{center}
\begin{tabularx}{\textwidth}{M{0.9cm} | M{0.6cm} | Y | Y | Y | Y | Y | Y | Y | Y }
\hline\hline
\multicolumn{2}{Y |}{} & \multicolumn{4}{ c |}{$\sigma_0=10^{-1}$} & \multicolumn{4}{c }{$\sigma_0=10^{-2}$}\Tstrut\Bstrut \\
\hline
$\mathrm{L}$ & $\gamma_0^2$ & $\tilde{s}_{\gamma \rightarrow 0}$ & $f_{\mathrm{RT}}$ & $\tilde{s}_{\mathrm{MMSE}}$ & $\tilde{s}_{\gamma \rightarrow \infty}$ & $\tilde{s}_{\gamma \rightarrow 0}$ & $f_{\mathrm{RT}}$ & $\tilde{s}_{\mathrm{MMSE}}$ & $\tilde{s}_{\gamma \rightarrow \infty}$\Tstrut\Bstrut\\
\hline
\multirow{5}{*}{$\mathrm{D}$}& $10^{-3}$ & $\mathbf{1.37 \times 10^{-5}}$ & $\mathbf{1.37 \times 10^{-5}}$ & $\mathbf{1.37 \times 10^{-5}}$ & $1.78$ & $\mathbf{8.40 \times 10^{-6}}$ & $\mathbf{8.40 \times 10^{-6}}$ & $\mathbf{8.40 \times 10^{-6}}$ & $2.94$\Tstrut\\
& $10^0$ & $\mathbf{1.17 \times 10^{-2}}$ & $\mathbf{1.17 \times 10^{-2}}$ & $\mathbf{1.17 \times 10^{-2}}$ & $1.66$ & $\mathbf{8.44 \times 10^{-3}}$ & $\mathbf{8.44 \times 10^{-3}}$ & $\mathbf{8.44 \times 10^{-3}}$ & $2.72$\\
& $10^3$ & $1.59 \times 10^{-1}$ & $1.56 \times 10^{-1}$ & $\mathbf{1.49 \times 10^{-1}}$ & $1.58 \times 10^{-1}$ & $1.05 \times 10^{-1}$ & $1.05 \times 10^{-1}$ & $\mathbf{9.96 \times 10^{-2}}$ & $1.21 \times 10^{-1}$\\
& $10^6$ & $1.61 \times 10^{-1}$ & $1.60 \times 10^{-1}$ & $\mathbf{1.43 \times 10^{-1}}$ & $\mathbf{1.43 \times 10^{-1}}$ & $1.07 \times 10^{-1}$ & $1.07 \times 10^{-1}$ & $\mathbf{9.11 \times 10^{-2}}$ & $\mathbf{9.11 \times 10^{-2}}$\\
& $10^9$ & $1.66 \times 10^{-1}$ & $1.66 \times 10^{-1}$ & $\mathbf{1.47 \times 10^{-1}}$ & $\mathbf{1.47 \times 10^{-1}}$ & $1.10 \times 10^{-1}$ & $1.10 \times 10^{-1}$ & $\mathbf{9.34 \times 10^{-2}}$ & $\mathbf{9.34 \times 10^{-2}}$\Bstrut\\
\hline
\multirow{5}{*}{$\mathrm{D^2}$} & $10^{-3}$ & $\mathbf{8.43 \times 10^{-7}}$ & $\mathbf{8.43 \times 10^{-7}}$ & $\mathbf{8.43 \times 10^{-7}}$ & $1.07$ & $3.12 \times 10^{-8}$ & $\mathbf{3.11 \times 10^{-8}}$ & $\mathbf{3.11 \times 10^{-8}}$ & $1.34$\Tstrut\\
& $10^0$ & $9.06 \times 10^{-4}$ & $9.06 \times 10^{-4}$ & $\mathbf{9.05 \times 10^{-4}}$ & $1.07$ & $\mathbf{3.34 \times 10^{-5}}$ & $\mathbf{3.34 \times 10^{-5}}$ & $\mathbf{3.34 \times 10^{-5}}$ & $1.33$\\
& $10^3$ & $4.04 \times 10^{-1}$ & $4.04 \times 10^{-1}$ & $\mathbf{3.61 \times 10^{-1}}$ & $7.1 \times 10^{-1}$ & $\mathbf{1.46 \times 10^{-2}}$ & $\mathbf{1.46 \times 10^{-2}}$ & $\mathbf{1.46 \times 10^{-2}}$ & $5.78 \times 10^{-1}$\\
& $10^6$ & $6.53 \times 10^{-1}$ & $6.53 \times 10^{-1}$ & $\mathbf{3.66 \times 10^{-1}}$ & $\mathbf{3.66 \times 10^{-1}}$ & $2.63 \times 10^{-2}$ & $2.63 \times 10^{-2}$ & $\mathbf{2.26 \times 10^{-2}}$ & $2.29 \times 10^{-2}$\\
& $10^9$ & $6.62 \times 10^{-1}$ & $6.62 \times 10^{-1}$ & $\mathbf{3.86 \times 10^{-1}}$ & $\mathbf{3.86 \times 10^{-1}}$ & $2.65 \times 10^{-2}$ & $2.65 \times 10^{-2}$ & $\mathbf{2.16 \times 10^{-2}}$ & $\mathbf{2.16 \times 10^{-2}}$\Bstrut\\
\hline
\multirow{5}{*}{$\mathrm{D^2+4 I}$} & $10^{-3}$ & $\mathbf{5.53 \times 10^{-7}}$ & $\mathbf{5.53 \times 10^{-7}}$ & $\mathbf{5.53 \times 10^{-7}}$ & $1.03$ & $\mathbf{1.71 \times 10^{-8}}$ & $\mathbf{1.71 \times 10^{-8}}$ & $\mathbf{1.71 \times 10^{-8}}$ & $1.22$\Tstrut\\
& $10^0$ & $5.56 \times 10^{-4}$ & $5.56 \times 10^{-4}$ & $\mathbf{5.55 \times 10^{-4}}$ & $1.04$ & $\mathbf{1.77 \times 10^{-5}}$ & $\mathbf{1.77 \times 10^{-5}}$ & $\mathbf{1.77 \times 10^{-5}}$ & $1.24$\\
& $10^3$ & $3.67 \times 10^{-1}$ & $3.67 \times 10^{-1}$ & $\mathbf{3.04 \times 10^{-1}}$ & $8.79 \times 10^{-1}$ & $1.21 \times 10^{-2}$ & $1.21 \times 10^{-2}$ & $\mathbf{1.20 \times 10^{-2}}$ & $8.71 \times 10^{-1}$\\
& $10^6$ & $1.52$ & $1.52$ & $\mathbf{4.63 \times 10^{-1}}$ & $\mathbf{4.63 \times 10^{-1}}$ & $3.94 \times 10^{-2}$ & $3.94 \times 10^{-2}$ & $\mathbf{2.98 \times 10^{-2}}$ & $3.04 \times 10^{-2}$\\
& $10^9$ & $1.47$ & $1.47$ & $\mathbf{4.87 \times 10^{-1}}$ & $\mathbf{4.87 \times 10^{-1}}$ & $4.67 \times 10^{-2}$ & $4.67 \times 10^{-2}$ & $\mathbf{3.18 \times 10^{-2}}$ & $\mathbf{3.18 \times 10^{-2}}$\Bstrut\\
\hline\hline
\end{tabularx}
\end{center}
\end{table*}
\begin{figure*}\vspace*{-0.15cm}
\center\hspace*{-0.1cm}
\subfigure[$\mathrm{L=D}$, $\gamma_0^2=1$, and $\sigma_0^2=10^{-2}$.]{
\includegraphics[width=0.62\columnwidth]{TableD.pdf}
}\hspace*{0.15cm}
\subfigure[$\mathrm{L=D^2+4 \pi^2 I}$, $\gamma_0^2=10^6$, and $\sigma_0^2=10^{-4}$.]{
\includegraphics[width=0.62\columnwidth]{TableEllipse.pdf}
}\hspace*{0.15cm}
\subfigure[$\mathrm{L=D^2}$, $\gamma_0^2=10^9$, and $\sigma_0^2=10^{-4}$.]{
\includegraphics[width=0.62\columnwidth]{TableD2.pdf}
}
\caption{Illustrations of $s \sim \mathcal{GB}(\mathrm{L}, \gamma_0^2)$, $\tilde{s}_{\mathrm{MMSE}}$, $f_{\mathrm{RT}}$, and $\tilde{s}_{\gamma \rightarrow \infty}$ for several operators and values of $\gamma_0^2$ and $\sigma_0^2$ . We used $M=30$ noisy measurements $\mathbf{y}=(y_1, \ldots, y_M)$.}
\label{fig: plot Table}\vspace{-0.2cm}
\end{figure*}
\begin{figure*}
\center\hspace*{+0.2cm}
\subfigure[$\gamma_0^2=1$.]{
\includegraphics[width=0.58\columnwidth]{plot_gamma20.pdf}
}\hspace*{0.8cm}
\subfigure[{$\gamma_0^2=10^3$.}]{
\includegraphics[width=0.58\columnwidth]{plot_gamma.pdf}
}\hspace*{0.8cm}
\subfigure[$\gamma_0^2=10^6$.]{
\includegraphics[width=0.58\columnwidth]{plot_gamma3.pdf}
}
\caption{Evolution of NMSE according to $\gamma$ for $s \sim \mathcal{GB}(\mathrm{D}, \gamma_0^2)$.}
\label{fig: plot gamma}\vspace{-0.5cm}
\end{figure*}
\section{Discussion}\label{sec:periodicvsline}
\subsection{Comparison with Inverse Problems on the Real Line}
It is worth noting that the periodic setting has important differences as compared to reconstruction methods over the complete real line, which motivated and played an important role in this paper.
\smallskip
\begin{itemize}[leftmargin=*]
\item The role of the Dirac impulse $\delta$ is played by the Dirac comb $\Sha$ in the periodic setting. It is indeed the neutral element of the periodic convolution~\eqref{eq: periodic convolution} and appears in the definition of the periodic $\mathrm{L}$-splines (Definition~\ref{def:Lspline}) and RKHS (Definition~\ref{def: RKHS}).
\smallskip
\item In the real-line setting, in addition to smoothness properties, functions are also characterized by their property of decay at infinity~\cite{Simon2003distributions}. For periodic functions, we only consider the smoothness properties, which brings substantial simplifications.
\smallskip
\item In general, a continuous LSI operator does not preserve the asymptotic behavior of the input function. For instance, a test function in the space $\mathcal{S}(\mathbb{R})$ of smooth and rapidly decaying functions is not necessarily mapped to a rapidly decaying function. In contrast, any continuous LSI operator maps the space of periodic test functions $\mathcal{S}(\mathbb{T})$ onto itself (see Section~\ref{subsec:LSIop}). This greatly simplifies the study of operators that act on periodic functions.
\smallskip
\item The null space of a continuous LSI operator can differ for the two cases. In particular, when acting on periodic functions, the null space of the $n$th derivative $\mathrm{D}^n$ is reduced to constant functions for every $n \geq 1$. This is crucial due to the role of the null space in Theorems~\ref{th: periodic representer thm} and~\ref{th: MMSE}.
\smallskip
\item In Proposition~\ref{prop: RKHS}, we give a necessary and sufficient condition for a continuous LSI operator of finite-dimensional null space to specify a RKHS in the sense of Definition~\ref{def: RKHS}. This is significantly more complicated over the real line, for which only partial results are known~\cite{Unser2016splines}.
\smallskip
\item We have seen that it is not always possible to find a periodic solution $s$ to the equation $\mathrm{L} s = w$, where $w$ is a periodic Gaussian white noise. This lead us to modify the stochastic differential equation (see~\eqref{eq:GB}) and to introduce the family of Gaussian bridges.
\smallskip
\item In Theorem~\ref{th: MMSE}, we give the MMSE estimator of the \emph{complete} process $s$, not only for the estimation of $s(t_0)$ at a fixed time $t_0$. In the non-periodic setting, however, solutions of stochastic differential equations are generally not square-integrable. For instance, if $s$ is a nontrivial stationary Gaussian process, then
\begin{align}
\mathbb{E} [\lVert s \rVert_{L_2(\mathbb{R})}^2] &= \sum_{k\in \mathbb{Z}} \mathbb{E}[\lVert 1_{[k,k+1)}\cdot s \rVert_{L_2(\mathbb{R})}^2]\nonumber \\
& \overset{(i)}{=} \sum_{k\in \mathbb{Z}} \mathbb{E}[\lVert 1_{[0,1)}\cdot s \rVert_{L_2(\mathbb{R})}^2] = \infty,
\end{align}
where $1_{[a,b)}$ is the indicator function on $[a,b)$ and $(i)$ exploits stationarity. Another example is the Brownian motion, whose supremum over $[0,t]$ grows faster than $t^p$ for any $p<1/2$ (almost surely) when $t$ goes to infinity~\cite{Karatzas2012brownian}, hence being of infinite energy.
As a consequence, it is irrelevant to consider the MMSE estimator of the complete process and one ought to, for instance, restrict to MMSE estimators of local values $s(t_0)$ of the process.
\end{itemize}
\subsection{Comparison with TV Regularization}\label{sec:TV}
A recent tendency in the field of signal reconstruction is to rely on sparsity-promoting regularization, motivated by the fact that many real-world signals are sparse in some adequate transform domain~\cite{Elad10,mallat2008wavelet,Unser2014sparse}.
The vast majority of works focuses on the finite-dimensional setting via $\ell_1$-type regularization. However, some authors have recently promoted the reconstruction of infinite-dimensional sparse signals~\cite{Adcock2015generalized,candes2014towards}. The adaptation of discrete $\ell_1$ methods to the continuous domain is based on the total-variation (TV) regularization norm, for which it is possible to derive representer theorems (see~\cite[Theorem 1]{Unser2016splines}). A comparison between Tikhonov and TV variational techniques is proposed in Gupta \textit{et al.}~\cite{gupta2018continuous} for non-periodic signals. In brief, at identical measurements and regularization operator $\mathrm{L}$, Tikhonov regularization favors smooth solutions restricted to a finite-dimensional space, while TV regularization allows for adaptive and more compressible solutions. In~\cite[Table I]{gupta2018continuous}, it was shown on simulations that Tikhonov methods perform better on fractal-type signals, while TV methods are better suited to sparse signals. We expect similar behaviors for the periodic setting.
At the heart of the present paper is the connection between $L_2$-regularization and the statistical formalism of MMSE estimation of Gaussian processes. A theoretical link between deterministic and stochastic frameworks is much harder to provide for sparsity-inducing priors. There is strong empirical evidence that sparse stochastic models are intimately linked to TV-based methods~\cite{Unser2014sparse}, but the extent to which such estimators approach the MMSE solution is still unknown.
\section{Conclusion}\label{sec:conslusions}
We have presented two approaches for the reconstruction of periodic continuous-domain signals from their corrupted discrete measurements.
The first approach is based on optimization theory and culminates with the specification of a periodic representer theorem (Theorem~\ref{th: periodic representer thm}).
In the second approach, a signal is modeled as a stationary periodic random process and the reconstruction problem is transformed into an estimation problem. Theorem~\ref{th: MMSE} then gives the optimal estimator (in the mean-square sense) for Gaussian bridges.
We have also provided theoretical and experimental comparisons of the two approaches and identified two main findings.
First, for invertible operators, the statistical and variational approaches are equivalent and correspond to an identical reconstruction scheme.
For noninvertible operators, however, this equivalence is not valid anymore, but the variational method corresponds to the statistical reconstruction when the parameter $\gamma$ vanishes. More importantly, for small values of $\gamma_0^2$, the variational method is practically equivalent to the optimal statistical reconstruction.
This demonstrates the efficiency of the representer theorem for reconstructing Gaussian bridges, even for noninvertible operators.
|
\section{Introduction}
In this note, we make further remarks on a question of Moshe Newman, which asked whether it is the case that if $H$ and $K$ are isomorphic subgroups of a finite solvable group $G$ and $H$ is maximal in $G$, then $K$ is also maximal. This continues work begun in [1], where it was proved that this is true if $H$ either has Abelian Sylow $2$-subgroups or a Sylow tower, and it was noted (in Theorem 3 of that paper) that in a minimal counterexample $G$, we have $O_{p}(G) = 1$, where $[G:H]$ is a power of the prime $p$.
\section{Notation, Assumed Background, and Preliminary Results}
\medskip
Recall that a finite group $L$ is said to be involved in the finite group $G$ if there is a subgroup $H$ of $G$ and a normal subgroup $K$ of $G$ such that $H/K \cong L.$ We will make frequent use of the first lemma throughout.
\medskip
\noindent {\bf Lemma 1:} \emph{ Let $X$ be a finite solvable group, and $\pi$ be a set of primes. Let $T$ be a finite $\pi$-group and let $Y$ be a Hall $\pi$-subgroup of $X.$ Then $T$ is involved in $X$ if and only if $T$ is involved in $Y$.}
\medskip
\noindent {\bf Proof:} It is clear that $T$ is involved in $X$ if $T$ is involved in $Y$. We prove the opposite implication by induction on $|X|$. Suppose that $U,V$ are subgroups of $X$ with $V \lhd U$ and $U/V \cong T.$
Then $T$ is certainly involved in $U$.
\medskip
If $|U| < |X|,$ then $T$ is involved in a Hall $\pi$-subgroup of $U$ by induction. But any Hall $\pi$-subgroup of $U$ is conjugate to a subgroup of $Y$, so that $T$ is involved in $Y$. Hence we may suppose that $U = X$ and that $V \lhd X$ with $X/V \cong T.$
\medskip
Let $M$ be a minimal normal subgroup of $X$ contained in $V.$ Then $M$ is either a $\pi$-group or a $\pi^{\prime}$-group. Now $T \cong (U/M)/(V/M)$ so that $T$ is involved in $X/M.$ By induction, $T$ is involved in the Hall $\pi$-subgroup $YM/M$ of $X/M.$
\medskip
If $M$ is a $\pi$-group then $Y/M$ is a Hall $\pi$-subgroup of $X/M$ and we are done since we already remarked that $T$ is involved in $Y/M.$ If $M$ is a $\pi^{\prime}$-group, then $YM/M \cong Y,$ so $T$ is involved in $Y,$ and the proof of the lemma is complete.
\medskip
Let $r$ be a prime. The group ${\rm Qd}(r)$ is the semi-direct product of the natural module for ${\rm SL}(2,r)$
with ${\rm SL}(2,r).$ Note that $|{\rm Qd}(r)| = r^{3}(r^{2}-1)$ and that ${\rm Qd}(2)$ is isomorphic to the symmetric group $S_{4}.$ Note also that ${\rm Qd}(3)$ is a $\{2,3\}$-group. We recall that ${\rm Qd}(r)$ is solvable if and only if $r < 5.$ In this note, the more elementary fact that $|{\rm Qd}(r)|$ has at least three prime divisors when $r > 3$ usually suffices for our purposes. For $r-1$ and $r+1$ can't both be powers of $2$ when $r$ is a prime greater than $3$.
\medskip
Using Lemma 1, we note that if ${\rm Qd}(r)$ is involved in a solvable group $X,$ then $r < 5$ and ${\rm Qd}(r)$ is involved in a Hall $\{2,3\}$-subgroup of $X.$
\medskip
Using Theorems of Stellmacher and Glauberman, and Lemma 1, for each finite $r$-group $R,$ there is a characteristic subgroup $W(R)$ of $R$ (which is non-trivial whenever $R$ is non-trivial) such that whenever $X$ is a finite solvable group with Sylow $r$-subgroup $R,$ then we have $X = O_{r^{\prime}}(X) N_{X}(W(R))$ and furthermore $O_{r^{\prime}}(X)W(R)$ is characteristic in $X,$ unless, perhaps, $r <5$ and ${\rm Qd}(r)$ is involved in a Hall $\{2,3\}$-subgroup of $X$. When $r$ is odd, we may take $W(R) = ZJ(R).$
\medskip
When $n$ is an integer, we let $\pi(n)$ denote the set of prime divisors of $n$.
We will make frequent use of the following easy lemma:
\medskip
\noindent {\bf Lemma 2:} \emph{ Let $X$ be a finite solvable group and $\pi$ be a set of primes.
Let $Y$ be a subgroup of $X$ with $\pi([X:Y]) \subseteq \pi$. Then $O_{\pi}(Y) \leq O_{\pi}(X).$ }
\medskip
\noindent {\bf Proof:} Let $U$ be a Hall $\pi$-subgroup of $X$ containing $O_{\pi}(Y).$ Since $Y$ contains a Hall $\pi^{\prime}$-subgroup of $X,$ we have $X = YU = UY.$ Now we have
$$O_{\pi}(X) = \cap_{x \in X} U^{x} = \cap_{y \in Y} U^{y} \geq O_{\pi}(Y),$$ as claimed.
\medskip
\noindent {\bf Corollary 3:} \emph{Let $X$ be a finite solvable group and $Y$ be a subgroup of $X$ whose index is a power of the prime $s.$ Let $r$ be a prime divisor of $X$ different from $s$. Then $O_{\{r,s\}}(Y)$ and $O_{\{r,s\}}(X)$ have a common Sylow $r$-subgroup.}
\medskip
\noindent {\bf Proof:} By Lemma 2, applied with $\pi = \{r,s\}$, we have $O_{\{r,s\}}(Y) \leq O_{\{r,s\}}(X).$ Note that $Y$ contains a Sylow $r$-subgroup of $X,$ so that $Y$ certainly contains a Sylow $r$-subgroup $R$ of $O_{\{r,s\}}(X).$ Now $$R \cap Y \leq O_{\{r,s\}}(X) \cap Y \leq O_{\{r,s\}}(Y),$$ so that
$O_{\{r,s\}}(Y)$ contains a Sylow $r$-subgroup of $O_{\{r,s\}}(X).$ On the other hand, since
$O_{\{r,s\}}(Y) \leq O_{\{r,s\}}(X),$ a Sylow $r$-subgroup of $O_{\{r,s\}}(Y)$ is contained
in some Sylow $r$-subgroup of $O_{\{r,s\}}(X).$
\section{ Statement and Proof of Theorem A}
\medskip
\noindent {\bf Theorem A:} \emph{ Let $H$ be a maximal subgroup of the finite solvable group $G$ and suppose that $[G:H] = p^{a}$ where $p$ is a prime and $a$ is a positive integer. Let $K$ be a subgroup of $G$ which is isomorphic to $H.$ Suppose that $K$ is not maximal in $G.$}
\medskip
\emph{Then $p \leq 3,$ and, for $q = 5-p,$ we have $$O_{q^{\prime}}(H) = O_{q^{\prime}}(G) = O_{q^{\prime}}(K)$$
and, for $G^{\ast} = G/O_{q^{\prime}}(G),$ etc., $H^{\ast}$ and $K^{\ast}$ are isomorphic subgroups
of $G^{\ast}$ with $H^{\ast}$ maximal and $K^{\ast}$ not maximal.}
\medskip
\noindent {\bf Proof:} Let $\phi: H \to K$ be an isomorphism.
\medskip
We proceed by induction on $|G|$. Note that if $N$ is a $\phi$-invariant normal subgroup of $H$ then
$$N = N\phi \lhd H\phi = K$$ and then $N \lhd \langle H,K \rangle = G.$ Then $\phi$ induces an isomorphism
between $H/N$ and $K/N$, and $H/N$ is maximal in $G/N$, but $K/N$ is not maximal in $G/N$.
\medskip
Using Theorem 3 of [1] and Lemma 2, we may conclude that $O_{p}(G) = 1.$ For if not, then we have
$O_{p}(G) \subseteq O_{p}(G) \cap H$ by the former result, and both $O_{p}(H) \leq O_{p}(G)$ and
$O_{p}(K) \leq O_{p}(G)$ by the latter. Then we have $$O_{p}(G) = O_{p}(H)$$ and
$$O_{p}(H)\phi = O_{p}(K) \leq O_{p}(G).$$ Since $$|O_{p}(G)| = |O_{p}(H)| = |O_{p}(K)|,$$
we have $$O_{p}(G)\phi = O_{p}(G),$$ so we may apply the argument above with $O_{p}(G)$ in the role of $N$.
\medskip
We may suppose by induction that the theorem holds for the triple $(G/N,H/N,K/N).$ Then $p \leq 3$
since $[G/N:H/N] = [G:H],$ and we note that $$G/O_{q^{\prime}}(G) \cong (G/N)/O_{q^{\prime}}(G/N)$$ for $q = 5-p$. Hence the theorem holds for $G$ in this case. Thus we may suppose that $O_{p}(G) = 1.$
\medskip
Since $O_{p}(G) = 1$ and (by Lemma 2) $O_{p}(H) \leq O_{p}(G)$, we see that $F(H)$ is a $p^{\prime}$-group, as is the isomorphic group $F(K)$. Also, there is a prime $r \neq p$ such that $O_{r}(H) \neq 1.$
\medskip
Suppose first that $r > 3.$ Then by Lemma 1, we have $$O_{\{r,p\}}(H) \leq O_{\{r,p\}}(G)$$ and likewise,
$$O_{\{r,p\}}(K) \leq O_{\{r,p\}}(G).$$ Furthermore, by Corollary 3 we may suppose (possibly after replacing $K$ by a conjugate) that $O_{\{r,p\}}(H),O_{\{r,p\}}(G)$ and $O_{\{r,p\}}(K)$ all have a common Sylow $r$-subgroup, say $S$.
\medskip
Since $O_{p}(H) = O_{p}(K) = O_{p}(G) = 1,$ we see that $W(S)$ is characteristic in each of
$O_{\{r,p\}}(H),O_{\{r,p\}}(K)$ and $O_{\{r,p\}}(G),$ where $W(S)$ is the Glauberman-Stellmacher characteristic subgroup of $S$ (for note that $O_{\{r,p\}}(G)$ does not involve ${\rm Qd}(2)$ or ${\rm Qd}(3)$ since $r > 3,$
and likewise for $O_{\{r,p\}}(H)$ and $O_{\{r,p\}}(K))$. Thus $W(S) \lhd G.$ Furthermore, $$O_{\{r,p\}}(H)\phi = O_{\{r,p\}}(K)$$ and composing $\phi$ with an inner automorphism of $K$ if necessary, we may suppose that $S\phi = S,$ in which case $W(S)\phi = W(S).$ Then $\phi$ induces an isomorphism between $H/W(S)$ and $K/W(S).$
\medskip
By induction, the theorem holds for $G/W(S).$ In particular, $p \leq 3,$ and, setting $q = 5-p,$ we see that $W(S)$ is a $q^{\prime}$-group, and that the Theorem therefore holds for $G.$
\medskip
Hence we may suppose that $F(H)$ is a $q$-group for some prime $q \leq 3.$\\
If $p \neq 5-q,$ we may argue as above that $W(T)$ is normal in each of\\ $O_{\{q,p\}}(H),O_{\{q,p\}}(G)$ and $O_{\{q,p\}}(K),$ where $T$ is a common Sylow $q$-subgroup of $O_{\{q,p\}}(H),O_{\{q,p\}}(G)$ and $O_{\{q,p\}}(K).$ This time, $O_{\{q,p\}}(G)$ does not involve ${\rm Qd}(2)$ or ${\rm Qd}(3)$ since $p > 3.$
But that leads to a contradiction, since the theorem holds for $G/W(T)$ and then we see that
$[G:H] = [G/W(T):H/W(T)]$ is either a power of $2$ or power of $3,$ so $p \leq 3,$ contrary to current assumptions.
\medskip
Now we may suppose that $p \leq 3$ and that $F(H)$ is a $q$-group, where $q = 5-p.$ Since $[G:H]$ is a power of $p$ and $O_{p}(G) = 1,$ we have $F(G) \leq H,$ so that $F(G) \leq F(H)$ and $F(G)$ is a $q$-group.
Since $H \cong K,$ we also see that $F(K)$ is a $q$-group. The proof of Theorem A is complete.
\section{Statement and Proof of Theorem B and some consequences}
\medskip
\medskip
\noindent {\bf Theorem B:} \emph{ Let $H$ be a maximal subgroup of the finite solvable group $G$ and suppose that $[G:H] = p^{a}$ where $p \leq 3$ is a prime and $a$ is a positive integer. Let $K$ be a subgroup of $G$ which is isomorphic to $H.$ Suppose that $K$ is not maximal in $G$ and that $F(H),F(K)$ and $F(G)$ are all
$q$-groups, where $q = 5-p.$ Let $Q$ be a Sylow $q$-subgroup of $H$.}
\medskip
\emph{ Then $G$ has a homomorphic image $G^{\ast}$ such that $H^{\ast}$ and $K^{\ast}$ (the respective images of $H$ and $K$) are isomorphic subgroups of $G^{\ast}$ with $H^{\ast}$ maximal and $K^{\ast}$ not maximal, and with $F(H^{\ast}),F(K^{\ast})$ and $F(G^{\ast})$ all $q$-groups. Furthermore, $O_{\{2,3\}}(H^{\ast})$ involves ${\rm Qd}(q)$ and no non-identity characteristic subgroup of $Q^{\ast}$ is normal in $H^{\ast}.$ }
\medskip
\noindent {\bf Proof:} Let $\phi:H \to K$ be the isomorphism as before. We may, and do, suppose that $Q\phi = Q,$
composing $\phi$ with an inner automorphism of $K$ if necessary.
\medskip
Using Theorem A, we may suppose that there is no non-trivial $\phi$-invariant normal subgroup of $H.$
For otherwise, if $1 \neq N \lhd H$ is $\phi$-invariant and chosen of maximal order subject to these conditions,
then $N = N \phi \lhd H\phi = K,$ so $N \lhd \langle H,K \rangle = G$ and $\phi$ induces an isomorphism between $H/N$ and $K/N$. By Theorem A, we have $$O_{q^{\prime}}(H/N) = O_{q^{\prime}}(G/N) = O_{q^{\prime}}(K/N)= 1$$ by the maximal choice of $N$ (hence $F(H/N),F(K/N)$ and $F(G/N)$ are all $q$-groups) and the theorem holds for $G/N$ by induction, so it holds for $G.$
\medskip
Suppose now that $O_{\{2,3\}}(H)$ does not involve ${\rm Qd}(q).$ Then as before, $Q$ contains a common Sylow $q$-subgroup $T$ of $O_{\{2,3\}}(H), O_{\{2,3\}}(K)$ and $O_{\{2,3\}}(G)$ such that $W(T) \lhd H$
and $W(T)\phi = W(T),$ contradicting the fact that there is no non-trivial $\phi$-invariant normal subgroup of $H$. Hence $O_{\{2,3\}}(H)$ involves ${\rm Qd}(q).$
\medskip
If there is a non-identity characteristic subgroup $S$ of $Q$ which is normal in $H,$ then we have $S = S\phi$ since $Q = Q\phi,$ again contradicting the fact that no non-trivial $\phi$-invariant normal subgroup of $H$. The proof of Theorem B is complete.
\medskip
We may combine Theorems A and B to deduce:
\medskip
\noindent {\bf Corollary C:} \emph{ Let $G$ be a finite solvable group and
$H,K$ be isomorphic subgroups of $G$ such that $H$ is maximal but $K$ is not, and with $[G:H] = p^{a}$ for some prime $p$ and positive integer $a$. Suppose (with no loss of generality) that $H \cap K$ contains a Hall $p^{\prime}$-subgroup of $G$ (this is just a matter of replacing $K$ by a conjugate if necessary). Then $p \leq 3$. Let $\phi: H \to K$ be an isomorphism, chosen so that $Q\phi = Q$ for a Sylow $q$-subgroup $Q$ of $H,$ where $q = 5-p$ (this can be achieved by composing $\phi$ with an inner automorphism of $K$ if necessary). Let $C$ be the unique maximal $\phi$-invariant normal subgroup of $G$ which is contained in $H \cap K,$ and let $G^{\ast} = G/C,$ etc.. Then $F(H^{\ast}),F(K^{\ast})$ and $F(G^{\ast})$ are all $q$-groups. Also, ${\rm Qd}(q)$ is involved in $O_{\{2,3\}}(H^{\ast}),$ and no non-identity characteristic subgroup of $Q^{\ast}$ is normal in $H^{\ast}.$ }
\medskip
\noindent {\bf Proof:} The proofs of Theorems A and B show that unless $G$ itself can play the role of $G^{\ast},$ there is a non-identity $\phi$-invariant normal subgroup $N$ of $G$ contained in both $H$ and $K$ such that $H/N$ and $K/N$ are isomorphic with $H/N$ maximal in $G/N$ and $K/N$ not maximal. In that case, since
$N \leq C,$ we may suppose that the Theorem holds in $G/N$ by induction, and then it holds in $G.$
\medskip
\noindent {\bf Corollary D:} \emph{Let $G$ be a finite solvable group containing isomorphic subgroups $H$ and $K$ such that $H$ is maximal in $G$ but $K$ is not. Then $[G:H] = [G:K]$ is a power of a prime $p \leq 3$ and a Hall $\{2,3\}$-subgroup $L$ of $H$ involves ${\rm Qd}(q),$ where $q = 5-p.$ In particular, $L$ involves $S_{3},A_{4}$ and a non-Abelian group of order $8,$ so that $L$ is neither $2$-closed nor $2$-nilpotent.}
\medskip
\begin{center}
{\bf References}
\end{center}
\medskip
\noindent [1] Isaacs, I.M. \& Robinson, G.R., \emph{Isomorphic subgroups of solvable groups}, Proc Amer. Math. Soc
{\bf 143} (2015) 3371-3376.
\end{document}
\section{Can we reduce to $\{2,3\}$-groups?}
Now let $(G,H,K)$ be a minimal violating triple for Newman's question, and suppose that $G$ is not a $\{2,3\}$-group. Then $G = O_{\{2,3\}}(G)H.$ For we have $O_{\{2,3\}}(H) \leq O_{\{2,3\}}(G)$ by Lemma 2.
If $O_{\{2,3\}}(G) \leq H,$ then we have $$O_{\{2,3\}}(H) \leq O_{\{2,3\}}(G) \geq O_{\{2,3\}}(K)$$
using Lemma 2 again, and then $$O_{\{2,3\}}(H) = O_{\{2,3\}}(G) = O_{\{2,3\}}(K)$$ is a $\phi$-invariant normal subgroup of $H$, contrary to Corollary C. Now we have $G = O_{\{2,3\}}(G)H$ by the maximality of $H$. Since we do not know whether $p =2$ or $3$ at present it is convenient to assume that $F(G) = O_{q}(G)$ where $q \leq 3$ and $p=5-q.$
\section{$O_{q,p}(G)$}
We know that $G = O_{q,p}(G)H$ and we know that in ${\bar G} = G/O_{q}(G)$ we have $O_{p}({\bar H}) \lhd {\bar G}.$ We know that $G^{\ast} ={\bar G}/O_{p}{\bar H}$ is the semi-direct product $S^{\ast}H^{\ast}$ where
$S^{\ast} = O_{p}(G^{\ast}).$
\medskip
Now $O_{q}(H)^{\ast}$ acts faithfully on $S^{\ast}$ since $\overline{O_{q}(H)}$ centralizes $O_{p}({\bar H}).$
\medskip
We know that $O_{\{p,q\}}(G)$ is neither $p$-closed nor $p$-nilpotent. Hence $$O_{q}(G^{\ast}/S^{\ast}) \neq 1.$$
Thus $$O_{q}(H^{\ast}) \neq 1.$$
\end{document}
|
\section{Summary of Fabrication and Measurements Methods}
\subsection{Fabrication details}
All devices are fabricated on x-cut single crystalline thin-film lithium niobate (LN) wafers (NANOLN). The wafer stack consists of a 600 nm thin-film LN layer, a 2 $\mu$m thermally grown SiO$_2$ layer and a 500 $\mu$m silicon handle layer. Standard electron-beam (e-beam) lithography is used to pattern optical waveguide and micro-racetrack resonators. The patterns are then transferred into the LN layer using argon (Ar$^+$) plasma etching in an inductively coupled plasma reactive ion etching (ICP-RIE) tool13. The etch depth is 350 nm, leaving a 250 nm thick LN slab behind, which enables efficient electric field penetration into the waveguide core. Gold contact patterns are then created using aligned e-beam lithography, and the metal is transferred using e-beam evaporation methods and lift-off processes. The chip is then diced and the facet is polished for end-fire optical coupling.
\subsection{Microwave driving circuitry }
The 10 GHz microwave drive signal is generated by a radio-frequency (RF) synthesizer and amplified by an electrical power amplifier. The amplified electrical signal is passed through a microwave circulator and delivered to the microelectrodes. As the microelectrodes represent a capacitive load, most of the electrical driving signal is reflected back to the circulator and terminated at the circulator output by a 50-$\Omega$ load.
In the dual-drive EO comb generation experiment, two RF synthesizers are phase-locked via a common 10 MHz clock and are free to operate at different frequencies. The two sinusoidal microwave signals are power balanced and combined using an RF power splitter and passed through the amplifier-circulator circuitry described previously.
\subsection{Optical characterization and detection}
Light from a tunable laser (SANTEC TS510) is launched into, and the comb output is collected from, the LN waveguides by a pair of lensed optical fibers. The output comb is passed to an optical spectrum analyser OSA having a minimum resolution of 20 pm. This finite resolution accounts for the limited signal-to-noise ratio observed in Fig 2b ($\sim$ 20 dB). The shot-noise-limited signal-to-noise ratio is much higher, as the comb shot noise lies below the OSA noise floor. Although the measurement in the paper is chosen to center at 1600 nm, the frequency comb center wavelength can be flexibly chosen between 1500 nm to 1680 nm of the tunable laser’s range without affecting much of the generated comb width.
In the dual-drive EO comb measurements, the modulated light is passed to a fast photodetector (New Focus 1544A) and the resulting electrical signal is sent to a RF spectrum analyzer to record the beating in the RF domain.
\subsection{Measurement and calculation of resonator parameters }
As demonstrated by Equation (4) of the Supplementary Information, there are four resonator parameters that fully characterize the EO comb spectrum: the internal round-trip transmission coefficient α, the power coupling coefficient k, the coupler insertion loss of the coupler γ, and the phase modulation index β. Finding each of these four parameters by fitting to the comb spectrum of Equation (4) is difficult because the output comb can be fully determined by a subset of these independent parameters (e.g., increasing the modulation index has the same effect as decreasing the loss in the resonator). Instead, each of the parameters must be measured separately.
We find $\alpha$ and $\kappa$ by measuring the total transmitted power without phase modulation (Figure 2b right inset). By fitting to the expected transmission of an all-pass ring resonator, we find $Q=1.5\times10^6, \alpha=0.95$ and $\kappa=0.027$. Then we perform a grid search optimization for $\gamma$ and $\beta$ comparing the measured output spectrum (Fig \ref{fig2}b) with the spectrum determined from the output time-domain electric field of Equation (3) of the Supplementary Information. We find a best fit for $\gamma=-0.004$ dB and $\beta=1.2 \pi$, where the average difference between experimental and theoretical comb line power is 0.6 dB.
The output power transmission for nonzero modulation indices (Fig. \ref{fig2}b right inset) is calculated by sampling the output electric field with Equation (3) of the Supplementary Information and averaging the power over more than 100 modulation periods.
\subsection{Dispersion simulations in thin-film LN waveguides}
The dispersion of the waveguide is simulated using finite-domain-time-difference methods (LUMERICAL). The simulation accounts for the LN material anisotropy and the finite waveguide etching angle (around $70^\circ$ from horizontal). The round-trip phase of the light inside the resonator is calculated by integrating the simulated group velocity dispersion twice to determine the total frequency-dependent phase-shift. We find that for an air-cladded waveguide with a 600 nm thin-film LN layer, 350 nm etch depth and 1.5 $\mu$m waveguide width, a comb spanning $\sim$ 1.2 octave can be generated, as shown in Fig \ref{fig3}e.
\section{Canonical EO Comb Generator Design}
\subsection{Resonant Operation}
A canonical waveguide-based comb generator is shown in Figure 1c of the main text. A single-frequency input with electric field $E_{in}(t)=\hat{E}_{in} e^{i \omega_0 t}$ is coupled, with power coupling coefficient $k$ and insertion loss $\gamma$, to a resonator having round trip time $T$ at center frequency $\omega_0$ and round trip power loss $\alpha$. The resonator contains a phase modulator driven with modulation index $\beta$ and frequency $\omega_m$. The output electric field is ~\cite{ho_optical_1993}:
\begin{equation}
\begin{aligned}
\label{eq:E_out1}
E_{out} = \sqrt{(1-\gamma)(1-k)} E_{in}(t) - k \sqrt{\frac{1-\gamma}{1-k}} \sum_{n=1}^{\infty} r^n e^{-i \beta F_n(\omega_m t)} E_{in} (t -n T),
\end{aligned}
\end{equation}
where $r = \sqrt{(1-\gamma)(1-k) \alpha}$ is the round trip electric field transmission and $ F_n(\omega_m t)= \sum_{i=1}^n \sin{\omega_m(t-iT)}$ is the modulator coherence function. The parameter $l = 1-r$, corresponding to the round-trip electric field loss, is used in the main text for simplicity. When the optical carrier is resonant ($\omega_0 T = 2 \pi m_1$) and the microwave drive signal is resonant ($\omega_m T = 2 \pi m_2$), the modulator coherence function becomes $F_n (\omega_m t) = n \sin{\omega_m(t-iT)}$ and the output electric field can be simplified to
\begin{equation}
\begin{aligned}
\label{eq:E_out2}
E_{out}(t) = \bigg[ \sqrt{(1-\gamma)(1-k)} - k \sqrt{\frac{1-\gamma}{1-k}} \frac{r e^{-i \beta \sin{\omega_m t}}}{1 - r e^{-i \beta \sin{\omega_m t}}} \bigg] E_{in}(t).
\end{aligned}
\end{equation}
This output electric field corresponds to an optical frequency comb spaced at the modulation frequency. The power in the $q$th comb line away from the center frequency can be found by rewriting Equation \ref{eq:E_out1} as
\begin{equation}
\begin{aligned}
\label{eq:E_out3}
E_{out}(t) = \sqrt{(1-\gamma)(1-k)} \hat{E}_{in} e^{i \omega_0 t} - k \sqrt{ \frac{1-\gamma}{1-k} } \hat{E}_{in} \sum_{q = -\infty}^{\infty} e^{i(\omega_0+q \omega_m)t} \sum_{n=1}^{\infty} r^n J_q(\beta n),
\end{aligned}
\end{equation}
where $J_q$ is the $q$th order Bessel function of the first kind. The power of the $q$th (nonzero) comb line is then
\begin{equation}
\begin{aligned}
\label{eq:Pq_r}
P_q = k^2 \frac{1-\gamma}{1-k} P_{in} \bigg| \sum_{n = 1}^{\infty} r^n J_q(\beta n) \bigg|^2 .
\end{aligned}
\end{equation}
~\cite{kourogi_wide-span_1993} found an approximation for the power of the $q$th comb as $P_q \propto e^{-\frac{|q|(1-r)}{\beta}}$.
\subsection{Nonresonant Operation}
In the presence of optical and microwave detuning from resonance, the comb spectrum can still be calculated. When the optical carrier is off resonance, the total round-trip phase is $\omega_0 T = 2 \pi m_1+ \phi_{opt}$. Similarly, when the microwave carrier is off resonance the total round-trip phase is $\omega_m T = 2 \pi m_2+ \phi_{micro}$. Using these expressions in Equation~\ref{eq:E_out1}, we can find the following expression for the power in the $q$th comb line:
\begin{equation}
\begin{aligned}
\label{eq:Pq_nr}
P_q = k^2 \frac{1-\gamma}{1-k} P_{in} \bigg| \sum_{p = -\infty}^{\infty} \sum_{n = 1}^{\infty} (r e^{i \phi_{opt}})^n e^{i p \frac{\pi}{2}} J_{q-p}(\beta_o(\phi_{micro},n)) J_p(\beta_e(\phi_{micro},n)) \bigg|^2 .
\end{aligned}
\end{equation}
The modified even and odd modulation indices ($\beta_e$ and $\beta_o$, respectively) are
\begin{equation}
\begin{aligned}
\label{eq:beta_e}
\beta_e(\phi_{micro},n) = \beta \bigg[ \frac{1}{2} \cot{\phi_{micro}/2} - \frac{\cos{(n + \frac{1}{2}) \phi_{micro}}}{2 \sin{\phi_{micro}/2}} \bigg]
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
\label{eq:beta_o}
\beta_o(\phi_{micro},n) = \beta \bigg[ -\frac{1}{2} + \frac{ \sin{(n + \frac{1}{2}) \phi_{micro}} }{2 \sin{\phi_{micro}/2}} \bigg].
\end{aligned}
\end{equation}
It is clear here that in the regime of low optical detuning, the slope of the comb decreases by a factor of $\cos{(\phi_{opt})}$. This effect has been studied and reported in~\cite{saitoh_modulation_1998}. The effect of microwave detuning is harder to visualize, but results in a destructive interference condition for large values of $q$ in Equation~\ref{eq:Pq_nr}. This effect is demonstrated experimentally and theoretically in Figure 3a and 3b of the main text.
\subsection{Noise Properties}
The optical phase noise of the comb lines is important in applications that require high optical signal-to-noise ratios, such as high-capacity optical communications. It is well known that the optical phase noise contribution from the pump laser does not increase with increasing comb line index $q$ \cite{ho_optical_1993}. By contrast, the phase noise contribution from the microwave modulation signal increases in power with comb line quadratically with $q$. This can be shown by modifying the modulator coherence function to include the effects of microwave modulation phase noise $\theta(t)$:
\begin{equation}
\begin{aligned}
\label{eq:F_n}
F_n(\omega_m t) = \sum_{i=1}^n \sin{\omega_m (t - i T + \theta(t - i T) )}.
\end{aligned}
\end{equation}
The output optical field can then be written as:
\begin{equation}
\begin{aligned}
\label{eq:E_out4}
E_{out}(t) = \sqrt{(1-\gamma)(1-k)} \hat{E}_{in} e^{i \omega_0 t} - k \sqrt{\frac{1-\gamma}{1-k}} \hat{E}_{in} \sum_{p = -\infty}^{\infty} \sum_{n = 1}^{\infty} r^n J_q(\beta n) e^{i(\omega_0 + q \omega_m)t + i q \theta(t)}.
\end{aligned}
\end{equation}
The phase noise amplitude increases linearly with increasing comb line index $q$, corresponding to a quadratic increase in phase noise power.
For applications that require few comb lines, this increase in microwave phase noise is often negligible because quartz crystal oscillators have very low phase noise. For applications requiring many comb lines, however, the effect of microwave phase noise may be noticeable. Recently, there has been experimental evidence of microwave phase noise suppression in EO comb generators \cite{kim_cavity_2016}, \cite{kim_cavity-induced_2017}. In these studies, the phase noise increase can be dramatically suppressed when the EO comb generator is driven off resonance, both optically and electrically. These experiments suggest that EO comb generators can generate low-noise comb lines over their entire dispersion-limited bandwidth. Integrated platforms, such as the one presented in the main text, enable additional filtering cavities and structures to be readily included in the resonator structure.
\section{Round-Trip Phase Model}
To include the effect of dispersion, we introduce a round-trip phase model. In particular, we consider the destructive interference that occurs due to the microwave detuning motivates a phase-based resonance approximation for the viable comb bandwidth. Previous analytical work~\cite{kourogi_limit_1995} provided a mathematical treatment of the dispersion limits of resonator-based EO comb generators. Here, as a complement to that work, we clarify the physical interpretation of the round-trip phase model and demonstrate its application to combs of arbitrary bandwidth within a given dispersion-limited window.
\subsection{Resonance Conditions}
The resonance condition of an optical frequency $\omega_q$ in a microresonator without EO modulation is $|\omega_q T - 2 \pi N | < 2 l$, where the total round-trip phase offset $\Delta \phi_q = \omega_q T - 2 \pi N$, $T=1/$FSR is the round-trip time and $N$ is the number of optical cycles per round-trip that ensures that $|\Delta \phi_q | < 2 \pi$. Frequency components outside of the resonance are attenuated by destructive interference, and thus do not resonate. When the resonance condition is satisfied, the optical fields constructively interfere inside the resonator at every time and spatial location.
In a resonator containing an EO phase modulator, the (now time-dependent) resonance condition becomes $|\Delta \phi_q + \beta \sin{2 \pi f_m t}| < 2l$, where $\beta$ is the modulation index and $f_m$ is the modulation frequency. Here, it is clear that the resonance condition can be satisfied for much larger round-trip phase offsets $\Delta \phi_q$ because within the round-trip resonator propagation time, the modulation term oscillates between negative and positive $\beta$ (i.e. $-\beta < \beta \sin{2 \pi f_m t} < \beta$).
This effect may be understood by plotting the total transmission of the EO comb generator for various $\beta$, as shown in Figure 2b, right inset, of the main text. The transmission is calculated by averaging the output power of a time-domain representation of the electric field given in Equation~\ref{eq:E_out1}. The optical power output depends primarily on the interference between the input optical field and the optical field inside the resonator. As in a microresonator without EO modulation, the dips in the transmission spectrum correspond to a large built-up field inside the resonator. For various values of $\beta$, the width of the resonance increases, indicating that for large modulation indices, the resonance condition can be satisfied for various detuning values. As shown in Figure 2b, the amount of detuning is approximately equal to the modulation index $\beta$, as is predicted by the phase model when $\Delta \phi_q= \phi_{opt}$.
\subsection{Frequency-Dependent Round-Trip Phase}
We can now determine the contributions to the optical phase offset $\Delta \phi_q$ as a function of frequency. The optical phase offset, as discussed in the previous section, does not induce frequency-dependent phase shifts. However, microwave signal detuning and dispersion effects are frequency dependent.
Once the resonator has reached steady state, the output field is an EO comb spaced at the modulation frequency $f_m$, such that the $q$th comb line frequency is $f_q = f_0+ q f_m$. A mismatch between the microwave frequency and the resonator free spectral range results in a frequency-dependent phase offset $\phi_{micro}(q) = 2 \pi q $(FSR)$ T$.
For an arbitrary dispersion profile, it is possible to find the frequency-dependent phase offset by integrating the group velocity dispersion profile of the waveguide. However, if the dispersion is approximately linear with frequency, the dispersion-related phase offset is $\Delta \phi_{disp}(q) = 2 \pi (q f_m)^2 \beta_2 L $ where $\beta_2 L$ is the round-trip group velocity dispersion in ps/nm.
To first order we then have a model for the total phase offset as a function of frequency, $\Delta \phi_q = \Delta \phi_{opt}+ \Delta \phi_{micro}(q) + \Delta \phi_{disp}(q)$. In fact, this model agrees exactly with the analytical model for the output comb shape developed in~\cite{kourogi_wide-span_1993}. In the case of maximum comb bandwidth, corresponding to zero microwave detuning and optical detuning satisfying $\phi_{opt} + \beta = 0$, the maximum dispersion-limited bandwidth is $ \Delta f_{comb} = \frac{1}{\pi} \sqrt{\frac{2 \beta}{\beta_2 L}} $, agreeing with~\cite{kourogi_limit_1995} up to a factor of $\sqrt{2}$ due to the difference in FSR of a Fabry-Pérot resonator and ring resonator of identical length.
Using this model, it is a straightforward optimization problem to start with the frequency-dependent round-trip resonance condition and alter the optical and microwave detuning so that the resonance condition is satisfied only for a desired frequency region, as is done to demonstrate the one-sided comb in Figure 3c of the main text.
\end{document}
|
\section{Introduction}
The region beyond Neptune is populated by hundreds of thousands of planetesimals. These trans-Neptunian objects (TNOs) are the fossils left over after the era of planet construction in our Solar System. Their orbits are a record of the outer Solar System's past evolution (Sections \ref{sec:dynamics}-\ref{sec:history}), and their surface composition is a window into the conditions in the early planet-forming disk (Sections \ref{sec:surfaces}-\ref{sec:motivations}). We developed the Colours of the Outer Solar System Origins Survey (Col-OSSOS) to map the surface properties of the trans-Neptunian populations, through near-simultaneous $g$-, $r$- and $J$-band observations with the 8.1-m Frederick C. Gillett Gemini North Telescope.
Col-OSSOS provides color measurements for a set of TNOs that were detected in a flux-limited survey with a well-measured detection efficiency. The survey targets TNOs brighter than $m_r = 23.6$ mag found in the Outer Solar System Origins Survey \cite[OSSOS;][]{Bannister2016,Bannister2018}.
Col-OSSOS affords the first opportunity to explore the true frequency of surface colors within the Kuiper belt, subdivided by dynamical classification.
Initial results are discussed in \citet{Fraser2017,Pike2017} and \citet{Marsset2018}.
We were also able to compare Col-OSSOS TNO measurements to the first interstellar object, 1I/`Oumuamua \citep{Bannister2017}.
In this paper, we provide an overview of the Col-OSSOS survey, detailing the observing strategy to obtain near-simultaneous optical and near-infrared (NIR) colors (Section \ref{sec:design}) and summarizing our data analysis strategies (Section \ref{sec:colours} and Appendixes~\ref{sec:optical} and ~\ref{sec:NIR}).
Our first data release is 35 TNOs (Section \ref{sec:sample}), which display three color-dynamical taxonomic groups (Section \ref{sec:first_colours}).
We infer the observed and debiased ratio of the two color groups of the red and neutral surfaces within the dynamically excited Kuiper belt population (Section \ref{sec:colour_fraction}).
We briefly examine the implications for the radial color distribution in the primordial planetesimal disk from which the excited TNOs originated (Section \ref{sec:disk}).
\section{The Orbital Structure and Surface Compositions of Trans-Neptunian Objects}
\subsection{Dynamical Populations}
\label{sec:dynamics}
The main Kuiper belt (trans-Neptunian region) can be defined as the conglomeration of minor planets on orbits with semimajor axes between 37 and $\sim50$ au\footnote{More precisely, between the $\nu$18 secular resonance and the 2:1 mean motion resonance with Neptune at 47.5~au, though the cold classicals extend a few au further.}. The TNO population can be split into two broad dynamical subgroups: the `cold classicals', on near-circular and low-inclination \emph{i} $\lesssim$ 5$^\circ$ orbits, and an overlapping dynamically excited population with \emph{i} $\gtrsim$ 5$^\circ$ \citep{Brown2001, Bernstein2004, Gladman2008, Fuentes2008,Fraser2010,Petit2011,Bannister2018}.
The cold classicals have very little dynamical excitation, with an inclination width of only $\simeq2^{\circ}$ \citep{Brown2001,2008ssbn.book...59K,Gulbis2010, Petit2011}.
In contrast, the dynamically excited Kuiper belt has an inclination width of $\sim$14-16$^{\circ}$ \citep{Petit2017}.
The dynamically excited population is a highly complex structure with several subclasses.
A fifth of the population is locked in mean motion resonances with Neptune \citep{Malhotra1995,Gladman2012}: their orbits have integer period ratios with Neptune's orbit.
Non-resonant TNOs with large eccentricities form the scattering disk, a slowly decaying population.
Their semimajor axes evolve due to active gravitational scattering with Neptune \citep{Gladman2008,Gomes2008}.
The scattering disk feeds into the short-lived Centaur population, which orbits between the giant planets for timescales of tens of Myr \citep{1999Icar..142..509D,2003AJ....126.3122T}, although it may not be the only population contributing to the Centaurs \citep{1999AJ....118.1873Y,2010A&A...519A.112D,2010MNRAS.402...13H}.
In contrast, TNOs on $q \gtrsim 38$~au orbits with large $a$ beyond the 2:1 mean motion resonance reside on moderately stable orbits with little or no direct gravitational interactions with Neptune \citep{Emelyanenko2003, Gomes2005,Gladman2008, Brasser2015}.
Orbits with $a>250$~au and $q \gtrsim 45$~au are difficult if not impossible for Neptune to directly scatter planetesimals onto during its outward migration \citep{Brasser2015}; these orbits have unclear origins, and may be emplaced by a different dynamical mechanism \citep{Brown2004, Morbidelli2004, 2006ApJ...643L.135G, 2006Icar..184...59B,2008Icar..197..221K, 2012Icar..217....1B,Trujillo2014,Batygin2016,Sheppard2016,Bannister2017diffusion,2018ApJ...863...45P}.
\subsection{Migration History and Population Emplacement}
\label{sec:history}
Overwhelming evidence indicates that the Kuiper belt did not form entirely in situ. Planetesimal-driven giant planet migration which scatters the early Solar System's protoplanetesimal disk is the overarching framework \citep{Malhotra1995,Thommes1999,Levison2008,Nesvorny2015b,Nesvorny2015} that best reproduces the observed orbital structure of the trans-Neptunian region.
The specifics of the dynamical instability result in different signatures in the Kuiper belt's present orbital distribution.
The dynamically excited population, including most of the resonant objects, were emplaced into the Kuiper belt through gravitational scattering from inward regions of the planetesimal disk.
Subsequent interactions with Neptune further sculpted the region and are imprinted on the dynamically excited population's present orbital distribution.
The physical properties of the cold classicals -- their photometric colors, slope of their size distribution, albedo distribution, and resolved binary fraction -- significantly differ from the rest of the Kuiper belt \citep[e.g.][]{Tegler2000, Doressoundiram2002, Peixinho2004, Noll2008, Peixinho2008, 2009Icar..201..284B,Fraser2014,Lacerda2014}.
Additionally, the widest cold classical binaries would be disrupted if scattered out into the classical belt by Neptune \citep{Parker2010}.
Together, this evidence suggests that the vast majority of the cold classicals formed in place. Thus, the cold classicals place unique constraints on Neptune's dynamical history, as their orbits must remain largely undisturbed by Neptune's migration as the planet reached its present-day orbit.
Recent work by \cite{Fraser2017} using Col-OSSOS measurements shows that this picture is slightly more complicated. Red colored cold classicals are thought to have originated at their present location, but ``blue binaries" (neutral-colored binaries) found within the cold classical belt are thought to be interlopers that formed further inward and were deposited into the region during the final stages of Neptune migration.
Recent studies have shown that the smoothness of Neptune's migration at late stages of planetesimal-driven migration produces differences in the predicted structure of Kuiper belt orbits.
\citet{Nesvorny2015b,Nesvorny2015} showed that Neptune migration with a slow and occasionally jumping ``grainy" motion through a cold disk of planetesimals implants objects on sufficiently excited inclinations and eccentricities in the dynamically excited population.
\citet{Lawler2018sub} find the best match to the near-resonant distant populations is from grainy migration. Work by \cite{2011ApJ...738...13B}, \cite{2018arXiv180802146R}, and \cite{2018Icar..306..319G} conclude that a moderately high eccentric phase during Neptune migration is also a viable scenario to reproduce the structure of the cold classical belt.
\subsection{Compositional Surveys}
\label{sec:surfaces}
The past two decades of observations have provided substantial insights into the surface composition of the bright $m_r < 22$ mag TNOs, which are readily studied via optical and NIR reflectance spectroscopy.
Their surfaces are divided into three categories: dwarf planets rich in volatile ices such as methane, ethane and water ice \citep{Schaller2007, Barkume2008, Brown2012}, the Haumea collisional family with strong water ice absorption \citep{Brown2007, Schaller2008, Snodgrass2010, Trujillo2011, Carry2012, Fraser2012}, and surfaces devoid of feature-imprinting volatiles other than water ice.
The majority of spectroscopically studied TNOs are spectrally featureless.
They exhibit a diversity of surfaces, with spectral gradients ranging from solar-neutral colors to redder than solar in optical wavelengths \citep{2008A&A...487..741A,Barucci2008,2008AJ....135...55B,2011Icar..214..297B}.
For the far more abundant $>$ 22 mag TNOs, optical and infrared spectroscopy is impossible with current ground and space-based facilities.
We must instead rely on what broad-band and narrow-band colors reveal by proxy about the optical and near-infrared spectral slopes.
Large surveys of TNO surface colors have used a variety of optical and infrared wavelength measurements to attempt to understand and classify TNO surfaces.
\citet{Doressoundiram2008, Peixinho2015}, and the MBOSS database\footnote{\url{http://www.eso.org/~ohainaut/MBOSS/}} \citep{Hainaut2012} provide a compilation of surface colors in the published literature.
The first surveys found that TNO surfaces have a broad range of surface colors, from nearly neutral solar colors to very red \citep{Luu1996}.
Some dynamical populations are confined to a color range, or exhibit a bimodality in color.
\cite{2000Natur.407..979T} and \cite{Doressoundiram2001} identified that more neutral surfaces exist at higher orbital inclinations; this trend was later identified as largely an effect of the vast majority of cold classical TNOs having red surfaces \citep{Doressoundiram2001, Tegler2003}.
Additional work confirmed the statistical significance of the cold classicals' color distribution as separate from that of other TNOs \citep{Doressoundiram2001, Doressoundiram2002,Tegler2003, Doressoundiram2007,Peixinho2008}.
The European Southern Observatory Large Program on Centaurs and TNOs found a continuum of surface colors in the optical and infrared, with relatively linear color slopes from $B$ to $J$ bands \citep{Boenhardt2002, Delsanti2004, Peixinho2004, Delsanti2006}.
Other photometric surveys identified a bimodality in surface colors of the Centaur population \citep[][]{Peixinho2003,Tegler2003I}.
As the $r > 22$ mag TNOs were surveyed, a bimodality in color became apparent in smaller-size TNOs \citep{Peixinho2012,Fraser2012,Peixinho2015,Fraser2015}, with a bimodality in $g-i$ in smaller excited TNOs \citep{Wong2017}.
\subsection{Surface Origins in the Protoplanetesimal Disk}
\label{sec:motivations}
Two distinct models have been put forth to explain the observed diversity of surfaces observed within the small ($<$ 500 km in diameter) TNOs. The two proposals suggest different scenarios for how the surface color variation was emplaced in the TNO population before Neptune migration occurred. Each scenario has its own implications for the structure of the early Solar System's planetesimal disk. Col-OSSOS aims to distinguish between these two ideas.
In the first model, summarized by \cite{DalleOre2013}, small TNO surfaces are divided into five discrete types, each with its own unique color, albedo, and composition.
The cold classicals fall into their own separate class, while the dynamically excited TNO population is divided into four types. In this scenario, the unique color classes are the direct result of $\sim$5 or more ice lines that existed in the primordial disk: each unique composition corresponds to formation beyond a specific ice line. In the \cite{DalleOre2013} model, these compositional classes were caused by a sharply striped primordial disk with five or more compositional boundaries and very little or no cross-mixing between the forming protoplanetesimals. In this scenario, individual compositional classes manifest as groups of objects clustered in different regions of the optical-NIR color space; at least 5 should be apparent with sufficiently accurate color measurements.
In the second model, of \citet{Fraser2012}, small TNOs fall into only three compositional classes, which also correlate with dynamical class: the red cold classicals, the neutral dynamically excited objects, and red dynamically excited objects. \cite{Fraser2012} suggest that the protoplanetesimal disk did not show a primordial compositional gradient, but rather, was compositionally homogenous between 15 and 45 au. All planetesimals in this region would have had relatively similar abundances of surface volatiles after formation. The observed compositional classes were thus a result of post-formation evolution, with some objects rapidly losing their light volatile species. What volatile was lost depended only on surface temperature and hence formation location, with each class of TNO predominantly residing at a different distance within the protoplanetesimal disk \citep{Wong2016,Wong2017}. The separate long-term chemical evolution pathways resulted in different surface colors dividing into 3 broad classes (the red cold classicals, the neutral dynamically excited objects, and red dynamically excited objects). In the \cite{Fraser2012} scenario, the range of surface colors seen in each class is the result of a range of mixing of unique surface materials. Only two taxons will be apparent in the optical and NIR space, each exhibiting a range of optical-NIR colors; with the cold classical objects sharing a taxon with the dynamically excited red objects.
A sample of sufficiently accurate optical and NIR TNO colors should be able to differentiate between the scenarios of \cite{DalleOre2013} and \cite{Fraser2012} and inform us of the compositional properties of the early planetesimal disk. Combining the dynamics of the ensemble Kuiper belt together with its physical and chemical properties would create a powerful probe of Neptune's migration and of the compositional structure of the primordial disk from which the TNOs originated.
To date, this has proven a challenging task.
Most TNO physical property studies examine the hodgepodge set of objects that were discovered by various surveys with different and varying detection biases.
Object size and dynamical classification are dependent properties: small TNOs become discoverable when they are near perihelion, at closer heliocentric distances.
For example, Centaurs with color measurements are on average much smaller in size than the hot classical objects with measured colors.
This has made it difficult, if not impossible, to accurately estimate the true frequency of the different surface color groups in the modern-day Kuiper belt.
Thus, a careful sample of TNOs with known discovery biases is necessary to disentangle the effects of observational biases from the color distribution of the intrinsic populations \citep{pike2013}.
\section{Survey Design}
\label{sec:design}
The goal of Col-OSSOS is to produce the first ever flux-limited optical and near-infrared color survey of TNOs, with well-characterized and well-quantified biases.
We aim to use this sample to test and probe the taxonomic classes within the Kuiper belt and to produce reliable intrinsic population statistics.
The Col-OSSOS program is governed by a set of overarching requirements:
\begin{itemize}[itemsep=-3pt]
\item a well understood, flux-limited TNO sample with quantified biases (Section \ref{sec:targets})
\item color measurements that can distinguish between the disk models of \citet{DalleOre2013} and \citet{Fraser2012} (Section \ref{sec:filters})
\item homogeneity in our observing scheme (Section \ref{sec:observing_strategy})
\item observations in different filters acquired as temporally close as possible (Section \ref{sec:observing_strategy})
\item consistent high photometric quality of the observations for all targets (Section \ref{sec:observing_strategy}).
\end{itemize}
\subsection{Target TNO Selection}
\label{sec:targets}
All Col-OSSOS targets are drawn from the sample of over 800 TNOs and Centaurs ($m_r = 21.8$-25.2 mag; $3.6 < H_r < 14.5$ mag) found by the Outer Solar System Origins Survey \citep[OSSOS:][]{Bannister2016,Bannister2018}, a survey with well-characterized survey biases.
OSSOS was a wide-field $r$-band survey with the 3.58-m Canada-France-Hawaii Telescope's (CFHT) MegaPrime square-degree field-of-view (FOV) imager (MegaCam) \citep{Boulade2003}.
Operating from 2013-2017, OSSOS searched eight $20$ deg$^2$ regions (blocks) of sky \citep{Bannister2018}, at locations where resonant TNOs come to perihelion and become optimally detectable.
Strong emphasis was placed on thorough recovery and tracking of discoveries in a dense observing cadence across two years, freeing the discovered TNOs from the challenges of ephemeris bias \citep{Jones2006}. 97\% of the OSSOS TNO sample have fractional semimajor axis uncertainty of $\sigma_a < 0.1$\% \citep{Bannister2016,Bannister2018}.
The biases that surveying imposed on the TNO sample are well-quantified for OSSOS discoveries
\citep[see][]{Bannister2016,Lawler2018}.
The high-quality orbit dynamics and population statistics of OSSOS form a framework on which to overlay the color information provided by Col-OSSOS. The Col-OSSOS sample consists of the 96 TNOs brighter than or equal to $m_r$ = 23.6 in the 13AE, 13AO, 13BL, 14BH, and 15BS OSSOS blocks
Col-OSSOS photometry for the 35 TNOs from 13BL and 14BH blocks are reported in this paper (see Section \ref{sec:sample}).
\subsection{Filter Selection and Color Precision}
\label{sec:filters}
The compositional classes described in Section \ref{sec:motivations} are identified by clusters in the optical and NIR reflectance colors exhibited by TNOs.
The predominantly linear spectra of small, spectrally featureless TNOs in optical wavelengths are fully characterized by their $g-r$ color (see \citealt{Doressoundiram2007}).
The neutral/red bifurcation of the dynamically excited TNOs exists only in the optical, and is most prominent in $(g-r)$ \citep{Doressoundiram2008,Fraser2012}.
With observations at wavelengths longer than $I$, additional surface classes become apparent in other small-body populations \citep{DeMeo2009, Emery2011}.
This is shown indirectly for TNOs by \citet{Doressoundiram2008}, who present $BVRI$ photometry of $\sim$100 sources, with only two surface types apparent.
The transition from the optical to near-infrared spectral gradients occurs at $\sim$1 $\mu$m; thus, only filters at longer wavelengths can provide the necessary slope information.
It should be noted that \citet{DalleOre2013} present no practical difference between the $J$, $H$, and $K$ bands in terms of identifying their proposed compositional classifications.
In order to distinguish between the compositional classes proposed by \citet{DalleOre2013} and \citet{Fraser2012}, we selected two optical broad-band filters, $g$ and $r$, and a near-infrared filter, $J$, for our observations.
These were the fewest filters that could define the optical and near-infrared slope of each TNO in the wavelength region of interest; $(g-r)$ characterizes the optical slope, and $(r-J)$ characterizes the near-infrared slope. Figure \ref{fig:pastdata} shows the color precision of all available optical+J-band color measurements as of the start of our survey \citep{Hainaut2012,Peixinho2015}, with appropriate conversions to $grJ$ \citep{Jester2005,Jordi2006}. These measurements are mean values, which do not necessarily account for the rotational variability of each TNO. These data demonstrate the importance of precision and temporal near-simultaneity in color measurements: their precision is insufficient to distinguish the color classes discussed in Section \ref{sec:motivations}.
Those classes have optical and NIR colors that differ by as little as 0.04-0.06 magnitudes from class to class.
The photometric uncertainty in previously published color studies ranges for 0.04 to 0.2 mag, with no measurements in the size/H range that Col-OSSOS aimed to sample \citep{Hainaut2012,Peixinho2015}. To achieve 0.06 mag or better photometric precision for the Col-OSSOS sample required the collecting area of an 8-10-m class telescope and the non-standard observing and analysis techniques that we describe in Sections \ref{sec:observing_strategy}, \ref{sec:colours} and Appendixes \ref{sec:optical} and \ref{sec:NIR}.
\begin{figure}[htbp]
\includegraphics[width=\textwidth]{pastdata.pdf}
\caption{\label{fig:pastdata} Non-simultaneous mean observed (converted) $grJ$ color-space measurements of trans-Neptunian populations existing at the conception of the Col-OSSOS survey. Plotted are derived colors for sources with optical and J-band photometry, as indexed by MBOSS \citep{Hainaut2012} or reported by \cite{Peixinho2015}. $(g-r)$ was estimated from the reported $(B-V)$ and the conversion of \citet{Jester2005}. $(r-J)$ was estimated by first estimating $(r-i)$ from the reported $(R-I)$ using the conversion reported by \citet{Jester2005} and then estimating $(R-r)$ using the conversion reported by \citet{Jordi2006}. $(R-r)$ is used to estimate $(r-J)$ as $(r-J)$ = $(R-J)$ - $(R-r)$. The Solar color, with g - r = 0.45 and r - J = 0.97, is shown by the yellow star. The dashed curve indicates the reddening line, a line of constant spectral slope through the $grJ$ spectral range, calculated using the pysynphot software package \citep{Lim+2015}. A broad trend of redness relative to the Sun exists, but detail within the population cannot be discerned.}
\end{figure}
As we observe in a filter system that is close to widely used bandpasses, but has subtle distinctions worth accounting for at our required level of photometric precision, throughout this paper we use the following nomenclature for our filters and corresponding colors:
\begin{description}[labelwidth=1cm, leftmargin=1cm, before={\renewcommand\makelabel[1]{\bfseries ##1}}, itemsep=-1pt]
\item[$g, r, z$] optical bandpasses, in contexts where the specific filter system does not need to be distinguished
\item[$J$] Maunakea Observatory (MKO) filter set $J$ band \citep{2002PASP..114..169S}, $\lambda$=12500 \AA, 11500-13300 \AA ~coverage
\end{description}
Where necessary for specific observations and color conversions, we specify the exact bandpass in $g$, $r$, or $z$ with appropriate subscripts, as shown here for $r$:
\begin{description}[labelwidth=1cm, leftmargin=1cm, before={\renewcommand\makelabel[1]{\bfseries ##1}}, itemsep=-1pt]
\item[$r_S$] Sloan Digital Sky Survey (SDSS) Photometric System \citep{Fukugita1996, Padmanabhan2008}
\item[$r_G$] Gemini Multi-Object Spectrograph (GMOS) North filters\footnote{See \url{http://www.gemini.edu/node/10420}}. Either acquired with the E2V or the Hamamatsu detectors (specified as needed; see Section \ref{sec:observing_strategy})
\item[$r_{PS}$] Pan-STARRS1 photometric system \citep{Tonry2012}
\end{description}
\subsection{Observing Strategy}
\label{sec:observing_strategy}
Our primary observing facility is the 8.1-m Frederick C. Gillett Gemini North Telescope located on Maunakea, Hawai`i.
In the optical, each target was observed with the imaging mode ($5.5\arcmin \times 5.5 \arcmin$ FOV) of the Gemini Multi-Object Spectrograph \citep[GMOS;][]{Hook2004} using the r$\_$G0303 ($\lambda$=6300 \AA, $\delta\lambda$=1360 \AA) and g$\_$G0301($\lambda$=4750 \AA, $\delta\lambda$=1540 \AA) filters, similar to the SDSS $r$ and $g$ bands.
The GMOS observations were obtained in 1$\times$1 binning mode. This had $0.0747 \arcsec$ pixels with the e2v deep depletion charge-coupled devices (CCDs) available in GMOS during the 2014-2016 observing semesters, and $0.0807 \arcsec$ pixels after the 2017 installation of Hamamatsu red-sensitive CCDs. GMOS observations were dithered by $3\arcsec$ to $5\arcsec$ from exposure to exposure. For the near-infrared measurements, we observed
with the Near-Infrared Imager \citep[NIRI;][]{Hodapp2003} in Maunakea Observatory $J$. NIRI observations were acquired using the f/6 camera ($0.116 \arcsec$ per pixel resolution) with a $119.9 \arcsec \times 119.9 \arcsec$ FOV. NIRI observations utilized a grid dither pattern with $8 \arcsec$ spacing between exposures to ensure accurate background measurement and removal.
For both NIRI and GMOS, these instrument configurations allow the best possible characterization of the point spread function (PSF), and hence, the best knowledge of the photometric curve of growth.
For our color precision requirement to detect and distinguish the \citet{DalleOre2013} classes, we aim for an 0.06 mag color precision in $(g-r)$ and $(r-J)$ for all targets. To achieve our desired color precision (Section \ref{sec:filters}), we require a signal-to-noise (SNR) $\geq 25$ in $g$ and $r$, and SNR $\geq 20$ in $J$. For each Col-OSSOS TNO, the total effective exposure time required in each filter was estimated using the mean OSSOS r-band discovery magnitude and assuming very red optical and only modestly red near-infrared colors relative to those typically exhibited by small TNOs (Figure \ref{fig:pastdata}): $g-r =1.1$, and $r-J=1.2$.
Ideally, most TNOs will have bluer optical colors and redder near-infrared colors, and their observations will thus meet our SNR requirements regardless of their true color.
At the time of the Gemini observations no light curves were yet measured for these TNOs.
If the 15-20 OSSOS CFHT observations of the TNO over 1-2 years showed significant $\pm 0.3$ mag photometric variability, additional frames in all filters were added to the Gemini observing sequence, in case observing happened during minimum TNO brightness.
We used Gemini's fast instrument switching abilities ($\lesssim4$~minutes) to provide near-simultaneous colors: all targets were visited in an \emph{unbroken} imaging sequence of $rgJgr$.
Anchoring the sequence with $r$ exposures lets us account for light curve variations during color estimates (discussed further in Section \ref{sec:colours}). Observations were executed to gather the desired cumulative SNR by taking half the required optical frames before and then after the required $J$-band observations. Individual GMOS exposures were set to 300s in duration.
Individual NIRI exposures were limited to 120s, both to minimize trailing losses and to mitigate the high sky background.
The telescope tracked at the sidereal rate, permitting use of calibration stars within the images (see Appendixes \ref{sec:optical} and \ref{sec:NIR}). Source trailing in each observation was minimal, as Col-OSSOS targets typically have on-sky motions of several arcseconds per hour or less. The timing of the observations was chosen to avoid the TNO passing over or close to bright or contaminating background stars and galaxies. A combination of SDSS observations \citep{York2000}, stacked OSSOS images \citep{Bannister2018}, and stacked Pan-STARRS1 \citep{2016arXiv161205560C,2016arXiv161205243F} images were used to identify times when the TNOs were moving through areas of sky devoid of background sources brighter than $\sim$24th magnitude. The majority of the GMOS observations were made at an airmass $< $2, and most of the NIRI photometry was obtained at an airmass $<$1.4. With Gemini's queue scheduling and priority visitor mode, all observations were made in photometric conditions during dark time, in Gemini's 50 percentile sky background SB 50) criteria.
The images were executed in Gemini's 70 percentile image quality (IQ 70) or better, achieving image quality typically 0.7$^{\prime\prime}$ or better in the optical and 0.5$^{\prime\prime}$ or better in the NIR for the majority of observations. Before and after each $rgJgr$ sequence, NIRI photometric calibrator frames were acquired with bright standard stars at different elevations, chosen to encompass the range of elevations spanned during the $J$ observations of each TNO.
Exposure times for the calibration stars were chosen so as to not saturate the NIRI detector, and exposures were taken in a 9-point dither pattern sequence. Associated NIRI lamp flats, NIRI dark frames, and GMOS bias observations were also obtained for each night of observing.
\subsubsection{$z$-band Imaging}
A subsample of the Col-OSSOS targets were also imaged with Gemini in the $z$\_$G0304$ filter (8500-10000 \AA~coverage).
The inclusion of $z$ observations was ad hoc and with no pre-defined minimum SNR requirement.
During exceptional sky conditions (IQ 20, c.f. $\sim 0.4\arcsec$ seeing), we reduced the number of $g$ and $r$ frames, maintaining SNR 25, and added GMOS $z_G$-band observations.
They bracketed the start and end of the optical sequences, with the same 300s exposure time.
We describe the $z_G$ data processing in Appendix \ref{sec:optical} and their implications in \citet{Pike2017}.
Additional simultaneous $z$ images were acquired with Suprime-Cam \citep{Miyazaki2002} on the Subaru Telescope in August 2014. These observations are reported in detail in \citet{Pike2017}.
\section{TNO Color Technique}
\label{sec:colours}
Detailed overviews of the optical and NIR data reduction, photometry, and calibration we performed on Col-OSSOS observations are described in Appendixes \ref{sec:optical} and \ref{sec:NIR}. Photometry was performed on each individual GMOS frame. For the NIRI observations, the sequences of images were divided and combined into two stacked images; photometry was performed on each. Our measurements were made with TRIPPy (Trailed Image Photometry in Python), a dedicated software for photometry of linearly trailed sources \citep{Fraser2016}. TRIPPy makes use of a pill-shaped aperture, an aperture elongated based on a Solar System object's predicted rate of motion. For trailed Solar System sources, this process correctly accounts for the flux that would be lost in making use of circular apertures, while maintaining the photometric precision found with use of small area apertures. For PSFs derived from sidereal tracked stars, aperture corrections can be determined to better than 0.01 mag for the pill aperture \citep{Fraser2016}. Our optical and NIR measurements were calibrated to SDSS \citep{York2000,Padmanabhan2008} filter system and Maunakea Observatory (MKO) filter set $J$ band \citep{2002PASP..114..169S} respectively. The estimation of the color transform between the SDSS and Gemini filter sets is described in Appendix \ref{sec:colorterm}.
The final step to estimating colors is to account for brightness variations of the TNO over the duration of the science sequence.
Any significant amplitude changes due to rotational variability of the TNO can cause difficulty combining measured broad-band colors if not accounted for.
Our full GMOS-NIRI-GMOS sequences span between 1 and 6 hours.
Small TNOs ($H > 5$) typically have 6- to 15-hour rotation periods, with peak-to-peak variations of $\sim$0.3 magnitudes \citep{2006AJ....131.1149T,Duffard2009,Benecchi2013}, though significantly larger variations have been observed \citep{Fraser2015,Thirouin2018}, including in the OSSOS sample \citep{Alexandersen2018}.\footnote{\cite{Alexandersen2018} examine the photometric variability for four targets in our first release sample: 2013 UL15, 2013 UP15, 2013 UM15, 2013 UN15.}
We took advantage of each sequence's bracketing $g$ and $r$ imagery to correct for light curves by fitting a linear model to all optical data. An example of a fitted light curve for a target that exhibited a variation in brightness over the Col-OSSOS observing sequence is presented in Figure~\ref{fig:light curve}. This model makes two assumptions: that the light curve variations observed across a given sequence are linear in nature, and that the object exhibits no significant color variations across the sequence.
Thus, the model has as free parameters: slope (change in brightness with time), a reference $r$ magnitude, a color $g-r$, and sometimes a color $r-z$, where $z$ images were available. This model was fit in a least-squares sense to the available data. To evaluate uncertainties on each optical color, we adopt a Monte-Carlo approach. Specifically, each individual photometric measurement was scattered by a Gaussian distribution with width equal to the photometric uncertainty on that point, and the scattered dataset was fit. This process was repeated 200 times, and the the quadrature sum of the standard deviation on the randomized color terms, and the uncertainty on the mean of the individual photometric measurements were adopted as the uncertainty on those points. The fitted linear light curve was then used to estimate the $r$ brightness at the midpoint of each $J$-band measurement so as to determine the mean $r-J$ color. Uncertainties in this color include the photometric uncertainties on $J$, and the uncertainties due to the light curve model parameters (both slope and reference $r$ value).
\begin{figure}
\includegraphics[width=\columnwidth]{o4h19.pdf}
\caption{\label{fig:light curve} {Example of observed photometry, and fitted light curve, and colors for object 2014 UK225. Observed $r$, $g$, $z$, and $J$ photometry are shown by circles. Those points adjusted to $r$-band based on the fitted colors are plotted as triangles. The target exhibits an $\sim$0.06 magnitude increase in brightness over the duration of the sequence. This change in brightness is comparable to the precision in measured colors of the source, and is accounted for by our light curve fitting technique.}}
\end{figure}
We note that the linear fitting process for intrinsic variability will only correctly remove any first derivatives in the light curve. It will not account for the full range of possible photometric variation. We are not able to correct for, nor are we able to meaningfully estimate the additional uncertainty in, our color estimates caused by these unknown non-linear light curve variations. There are undoubtedly shape and albedo effects that impact our color estimates which are not accounted for in the linear fit. However, With the small number of photometric measurements we have for each target, we have no resolution to further constrain these effects. We discuss the impact of non-linear light curve variations on our results in Section \ref{sec:lc}
\section{Col-OSSOS First Release: The OSSOS 13BL and 14BH Sample}
\label{sec:sample}
The observations we present here comprise a complete flux-limited sample of TNOs: all the $m_r < 23.6$ mag discoveries from the 13BL and 14BH OSSOS survey blocks \citep{Bannister2016,Bannister2018}. In this section, we also present the orbital properties of our TNO release sample and summarize specific observing circumstances and data analysis unique to these targets compared to the overall Col-OSSOS survey strategy.
\subsection{Observational and Orbital Properties of the TNO Sample from 13BL and 14BH Blocks}
The 35 TNOs in this first Col-OSSOS release were found in 2013 and 2014 in two OSSOS survey regions of sky near the ecliptic, in a latitude range from the invariant plane up to $5\degr$ off-plane.
They comprise the 18 $m_r < 23.6$ mag discoveries from the 13BL block of OSSOS (\textit{o3l}-designated targets), and the 17 $m_r < 23.6$ mag TNOs from the 14BH block of OSSOS (\textit{o4h}-designated targets). 13BL block is a 20 deg$^2$ region overlaying the invariant plane, centered at R.A. $0^{h}54^{m}$, decl. $+3^{\circ}50$'.
14BH block is a 21 deg$^2$ region $2-5^{\circ}$ off the invariant plane, centered at R.A. $1^{h}35^{m}$, decl. +$13^{\circ}28$'.
The detection efficiency of OSSOS for moving objects in each survey region is thoroughly characterized \citep{Bannister2018}.
Table \ref{tab:sample} lists the orbital information for the 35 TNOs. The orbital distribution is shown in Figure \ref{fig:sample}. We tally the dynamical classes in Table \ref{tab:orbit_numbers}. The TNOs in this first Col-OSSOS release predominantly sample the classical Kuiper belt and the resonant populations. The barycentric orbital properties of the 35 TNOs are derived from 2-5 years of densely sampled observation \citep{Bannister2018}. A full search for binarity of our sample is beyond the scope of this work, but \cite{Fraser2017}, identified three cold classicals in our sample as binaries: 2016 BP81, 2014 UD225, and 2013 SQ99. One hot classical Kuiper belt object, 2013 UQ15, has a semimajor axis, inclination, and eccentricity consistent with being in the dynamical cloud of the Haumea collisional family \citep{Brown2007, Ragozzine2007}.
5 of the 14 resonant TNOs are in the 3:2 mean motion resonance with Neptune, with the rest in the 5:3, 4:3, 7:4, 5:2, 11:6 and 9:5 resonances.
One TNO, 2007 TC434, is securely in the most distant resonance with Neptune yet confirmed, the 9:1 at $a \sim 130$~au \citep{Volk2018}.
There are also 3 objects from the more transient Centaur and scattering populations. The majority of our TNO sample are new discoveries by OSSOS. No targets in our release sample have previous near-simultaneous multi-filter photometry to the precision that we discuss here. Of the few targets with observations earlier than when OSSOS began in 2013, none have published colors \citep[][Kavelaars private communication]{Hainaut2012}.
All of our 35 targets have absolute magnitudes $H_r > 5$ and are thus much smaller than the dwarf planet size transition of $H\simeq4$ \citep{Brown2008,Tancredi2008}.
Most have $5 < H_r < 8$, brighter than the break in the measured TNO luminosity function at $H_r \sim 8$ \citep{Bernstein2004, Fuentes2008, Fraser2009, Fraser2014,2014AJ....148...55A}.
Only two have $H_r > 10$.
\begin{figure}
\includegraphics[width=0.91\columnwidth]{colossos_aei.pdf}
\caption{\label{fig:sample} Barycentric orbital parameters, derived from \citet{Bannister2018}, of TNOs with Col-OSSOS color measurements presented in this paper. One TNO at $a=130$~au is omitted for better resolution. The 1-$\sigma$ uncertainties are smaller than the size of the plot symbol.}
\end{figure}
\begin{deluxetable*}{llrrrrrrccclll}
\tabletypesize{\scriptsize}
\tablecaption{\label{tab:sample} Orbital parameters and optical and near-infrared colors of the o3l and o4h Col-OSSOS TNO sample}
\tablehead{
\colhead{MPC} & \colhead{OSSOS} & \colhead{a} &\colhead{e} & \colhead{ inc} & \colhead{$\Delta$} & \colhead{r$_{helio}$} & \colhead{Mean $m_r$} & \colhead{$H{_r}$ } & \colhead{Orbit} & \colhead{$g-r$} & \colhead{$r-J$} & \colhead{$r-z$} \\
\colhead{ID} & \colhead{ID} & \colhead{(au)} & \colhead{} & \colhead{ ($\degr$)} & \colhead{(au)} & \colhead{(au)} & \colhead{(SDSS)} & \colhead{} & \colhead{Class} & \colhead{(SDSS)} & \colhead{(SDSS-MKO)} & \colhead{(SDSS)} }
\startdata
2013 UR15$^H$ & o3l01 & 55.82 & 0.719 & 22.25 & 15.69 & 16.31 & 23.40 $\pm$ 0.17 & 11.36 & sca & 0.67 $\pm$ 0.02 & 1.49 $\pm$ 0.08 & -- $\pm$ -- \\
2001 QF331$^{F,H}$ & o3l06PD & 42.25 & 0.252 & 2.67 & 31.88 & 32.60 & 22.92 $\pm$ 0.03 & 7.84 & 5:3 & 0.88 $\pm$ 0.03 & 1.67 $\pm$ 0.05 & -- $\pm$ -- \\
2013 US15$^H$ & o3l09 & 36.38 & 0.070 & 2.02 & 33.66 & 34.40 & 23.17 $\pm$ 0.02 & 7.85 & 4:3 & 1.03 $\pm$ 0.02 & 1.49 $\pm$ 0.05 & -- $\pm$ -- \\
2003 SR317 & o3l13PD & 39.43 & 0.166 & 8.35 & 36.27 & 37.23 & 23.42 $\pm$ 0.08 & 7.77 & 3:2 & 0.64 $\pm$ 0.01 & 1.33 $\pm$ 0.05 & -- $\pm$ -- \\
2013 SZ99$^H$ & o3l15 & 38.28 & 0.017 & 19.84 & 37.96 & 38.75 & 23.80 $\pm$ 0.06 & 7.96 & cla & 0.68 $\pm$ 0.02 & 1.35 $\pm$ 0.07 & -- $\pm$ -- \\
2010 RE188 & o3l18 & 46.01 & 0.147 & 6.75 & 38.90 & 39.64 & 22.34 $\pm$ 0.02 & 6.40 & cla & 0.58 $\pm$ 0.02 & 1.43 $\pm$ 0.06 & -- $\pm$ -- \\
2013 SP99$^G$ & o3l32 & 43.78 & 0.060 & 0.79 & 41.09 & 42.01 & 23.53 $\pm$ 0.05 & 7.35 & cla & 1.00 $\pm$ 0.02 & 1.61 $\pm$ 0.05 & -- $\pm$ -- \\
2016 BP81$^{B,G,H}$ & o3l39 & 43.68 & 0.076 & 4.18 & 41.81 & 42.54 & 22.83 $\pm$ 0.11 & 6.58 & 7:4I & 0.59 $\pm$ 0.03 & 1.60 $\pm$ 0.07 & -- $\pm$ -- \\
2013 UL15$^{F,G,H}$ & o3l43 & 45.79 & 0.097 & 2.02 & 42.38 & 43.10 & 23.12 $\pm$ 0.10 & 6.82 & cla & 0.90 $\pm$ 0.04 & 1.51 $\pm$ 0.06 & -- $\pm$ -- \\
2013 UP15$^{G,H}$ & o3l46 & 46.61 & 0.079 & 2.47 & 42.70 & 43.38 & 23.92 $\pm$ 0.10 & 7.58 & cla & 0.90 $\pm$ 0.02 & 1.78 $\pm$ 0.07 & -- $\pm$ -- \\
2013 UO15$^G$ & o3l50 & 43.33 & 0.049 & 3.73 & 43.08 & 43.99 & 23.28 $\pm$ 0.00 & 6.89 & cla & 0.96 $\pm$ 0.02 & 1.75 $\pm$ 0.04 & -- $\pm$ -- \\
2013 UM15$^{F,G}$ & o3l57 & 45.04 & 0.075 & 1.84 & 43.77 & 44.46 & 23.39 $\pm$ 0.00 & 6.95 & 11:6 & 1.08 $\pm$ 0.01 & 1.56 $\pm$ 0.06 & -- $\pm$ -- \\
2006 QF181$^G$ & o3l60 & 44.82 & 0.076 & 2.66 & 44.20 & 44.54 & 23.56 $\pm$ 0.02 & 7.08 & cla & 0.89 $\pm$ 0.03 & 1.65 $\pm$ 0.05 & -- $\pm$ -- \\
2013 UN15$^{D,G,H}$ & o3l63 & 45.13 & 0.054 & 3.36 & 44.39 & 45.14 & 24.13 $\pm$ 0.09 & 7.62 & cla & 1.04 $\pm$ 0.03 & 1.57 $\pm$ 0.08 & 0.37 $\pm$ 0.09 \\
2013 UN15$^{C,E,G,H}$ & o3l63 & 45.13 & 0.054 & 3.36 & 44.20 & 45.19 & 23.62 $\pm$ 0.00 & 7.11 & cla & 1.08 $\pm$ 0.03 &1.82 $\pm$ 0.04 & 0.73 $\pm$ 0.06 \\
2013 UX18 & o3l69 & 43.60 & 0.057 & 2.89 & 44.78 & 45.73 & 23.93 $\pm$ 0.00 & 7.37 & cla & 0.93 $\pm$ 0.01 & 1.64 $\pm$ 0.08 & -- $\pm$ -- \\
2013 SQ99$^{B,G,H}$ & o3l76 & 44.15 & 0.093 & 3.47 & 46.60 & 47.34 & 23.17 $\pm$ 0.04 & 6.45 & cla & 0.98 $\pm$ 0.03 & 1.51 $\pm$ 0.06 & 0.56 $\pm$ 0.03 \\
2013 UQ15$^{A,F,H}$ & o3l77 & 42.77 & 0.113 & 27.34 & 46.84 & 47.54 & 23.02 $\pm$ 0.22 & 6.28 & cla & 0.53 $\pm$ 0.04 & 0.92 $\pm$ 0.11 & -- $\pm$ -- \\
2013 SA100$^{D,H}$ & o3l79 & 46.30 & 0.166 & 8.48 & 49.69 & 50.43 & 23.02 $\pm$ 0.03 & 6.03 & cla & 0.63 $\pm$ 0.02 & 1.66 $\pm$ 0.08 & 0.44 $\pm$ 0.02 \\
2013 SA100$^{E,H}$ & o3l79 & 46.30 & 0.166 & 8.48 & 49.30 & 50.29 & 22.78 $\pm$ 0.03 & 5.81 & cla & 0.67 $\pm$ 0.02 & 1.48 $\pm$ 0.04 & 0.42 $\pm$ 0.02 \\
2014 UJ225$^H$ & o4h01 & 23.20 & 0.378 & 21.32 & 17.26 & 17.83 & 23.05 $\pm$ 0.08 & 10.61 & cen & 0.65 $\pm$ 0.01 & 1.08 $\pm$ 0.09 & -- $\pm$ -- \\
2014 UQ229 & o4h03 & 49.90 & 0.779 & 5.68 & 20.84 & 21.83 & 22.80 $\pm$ 0.13 & 9.51 & sca & 1.03 $\pm$ 0.02 & 1.94 $\pm$ 0.05 & -- $\pm$ -- \\
2014 UX229 & o4h05 & 39.63 & 0.335 & 15.97 & 25.76 & 26.41 & 22.50 $\pm$ 0.07 & 8.34 & 3:2 & 0.64 $\pm$ 0.02 & 1.53 $\pm$ 0.06 & -- $\pm$ -- \\
2010 TJ182 & o4h07 & 39.65 & 0.276 & 9.50 & 27.88 & 28.86 & 22.55 $\pm$ 0.02 & 8.02 & 3:2 & 0.60 $\pm$ 0.02 & 1.37 $\pm$ 0.04 & -- $\pm$ -- \\
2014 UV228 & o4h09 & 39.49 & 0.228 & 10.13 & 30.78 & 31.75 & 23.57 $\pm$ 0.05 & 8.62 & 3:2 & 0.65 $\pm$ 0.02 & 1.51 $\pm$ 0.04 & -- $\pm$ -- \\
2014 UO229 & o4h11 & 39.45 & 0.161 & 10.09 & 33.77 & 34.03 & 23.82 $\pm$ 0.03 & 8.52 & 3:2 & 0.72 $\pm$ 0.02 & 1.15 $\pm$ 0.06 & -- $\pm$ -- \\
2014 UD229 & o4h13 & 36.39 & 0.145 & 6.85 & 33.58 & 34.31 & 23.66 $\pm$ 0.02 & 8.35 & 4:3 & 0.71 $\pm$ 0.02 & 1.23 $\pm$ 0.06 & -- $\pm$ -- \\
2014 US229 & o4h14 & 55.26 & 0.398 & 3.90 & 32.32 & 33.31 & 23.47 $\pm$ 0.01 & 8.31 & 5:2 & 0.63 $\pm$ 0.02 & 1.50 $\pm$ 0.05 & -- $\pm$ -- \\
2014 UX228 & o4h18 & 36.35 & 0.167 & 20.66 & 37.01 & 37.99 & 23.20 $\pm$ 0.03 & 7.46 & 4:3 & 0.56 $\pm$ 0.03 & 1.44 $\pm$ 0.05 & -- $\pm$ -- \\
2014 UK225$^H$ & o4h19 & 43.52 & 0.127 & 10.69 & 37.09 & 38.06 & 23.32 $\pm$ 0.04 & 7.57 & cla & 0.95 $\pm$ 0.02 & 1.60 $\pm$ 0.04 & 0.69 $\pm$ 0.02 \\
2014 UL225$^H$ & o4h20 & 46.34 & 0.199 & 7.95 & 37.36 & 37.96 & 23.33 $\pm$ 0.11 & 7.57 & cla & 0.55 $\pm$ 0.03 & 1.03 $\pm$ 0.09 & -- $\pm$ -- \\
2014 UH225$^H$ & o4h29 & 38.64 & 0.037 & 29.53 & 39.08 & 40.06 & 23.48 $\pm$ 0.11 & 7.50 & cla & 0.57 $\pm$ 0.02 & 1.69 $\pm$ 0.04 & 0.38 $\pm$ 0.03 \\
2014 UM225$^H$ & o4h31 & 44.48 & 0.098 & 18.30 & 39.50 & 40.16 & 23.53 $\pm$ 0.06 & 7.52 & 9:5 & 0.80 $\pm$ 0.02 & 1.60 $\pm$ 0.04 & -- $\pm$ -- \\
2007 TC434 & o4h39 & 129.92 & 0.695 & 26.47 & 39.74 & 40.60 & 23.47 $\pm$ 0.05 & 7.43 & 9:1 & 0.64 $\pm$ 0.02 & 1.59 $\pm$ 0.05 & -- $\pm$ -- \\
2014 UD225$^{B,G,H}$ & o4h45 & 43.36 & 0.130 & 3.66 & 43.67 & 44.29 & 22.98 $\pm$ 0.05 & 6.55 & cla & 0.74 $\pm$ 0.02 & 1.42 $\pm$ 0.06 & -- $\pm$ -- \\
2001 RY143 & o4h48 & 42.08 & 0.155 & 6.91 & 46.34 & 47.32 & 23.66 $\pm$ 0.08 & 6.95 & cla & 0.92 $\pm$ 0.03 & 1.88 $\pm$ 0.06 & -- $\pm$ -- \\
2014 UE225$^{G,H}$ & o4h50 & 43.71 & 0.066 & 4.49 & 45.96 & 46.56 & 22.92 $\pm$ 0.01 & 6.27 & cla & 1.03 $\pm$ 0.02 & 1.87 $\pm$ 0.05 & -- $\pm$ -- \\
\enddata
\tablenotetext{^A}{Orbit consistent with the Haumea collisional family cluster.}
\tablenotetext{^B}{Confirmed as a binary in \citet{Fraser2017}.}
\tablenotetext{^C}{Photometry includes CFHT measurements.}
\tablenotetext{^D}{2014B observations}
\tablenotetext{^E}{2015B observations}
\tablenotetext{^F}{Telescope tracked non-sidereally at the TNO's rate of motion on-sky.}
\tablenotetext{^G}{Targets with previously-published optical spectral slopes in \cite{Fraser2017} reprocessed here using the latest version of TRIPPy
\citep{Fraser2016}, the most recent data analysis pipeline, and improved SDSS color terms.}
\tablenotetext{^H}{Targets with previously-published optical colors in \cite{Pike2017} reprocessed here using the latest version of TRIPPy
\citep{Fraser2016}, the most recent data analysis pipeline, and improved SDSS color terms. Additional Subaru $r$ and $z$ photometry for some of the highlighted targets is reported in \cite{Pike2017}. Only $(r-z)$ colors obtained from Gemini observing sequences if available are reported here. }
\tablenotetext{}{Orbit Class: Dynamical classification of barycentric orbits from a 10 Myr integration: cen=centaur, sca = scattering disk, cla = classical belt, $N$:$M$= mean motion resonance with Neptune. An `I' after the resonant identifier signifies an insecure resonance classification. See \citet{Bannister2018} for further details.}
\tablenotetext{}{Geometric parameters and derived $H_r$ are reported for the time of the Col-OSSOS observation (see Table \ref{tab:sample_obs}).}
\tablenotetext{}{Mean $m_r$ is the mean of the measured Col-OSSOS $r$-band photometry. Values for individual frames are reported in Table \ref{tab:sample_obs}.}
\tablenotetext{}{A machine-readable version of this table can be found in the online supplemental files. Those targets with two observation epochs have a separate entry for the color measurements derived at each epoch, ordered chronologically.}
\end{deluxetable*}
\begin{deluxetable}{lll}[h]
\tablecolumns{3}
\tablecaption{\label{tab:orbit_numbers} Dynamical classifications of the Col-OSSOS 13BL and 14BH block targets}
\tablehead{\colhead{Orbital Class} & \colhead{$\#$} & \colhead{Comment} }
\startdata
Centaurs & 1 & \\
Cold Classicals & 10 & $i<5\degr$ inclination, all in main belt \\
Hot Classicals & 8 & 1 consistent with the Haumea family \\
Resonant & 14 & 3:2, 5:3, 4:3, 7:4$^*$, 5:2, 11:6, 9:5, 9:1 \\
Scattering & 2 & \\
\enddata
\tablenotetext{^*}{7:4 resonance identification of 2016 BP81 is insecure. See \citet{Bannister2018} for further details.}
\end{deluxetable}
\subsection{Observations and Data Analysis}
\label{sec:uniqobs}
The Gemini observations were acquired as described in Section~\ref{sec:observing_strategy}, during 2014--2015 (under programs GN-2014B-LP-1 and GN-2015B-LP-1) with some minor exceptions described below. We also note here any details where the specific analysis differed from the general analysis for the whole survey given in Section \ref{sec:colours} and Appendixes~\ref{sec:optical} and~\ref{sec:NIR}. Table \ref{tab:sample_obs} gives a detailed summary of the GMOS-N and NIRI observations for each target TNO. For all observations, GMOS-N was equipped with the e2v deep depletion detectors. Table \ref{tab:sample_obs} lists the calculated SDSS magnitude of the observed TNO and associated Gemini-filter zero point for each GMOS and NIRI exposure, as detailed in Appendix \ref{sec:colorterm}. We note that our reported uncertainty in the optical magnitudes combines the uncertainties in the SDSS transformation, calculated zero point, and flux measurement. The full photometry sequences for all target TNOs in this paper are presented in the supplemental material, and a representative sample are plotted in Appendix \ref{sec:aphotplots}.
Exceptional sky conditions occurred during the 2014 August priority visitor run, where many observations were acquired in 0.4-0.5\arcsec seeing. The 2015 October priority visitor run had IQ 20 (0.3--0.6$\arcsec$ seeing). When observing in IQ 20 conditions, we shortened the individual GMOS frames to be less than 300s for 10 TNOs (2001 RY143, 2010 RE188, 2010 TJ182, 2013 UR15, 2014 UE225, 2014 UH225, 2014 UK225, 2014 UL225, 2014 UV228, and 2016 BP81) while preserving our desired SNR goals.
Two TNOs in the sample (2013 SA100 and 2013 UN15) were observed twice. We report both color measurements in Tables \ref{tab:sample} and \ref{tab:sample_obs}.
The dynamically excited TNO 2013 SA100 was observed in both the 2014B and 2015B semesters, with a full $grzJ$ sequence at each epoch.
The cold classical 2013 UN15 was observed in both the 2014B and 2015B semesters. As it appeared very red in the 2014B observations ($r-J \sim 1.5$; Table \ref{tab:sample}), additional time was spent on the 2015 $J$ observations, with no Gemini $g$ observations.
The $g$ observations were instead acquired simultaneously by MegaCam on CFHT with an $rgr$ sequence of ten 300 s g.MP9402 filter images bracketed by two 300 s r.MP9602 filter images before and after. We also report the CFHT photometry and calculated zero points in Table \ref{tab:sample_obs}.
For 4 of the 35 objects (2001 QF331, 2013 UL15, 2013 UM15, and 2013 UQ15), the telescope instead tracked at the on-sky rate of motion for the target TNO rather than standard sidereal tracking. Given their slow rate of motion of a few arcseconds per hour, the PSF of both the stars and TNO were still quite round in the observations.
Thus, the photometry was measured with the same procedure as the sidereally tracked targets, but the uncertainty in the aperture correction was doubled to 0.02 magnitudes to reflect the small errors induced by the non-sidereal tracking. Additionally, \citet{Fraser2017} identified three of the cold classicals in our sample as binaries: 2016 BP81, 2013 SQ99, and 2014 UD225.
In our GMOS observations, the objects were elongated, but the components were not fully separated.
Photometric apertures with radii of 2.5$\times$FWHM and appropriate aperture corrections were used to ensure the flux of both sources was included in the aperture, for a combined photometric measurement (Table \ref{tab:sample_obs}).
For 2013 UX18, the standard $rgJgr$ sequence is incomplete: the first r$_G$-band observation fell on a faint star and was rejected. For the 2015B observations of 2013 UN15, the CFHT data were reduced with the OSSOS procedures described in \citet{Bannister2018}.
Photometry was measured using TRIPPy, in the same fashion as for the Gemini data (see Appendix~\ref{sec:optical}). The $g-r$ color was extracted from the CFHT photometry using the line fitting technique described in Section~\ref{sec:colours}, and converted to the SDSS system using conversions provided as part of the MegaPipe pipeline \citep{2008PASP..120..212G}.
Separately, the line fitting technique was applied to the Gemini photometry to extract $r-z$ and $r-J$ colors in the Gemini system. Finally, the colors were converted to the SDSS system using the $g-r$ color found from the CFHT observations.
\subsubsection{Overlap with previous Col-OSSOS Publications}
Optical colors and optical slopes derived from preliminary analysis of Col-OSSOS observations have been previously published in \cite{Fraser2017} and \cite{Pike2017} for 22 TNOS in our release sample. These targets are identified in Table \ref{tab:sample}. The same Gemini observations used in those publications are analyzed in this Paper. We only report $(r-z)$ colors that were obtained during the Gemini sequences; the full $(r-z)$ sample including near-simultaneous supplemental Subaru $z$ observations can be found in \cite{Pike2017}. We were able to acquire Gemini $z$ observations for five targets: 2013 SA100, 2013 SQ99, 2013 UN15, 2014 UH225, and 2014 UK225. We report the photometry and $(r-z)$ colors for those targets in Table \ref{tab:sample_obs}. The $(r-J)$ color values of our release sample have not been previously published, but we note that preliminary $(g-r)$ and $(J-r)$ colors for the 9 TNOs that overlap with the \cite{Pike2017} sample were plotted in Figure 3 of \cite{Bannister2017} to compare to the near simultaneous $g,r,$ and $J$ photometry obtained for interstellar object 'Oumuamua. \cite{Marsset2018} utilizes Col-OSSOS optical colors of different OSSOS blocks that are not part of this first full data release.
The colors and photometry reported here were reprocessed using the latest version of TRIPPy \citep{Fraser2016}, the most recent data analysis pipeline. We note that small differences in the optical colors and slopes of targets reported in \cite{Fraser2017}, \cite{Pike2017}, and this work are due to updates to TRIPPY, improved light curve fitting, and improved estimates of the color transformations from the Gemini filters to the SDSS photometric system, with the inclusion of additional GMOS observations. Also, during the image reductions of the 2015B Gemini observations of 2013 UN15 that were published in \cite{Pike2017}, the source was contaminated by a background source in the second half of the GMOS sequence. Upon re-reduction for this work, the last image was found to be useable with a sufficiently small $r=0.8\arcsec$ aperture, which avoided the background star. This has caused a small adjustment in the color, and a large improvement in overall color accuracy, which is reflected in the values reported in Tables \ref{tab:sample} and \ref{tab:sample_obs}.
\begin{deluxetable*}{llllllllll}
\tabletypesize{\scriptsize}
\tablecaption{\label{tab:sample_obs} o3l and o4h Col-OSSOS TNO Sample Observations}
\tablehead{\colhead{MPC} & \colhead{OSSOS} & \colhead{Header} & \colhead{Reduced} & \colhead{Filter} &\colhead{MJD} & \colhead{Gemini mag} & \colhead{zero point} & \colhead{SDSS mag} & \colhead{Exposure$^*$}\\ \colhead{ID} & \colhead{ID} & \colhead{ID} & \colhead{Filename} & & & & & & \colhead{(s)}}
\startdata
2013 UR15 & o3l01 & O13BL3RQ & N20140825S0315.fits & r$\_$G0303 & 56894.43641 & 23.234 $\pm$ 0.022 & 28.251 $\pm$ 0.005 & 23.275 $\pm $0.022 & 300 \\
2013 UR15 & o3l01 & O13BL3RQ & N20140825S0316.fits & g$\_$G0301 & 56894.44091 & 23.937 $\pm$ 0.033 & 28.136 $\pm$ 0.005 & 24.03 $\pm $ 0.033 & 300 \\
2013 UR15 &o3l01 & O13BL3RQ & O13BL3RQ$\_$0.fits & J & 56894.46418 & 21.953 $\pm$ 0.117& 23.905 $\pm$ 0.02 & -- $\pm$ -- & 1200 \\
2013 UR15 & o3l01 & O13BL3RQ & O13BL3RQ$\_$1.fits & J &56894.47983 & 21.912 $\pm$ 0.094 & 23.907 $\pm$ 0.02 & -- $\pm$ -- & 1320 \\
2013 UR15 & o3l01 &O13BL3RQ & N20140825S0341.fits & g$\_$G0301 & 56894.49351 & 24.12 $\pm$ 0.035 & 28.175 $\pm$0.005 & 24.213 $\pm$ 0.036 & 225 \\
2013 UR15 & o3l01 & O13BL3RQ & N20140825S0342.fits & g$\_$G0301 & 56894.49706 & 23.977 $\pm$0.032 & 28.18 $\pm$ 0.006 & 24.07 $\pm$ 0.033 & 225 \\
2013 UR15 & o3l01 & O13BL3RQ & N20140825S0343.fits & r$\_$G0303 & 56894.50069 & 23.584 $\pm$ 0.029 & 28.271 $\pm$ 0.007 & 23.625 $\pm$ 0.029 & 225 \\
2001 QF331 & o3l06PD & O13BL3SH & N20140823S0289.fits & r$\_$G0303 & 56892.41022 & 22.839 $\pm$ 0.034 & 28.22 $\pm$ 0.005 & 22.892 $\pm$ 0.034 & 300 \\
2001 QF331 & o3l06PD & O13BL3SH & N20140823S0290.fits & g$\_$G0301 & 56892.41473 & 23.702 $\pm$ 0.039 & 28.127 $\pm$ 0.004 & 23.825 $\pm$ 0.039 & 300 \\
2001 QF331 & o3l06PD & O13BL3SH & N20140823S0291.fits & g$\_$G0301 & 56892.41916 & 23.621 $\pm$ 0.038 & 28.133 $\pm$ 0.005 & 23.745 $\pm$ 0.038 & 300 \\
2001 QF331 & o3l06PD & O13BL3SH & O13BL3SH$\_$0.fits & J & 56892.43259 &21.32 $\pm$ 0.063 & 23.852 $\pm$ 0.02 & -- $\pm$ -- & 840 \\
2001 QF331 & o3l06PD & O13BL3SH & O13BL3SH$\_$1.fits & J & 56892.44375 & 21.197 $\pm$ 0.055 & 23.851 $\pm$ 0.02 & -- $\pm$ -- & 960 \\
2001 QF331 & o3l06PD & O13BL3SH & N20140823S0308.fits & g$\_$G0301 & 56892.45469 & 23.716 $\pm$0.037 & 28.175 $\pm$ 0.004 & 23.839 $\pm$ 0.037 & 300 \\
2001 QF331 & o3l06PD & O13BL3SH & N20140823S0309.fits & r$\_$G0303 & 56892.45919 & 22.895 $\pm$ 0.033 & 28.263 $\pm$ 0.004 & 22.948 $\pm$ 0.033 & 300 \\
\enddata
\tablenotetext{}{A machine-readable version of this table in its entirety can be found in the online supplemental files. A portion is reproduced here for guidance regarding its form and content.}
\tablenotetext{}{All raw Gemini data files and calibration files associated with these observations are available via the Gemini Observatory Archive (\url{https://archive.gemini.edu}). Gemini program IDs for these observations are GN-2014B-LP-1 and GN-2015B-LP-1. }
\tablenotetext{}{All raw CFHT data files and calibration files associated with these observations are available via the CFHT Science Archive (\url{http://www.cadc-ccda.hia-iha.nrc-cnrc.gc.ca/en/cfht/}). CFHT program IDs for these observations are 14BP05 and 15BP05. }
\tablenotetext{}{The Header ID column reflects the OBJECT keyword in the Gemini/CFHT raw FITS headers, as the internal survey designation for target TNOs on occasion may have changed over the course of the OSSOS survey.}
\tablenotetext{}{The reduced FITS files produced in this analysis are available for download at the Canadian Astronomy Data Centre (\url{http://apps.canfar.net/storage/list/ColOSSOS}). The online repository will be available at the time the manuscript is published.}
\tablenotetext{^*}{For $J$-band, the reported exposure time is the total effective exposure time of the stacked image}
\end{deluxetable*}
\subsubsection{Light Curve Effects}
\label{sec:lc}
The range of brightness variations we observed across a Col-OSSOS sequence (between approximately 1 and 6 hours duration) was 0 to 0.5 mag. The photometric variability of our first release sample is presented in Appendix \ref{sec:aphotplots} and the supplemental material. As described in Section \ref{sec:colours}, a linear fit was used to remove light curve effects from our color estimates. We checked for violations of our linear model assumptions that may impact the color measurements presented here. In all cases linearity was sufficient to describe the variability we observed. By comparing the color inferred from the first half and last half of the full $rgJgr$ sequences, we found that no objects exhibited detectable spectral differences over the span of our observations. That is, the $(r-J)$ and $(g-r)$ colors that were inferred from the first half, and last half of each sequence were consistent at better than 2$-\sigma$ in all cases. There still may be other brightness variations present (sinusoidal being one possibility of many) that we cannot determine from our photometry and have not been accounted for in our analysis. Such variations within our observations is unlikely to be correlated with the surface properties of the object in such a way such that it then creates, artificially, the correlations between orbit and surface properties that we report in Section \ref{sec:results}.
\section{Results and Discussion}
\label{sec:results}
The $g-r$ and $r-J$ colors of the 35 TNOs in our sample are presented in Figure~\ref{fig:colours} and are reported in Table \ref{tab:sample}. There are some notable features of the optical-NIR color distribution which we discuss in this section, along with a discussion of clear outliers to the majority of the sample. Additionally, we present an analysis of the intrinsic population of objects that belong to the neutral and red classes of dynamically excited Kuiper belt populations.
\subsection{Colors of the First Release Sample}
\label{sec:first_colours}
\begin{figure*}
\includegraphics[width=\textwidth]{HL_g-J_singlepanel.pdf}
\caption{\label{fig:colours} Optical and near-infrared colors of the 35 TNOs in the Col-OSSOS first release sample (all $m_r < 23.6$ TNOs in the OSSOS 13BL and 14BH survey blocks). The dashed curve indicates the reddening line; see Section \ref{sec:first_colours} for details. Three objects, 2013 SQ99, 2014 UD225, and 2016 BP81 (blue squares), have cold classical orbits (main Kuiper belt with $i < 5\degr$; red squares) and have been previously identified as widely separated binaries. The object 2013 UQ15 (magenta triangle) is dynamically consistent with the Haumea family. Excited TNOs (black dots) belong to the centaurs, scattering, resonant, and hot classical dynamical populations. The two measurements of the two re-observed targets, neutral hot classical 2013 SA100 and red cold classical 2013 UN15, are linked by dashed grey lines (the color measurements at each of the two observation epochs are reported in Table \ref{tab:sample}). The Solar color, with $g-r = 0.45$ and $r-J = 0.97$, is shown by the yellow star.}
\end{figure*}
We present the optical and NIR color distribution of Col-OSSOS targets in the 13BL and 14BH OSSOS blocks in Figure~\ref{fig:colours}.
We include the so-called reddening line, or line of constant spectral slope through the $grJ$ spectral range. This line was calculated using the pysynphot software package \citep{Lim+2015}, using the known bandpass measurements for the SDSS $g$ and $r$ filters, and the Maunakea $J$ filter. We note that the available $J$ bandpass data was measured in the laboratory at room-temperature conditions, rather than at the temperature experienced inside the NIRI dewar\footnote{\url{http://www.gemini.edu/sciops/instruments/niri/imaging/filters}}. This may cause a small deviation of the estimated $J$-band throughput away from the measured curve, and hence the calculated reddening line away from the true curve.
Sub-structure is apparent in the optical-NIR color space of our 35 TNOs (Figure \ref{fig:colours}). The most notable feature is the bifurcation of the dynamically excited populations into two separate color classes, as seen previously (Section \ref{sec:surfaces}). In our sample, the bifurcation into red and neutral clumps occurs at $g-r \sim0.75$. To test for the presence of the bimodality in the Col-OSSOS observations, we apply a multi-dimensional test, the \emph{F~optimal~plane} (FOP) test developed in \citet{Fraser2012}. This test uses minimal spanning tree clustering in Euclidean color space to test for the significance of potential sub-populations within a dataset. We apply the FOP test to our ($g-r$) and ($r-J$) observations. The FOP test divides the population into two separate classes, that can be approximately divided in just the optical color, that is at ($g-r$)$=0.75$, with only a 2\% chance that such a division would occur by chance. We present the minimum spanning tree generated by the FOP test in Figure \ref{fig:FOP}.
\begin{figure*}
\includegraphics[width=\textwidth]{FOP.pdf}
\caption{\label{fig:FOP} The minimal spanning tree, and sub-trees determined by the FOP-test applied to the optical and NIR colors as shown in Figure~\ref{fig:colours}. The branches connecting members of the two subclasses identified by the test are shown in blue and red lines respectively. The one branch of the full tree that is cut to result in the two sub-trees is shown by the dashed black line.}
\end{figure*}
Our sample has two objects that stand as clear outliers from the broad trend of TNO $r-J$ colors being redder than Solar: 2014 UL225 and 2013 UQ15, both of which have $g-r \sim 0.55$ and $r-J \sim 0.95$ (Figure \ref{fig:colours}).
2013 UQ15 is a hot classical object with orbital elements that place it well within the cloud of objects belonging to the Haumea collisional family \citep{Brown2007,Ragozzine2007}. The slightly red optical color and neutral NIR color of 2014 UL225 is very similar to those exhibited by known Haumea family members, including 2005 RR43 and 1995 SM55 \citep{Snodgrass2010}, but the deep water ice absorption signature that is characteristic of Haumea family members cannot be identified with the Col-OSSOS observations alone. While 2014 UL225 exhibits similar spectral properties to 2013 UQ15, its orbital inclination of $7.9^{\circ}$ is significantly lower than the Haumea cloud. We consider if resonant diffusion, as experienced by Haumea itself \citep{Ragozzine2007}, could move 2014 UL225 so far away from the orbital phase space occupied by the majority of known family members. Resonant diffusion is most effective in changing the eccentricity of an orbit. 2014 UL225's eccentricity is consistent with having been affected by diffusion, but its low orbital inclination makes resonant diffusion unlikely. \cite{2012Icar..221..106V} examine the long-term orbital evolution of hypothetical Haumea family members, and there are no instances of resonant diffusion down to inclinations less than 20$^\circ$ for stable orbits. 2014 UL225's colors make it an outlier compared to the bulk of the neutral class. Further study is warranted to determine if this object has a water-rich surface and if it is also consistent with the Haumea family.
We find that the optical/NIR colors vary for the two TNOs (2013 SA100 and 2013 UN15) that we observed twice. We report the objects' two sets of colors measurements in Figure \ref{fig:colours} and Table \ref{tab:sample}. Spectral variability at this level has been observed in other small TNOs \citep[see ][]{Fraser2015}. We note that in neither case do the variations in colors between epochs shift the TNOs from their color class: they are consistent in color class despite the variability. For the hot classical 2013 SA100, the colors measured at each epoch both place the TNO firmly within the neutral class, but the colors are discrepant at the $1$-$\sigma$ level for each of the repeated $g-r$, $r-J$, and $r-z$ colors
Both measurements of cold classical 2013 UN15, place it in the red cloud but the optical and near infrared colors are not consistent within the $1$-$\sigma$ measurement uncertainty. We find 2013 UN15 varies by 0.2 magnitudes or more in $r-J$ and $r-z$ \citep[also noted by ][]{Pike2017}.
The bulk of the objects with $g-r <0.75$ (the neutral class) appear to exhibit an inverse correlation in their optical and NIR colors (see Figure \ref{fig:colours}). We apply the Spearman rank test to examine this further. When we exclude the candidate Haumea collisional fragment 2013 UQ15 as it is a surface type produced via collision and not intrinsic to the planetesimal disk, the Spearman rank test suggests that there is a 30\% chance that the observed correlation would occur by chance, finding no evidence that the correlation is statistically significant. The correlation becomes much stronger if we exclude 2013 UQ15 and 2014 UL225, both outliers from the excited TNO distribution. Excluding 2013 UQ15 and 2014 UL225, the Spearman rank test suggests that there is only a 2\% chance that the observed correlation would occur by chance, but it is not clear that 2014 UL225 surface colors are due to some process similar to the Haumea family formation that would justify excluding it from the analysis. Correlations in the optical and NIR colors of dynamically excited TNOs have been identified previously. In particular, the NIR colors of both the neutral and red class members correlate \emph{positively} with their optical colors in the Hubble Space Telescope (HST) Wide-Field Camera 3 broadband filters centered at $\sim$0.6 $\mbox{$\mu$m}$ (F606w), $\sim$0.8$ \mbox{$\mu$m}$ (F814w), and $\sim1.39 \mbox{ $\mu$m}$ (F139m) \citep{Fraser2012}. That result is in stark contrast to the potential inverse optical-NIR color correlation found here in the $g$, $r$, and $J$ filters, with band centers at $0.48 \mbox{ $\mu$m}$, $0.62 \mbox{ $\mu$m}$, and $1.25 \mbox{ $\mu$m}$, respectively. For the dynamically excited objects in the red class, optical and NIR colors exhibit a positive correlation in the HST filter set. Yet strangely, no correlation of any kind is present in the Col-OSSOS $grJ$ observations of 7 the dynamically excited red objects in our sample (as shown in Figure \ref{fig:colours}). This may suggest that the $J$-band samples a different part of TNO surface reflectance than the F139m HST band, but we cannot demonstrate that the correlation at the level seen in the HST band can be excluded by this sample of TNOs. A large $grJ$ sample is required; this topic will be further investigated in future Col-OSSOS data releases.
The color trends observed in the HST filters have been used to provide compositional constraints for the bulk ice and silicate components of small TNOs \citep{Fraser2012}. For example, organic materials have been suggested to account for the positively correlated optical and NIR colors through a simple compositional mix: higher organic content leads to redder optical and NIR colors. If the inverse correlation that we have detected for the neutral class of dynamically excited TNOs is confirmed, then the material that is responsible for the optical-NIR color correlations of the neutral class must exhibit an absorption feature that overlaps $J$, and not the HST F139m filter, so as to account for the signs of the correlations seen in the two different filter sets. Clearly, confirmation of the inverse color correlation in the neutral class is important.
To date, no correlation between the optical and NIR colors of the cold classical objects has been detected. Rather, the bulk of cold classical TNOs exhibits a range of red optical and NIR colors, broadly spanning nearly the full range of colors exhibited by the red dynamically excited objects, though in an uncorrelated fashion. The only objects for which this does not appear to be true are the blue binaries \citep{Fraser2017}. All members of this recently discovered class of cold classical TNOs exhibit colors compatible with the neutral dynamically excited class throughout the $grzJ$ wavelength range \citep{Pike2017}, and exhibit a nearly 100\% binary fraction \citep{Fraser2017}. If these objects are indeed survivors of a soft push-out via mean motion resonance sweep-up during the smooth phases of Neptune's outward migration \citep{Fraser2017}, then it follows that the binary cold classical objects should also exhibit a bimodal optical color distribution, like the dynamically excited TNOs. As yet, insufficient data are available to test this assertion.
\subsection{The Color Fraction of Red/Neutral Surfaces in the Dynamically Excited Kuiper Belt}
\label{sec:colour_fraction}
To estimate the intrinsic fraction of objects in the neutral and red classes of dynamically excited TNOs, we consider an analytic derivation of the number of objects observed within a given OSSOS survey block. To derive that number, we consider a Kuiper belt comprised of bodies that can be described by object radius $R$ , heliocentric distance $r$, and albedo $a$. Distributions in those parameters within the Kuiper belt are thus given by $f(R)$, $g(r)$, and $h(a)$, respectively. Here, $g(r)$ will largely depend on the distribution of resonant TNOs, which are preferentially found at certain longitudes with respect to Neptune \citep[see][for example]{Gladman2012}. For resonators and other excited TNOs, no correlations between optical colors and perihelion/argument of perihelion/longitude of ascending node have been observed \citep{Peixinho2015}. Therefore the explicit longitudinal and latitudinal structure will only affect the absolute number of observed objects at a given sky location, and not the fraction of objects in a color class at that pointing. Thus, for clarity of our derivation, we avoid writing $g$ as a function of latitude and longitude.
The number of objects with albedos between $a$ and $a+da$, radii $R$ and $R+dR$ and distances $r$ and $r+dr$ is given by:
\begin{equation}
n(R,r,a) = A\, f(R)\, g(r)\, h(a)\, da\, dr\, dR
\label{eq:n}
\end{equation}
\noindent
where $A$ is a convenience constant to determine the desired density unit, we adopt objects per square degree.
The magnitude of an object is $m = K - 2.5 \log(a) + 5 \log(r\,\Delta) - 5 \log(R)$ where $K$ is a constant related to the Solar luminosity, $\Delta$ is the geocentric distance to the object, and is a function of $r$, and we have ignored phase effects. Writing $R$ in terms of $m$ and its derivative with respect to $m$, we have
\begin{eqnarray}
R & = & \frac{r\, \Delta}{\sqrt a}10^{\frac{K-m}{5}} \\
dR & = & \frac{-\ln 10}{5} R dm
\label{eq:R}
\end{eqnarray}
The majority of Col-OSSOS targets have r-band absolute magnitudes brighter than $H_{r}\sim8$, the approximate magnitude at which absolute magnitude distribution of the dynamically excited objects transitions from a steep power-law, to a shallower slope \citep{Bernstein2004, Fuentes2008, Fraser2009, Fraser2014,2014AJ....148...55A}. As such, we will consider only objects with $H_{r}<8$, and approximate the size distribution is a power-law of the form $f(R) = C\, R^{-q}$, where $C$ is a normalization constant and $q$ is the power-law slope. Substituting Equation \ref{eq:R} by this size distribution into Equation \ref{eq:n}, $n$ can be defined as:
\begin{equation}
n(R,r,a) = \frac{\ln10\, C\, A}{5} \, R^{1-q}\, g(r)\, h(a)\, da\, dr\, dm
\end{equation}
Assuming that the Kuiper belt is bounded by distances $r_0\leq r \leq r_1$, and objects in it have albedos with values $a_0 \leq a \leq a_1$, the number of objects between magnitudes $m_0$ and $m_1$ is shown by Equation \ref{eq:16}.
\begin{equation}
N = \frac{ \ln10\, C\, A }{5}\int_{a_o}^{a_1} \int_{r_o}^{r_1} \int_{m_o}^{m_1} h(a) a^{\left(\frac{q-1}{2}\right)} g(r) r^{1-q} \Delta^{1-q} 10^{\frac{(1-q)(K-m)}{5}} dm\, dr\, da.
\label{eq:16}
\end{equation}
If we substitute $q = 5\alpha+1$, where $\alpha$ is the logarithmic slope of the power law, we find Equation \ref{eq:17},
\begin{eqnarray}
N & = & \frac{ \ln10\, C\,A }{5}\int_{a_o}^{a_1} \int_{r_o}^{r_1} \int_{m_o}^{m_1} h(a) a^{\frac{5\alpha}{2}} g(r) r^{-5\alpha} \Delta^{-5\alpha} 10^{\alpha(m-K)} dm\, dr\, da \nonumber \\
& = & \frac{C\,A\, 10^{-\alpha K} }{5\alpha}\int_{a_o}^{a_1} \int_{r_o}^{r_1} h(a) a^{\frac{5\alpha}{2}} g(r) r^{-5\alpha} \Delta^{-5\alpha} dr da \left[10^{\alpha m_1} - 10^{\alpha m_o}\right]
\label{eq:17}
\end{eqnarray}
We consider a simple survey in which $m_2 >> m_1$ and has a constant efficiency $\eta$ that goes to zero at magnitude $m$. Then the number of observed objects is given by Equation \ref{eq:18}, and we have arrived at the general form of the cumulative luminosity function $N(<m) = 10^{\alpha\left(m-m_o\right)}$, where $\alpha\sim0.7$ \citep{Fraser2008,Fuentes2008,Petit2011}.
\begin{equation}
N (<m) = \frac{C\,A\, 10^{-\alpha K} }{5\alpha}\int_{a_o}^{a_1} h(a) a^{\frac{5\alpha}{2}} da \int_{r_o}^{r_1} g(r) r^{-5\alpha} \Delta^{-5\alpha} dr 10^{\alpha m}
\label{eq:18}
\end{equation}
Now consider that the Kuiper belt exhibits two main color populations: the red and neutral objects. These populations differ in their albedo distributions \citep{Stansberry2008,Fraser2014,Lacerda2014}, and overall number density. If, within a given survey pointing, we assume they share the same size and radial distributions, we can derive the observed red:neutral population ratio for a given intrinsic ratio. While the latter assumption hasn't been tested, no detectable size distribution differences, other than absolute number, have been detected over the observable range of the dynamically excited neutral and red classes in surveys that are sensitive to those differences \citep[e.g.][]{Wong2017}. With these assumptions, if the intrinsic ratio of objects in the red and neutral populations are given by $A_{\textrm{n}} = \gamma A_{\textrm{r}}$, then the \emph{observed} ratio of the red and neutral populations, $R_{\textrm{r,n}}(<m)$), is:
\begin{equation}
R_{\textrm{r,n}}(<m) = \frac{1}{\gamma} \frac{\int_{a_o}^{a_1} h_{\textrm{r}}(a) a^{\frac{5\alpha}{2}} da}{\int_{a_o}^{a_1} h_{\textrm{n}}(a) a^{\frac{5\alpha}{2}} da}
\label{eq:ratio}
\end{equation}
\noindent
For our observations, the limiting magnitude, $m$ is $m_r = 23.6$ though we note that there is no explicit dependence of Equation~\ref{eq:ratio} on $m$.
While it is certainly true that the red and neutral populations exhibit a range of albedos, the true distribution is currently unknown. Thus, for simplicity and a basic first estimate of the intrinsic red:neutral fraction, we model the two populations as having a single unique albedo, $a_{\textrm{r}}$ and $a_\textrm{n}$. That is, $h_\textrm{r}(a) = \delta(a_\textrm{r} - a)$ and $h_\textrm{n}(a) = \delta(a_\textrm{n} - a)$, where $\delta$ is the Dirac delta function. Then we are presented with the simple red:neutral fraction relation
\begin{equation}
R_{\textrm{r,n}}(<m) = \frac{1}{\gamma} \frac{ a_\textrm{r}^{\frac{5\alpha}{2}}}{a_\textrm{n}^{\frac{5\alpha}{2}}}
\label{eq:redneutral}
\end{equation}
\noindent
which, importantly, is independent of limiting magnitude. The mean albedos for the red and neutral populations are $a_\textrm{r}=12\%$ and $a_\textrm{n}=6\%$ respectively \citep{Fraser2014,Lacerda2014}. Thus, we find $R_{r,n} \sim \frac{3.4}{\gamma}$, where $\gamma$ is the \emph{intrinsic} ratio of neutral to red objects in the dynamically hot TNO population.
As we are considering only the bulk of the dynamically excited TNOs, we exclude the potential Haumea family member 2013 UQ15 due its unique surface properties attributed to its collisional origin. We avoid counting the three targets (2014 UJ225, 2014 UQ229, and 2013 UR15) which have $H_{r}>8$, and therefore avoid the region where a single power-law size distribution is not satisfied \citep{Bernstein2004, Fuentes2008, Fraser2009, Fraser2014}. Thus, in the remaining bulk sample of dynamically excited objects, there are 9 neutral and 3 red class objects in 14BH block, and 4 neutral and 3 red objects in 13BL block. The red:neutral ratios of each block are consistent at the $2$-$\sigma$ level. Together, the observed ratio is $R_{r,n}=6/13$. Considering the $1$-$\sigma$
range on the observed ratio and Equation~\ref{eq:redneutral}, we find $\gamma=7.4_{-3}^{+3.6}$. Thus, the observed population implies that in the intrinsic population, the neutral class outnumbers the red class, by a factor of 4.4-11.0. We further note that adoption of a distribution of albedos for each class has a tendency to increase this factor substantially. For example, if we adopt uniform albedo distributions that span the observed range of albedos of the neutral and red classes ($0.04\leq a \leq 0.08$ and $0.08\leq a \leq 0.22$), the inferred intrinsic ratio would be a factor of $\sim3$ higher than what we infer using the mean albedos of each class. Thus, our result should be interpreted as a lower limit.
\citet{Wong2017} adopt a different approach to determining the relative neutral:red population fraction. We note that the two color categories in \citet{Wong2017}, ``red" and ``very red", are similar to our ``neutral" and ``red" color categories respectively. Instead of integrating to a certain limiting magnitude, \citet{Wong2017} consider only objects detected in their survey to a given size, with appropriate assumptions on albedo of each of the red and neutral classes. Over a similar size range as that discussed here, they find that the intrinsic neutral to red number ratio is $\gamma=3.6\pm1.2$ where the uncertainty on this number is derived from the $1$-$\sigma$ Poisson range on the observed number of objects. This number is in $2$-$\sigma$ agreement with our measured value.
\subsection{The Structure of the Protoplanetesimal Disk}
\label{sec:disk}
If the separate classes of TNOs reflect the compositional structure of the protoplanetesimal disk from which they originated \citep[see][for example]{Fraser2012}, the presence of only two classes of dynamically excited TNO argues for a moderately compositionally homogenous disk, up to the level of our measurement uncertainty.To explain our observed optical-NIR color distribution of excited TNOs, only one compositional division would be needed, between $\sim20$ and $\sim30$~au where the majority of dynamically excited TNOs originated \citep[e.g.][]{2005Natur.435..466G,Levison2008,2013Icar..225...40B,Nesvorny2015b,2016ApJ...825...94N}. We note that we cannot rule out further finer color-composition structure below our measurement precision ($\sim$0.04 mag ) that may exist in the dynamically excited TNO source population.
From our calculated $\gamma$, we can estimate where the division between the neutral and red classes occurred. To that end, we assume a simple disk, with a surface density described as $\Sigma(r)\propto r^{-\beta}$ with inner and outer extents of $r_{\textrm min}$ and $r_{\textrm max}$. To gauge the radial extent, we turn to models of the Solar System's large scale dynamical restructuring (Section \ref{sec:history}).
The currently favored scenario for giant planet migration and dispersal of the planetesimal disk is that of \citet{Nesvorny2015}, in which Neptune originates at 22 au, implying a disk inner edge of $r_{\textrm min}\sim23$~au. Objects in the dynamically excited populations originate inside the final location of the 3:2 mean motion resonance with Neptune, or $\sim39$~au.
We adopt that value for the outer edge of the disk from which dynamically excited objects originated.
In the disk, we hypothesize a sharp transition distance, $r_{\textrm s}$ for the original locations of the neutral and red populations. We note that our disk model is based on a simple assumption about the radial surface density distribution that at some level is not correct. For example, it is generally accepted that a sharp density gradient at $\sim$30 au is required to halt Neptune during its late stages of migration \citep[see for example][]{2004Icar..170..492G}. Such a gradient is likely steeper than reflected in our simple model. As a result, the distance $r_s$ is likely interior to the value we estimate below.
Additional observational constraints also help inform the portrait of the protoplanetesimal disk. The dynamically quiescent cold classicals exhibit a different range of $r-z$ colors distinct from red dynamically excited TNOs even though they exhibit similar $r-J$ and $g-r$ colors \citep{Pike2017}. Thus, \cite{Pike2017} infer that the cold classicals are their own unique TNO surface type.
It follows from this result that the disk had a second division beyond which the cold classicals originated.
This second compositional division/boundary must have been near the current inner edge of the cold classical objects, to explain why cold classical-like surfaces are rare (or not present) in the dynamically excited populations. \cite{Fraser2017} found that the blue binary cold classicals, which have neutral colors consistent with the neutral excited TNO surfaces, are interlopers emplaced during Neptune migration. This places an additional constraint on where this cold classical surface boundary can be. Dynamical modeling by \cite{Fraser2017} find that in order to deliver the blue binaries onto cold classical orbits during Neptune migration, neutral surfaces were present up to the inner edge of the cold classical belt. Thus, the transition to red cold classical surfaces would be expected near the start of the present-day cold classical belt with red excited TNOs originating more inward than the neutral TNO surfaces.
Combining our results with these additional observational constraints, we can explore the red/neutral transition region for the source of the excited TNOs. Despite not knowing $\beta$, the power-law slope of the disk surface density, we can use the inferred intrinsic neutral to red population ratio to place some constraints on the transition distance. For $0\leq\beta\leq3$, our values of gamma imply $37.4\leq r_{\textrm s}\leq 38.5$ au if the neutral class originated inside $r_{\textrm s}$, or $32.5\leq r_{\textrm s}\leq 33.4$ au if the neutral class originated outside $r_{\textrm s}$.
This estimate fails to account for any variation in efficiency of scattering from certain regions of the disk into different dynamical classes within the Kuiper belt.
It also fails to account for the currently unexplained sharp transition in surface density of the protoplanetesimal disk at $\sim30$~au, which is seemingly required to halt Neptune's migration at the correct distance. As such, this estimate should be taken only as a rough guide for the location of $r_{\textrm s}$. A more thorough estimate will be made through the use of the OSSOS survey simulator \citep{Lawler2018}, and forward-modeling migration model output, when a larger sample of Col-OSSOS photometry is complete.
\begin{figure*}
\includegraphics[width=\textwidth]{disk_structure_cartoon.pdf}
\caption{\label{fig:disk} The implied planetesimal disk structure, not to scale, under the assumption of a moderately compositionally homogenous disk.}
\end{figure*}
Synthesizing all the observational constraints from \cite{Pike2017} and \cite{Fraser2017} with our results based on the assumptions described in Section \ref{sec:colour_fraction}, we find a protoplanetesimal disk with a red-blue-red structure, as shown in Figure \ref{fig:disk}.
Closest to the Sun, today's dynamically excited red class originates at a point interior to the neutral class, with the division between the two at $\sim33$~au.
The dynamically excited neutral class starts interior to the cold classical objects, with a division between the two only a few au inside of the current inner edge of the cold classical region, at $\sim40$~au.
The higher inclinations of the neutral dynamically excited TNOs \citep{Marsset2018} imply they have experienced a more agitated dynamical history than the red ones \citep{Gomes2003}, potentially complicating this picture.
A more detailed comparison between Col-OSSOS observations and output of dynamically compatible migration simulations will test the viability of this overall compositional picture. We also note that finer $grJ$ color-composition structure that is not resolvable by our measurements may exist within each of these composition classes, further complicating this picture.
\section{Conclusions}
We present optical and NIR colors of 35 TNOs, found in the 13BL and 14BH OSSOS discovery blocks with magnitudes brighter than $m_r=23.6$. In $g-r$ and $r-J$, the dynamically excited TNOs, exhibit two classes of objects: the neutral and red classes. We find a tentative negative correlation between the $g-r$ and $r-J$ colors for the neutral class, but additional observations are needed to confirm.
We find no evidence for a correlation in the colors of the dynamically excited red class. Assuming a population density that is a separable function of distance, size, and albedo, we find that the neutral class outnumbers the red class by at least 4.4:1.0 but could be as high as 11.0:1.0.
We find that the cold classical TNOs predominantly occupy the same range of colors in $g-r$ and $r-J$ as the dynamically excited, red class of TNO. As shown by \citet{Pike2017} however, the cold classical TNOs occupy a different range of $r-z$, demonstrating that the cold classicals present a different surface than the equivalently optically red excited objects. Combining this observation with our data, we find that within our measurement uncertainty, our observations are consistent with the bulk of TNOs: dynamically excited neutral, dynamically excited red, and cold classical. This excludes rare objects such as the Haumea family members \citep{Brown2007}, the volatile-bearing dwarf planet-sized bodies \citep{Schaller2007,Brown2008}, or the silicate-rich TNO 2004 EW95 \citep{Seccull2018}.
Based on the assumptions and simple TNO model described in Section \ref{sec:first_colours}, we find our observations are consistent with a planetesimal disk with two compositional divisions separating three separate classes of objects occurred at roughly 33~au, and at just a few astronomical units inside $40$~au, the current inner edge of the cold classical region. We note that further finer color-composition structure in the planetesimal disk in $grJ$ color space that are not resolvable with our measurement cannot be ruled out. To probe the possibility of finer structure in the TNO color/composition space will require measurement uncertainties smaller than 0.01 magnitude in $g,r,$ and $J$. Our observations and past TNO color measurements are consistent with the 3 surface type model for the bulk of the TNO population. The presence of neutral class interlopers in the cold classical region suggests that the neutral objects bordered the inner primordial edge of the cold classical objects, and the red TNOs began interior to the neutral class. The complete Col-OSSOS sample is expected to include 96 objects from five OSSOS blocks, and will include additional $u$-band photometry from CFHT, acquired simultaneously alongside the Gemini observations. This future four-band dataset will be used to generate a robust taxonomic system for TNOs that accounts for the correlated optical and NIR colors they exhibit.
\section*{Acknowledgements}
The authors acknowledge the sacred nature of Maunakea, and appreciate the opportunity to observe from the mountain.
This work is based on observations from the Large and Long Program GN-2014B-LP-1 and GN-2015B-LP-1, at the Gemini Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under a cooperative agreement with the NSF on behalf of the Gemini partnership: the National Science Foundation (United States), the National Research Council (Canada), CONICYT (Chile), Ministerio de Ciencia, Tecnolog\'{i}a e Innovaci\'{o}n Productiva (Argentina), and Minist\'{e}rio da Ci\^{e}ncia, Tecnologia e Inova\c{c}\~{a}o (Brazil). This work is also based on observations obtained with MegaPrime/MegaCam, a joint project of CFHT and CEA/DAPNIA, at the Canada-France-Hawaii Telescope (CFHT) which is operated by the National Research Council (NRC) of Canada, the Institut National des Sciences de l'Univers of the Centre National de la Recherche Scientifique of France, and the University of Hawaii. We thank the Gemini North staff for their support of the Col-OSSOS program. In particular, we acknowledge the Gemini North queue coordinators, program contact scientists, science operations specialists, and the NIRI and GMOS instrument teams for their assistance. We also thank the CFHT staff for their support of the Col-OSSOS program. The authors also thank the anonymous reviewer for the careful and constructive review that improved this manuscript.
MES was supported by Gemini Observatory and also in part by an Academia Sinica Postdoctoral Fellowship.
M.T.B. appreciates support during Col-OSSOS from UK STFC grants ST/P0003094/1 and ST/L000709/1, the National Research Council of Canada, and the National Science and Engineering Research Council of Canada.
N.P. acknowledges funding from the Portuguese FCT --- Foundation for Science and Technology (ref: SFRH/BGCT/113686/2015). CITEUC is funded by National Funds through FCT --- Foundation for Science and Technology (project: UID/ Multi/00611/2013) and FEDER - European Regional Development Fund through COMPETE 2020 - Operational Programme Competitiveness and Internationalisation (project: POCI-01-0145-FEDER-006922).
KV acknowledges support from NASA grants NNX15AH59G and NNX14AG93G.
This research used the facilities of the Canadian Astronomy Data Centre operated by the National Research Council of Canada with the support of the Canadian Space Agency.
This work also made use of the Gemini Observatory Archive, NASA's Astrophysics Data System Bibliographic Services, the JPL HORIZONS web interface (\url{https://ssd.jpl.nasa.gov/horizons.cgi}), and data and services provided by the International Astronomical Union's Minor Planet Center. This research made use of Astropy, a community-developed core Python package for Astronomy \citep{Astropy2013, Astropy2018}. PyRAF are products of the Space Telescope Science Institute, which is operated by AURA for NASA.
The Pan-STARRS1 Surveys (PS1) and the PS1 public science archive have been made possible through contributions by the Institute for Astronomy, the University of Hawaii, the Pan-STARRS Project Office, the Max-Planck Society and its participating institutes, the Max Planck Institute for Astronomy, Heidelberg and the Max Planck Institute for Extraterrestrial Physics, Garching, The Johns Hopkins University, Durham University, the University of Edinburgh, the Queen's University Belfast, the Harvard-Smithsonian Center for Astrophysics, the Las Cumbres Observatory Global Telescope Network Incorporated, the National Central University of Taiwan, the Space Telescope Science Institute, the National Aeronautics and Space Administration under Grant No. NNX08AR22G issued through the Planetary Science Division of the NASA Science Mission Directorate, the National Science Foundation Grant No. AST-1238877, the University of Maryland, Eotvos Lorand University (ELTE), the Los Alamos National Laboratory, and the Gordon and Betty Moore Foundation.
Funding for the Sloan Digital Sky Survey IV has been provided by the Alfred P. Sloan Foundation, the U.S. Department of Energy Office of Science, and the Participating Institutions. SDSS-IV acknowledges support and resources from the Center for High-Performance Computing at the University of Utah. The SDSS web site is \url{www.sdss.org}. SDSS-IV is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS Collaboration including the Brazilian Participation Group, the Carnegie Institution for Science,
Carnegie Mellon University, the Chilean Participation Group, the French Participation Group, Harvard-Smithsonian Center for Astrophysics,
Instituto de Astrof\'isica de Canarias, The Johns Hopkins University,
Kavli Institute for the Physics and Mathematics of the Universe (IPMU) /
University of Tokyo, Lawrence Berkeley National Laboratory,
Leibniz Institut f\"ur Astrophysik Potsdam (AIP),
Max-Planck-Institut f\"ur Astronomie (MPIA Heidelberg),
Max-Planck-Institut f\"ur Astrophysik (MPA Garching),
Max-Planck-Institut f\"ur Extraterrestrische Physik (MPE),
National Astronomical Observatories of China, New Mexico State University,
New York University, University of Notre Dame,
Observat\'ario Nacional / MCTI, The Ohio State University,
Pennsylvania State University, Shanghai Astronomical Observatory,
United Kingdom Participation Group,
Universidad Nacional Aut\'onoma de M\'exico, University of Arizona,
University of Colorado Boulder, University of Oxford, University of Portsmouth,
University of Utah, University of Virginia, University of Washington, University of Wisconsin,
Vanderbilt University, and Yale University.
|
\section{\bf Introduction}
In this work we study the Phragm\'en-Lindel\"of property of viscosity solutions $u(x,t)$ for a class of nonlinear parabolic equations on the infinite strip $\mathbb{R}^n_T=\mathbb{R}^n\times (0,T)$, where $n\ge 2$ and $0<T<\infty$. The current work may be viewed as partly complementing the work \cite{BM6}. See also, \cite{BM3}.
Set $\mathbb{R}^n_T=\mathbb{R}^n\times (0,T)$ and let $g:\mathbb{R}^n\rightarrow (0,\infty)$ be continuous and $f:[0,\infty)\rightarrow [0,\infty)$ be an increasing continuous function. As described in \cite{BM6}, the motivation for this work arises from the study of doubly nonlinear equations of the kind
\begin{eqnarray}\label{sec1.1}
H(Du, D^2u)-f(u)u_t=0, \;\;\mbox{in $\mathbb{R}^n_T,$}\;\;\mbox{with $u(x,0)=g(x),\;\forall x$ in $\mathbb{R}^n$,}
\end{eqnarray}
where $H$ satisfies certain homogeneity conditions and $u\in C(\mathbb{R}^n\times [0,T))$ is a viscosity solution. See Section 2 for more details.
As noted in \cite{BM5, BM6}, if $f$ satisfies certain conditions then there is an increasing function $\phi$ and a non-increasing function $Z\ge 0$ such that the change of variable $u=\phi(v)$ transforms the differential equation in (\ref{sec1.1}) to
\eqRef{sec1.2}
H(Dv, D^2v+Z(v)Dv\otimes Dv)-v_t=0, \;\mbox{in $\mathbb{R}^n_T$ with $v(x,0)=\phi^{-1}(g(x)),\;\forall x$ in $\mathbb{R}^n$.}
\end{equation}
It follows that the solutions of (\ref{sec1.2}) and hence, the solutions of (\ref{sec1.1}), satisfy a comparison principle, see \cite{BM1, BM2, BM5}. Incidentally, we do not require that $Z$ be defined in all of $\mathbb{R}$, a matter that will be discussed later. For purposes of the current discussion, we will overlook this issue.
As done in \cite{BM6}, we consider a some what more general setting and study Phragm\'en-Lindel\"of type results for equations of the kind
\begin{eqnarray}\label{sec1.3}
&&H(Dv, D^2v+Z(v)Dv\otimes Dv)+\chi(t) |Dv|^{\sigma}-v_t=0, \;\;\mbox{in $\mathbb{R}^n_T$},\nonumber\\
&&\qquad\mbox{$v(x,0)=h(x)$}, \;\forall x\in \mathbb{R}^n,
\end{eqnarray}
where $\sigma\ge 0$ and $\chi:(0,T)\rightarrow \mathbb{R}$ and $h:\mathbb{R}^n\rightarrow \mathbb{R}$ are both continuous and bounded.
In \cite{BM6}, we assumed that $\sup_\lambda \left[ \min_{|e|=1}H(e, \lambda e\otimes e+I) \right]=\infty$, where $e$ is a unit vector, $I$ is the $n\times n$ identity matrix and $\lambda$ is a real valued parameter. We showed that the maximum principle was valid for solutions that satisfied certain growth rates for large $x$. The class of operators, we considered, included, among others, the $p$-Laplacian ($p\ge 2$), the infinity-Laplacian and the Pucci operators.
The current work addresses the case $\sup_\lambda \left[ \max_{\{e=1\}}|H(e, \lambda e\otimes e\pm I)| \right]<\infty$ and, in a sense, complements \cite{BM6}. In Section 2, we have listed some examples of operators that satisfy this condition.
We remark also that, much like \cite{BM6}, the imposed growth rates are influenced by the dueling terms $Z(v)Dv\otimes Dv)$ and $\chi(t)|Dv|^\sigma$ and the power $\sigma$.
Since $Z\ge 0$, by ellipticity, $H(Du, D^2u)\le H(Du, D^2u+Z(u)Du\otimes Du)$. Our work will show that, unlike \cite{BM6}, $Z(s)$ can be allowed to vanish, i.e,
$Z(s)=0,\;\forall s\ge s_0$, for some $s_0$. The value of $Z$ does not influence the bound on $H(e, \lambda e\otimes e\pm I)$.
We have divided our work as follows. In Section 2, we introduce more notation and state the main results. Section 3 contains preliminary calculations and previously proven lemmas, useful for the current work, In Sections 4 and 5, we present the constructions of super-solutions and sub-solutions respectively. Section 6 addresses some special situations. The proofs of the main results appear in Section 7.
As a final note, we do not address questions of existence and uniqueness and nor do
we address optimality of the growth rates stated in the theorems. Also, we direct the
reader to \cite{AJK, ED, JL, TR, Tych} for related questions and discussion.
\vskip 2em
\section {Notation and main results}
\vskip 1em
In this work, sub-solutions, super-solutions and solutions are meant in the sense of viscosity. For definitions, we direct the reader to \cite{BM5,CIL}.
We introduce notation that are used throughout this work. We address the problems in (\ref{sec1.1}) and (\ref{sec1.3}) on infinite strips in $\mathbb{R}^{n+1}$ where $n\ge 2$. The letter $o$ denotes the origin in $\mathbb{R}^n$ and $e$ denotes a unit vector in $\mathbb{R}^n$. Let $S^{n\times n}$ be the set of all symmetric $n\times n$ real matrices. Let $I$ be the identity matrix and $O$ the $n\times n$ zero matrix. The expressions $usc$ and $lsc$ stand for {\it upper semi-continuous} and {\it lower semi-continuous} respectively.
Through out this work, we assume that $H$ satisfies the following conditions.
\vskip 1em
{\bf Condition A (Monotonicity):} Let $H:\mathbb{R}^n\times S^{n\times n}\rightarrow \mathbb{R}$ is continuous for any $(q,X)\in \mathbb{R}^n\times S^{n\times n}$. We require that
\begin{eqnarray}\label{sec2.1}
&&\mbox{(i)}\;H(q,X)\le H(q,Y),\;\mbox{$\forall\;q\in \mathbb{R}^n$ and $\forall\;X,\;Y$ in $S^{n\times n}$, with $X\le Y$}, \nonumber\\
&&\mbox{(ii)}\;H(q,O)=0,\;\mbox{$\forall\;q\in \mathbb{R}^n$.}
\end{eqnarray}
Clearly, for any $q\in \mathbb{R}^n$ and $X\in S^{n\times n}$, $H(q,X)\ge 0$ if $X\ge O$.
\vskip 1em
{\bf Condition B (Homogeneity):} There is a constant $k_1\ge 0,$ such that for any $(q,X)\in \mathbb{R}^n\times S^{n\times n}$,
\begin{eqnarray}\label{sec2.2}
&&\mbox{(i)}\;H(\theta q, X)=|\theta|^{k_1}H(q, X),\;\mbox{$\forall\;\theta\in \mathbb{R}$, and}\nonumber\\
&&\mbox{(ii)}\;H(q, \theta X)=\theta H(q, X),\;\mbox{$\forall\;\theta>0.$}
\end{eqnarray}
\vskip 1em
We introduce two quantities before stating the next condition. For any unit vector $e\in \mathbb{R}^n$, we recall that $(e\otimes e)_{ij}=e_ie_j,$ for any $i,j,=1,2, \cdots, n.$ Moreover, $e\otimes e\ge O$. For $\lambda\in \mathbb{R}$, set
\begin{eqnarray}\label{sec2.3}
\Lambda_{\min}(\lambda)=\min_{|e|=1}H(e, \lambda e\otimes e-I)\;\;\mbox{and}\;\;\Lambda_{\max}(\lambda)=\max_{|e|=1}H(e, \lambda e\otimes e+ I).
\end{eqnarray}
By Condition A, both $\Lambda_{\min}(\lambda)$ and $\Lambda_{\max}(\lambda)$ are non decreasing functions of $\lambda$.
\vskip 1em
{\bf Condition C(Growth at Infinity):} We impose that
$$\max_{|e|=1}H(e, -I)<0<\min_{|e|=1}H(e, I).$$
Set $\Lambda^{\sup}=\sup_{\lambda} \Lambda_{\max}(\lambda)$ and $\Lambda^{\inf}=\inf_{\lambda}\Lambda_{\min}(\lambda)$.
Assume further that
\eqRef{sec2.4}
\Lambda^{\sup}<\infty.
\end{equation}
It follows easily from (\ref{sec2.4}), Condition A and Condition B (ii) that $H(e, e\otimes e)=0.$
\vskip 1em
In this work, the requirement (\ref{sec2.4}) will apply through out. For some of the results, we will require additionally that
$$\Lambda^{\inf}>-\infty.\qquad\qquad \Box$$
\vskip 1em
We now present examples of operators that satisfy Conditions A, B and C, and include some observations.
\vskip 1em
\begin{rem}\label{sec2.40} (i) An example of an operator that satisfies
Conditions A, B and C is
$$
H_p(q, X)=|q|^p \{ |q|^2Tr(X)-q_iq_jX_{ij} \},\;p\ge 0,\;\;\forall (q, X)\in \mathbb{R}^n\times S^{n\times n},$$
where $Tr(X)$ is the trace of $X$. Clearly,
$$H_p(Du, D^2u)=|Du|^p\left( |Du|^2\Delta u-\Delta_\infty u \right).$$
Thus, for any $c\in \mathbb{R}$,
\begin{eqnarray*}
H_p(q, X+cq\otimes q)=|q|^p\left[ |q|^2Tr(X)+c|q|^4-q_iq_jX_{ij}-c|q|^4 \right]=H_p(q, X).
\end{eqnarray*}
In particular,
\begin{eqnarray*}
H_p(e, \lambda e\otimes e \pm I)=H_p(e, \pm I)=\pm(n-1),\;\;\mbox{for any}\; \lambda\in \mathbb{R}.
\end{eqnarray*}
Note that $k=k_1+1\ge 1$, see (\ref{sec2.51}) below. A closely allied example is $H(Du, D^2u)=|Du|^4\Delta_pu-(p-1)|Du|^p\Delta_\infty u.$
(ii) A second example can be constructed as follows. Let $\mu_i=\mu_i(X),\;i=1,2,\cdots,n$ be the eigenvalues of any $X\in S^{n\times n}$. We order these as $\mu_1\ge \mu_2\ge \cdots\ge \mu_n$. Define
$$H^m_{p}(q,X)=|q|^p \left(\sum_{i=m}^n \mu_i(X) \right),\quad \mbox{$p\ge 0$ and $2\le m<n$.} $$
Clearly, $H$ satisfies Conditions A and B, $H^m_{p}(e,\pm I)=\pm(n-m+1)$.
Observe that det$(e\otimes e)=0$ and $(e\otimes e)^2=e\otimes e$ and $(e\otimes e-\mu I)x=0\, (x\perp e)$ if and only if $\mu=0$ or $\mu=1\, (x\parallel e)$ implying that the eigenvalues of $e\otimes e$ are $0$ (multiplicity $n-1$) and $1$. Thus,
the eigenvalues of $\lambda e\otimes e+I$ are $1$ (multiplicity $n-1$) and $\lambda+1$. Similarly, the eigenvalues of $\lambda e\otimes e-I$ are $-1$ (multiplicity $n-1$) and $\lambda-1.$ Thus,
$$\left\{ \begin{array}{ccc}H^m_{p}(e,\lambda e\otimes e+I)=n+1-m,&\lambda\ge 0,\\ H^m_p(e, \lambda e\otimes e-I)=\lambda-(n-m+1), & \lambda\le 0. \end{array}\right. $$
Some of our results, in particular, the maximum principle in Theorem \ref{sec2.6} given below, hold for this operator. The case $m=1$ (Laplacian) is included in \cite{BM6}.
Observe that $k=p+1\ge 1$ in this case.
\vskip 1em
(iii) If $H$ is odd in $X$ i.e., $H(q,-X)=-H(q,X)$ then (\ref{sec2.4}) shows that $H(e, \lambda e\otimes e+I)=-H(e, -\lambda e\otimes e-I)$ and $\Lambda^{\sup}=-\Lambda^{\inf}<\infty$. Clearly,
$H(e,\pm e\otimes e)=0$.
\vskip 1em
(iv) We record a simple observation. If $k_1=0$ i.e., $k=1$, then $H(e, X)= H(e/s, X)$, for any $s>0$. Thus, $H(q, X)=H(0,X)=H(X)$. \quad $\Box$
\end{rem}
\vskip 1em
We introduce some further notation. Set $\mathbb{R}^n_T=\mathbb{R}^N\times (0,T)$. Let $\chi:(0,T)\rightarrow \mathbb{R}$ be a bounded continuous function and, for some $m\in\mathbb{R}$ (to be specified later)
$Z:[m, \infty)\rightarrow [0,\infty)$ be a non-increasing continuous function.
For $\sigma\ge 0$, set
\eqRef{sec2.5}
\mathcal{P}_\sigma(t,u, u_t, Du, D^2u)=H(Du, D^2u+Z(u)Du\otimes Du)+\chi(t)|Du|^\sigma-u_t.
\end{equation}
We assume through out that $H$ satisfies Conditions A, B and C.
Define
\eqRef{sec2.51}
k=k_1+1\quad\mbox{and}\quad \gamma=k+1=k_1+2.
\end{equation}
Clearly, $\gamma\ge 2$ and if $k=1$ then $k_1=0$ and $\gamma=2$. Next, define
\eqRef{sec2.52}
\forall \;\sigma>1,\;\sigma^*=\frac{\sigma}{\sigma-1},\;\;\mbox{and,}\;\;\forall\;k>1,\;\gamma^*=\frac{\gamma} {k-1}=\frac{\gamma}{\gamma-2}.
\end{equation}
\vskip 1em
For a fixed $z\in \mathbb{R}^n$ and $\forall x\in \mathbb{R}^n$, set $r=|x-z|.$ Also, define $B^R_T=\{(x,t): |x-z|\le R,\;0<t<T\}.$ Let $\mathcal{P}_\sigma$ be as defined in (\ref{sec2.5}).
\vskip 1em
We first state the results for $k>1$ or equivalently for $\gamma>2$.
\vskip 1em
\begin{thm}\label{sec2.6}{(Maximum Principle)} Let $0<T<\infty$, $h:\mathbb{R}^n\rightarrow \mathbb{R}$ be continuous with $\sup_{\mathbb{R}^n} h(x)<\infty$ and, for some $m$, $Z:[m, \infty)\rightarrow [0,\infty)$ be non-increasing and continuous.
Suppose that (\ref{sec2.4}) holds, i.e., $\Lambda^{\sup}<\infty$. Let $u\in usc(\mathbb{R}^n_T),\;\inf u>m,$ solve
$$\mathcal{P}_\sigma(t,u, u_t, Du, D^2u)\ge 0\;\mbox{in $\mathbb{R}^n_T,$ and $u(x)\le h(x),\;\forall x\in \mathbb{R}^n$}.$$
Let $\gamma^*$ and $\sigma^*$ be as in (\ref{sec2.52}). Suppose that $\sup_{B^R_T}u(x,t)= o( R^{\beta}),$ as $R\rightarrow \infty$.
Then the following hold.
\vskip 1em
(a) If $0\le \sigma\le \gamma/2$ and $\beta=\gamma^*$ then
$$\sup_{\mathbb{R}^n_T} u(x,t)\le \left\{\begin{array}{ccc}\sup_{\mathbb{R}^n} h(x)+t(\sup_{[0,T]}|\chi(t)| ), & \sigma=0,\\ \sup_{\mathbb{R}^n} h(x),& 0<\sigma\le \gamma. \end{array}\right.
$$
(b) If $\sigma>\gamma/2$ and $\beta=\sigma^*$ then
$$\sup_{\mathbb{R}^n_T} u(x,t)\le \sup_{\mathbb{R}^n} h(x).\;\;\;\;\;\;\;\;\Box$$
\end{thm}
Observe that if $m=-\infty$ then the restriction $\inf u>m$ may be dropped. Also, note that if $\sigma=\gamma/2$ we get $\sigma^*=\gamma/(\gamma-2)=\gamma^*$.
\vskip 2em
\begin{thm}\label{sec2.7}{(Minimum Principle)} Let $0<T<\infty$, $h:\mathbb{R}^n\rightarrow \mathbb{R}$ be a continuous function, with $\inf_{\mathbb{R}^n} h(x)>-\infty$, and $Z:(-\infty, \infty)\rightarrow [0,\infty)$ be a non-increasing continuous function. We assume that $\Lambda^{\sup}<\infty$.
Let $u\in lsc(\mathbb{R}^n_T)$ solve
$$\mathcal{P}_\sigma(t,u, u_t, Du, D^2u)\le 0,\;\;\mbox{in $\mathbb{R}^n_T$ and $u(x)\ge h(x),\;\forall x\in \mathbb{R}^n$}.$$
Let $\gamma^*$ and $\sigma^*$ be as in (\ref{sec2.52}). Suppose that $\sup_{ B_T^R}(-u(x,t))= o( R^\beta)$ as $R\rightarrow \infty$. Then the following hold.
(a) If $0\le \sigma\le \gamma/2$ and $\beta=\gamma^*$ then
$$\inf_{\mathbb{R}^n_T} u(x,t)\ge \left\{ \begin{array}{ccc} \inf_{\mathbb{R}^n} h(x)-t\left(\sup_{[0,t]}| \chi(t)| \right),& \sigma=0,\\ \inf_{\mathbb{R}^n} h(x), & 0<\sigma\le \gamma. \end{array}\right.$$
(b) If $\sigma>\gamma/2$ and $\beta=\sigma^*$ then
$$\inf_{\mathbb{R}^n_T} u(x,t) \ge \inf_{\mathbb{R}^n} h(x).\;\;\;\;\;\;\;\;\Box$$
\end{thm}
We impose no restrictions on $\Lambda^{\inf}$ for Theorem \ref{sec2.7}.
\vskip 2em
We now state analogous results for $k=1$ i.e, $\gamma=2$. See Remark \ref{sec2.40} (iv).
The statement that, for some $s>0$, $w(r)=e^{o(r^s)}$ as $r\rightarrow \infty$, will mean
that $\log v^+=o(r^s)$ as $r\rightarrow \infty$, where $v^+=\max(v,0).$
\vskip 1em
\begin{thm}\label{sec2.9}{(Maximum Principle)} Let $0<T<\infty$, $h:\mathbb{R}^n\rightarrow \mathbb{R}$ be continuous with $\sup_{\mathbb{R}^n} h(x)<\infty$.
For some $m$, let $Z:[m, \infty)\rightarrow [0,\infty)$ be non-increasing and continuous. Suppose that (\ref{sec2.4}) holds, i.e., $\Lambda^{\sup}<\infty$.
Let $u\in usc(\mathbb{R}^n_T)$, $\inf u>m$, solve
$$H(D^2u+Z(u)Du\otimes Du)-u_t\ge 0\;\mbox{in $\mathbb{R}^n_T,$ and $u(x)\le h(x),\;\forall x\in \mathbb{R}^n$}.$$
Let $\sigma^*$ be as in (\ref{sec2.52}). Then the following hold
\vskip 1em
(a) Suppose that $\sigma=0$. If $\sup_{B^R_T}u(x,t)= e^{o(R^2)},$ as $R\rightarrow \infty$, then
$$u(x,t)\le \sup_{\mathbb{R}^n}h(x)+\left( \sup_{(0,T)}\chi(t) \right) t,\;\;\forall (x,t)\in \mathbb{R}^n_T.$$
\vskip 1em
(b) Let $0<\sigma\le 1$. If $\sup_{B^R_T}u(x,t)= e^{o(R)},$ as $R\rightarrow \infty$ then
$$u(x,t)\le \sup_{\mathbb{R}^n}h(x)+K(1-\sigma)\left( \sup_{(0,T)} \chi(t)\right),$$
where $K=K(\alpha, \Lambda^{\sup},\sigma, T)$.
\vskip 1em
(c) Let $1<\sigma<\infty$ and assume that $\sup_{ B_T^R}u(x,t)= o(R^{\sigma^*}),$ as $R\rightarrow \infty$. Then
$$u(x,t)\le \sup_{\mathbb{R}^n}h(x). \qquad \Box$$
\end{thm}
\vskip 1em
We now present a minimum principle. Note that the condition $\Lambda^{\inf}>-\infty$ is needed only for parts (a) and (b) of the theorem. Part (c) of the theorem holds without this restriction.
\vskip 1em
\begin{thm}\label{sec2.10}{(Minimum Principle)} Let $0<T<\infty$, $h:\mathbb{R}^n\rightarrow \mathbb{R}$ be continuous, with $\sup_{\mathbb{R}^n} h(x)<\infty$, and
$Z:(-\infty, \infty)\rightarrow [0,\infty)$ be non-increasing and continuous. We assume that $\Lambda^{\sup}<\infty$.
Let $u\in usc(\mathbb{R}^n_T)$ solve
$$H(D^2u+Z(u)Du\otimes Du)-u_t\le 0\;\mbox{in $\mathbb{R}^n_T,$ and $u(x)\ge h(x),\;\forall x\in \mathbb{R}^n$}.$$
Assume for parts (a) and (b) that $\Lambda^{\inf}>-\infty$. Let $\sigma^*$ be as in (\ref{sec2.52}). Then the following hold.
\vskip 1em
(a) Suppose that $\sigma=0$. If $\sup_{B^R_T}(-u(x,t))= e^{o(R^2)},$ as $R\rightarrow \infty$, then
$$u(x,t)\ge \inf_{\mathbb{R}^n}h(x)-t\left( \sup_{(0,T)}\chi(t) \right) ,\;\;\forall (x,t)\in \mathbb{R}^n_T.$$
\vskip 1em
(b) Let $0<\sigma\le 1$. If $\sup_{B^R_T}(-u(x,t))= e^{o(R)},$ as $R\rightarrow \infty$ then
$$u(x,t)\ge \inf_{\mathbb{R}^n}h(x).$$
\vskip 1em
(c) Let $1<\sigma<\infty$ and assume that $\sup_{B_T^R}(-u(x,t))= o(R^{\sigma^*}),$ as $R\rightarrow \infty$. Then
$$u(x,t)\ge \inf_{\mathbb{R}^n}h(x). \qquad \Box$$
\end{thm}
\vskip 1em
Finally, we present similar results for a class of doubly nonlinear equations of the type
$$H(Du, D^2u)-f(u)u_t=0,\;\;\mbox{in $\mathbb{R}^n_T,$\; with $u(x,0)=g(x),\;\forall x\in \mathbb{R}^n$.} $$
If $k=1$, we assume that $f\equiv 1$ and the differential equation then reads
\eqRef{sec2.70}
H(D^2u)-u_t=0,\;\;\mbox{in $\mathbb{R}^n_T$ with $u(x,0)=g(x),\;\forall x\in \mathbb{R}^n$.}
\end{equation}
The above is not doubly nonlinear but is contained in our work.
It is to be noted that the afore stated theorems are used to obtain a maximum principle for these equations. The minimum principle, however, requires a different treatment.
If $k>1$ we take $f: [0,\infty)\rightarrow[0,\infty)$ to be an increasing $C^1$ function such that
$f^{1/(k-1)}$ is concave and consider equations of the type
\eqRef{sec2.71}
H(Du,D^2u)-f(u) u_t=0,\;\;\mbox{in $\mathbb{R}^n_T$ with $u(x,0)=g(x),\;\forall x\in \mathbb{R}^n$,}
\end{equation}
where $u>0$.
For $k>1$, let $F$ be a primitive of $f^{-1/(k-1)}$. Since $f(s)>f(0)\ge 0,\;\forall s>0$, we consider the following two situations:
\eqRef{sec2.71.0}
\mbox{(i)}\;\; \lim_{\varepsilon \rightarrow 0^+} F(1)-F(\varepsilon) <\infty, \quad\mbox{and}\quad \mbox{(ii)}\;\;\lim_{\varepsilon \rightarrow 0^+} F(1)-F(\varepsilon) =\infty.
\end{equation}
\vskip 1em
We set $\chi(t)\equiv0$ in Theorems \ref{sec2.6} and \ref{sec2.10}.
\vskip 1em
\begin{thm}\label{sec2.8}
Let $f: [0,\infty)\rightarrow [0,\infty)$ be a $C^1$ increasing function and $g:\mathbb{R}^n\rightarrow (0,\infty)$, continuous, be such that $0<\inf_x g(x)\le \sup_x g(x)<\infty.$ Assume that
$\Lambda^{\sup}<\infty$.
\noindent (a) Maximum Principle: Let $k>1$ and $f^{1/(k-1)}$ be a concave function. Suppose that
$\phi:\mathbb{R}\rightarrow [0,\infty)$ is a $C^2$ increasing function such that
$\phi^{\prime}(\tau)= f(\phi(\tau))^{1/(k-1)}.$ Recall $\gamma^*$ from (\ref{sec2.52}).
If $u\in usc( \overline{\mathbb{R}^n_T}),\;u>0,$ solves
$$H(Du, D^2u)-f(u) u_t\ge 0,\;\mbox{in $\mathbb{R}^n_T$ and $u(x,0)\le g(x),\;\forall x\in \mathbb{R}^n$,}$$
and $\sup_{B^R_T} u(x,t)\le \phi( o(R^{\gamma^*}))$, as $R\rightarrow \infty$, then
$$\sup_{ \mathbb{R}^n_T}u(x,t)\le \sup_{\mathbb{R}^n} g(x).$$
Let $k=1$ and $f\equiv 1$, i.e, $H(D^2u)-u_t\ge 0$. If $\sup_{B_T^R} u(x,t)\le e^{o(R^2)}$, as $R\rightarrow \infty$, then
$\sup_{ \mathbb{R}^n_T}u(x,t)\le \sup_{\mathbb{R}^n}g(x).$
\vskip 1em
\noindent (b) Minimum Principle: Let $k>1$, $f$ and $\phi$ be as in part (a).
Suppose that $u\in lsc( \overline{\mathbb{R}^n_T}),\;u>0,$ solves
$$H(Du, D^2u)-f(u) u_t\le 0,\;\mbox{in $\mathbb{R}^n_T$ and $u(x,0)\ge g(x),\;\forall x\in \mathbb{R}^n$.}$$
If condition (\ref{sec2.71.0})(i) holds, i.e, $\lim_{\varepsilon\rightarrow 0^+}F(1)-F(\varepsilon)<\infty$ then
$$u(x,t)\ge \inf_{\mathbb{R}^n} g(x),\;\forall(x,t)\in \mathbb{R}^n_T.$$
If condition (\ref{sec2.71.0})(ii) holds, i.e., $\lim_{\varepsilon\rightarrow 0^+} F(1)-F(\varepsilon)=\infty$, and $\inf_{B_T^R}u(x,t)\ge \phi(-o(R^{\gamma^*}))$ as $R\rightarrow \infty$ then
$$u(x,t)\ge \inf_{\mathbb{R}^n} g(x),\;\forall(x,t)\in \mathbb{R}^n_T.$$
\vskip 1em
Suppose $k=1$ and $f\equiv 1$, i.e., $H(D^2u)-u_t\le 0$. If $\inf_{B^R_T} u(x,t)\ge -e^{o(R^2)}$, as $R\rightarrow \infty$, then
$$u(x,t)\ge \inf_{\mathbb{R}^n}g(x),\;\;\forall(x,t)\in \mathbb{R}^n_T.$$
\end{thm}
\vskip 2em
\section{Preliminaries}
\vskip 1em
In this section, we present some definitions, lemmas and remarks we will use to prove the main results. Fix $z\in \mathbb{R}^n$ and set $r=|x-z|,\;\forall x\in \mathbb{R}^n$. A unit vector in $\mathbb{R}^n$ is denoted by $e=(e_1,e_2,\cdots,e_n)$.
\vskip 1em
We begin with an elementary remark that will be used frequently in our work.
\begin{rem}\label{sec3.2} Assume that $w:\mathbb{R}^n\times [0,\infty)\rightarrow \mathbb{R}$ is a $C^{1}$ function in $x$ and $t$ and $C^2$ in $x$ except, perhaps, at $x\ne z $. We get, for $r\ne 0$,
\begin{eqnarray}\label{sec3.3}
&&H(Dw, D^2w+Z(w) Dw\otimes Dw)\nonumber\\
&&\quad\qquad\qquad\qquad\qquad\qquad=H\left( w_r e, \; \left( \frac{w_r}{r} \right) I +\left(w_{rr} -\frac{w_r}{r}+ (w_r)^2 Z(w) \right)e\otimes e \; \right),
\end{eqnarray}
where $e=(e_1,e_2,\cdots, e_n)$ with $e_i=(x-z)_i/r,\;\forall i=1,2,\cdots, n.$ Let $\kappa:(0,T)\rightarrow [0,\infty)$ be a $C^1$ function.
{\bf Case (a) ($w_r\ge 0$):} We apply Condition B, in (\ref{sec2.2}), to (\ref{sec3.3}). Factor $w_r$ from the first entry, $w_r/r$ from the second and use $k=k_1+1$ to get
\begin{eqnarray}\label{sec3.4}
&&H(Dw, D^2w+Z(w) Dw\otimes Dw) \nonumber \\
&&\qquad\qquad\qquad\qquad= \left( \frac{ w_r^k}{ r }\right) H\left( e, I +\left( \frac{r w_{rr}}{w_r}-1 + r w_r Z(w)\right)e\otimes e\right), \;\forall r>0.
\end{eqnarray}
\vskip 1em
If $w(x,t)=\kappa(t) v(r)$, with $v^\prime(r)\ge 0$, then (\ref{sec3.4}) implies that, in $r>0$,
\begin{eqnarray}\label{sec3.6}
&&H(Dw, D^2w+Z(w) Dw\otimes Dw) \nonumber\\
&&\qquad\qquad=\frac{ (\kappa(t) v^\prime(r))^k}{r} H\left( e, \;I +\left( \frac{r v^{\prime\prime}(r)}{v^\prime(r)}-1 + r\kappa(t) v^{\prime}(r) Z(w) \; \right)e\otimes e\;\right).
\end{eqnarray}
\vskip 1em
{\bf Case (b) ($w_r\le 0$):} Clearly, (\ref{sec3.3}) leads to
\begin{eqnarray}\label{sec3.7}
&&H(Dw, D^2w+Z(w) Dw\otimes Dw) \nonumber\\
&&\qquad\qquad\qquad\qquad=\frac{ |w_r|^k}{ r } H\left( e, \left( 1-\frac{r w_{rr}}{w_r} + r |w_r| Z(w)\right)e\otimes e-I\right),\;\;\forall r>0.
\end{eqnarray}
If $w(x,t)=\kappa(t)v(r)$ and $v^{\prime}(r)\le 0$ then (\ref{sec3.7}) leads to the following analogue of (\ref{sec3.6}):
\begin{eqnarray}\label{sec3.8}
&&H(Dw, D^2w+Z(w) Dw\otimes Dw) \nonumber\\
&&\quad\qquad=\frac{ (\kappa(t)|v^\prime(r)|)^k}{ r } H\left( e, \;\left(\; r|v^{\prime}(r)|\kappa(t)Z(w)\;+1 - \frac{rv^{\prime\prime}(r)}{v^\prime(r)}\right)e\otimes e -I \right).
\quad \Box
\end{eqnarray}
\end{rem}
\vskip 1em
The following lemma was proven in \cite{BM6}.
\begin{lem}\label{sec3.10} Let $\beta,\;\bar{\beta}$ be such that $1<\bar{\beta}<\beta$ and $R>0$. Fix $z\in \mathbb{R}^n$, set $r=|x-z|$ and
define
$$v(r)=\int_0^{ r^\beta } \frac{1}{1+\tau^p} \;d\tau,\quad \mbox{where}\;\;p=\frac{\beta-\bar{\beta}}{\beta}. $$
\vskip 1em
Then (i)\;$0<p<1$,\quad (ii)\;$(1-p)\beta=\bar{\beta}$, and
\begin{eqnarray*}
\mbox{ (iii)\;\;$\forall r\ge 0$},\quad \frac{ r^{\beta} }{ 1+r^{\beta p} }\le v(r)\le \min\left( r^{ \beta},\; \frac{ r^{\bar{\beta}} }{1-p} \right).
\end{eqnarray*}
If $R>1$ then
\begin{eqnarray*}
\mbox{(iv)}\;\;\frac{\beta}{2\bar{\beta}}=\frac{1 }{2(1-p)}\le \frac{ v(r)-v(R) }{r^{\bar{\beta}}-R^{\bar{\beta}}} \le \frac{1 }{1-p}=\frac{\beta}{\bar{\beta}}, \quad \forall r\ge R.
\end{eqnarray*}
Moreover, $v^{\prime}(r)=\beta r^{\beta-1}/(1+ r^{p\beta})$ implying that
\begin{eqnarray*}
&& \mbox{(v)}\;\;v^{\prime}(r)\le\beta \min\left( r^{ \bar{\beta}-1},\;r^{\beta-1}\right),\;\mbox{(vi)}\;\;\frac{ (v^{\prime}(r))^k }{ r}\le \beta^k\min\left( r^{k\beta-\gamma},\;r^{k\bar{\beta} -\gamma} \right),\\
&&\mbox{and (vii)}\;\;v^{\prime\prime}(r) =\beta r^{\beta-2}
\left( \frac{ (\beta-1)+(\bar{\beta}-1) r^{p\beta} }{ ( 1+r^{p\beta})^2 } \right).\
\end{eqnarray*}
\end{lem}
{\bf Comment:} Parts (iii) and (iv) of Lemma \ref{sec3.10} show that $v(r)$ grows like $r^\beta$ near $r=0$ and like $r^{\bar{\beta}}$ for large values of $r$. Since $\beta\ge \bar{\beta}$, one can design the function to decay fast enough at $r=0$ so as to be differentiable while its growth rate for large values of $r$ may be slower.
\begin{proof} Parts (i)-(iii) follow quite readily. For part (iv), we take $R>1$ and write
\begin{eqnarray*}
v(r)=\int_0^{r^\beta} \left(1+\tau^p \right)^{-1} d\tau=v(R)+\int_{R^\beta}^{r^\beta} \left( 1+\tau^p \right)^{-1} d\tau
\end{eqnarray*}
We estimate $(2\tau^p)^{-1}\le (1+\tau^p)^{-1}\le \tau^{-p}$, for $\tau\ge 1$, and use this in the second integral to obtain part (iv). For part (v), note that
$1+r^{p\beta}\ge \min(1,\;r^{p\beta}).$ Using part (ii) yields the claim. Part (vi) follows by recalling that $\gamma=k+1=k_1+2$.
Next,
$$v^{\prime\prime}(r)=\beta\left[ \frac{ (\beta-1)r^{\beta-2} }{1+r^{p\beta} }- \frac{ p\beta\; r^{p \beta +\beta-2} } { \left( 1+r^{p\beta} \right)^2 } \right].$$
A simple calculation leads to part (vii).
\end{proof}
\vskip 1em
The following remark is useful for the construction of the auxiliary functions. The values of $\bar{\beta}$ and $\beta$, used in the remark, are motivated by the work in Sections 4 and 5.
\begin{rem}\label{sec3.11} For Sub-Part (iv) of Part I in Section 4, we take $k>1$ (i.e, $\gamma>2$) and $\sigma>\gamma/2$. We set
$$\beta=\gamma^*=\gamma/ (\gamma-2)\quad\mbox{ and}\quad \bar{\beta}=\sigma^*=\sigma/(\sigma-1).$$
Then $p=(\gamma^*-\sigma^*)/\gamma^*=(2\sigma-\gamma)/\gamma(\sigma-1)>0.$ Clearly, $0<p<1$.
We take
\begin{eqnarray*}
v(r)=\int_0^{ r^{\gamma^*}} \frac{1}{1+\tau^p}\;d\tau,\quad\mbox{where}\; p=1-\frac{\sigma^*}{\gamma^*}=\frac{2\sigma-\gamma}{\gamma(\sigma-1)}.
\end{eqnarray*}
From Lemma \ref{sec3.10}, (i) $0<p<1$, \;\;(ii) $(1-p)\gamma^*=\sigma^*$,
\begin{eqnarray*}
&&\mbox{(iii) for $r\ge 0$,}\;\;\; \frac{ r^{\gamma^*} }{ 1+r^{\gamma^* p} }\le v(r)\le \min\left( r^{ \gamma^* },\; \frac{\gamma^* r^{\sigma^*} }{\sigma^*} \right), \\
&&\mbox{(iv) for any $R>1$,}\;\;\frac{\gamma^*}{2\sigma^*}=\frac{1 }{2(1-p)}\le \frac{v(r)-v(R)}{r^{\sigma^*}-R^{\sigma^*}}\le \frac{1 }{1-p}=\frac{\gamma^*}{\sigma^*}, \;\; \forall r\ge R.
\end{eqnarray*}
Moreover, $v^{\prime}(r)=\gamma^* r^{\gamma^*-1}/(1+ r^{p\gamma^*}),$
\begin{eqnarray*}
&&\mbox{(v)}\;\;v^{\prime}(r)\le \gamma^* \min\left( r^{ \sigma^*-1},\;r^{\gamma^*-1}\right),\;\;\mbox{(vi)}\;\frac{ (v^{\prime}(r))^k }{ r}\le {\gamma^*}^k\min\left( r^{k\gamma^*-\gamma},\;r^{k\sigma^* -\gamma} \right),\\\ &&\mbox{and (vii)}\;\;v^{\prime\prime}(r) =\gamma^* r^{\gamma^*-2}
\left( \frac{ (\gamma^*-1)+(\sigma^*-1) r^{p\gamma^*} }{ ( 1+r^{\gamma^* p})^2 } \right).\quad \Box
\end{eqnarray*}
\end{rem}
\vskip 1em
\begin{rem}\label{sec3.201} The super-solutions and sub-solutions make use of functions that involve a $C^1$ function of $t$ and a $C^{1,\alpha}$ (for some $\alpha>0$) function of $v(r)$. See the functions discussed in Remark \ref{sec3.11}. The calculations done in the remark hold in the sense of viscosity at $r=0.$ The verification can be found in \cite{BM6}.
\quad $\Box$
\end{rem}
We recall a comparison principle needed for our work, see \cite{CIL}. See also \cite{BM5} and \cite{BM6}.
\vskip 1em
Let $F:\mathbb{R}^+\times \mathbb{R}\times \mathbb{R}^n\times S^n\rightarrow\mathbb{R}$ be continuous. Suppose that $F$ satisfies $\forall X,\;Y\in S^n$, with $X\le Y$, that
\begin{eqnarray}\label{sec3.17}
\mbox{$F(t, r_1, q, X)\le F(t, r_2, q, Y),$ $\forall (t,q)\in \mathbb{R}^+\times \mathbb{R}^n$ and $r_1\ge r_2$.}
\end{eqnarray}
In this work, $F(t, r, q, X)=H(q, X+Z(r) q\otimes q)+\chi(t)|q|^\sigma,$ where $Z$ is a non-increasing continuous function, $\sigma\ge 0$ and $H$ satisfies Conditions A, B and C.
\vskip 1em
Let $\Omega\subset \mathbb{R}^n$ be a bounded domain, $\Omega_T=\Omega\times (0,T)$ and $P_T$ be the parabolic boundary of $\Omega_T$.
\begin{lem}\label{sec3.18}{(Comparison principle)} Let $F$ satisfy (\ref{sec3.17}) and $\hat{f}:\mathbb{R}^+\rightarrow \mathbb{R}^+$ be a bounded continuous function. For some $m$, let
$Z:[m ,\infty)\rightarrow \mathbb{R}$ be a non-increasing continuous function.
Let $u\in usc(\Omega_T\cup P_T)$ and $v\in lsc(\Omega_T\cup P_T)$ such that
$\inf(\inf u, \inf v)>m$. Suppose that $u$ and $v$ solve
\begin{eqnarray*}
&& F(t,u, Du, D^2u+Z(u)Du\otimes Du)- \hat{f}(t)u_t\ge 0,\\
&&\mbox{and}\quad F(t,v, Dv, D^2v+Z(v)Dv\otimes Dv)- \hat{f}(t)v_t\le 0,\quad \mbox{in $\Omega_T$.}
\end{eqnarray*}
If $\sup_{P_T}v<\infty$ and $u\le v$ on $P_T$ then $u\le v$ in $\Omega_T$. \quad $\Box$
\end{lem}
Next, we discuss a change of variables that is used in the proof of Theorem \ref{sec2.8} for doubly nonlinear equations of the kind
\eqRef{sec3.190}
H(Du, D^2u)-f(u) u_t=0,\;\;\mbox{in $\mathbb{R}^n_T$, $u>0$, with $u(x,0)=g(x),\;\forall x\in \mathbb{R}^n$. }
\end{equation}
\begin{rem}\label{sec3.19} Let $f:[0,\infty)\rightarrow [0,\infty)$ be an increasing $C^1$ function. Suppose that $k>1$ and
$f^{1/(k-1)}$ is concave.
Let $I$ be either $[0,\infty)$ or $(-\infty, \infty)$, see (b) and (c) below. We select
$\phi: I\rightarrow [0,\infty)$, an increasing $C^2$ function, such that
$$\phi^{\prime}(\tau)= f(\phi(\tau))^{1/(k-1)},\;\;\forall \tau \in I,\;\;\mbox{or}\;\;\int ^{\phi(\tau)}_{\phi(\tau_0)} f^{-1/(k-1)}(\theta) \; d\theta=\tau-\tau_0.$$
We define the change of variable $u=\phi(v)$ by
\eqRef{sec2.71}
v(u)-v(u_0)=\phi^{-1}(u)=\int^u_{u_0} f^{-1/(k-1)}(\theta)\;d\theta,\;\;u\ge u_0,
\end{equation}
for some $u_0\ge 0$.
We discuss some examples. Let $\alpha>0$, $a\ge 0$ and $f(s)=(s+a)^\alpha,\;\forall s\ge 0.$ Then $f(s)^{1/(k-1)}$ is concave if $\alpha\le k-1$. Set $c_k=(k-1-\alpha)/(k-1).$ We may take $u_0=0$ in (\ref{sec2.71}), we get that
$$u=\phi(v)=\left\{ \begin{array}{ccc} \left[ c_kv+a^{c_k})\right]^{1/c_k }-a ,& 0<\alpha<k-1,\;a\ge 0,\\ a e^v-a, & \alpha =k-1,\;a>0. \end{array}\right.$$
See also part (b) below.
If $a=0$, take $f(s)=s^{k-1}$ then $u=b e^v$ for any $b>0$. But, $u_0\ne 0$, see part (c).
\vskip 1em
We make some observations about (\ref{sec2.71}).
{\bf (a)} It is clear that $v$ is an increasing concave function of $u$. The concavity follows since $f$ is non-decreasing. Since $v$ is increasing, $u$ is a convex function of $v$.
{\bf (b)} If the integral in (\ref{sec2.71}) is convergent for $u_0=0$ we then define
$$v=\phi^{-1}(u)=\int^u_0 f^{-1/(k-1)}(\theta)\;d\theta.$$
Thus, $v(0)=0$ and $v>0$.
We choose $I=[0,\infty)$ and $\phi:[0,\infty)\rightarrow [0,\infty)$.
This applies to examples like
$$f(s)=\left\{ \begin{array}{ccc} s^\alpha, & 0\le \alpha<k-1,\\ (s+a)^\alpha, & 0\le \alpha \le k-1,\end{array}\right.$$
where $a>0$.
\vskip 1em
{\bf (c)} If the integral in (\ref{sec2.71}) is divergent for $u_0=0$ then $v(u_0)\rightarrow -\infty$ as $u_0\rightarrow 0^+$.
In this case, we select a primitive
$$v=\phi^{-1}(u)=\int^u f^{-1/(k-1)}(\theta)\;d\theta.$$
We choose $I=(-\infty, \infty)$ and $\phi:(-\infty, \infty)\rightarrow (0,\infty)$.
This includes examples such as $f(s)=s^{k-1},\;(s+\log(s+1))^{k-1}$ etc.
\vskip 1em
{\bf (d)} We show that in parts (b) and (c), $v\rightarrow \infty$ if $u\rightarrow \infty$. Set $\nu(s)=f^{1/(k-1)}(s)$. Since $\nu(s)$ is concave in $(0,\infty)$, it is clear that,
for a fixed $\varepsilon>0$,
$$\nu(s)\le \nu(\varepsilon)+(s-\varepsilon) \nu^{\prime}(\varepsilon),\;\;s\ge \varepsilon.$$
Using (\ref{sec2.71}), we get that
$$v(u)=v(\varepsilon)+ \int_\varepsilon^u \frac{1}{\nu(s)}ds\ge v(\varepsilon)+ \int_\varepsilon^u \frac{1}{ \nu(\varepsilon)+(s-\varepsilon) \nu^{\prime}(\varepsilon) } ds.$$
The claim holds.
\vskip 1em
{\bf (e)} It is clear from (\ref{sec2.71}) that
$$\frac{\phi^{\prime\prime}(v)}{\phi^{\prime}(v)}= \left.\left( \frac{ d}{ds} f^{1/(k-1)}(s) \right) \right|_{\phi(v)},$$
and $\phi^{\prime\prime}(v)/\phi^{\prime}(v)$ is non-increasing in $v$ since $f^{1/(k-1)}$ is concave and $\phi(v)$ is increasing in $v$.
Suppose that there are constants $0<\omega_1\le \omega_2<\infty$ such that
\begin{equation}\label{eq:add}\omega_1\le \phi^{\prime\prime}(v)/ \phi^{\prime}(v) \le \omega_2.\end{equation}
Integrating from $s=0$ to any $s>0$, we get that,
$$\omega_1 s\le f^{1/(k-1)}(s)-f^{1/(k-1)}(0)\le \omega_2 s,\;\;\forall s\ge 0.$$
Since $f(0)\ge 0$, we get that, for some $\omega\ge 0$, $(\omega_1s+\omega)^{k-1}\le f(s)\le (\omega_2 s+\omega)^{k-1},\;\forall s\ge 0$.
If $\omega>0$ then we use $v$ as in part (b). If $\omega=0$ then we use part (c).
\vskip 1em
{\bf (f)} The change of variable $u=\phi(v)$, as given by (\ref{sec2.71}), transforms (\ref{sec3.190}) into
\begin{eqnarray*}
H(Dv, D^2v+Z(v) Dv\otimes Dv)-v_t=0\;\;\mbox{in $\mathbb{R}^n_T$ with $v(x,0)=\phi^{-1}(g(x)),\;\forall x\in \mathbb{R}^n$,}
\end{eqnarray*}
where $Z(v)=\phi^{\prime\prime}(v)/\phi^{\prime}(v)$, see Lemma 2.3 in \cite{BM5}. By part (e), $Z(v)$ is non-increasing in $v$ and the domain of $Z$ contains either $(0,\infty)$ or $(-\infty, \infty)$. $\Box$
\end{rem}
We now state a comparison principle for doubly nonlinear equations.
\begin{lem}\label{sec3.20} Let $T>0$ and $\Omega\subset \mathbb{R}^n$ be a bounded domain. Suppose that $k>1$ and $f:[0,\infty)\rightarrow [0,\infty)$ is a non-decreasing $C^1$ function such that $f^{1/(k-1)}$ is concave. Set $\Omega_T=\Omega\times (0,T)$ and $P_T$ to be the parabolic boundary of $\Omega_T$.
Let $u \in usc(\Omega_T)$, $v\in lsc(\Omega_T)$ and $u>0$ and $v>0$. Suppose that
\begin{eqnarray*}
H(Du, D^2u)-f(u)u_t\ge 0,\;\;\mbox{in $\Omega_T$, and}\;\;H(Dv, D^2v)-f(v) v_t\le 0,\;\;\mbox{in $\Omega_T$,}
\end{eqnarray*}
where $H$ satisfies conditions $A,\;B$ and $C$.
If $u\le v$ on $P_T$ then $u\le v$ in $\Omega_T$.
\end{lem}
\begin{proof} We employ Lemma \ref{sec3.18} and Remark \ref{sec3.19}. Let $u$ and $v$ be as in the statement of the theorem.
Set
$$F(\hat{s},s)=\int_s^{\hat{s}} f^{-1/(k-1)}(\theta) d\theta,\;\;\forall \hat{s}\ge s\ge 0.$$
We define $F(\hat u,0)=\lim_{s\rightarrow 0^+}F(\hat u,s)$, if it exists.
{\bf (i)} Suppose that $F(1,0)<\infty$ then we define
$$\bar{u}=\phi^{-1}(u)=F(u, 0)\;\;\mbox{and}\;\;\bar{v}=\phi^{-1}(v)=F(v,0).$$
By parts (a) and (b) of Remark \ref{sec3.19}, $\bar{u}>0$ and $\bar{v}>0$. Also, by part (f) of Remark \ref{sec3.19},
\begin{eqnarray*}
H(D\bar{u}, D^2\bar{u}+Z(\bar{u}) D\bar{u}\otimes D\bar{u})-\bar{u}_t\ge 0\;\;\mbox{and}\;\; H(D\bar{v}, D^2\bar{v}+Z(\bar{v}) D\bar{v}\otimes D\bar{v})-\bar{v}_t\le 0,
\end{eqnarray*}
in $\Omega_T$, where $Z(s)=\phi^{\prime\prime}(s)/\phi^{\prime}(s)$ is non-increasing in $s$. Note that the domain of $Z$ contains $(0,\infty)$. Using Lemma \ref{sec3.18}, $\bar{u}\le \bar{v}$ in $\Omega_T$ thus implying that $u\le v$ in $\Omega_T$.
\vskip 1em
{\bf (ii)} Suppose now that $F(1, 0)$ is divergent, see part (c) of Remark \ref{sec3.19}. Fix a primitive
$$F(s)=\int^s f^{-1/(k-1)}(\theta) d\theta,\;\;s>0.$$
Define
$\bar{u}=\phi^{-1}(u)=F(u)\;\;\mbox{and}\;\;\bar{v}=\phi^{-1}(v)=F(v).$
Then $-\infty<\bar{u},\;\bar{v}<\infty$ and by parts (e) and (f) of Remark \ref{sec3.19}, we get in $\Omega_T$,
\begin{eqnarray*}
H(D\bar{u}, D^2\bar{u}+Z(\bar{u}) D\bar{u}\otimes D\bar{u})-\bar{u}_t\ge 0\;\;\mbox{and}\;\; H(D\bar{v}, D^2\bar{v}+Z(\bar{v}) D\bar{v}\otimes D\bar{v})-\bar{v}_t\le 0,
\end{eqnarray*}
where the domain of $Z$ is $(-\infty,\infty)$. Using Lemma \ref{sec3.18}, $\bar{u}\le \bar{v}$ in $\Omega_T$ thus implying that $u\le v$ in $\Omega_T$.
\end{proof}
\section{ \bf Super-solutions}
\vskip 1em
In this section, we construct super-solutions of (\ref{sec1.3}) and these are used to prove Theorems \ref{sec2.6}, \ref{sec2.9} and \ref{sec2.8}. We have divided our work into two parts. Part I addresses the case $k>1$(or $\gamma>2$) and Part II discusses the case $k=1$ or $\gamma=2$. In each part, the work is further sub-divided to address various situations based on the values of $\sigma$. Since the auxiliary functions are non-negative, we assume that the domain of $Z$ is at least $(0,\infty)$, see discussion below.
Part I has four sub-parts:
(i) $\sigma=0$, (ii) $0<\sigma<\gamma/2$, (iii) $\sigma=\gamma/2$ and (iv) $\sigma>\gamma/2$, and Part II has three sub-parts: (i) $0\le \sigma\le 1$, (ii) $1<\sigma\le 2$, and (iii) $\sigma>2$.
We recall from (\ref{sec2.5}) that
\eqRef{sec4.0}
\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w):=H(Dw, D^2w+Z(w)Dw\otimes Dw)+\chi(t)|Dw|^\sigma-w_t,
\end{equation}
where $\sigma\ge 0$, and $Z(s)\ge 0$ and is a non-increasing continuous function of $s$.
Let $m<\min (0, \inf_{\mathbb{R}^n} h,\inf_{\mathbb{R}^n_T}u )$, where $h$ is the initial data in (\ref{sec1.3}) and $u$ is the given sub-solution. We assume that the domain of $Z$ is at least $[m,\infty)$.
Recall from (\ref{sec2.3}) and (\ref{sec2.4}) that $\Lambda^{\sup}=\sup_{\lambda} \left( \max_{|e|=1}H(e, I+\lambda e\otimes e) \right).$ We set
\eqRef{sec4.1}
\alpha=\sup_{[0,T]} |\chi(t)| \;\;\mbox{and}\;\; M=\max\left( \Lambda^{\sup}, 1\right).
\end{equation}
We also recall from (\ref{sec2.51}) and (\ref{sec2.52}) that
$$k=k_1+1,\;\; \gamma=k_1+2=k+1,\;\;\gamma\ge 2\;\;\mbox{and}\;\; \gamma^*=\frac{\gamma}{\gamma-2}\;\;\mbox{if $\gamma>2$.}$$
Moreover, $\gamma=2$ if and only if $k=1$($k_1=0$).
\vskip 1em
\noindent{\bf Super-solutions:}
\vskip 1em
\noindent {\bf Part I ($k>1$):} Since $\gamma>2$, we see that
\eqRef{sec4.2}
\gamma^*-1=\frac{2}{\gamma-2},\quad \gamma^*-2=\frac{4-\gamma}{\gamma-2}>-1\quad\mbox{and}\quad k\gamma^*-\gamma=\frac{\gamma}{\gamma-2}=\gamma^*.
\end{equation}
\vskip 1em
We start with the case $0\le \sigma\le \gamma/2$ and first carry out some calculations that will hold for the entire interval $[0,\gamma/2]$. We will then discuss the cases $\sigma=0$, $0<\sigma<\gamma/2$ and
$\sigma=\gamma/2$ separately.
Let $z\in \mathbb{R}^n$ be fixed, set $r=|x-z|,\;\forall x\in \mathbb{R}^n$, and define
\begin{eqnarray}\label{sec4.4.0}w(x,t)=at+ b(1+t) v(r), \;v^{\prime}(r)\ge 0,\;\;\forall(r,t)\in \mathbb{R}^n_T,\end{eqnarray}
where $a\ge 0$ and $0<b\le 1$ are to be determined. We do this in each of the three cases listed above and also calculate $\lim_{b\rightarrow 0} a,$ wherever it is meaningful.
Using (\ref{sec3.6}), (\ref{sec4.0}) and (\ref{sec4.1}), we get
\begin{eqnarray}\label{sec4.4}
&&\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w)\nonumber \\
&&\quad\qquad=\frac{ [ b(1+t) v^\prime(r) ]^k}{r} H\left( e, \;I +\left( \frac{r v^{\prime\prime}(r)}{v^\prime(r)}-1 + b(1+t) r v^{\prime}(r) Z(w) \; \right)e\otimes e\;\right) \nonumber\\
&&\quad\qquad\qquad\qquad+\chi(t) [ b(1+t) v^\prime(r) ]^\sigma-a-bv(r) \nonumber \\
&&\quad\qquad\le \frac{M [ b(1+T) ]^k v^\prime(r)^k}{r} + \alpha \left [ b(1+T) \right ]^\sigma (v^\prime(r))^\sigma-a-bv(r).
\end{eqnarray}
We use the above inequality in both Parts I and II.
\vskip 1em
For Part I, we take $v(r)=r^{\gamma^*}$. Using (\ref{sec4.2}) and $k=\gamma-1$ in (\ref{sec4.4}), we find that
\begin{eqnarray}\label{sec4.5.1}
&&\mathcal{P}_\sigma(t,w,w_t, Dw, D^2w)\nonumber\\
&&\qquad\qquad \le M [ b\gamma^*(1+T) ]^k \frac{(r^{\gamma^*-1} )^k }{r} + \alpha\left [ b \gamma^* (1+T)\right ]^\sigma \left( r^{\gamma^*-1}\right)^\sigma-a- br^{\gamma^*}\nonumber \\
&&\qquad\qquad\le M\left[ \gamma^* (1+T)\right ]^k ( b^k r^{\gamma^*} )
+ \alpha \left [ \gamma^* (1+T) \right ]^\sigma ( b^\sigma r^{ 2\sigma/(\gamma-2)} )- a- ( br^{\gamma^*} ).
\end{eqnarray}
In order to write more compactly, we set
$$E=M\left [\gamma^* (1+T)\right ]^k\quad \mbox{and}\quad F= \left[ \gamma^* (1+T) \right ]^\sigma.$$
Thus, (\ref{sec4.5.1}) reads
\begin{eqnarray}\label{sec4.40}
\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w)\le E (b^k r^{\gamma^*})+\alpha F ( b^\sigma r^{2\sigma/(\gamma-2)}) -a- (br^{\gamma^*}).
\end{eqnarray}
\vskip 1em
{\bf Sub-Part (i) ($\sigma=0$):} Taking $\sigma=0$ in (\ref{sec4.40}), we get that $F=1$ and
$$\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w)\le b(Eb^{k-1}-1) r^{\gamma^*}+\alpha-a.$$
Select $a=\alpha$ and $0<b<\min(1,E^{1-k})$. Clearly, $w(x,t)$ is a super-solution in $\mathbb{R}^n_T$ and
\eqRef{sec4.41}
w(x,t)=\alpha t+b(1+t)r^{\gamma^*}. \quad\qquad \Box
\end{equation}
\vskip 1em
{\bf Sub-Part (ii) ($0<\sigma<\gamma/2$):} Since $\gamma^*=\gamma/(\gamma-2)$, (\ref{sec4.40}) yields that
\begin{eqnarray}\label{sec4.42}
\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w)&\le& E b^k r^{\gamma^*}-br^{\gamma^*}+\alpha F b^\sigma r^{2\sigma/(\gamma-2)}-a \nonumber\\
&=&b r^{\gamma^*} \left( Eb^{k-1}-1+ \frac{ \alpha F b^{\sigma-1} }{ r^{(\gamma-2\sigma)/(\gamma-2) } } \right)-a.
\end{eqnarray}
We choose
\eqRef{sec4.43}
\left\{ \begin{array}{lcr} 0<b^{k-1}<\min\left(1, \;\; (4E)^{-1}\right),\;\;R= \left( 4 \alpha F b^{\sigma-1}\right)^{(\gamma-2)/(\gamma-2\sigma)},\\
\mbox{and}\;\;a= E b^k R^{\gamma^*}+\alpha F b^\sigma R^{2\sigma/(\gamma-2)}.
\end{array}\right.
\end{equation}
The choice for $a$ shows that $w$ is a super-solution in $B_R(z)\times (0,T)$. In $r\ge R$, using $0<\sigma<\gamma/2$ and
the selections for $b$ and $R$, stated in (\ref{sec4.43}), in (\ref{sec4.42}), we get
$$Eb^{k-1}-1+ \frac{\alpha F b^{\sigma-1} }{ r^{(\gamma-2\sigma)/(\gamma-2) } }\le -\frac{3}{4}+ \frac{ \alpha F b^{\sigma-1} }{ R^{(\gamma-2\sigma)/(\gamma-2) } }= -\frac{3}{4}+\frac{1}{4}=-\frac{1}{2}.$$
Thus, $w$ is a super-solution in $\mathbb{R}^n_T$ for any $a$ and $b>0$ satisfying the requirement in (\ref{sec4.43}).
\vskip 1em
We now evaluate $\lim_{b \rightarrow 0} a$. If $\sigma\ge 1$, it is clear from (\ref{sec4.43}) that $\lim_{b\rightarrow 0} a=0$.
Let $0<\sigma<1$. Using (\ref{sec4.43}), $\gamma^*=\gamma/(\gamma-2)$ and $k=\gamma-1$, we obtain that, for some $K_1$ and $K_2$, independent of $b$,
\begin{eqnarray*}
b^k R^{\gamma^*}&=&K_1 b^{\gamma-1} \left( b^{ (\sigma-1) (\gamma-2)/(\gamma-2\sigma)} \right)^{\gamma/(\gamma-2)} =K_1 b^{(\gamma-\sigma)(\gamma-2)/(\gamma-2\sigma)},\\
\mbox{and}\;\;b^\sigma R^{2\sigma/(\gamma-2)}&=&K_2 b^\sigma \left( b^{(\sigma-1)(\gamma-2)/(\gamma-2\sigma)} \right)^{2\sigma/(\gamma-2)}=K_1b^{\sigma (\gamma-2)/(\gamma-2\sigma)}.
\end{eqnarray*}
It is clear that
\eqRef{sec4.44}
\lim_{b \rightarrow 0} a=0. \quad\qquad \Box
\end{equation}
\vskip 1em
{\bf Sub-Part (iii) ($\sigma=\gamma/2$):} We modify $w$ as follows. Take
\eqRef{sec4.5}
w(x,t)=b (t+1) r^{\gamma^*},
\end{equation}
where $b>0$ is to be determined. Note that
$$\gamma^*=\frac{\gamma}{\gamma-2}=\frac{2\sigma}{ \gamma-2 }.$$
Taking $a=0$ in (\ref{sec4.40}) and observing that $k>1$ and $\gamma>2$, we get
\begin{eqnarray*}
&&\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w) \\
&&\quad\qquad\qquad\qquad \le E b^k r^{\gamma^*}+\alpha F b^\sigma r^{2\sigma/(\gamma-2)}-br^{\gamma^*}=Eb^kr^{\gamma^*}+\alpha F b^{\gamma/2} r^{\gamma^*}-b r^{\gamma^*}\nonumber\\
&&\quad\qquad\qquad\qquad =b r^{\gamma^*} \left( Eb^{k-1}+\alpha F b^{(\gamma-2)/2}-1\right)\le 0,\nonumber
\end{eqnarray*}
if $0<b\le b_0$, for some $b_0=b_0(\alpha,k,\gamma, E, F)$ chosen small enough. Thus,
\eqRef{sec4.50}
w(x,t)=b(1+t) r^{\gamma^*},\;\;\forall 0<b\le b_0,
\end{equation}
is a super-solution in $\mathbb{R}^n_T$. $\Box$
\vskip 1em
{\bf Sub-part (iv) ($\sigma>\gamma/2$):} We use Remark \ref{sec3.11} and take
\eqRef{sec4.6}
w(x,t)=at+b(1+t)v(r),
\end{equation}
where
$$v(r)=\int_0^{r^{\gamma^*} } \frac{1}{1+\tau^p}\;d\tau,\;\;\; p=\frac{\gamma^*-\sigma^*}{\gamma^*},\;\;\gamma^*=\frac{\gamma}{\gamma-2}\;\;\mbox{and}\;\;\sigma^*=\frac{\sigma}{\sigma-1}.$$
Here $a>0$ and $0<b\le 1$ are to be determined. Note that $v(r)$ grows like $r^{\gamma^*}$ near $r=0$ and like $r^{\sigma^*}$ for large $r$.
Recall (\ref{sec4.4}) i.e.,
\eqRef{sec4.7}
\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w)\le \frac{M [ b(1+T)]^k v^\prime(r)^k}{r} + \alpha\left[ b(1+T) \right]^\sigma v^{\prime}(r)^\sigma-a-bv(r).
\end{equation}
\vskip 1em
We use parts (ii)-(viii) of Remark \ref{sec3.11}, $k=\gamma-1$ and $(\sigma^*-1)\sigma=\sigma^*$. Note that
\begin{eqnarray*}
&&( v^{\prime}(r) )^\sigma\le (\gamma^*)^\sigma \min\left( r^{\sigma^*-1},\;r^{\gamma^*-1} \right)^\sigma=(\gamma^*)^\sigma \min \left( r^{\sigma^*},\;r^{2\sigma/(\gamma-2)} \right),\\
&&\mbox{and}\quad \frac{ (v^{\prime}(r))^k}{r}\le \min (\gamma^*)^k \left( r^{k\sigma^*-\gamma},\;r^{k\gamma^*-\gamma}\right)=(\gamma^*)^k \min \left( r^{(\gamma-\sigma)/(\sigma-1)},\;r^{\gamma^*}\right).
\end{eqnarray*}
Using the above in (\ref{sec4.7}) and recalling the definitions of $E$, $F$ (see the line following (\ref{sec4.5.1})) we get that
\begin{eqnarray}\label{sec4.80}
\mathcal{P}_\sigma(t, w,w_t, Dw, D^2w) \le Eb^k r^{(\gamma-\sigma)/(\sigma-1)} + \alpha Fb^\sigma r^{\sigma^*}-a-b v(r).
\end{eqnarray}
A lower bound for $v(r)$ is obtained by setting $R=1$ in Remark \ref{sec3.11}(iv) and ignoring $v(1)$. Taking $r\ge 1$, (\ref{sec4.80}) yields that
\begin{eqnarray}\label{sec4.8}
\mathcal{P}_\sigma(t, w,w_t, Dw, D^2w) &\le& Eb^k r^{(\gamma-\sigma)/(\sigma-1)} + \alpha Fb^\sigma r^{\sigma^*}-a-\frac{b \gamma^* \left( r^{\sigma^*} - 1 \right)}{2\sigma^*} \nonumber\\
&=&Eb^k r^{(\gamma-\sigma)/(\sigma-1)} + \alpha Fb^\sigma r^{\sigma^*}+\frac{b\gamma^* }{2\sigma^*}-a- \frac{ b\gamma^* r^{\sigma^*}}{2\sigma^*},
\end{eqnarray}
where we have used that $1-p=\sigma^*/\gamma^*$.
We select
\eqRef{sec4.85}
a= Eb^k+\alpha F b^\sigma +\frac{b\gamma^* }{\sigma^*},
\end{equation}
From (\ref{sec4.8}) and (\ref{sec4.85}), it follows that
$w$ is a super-solution in $B_1(o)\times [0,T]$.
Since $r^{(\gamma-\sigma)/(\sigma-1)}\le r^{\sigma^*}$, in $r\ge 1$, using (\ref{sec4.85}) in (\ref{sec4.8}) implies that
\begin{eqnarray*}
\mathcal{P}_\sigma(t, w_t, Dw, D^2w)& \le &Eb^k r^{\sigma^*}+ \alpha Fb^\sigma r^{\sigma^*}+\frac{b\gamma^* }{2\sigma^*}-a- \frac{ b r^{\sigma^*}}{2\sigma^*}\\
&\le &b r^{\sigma^*} \left( E b^{k-1}+\alpha F b^{\sigma-1} -\frac{\gamma^*}{2\sigma^*} \right)\le 0,
\end{eqnarray*}
if we select $0<b\le b_0$, where $b_0$ depends only on $\alpha,\gamma,\;\sigma,\;E$ and $F$, and is chosen small enough.
Thus, $w$ is super-solution in $\mathbb{R}^n_T$ and
\eqRef{sec4.9}
\lim_{b\rightarrow 0} a=0.\qquad \Box
\end{equation}
\vskip 1em
{\bf Part II ($k=1$):} In this case, $\gamma=2$ and $k_1=0$.
By Remark \ref{sec2.40}(iv), $H(q, X)=H(X),\;\forall (q,X)\in \mathbb{R}^n\times S^{n\times n}$. Thus, we work with
$$H(D^2u+Z(u)Du\otimes Du)+\chi(t)|Du|^\sigma-u_t\ge 0,\;\;\mbox{in $\mathbb{R}^n_T$ with $u(x,0)\le h(x),\;\forall x\in \mathbb{R}^n$.}$$
We treat separately the three possibilities: (i) $0\le \sigma\le 1$, (ii) $1<\sigma\le 2$ and (iii) $2<\sigma<\infty$.
\vskip 1em
{\bf Sub-Part (i) ($0\le \sigma\le1$):} Take
\eqRef{sec4.11}
w(x,t)=at+b(1+t)v(r),\quad \forall (x,t)\in \mathbb{R}^n_T,
\end{equation}
where $a\ge 0$ and $0<b\le1$ are to be determined.
\vskip 1em
{\bf (a) ($\sigma=0$):} We choose
$$v(r)=e^{cr^2}.$$
where $c>0$ is to be determined. We note the following elementary facts.
\begin{eqnarray*}
v^{\prime}(r)=2cre^{cr^2},\;\; \frac{v^{\prime}(r)}{r}=2c e^{cr^2},\;\;\;\mbox{and}\;\; \frac{ r v^{\prime\prime}(r)}{v^{\prime}(r)}=1+2cr^2.
\end{eqnarray*}
Using these in (\ref{sec4.4}) and using $\sigma=0$, we get
\begin{eqnarray*}
\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w)& \le& b(1+T)M \left( \frac{v^\prime(r)}{r} \right)+\alpha -a-bv(r)\\
&=& 2bc(1+T)M e^{cr^2}+\alpha -a-b e^{cr^2}.
\end{eqnarray*}
Set $a=\alpha$, $\bar{E}=2(1+T)M$ and $c=1/\bar{E}$ to obtain $\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w) \le 0$ in $\mathbb{R}^n_T$.
Thus,
$$ w(x,t)=\alpha t+ b(1+t) e^{r^2/\bar{E}},\quad \forall (x,t)\in \mathbb{R}^n_T,$$
is a super-solution in $\mathbb{R}^n_T$ for any $b>0$. Moreover,
\eqRef{sec4.131}
\lim_{b\rightarrow 0} w(x,t)=\alpha t. \qquad \Box
\end{equation}
\vskip 1em
{\bf (b) ($0<\sigma\le1$):} For $a>0$, $0<b\le 1$ and $c>0$ (to be determined), we define
\eqRef{sec4.20}
w(x,t)=at+b(1+t) v(r),\;\;\mbox{in $\mathbb{R}^n_T$},\;\; \mbox{where}\;\; v(r)=e^{cr}-(1+cr).
\end{equation}
Thus,
$$v^{\prime}(r) =c \left( e^{cr}-1\right),\quad c^2\le \frac{v^{\prime}(r)}{r}\le c^2 e^{cr},\quad \mbox{and}\quad 1\le \frac{r v^{\prime\prime}(r)}{v^{\prime}(r)}\le \frac{ e{\max(1, cr)}}{e-1}.$$
In the last estimate, for $0 < \theta < 1$ we used that $\theta e^\theta/(e^\theta-1)$ is increasing and for $1<\theta,$ we used
that $e^\theta/(e^\theta-1)$ is decreasing.
Applying the above to (\ref{sec4.4}), we obtain
\begin{eqnarray}\label{sec4.200}
&&\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w)\nonumber \\
&&\quad\qquad\le b(1+T)M\left( \frac{v^\prime(r)}{r} \right)+ \alpha \left[ b(1+T) v^\prime(r) \right]^\sigma-a-bv(r) \\
&&\quad\qquad\le bc^2(1+T) M e^{cr}+\alpha \left[ bc (1+T) \left( e^{cr}-1\right) \right]^\sigma -a- b\left( e^{cr}-1-cr\right).\nonumber
\end{eqnarray}
Set $\bar{E}=(1+T)M$ and $\bar{F}=\alpha (1+T)^\sigma$. A rearrangement of the above leads to
\eqRef{sec4.141}
\mathcal{P}_\sigma(t, w, w_t, Dw, D^2w)\le b(1+cr)+ \left(c^2\bar{E}\right) b e^{cr}+\left( c^\sigma \bar{F}\right) (be^{cr})^\sigma-b e^{cr}-a.
\end{equation}
Applying Young's inequality $(be^{cr})^\sigma\le (1-\sigma)+\sigma b e^{cr}$, (\ref{sec4.141}) implies that
\begin{eqnarray*}
\mathcal{P}_\sigma(t, w, w_t, Dw, D^2w)&\le& b(1+cr)+e^{cr}\left( c^2\bar{E}+\sigma c^\sigma \bar{F}-1\right)+(1-\sigma)c^\sigma\bar{F}-a\\
&\le& \left[ (1-\sigma)c^\sigma \bar{F}-a \right]+b\left[ (1+cr)+ e^{cr}\left( c^2\bar{E} +\sigma c^\sigma \bar{F}-1 \right) \right].
\end{eqnarray*}
Select $c>0$ such that $c^2\bar{E}+\sigma c^\sigma \bar{F}=1-\varepsilon$, for a fixed small $0<\varepsilon<1$. Hence,
\eqRef{sec4.201}
\mathcal{P}_\sigma(t, w, w_t, Dw, D^2w)\le\left[ (1-\sigma)c^\sigma \bar{F}-a \right]+b\left[ (1+cr)-\varepsilon e^{cr} \right].
\end{equation}
\vskip 1em
The maximum of the function $1+cr-\varepsilon e^{cr}$ occurs at $r_0=c^{-1} \log(1/\varepsilon)$ and the maximum value is $\log(1/\varepsilon)$. Select
$$a= b \log(1/\varepsilon)+(1-\sigma)c^\sigma \bar{F}.$$
Using the choice for $a$ in (\ref{sec4.201}), we get that $\mathcal{P}_\sigma(t,w_t, Dw, D^2w)\le 0$, in $\mathbb{R}^n_T$.
Thus, $w$ is a super-solution in $\mathbb{R}^n_T$ and
\eqRef{sec4.202}
\lim_{b \rightarrow 0} a=(1-\sigma)c^{\sigma}\bar{F}, \;\;0<\sigma\le 1.\qquad \Box
\end{equation}
\vskip 1em
Observe that if $\sigma=0$ then $\lim_{b \rightarrow 0}a=\alpha$. While this agrees with part (a), the growth rate allowed in part (a) is greater. Also, if we take $\sigma=1$, $\lim_{b\rightarrow 0} a=0$.
\vskip 1em
{\bf Sub-Part (iii) ($1<\sigma\le 2$):} For $a>0$ and $0<b\le 1$ (to be determined), we select
\eqRef{sec4.30}
w(x,t)=at+b(1+t) r^{\sigma^*},\;\;\forall(x,t)\in \mathbb{R}^n_T,\;\;\mbox{where}\;\; \sigma^*=\frac{\sigma}{\sigma-1}.
\end{equation}
\vskip 1em
Note that $\sigma^*\ge 2$. Setting $v(r)=r^{\sigma^*}$, we find that
$$\frac{v^{\prime}(r)}{r}=\sigma^* r^{\sigma^*-2}=\sigma^* r^{(2-\sigma)/(\sigma-1)},\quad v^{\prime}(r)^\sigma=(\sigma^*)^\sigma r^{\sigma^*}\quad\mbox{and}\quad \frac{rv^{\prime\prime}(r)}{v^{\prime}(r)}=\sigma^*-1. $$
Using the above in (\ref{sec4.200}) or (\ref{sec4.4}) and recalling the definitions of $\bar{E}$ and $\bar{F}$ (see Sub-Part (ii)) we obtain that
\begin{eqnarray}\label{sec4.31}
\mathcal{P}_\sigma(t,w,w_t,Dw, D^2w)&\le& \bar{E} \left(\frac{b v^\prime(r)}{r}\right)+ \bar{F} (bv^\prime(r))^\sigma-a-bv(r) \nonumber\\
&=& (\sigma^* \bar{E} )b r^{\sigma^*-2}+( {\sigma^*}^{\sigma} \bar{F} ) b^{\sigma} r^{\sigma^*}-a-br^{\sigma^*}.
\end{eqnarray}
Choose
\begin{eqnarray*}
R=\sqrt{4\sigma^* \bar{E}},\;\; 0<b< \left( \frac{1}{4 {\sigma^*}^\sigma \bar{F} } \right)^{1/(\sigma-1)} \;\;\mbox{and}\;\;a=(\sigma^* \bar{E}) bR^{\sigma^*-2}+ ( {\sigma^*}^{\sigma} \bar{F} )b^{\sigma} R^{\sigma^*}.
\end{eqnarray*}
Employing the above values in (\ref{sec4.31}) and noting that $\sigma^*\ge 2$, we see that $w$ is super-solution in $[0,R]\times [0,T].$ In $r\ge R$,
\begin{eqnarray*}
\mathcal{P}_\sigma(t,w,w_t,Dw, D^2w)&\le& (\sigma^* \bar{E}) br^{\sigma^*-2}+( {\sigma^*}^\sigma \bar{F} )b^\sigma r^{\sigma^*}-a-br^{\sigma^*}\\
&=&b r^{\sigma^*}\left( \frac{\sigma^* \bar{E}}{r^2}+ ({\sigma^*}^{\sigma} \bar{F} ) b^{\sigma-1} -1 \right)-a.
\end{eqnarray*}
Using the values of $R$ and $b$, it is clear that $w$ is super-solution in $\mathbb{R}^n_T$. Moreover,
\eqRef{sec4.32}
\lim_{b\rightarrow 0} a=0.\qquad \Box
\end{equation}
\vskip 1em
{\bf Sub-Part (iv) $2<\sigma<\infty$:} We choose
\eqRef{sec4.60}
w(x,t)=at+b(1+t) v(r),\;\;\forall (x,t)\in \mathbb{R}^n_T,
\end{equation}
where
$$v(r)=\int_0^{r^2} \frac{1}{1+\tau^p}\;d\tau\quad\mbox{with}\;p=1-\frac{\sigma^*}{2}=\frac{\sigma-2}{2(\sigma-1)}.$$
Observe that $2(1-p)=\sigma^*$ and also, that $v(r)$ is like $r^2$ near $r=0$ and like $r^{\sigma^*}$ for large $r$.
In Lemma \ref{sec3.10}, we set $\beta=2$ and $\bar{\beta}=\sigma^*$. Thus, parts (iv), (v) and (vi) yield
\begin{eqnarray*}
&&\mbox{(iv)}\;\;\frac{1}{\sigma^*}\le \frac{v(r)-v(1)}{r^{\sigma^*}-1}\le \frac{2}{\sigma^*},\;\forall r\ge 1,\\
&&\mbox{(v)}\;\;v^{\prime}(r)\le 2\min\left( r^{1/(\sigma-1)},\;r\right)\;\;\mbox{and (vi)}\;\;\frac{v^{\prime}(r)}{r}\le 2.
\end{eqnarray*}
Using the above values and expressions in (\ref{sec4.200}) or (\ref{sec4.4}) and recalling $\bar{E}$ and $\bar{F}$, we get
\begin{eqnarray}\label{sec4.61}
\mathcal{P}_\sigma(t, w, w_t, Dw, D^2w)&\le & b(1+T)M \left(\frac{v^\prime(r)}{r}\right)+ \left[ \alpha ( b(1+T) )^\sigma\right] (v^\prime(r))^\sigma-a-bv(r) \nonumber\\
&\le&2b \bar{E}+(2b)^{\sigma}\bar{F}\min\left( r^{\sigma^*},\;r^\sigma\right)-a -b v(r).
\end{eqnarray}
We choose
\begin{eqnarray*}
a=2b \bar{E}+(2b)^{\sigma}\bar{F}+\frac{b}{\sigma^*}\quad\mbox{and}\quad 0<b< \left( \frac{1}{2^{\sigma}\sigma^* \bar{F}}\right)^{1/(\sigma-1)}.
\end{eqnarray*}
Using the above, $w$ is a super-solution in $0\le r\le 1$ and $0\le t\le T$.
In $r\ge 1$, we employ values of $a$, $b$ and the bound $v(r)\ge (r^{\sigma^*}-1)/\sigma^*$ in (\ref{sec4.61}) to find that
\begin{eqnarray*}
\mathcal{P}_\sigma(t, w, w_t, Dw, D^2w)&\le&2b \bar{E}+(2b)^{\sigma}\bar{F}\min\left( r^{\sigma^*},\;r^\sigma\right)-a -\frac{ b \left( r^{\sigma^*}-1\right) }{\sigma^*} \nonumber \\
&\le&(2b)^{\sigma}\bar{F} r^{\sigma^*}-\frac{ b r^{\sigma^*}}{\sigma^*}
\le b r^{\sigma^*} \left( 2^\sigma b^{\sigma-1} \bar{F}-\frac{1}{\sigma^*}\right)\le 0.
\end{eqnarray*}
Thus, $w$ is super-solution in $\mathbb{R}^n_T$. Moreover,
\eqRef{sec4.63}
\lim_{b\rightarrow 0} a=0.\qquad \Box
\end{equation}
We summarize: select $w(x,t)=at+b(1+t)v(r)$ where $v(r)$ is as follows
\begin{eqnarray}\label{sec4.64}
&&\qquad\qquad\mbox{}\\
&&\mbox{(I) $k>1$:}\;\;v(r)=\left\{ \begin{array}{ccc} r^{\gamma^*},& 0\le \sigma\le \gamma/2,\\ \int_0^{r^{\gamma^*}}(1+\tau^p)^{-1}d\tau,& \sigma>\gamma/2, \end{array}\right.
\;\;\lim_{b\rightarrow 0}a=\left\{ \begin{array}{ccc} \alpha,& \sigma=0,\\ 0,& \sigma>0, \end{array}\right. \nonumber\\
&&\quad \mbox{where}\;\;p=1-(\sigma^*/\gamma^*), \nonumber\\
&&\;\mbox{(II) $k=1$:}\;v(r)=\left\{ \begin{array}{ccc} e^{cr^2},& \sigma=0,\\ e^{cr}-1-cr,& 0<\sigma\le 1,\\ r^{\sigma^*},& 1<\sigma\le 2,\\
\int_0^{r^2}(1+\tau^p)^{-1}d\tau,& \sigma>2, \end{array}\right. \;\;\lim_{b\rightarrow 0}a=\left\{ \begin{array}{ccc} \alpha,& \sigma=0,\\ (1-\sigma)c^\sigma \bar{F},& 0<\sigma\le 1,\\
0,& \sigma>1. \end{array}\right. \nonumber\\
&&\qquad \;\mbox{where}\;p=1-(\sigma^*/2). \nonumber
\end{eqnarray}
See (\ref{sec4.41}), (\ref{sec4.44}) and (\ref{sec4.9}), (\ref{sec4.131}), (\ref{sec4.202}), (\ref{sec4.32}) and (\ref{sec4.63}). Recall that $v(r)$ grows like $r^{\sigma^*}$ in (I) (for $\sigma>\gamma/2$) and in (II) (for $\sigma>2$).
\vskip 1em
\section{Sub-solutions}
\vskip 1em
The work in this section is quite similar to that in Section 4. Although, $H$ is not assumed to be odd in $X$, the auxiliary functions used in Section 4 continue to apply here. We will not repeat the calculations done in Section 4, instead, provide an outline as to how to use them to obtain sub-solutions. We require that the domain for $Z$ be $(-\infty, \infty).$
\vskip 1em
We use functions of the type $w(x,t)=-\left[at+b(1+t) v(r)\right],$
where $a>0$ and $b>0$, small, $v(r)>0$ and $v^{\prime}(r)\ge 0$. Recalling (\ref{sec3.8}), we see that
\begin{eqnarray}\label{sec4.14}
&&\mathcal{P}_\sigma(t,w, w_t, Dw, D^2w) \nonumber\\
&&\qquad\qquad=\frac{ [ b(1+t)v^{\prime}(r) ]^k}{ r } H\left( e, \left(1- \frac{r v^{\prime\prime}(r)}{v^\prime(r)} + b(1+t)rZ(w) v^\prime(r)\right)e\otimes e -I \right) \nonumber\\
&&\qquad\qquad\qquad\qquad+\chi(t) [ b (1+t) v^{\prime}(r)]^\sigma+a+b v(r).
\end{eqnarray}
We set
$$\alpha=\sup_{[0,T]}|\chi(t)|\;\;\;\;\mbox{and}\;\;\;N=\inf_\lambda\left( \min_{|e|=1}H(e, \lambda e\otimes e-I) \right).$$
We note that $N\le 0$ since $H(e, -I)\le 0$, see Condition C in Section 2.
\vskip 1em
As done in Section 4, we take $v(r)$ to be either a power of $r$(power greater than $1$) or $e^{cr^2}$ or $e^{cr}$. For the exponential type functions, since $1-(rv^{\prime\prime}(r))/v^{\prime}(r)$ could become unbounded, a lower bound on $H$ is needed. However, $1-(rv^{\prime\prime}(r))/v^{\prime}(r)$ is bounded from below if $v(r)$ is a power of $r$ and the bound depends on the power. Since $H$ is continuous and non-decreasing in $X$, we get a natural lower bound depending on the power of $r$. We use $N$ to denote the lower bound in both situations.
With the above discussion in mind, (\ref{sec4.14}) implies
\begin{eqnarray}\label{sec4.140}
&&\mathcal{P}_\sigma(t,w,w_t, Dw, D^2w)\ge \frac{ [ b(1+T)v^{\prime}(r) ]^k N}{ r}-\alpha [ b (1+T) v^{\prime}(r)]^\sigma+a +b v(r) \nonumber\\
&&\quad\qquad\qquad\qquad\qquad=-\left( \frac{ [ b(1+T)v^{\prime}(r) ]^k |N|}{ r } +\alpha [b (1+T) v^{\prime}(r)]^\sigma-a-b v(r) \right).
\end{eqnarray}
\vskip 1em
We now use auxiliary functions $v(r)$ that are similar to those in Section 4. The goal is to choose $a\ge 0$ and $0<b<1$ such that the expression in (\ref{sec4.140}) is non-positive i.e,
$$ \frac{ [ b(1+T)v^{\prime}(r) ]^k}{ r } |N|+\alpha [ b (1+T) v^{\prime}(r)]^\sigma-a-b v(r)\le 0.$$
The analysis is almost identical to Section 4. We list the choice for $w(x,t)$ for the various values of $\sigma$.
{\bf Part I\;\;$k>1$:} Recall that $\gamma>2$ and $\gamma^*=\gamma/(\gamma-2)$. Set $r=|x-z|$, for some fixed $z\in \mathbb{R}^n$, and take
\begin{eqnarray*}
&&w(x,t)=\left\{ \begin{array}{ccc} -at-b(1+t) r^{\gamma^*}, & 0\le \sigma<\gamma/2,\\ -b(1+t) r^{\gamma^*},& \sigma=\gamma/2,\\ -at-b(1+t)v(r), & \sigma>\gamma/2, \end{array}\right.\\
&&\mbox{where}\quad v(r)=\int_0^{r^{\gamma^*} } \frac{1}{1+\tau^p}\; d\tau\;\; \mbox{with}\;\; \sigma^*=\frac{\sigma}{\sigma-1},\;p=1-\frac{\sigma^*}{\gamma^*}=\frac{2\sigma-\gamma}{\gamma(\sigma-1)}.
\end{eqnarray*}
It is easy to check that (see Remark \ref{sec3.11}) that
$$1-\frac{rv^{\prime\prime}(r)}{v^{\prime}(r)}=\frac{ 2-\gamma^*+(2-\sigma^*) r^{p\gamma^*} }{1+r^{p\gamma^*}} \ge 2-\sigma^*>-\infty. $$
We choose $N$ to be an appropriate lower bound for $H$, see the right hand side of (\ref{sec4.14}). Thus, (\ref{sec4.140}) holds without any restrictions on $\inf_\lambda\left[ \min_{|e|=1}H(e, \lambda e\otimes e-I) \right].$
Moreover, from (\ref{sec4.64}),
\eqRef{sec4.15}
\lim_{b\rightarrow 0} a=\left\{ \begin{array}{ccc} \alpha, & \sigma=0,\\ 0,& \sigma>0. \end{array}\right. \qquad \Box
\end{equation}
\vskip 1em
{\bf Part II\;\;$k=1$:} In this case, $\gamma=2$ and $k_1=0$. Set $\sigma^*=\sigma/(\sigma-1)$. We choose $a\ge 0$, $0<b<1$ and $c>0$ such that (\ref{sec4.140}) in non-positive. We select
\begin{eqnarray*}
&&w(x,t)=\left\{ \begin{array}{ccc} -a-b(1+t)e^{cr^2},& \sigma=0,\\ -at-b(1+t)\left( e^{cr}-1-cr\right),& 0<\sigma\le 1,\\ -at-b(1+t) r^{\sigma^*},& 1<\sigma\le 2,\\
-at-b(1+t)v(r),& 2<\sigma<\infty, \end{array}\right.\\
&&\mbox{where}\;\;v(r)=\int_0^{r^2} \frac{1}{1+\tau^p}\;d\tau\;\;\mbox{with}\;\; p=1-\frac{\sigma^*}{2}=\frac{\sigma-2}{2(\sigma-1)}.
\end{eqnarray*}
\vskip 1em
If $0\le \sigma\le 1$ then $1-rv^{\prime\prime}(r)/v^{\prime}(r)\le 0$ becomes unbounded as $r\rightarrow \infty$. Thus, we impose that $|\inf_\lambda\left[ \min_{|e|=1}H(e, \lambda e\otimes e-I) \right] |<\infty.$ For $\sigma>1$, however, no such requirement is made.
Moreover, from (\ref{sec4.64}),
\eqRef{sec4.16}
\lim_{b\rightarrow 0} a=\left\{ \begin{array}{ccc} \alpha, & \sigma=0,\\ (1-\sigma) \alpha (c(1+T))^\sigma,& 0<\sigma\le 1,\\ 0, & \sigma>1. \end{array}\right.\qquad \Box
\end{equation}
\vskip 2em
\section{Some Special cases}
\vskip 1em
In this section we consider some special cases. Recall that
\eqRef{sec6.0}
\mathcal{P}_\sigma(t,w, w_t, Dw, D^2w)=H(Dw, D^2w+Z(w)Dw\otimes Dw)+\chi(t) |Dw|^\sigma-w_t.
\end{equation}
As before, set
$$N=\inf_\lambda \left[ \min_{|e|=1}H(e,\lambda e\otimes e-I) \right].$$
\vskip 1em
We discuss the following three cases.
{\bf Case (i):} $k\ge 1$ and $\chi\equiv 0$. The equations reads
$$H(Dv, D^2v+Z(v)Dv\otimes Dv)-v_t=0,\;\;\mbox{in $\mathbb{R}^n_T$, $v>0$, with $v(x,0)=h(x),\;\forall x\in \mathbb{R}^n$.}$$
As observed in (\ref{sec1.1}), (\ref{sec1.2}) and part (f) of Remark \ref{sec3.19}, this applies to the doubly nonlinear case by employing a change of variables. Moreover, as noted in
Remark \ref{sec3.19} and Lemma \ref{sec3.20}, the convergence or the divergence of the integral
$$I=\int_0^1 f^{-1/(k-1)}(\theta) \; d\theta,\;\;k>1, $$
determines the domain of $Z$. In particular, if $I<\infty$ then the domain of $Z$ is $(0,\infty)$ or $[0,\infty)$, and if $I=\infty$ then the domain is $(-\infty, \infty)$.
The super-solutions in Section 4, (in particular, the one in Sub-Part (i) or Part I) being positive, are also super-solutions of (\ref{sec6.0}) regardless the domain of $Z$. However, the domain of $Z$ needs to be stated more precisely for sub-solutions. If the integral $I$ diverges then the work in Section 5, in particular, Part I applies since the domain of $Z$ is $(-\infty, \infty)$. If $I$ converges then the domain is $(0,\infty)$ or $[0,\infty)$ and a different sub-solution needs to be calculated. We do this in this section.
We also include here the case $k=1$ where $Z$ is defined on $(0,\infty)$ or $[0,\infty)$. The two Part II's in Sections 4 and 5 address the case where the domain is $(-\infty, \infty)$.
\vskip 1em
The next two cases bring out the influence of the sign of $\chi$.
{\bf Case (ii):} We discuss super-solutions in the case $\chi\le 0$ and we derive a maximum principle.
{\bf Case (iii):} We study sub-solutions for
$\chi\ge 0$ and this leads to a minimum principle.
The cases {\bf (ii)} and {\bf (iii)} are related.
\vskip 1em
Let $z\in \mathbb{R}^n$ be a fixed and set $r=|x-z|,\;\forall x\in \mathbb{R}^n$.
We begin with Case (i).
\vskip 1em
{\bf Case (i-1):} We take $k>1$, $\chi\equiv 0$, $\sigma=0$ and assume that the domain of $Z$ contains $(0,\infty)$. Thus, the equation reads
$$\mathcal{P}_0(t,w, w_t, Dw, D^2w)=H(Dw, D^2w+Z(w)Dw\otimes Dw)-w_t.$$
Since our goal is to construct positive sub-solutions $w$, it suffices to find a $w$ such that
$H(Dw, D^2w)-w_t\ge 0$ since ellipticity ($Z\ge 0$) implies the desired conclusion.
Let $R>0$ and set $B^R_T=B_R(z)\times (0,T)$. We construct a sub-solution $w$ for any large $R$. More precisely, $w\ge 0$ solves
$$H(Dw, D^2w)-w_t\ge 0,\;\;\mbox{in $B^R_T$ and $w(x,0)\le g(x),\;\forall x\in B_R(z)$.}$$
We define
\eqRef{sec6.900}
w(x,t)=\psi(t)v(r)= \frac{D \left[ R^{ (k+1)/k}-r^{(k+1)/k} \right]^{k/(k-1)} } { (E+t)^{1/(k-1)} },\;\;\forall(x,t)\in B^R_T,
\end{equation}
where $D,\; E>0$ are to be determined.
One recalls from (\ref{sec3.7}) that if $w=\psi(t) v(r)$, with $w_r\le 0$, then
\begin{eqnarray}\label{sec6.90}
&&H(Dw, D^2w)-w_t=\frac{ \left( |\psi(t) v^{\prime}(r)| \right)^k }{r} H\left(e, \left( 1-\frac{rv^{\prime\prime}(r)}{v^{\prime}(r) } \right)e\otimes e-I \right)-v(r)\psi^{\prime}(t) \nonumber\\
&&\qquad\qquad\qquad\qquad\ge -\frac{|N| \left( |\psi(t) v^{\prime}(r)| \right)^k }{r}-v(r)\psi^{\prime}(t).
\end{eqnarray}
Using the expression for $w$ and setting $c_k= \left[ (k+1)/(k-1) \right]^k$, we see that
\begin{eqnarray*}
&&-v(r)\psi^{\prime}(t)- \frac{ |N| \left( |\psi(t) v^{\prime}(r)| \right)^k }{r}=\frac{ D v(r)}{ (k-1) ( E+t)^{k/(k-1)} }-\frac{ c_k |N| D^k v(r)}{ (E+t)^{k/(k-1)} } \\
&&\qquad\qquad\qquad\qquad\qquad\quad=\frac{ D v(r)}{ (k-1) ( E+t)^{k/(k-1)} } \left[ 1- (k-1) c_k |N| D^{k-1} \right].
\end{eqnarray*}
Choosing
$$D= \left(\frac{ 1}{c_k(k-1) |N| } \right)^{1/(k-1)},$$
and using the above in (\ref{sec6.90}), we get a sub-solution $w\ge 0$ in $B^T_R$ such that $w(R,t)=0$. Next, we calculate $E$ by requiring that
$$w(z,0)=w(0,0)=\frac{ D R^{(k+1)/(k-1)} }{ E^{1/(k-1)} }=\inf_x h(x)=\mu.$$
Thus,
\begin{eqnarray*}
w=\frac{ D R^{(k+1)/(k-1)}} {E^{1/(k-1)} } \frac{ \left[ 1-(r/R)^{(k+1)/k} \right]^{k/(k-1)} } { (1+(t/E) )^{1/(k-1)} } = \frac{ \mu \left[ 1-(r/R)^{(k+1)/k} \right]^{k/(k-1)} } { (1+(t/E) )^{1/(k-1)} }.
\end{eqnarray*}
Note that $E=O(R^{k+1})$ and
$$
w(z,t)=w(0,t)= \frac{\mu }{ (1+(t/E) )^{1/(k-1) } }\rightarrow \mu \;\;\mbox{as $R\rightarrow \infty$}.
$$
We record that in $0\le r<R$,
\eqRef{sec6.91}
w(x,t)= \frac{ \mu \left[ 1-(r/R)^{(k+1)/k} \right]^{k/(k-1)} } { (1+(t/E) )^{1/(k-1)} },\;\;\mbox{where}\;\;E=\frac{R^{k+1}}{ c_k \mu^{k-1}(k-1)|N|}.
\end{equation}
\vskip 1em
{\bf Case (i-2):} We now study $k=1$. We take $w(x,t)=D e^{-Er^2}e^{-Ft}$ and recall (\ref{sec6.90}). We get
\begin{eqnarray*}
&&-|N| \frac{ \psi(t)| v^{\prime}(r)| }{r}-v(r)\psi^{\prime}(t)\\
&&\qquad\qquad\qquad\qquad=DF e^{-Er^2}e^{-Ft}-|N| 2DE e^{-Er^2} e^{-Ft}=De^{-Er^2} e^{-Ft} \left( F-2|N| E \right)
\end{eqnarray*}
We take $F=2|N|E$ and $D=\mu$ and obtain a sub-solution
\eqRef{sec6.92}
w(x,t)=\mu e^{-Er^2} e^{-2|N| E t},\;\;\forall E>0.\end{equation}
It is clear that $W\rightarrow \mu$ as $E\rightarrow 0.$ $\Box$
\vskip 2em
{\bf Case (ii):} We consider
\eqRef{sec6.1}
\mathcal{P}_\sigma(t, w, w_t, Dw, D^2w)=H(Dw, D^2w+Z(w) Dw\otimes Dw)+\chi(t)|Dw|^\sigma-w_t
\end{equation}
where $\chi\le 0$. We set
$$\hat{\alpha}=\sup_{(0,T)}\chi(t)$$
and assume that $\hat{\alpha}<0$. We further assume that
\eqRef{sec6.2}
k\ge 1\quad \mbox{and}\quad \sigma\ge k.
\end{equation}
Our goal here is to construct super-solutions $w\ge 0$, i.e., $\mathcal{P}_\sigma(t, w, w_t, Dw, D^2w)\le 0$ in cylinders $B^R_T$.
Selecting $w(x,t)=at+(1+t)v(r),\;v^\prime\ge 0$, setting $M=\sup_\lambda \left[ \max_{|e|=1} H(e, I+\lambda e\otimes e) \right]$ and recalling (\ref{sec3.6}) and (\ref{sec6.1}) we find that
\begin{eqnarray}\label{sec6.3}
&&\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w)\nonumber\\
&&\qquad\le \frac{ [(1+t) v^\prime(r)]^k}{r} H\left( e, \;I +\left( \frac{r v^{\prime\prime}(r)}{v^\prime(r)}-1 + (1+t) r v^{\prime}(r) Z(w) \; \right)e\otimes e\;\right) \nonumber\\
&&\quad\qquad\qquad\qquad-|\hat{\alpha}| [(1+t) v^\prime(r)]^\sigma-a-v(r) \nonumber\\
&&\qquad\le \frac{ [(1+t) v^\prime(r)]^k }{r} \left\{ M -|\hat{\alpha}| [(1+t) v^\prime(r)]^{\sigma-k} r \right\}-a- v(r).
\end{eqnarray}
For $R>0$, set
$$v(r)=\left( R^2-r^2\right)^{-1},\;\;0\le r<R.$$
Since
$$v^{\prime}(r)=(2 r) ( R^2-r^2 )^{-2},$$
(\ref{sec6.3}) yields that, in $0\le r<R$,
\begin{eqnarray}\label{sec6.4}
&&\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w) \le \frac{ [2(1+t)]^k r^{k-1} }{ (R^2-r^2)^{2k}} \left( M- |\hat{\alpha}| \left( \frac{ 2 (1+t) }{ (R^2-r^2)^2 } \right)^{\sigma-k} r^{\sigma-k+1} \right)-a \nonumber \\
&&\qquad\qquad\qquad\qquad\qquad\qquad-\frac{1}{ R^2-r^2 }.
\end{eqnarray}
\vskip 1em
{\bf Sub-Case (ii-1) ($\sigma=k$:)} Set $r^*=M/\hat{\alpha}$ and take $R>r^*$. Then (\ref{sec6.4}) yields that
\eqRef{sec6.5}
\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w) \le \left( \frac{ 2(1+t)}{ (R^2-r^2)^{2} }\right)^k \left( M - |\hat{\alpha}| r \right)r^{k-1}-a-\frac{1}{ R^2-r^2}.
\end{equation}
Select
$$a=M\left( \frac{ 2(1+T) }{ (R^2-(r^*)^2)^{2} }\right)^k (r^*)^{k-1}.$$
With this choice (\ref{sec6.5}) shows that $w$ is a super-solution in $B^R_T$.
Thus,
\eqRef{sec6.6}
w(x,t)=at+(1+t) v(r)\;\;\mbox{and}\;\;\lim_{R\rightarrow \infty}a=0.\qquad \Box
\end{equation}
\vskip 1em
{\bf Case (ii-2) ($\sigma>k$:)} From (\ref{sec6.4}) we have that
\begin{eqnarray*}
&&\mathcal{P}_\sigma(t,w,w_t,Dw,D^2w) \nonumber\\
&&\qquad \qquad \le \left( \frac{ 2(1+t) }{ (R^2-r^2)^{2} }\right)^k \left( M - |\hat{\alpha}| \left( \frac{2 }{ (R^2-r^2)^{ 2 } } \right)^{\sigma-k} r^{\sigma-k+1} \right)r^{k-1}-a.
\end{eqnarray*}
Since the function $f(r)=r^{\sigma-k+1} \left( R^2-r^2 \right)^{2(k-\sigma)}$ is continuous and increasing in $0\le r<R,$ f(0)=0 and $f(r)\rightarrow \infty$, as $r\rightarrow R$, there is an $r^*=r^*(R)<R$ such that
$2^{\sigma-k} |\hat{\alpha}| f(r^*)=M$. Choose
$$a=M\left( \frac{ 2 (1+T) }{ (R^2-(r^*)^2)^{2} }\right)^k (r^*)^{k-1}.$$
Clearly, $w$ is super-solution in $0\le r<R$.
Next, we recall that
$$f(r^*)= \frac{ (r^*)^{\sigma-k+1} }{ [R^2-(r^*)^2]^{2(\sigma-k)} } =\frac{ M}{ 2^{\sigma-k} |\hat{\alpha}| } \; .$$
Clearly, $r^*\rightarrow \infty$, as $R\rightarrow \infty$. For calculating $\lim_{R\rightarrow \infty} a$, we use the formula for $f(r^*)$ and observe that for an appropriate constant $D$, we have
$$ \frac{ (r^*)^{k-1} } { [ R^2-(r^*)^2]^{2k} }=\frac{ D (r^*)^{k-1}}{ (r^*)^ { k(\sigma-k+1)/(\sigma-k) } }=\frac{D}{ (r^*)^{1+k/(\sigma-k)}}.$$
Thus,
\eqRef{sec6.7}
\lim_{R\rightarrow \infty}a=0.\qquad \Box
\end{equation}
\vskip 2em
{\bf Case (iii) Sub-solution:}
We construct a function $w(x,t)$ such that
\begin{eqnarray*}
&&\mathcal{P}_\sigma(t, w, w_t, Dw, D^2w)\\
&&\qquad\qquad=H(Dw, D^2w+Z(w)Dw\otimes Dw)+\chi(t)|Dw|^\sigma-w_t\ge 0,\;\;\mbox{in $\mathbb{R}^n_T$,}
\end{eqnarray*}
where $\chi\ge 0$. Set
$$N=\inf_\lambda \left[ \min_{|e|=1} H(e, \lambda e\otimes e-I) \right]\qquad\mbox{and}\qquad \hat{\alpha}=\inf \chi(t).$$
Select $w(x,t)=-at-(1+t)v(r),\;v^{\prime}\ge 0,$ and recall (\ref{sec3.8}):
\begin{eqnarray}\label{sec6.8}
&&\mathcal{P}_\sigma(t,w, w_t, Dw, D^2w) \nonumber\\
&&\qquad\qquad=\frac{ [ (1+t)v^\prime(r)]^k}{ r } H\left( e, \;\left(\; (1+t)rv^{\prime}(r)Z(w)\;+1 - \frac{rv^{\prime\prime}(r)}{v^\prime(r)}\right)e\otimes e -I \right) \nonumber\\
&&\quad\qquad\qquad\qquad +\chi(t) [ (1+t) v^{\prime}(r) ]^\sigma+a+ v(r) \nonumber\\
&&\qquad\qquad \ge - \left[ \frac{ [ (1+t)v^\prime(r)]^k}{ r } |N| -\chi(t) [ (1+t) v^{\prime}(r) ]^\sigma-a- v(r) \right].
\end{eqnarray}
\vskip 1em
Defining
$$v(r)=\frac{1}{ R^2-r^2 },\;\;\forall\,\, 0\le r<R,$$
and proceeding as in Case (ii), one can construct a sub-solution $w$ with the same properties. $\Box$
\begin{rem}\label{sec6.100} We point out that, except for Case (i) in this section all the auxiliary functions in this work are of the kind $w(x,t)=at+b(1+t)v(r)$, where $v(r)$ is an appropriately chosen function, $r=|x-z|,\;\forall x\in \mathbb{R}^n$ and $z\in \mathbb{R}^n$ is fixed.
Case (i) is used in proving the minimum principle in Theorem. For $k>1$ we utilize $w$ in (\ref{sec6.91}) and for $k=1$ we use $w$ in (\ref{sec6.92}). Note that $k>1$ requires no lower bound except $u>0$, however, for $k=1$ we assume a lower bound.
Case (ii) implies a maximum principle without any imposition of an upper bound. Case (iii) leads to a minimum principle without requiring any lower bound.
We provide details in Section 7. $\Box$
\end{rem}
\vskip 1em
\section{Proofs of the main results} \label{Section 7}
\vskip 1em
Assume that $-\infty<\inf_{\mathbb{R}^n} g\le \sup_{\mathbb{R}^n} g<\infty$ and set
$$\mu=\inf_{\mathbb{R}^n}g\;\;\;\mbox{and}\;\;\;\nu=\sup_{\mathbb{R}^n}g.$$
\vskip 1em
\noindent{\bf Proofs of Theorems \ref{sec2.6} and \ref{sec2.7}: ($k>1$)}
\vskip 1em
We first present the proof of Theorem \ref{sec2.6}.
Select $\varepsilon>0$ small and $R_0>0$ such that
\eqRef{sec7.1}
\sup_{ [0,R]\times[0,T]}u(x,t)\le \varepsilon R^{\beta},\;\;\;\forall\;R\ge R_0.
\end{equation}
where $\beta$ is as described in the statements of the theorem.
Recall from (\ref{sec4.4.0}) and (\ref{sec4.64}) that the super-solution $w(x,t)$ can be written as
\begin{eqnarray*}
w(x,t)=at+b(1+t){v}(r),
\end{eqnarray*}
for an appropriate ${v}(r)>0$. Observe that $w$ is a super-solution for any small $b>0$. Also, ${v}$ grows like $r^\beta$, see (\ref{sec4.64}) and the constructions of the super-solutions in Section 4. Define
$$W(x,t)=\nu+w(x,t).$$
Let $\hat{k}>2$ be a constant so that $\hat{k}v\ge r^\beta$ for $r\ge R_1$, where $R_1$ is large. We take $b=\hat{k}\varepsilon$ in $W(x,t)$ and consider the cylinder $B_{R}(z)\times [0,T]$,
where $R\ge$ max$(R_0, R_1)$. At $t=0$, $W(x,0)=\nu+\hat{k}\varepsilon {v}(r)\ge \nu \ge u(x,0)$. On $|x-z|=R$,
$$W(x,t)\ge \hat{k}\varepsilon {v}(R)\ge \varepsilon R^\beta.$$
Thus, $W\ge u$ on the parabolic boundary of $B_{R}(z)\times (0,T).$ We apply Lemma \ref{sec3.18} to conclude that $W\ge u$ in $B_R(z)\times(0,T)$ for any $R$, i.e.,
\begin{eqnarray*}
u(x,t)\le at+\hat{k}\varepsilon(1+t)v(r)+\nu,\;\;\forall |x-z|\le R.
\end{eqnarray*}
Taking $x=z$, we get that $u(z,t)\le at+\nu$. Letting $R\rightarrow \infty$ and then $\varepsilon\rightarrow 0$ (i.e. $b\rightarrow 0$) and using (\ref{sec4.64}) (employ $\lim_{b\rightarrow 0} a$) we obtain the conclusion of the Theorem.
The proof of Theorem \ref{sec2.7} can be obtained by using Part I of Section 5 and arguing analogously. We omit the details.
\hfill $\Box$
\vskip 1em
\noindent{\bf Proofs of Theorems \ref{sec2.9} and \ref{sec2.10}: ($k=1$)}
We first prove Theorem \ref{sec2.9}. We recall (II) in (\ref{sec4.64}).
We take $\sigma=0$. Let $0<\varepsilon<c/10$ be small and fixed. Set
$$W(x,t)={\nu}+\alpha t+{ \varepsilon}(1+t)e^{cr^2},\;\forall (x,t)\in \mathbb{R}^n_T.$$
Then $W$ is super-solution for any small {$\varepsilon>0$. }
Choose $R_0>0$ such that $\sup_{B_R(z)\times [0,T]} u(x,t)\le e^{\varepsilon R^2}$ and
$\varepsilon e^{cR^2}>e^{\varepsilon R^2},\;\forall R>R_0$.
We apply the comparison principle Lemma \ref{sec3.18} to prove the claim in the theorem.
Observe that $W(x,0)\ge { \nu}\ge u(x,0),\;\forall x\in \mathbb{R}^n_T.$ On $|x-z|=R>R_1$, $W(x,t)\ge \varepsilon e^{cR^2}\ge e^{\varepsilon R^2}.$ By Lemma \ref{sec3.18}, $u(x,t)\le W(x,t),\;\forall (x,t)\in B_R(z)\times (0,T),$ for any $R>R_0$. Hence,
$$u(z,t)\le W(z,t)={ \nu +}\alpha t+\varepsilon(1+t){ e^{\varepsilon r^2}}.$$
Since the above holds for any large $R$, we let $\varepsilon\rightarrow 0$ to obtain the claim in part (a).
Part (b) may now be shown by arguing as above. Part (c) may be shown by following the ideas in the Proof of Theorem \ref{sec2.6}. Theorem \ref{sec2.10} follows analogously, see Part II in Section 5. $\Box$
\vskip 1em
We now present the proof of Theorem \ref{sec2.8}. We start with the maximum principle.
\vskip 1em
{\bf Proof of Theorem \ref{sec2.8}(a): (Maximum principle)} We refer to Remark \ref{sec3.19} and the comparison principle in Lemma \ref{sec3.20}.
We set $\alpha=0$ in part (a) of Theorem \ref{sec2.6}. Suppose that $u=\phi(v)$ where the change of variable is as in Remark \ref{sec3.19}. If
$$H(Du, D^2u)-f(u)u_t\ge 0,\;\;\mbox{in $\mathbb{R}^n_T$, $u>0$, with $u(x,0)\le g(x),\;\forall x\in \mathbb{R}^n$,}$$
then
$$H(Dv, D^2v+Z(v) DV\otimes Dv)-v_t\ge 0,\;\;\mbox{in $\mathbb{R}^n_T$, with $v(x,0)\le \phi^{-1}(g(x)),\;\forall x\in \mathbb{R}^n$,}$$
where $Z(s)=\phi^{\prime\prime}(s)/\phi^{\prime}(s)$ and the domain of $Z$ contains $(0,\infty)$. See Remark \ref{sec3.19}.
The super-solution $w$ used in the proof of Theorem \ref{sec2.6} is positive. Clearly, $Z(w)$ is well-defined. Using Lemma \ref{sec3.18} (or Lemma \ref{sec3.20}) and
arguing as in the proof of Theorem
\ref{sec2.6} we get that $v\le \sup_x \phi^{-1}(g(x))$, if
$$ \sup_{B_R(z)\times (0,T)} v(x,t)=o( R^{\gamma^*})\;\;\mbox{as $R\rightarrow \infty$.}$$
Thus, the claim holds for $u$.
\vskip 1em
{\bf Proof of Theorem \ref{sec2.8}: (Minimum principle)}
(i) Suppose that
$$\lim_{\delta\rightarrow 0^+}F(1)-F(\delta)<\infty.$$
We choose
$$v=\phi^{-1}(u)=\int_0^u f^{-1/(k-1)}(\theta)d\theta,\;\;u>0.$$
Then
$$H(Dv, D^2v+Z(v) Dv\otimes Dv)-v_t\le 0,\;v>0,\;\mbox{in $\mathbb{R}^n_T$ with $v(x,0)\ge \phi^{-1}(g(x)),\;\forall x\in \mathbb{R}^n,$}$$
where the domain of $Z$ contains $(0, \infty)$. We recall Case(i-1) from Section 6 and (\ref{sec6.91}) i.e.,
$$w(x,t)= \frac{ \hat{\mu} \left[ 1-(r/R)^{(k+1)/k} \right]^{k/(k-1)} } { (1+(t/E) )^{1/(k-1)} },\;\;\mbox{where}\;\;E=\frac{R^{k+1}}{ c_k {\hat\mu}^{k-1}(k-1)|N|},$$
for any large $R>0$. Here $\hat{\mu}=\phi^{-1}(\mu).$
We use comparison in $B_R(z)\times [0,T]$. It is clear that $v(x,0)\ge \phi^{-1}(g(x)){ \ge} w(x,0),$ $\forall |x-z|<R.$ Since $v>0$ in $\mathbb{R}^n_T$, working with $R^{'}<R$, close to $R$, we see that
$v(x,t)\ge w(x,t)$. Applying Lemma \ref{sec3.18} to the parabolic boundary of $B_{R^{'}}(z)\times (0,T)$, we get that $v(x,t)\ge w(x,t)$ in $B_{R^{'}}(z)\times (0,T)$. Thus,
$$v(z,t)\ge w(z,t)=\frac{ \hat{\mu} } { (1+(t/E) )^{1/(k-1)} },$$
Letting $R\rightarrow \infty$ (i.e. $E\rightarrow \infty$), we get that $v(z,t)\ge \hat{\mu}$ and the claim follows for $u$.
For $k=1$ and $f\equiv 1$, we use Case (i-2) in Section {{6}} and (\ref{sec6.92}) and assume that $\inf_{B_R(z)\times [0,T]} u(x,t)\ge { \mu} e^{-\varepsilon R^2}$, where $R>0$ is large enough and $\varepsilon>0$ is small but fixed. Recall from (\ref{sec6.92}) that
$$w(x,t)=\mu e^{-Er^2} e^{-2|N| E t},\;\;\forall E>0,$$
is a sub-solution in $\mathbb{R}^n_T$. Working in cylinders $B_R(z)\times (0,T)$, for large $R$, we find that $u(x,0)\ge \mu\ge w(x,0)$, for any $E>0$. Fix an $E>\varepsilon$. On $|x-z|=R$,
$w(x,t)\le u(x,t)$ implying that $w(x,t)\le u(x,t)$ in $B_R\times (0,T)$, for any large $R$, and, hence, in $\mathbb{R}^n_T$. Thus,
$$w(z,t)=\mu e^{-2|N|E t}\le u(z,t),\;\forall E>\varepsilon.$$
Since the above holds for any $R$ and, hence, for any $\varepsilon>0$, we get that the above estimate holds for any $E>0$. Clearly, the claim holds.
\vskip 1em
(ii) Suppose that
$$\int_0 ^1 f^{-1/(k-1)}(\theta)d\theta<\infty.$$
We choose
$$v=\phi^{-1}(u)=\int_0^u f^{-1/(k-1)}(\theta)d\theta,\;\;u>0.$$
Then
$$H(Dv, D^2v+Z(v) Dv\otimes Dv)-v_t\le 0,\;\;\mbox{in $\mathbb{R}^n_T$ with $v(x,0)\ge \phi^{-1}(g(x)),\;\forall x\in \mathbb{R}^n,$}$$
where the domain of $Z$ contains $(0, \infty)$. This is similar to the proof of Theorem \ref{sec2.10}.
The case $k=1$ and $f\equiv 1$ also follows in an analogous way. $\Box$
|
\section{Introduction}
\label{s:intro}
On a closed manifold of dimension at least 2, a Riemannian metric is called Besse when all of its geodesics are closed. It is called Zoll when all its unit-speed geodesics are closed with the same minimal period, and simple Zoll when they are also without self-intersections. As usual, by closed geodesic we mean a non-constant periodic orbit of the geodesic flow. Riemannian metrics in these three classes are of great interest in Riemannian geometry, see \cite{Besse:1978pr}.
The only known closed manifolds admitting Zoll Riemannian metrics are the compact rank-one symmetric spaces, that is, $S^n$, $\mathds{R} \mathds{P}^n$, $\mathds{C} \mathds{P}^n$, $\mathds{H} \mathds{P}^n$, or $\mathrm{Ca}\mathds{P}^2$, whose canonical Riemannian metrics are simple Zoll. Actually a result of Bott and Samelson \cite{Bott:1954aa, Samelson:1963aa} implies that any closed manifold admitting a simple Zoll Riemannian metric has the integral cohomology ring of a compact rank-one symmetric space.
Conjecturally, on simply connected closed manifolds $M$, the notions of Besse and Zoll Riemannian metrics are equivalent. This conjecture has been recently established for $M=S^n$ with $n\geq4$ by Radeschi and Wilking \cite{Radeschi:2017dz}, and was earlier established for $M=S^2$ by Gromoll and Grove \cite{Gromoll:1981kl}, who also showed that on $S^2$ the condition of being simple Zoll is equivalent to the other two ones.
The geodesic flow of a Riemannian manifold is a classical autonomous Hamiltonian flow in its tangent bundle. This implies that the closed geodesics parametrized with constant speed are the non-trivial critical points of the energy functional on the space of loops, whereas the closed geodesics with any parametrization are the critical points of the length energy functional. A result claimed by Lusternik \cite{Ljusternik:1966tk} and proved recently by the authors \cite{Mazzucchelli:2017aa} implies that a Riemannian 2-sphere is Zoll if and only if the min-max values of the length functional over three suitable homology classes of the space of unparametrized simple loops coincide. If instead only two among these three values coincide, the Riemannian metric may not be Zoll, but the geodesic dynamics is still rather special: any point of the 2-sphere must lie on a closed geodesic. For $n$-spheres of arbitrary dimension $n\geq2$ with sectional curvature pinched inside $[1/4,1]$, analogous results were proved by Ballmann-Thorbergsson-Ziller \cite{Ballmann:1983fv}. The goal of this paper is to provide further results along this line for more general closed manifolds, and in particular for higher dimensional ones, without any assumption on the curvature.
In order to state our theorems, let us quickly recap the variational theory for the closed geodesic problem. Let $M$ be a closed orientable manifold of dimension $n\geq 2$ admitting a simple Zoll Riemannian metric. The manifold $M$ will be implicitly identified with the submanifold of constant loops in the free loop space $\Lambda M=W^{1,2}(\mathds{R}/\mathds{Z},M)$. For this class of manifolds, we have the explicit cohomology computation
\begin{align}
\label{e:cohomology_loop_space}
H^*(\Lambda M,M)
\cong
\bigoplus_{m\geq1}
H^{*-m\,i(M)-(m-1)(n-1)}(SM),
\end{align}
where $SM$ denotes the unit tangent bundle of $M$, and $i(M)$ is a suitable positive integer only depending on the integral cohomology ring of $M$, according to an argument due to Radeschi and Wilking \cite[page~942]{Radeschi:2017dz}. Throughout this paper, the singular cohomology $H^*$ and the singular homology $H_*$ will always be intended with $\mathds{Z}$ coefficients unless we explicitly state otherwise. Since $M$ admits a simple Zoll Riemannian metric, Bott and Samelson's theorem \cite[Theorem~7.23]{Besse:1978pr} implies that it is simply connected and with vanishing Euler characteristic if and only if it is homeomorphic to an odd dimensional sphere $S^n$. In this case, the relative cohomology group $H^{*}(\Lambda M,M)$ has rank at most one in every degree, and for each integer $m\geq1$ we choose two generators
\begin{equation}
\label{e:classes_thm1}
\alpha_m \in H^{(2m-1)(n-1)}(\Lambda M,M),
\qquad
\beta_m \in H^{2m(n-1)+1}(\Lambda M,M).
\end{equation}
For each Riemannian metric $g$ on $M$, the associated energy functional is
\[
E:\Lambda M\to[0,\infty),\qquad E(\gamma)=\int_0^1 \|\dot\gamma(t)\|^2_g\,\mathrm{d} t.
\]
For each $b>0$, we consider the energy sublevel set $\Lambda M^{<b}:=\{\gamma\in\Lambda M\ |\ E(\gamma)<b\}$, and denote by $\iota_b:(\Lambda M^{<b},M)\hookrightarrow(\Lambda M,M)$ the inclusion. Given a non-trivial cohomology class $\mu\in H^d(\Lambda M,M)$, the associated min-max
\begin{align*}
c_g(\mu)=c_g(-\mu):=\inf\{b>0\ |\ \iota_b^*\mu\neq 0\}
\end{align*}
is a critical value of $E$, and thus the energy of a closed geodesic.
One can easily verify that, if $g$ is a Zoll Riemannian metric with unit-speed geodesics of minimal period $\ell>0$, then $c_g(\alpha_m)=c_g(\beta_m)=m^2\ell^2$ for all $m\in\mathds{N}$. Conversely, we will prove the following theorem.
\begin{thm}
\label{t:every_point}
Let $M$ be a manifold homeomorphic to an odd dimensional sphere $S^n$, $n\geq 3$, and $g$ a Riemannian metric on $M$.
If $c_g(\alpha_m)=c_g(\beta_m)$ for some $m\geq 1$, then for each $q\in M$ there exists a (possibly iterated) closed geodesic $\gamma\in\mathrm{crit}(E)$ with $\gamma(0)=q$ and $E(\gamma)=c_g(\alpha_m)$.
\end{thm}
Our second result provides a min-max characterization of Zoll Riemannian metrics on simply connected spin closed manifolds. The statement requires a new finite dimensional reduction of the variational settings for the energy, which goes as follows. Let $(M,g)$ be a closed Riemannian manifold of dimension $n\geq2$, with associated energy functional $E:\Lambda M\to[0,\infty)$. We denote by $\rho=\mathrm{injrad}(M,g)>0$ the injectivity radius, and by $d:M\times M\to[0,\infty)$ the Riemannian distance. For each $\delta\in(0,\rho)$ and $k\in\mathds{N}$, we consider the space
\begin{align*}
\Upsilon M = \Upsilon_{\delta,k}M:=\left\{ \bm{q}=(q_0,...,q_{k-1})\in M\times...\times M\
\left|\
\begin{array}{@{}l@{}}
d(q_0,q_1)=\delta, \vspace{5pt} \\
\displaystyle\sum_{i\in\mathds{Z}_k\setminus\{0\}}\!\!\! d(q_i,q_{i+1})^2<\rho^2
\end{array}
\right.\right\}.
\end{align*}
For each $\bm{q}\in\Upsilon M$, we define the quantities
\begin{align}
\label{e:sigma}
\sigma(\bm{q})^2&:=(k-1)\!\!\!\sum_{i\in\mathds{Z}_k\setminus\{0\}}\!\!\!d(q_i,q_{i+1})^2>0,\\
\label{e:tau_0}
\tau_0(\bm{q})&:=0,\\
\label{e:tau_1}
\tau_1(\bm{q})&:=\frac{\delta}{\delta+\sigma(\bm{q})},\\
\label{e:tau_i}
\tau_{i}(\bm{q})&:=\tau_1(\bm{q})+ \frac{(i-1)(1-\tau_1(\bm{q}))}{k-1},\quad\forall i=2,...,k,
\end{align}
so that $0=\tau_0<\tau_1<...<\tau_k=1$.
We consider $\Upsilon M$ as a finite dimensional submanifold of $\Lambda M$, by identifying each $\bm{q}\in\Upsilon M$ with the unique periodic curve $\gamma_{\bm{q}}\in\Lambda M$ such that, for each $i=0,...,k-1$, the restriction $\gamma_{\bm{q}}|_{[\tau_i(\bm{q}),\tau_{i+1}(\bm{q})]}$ is the unique shortest geodesic joining $q_i$ and $q_{i+1}$.
A feature of $\Upsilon M$ that was missing in $\Lambda M$ is the smooth evaluation map
\begin{align}
\label{e:evaluation}
\mathrm{Ev}: \Upsilon M\to SM,
\qquad
\mathrm{Ev}(\bm{q})
:=
\exp_{q_0}^{-1}(q_1),
\end{align}
where we have denoted by $SM$ the unit tangent bundle of $(M,\delta^{-2}g)$. This map is injective in cohomology (Lemma~\ref{l:injectivity}). We choose two generators
\begin{align}
\label{e:omega}
&\omega\in \mathrm{Ev}^*(H^{2n-1}(SM))\cong\mathds{Z},\\
\label{e:alpha}
&\alpha\in H^{i(M)}(\Lambda M,\Lambda M^{<4\rho^2})\cong H^{i(M)}(\Lambda M,M)\cong\mathds{Z},
\end{align}
We will always fix a sufficiently small parameter $\delta\in(0,\rho)$ and a sufficiently large $k\in\mathds{N}$ for the space $\Upsilon M=\Upsilon_{\delta,k}M$ so that, according to Lemma~\ref{l:cohomology}, $\omega\smile j^*\alpha\neq 0$ in $H^{i(M)}(\Upsilon M,\Upsilon M^{<4\rho^2})$, where $j:(\Upsilon M,\Upsilon M^{<4\rho^2})\hookrightarrow(\Lambda M,\Lambda M^{<4\rho^2})$ is the inclusion. For each $b>0$, we set $\Upsilon M^{<b}:=\Lambda M^{<b}\cap\Upsilon M$, and denote by
$j_{b}:(\Upsilon M^{<b},\Upsilon M^{<4\rho^2})
\hookrightarrow
(\Upsilon M,\Upsilon M^{<4\rho^2})$
the inclusion. Given a non-trivial cohomology class $\mu\in H^d(\Upsilon M,\Upsilon M^{<4\rho^2})$ the associated min-max
\begin{align}
\label{e:minmax_delta_k}
c_g(\mu)=c_g(-\mu):=\inf\{b>0\ |\ j_b^*\mu\neq 0\}
\end{align}
is a critical value of the restricted energy $E|_{\Upsilon M}$, see Section~\ref{e:loop_space}. We can now state our second main theorem.
\begin{thm}
\label{t:main}
Let $M$ be a simply connected spin closed manifold of dimension $n\geq 2$ admitting a simple Zoll Riemannian metric, and $g$ a Riemannian metric on $M$. Then $c_g(j^*\alpha)=c_g(\omega\smile j^*\alpha)=:\ell^2$ if and only if $g$ is Zoll and the unit-speed geodesics of $(M,g)$ have minimal period $\ell$.
\end{thm}
In this theorem, if we drop the assumptions of $M$ being simply connected and spin, it is still true that $c_g(j^*\alpha)=c_g(\omega\smile j^*\alpha)=:\ell^2$ if $g$ is Zoll with unit-speed geodesics of minimal period $\ell$; however, if $c_g(j^*\alpha)=c_g(\omega\smile j^*\alpha)=:\ell^2$, our proof would only imply that $g$ is Besse and either $\ell$ or $\ell-2\delta$ is a common period of the unit-speed closed geodesics.
Actually, the only simply connected closed manifolds that are not spin and admit a Besse Riemannian metric have the same integral cohomology of even dimensional complex projective spaces $\mathds{C}\mathds{P}^{2m}$. Therefore, Theorem~\ref{t:main} applies to $S^n$, $\mathds{C} \mathds{P}^{2m+1}$, $\mathds{H} \mathds{P}^n$, and $\mathrm{Ca}\mathds{P}^2$. The theorem would also apply to all those closed manifolds admitting a simple Zoll Riemannian metric and having the integral cohomology of $S^n$, $\mathds{C} \mathds{P}^{2m+1}$, $\mathds{H} \mathds{P}^n$, or $\mathrm{Ca}\mathds{P}^2$; however, as we already mentioned, there is no known example of simply connected, spin, closed manifold different from $S^n$, $\mathds{C} \mathds{P}^{2m+1}$, $\mathds{H} \mathds{P}^n$, and $\mathrm{Ca}\mathds{P}^2$, and admitting a Besse Riemannian metric.
\subsection{Organization of the paper}
In Section~\ref{s:preliminaries} we provide some background on the energy functional of Besse Riemannian manifolds. In Section~\ref{s:covering} we prove Theorem~\ref{t:every_point}. In Section~\ref{e:loop_space} we study the variational theory of the energy functional in the finite dimensional loop space $\Upsilon M$. Finally, in Section~\ref{s:Zoll} we prove Theorem~\ref{t:main}.
\footnotesize
\subsection*{Acknowledgments}
Marco Mazzucchelli is grateful to Viktor Ginzburg, Basak G\"urel, and Nancy Hingston for discussions concerning the topic of this paper.
This material is based upon work supported by the National Science Foundation under Grant No.~DMS-1440140 while Marco Mazzucchelli was in residence at the Mathematical Sciences Research Institute in Berkeley, California, during the Fall 2018 semester. Stefan Suhr is supported by the SFB/TRR 191 ``Symplectic Structures in Geometry, Algebra and Dynamics'', funded by the Deutsche Forschungsgemeinschaft.
\normalsize
\section{Preliminaries}
\label{s:preliminaries}
Let $E:\Lambda M\to[0,\infty)$ be the energy functional of a closed Besse manifold $(M,g)$ of dimension $n\geq 2$. By a theorem due to Wadsley \cite{Wadsley:1975sp}, all unit-speed geodesics have a minimal common period $\ell>0$. In particular, each critical manifold
\[K^m:=\mathrm{crit}(E)\cap E^{-1}(m^2\ell^2),\]
where $m\in\mathds{N}=\{1,2,3,...\}$, is diffeomorphic to the unit tangent bundle $SM$.
As usual, we denote by $\mathrm{ind}(E,K^m)$ the Morse index of $E$ at any $\gamma\in K^m$, which is the number of negative eigenvalues of the symmetric operator associated to the Hessian $\mathrm{d}^2E(\gamma)$. We also denote by $\mathrm{nul}(E,K^m)=\dim\ker\mathrm{d}^2E(\gamma)-1$ the Morse nullity of $E$ at any $\gamma\in K^m$. Both $\mathrm{ind}(E,K^m)$ and $\mathrm{nul}(E,K^m)$ are independent of the choice of $\gamma$ within $K^m$.
The nullity of a closed geodesic is always bounded from above by $2n-2$; since $\dim(K^m)=2n-1\leq\mathrm{nul}(E,K^m)+1$, we readily infer that $\mathrm{nul}(E,K^m)=2n-2=\dim(K^m)-1$. Therefore, each critical manifold $K^m$ is non-degenerate, meaning that the restriction of the energy functional to the fibers of its normal bundle has non-degenerate Hessian. This, together with Bott's iteration theory \cite{Bott:1956sp}, implies that
\begin{equation}
\label{e:iter_idx}
\begin{split}
\mathrm{ind}(E ,K^m) & = m\,\mathrm{ind}(E ,K) + (m-1)(n-1),\\
\mathrm{nul}(E, K^m) & =2n-2,
\end{split}
\end{equation}
see \cite[Eq.~(13.1.1)]{Goresky:2009fq}. Actually, Wilking \cite{Wilking:2001pi} showed that the energy functional $E$ is always Morse-Bott when $(M,g)$ is Besse, meaning that the critical points are organized in critical manifolds $J\subset \mathrm{crit}(E)$ such that $\mathrm{nul}(E,J)=\dim(J)-1$. Moreover, Radeschi and Wilking \cite[page 941]{Radeschi:2017dz} proved that the minimal index
\begin{align*}
i(M):=\min\big\{\mathrm{ind}(E,\gamma)\ \big|\ \gamma\in\mathrm{crit}(E)\cap E^{-1}(0,\infty)\big\}
\end{align*}
is independent of the choice of a Besse Riemannian metric on $M$, and indeed only depends on the integral cohomology ring of $M$.
Now, let us assume that $(M,g)$ is an orientable Zoll Riemannian manifold. This readily implies that $i(M)=\mathrm{ind}(E,K)$ and
\begin{align*}
\mathrm{crit}(E)\cap E^{-1}(0,\infty) = \bigcup_{m\geq 1} K^m.
\end{align*}
A result of Goresky and Hingston \cite[Theorem~13.4(1)]{Goresky:2009fq} implies that, if one further assumes that $g$ is simple Zoll, the energy functional $E$ is perfect for the integral singular homology, and equivalently for the integral singular cohomology. This means that, for every integer $m\geq1$ and for all $\epsilon>0$ small enough, the homomorphism
\begin{align*}
H^*(\Lambda M,\Lambda M^{<m^2\ell^2})
\ttoup^{\mathrm{incl}^*}
H^*(\Lambda M^{<m^2\ell^2+\epsilon},\Lambda M^{<m^2\ell^2})
\end{align*}
is surjective. Moreover, by \cite[Proposition~13.2]{Goresky:2009fq}, the negative bundle of every critical manifold $K^m$ is oriented, which implies that
\begin{align*}
H^*(\Lambda M^{<m^2\ell^2+\epsilon},\Lambda M^{<m^2\ell^2})
\cong
H^{*-\mathrm{ind}(E ,K^m)}(SM).
\end{align*}
This, together with a usual gradient flow argument from Morse Theory, implies that cohomology of the free loop space $\Lambda M$ relative to the constant loops $M\subset \Lambda M$ is given by~\eqref{e:cohomology_loop_space}. Actually, for any Zoll Riemannian metric $g$ on $M$, the Morse index formula in~\eqref{e:cohomology_loop_space} becomes $\mathrm{ind}(E,K^m)=m\,i(M)+(m-1)(n-1)$. This readily implies that, on any closed manifold $M$ admitting a simple Zoll Riemannian metric, the energy functional $E:\Lambda M\to[0,\infty)$ of any Zoll Riemannian metric $g$ on $M$ is perfect even if $g$ is not simple Zoll.
\section{A min-max condition for covering with closed geodesics}
\label{s:covering}
Let $M$ be a manifold homeomorphic to an odd dimensional sphere $S^n$, $n\geq 3$. The inclusion $M\subset\Lambda M$ of the constant loops admits the evaluation map \[\mathrm{ev}:\Lambda M\to M,\qquad \mathrm{ev}(\gamma)=\gamma(0)\] as left inverse (this map should not be confused with the velocity evaluation map defined in~\eqref{e:evaluation}). This implies that the cohomology homomorphism
\begin{align*}
\mathrm{ev}^*:H^*(M)\hookrightarrow H^*(\Lambda M)
\end{align*}
is injective.
Ziller's computation \cite{Ziller:1977rp} of the homology of the free loop space of the compact rank-one symmetric spaces gives
\begin{align*}
H^d(\Lambda M,M)\cong
\left\{
\begin{array}{ll}
\mathds{Z}, & \mbox{if }d=m(n-1)\mbox{ or }d=m(n-1)+n,\mbox{ for }m\in\mathds{N}\cup\{0\}, \\
0, & \mbox{otherwise}.
\end{array}
\right.
\end{align*}
In particular, we have two non-trivial generators $\alpha_m\in H^{(2m-1)(n-1)}(\Lambda M,M)$ and $\beta_m\in H^{2m(n-1)+1}(\Lambda M,M)$, as claimed in~\eqref{e:classes_thm1}.
\begin{lem}\label{l:subordinated}
For each $m\in\mathds{N}$, we have \[\alpha_m\smile\mathrm{ev}^*\nu=\beta_m\] for a suitable generator $\nu\in H^n(M)$.
\end{lem}
\begin{proof}
Let $g$ be a Zoll Riemannian metric on $M$, and $E:\Lambda M\to[0,\infty)$ the associated energy functional. If $K=\mathrm{crit}(E)\cap E^{-1}(\ell^2)$ is the critical manifold of the prime closed geodesics, the other critical manifolds are $M=E^{-1}(0)$ and
\begin{align*}
K^m=\{\gamma^m\ |\ \gamma\in K\},\qquad
\forall m\in\mathds{N}.
\end{align*}
Here, as usual, we have denoted by $\gamma^m\in\Lambda M$ the $m$-th iterate of $\gamma$, which is defined by $\gamma^m(t)=\gamma(mt)$. The associated critical values are $m^2\ell^2=E(K^m)$. Let $G$ be any complete Riemannian metric on $\Lambda M$ (for instance the usual $W^{1,2}$-one induced by $g$). We denote by $\pi:N_m\to K^m$ the negative bundle of $K^m$, which is an orientable vector bundle of rank $\mathrm{ind}(E,K^m)$ whose fibers $\pi^{-1}(\gamma)$ are the negative eigenspaces of the self-adjoint Fredholm operator $H_\gamma$ on $\mathrm{T}_\gamma\Lambda M$ defined by
$G(H_\gamma \cdot,\cdot)=\mathrm{d}^2E(\gamma)$.
By means of the exponential map of $(\Lambda M,G)$, we can see the total space $N_m$ as a submanifold of $\Lambda M$ containing the critical manifold $K^m$ in its interior. Since $E$ is a Morse-Bott function, for all $\epsilon>0$ small enough the inclusion induces a cohomology isomorphism
\begin{align}
\label{e:local_homology}
H^*(\Lambda M^{<m^2\ell^2+\epsilon} , \Lambda M^{<m^2\ell^2})
\ttoup^{\mathrm{incl}_*}_{\cong}
H^*(N_m,\partial N_m).
\end{align}
The critical manifold $K^m$ is homeomorphic to $SM$ via the map $\gamma\mapsto\dot\gamma(0)/\|\dot\gamma(0)\|_g$. Since both the Euler characteristic $\chi(M)=\chi(S^n)$ and the cohomology group $H^1(M)\cong H^1(S^n)$ vanish, the Gysin sequence of $SM$ implies that $\pi^*:H^n(M)\rightarrow H^n(SM)$ is an isomorphism. Therefore, the evaluation map induces a cohomology isomorphism
\begin{align*}
\mathrm{ev}|_{K^m}^*:H^n(M)\toup^{\cong} H^n(K^m).
\end{align*}
Since the inclusion $K^m\subset N_m$ is a homotopy equivalence, we also have a cohomology isomorphism
\begin{align*}
\mathrm{ev}|_{N_m}^*:H^n(M)\toup^{\cong} H^n(N_m),
\end{align*}
which fits into the commutative diagram
\begin{equation*}
\begin{split}
\xymatrix{
H^n(\Lambda M) \ar[rr]^{\mathrm{incl}^*} & & H^n(N_m) \\ \\
\Big. H^n(M)\ar[uurr]_{\mathrm{ev}|_{N_m}^*}^{\cong} \ar@{^{(}->}[uu]^{\mathrm{ev}^*} & &
}
\end{split}
\end{equation*}
We set $\mathds{I}:=\{(2m-1)(n-1),2m(n-1)+1\}$.
The index formulas~\eqref{e:iter_idx}, together with $i(S^n)=n-1$, imply that
\begin{align*}
H^d(N_m,\partial N_m)\cong H^{d-\mathrm{ind}(E,K^m)}(K^m)\cong H^{d-(2m-1)(n-1)}(SS^n)\cong\mathds{Z},
\qquad \forall d\in\mathds{I}.
\end{align*}
This, together with the fact that $E$ is a perfect functional and that the relative cohomology groups $H^*(\Lambda M,M)$ have rank at most rank 1 in each degree, implies that the inclusion induces the cohomology isomorphism
\begin{align}
\label{e:incl_*}
\kappa_1^*: H^d(\Lambda M,\Lambda M^{<m^2\ell^2})
\ttoup^{\mathrm{incl}^*}_{\cong}
H^d(N_m,\partial N_m),
\qquad
\forall d\in\mathds{I}.
\end{align}
Let $\nu$ be a generator of $H^n(M)$.
We set
\begin{align*}
\mu:=\mathrm{ev}^*\nu\in H^n(\Lambda M),
\qquad
\mu':=\mathrm{ev}|_{N_m}^*\nu\in H^n(N_m).
\end{align*}
Let $\alpha'\in H^{(2m-1)(n-1)}(N_m,\partial N_m)$ be the Thom class of the orientable vector bundle $N_m\to K^m$ corresponding to an arbitrary orientation. The Thom isomorphism implies that $\alpha'\smile\mu'$ is a generator of $H^{2m(n-1)+1}(N_m,\partial N_m)$. Once again, since $E$ is a perfect functional, we have an isomorphism
\begin{align*}
\kappa_2^*:H^d(\Lambda M,\Lambda M^{<m^2\ell^2})
\ttoup^{\mathrm{incl}^*}_{\cong}
H^d(\Lambda M,M),
\qquad
\forall d\in\mathds{I},
\end{align*}
and we infer
\begin{align*}
(\kappa_2^*)^{-1}\beta_m
&=
(-1)^h
(\kappa_1^*)^{-1}(\alpha'\smile\mathrm{ev}|_{N_m}^*\nu)\\
&=
(-1)^h
(\kappa_1^*)^{-1}\alpha' \smile \mathrm{ev}^*\nu\\
&=
(-1)^h
(\kappa_2^*)^{-1}\alpha_m \smile \mathrm{ev}^*\nu\\
&=
(-1)^h
(\kappa_2^*)^{-1}(\alpha_m \smile \mathrm{ev}^*\nu).
\end{align*}
for some $h\in\{0,1\}$. Up to replacing $\nu$ with $-\nu$, we can assume that $h=0$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{t:every_point}]
Let us assume by contradiction that $\ell^2:=c_g(\alpha_m)=c_g(\beta_m)$ for some $m\in\mathds{N}$, but that for some $q\in M$ there is no $\gamma\in\mathrm{crit}(E)$ of energy $E(\gamma)=\ell^2$ with $\gamma(0)=q$. Under this latter assumption, the open subset
$
U:=\Lambda M\setminus \mathrm{ev}^{-1}(q)
$
is a neighborhood of the critical set $\mathrm{crit}(E)\cap E^{-1}(\ell^2)$. By Lemma~\ref{l:subordinated}, $\beta_m=\alpha_m\smile\mathrm{ev}^*\nu$ for some generator $\nu$ of $H^n(M)$, and therefore the classical Lusternik-Schnirelmann theorem (see, e.g., \cite[Theorem~1.1]{Viterbo:1997pi} for a modern account) implies that $(\mathrm{ev}^*\nu)|_{U}\neq0$ in $H^n(U)$.
Now, consider the commutative diagram
\begin{equation*}
\xymatrix{
H^n(M) \ar[rr]^{\mathrm{ev}^*}\ar[dd]_{\mathrm{incl}^*} && H^n(\Lambda M) \ar[dd]_{\mathrm{incl}^*}\\\\
H^n(M\setminus\{q\}) \ar[rr]^{\mathrm{ev}|_U^*} & & H^n(U)
}
\end{equation*}
Since $M$ has dimension $n$, the punctured manifold $M\setminus\{q\}$ has trivial cohomology group $H^n(M\setminus\{q\})$. This, together with the above commutative diagram, implies that $(\mathrm{ev}^*\nu)|_{U}=\mathrm{ev}^*(\nu|_{M\setminus\{q\}})=0$, contradicting Lusternik-Schnirelmann theorem.
\end{proof}
\section{The finite dimensional loop space $\Upsilon M$}
\label{e:loop_space}
Let $(M,g)$ be a closed Riemannian manifold of dimension $n\geq2$ with associated Riemannian distance $d:M\times M\to[0,\infty)$, injectivity radius $\rho=\mathrm{injrad}(M,g)>0$, and energy functional $E:\Lambda M\to[0,\infty)$.
For each $\delta\in(0,\rho)$ and $k\in\mathds{N}$, we consider the space $\Upsilon M=\Upsilon_{\delta,k}M$, which we identify with a subspace of $\Lambda M$ as we explained in the introduction.
The restriction of the energy functional $E_{\delta,k}:=E|_{\Upsilon M}$ can be expressed as
\begin{align*}
E_{\delta,k}(\bm{q})
&=
\int_0^1 \|\dot\gamma_{\bm{q}}(t)\|^2_g\,\mathrm{d} t
=
\sum_{i=0}^{k-1} \frac{d(q_i,q_{i+1})^2}{\tau_{i+1}(\bm{q})-\tau_i(\bm{q})}\\
&=
\frac{\delta^2}{\tau_1(\bm{q})}
+
\frac{k-1}{1-\tau_1(\bm{q})}
\sum_{i\in\mathds{Z}_k\setminus\{0\}} d(q_i,q_{i+1})^2\\
&=
\left(\delta+\sqrt{(k-1)\sum_{i\in\mathds{Z}_k\setminus\{0\}} d(q_i,q_{i+1})^2}\right)^2
.
\end{align*}
Here, the times $0=\tau_0(\bm{q})<...<\tau_{k}(\bm{q})=1$ are those defined in Equations \eqref{e:tau_0}, \eqref{e:tau_1} and \eqref{e:tau_i}.
For each $i\in\mathds{Z}_k$, we define $v_i^{\pm}(\bm{q})\in\mathrm{T}_{q_i}M$ by
\begin{align*}
v_i^{\pm}(\bm{q}):=\dot\gamma_{\bm{q}}(\tau_i(\bm{q})^{\pm}).
\end{align*}
The choice of $\tau_1$ that we made in~\eqref{e:tau_1} is such that, for all $\bm{q}\in\mathrm{crit}(E_{\delta,k})$, the corresponding curve $\gamma_{\bm{q}}$ has constant speed (even though $\dot\gamma_{\bm{q}}$ may not be smooth at times $\tau_0(\bm{q})=0$ and $\tau_1(\bm{q})$). More precisely, we have the following statement.
\begin{prop}
\label{p:critical_points}
The critical points of $E_{\delta,k}$ are precisely those $\bm{q}\in\Upsilon M$ such that $v_{0}^{-}(\bm{q})\in\{v_{0}^+(\bm{q}),-v_{0}^+(\bm{q})\}$, $v_{1}^{+}(\bm{q})\in\{v_{1}^-(\bm{q}),-v_{1}^-(\bm{q})\}$, and $v_i^-(\bm{q})=v_i^+(\bm{q})$ for all $i\in\mathds{Z}_k\setminus\{0,1\}$.
\end{prop}
\begin{proof}
Consider the functional
\begin{gather*}
F:\underbrace{M\times...\times M}_{\times k}\times(0,1)\to[0,\infty),
\\
F(\bm{q},\tau)=
\frac{1}{\tau} d(q_0,q_1)^2
+
\frac{k-1}{1-\tau}
\sum_{i\in\mathds{Z}_k\setminus\{0\}} d(q_i,q_{i+1})^2,
\end{gather*}
which is smooth on the subset $U\subset M^{\times k}\times (0,1)$ of all those points $(\bm{q},\tau)$ such that $d(q_i,q_{i+1})<\rho$ for all $i\in\mathds{Z}_k$. Notice that, for all $\bm{q}\in\Upsilon M$, we have $E(\gamma_{\bm{q}})=F(\bm{q},\tau_1(\bm{q}))$, and one can easily verify that $\tau_1(\bm{q})$ is the unique critical point and the global minimizer of the function $\tau\mapsto F(\bm{q},\tau)$. In the following, for each $\tau\in(0,1)$ we set
\begin{align*}
F_\tau:=F(\cdot,\tau).
\end{align*}
We denote by $SM$ the unit tangent bundle of $(M,\delta^{-2}g)$, that is,
\begin{align*}
SM=\big\{ (q,v)\in\mathrm{T} M\ \big|\ \|v\|_g=\delta \big\}.
\end{align*}
The space $\Upsilon M$ is diffeomorphic to the space $\Upsilon' M$ of those
\begin{align*}
\bm{q}'=(q_0,v_0,q_2,...,q_{k-2},q_{k-1})\in SM\times\underbrace{M\times...\times M}_{\times k-2},
\end{align*}
such that, if we set $q_1:=\exp_{q_{0}}(v_0)$, we have
\begin{align*}
\sum_{i\in\mathds{Z}_k\setminus\{0\}} \!\!\! d(q_i,q_{i+1})^2<\rho^2.
\end{align*}
The explicit diffeomorphism is
\begin{align*}
\iota:\Upsilon' M\toup^{\cong}\Upsilon M,
\qquad
\iota(q_0,v_0,q_2,...,q_{k-1})=\bm{q}=(q_0,q_1,q_2,...,q_{k-1}),
\end{align*}
and we have
\[
\frac{v_0}{\|v_0\|_g}=\frac{\dot\gamma_{\bm{q}}(0^+)}{\|\dot\gamma_{\bm{q}}(0^+)\|_g}.
\]
We consider the submersion $Q:SM\to M$, $Q(q,v)=\exp_q(v)$. The differential of $\iota$ is given by
\begin{align*}
\mathrm{d}\iota(\bm{q}')(z,\bm{w})=(\mathrm{d}\pi(q_{0},v_{0})z,\mathrm{d} Q(q_{0},v_{0})z,\bm{w}),
\end{align*}
where $\bm{w}=(w_2,...,w_{k-2})\in\mathrm{T}_{q_2}M\times...\times\mathrm{T}_{q_{k-1}}M$, and $z\in\mathrm{T}_{(q_0,v_0)}SM$. We set
\begin{align*}
\bm{q}=(q_0,q_1,q_2,...,q_{k-1}):=\iota(\bm{q}'),
\qquad
\tau_i:=\tau_i(\bm{q}),
\qquad
v_i^{\pm}:=\dot\gamma_{\bm{q}}(\tau_i^\pm),
\end{align*}
so that
\begin{equation}
\label{e:diff_E_iota}
\begin{split}
\mathrm{d}(F_{\tau_1}\circ\iota)(\bm{q}')(\bm{w},z)
=
&
\sum_{i\in\mathds{Z}_k\setminus\{0,1\}} 2 g( v_i^- - v_i^+, w_i)\\
& + 2 g( v_1^- - v_1^+, \mathrm{d} Q(q_{0},v_{0})z)\\
& + 2 g( v_{0}^- - v_{0}^+, \mathrm{d} \pi(q_{0},v_{0})z).
\end{split}
\end{equation}
By Equation~\eqref{e:diff_E_iota}, $\mathrm{d}(F_{\tau_1}\circ\iota)(\bm{q}')(0,\bm{w})=0$ for all $\bm{w}$ if and only if $v_i^- = v_i^+$ for all $i\in\mathds{Z}_k\setminus\{0,1\}$.
Notice that
\[\mathrm{d} Q(q_{0},v_{0})\big(\ker\mathrm{d}\pi(q_{0},v_{0})\big)=\mathrm{span}\{v_{1}^-\}^{\bot}.\]
Therefore, by Equation~\eqref{e:diff_E_iota}, $\mathrm{d}(F_{\tau_1}\circ\iota)(\bm{q}')(z,0)=0$ for all $z\in\ker\mathrm{d}\pi(q_{0},v_{0})$ if and only if $v_1^- - v_1^+ \bot \mathrm{span}\{v_{1}^-\}^{\bot}$, that is, $v_{1}^{+}\in\mathrm{span}\{v_{1}^-\}$.
Now, fix an arbitrary tangent vector
\[v\in\mathrm{span}\{v_{0}\}^{\bot}=\mathrm{span}\{v_{0}^+\}^{\bot},\]
and choose any smooth curve $\zeta:(-\epsilon,\epsilon)\to M$ such that $\zeta(0)=q_{0}$, $\dot\zeta(0)=v$, and $d(\zeta(t),q_1)=\delta$ for all $t\in(-\epsilon,\epsilon)$. We set $\xi(t):=\exp_{\zeta(t)}^{-1}(q_1)$, and \[z:=\tfrac{\mathrm{d}}{\mathrm{d} t}|_{t=0}(\zeta(t),\xi(t)).\] Notice that
\begin{align*}
\mathrm{d}\pi(q_{0},v_{0})z=v,\qquad \mathrm{d} Q(q_{0},v_{0})z=0.
\end{align*}
Therefore, by Equation~\eqref{e:diff_E_iota}, $\mathrm{d}(F_{\tau_1}\circ\iota)(\bm{q}')(z,0)=0$ for all $z$ of this form if and only if $v_{0}^- - v_{0}^+ \bot \mathrm{span}\{v_{0}^+\}^{\bot}$, that is, $v_{0}^{-}\in\mathrm{span}\{v_{0}^+\}$.
It remains one last case in order to cover all the possible choices of tangent vectors $z\in\mathrm{T}_{(q_{0},v_{0})}SM$, namely when $z$ is the value of the geodesic vector field at $(q_{0},v_{0})$. In this case,
\begin{align*}
\mathrm{d}\pi(q_{0},v_{0})z=v_{0}=\frac{\delta}{\|v_0^-\|_g}v_0^+,\qquad \mathrm{d} Q(q_{0},v_{0})z=\frac{\delta}{\|v_1^-\|_g} v_1^-=\frac{\delta}{\|v_0^-\|_g} v_1^-.
\end{align*}
Therefore, by Equation~\eqref{e:diff_E_iota}, $\mathrm{d}(F_{\tau_1}\circ\iota)(\bm{q}')(z,0)=0$ if and only if
\begin{align*}
g( v_1^- - v_1^+, v_1^-) - g( v_{0}^- - v_{0}^+, v_{0}^+)=0.
\end{align*}
Since $\|v_1^-\|_g=\|v_{0}^+\|_g$, this latter equation is verified if and only if
\begin{align}
\label{e_redundant}
g( v_1^+, v_1^-) = g( v_{0}^-, v_{0}^+).
\end{align}
Notice however that condition~\eqref{e_redundant} for the critical points of $E_{\delta,k}$ is redundant: indeed, any point $\bm{q}\in\Upsilon M$ such that $v_{0}^{-}\in\mathrm{span}\{v_{0}^+\}$, $v_{1}^{+}\in\mathrm{span}\{v_{1}^-\}$, $v_i^-=v_i^+$ for all $i\in\mathds{Z}_k\setminus\{0,1\}$, and $g( v_1^+, v_1^-) \neq g( v_{0}^-, v_{0}^+)$ would define a geodesic cusp $\gamma_{\bm{q}}$ and thus violate the uniqueness of the solution to the geodesic equation.
Summing up, we have proved that $\bm{q}\in\mathrm{crit}(E_{\delta,k})$ if and only if $v_{0}^{-}\in\mathrm{span}\{v_{0}^+\}$, $v_{1}^{+}\in\mathrm{span}\{v_{1}^-\}$, $v_i^-=v_i^+$ for all $i\in\mathds{Z}_k\setminus\{0,1\}$. In this case, we have
\begin{align*}
\sigma=\sigma(\bm{q})=\int_{\tau_1}^1 \|\dot\gamma_{\bm{q}}\|_g\,\mathrm{d} t,
\end{align*}
and therefore
\begin{align*}
\|v_0^+\|_g=\frac{\delta}{\tau_1}=\delta+\sigma=\int_{0}^1 \|\dot\gamma_{\bm{q}}\|_g\,\mathrm{d} t = \tau_1\|v_0^+\|_g+ (1-\tau_1)\|v_1^+\|_g,
\end{align*}
which implies that $\|v_0^+\|_g=\|v_1^+\|_g$.
\end{proof}
\begin{figure}
\begin{center}
\begin{small}
\input{_critical.pdf_tex}
\caption{\textbf{(a)} A global minimizer of $E_{\delta,k}$. \textbf{(b)} A critical point of $E_{\delta,k}$ corresponding to a closed geodesic. \textbf{(c)} The ``zig-zag'' critical point of $E_{\delta,k}$ corresponding to the same closed geodesic.}
\label{f:critical_points}
\end{small}
\end{center}
\end{figure}
Proposition~\ref{p:critical_points} shows that, beside the global minimizers $E_{\delta,k}^{-1}(4\delta^2)$ (Figure~\ref{f:critical_points}(a)) there are two other kinds of critical points of $E_{\delta,k}$: the closed geodesics smoothly parametrized with constant speed (Figure~\ref{f:critical_points}(b)), and the closed geodesics pa\-ram\-e\-trized with constant speed but non-smoothly with a zig-zag at times $\tau_0=0$ and $\tau_1$ (Figure~\ref{f:critical_points}(c)). If $\bm{q}'\in\mathrm{crit}(E_{\delta,k})$ corresponds to a smoothly parametrized closed geodesic $\gamma_{\bm{q}'}\in\mathrm{crit}(E)$ and $\bm{q}''\in\mathrm{crit}(E_{\delta,k})$ corresponds to the same closed geodesic parametrized with a zig-zag, their energies are related by
\begin{align}
\label{e:energy_zigzag}
E_{\delta,k}(\bm{q}'')^{1/2}=E_{\delta,k}(\bm{q}')^{1/2}+2\delta.
\end{align}
We partition the critical point set $\mathrm{crit}(E_{\delta,k})$ as the disjoint union
\begin{align*}
\mathrm{crit}(E_{\delta,k})
=
E_{\delta,k}^{-1}(4\delta^2)
\cup
K'
\cup
K'',
\end{align*}
where $K'=\mathrm{crit}(E)\cap E^{-1}(0,\infty)$, while $K''$ contains zig-zag closed geodesics.
The functional setting of the energy $E_{\delta,k}:\Upsilon M\to[4\delta^2,\infty)$ is suitable for Morse theory. Indeed, $E_{\delta,k}$ can be continuously extended to the boundary $\partial \Upsilon M\subset M\times...\times M$, and we have
\begin{align}
\label{e:max_E_delta_k}
E_{\delta,k}|_{\partial \Upsilon M}
\equiv
\sup E_{\delta,k}
=
\big( \delta + \sqrt{(k-1)}\rho \big)^2.
\end{align}
In particular, every sublevel set $\Upsilon M^{\leq b}$, for $b<\big( \delta + \sqrt{(k-1)}\rho \big)^2$, is a compact subset of $\Upsilon M$. Therefore, the classical min-max theorem is available in this setting: for each non-trivial cohomology class $\mu\in H_*(\Upsilon M,\Upsilon M^{<4\rho^2})$, the min-max value $c_g(\mu)$ defined in~\eqref{e:minmax_delta_k} is a critical value of $E_{\delta,k}$.
Actually, each closed geodesic $\gamma\in\mathrm{crit}(E)\cap E^{-1}(\ell^2)$, with $\ell>0$, is contained in $\Upsilon M=\Upsilon_{\delta,k}M$ if and only if
\begin{align*}
k
>
\overline k(\ell,\delta)
:=
1 + \frac{(\ell-\delta)^2}{\rho^{2}}
\end{align*}
Indeed, if we define
\begin{align*}
\tau_0 & :=0,\\
\tau_{i} & :=\frac{\delta}{\ell}+(i-1)\frac{ \ell-\delta}{ (k-1)\ell },
\qquad i=1,...,k-1,
\end{align*}
we readily verify that $\bm{q}=(\gamma(\tau_0),...,\gamma(\tau_{k-1}))$ belongs to $\Upsilon M$, and $\gamma_{\bm{q}}=\gamma$.
The following two lemmas compare the Morse indices in the settings $\Lambda M$ and $\Upsilon M$. The reader may skip their rather technical proofs on a first reading.
\begin{lem}
\label{l:indices}
Let $\gamma\in\mathrm{crit}(E)\cap E^{-1}(0,\infty)$ be a closed geodesic. For each $\delta\in(0,\rho)$ and $k> \overline k(E(\gamma)^{1/2},\delta)$, if $\bm{q}\in \mathrm{crit}(E_{\delta,k})$ is such that $\gamma_{\bm{q}}=\gamma$, then
\[\mathrm{ind}(E,\gamma)=\mathrm{ind}(E_{\delta,k},\bm{q}),
\qquad
\mathrm{nul}(E,\gamma)=\mathrm{nul}(E_{\delta,k},\bm{q}).\]
\end{lem}
\begin{proof}
We set $\theta_0:=0$, $\theta_k:=1$, and, for each $i=1,...,k-1$, we choose a time value $\theta_i\in(\tau_i(\bm{q}),\tau_{i+1}(\bm{q}))$ sufficiently close to $\tau_i(\bm{q})$ so that $d(\gamma(\theta_i),\gamma(\theta_{i+1}))<\rho$ for all $i\in\mathds{Z}_k$. Notice that
\begin{align*}
0
=
\tau_0(\bm{q})=\theta_0
<
\tau_1(\bm{q})<\theta_1
<
\tau_2(\bm{q})<\theta_2
<
...
<
\tau_k(\bm{q})=\theta_k
=1.
\end{align*}
We set $q_i':=\gamma(\theta_i)$ and $\bm{q}':=(q_0',...,q_{k-1}')$. We consider the function
\begin{align*}
F:M^{\times k}\to[0,\infty),
\qquad
F(\bm{p}')
=
\sum_{i=0}^{k-1}
\frac{d(p_i',p_{i+1}')^2}{\theta_{i+1}-\theta_i},
\end{align*}
which is smooth on an open neighborhood of $\bm{q}'$. Since $\bm{q}'$ is obtained by sampling the closed geodesic $\gamma$ at times $\theta_i$, it is a critical point of $F$. Since $d(\gamma(\theta_i),\gamma(\theta_{i+1}))<\rho$ and for all $i\in\mathds{Z}_k$, it is well known that
\begin{align*}
\mathrm{ind}(E,\gamma)=\mathrm{ind}(F,\bm{q}'),
\qquad
\mathrm{nul}(E,\gamma)=\mathrm{nul}(F,\bm{q}'),
\end{align*}
see, e.g., \cite[Theorem~16.2]{Milnor:1963rf}. For each $i\in\mathds{Z}_k$, we denote by $\Sigma_i\subset\mathrm{T}_{q_i'}M$ the hyperplane orthogonal to $\dot\gamma(\theta_i)$. By the definition of the Morse indices, there exist vector subspaces $\mathds{V},\mathds{W}\subset \Sigma_0\times...\times\Sigma_{k-1}$ of dimensions $\mathrm{ind}(E,\gamma)$ and $\mathrm{ind}(E,\gamma)+\mathrm{nul}(E,\gamma)$ respectively such that
\begin{equation}
\label{e:negative_spaces}
\begin{split}
\mathrm{d}^2F(\bm{q}')[\bm{v}',\bm{v}']<0,&
\qquad \forall \bm{v}'\in\mathds{V}\setminus\{0\},\\
\mathrm{d}^2F(\bm{q}')[\bm{v}',\bm{v}']\leq0,&
\qquad \forall \bm{v}'\in\mathds{W}.
\end{split}
\end{equation}
Now, we choose an open neighborhood $U\subset M^{\times k}$ of $\bm{q}'$ that is small enough so that, for all $\bm{p}'=(p_0',...,p_{k-1}')\in U$, we have
\begin{align*}
&\delta<d(p_0',p_1')<\rho,\\
&d(p_i',p_{i+1}')<\rho,\qquad \forall i\in\mathds{Z}_{k}\setminus\{0\}.
\end{align*}
We define $\zeta_{\bm{p}'}\in\Lambda M$ to be the piecewise broken geodesic such that each restriction $\zeta_{\bm{p}'}|_{[\theta_i,\theta_{i+1}]}$ is the shortest geodesic joining $p_i'$ and $p_{i+1}'$. Notice that $\zeta_{\bm{q}'}=\gamma_{\bm{q}}=\gamma$. We set
\begin{align*}
\nu_0(\bm{p}') &:= 0,\\
\nu_1(\bm{p}') &:= \frac{\theta_1 \delta}{d(p_0',p_1')},\\
\nu_i(\bm{p}') &:= \nu_1(\bm{p}') + (i-1)\frac{1-\nu_1(\bm{p}')}{k-1},
\qquad\forall i=2,...,k.
\end{align*}
Notice that $\nu_i(\bm{q}')=\tau_i(\bm{q})$. Therefore, up to replacing $U$ with a smaller neighborhood of $\bm{q}'$, for each $\bm{p}'\in U$, the curve $\zeta_{\bm{p}'}$ is smooth at each time $\nu_i(\bm{p}')$. This implies that the map
\begin{align*}
\psi:U\to\Upsilon M=\Upsilon_{\delta,k} M,
\qquad
\psi(\bm{p}')=(\zeta_{\bm{p}'}(\nu_0(\bm{p}')),...,\zeta_{\bm{p}'}(\nu_{k-1}(\bm{p}'))).
\end{align*}
is smooth. Notice that $\psi(\bm{q}')=\bm{q}$ and
\begin{align*}
E_{\delta,k}(\psi(\bm{p}'))\leq F(\bm{p}'),
\qquad
\forall \bm{p}'\in U,
\end{align*}
with equality if $\bm{p}'=\bm{q}'$. This, together with~\eqref{e:negative_spaces}, implies that
\begin{equation}
\label{e:negative_spaces_2}
\begin{split}
\mathrm{d}^2E_{\delta,k}(\bm{q})[\mathrm{d}\psi(\bm{q}')\bm{v}',\mathrm{d}\psi(\bm{q}')\bm{v}']\leq \mathrm{d}^2F(\bm{q}')[\bm{v}',\bm{v}']<0,&
\qquad \forall \bm{v}'\in\mathds{V}\setminus\{0\},\\
\mathrm{d}^2E_{\delta,k}(\bm{q})[\mathrm{d}\psi(\bm{q}')\bm{v}',\mathrm{d}\psi(\bm{q}')\bm{v}']\leq \mathrm{d}^2F(\bm{q}')[\bm{v}',\bm{v}']\leq0,&
\qquad \forall \bm{v}'\in\mathds{W}.
\end{split}
\end{equation}
Each $\bm{v}'=(v_0',...,v_{k-1}')\in\mathrm{T}_{\bm{q}'}M^{\times k}$ defines a unique continuous and piecewise smooth vector field $\xi_{\bm{v}'}$ along $\gamma$ such that, for all $i=0,...,k-1$, $\xi_{\bm{v}'}(\theta_i)=v_i'$ and the restriction $\xi_{\bm{v}'}|_{[\theta_i,\theta_{i+1}]}$ is a Jacobi vector field. Since $d(q_i,q_{i+1})<\rho$, the geodesic $\gamma|_{[\tau_i(\bm{q}),\tau_{i+1}(\bm{q})]}$ is the shortest one joining $q_i$ and $q_{i+1}$. This readily implies that the map
\begin{align*}
\Psi:\mathrm{T}_{\bm{q}'}M^{\times k}\to \mathrm{T}_{\bm{q}}M^{\times k},
\qquad
\Psi(\bm{v}')=(\xi_{\bm{v}'}(\tau_0(\bm{q})),...,\xi_{\bm{v}'}(\tau_{k-1}(\bm{q})))
\end{align*}
is injective on $\mathds{V}$ and on $\mathds{W}$. The differential of $\psi$ at $\bm{q}'$ is given by
\begin{align*}
\mathrm{d}\psi(\bm{q}')\bm{v}'
=
\Psi(\bm{v}') + \big(\dot\gamma(\tau_0(\bm{q}))\mathrm{d}\nu_0(\bm{q}')\bm{v}',...,\dot\gamma(\tau_{k-1}(\bm{q}))\mathrm{d}\nu_{k-1}(\bm{q}')\bm{v}' \big)
\end{align*}
Consider a non-zero $\bm{v}'\in\mathds{V}\cup\mathds{W}$, and set $\bm{v}=(v_0,...,v_{k-1}):=\Psi(\bm{v}')$. By the injectivity of $\Psi$, at least one component of $\bm{v}$, say $v_i$, is non-zero. Since both tangent vectors $v_{i-1}'$ and $v_i'$ are orthogonal to $\dot\gamma$, the whole Jacobi field $\xi_{\bm{v}'}|_{[\theta_{i-1},\theta_i]}$ is pointwise orthogonal to $\dot\gamma$, and so is $v_i=\xi_{\bm{v}'}(\tau_i(\bm{q}))$. Therefore, $v_i+\dot\gamma(\tau_{i}(\bm{q}))\mathrm{d}\nu_{i}(\bm{q}')\bm{v}'$ is non-zero, which shows that the differential $\mathrm{d}\psi(\bm{q}')$ is injective on both $\mathds{V}$ and $\mathds{W}$. This, together with~\eqref{e:negative_spaces_2}, implies that
\begin{align*}
\mathrm{ind}(E_{\delta,k},\bm{q}) & \geq \mathrm{ind}(E,\gamma),\\
\mathrm{ind}(E_{\delta,k},\bm{q})+\mathrm{nul}(E_{\delta,k},\bm{q}) & \geq \mathrm{ind}(E,\gamma)+\mathrm{nul}(E,\gamma).
\end{align*}
Since $\Upsilon M$ is a subspace of $\Lambda M$, the opposite inequalities hold as well.
\end{proof}
\begin{lem}
\label{l:indices_zigzag}
Let $\gamma\in\mathrm{crit}(E)\cap E^{-1}(0,\infty)$ be a closed geodesic. For each $\delta\in(0,\rho)$ and integer $k> \overline k(E(\gamma)^{1/2}+2\delta,\delta)$, let $\bm{q}'\in \mathrm{crit}(E_{\delta,k})$ be such that $\gamma_{\bm{q}'}=\gamma$, and $\bm{q}''\in \mathrm{crit}(E_{\delta,k})$ be the associated zig-zag critical point, i.e.\ $q_0''=q_0'$, $q_1''=q_1'$, and $E_{\delta,k}(\bm{q}'')^{1/2}=E_{\delta,k}(\bm{q}'')^{1/2}+2\delta$. Then
\begin{align*}
\mathrm{ind}(E_{\delta,k},\bm{q}') & \leq \mathrm{ind}(E_{\delta,k},\bm{q}''),
\\
\mathrm{ind}(E_{\delta,k},\bm{q}')+\mathrm{nul}(E_{\delta,k},\bm{q}') & \leq \mathrm{ind}(E_{\delta,k},\bm{q}'')+\mathrm{nul}(E_{\delta,k},\bm{q}'').
\end{align*}
\end{lem}
\begin{proof}
The proof is somewhat analogous to the one of Lemma~\ref{l:indices}, but requires some extra ingredients. We consider the time values
$1+\tau_1(\bm{q}') =: \sigma_1
>
\sigma_2
>
...
>\sigma_k :=0$
such that
\begin{align*}
\gamma_{\bm{q}''}(t\tau_{i+1}(\bm{q}'')+(1-t)\tau_i(\bm{q}''))
=
\gamma(t\sigma_{i+1}+(1-t)\sigma_i),
\qquad
\forall t\in[0,1].
\end{align*}
We choose arbitrary values
$
0=:\theta_0<\theta_1<...<\theta_k:=1
$
such that
\begin{align*}
&\{\theta_1,...,\theta_{k-1}\}
\cap
\{\sigma_1\ \mathrm{mod}\ 1,...,\sigma_{k-1}\ \mathrm{mod}\ 1\}=\varnothing,\\
&d(\gamma_{\bm{q}'}(\theta_0),\gamma_{\bm{q}'}(\theta_1))>\delta,\\
&d(\gamma_{\bm{q}'}(\theta_i),\gamma_{\bm{q}'}(\theta_{i+1}))<\rho,\quad
\forall i=0,...,k-1.
\end{align*}
The function
\begin{align*}
F:M^{\times k}\to[0,\infty),
\qquad
F(\bm{p})
=
\sum_{i=0}^{k-1}
\frac{d(p_i,p_{i+1})^2}{\theta_{i+1}-\theta_i},
\end{align*}
is smooth on an open neighborhood of $\bm{q}:=(\gamma_{\bm{q}'}(\theta_0),...,\theta_{\bm{q}'}(\theta_{k-1}))$. Since $\bm{q}$ is obtained by sampling the closed geodesic $\gamma$ at times $\theta_i$, it is a critical point of $F$. By Lemma~\ref{l:indices} and \cite[Theorem~16.2]{Milnor:1963rf}, we have
\begin{align*}
\mathrm{ind}(E_{\delta,k},\bm{q}')
=
\mathrm{ind}(E,\gamma)
=
\mathrm{ind}(F,\bm{q}).
\end{align*}
Therefore, if we denote by $\Sigma_i\subset\mathrm{T}_{q_i}M$ the hyperplane orthogonal to $\dot\gamma(\theta_i)$, we can find vector subspaces $\mathds{V},\mathds{W}\subset \Sigma_1\times...\times\Sigma_{k-1}$ of dimensions $\mathrm{ind}(E,\gamma)$ and $\mathrm{ind}(E,\gamma)+\mathrm{nul}(E,\gamma)$ respectively such that
\begin{equation}
\label{e:negative_eigespaces_zigzag}
\begin{split}
\mathrm{d}^2F(\bm{q})[\bm{v},\bm{v}]<0,&
\qquad \forall \bm{v}\in\mathds{V}\setminus\{0\},\\
\mathrm{d}^2F(\bm{q})[\bm{v},\bm{v}]\leq0,&
\qquad \forall \bm{v}\in\mathds{W}.
\end{split}
\end{equation}
We choose an open neighborhood $U\subset M^{\times k}$ of $\bm{q}$ that is small enough so that $d(p_0,p_{1})>\delta$ and $d(p_i,p_{i+1})<\rho$ for all $\bm{p}=(p_0,...,p_{k-1})\in U$ and $i\in\mathds{Z}_{k}\setminus\{0\}$.
We define $\overline\beta_{\bm{p}}\in\Lambda M$ to be the piecewise broken geodesic such that each restriction $\overline\beta_{\bm{p}}|_{[\theta_i,\theta_{i+1}]}$ is the shortest geodesic joining $p_i$ and $p_{i+1}$. Notice that $F(\bm{p})=E(\overline\beta_{\bm{p}})$. Moreover, $\overline\beta_{\bm{q}}=\gamma_{\bm{q}'}=\gamma$, which implies
\begin{align*}
F(\bm{q})=E(\overline\beta_{\bm{q}})=E_{\delta,k}(\bm{q}').
\end{align*}
We set $\overline\nu(\bm{p}):=\delta\theta_1 d(p_0,p_1)^{-1}\in(0,\theta_1)$ and notice that, since the restriction $\overline\beta_{\bm{p}}|_{[0,\theta_1]}$ is a geodesic, we have
\[
d\big(\overline\beta_{\bm{p}}(0),\overline\beta_{\bm{p}}(\overline\nu(\bm{p}))\big)
=
\delta.
\]
We set
\begin{align*}
\nu(\bm{p})
:=
\frac
{\delta}
{\delta+\sqrt{(1-\overline\nu(\bm{p}))\int_{\overline\nu(\bm{p})}^{1} \|\dot{\overline\beta}_{\bm{p}}(t)\|_g^2\,\mathrm{d} t}}.
\end{align*}
We define $\beta_{\bm{p}}\in\Lambda M$ so that the restrictions $\beta_{\bm{p}}|_{[0,\nu(\bm{p})]}$ and $\beta_{\bm{p}}|_{[\nu(\bm{p}),1]}$ are affine reparametrizations of $\overline\beta_{\bm{p}}|_{[0,\overline\nu(\bm{p})]}$ and $\overline\beta_{\bm{p}}|_{[\overline\nu(\bm{p}),1]}$ respectively; namely,
\begin{align*}
\beta_{\bm{p}}(t)
:=
\left\{
\begin{array}{lll}
\overline\beta_{\bm{p}}\big(t\,\tfrac{\overline\nu(\bm{p})}{\nu(\bm{p})}\big), & & t\in[0,\nu(\bm{p})], \vspace{5pt}\\
\overline\beta_{\bm{p}}\big(\overline\nu(\bm{p})+ (t-\nu(\bm{p}))\,\tfrac{1-\overline\nu(\bm{p})}{1-\nu(\bm{p})}\big), & & t\in[\nu(\bm{p}),1].
\end{array}
\right.
\end{align*}
The choice of this reparametrization guarantees that
\begin{align*}
E(\beta_{\bm{p}})\leq E(\overline\beta_{\bm{p}}) = F(\bm{p}).
\end{align*}
Moreover, $\nu(\bm{q})=\overline\nu(\bm{q})=\tau_1(\bm{q}')$ and $\beta_{\bm{q}}=\overline\beta_{\bm{q}}=\gamma_{\bm{q}'}=\gamma$, and therefore
\begin{align*}
E(\beta_{\bm{q}})=F(\bm{q}).
\end{align*}
We define $\overline\alpha_{\bm{p}}:\mathds{R}/(1+2\nu(\bm{p}))\mathds{Z}\to M$ by suitably adding a zig-zag to $\beta_{\bm{q}}$ as follows
\begin{align*}
\overline\alpha_{\bm{p}}(t)
:=
\left\{
\begin{array}{lll}
\beta_{\bm{p}}(t), & & t\in[0,\nu(\bm{p})], \vspace{5pt}\\
\beta_{\bm{p}}(2\nu(\bm{p})-t), & & t\in[\nu(\bm{p}),2\nu(\bm{p})],\vspace{5pt}\\
\beta_{\bm{p}}(t-2\nu(\bm{p})), & & t\in[2\nu(\bm{p}),1+2\nu(\bm{p})],
\end{array}
\right.
\end{align*}
and we define $\alpha_{\bm{p}}\in\Lambda M$ by $\alpha_{\bm{p}}(t)=\overline\alpha_{\bm{p}}(t\,(1+2\nu(\bm{p})))$. The energies of $\alpha_{\bm{p}}$ and $\beta_{\bm{p}}$ are related by
\begin{align*}
E(\alpha_{\bm{p}}) = (2\delta + E(\beta_{\bm{p}})^{1/2})^2.
\end{align*}
Moreover, $\alpha_{\bm{q}}=\gamma_{\bm{q}''}$.
We set
\begin{align*}
\eta_i(\bm{p}) &
:=
\frac{k-i}{k-1}
\left(
1 + \nu(\bm{p})
\right),
\qquad\forall i=1,...,k,
\end{align*}
and notice that $\eta_i(\bm{q})=\sigma_i$.
Up to replacing $U$ with a smaller neighborhood of $\bm{q}$, for each $\bm{p}\in U$ the curve $\beta_{\bm{p}}$ is smooth at each time $\eta_i(\bm{p})$. Therefore, the map
\begin{align*}
\psi:U\to\Upsilon M=\Upsilon_{\delta,k} M,
\qquad
\psi(\bm{p})=(\beta_{\bm{p}}(0),\beta_{\bm{p}}(\eta_1(\bm{p})),\beta_{\bm{p}}(\eta_2(\bm{p})),...,\beta_{\bm{p}}(\eta_{k-1}(\bm{p})))
\end{align*}
is smooth, and satisfies
\[\psi(\bm{q})=\bm{q}''\in\mathrm{crit}(E_{\delta,k}).\] Notice that
$E_{\delta,k}(\psi(\bm{p}))\leq E(\alpha_{\bm{p}})$
with equality if $\bm{p}=\bm{q}$. This, together with the other energy inequalities pointed out so far, provides
\begin{align*}
E_{\delta,k}(\psi(\bm{p})) & \leq (2\delta+F(\bm{p})^{1/2})^2,\quad \forall\bm{p}\in U,\\
E_{\delta,k}(\psi(\bm{q})) & = E_{\delta,k}(\bm{q}'') = (2\delta+F(\bm{q})^{1/2})^2.
\end{align*}
This, together with the fact that $\psi(\bm{q})=\bm{q}''$ and $\bm{q}$ are critical points of $E_{\delta,k}$ and $F$ respectively, implies
\begin{align*}
\mathrm{d}^2E_{\delta,k}(\bm{q}'')[\mathrm{d}\psi(\bm{q})\bm{v},\mathrm{d}\psi(\bm{q})\bm{v}]
\leq
\frac{2\delta+F(\bm{q})^{1/2}}{F(\bm{q})^{1/2}} \,\mathrm{d}^2 F(\bm{q})[v,v].
\end{align*}
Therefore, by~\eqref{e:negative_eigespaces_zigzag}, we infer
\begin{align*}
\mathrm{d}^2E_{\delta,k}(\bm{q}'')[\mathrm{d}\psi(\bm{q})\bm{v},\mathrm{d}\psi(\bm{q})\bm{v}]<0,&
\qquad \forall \bm{v}\in\mathds{V}\setminus\{0\},\\
\mathrm{d}^2E_{\delta,k}(\bm{q}'')[\mathrm{d}\psi(\bm{q})\bm{v},\mathrm{d}\psi(\bm{q})\bm{v}]\leq0,&
\qquad \forall \bm{v}\in\mathds{W},
\end{align*}
which provides the following lower bounds for the Morse indices
\begin{align*}
\mathrm{ind}(E_{\delta,k},\bm{q}'') \geq \dim(\mathrm{d}\psi(\bm{q})\mathds{V}),
\qquad
\mathrm{ind}(E_{\delta,k},\bm{q}'')+\mathrm{nul}(E_{\delta,k},\bm{q}'') \geq \dim(\mathrm{d}\psi(\bm{q})\mathds{W}).
\end{align*}
Finally, the same argument as in the proof of Lemma~\ref{l:indices} implies that $\mathrm{d}\psi(\bm{q})$ is injective on both $\mathds{V}$ and $\mathds{W}$, i.e.
\begin{align*}
\dim(\mathrm{d}\psi(\bm{q})\mathds{V})&= \dim(\mathds{V}) = \mathrm{ind}(E_{\delta,k},\bm{q}'),
\\
\dim(\mathrm{d}\psi(\bm{q})\mathds{W})&=\dim(\mathds{W})=\mathrm{ind}(E_{\delta,k},\bm{q}')+\mathrm{nul}(E_{\delta,k},\bm{q}').
\qedhere
\end{align*}
\end{proof}
\section{Zoll Riemannian metrics}
\label{s:Zoll}
\subsection{The evaluation map on $\Upsilon M$}
\label{s:evaluation}
Let us quickly prove the following property of the evaluation map $\mathrm{Ev}:\Upsilon M\to SM$ defined in~\eqref{e:evaluation}. Here, as before, we denote by $SM$ the unit tangent bundle of $(M,\delta^{-2}g)$.
\begin{lem}
\label{l:injectivity}
For each $b\geq4\delta^2$, the cohomology homomorphism \[\mathrm{Ev}^*:H^*(SM)\hookrightarrow H^*(\Upsilon M^{\leq b})\] is injective.
\end{lem}
\begin{proof}
We define the homeomorphism
\begin{align*}
\iota:SM\to E_{\delta,k}^{-1}(4\delta^2)\subset\Upsilon M,\qquad \iota(q_0,v_0)=\bm{q},
\end{align*}
where $\bm{q}=(q_0,...,q_{k-1})$ is the unique element in $E_{\delta,k}^{-1}(4\delta^2)$ such that $\exp_{q_{0}}(v_0)=q_1$.
Namely, $\gamma_{\bm{q}}$ is the periodic curve such that $\tau_1(\bm{q})=1/2$, $\dot\gamma_{\bm{q}}(0^+)=2v_{0}$, and
$\gamma_{\bm{q}}(t)=\gamma_{\bm{q}}( 1-t)$ for all $t\in[0,1/2]$, as in Figure~\ref{f:critical_points}(a). Since the composition $\mathrm{Ev}\circ\iota$ is the identity, the lemma follows for $b=4\delta^2$. The lemma readily follows also for any $b>4\delta^2$, since $E_{\delta,k}^{-1}(4\delta^2)\subset\Upsilon M^{\leq b}$.
\end{proof}
As in~\eqref{e:omega}, we denote by $\omega$ the generator of $\mathrm{Ev}^*(H^{2n-1}(SM))$. The following lemma is the main ingredient for the proof of Theorem~\ref{t:main}.
\begin{lem}
\label{l:main_lemma}
Assume that there exists a cohomology class $\mu\in H^d(\Upsilon M,\Upsilon M^{\leq4\rho^2})$ such that $\omega\smile\mu\neq0$ in $H^{d+2n-1}(\Upsilon M,\Upsilon M^{\leq4\rho^2})$. If
\[c_g(\mu)=c_g(\omega\smile\mu)=:\ell^2,\]
then $g$ is a Besse manifold, and either $\ell$ or $\ell-2\delta$ is a common multiple of the periods of the unit-speed closed geodesics of $(M,g)$. Moreover, the critical set \[K:=\mathrm{crit}(E)\cap(E^{-1}(\ell^2)\cup E^{-1}((\ell-2\delta)^2)),\]
has Morse index $\mathrm{ind}(E,K)\leq d$.
\end{lem}
\begin{proof}
Assume by contradiction that $c_g(\mu)=c_g(\omega\smile\mu)=:\ell^2$, but there exists $(q,v)\in SM$ such that the unit-speed geodesic \[\gamma(t)=\exp_q(tv/\|v\|_g)\] is either not periodic, or it is periodic but neither $\ell$ nor $\ell-2\delta$ are multiples of its minimal period. By~\eqref{e:energy_zigzag}, the condition on $\ell-2\delta$ implies that none of the zig-zag critical points $\bm{q}\in K'' \cap E^{-1}(\ell^2)$ satisfies $\mathrm{Ev}(\bm{q})=(q,v)$.
Therefore, the open subset
\begin{align*}
U:=\big\{\bm{q}=(q_0,q_1,...,q_{k-1})\in\Upsilon M\ \big|\ (q_0,\exp_{q_0}^{-1}(q_1))\neq(q,v) \big\}
\end{align*}
contains the set of critical points $\mathrm{crit}(E_{\delta,k})\cap E^{-1}(\ell^2)$, and the classical Lusternik-Schnirelmann's theorem \cite[Theorem~1.1]{Viterbo:1997pi} implies that the cohomology class $\omega|_{U}\in H^{2n-1}(U)$ is non-zero. Consider the commutative diagram
\begin{equation*}
\xymatrix{
U \ar@{^{(}->}[rr]^{\mathrm{incl}}\ar[drr]_{\mathrm{Ev}|_U} & & \Upsilon M
\ar[d]^{\mathrm{Ev}} \\
& & SM
}
\end{equation*}
Since $\omega$ is the generator of the image $\mathrm{Ev}^*(H^{2n-1}(SM))$, $\omega|_{U}$ is the generator of the image $\mathrm{Ev}|_U^*(H^{2n-1}(SM))$. However, the homomorphism $\mathrm{Ev}|_U^*:H^{2n-1}(SM)\to H^{2n-1}(U)$ is zero, since the map $\mathrm{Ev}|_U$ is not surjective. This implies that $\omega|_U=0$ in $H^{2n-1}(U)$, which is a contradiction.
So far, we have proved that $g$ is Besse, and $\ell$ or $\ell-2\delta$ is a common multiple of the periods of the unit-speed geodesics.
Now, consider the critical sets
\begin{align*}
K & :=\mathrm{crit}(E)\cap\big(E^{-1}(\ell^2)\cup E^{-1}((\ell-2\delta)^2)\big),\\
K_{\delta,k} & :=\mathrm{crit}(E_{\delta,k})\cap E_{\delta,k}^{-1}(\ell^2)
\end{align*}
Let $\epsilon>0$ be small enough so that $(\ell^2,\ell^2+\epsilon)$ does not contain critical values of $E_{\delta,k}$. Since $c_g(\mu)=\ell$, the relative cohomology group
\begin{align*}
H^d(\Upsilon M^{<\ell^2+\epsilon},\Upsilon M^{<\ell^2})\cong H^{d-\mathrm{ind}(E_{\delta,k},K_{\delta,k})}(K_{\delta,k})
\end{align*}
is nontrivial. In particular
\begin{align}
\label{e:ind_K_delta_k}
\mathrm{ind}(E_{\delta,k},K_{\delta,k})\leq d.
\end{align}
Since $\delta<\rho$, $2\delta$ is smaller than the minimal period of the unit-speed geodesics of $(M,g)$. This readily implies that the values $\ell$ and $\ell-\delta$ cannot both be common periods for the unit-speed geodesics of $(M,g)$, and we have two possible cases:
\begin{itemize}
\item If $\ell$ is a common period for the unit-speed geodesics of $(M,g)$, then $K_{\delta,k}$ does not contain zig-zag closed geodesics, and indeed $K=K_{\delta,k}$.
\item If $\ell-2\delta$ is a common period for the unit-speed geodesics of $(M,g)$, then $K_{\delta,k}$ contains only zig-zag closed geodesics, and more precisely of those closed geodesics contained in $K$.
\end{itemize}
In both cases, Lemmas~\ref{l:indices} and~\ref{l:indices_zigzag}, together with the inequality~\eqref{e:ind_K_delta_k}, imply
\[
\mathrm{ind}(E,K)
\leq
\mathrm{ind}(E_{\delta,k},K_{\delta,k})
\leq
d.
\qedhere
\]
\end{proof}
\subsection{Two subordinated homology classes in the Zoll case}
In this subsection, we will consider a Zoll Riemannian manifold, and prove the ``if'' claim in Theorem~\ref{t:main}.
\begin{lem}\label{l:Zoll}
Let $M$ be a closed manifold of dimension $n\geq 2$ admitting a simple Zoll Riemannian metric, and $g$ a Zoll Riemannian metric on $M$ whose unit-speed closed geodesics have minimal period $\ell>0$. For each $\delta\in(0,\rho)$ and for each integer $k>\overline k(\ell,\delta)$, we consider the space $\Upsilon M=\Upsilon_{\delta,k}M$. For each $\epsilon\in(4\delta^2,\ell^2)$,
there exists a relative homology class
\[h\in H_{i(M)+2n-1}(\Upsilon M^{\leq \ell^2},\Upsilon M^{< \epsilon})\]
such that $h$ and $h\frown \omega|_{\Upsilon M^{\leq\ell^2}}$ are not in the kernel of the homomorphism
\begin{align*}
H_{*}(\Upsilon M^{\leq \ell^2},\Upsilon M^{<\epsilon})
\ttoup^{\mathrm{incl}_*}
H_{*}(\Lambda M,\Lambda M^{<\epsilon}).
\end{align*}
\end{lem}
\begin{proof}
Let $K:=\mathrm{crit}(E)\cap E^{-1}(\ell^2)$ be the critical manifold of the non-iterated closed geodesics. By Lemma~\ref{l:indices}, for each $\delta\in(0,\rho)$ and integer $k>\overline k(\ell,\delta)$, we have $K\subset\Upsilon M:=\Upsilon_{\delta,k}M$ and, for each $\gamma_{\bm{q}}\in K$,
\begin{equation}
\label{e:same_indices}
\begin{split}
\mathrm{ind}(E,\gamma_{\bm{q}})&=\mathrm{ind}(E_{\delta,k},\bm{q})=i(M),
\\
\mathrm{nul}(E,\gamma_{\bm{q}})&=\mathrm{nul}(E_{\delta,k},\bm{q})=2n-2.
\end{split}
\end{equation}
We denote by $G$ the Riemannian metric on $\Upsilon M$ induced by $g$, i.e.
\begin{align*}
G(\bm{v},\bm{w})=\sum_{i\in\mathds{Z}_{k}} g(v_i,w_i),
\qquad
\forall\bm{v},\bm{w}\in\mathrm{T}_{\bm{q}}\Upsilon M.
\end{align*}
Let $\pi:N\to K$ be the negative bundle of $E_{\delta,k}$ at $K$. Namely, for each $\bm{q}\in K$, the fiber $\pi^{-1}(\bm{q})\subset\mathrm{T}_{\bm{q}}\Upsilon M$ is the negative eigenspace of the symmetric linear map $H_{\bm{q}}:\mathrm{T}_{\bm{q}}\Upsilon M\to\mathrm{T}_{\bm{q}}\Upsilon M$ defined by
$G( H_{\bm{q}}\,\cdot,\cdot)= \mathrm{d}^2E_{\delta,k}(\bm{q})$.
The rank of this vector bundle is $i(M)$, according to~\eqref{e:same_indices}.
For each $r>0$, we set $N_r\subset N$ to be the $r$-neighborhood of the 0-section, measured with respect to $G$. With a slight abuse of notation, we still denote by $\exp$ the exponential map of $(\Upsilon M,G)$. We choose $r>0$ to be small enough so that $\exp|_{N_r}$ is a well defined diffeomorphism onto a neighborhood of $K$ in $\Upsilon M$, and $E_{\delta,k}(\exp_{\bm{q}}(\bm{v}))<E_{\delta,k}(\bm{q})$ for all $(\bm{q},\bm{v})\in N_r$ with $\bm{v}\neq0$. Since $E$ has no critical values in the interval $(\ell^2, (\ell+\delta)^2)$, the arrows in the following commutative diagram are isomorphisms
\begin{align*}
\xymatrix{
H_*(N_r,\partial N_r) \ar[rr]^{\exp_*\ \ \ \ \ \ \ \ \ }_{\cong\ \ \ \ \ \ \ \ \ }\ar[ddrr]_{\exp_*}^{\cong} & & H_{*}(\Upsilon M^{\leq\ell^2},\Upsilon M^{<\ell^2})\ar[dd]^{\mathrm{incl}_*}_{\cong} \\\\
& & H_{*}(\Lambda M^{<(\ell+\delta)^2},\Lambda M^{<\ell^2})
}
\end{align*}
see \cite[Theorem~D.2]{Goresky:2009fq}. Since $E$ is a perfect functional, the exponential map also induces an injective homomorphism
\begin{align*}
\exp_*:H_*(N_r,\partial N_r)\hookrightarrow H_{*}(\Lambda M,\Lambda M^{<\ell^2})
\end{align*}
Since $\ell^2$ is the smallest positive critical value of $E$, the restriction $E_{\delta,k}$ has no critical values in the interval $(4\delta^2,\ell^2)$. For each $\epsilon\in(4\delta^2,\ell^2)$, if we denote by $\phi_t$ the anti-gradient flow of $E_{\delta,k}$, we can fix $t>0$ large enough so that
\[\phi_t\circ\exp(\partial N_r)\subset\Upsilon M^{<\epsilon}.\]
If we set $\iota:=\phi_t\circ\exp$, the induced homomorphisms $\iota_*$ and $\mathrm{incl}_*\circ \iota_*$ in the following commutative diagram must be injective
\begin{align*}
\xymatrix{
\Big. H_*(N_r,\partial N_r)\ \ar@{^{(}->}[rr]^{\exp_*} \ar@{^{(}->}[dd]_{\iota_*} & & H_{*}(\Lambda M,\Lambda M^{<\ell^2}) \\\\
H_{*}(\Upsilon M^{\leq\ell^2},\Upsilon M^{<\epsilon})\ \ar[rr]^{\mathrm{incl}_*}
& & H_{*}(\Lambda M,\Lambda M^{<\epsilon}) \ar[uu]_{\mathrm{incl}_*}^{\cong}
}
\end{align*}
Since the closed geodesics in $K$ are not iterated, the negative bundle $N\to K$ is orientable. If $\tau\in H^{i(M)}(N_r,\partial N_r)$ denotes its Thom class with respect to any orientation, we have a Thom isomorphism
\begin{align*}
H^*(N_r)\to H^{*+i(M)}(N_r,\partial N_r),
\qquad
\mu\mapsto \tau\smile \mu.
\end{align*}
If we denote by $\omega'$ the generator of $H^{2n-1}(N_r)\cong H^{2n-1}(SM)$, and by $h'$ the generator of $H_{i(M)+2n-1}(N_r,\partial N_r)$, then $h'\frown\omega'$ is the generator of $H_{i(M)}(N_r,\partial N_r)$. Consider the evaluation map $\mathrm{Ev}:\Upsilon M^{\leq\ell^2}\to SM$ of Equation~\eqref{e:evaluation}, which is injective in cohomology according to Lemma~\ref{l:injectivity}. If we denote by $0_N\subset N$ the 0-section of $N$, the composition $\mathrm{Ev}\circ \iota|_{0_N}:0_N\to SM$ is clearly a homeomorphism. Therefore, up to changing the sign of $\omega'$,
\[\omega'=\iota^*(\omega|_{\Upsilon M^{\leq\ell^2}}).\]
We set
\[h:=\iota_*h'\in H_{i(M)+2n-1}(\Upsilon M^{\leq\ell^2},\Upsilon M^{<\epsilon}),\]
and notice that
\begin{align*}
h\frown\omega|_{\Upsilon M^{\leq\ell^2}}=(\iota_*h')\frown\omega|_{\Upsilon M^{\leq\ell^2}}=\iota_*(h' \frown \iota^*(\omega|_{\Upsilon M^{\leq\ell^2}}))= \iota_*(h'\frown\omega')\neq 0
\end{align*}
in $H_{i(M)}(\Upsilon M^{\leq\ell^2},\Upsilon M^{<\epsilon})$.
\end{proof}
In the following lemma, we will employ the notation of the introduction, and consider the cohomology classes $\alpha$ and $\omega\smile j^*\alpha$ from Equations \eqref{e:omega} and \eqref{e:alpha}.
\begin{lem}
\label{l:Zoll_minimax}
Let $M$ be a closed manifold of dimension $n\geq 2$ admitting a simple Zoll Riemannian metric, and $g$ a Zoll Riemannian metric on $M$ whose unit-speed closed geodesics have minimal period $\ell>0$. For each $\delta\in(0,\rho)$ and integer $k>\overline k(\ell,\rho/\sqrt2)$, consider the space $\Upsilon M=\Upsilon_{\delta,k}M$. Then $c_g(j^*\alpha)=c_g(\omega\smile j^*\alpha)=\ell^2$.
\end{lem}
\begin{proof}
By Lemma~\ref{l:Zoll}, there exists a homology class \[h\in H_{i(M)+2n-1}(\Upsilon M^{\leq \ell^2},\Upsilon M^{< \epsilon})\]
such that both $h$ and $h\frown \omega|_{\Upsilon M^{\leq\ell^2}}$ are mapped to non-zero homology classes under the homomorphism
\begin{align*}
(j_{\ell^2})_*=\mathrm{incl}_*:
H_{*}(\Upsilon M^{\leq \ell^2},\Upsilon M^{<\epsilon})
\to
H_{*}(\Lambda M,\Lambda M^{<\epsilon}).
\end{align*}
Equation~\eqref{e:cohomology_loop_space} implies that
\begin{align*}
H^{i(M)}(\Lambda M,\Lambda M^{<4\rho^2})&\cong H^{i(M)}(\Lambda M,M)\cong\mathds{Z},\\
H^{i(M)-1}(\Lambda M,\Lambda M^{<4\rho^2})&\cong H^{i(M)-1}(\Lambda M,M)=0.
\end{align*}
Therefore, by the universal coefficient theorem,
\begin{align*}
H^{i(M)}(\Lambda M,\Lambda M^{<4\rho^2})\cong\mathrm{Hom}\big(H_{i(M)}(\Lambda M,\Lambda M^{<4\rho^2}),\mathds{Z}\big),
\end{align*}
and the generator $\alpha\in H^{i(M)}(\Lambda M,\Lambda M^{<4\rho^2})$ must satisfy
\[
(\omega \smile j_{\ell^2}^*\alpha)h
=
(j_{\ell^2}^*\alpha)(h\frown\omega)
=
\alpha((j_{\ell^2})_*(h\frown\omega))
\neq0.
\]
This implies that $c_g(j^*\alpha)\leq c_g(\omega\smile j^*\alpha)\leq\ell^2$. On the other hand, $\ell$ is the smallest critical value of the energy $E|_{\Upsilon M}$ above the global minimum $4\rho^2$, and therefore we have the opposite inequality $c_g(j^*\alpha)\geq\ell^2$.
\end{proof}
\subsection{Two subordinated homology classes for arbitrary metrics}
Let $M$ be a closed Riemannian manifold of dimension $n\geq2$ equipped with a Zoll Riemannian metric $g_0$ and with an arbitrary Riemannian metric $g_1$. Their convex combinations
\[
g_s:=(1-s)g_0+sg_1, \qquad s\in[0,1],\]
give a path of Riemannian metrics. We will denote with a subscript or superscript $s$ the usual Riemannian objects associated with the Riemannian metric $g_s$: the exponential map $\exp^{(s)}:\mathrm{T} M\to M$, the Riemannian distance $d_s:M\times M\to[0,\infty)$, the injectivity radius $\rho_s=\mathrm{injrad}(M,g_s)$, and the energy $E_s:\Lambda M\to[0,\infty)$. We set
\begin{align*}
d_{\max}(q_0,q_1) & :=\max_{s\in[0,1]} d_s(q_0,q_1),\quad \forall q_0,q_1\in M,\\
\rho_{\min} & :=\min\{\rho_s\ |\ s\in[0,1]\}>0,\\
c & :=\min\big\{\|v\|_{g_0}\|v\|_{g_s}^{-1}\ \big|\ v\in \mathrm{T} M\setminus0\mbox{-}\mathrm{section},\ s\in[0,1]\big\}\in(0,1],\\
\delta_{\max} & :=\frac{c\,\rho_{\min}}{2}.
\end{align*}
We fix $\delta_0\in(0,\delta_{\max})$ small enough, $\epsilon_0:=8\delta_0^2$, and $\epsilon_1:=4\rho_1^2$ so that we have the inclusion of sublevel sets
\[\{E_0<\epsilon_0\}\subseteq\{E_1<\epsilon_1\}\subset\Lambda M.\]
Since both these sublevel sets can be deformed onto the space of constant loops $M\subset \Lambda M$, the inclusion induces a homology isomorphism
\begin{align*}
H_*(\Lambda M,\{E_0<\epsilon_0\})
\ttoup^{\mathrm{incl}_*}_{\cong}
H_*(\Lambda M,\{E_1<\epsilon_1\}).
\end{align*}
We denote by $\ell_0>0$ the minimal period of the unit-speed geodesics of the Zoll metric $g_0$. By Lemma~\ref{l:Zoll}, for each integer $k_0>\overline k_0(\ell_0,\delta_0)$, if we set
\[\Upsilon^{(0)} M=\Upsilon_{\delta_0,k_0}^{(0)}M,\]
there exists a relative homology class
$h\in H_{i(M)+2n-1}(\Upsilon^{(0)} M,\Upsilon^{(0)} M^{< \epsilon_0})$
such that $h$ and $h\frown \omega_{0}$ are not in the kernel of the homomorphism
\begin{align}
\label{e:inclusion_last}
H_{*}(\Upsilon^{(0)} M,\Upsilon^{(0)} M^{<\epsilon_0})
\ttoup^{\mathrm{incl}_*}
H_{*}(\Lambda M,\{E_0<\epsilon_0\}).
\end{align}
Here, $\omega_{0}\in H^{2n-1}(\Upsilon^{(0)}M)$ is the cohomology class~\eqref{e:omega} for the Riemannian metric $g_0$. Let $\sigma$ be a relative cycle representing $h$, which we can see as a continuous map of the form
\[\sigma:(\Sigma,\partial \Sigma)\to(\Upsilon^{(0)} M,\Upsilon^{(0)} M^{<\epsilon_0})\subset (\Lambda M,\{E_0<\epsilon_0\})\]
for a suitable simplicial complex $\Sigma$ with simplicial boundary $\partial \Sigma$. Hereafter, we will treat the points $\sigma(z)$ as elements of the loop space $\Lambda M$.
\begin{lem}
\label{l:T}
For each $\delta_1>0$ small enough, there exists a continuous function $T:\Sigma\to(0,1)$ such that
\begin{align*}
d_{1}(\sigma(z)(0),\sigma(z)(T(z)))=\delta_1,
\qquad\forall z\in\Sigma.
\end{align*}
\end{lem}
\begin{proof}
We will denote by
$\tau_1:\Upsilon^{(0)} M\to(0,1)$
the function \eqref{e:tau_1} associated to $g_0$, and by $SM$ the unit tangent bundle of $(M,g_0)$. If $\tau''\in(0,\delta_0)$ is sufficiently small, the function
\begin{gather*}
F:SM\times[0,\tau'')\to [0,\infty),
\\
F(q,v,t)
=
d_1(q,\exp_q^{(0)}(tv))^2
=
\|(\exp_q^{(1)})^{-1}\circ\exp_q^{(0)}(tv)\|_{g_1}^2
\end{gather*}
is smooth. For each $(q,v)\in SM$, the function $F(q,v,\cdot)$ has a unique global minimizer at $t=0$, and $F(q,v,0)=0$. Since
\begin{align*}
(\exp_x^{(1)})^{-1}\circ\exp_x^{(0)}(tv)=tv+o(t),
\end{align*}
we readily see that there exists $\tau'\in(0,\tau'')$ such that
\begin{align*}
\tfrac{\mathrm{d}}{\mathrm{d} t} F(q,v,t)>0,
\qquad\forall (q,v)\in SM,\ t\in(0,\tau').
\end{align*}
By the implicit function theorem, for each $\delta_1>0$ small enough there exists a smooth function $T':SM\to(0,\tau')$ such that
\begin{align*}
F(q,v,T'(q,v))=\delta_1,\qquad
\forall (q,v)\in SM.
\end{align*}
Now, for each $z\in\Sigma$, the curve $\gamma_z:=\sigma(z)|_{[0,\tau_1(\sigma(z))]}$ is a geodesic of $(M,g_0)$ with speed
$\|\dot\gamma(0^+)\|_{g_0}=\delta_0/\tau_1(\sigma(z))$. If we set $q_z:=\gamma_z(0)$ and $v_z:=\dot\gamma_z(0^+)/\|\dot\gamma_z(0^+)\|_{g_0}$, we have
\[
F(q_z,v_z,t)=d_1\big(\gamma_z(0),\gamma_z(t\,\tau_1(\sigma(z))/\delta_0)\big)^2.
\]
The desired continuous function is given by
$T(z):=T'(q_z,v_z)\,\tau_1(\sigma(z))/\delta_0$.
\end{proof}
For each $\tau\in(0,1)$, we introduce the subspace
\begin{align*}
U_\tau:=\left\{\gamma\in\Lambda M
\ \left|\
\begin{array}{@{}l@{}}
\gamma(0)\neq\gamma(t)\ \ \forall t\in[0,\tau] \vspace{5pt}\\
\displaystyle
\max_{t\in(0,\tau]}d_{\max}(\gamma(0),\gamma(t))<\rho_{\min}
\end{array}
\right.\right\}.
\end{align*}
We fix $\tau\in(0,1)$ small enough to that the support of our cycle $\sigma(\Sigma)$ is contained in $U_\tau$.
\begin{lem}
\label{l:sigma_s}
For each $\delta_1\in(0,\rho_{\min})$ small enough, integer $k_1\in\mathds{N}$ large enough, and $\tau>0$ small enough, if we set
\[\Upsilon^{(1)}M:=\Upsilon_{\delta_1,k_1}^{(1)}M,\]
there exists a homotopy $\sigma_s:\Sigma\to U_\tau$, with $s\in[0,1]$, such that $\sigma_0=\sigma$, $\sigma_1(\Sigma)\subset U_\tau\cap\Upsilon^{(1)}M$, and $s\mapsto E_1(\sigma_s(z))$ is monotonically decreasing for all $z\in\Sigma$.
\end{lem}
\begin{proof}
We fix a small enough $\delta_1\in(0,\rho_{\min})$ so that Lemma~\ref{l:T} holds with an associated function $T:\Sigma\to(0,1)$. We also fix $\tau\in(0,\min T)$, and a large enough $k_1\in\mathds{N}$ so that, if we set \[\Upsilon^{(1)}M:=\Upsilon^{(1)}_{\delta_1,k_1}M,\] we have
\begin{align*}
\max E_1\circ\sigma < \sup_{\Upsilon^{(1)}M} E_1,
\end{align*}
see~\eqref{e:max_E_delta_k}. We fix $z\in\Sigma$ and $\gamma_0:=\sigma(z)$, and define its deformation $\gamma_s=\sigma_s(z)\in U_\tau$, for $s\in[0,1]$, as follows. We set
\begin{align*}
t_i=t_i(z):=T(z) + \frac{1-T(z)}{k-1}(i-1),
\qquad
i=1,...,k,
\end{align*}
so that $0=:t_0<t_1<...<t_k=1$. The first half of the deformation, for $s\in[0,1/2]$, is the usual Morse shortening process: we set $r_{i,s}:=(1-2s)t_i+2st_{i+1}$ for $i=0,...,k-1$; we define $\gamma_s|_{[t_i,r_{i,s}]}$ the be the shortest $g_1$-geodesic such that $\gamma_s(t_i)=\gamma_0(t_i)$ and $\gamma_s(r_{i,s})=\gamma_0(r_{i,s})$, and we set $\gamma_s|_{[r_{i,s},t_{i+1}]}=\gamma_0|_{[r_{i,s},t_{i+1}]}$.
The curve $\gamma_{1/2}$ is a broken geodesic whose first portion $\gamma_{1/2}|_{[t_0,t_1]}$ has $g_1$-length $\delta_1$. The second half of the deformation, for $s\in[1/2,1]$, is just a time reparametrization of $\gamma_{1/2}$ that will make it belong to $\Upsilon^{(1)}M$. We set
\[
\bm{q}=(q_0,...,q_{k_1-1}):=(\gamma_{1/2}(t_0),\gamma_{1/2}(t_1),...,\gamma_{1/2}(t_{k-1})).
\]
We will denote by
$\tau_i:=\tau_i(\sigma(z))$
the times values \eqref{e:tau_0}, \eqref{e:tau_1}, and \eqref{e:tau_i} associated to $g_1$. For each $s\in[1/2,1]$, $i=0,...,k-1$, and $r\in[0,1]$, we set
\begin{align*}
\gamma_s\big( (2s-1)((1-r)\tau_i + r\tau_{i+1}) + (2-2s)((1-r)t_i + r t_{i+1}) \big)\qquad &\\
:=
\gamma_{1/2}\big((1-r)t_i + r t_{i+1} \big).&
\qedhere
\end{align*}
\end{proof}
For each $s\in[0,1]$ and $t\in(0,\tau]$, we have the evaluation map
\begin{align*}
\mathrm{Ev}_{s,t}:U_\tau\to \mathrm{T} M\setminus 0\mbox{-}\mathrm{section},
\qquad
\mathrm{Ev}_{s,t}(\gamma)=(\exp_{\gamma(0)}^{(s)})^{-1}(\gamma(t)).
\end{align*}
Since $\mathrm{Ev}_{s,t}$ depends continuously on the pair $(s,t)$, the cohomology homomorphism $\mathrm{Ev}_{s,t}^*:H^{2n-1}(\mathrm{T} M\setminus 0\mbox{-}\mathrm{section})\to H^{2n-1}(U_\tau)$ is actually independent of $(s,t)$. We denote a generator of its image by
\[
\Omega_\tau
\in
\mathrm{Ev}_{s,t}^*(H^{2n-1}(\mathrm{T} M\setminus 0\mbox{-}\mathrm{section}))
\subset
H^{2n-1}(U_\tau).
\]
\begin{lem}
\label{l:Omega}
Up to changing the sign of $\omega_0$ and $\omega_1$, we have
\[\Omega_\tau|_{U_\tau\cap\Upsilon^{(s)}M}=\omega_s|_{U_\tau\cap\Upsilon^{(s)}M},
\qquad\forall s\in\{0,1\}.\]
\end{lem}
\begin{proof}
We fix $s\in\{0,1\}$, and denote by
$\tau_1:\Upsilon^{(s)} M\to(0,1)$
the function \eqref{e:tau_1} and by $\mathrm{Ev}:\Upsilon^{(s)}M\to \mathrm{T} M\setminus0\mbox{-}\mathrm{section}$ the evaluation map~\eqref{e:evaluation} associated to $g_s$. Notice that the codomain of $\mathrm{Ev}$ in~\eqref{e:evaluation} is the unit tangent bundle $SM$ of $(M,\delta_s^{-2}g_s)$. However, since the the inclusion $SM\hookrightarrow\mathrm{T} M\setminus0\mbox{-}\mathrm{section}$ is a homotopy equivalence, the cohomology class $\omega_s$ will also be the generator of $\mathrm{Ev}^*(H^*(\mathrm{T} M\setminus0\mbox{-}\mathrm{section}))$. Since
\begin{align*}
\mathrm{Ev}(\gamma_{\bm{q}})=\mathrm{Ev}_{s,\tau_1(\bm{q})}(\gamma_{\bm{q}}),
\qquad\forall\gamma_{\bm{q}}\in U_\tau\cap\Upsilon^{(s)}M,
\end{align*}
we readily see that $\mathrm{Ev}|_{U_\tau\cap\Upsilon^{(s)}M}$ and $\mathrm{Ev}_{s,t}|_{U_\tau\cap\Upsilon^{(s)}M}$, for all $t\in(0,\tau]$, are homotopic maps. Therefore, both the restrictions $\Omega_\tau|_{U_\tau\cap\Upsilon^{(s)}M}$ and $\omega_s|_{U_\tau\cap\Upsilon^{(s)}M}$ are generators of the image $\mathrm{Ev}|_{U_\tau\cap\Upsilon^{(s)}M}^*(\mathrm{T} M\setminus0\mbox{-}\mathrm{section})$, and, up to changing the sign of $\omega_s$, they coincide.
\end{proof}
Since $h$ and $h\frown\omega_0$ are mapped to non-zero classes in $H_*(\Lambda M,\{E_0<\epsilon_0\})$ by the homomorphism~\eqref{e:inclusion_last}, Lemmas~\ref{l:sigma_s} and~\ref{l:Omega} imply that, for each $s\in\{0,1\}$, $[\sigma_s]$ and $[\sigma_s]\cap\omega_s$ are non-trivial relative homology classes in
\[
H_*(U_\tau\cap\Upsilon^{(s)}M,U_\tau\cap\Upsilon^{(s)}M^{<\epsilon_s}).
\]
If we consider the homomorphism induced by the inclusion
\begin{align*}
\iota^{(s)}_*:H_*(U_\tau\cap\Upsilon^{(s)}M,U_\tau\cap\Upsilon^{(s)}M^{<\epsilon_s})
\ttoup^{\mathrm{incl}_*}
H_*(U_\tau,U_\tau\cap\{E_1<\epsilon_1\}),
\end{align*}
we have
\begin{align*}
\iota^{(0)}_*[\sigma_0]=\iota^{(1)}_*[\sigma_1]=[\sigma],
\qquad
\iota^{(0)}_*([\sigma_0]\frown\omega_0)=\iota^{(1)}_*([\sigma_1]\frown\omega_1)=[\sigma]\frown\Omega_\tau.
\end{align*}
In particular, if we see $\sigma_s$ and $\sigma_s\frown\omega_s$ as relative cycles in $(\Lambda M,\{E_1<\epsilon_1\})$, we have $[\sigma_0]=[\sigma_1]$ and $[\sigma_0\frown\omega_0]=[\sigma_1\frown\omega_1]$ in $H_*(\Lambda M,\{E_1<\epsilon_1\})$.
Now, consider the generator
$\alpha\in H^{i(M)}(\Lambda M,\{E_1<\epsilon_1\})\cong\mathds{Z}$, and the inclusion
\[j_1:(\Upsilon^{(1)} M,\Upsilon^{(1)} M^{<\epsilon_1})\hookrightarrow(\Lambda M,\{E_1<\epsilon_1\}).\]
\begin{lem}
\label{l:cohomology}
For each $\delta_1\in(0,\rho_1)$ small enough and $k_1\in\mathds{N}$ large enough, if we set
\[\Upsilon^{(1)}M:=\Upsilon_{\delta_1,k_1}^{(1)}M,\]
we have $\omega_1\smile j_1^*\alpha\neq0$ in $H^{i(M)}(\Upsilon^{(1)} M,\Upsilon^{(1)} M^{<\epsilon_1})$.
\end{lem}
\begin{proof}
Let $\delta_1\in(0,\rho_1)$ be small enough and $k_1\in\mathds{N}$ large enough so that Lemma~\ref{l:sigma_s} holds. Since $H^{i(M)-1}(\Lambda M,\{E_1<\epsilon_1\})$ is trivial, the universal coefficient theorem implies that
\begin{align*}
H^{i(M)}(\Lambda M,\{E_1<\epsilon_1\})\cong\mathrm{Hom}\big( H_{i(M)}(\Lambda M,\{E_1<\epsilon_1\}),\mathds{Z} \big).
\end{align*}
This, together with the facts that $\alpha$ is the generator of $H^{i(M)}(\Lambda M,\{E_1<\epsilon_1\})$ and that $(j_0)_*([\sigma_0]\frown\omega_0)$ is non-zero in $H_{i(M)}(\Lambda M,\{E_1<\epsilon_1\})$, implies that
\begin{align*}
\alpha((j_0)_*([\sigma_0]\frown\omega_0))
\neq0.
\end{align*}
Therefore, we conclude
\begin{align*}
(\omega_1\smile j_1^*\alpha)[\sigma_1]
&=
(j_1^*\alpha)([\sigma_1]\frown\omega_1)
=
\alpha((j_1)_*([\sigma_1]\frown\omega_1))\\
&=
\alpha((j_0)_*([\sigma_0]\frown\omega_0))
\neq0.
\qedhere
\end{align*}
\end{proof}
\begin{proof}[Proof of Theorem~\ref{t:main}]
Let $M$ be a closed manifold of dimension $n\geq2$ admitting a simple Zoll Riemannian metric, and $g$ a Riemannian metric on $M$.
Lemma~\ref{l:cohomology} implies that $\omega\smile j^*\alpha\neq0$ in the relative homology group $H^{i(M)}(\Upsilon M,\Upsilon M^{<4\rho^2})$.
If $g$ is a Zoll Riemannian metric whose unit-speed geodesics have minimal period $\ell$, then Lemma~\ref{l:Zoll_minimax} implies that $c_g(j^*\alpha)=c_g(\omega\smile j^*\alpha)=\ell^2$. Conversely, assume that $c_g(j^*\alpha)=c_g(\omega\smile j^*\alpha)=:\ell^2$.
We can apply Lemma~\ref{l:main_lemma} with $d=i(M)$ and $\mu=j^*\alpha$, and infer that $g$ is a Besse Riemannian metric, and either $\ell$ or $\ell-2\delta$ is a common multiple of the periods of the unit-speed geodesics of $(M,g)$. Moreover, the critical set
\[K:=\mathrm{crit}(E)\cap\big(E^{-1}(\ell^2)\cup E^{-1}((\ell-\delta)^2)\big)\cong SM,\]
has Morse index $\mathrm{ind}(E,K)\leq i(M)$. Since $i(M)$ is the minimal Morse index of a closed geodesic, we have
\begin{align}
\label{e:ind_K}
\mathrm{ind}(E,K) = i(M).
\end{align}
Now, let us further require $M$ to be simply connected and spin, and assume by contradiction that $g$ is not a Zoll Riemannian metric. We are now going to employ two results due to Radeschi and Wilking. Since $(M,g)$ is a simply connected Besse manifold, by \cite[Theorem~D]{Radeschi:2017dz} the energy functional $E:\Lambda M\to[0,\infty)$ is perfect for the $S^1$-equivariant singular cohomology with rational coefficients $H^*_{S^1}(\cdot\,;\mathds{Q})$. Moreover, since $(M,g)$ is an orientable and spin Besse manifold, by \cite[Corollary~C]{Radeschi:2017dz} the negative bundles of all the critical manifolds of $E$ are orientable. This, in turn, implies that all critical manifolds of $E$ are homologically visible, and, if we set
\[
K_{i(M)}
:=
\big\{ \gamma\in\mathrm{crit}(E)\ \big|\ \mathrm{ind}(E)=i(M) \big\},
\]
we have
\begin{align*}
H^{i(M)}_{S^1}(\Lambda M,M;\mathds{Q})\cong H^{0}_{S^1}(K_{i(M)};\mathds{Q}).
\end{align*}
Namely, the rank of $H^{i(M)}_{S^1}(\Lambda M,M;\mathds{Q})$ is the number of path-connected components of $K_{i(M)}$.
Clearly, $K$ is a path-connected component of $K_{i(M)}$. Since $g$ is Besse but not Zoll, Equation~\eqref{e:ind_K} implies that $K_{i(M)}\setminus K$ is not empty, and therefore
\begin{align}
\label{e:rank_geq_2}
\rank\big(H^{i(M)}_{S^1}(\Lambda M,M;\mathds{Q})\big)\geq2.
\end{align}
On the other hand, if we repeat the whole argument with a Zoll Riemannian metric $g_0$ instead of $g$, the critical set $K_{i(M)}$ becomes diffeomorphic to the unit tangent bundle $SM$, which is path-connected. This implies that
\[\rank\big(H^{i(M)}_{S^1}(\Lambda M,M;\mathds{Q})\big)=1,\] and contradicts~\eqref{e:rank_geq_2}.
\end{proof}
|
\section*{References%
\title{Exponential Convergence Time of Gradient Descent\\ for One-Dimensional
Deep Linear Neural Networks}
\date{}
\author{Ohad Shamir\\Weizmann Institute of Science}
\begin{document}
\maketitle
\begin{abstract}
We study the dynamics of gradient descent on objective
functions of the form $f(\prod_{i=1}^{k} w_i)$ (with respect to scalar
parameters $w_1,\ldots,w_k$), which arise in the context of
training depth-$k$ linear neural networks. We prove that for standard
random initializations, and under mild assumptions on $f$, the number of
iterations required for convergence scales exponentially with the depth
$k$. We also show empirically that this phenomenon can occur in higher
dimensions, where each $w_i$ is a matrix. This highlights a potential
obstacle in understanding the convergence of gradient-based methods for
deep linear neural networks, where $k$ is large.
\end{abstract}
\section{Introduction}
One of the biggest open problems in theoretical machine learning is to explain
why deep artificial neural networks can be efficiently trained in practice,
using
simple gradient-based methods. Such training requires optimizing complex,
highly non-convex objective functions, which seem intractable from a worst-case
viewpoint. Over the past few years, much research has been devoted to this
question, but it remains largely unanswered.
Trying to understand simpler versions of this question, significant attention
has been devoted to \emph{linear} neural networks, which are predictors
mathematically defined as $\mathbf{x}\mapsto \prod_{i=1}^{k}W_i\mathbf{x}$,
with $W_1,\ldots,W_k$ being a set of parameter matrices, and $k$ being the
depth parameter
(e.g.
\citet{saxe2013exact,kawaguchi2016deep,hardt2016identity,lu2017depth,bartlett2018gradient,
laurent2018deep}).
The optimization problem associated with training such networks can be
formulated as
\begin{equation}\label{eq:linopt}
\min_{W_1,\ldots,W_k} F(W_1,\ldots,W_k) :=
f\left(\prod_{i=1}^{k}W_i\right)
\end{equation}
for some matrix-valued function $f$. Although much simpler than general
feedforward neural networks (which involve additional non-linear functions), it
is widely believed that \eqref{eq:linopt} captures important aspects of
neural network optimization problems. Moreover,
\eqref{eq:linopt} has a simple algebraic structure, which makes it more
amenable
to analysis. In particular, it is known that when $f$ is convex and
differentiable, \eqref{eq:linopt} has no local minima except global ones (see
\citet{laurent2018deep} and references therein). In other words, if an
optimization algorithm converges to some local minimum, then it must converge
to a global minimum.
Importantly, this no-local-minima result \emph{does not} imply that
gradient-based methods indeed solve \eqref{eq:linopt} efficiently: Even when
they converge to local minima (which is not always guaranteed, say in case the
parameters diverge), the number of required iterations might be arbitrarily
large. To study this question, \citet{bartlett2018gradient} recently considered
the special case where
$F(W_1,\ldots,W_k)~:=~\frac{1}{2}\|\prod_{i=1}^{k}W_i-Y\|_{F}^2$
(where $\norm{\cdot}_{F}$ is the Frobenius norm) for square matrices
$W_1,\ldots,W_k,Y$, using gradient descent starting specifically from $W_i=I$
for all $i$. In this setting, the authors prove a polynomial-time convergence
guarantee when $Y$ is positive semidefinite. On the other hand, when $Y$ is
symmetric and with negative eigenvalues, it is shown that gradient descent with
this initialization will never converge. Although these results provide
important insights, they crucially assume
that each $W_i$ is initialized \emph{exactly} at the identity $I$. Since in
practice parameters are initialized randomly, it is natural to ask whether
such results hold with random initialization. Indeed, even though
gradient descent might fail to converge with a specific initialization, it
could be that even a tiny random perturbation is sufficient for polynomial-time
convergence\footnote{For example, consider the
objective $F(w_1,w_2)=(w_1 w_2+1)^2$ where $w_1,w_2\in \mathbb{R}$. It is an easy
exercise to show that gradient descent starting from any $w_1=w_2>0$
(and sufficiently small step sizes) will converge to the suboptimal saddle
point $(0,0)$.
On the other hand, polynomial-time convergence holds with random initialization
(see \citet{du2018algorithmic}).}. More recently, \citet{arora2018convergence}
considered gradient descent on a similar objective, and managed to prove strong
polynomial-time
convergence guarantees under certain assumptions about the initialization.
However, as the authors discuss (in section 3.2.1), these assumptions are not
generally satisfied for standard initialization approaches. In another recent
related
work, \citet{ji2018gradient} show that for certain classification problems on
linearly separable data (corresponding to a suitable choice of $f$ in
\eqref{eq:linopt}), gradient descent asymptotically
converges to a globally optimal objective value. However, the result only
applies to particular choices of $f$, and more importantly, is asymptotic and
hence does not imply a finite-time convergence guarantee. Thus, analyzing the
finite-time convergence of gradient descent on \eqref{eq:linopt}, with
standard random initializations, remains a challenging open problem.
In this paper, we consider a simpler special case of \eqref{eq:linopt}, where
the matrices $W_1,\ldots,W_k$ are all scalars:
\begin{equation}\label{eq:linopt1}
\min_{\mathbf{w}\in\mathbb{R}^k} F(\mathbf{w}) := f\left(\prod_{i=1}^{k} w_i\right)~.
\end{equation}
Our main and perhaps surprising result is that even in
this relatively simple setting, gradient descent with
random initialization can
require $\exp(\Omega(k))$ iterations to converge. This holds under mild
conditions on the function $f$, and with standard initializations (including
Xavier initialization and any reasonable initialization close to
$(1,\ldots,1)$). We complement this by showing that
$\exp(\tilde{\mathcal{O}}(k))\cdot \max\{1,\log(1/\epsilon)\}$ iterations are also
sufficient for convergence to an $\epsilon$-optimal point. Moreover, in
\secref{sec:experiments} we present experiments which
strongly suggest that this phenomenon is not unique to one-dimensional
networks, and at least in some cases, the same exponential dependence can also
occur in multi-dimensional networks (i.e., \eqref{eq:linopt} where each $W_i$
is a $d\times d$ matrix, $d>1$). The take-home message is that even if we focus
on linear
neural networks, natural objective functions without any spurious local minima,
and random initializations, the associated optimization problems can sometime
be intractable for gradient descent to solve, when the depth is large.
Before continuing, we emphasize that our results do not imply that
gradient-based methods cannot learn deep linear networks in general. What they
do imply is that one would need to make additional assumptions or algorithmic
modifications to circumvent these negative results: For example, explicitly
using the fact that the matrix sizes are larger than $1$ -- something which is
not clear how to do with current analyses -- or having a fine-grained
dependency on the variance of the random initialization, as further discussed
in \secref{sec:experiments}. Alternatively, our results might be circumvented
using other gradient-based algorithms (for
example, by adding random noise to the gradient updates or using adaptive
step sizes), or other initialization strategies. However, that would not
explain why plain gradient descent with standard random initializations is
often practically effective on these problems. Overall, we believe our results
point to a potential obstacle in understanding the convergence of
gradient-based methods for linear networks: At the very least, one would have
to rule out one-dimensional layers, or consider algorithms other than plain
gradient descent with standard initializations, in order
to establish polynomial-time convergence guarantees for deep linear networks.
Finally, we note that our results provide a possibly interesting contrast to
the
recent work of \citet{arora2018optimization}, which suggests
that increasing depth can sometimes accelerate the optimization process. Here
we show that at least in some cases, the opposite occurs: Adding depth can
quickly turn a trivial optimization problem into an intractable one for
gradient descent.
\section{Preliminaries}
\textbf{Notation.} We use bold-faced letters to denote vectors.
Given a vector $\mathbf{w}$, $w_j$ refers to its $j$-th coordinate.
$\norm{\cdot}$,
$\norm{\cdot}_1$ and
$\norm{\cdot}_{\infty}$ refer to the Euclidean norm, the $1$-norm and the
infinity norm respectively. We let
$\prod_{i=1}^{k}w_i$ and $\prod_i w_i$ be a
shorthand for $w_1\cdot w_2\cdots w_k$. Also, we define a product
over an empty set as being
equal to $1$. Since our main focus is to
study the dependence on the
network
depth $k$, we use the standard notation
$\mathcal{O}(\cdot),\Omega(\cdot),\Theta(\cdot)$ to hide constants independent of
$k$, and $\tilde{\mathcal{O}}(\cdot),\tilde{\Omega}(\cdot),\tilde{\Theta}(\cdot)$ to
hide constants and factors logarithmic in $k$.
\textbf{Gradient Descent.} We consider the standard gradient descent method for
unconstrained optimization
of functions $F$ in Euclidean space, which given an initialization point
$\mathbf{w}(1)$, performs repeated iterations of the form $\mathbf{w}(t+1):=\mathbf{w}(t)-\eta
\nabla F(\mathbf{w}(t))$ for $t=1,2,\ldots$ (where $\nabla F(\cdot)$ is the gradient,
and $\eta>0$ is a step size parameter). For objectives as in
\eqref{eq:linopt1},
we
have $\frac{\partial}{\partial w_j} F(\mathbf{w})=
f'(\prod_i w_i)\prod_{j\neq i}w_i$, and gradient descent takes the form
\[
\forall j,~ w_j(t+1) = w_j(t)-\eta f'\left(\prod_i w_i(t)\right)\prod_{j\neq
i}w_i(t)~.
\]
\textbf{Random Initialization.} One of the most common initialization methods
for neural networks is \emph{Xavier} initialization
\citep{glorot2010understanding}, which in the setting of \eqref{eq:linopt}
corresponds to choosing each entry of each $d\times d$ matrix $W_i$
independently from a zero-mean distribution with variance $1/d$ (usually
uniform or Gaussian). This ensures
that the variance of the network outputs (with respect
to the initialization) is constant irrespective of the network size. Motivated
by residual networks,
\citet{hardt2016identity} and \citet{bartlett2018gradient} consider
initializing each $W_i$ independently at $I$, possibly with some random
perturbation. In this paper we denote such an initialization scheme as a
\emph{near-identity} initialization. Since we focus here on the case $d=1$ as
in \eqref{eq:linopt1},
Xavier initialization corresponds to choosing each $w_i$ independently from
a zero-mean, unit-variance distribution, and
near-identity initialization corresponds to choosing each $w_i$ close to $1$.
\section{Exponential Convergence Time for Gradient Descent}\label{sec:main}
For our negative results, we impose the following mild conditions on the
function $f$ in \eqref{eq:linopt1}:
\begin{assumption}\label{assump:f}
$f:\mathbb{R}\rightarrow\mathbb{R}$ is differentiable, Lipschitz continuous
and strictly monotonically
increasing on any interval $[-\frac{1}{2},z)$ where $z>0$. Moreover,
$\inf_{p\in [-\frac{1}{2},\infty)}f(p)-\inf_{p\in \mathbb{R}}f(p) > 0$.
\end{assumption}
Here, we assume that $f$ is fixed, and our goal is to study the convergence
time of gradient descent on \eqref{eq:linopt1} as a function of the depth $k$.
Some simple examples satisfying Assumption \ref{assump:f} in the context of
machine learning include $f(x) = (x+1)^2$ and $f(x) = \log(1+\exp(x))$ (e.g.,
squared
loss and logistic loss with respect to the input/output pair $(1,-1)$,
respectively). We note that this non-symmetry with respect to positive/negative
values is completely arbitrary, and one can prove similar results if their
roles are reversed.
\subsection{Xavier Initialization}
We begin with the case of Xavier initialization, where we initialize all
coordinates of $\mathbf{w}$ in \eqref{eq:linopt1} independently from a zero-mean, unit
variance distribution. We will consider any distribution which satisfies the
following:
\begin{assumption}\label{assump:initzero}
$w_1(1),\ldots,w_k(1)$ are drawn i.i.d. from a zero-mean, unit variance
distribution such that
\begin{enumerate}
\item $\Pr(w_1(1)\in [-a,a])\leq c_1 a$ for all $a\geq 0$
\item $\mathbb{E}[|w_1(1)|]\leq 1-c_2$~
\end{enumerate}
where $c_1,c_2>0$ are absolute constants independent of $k$.
\end{assumption}
The first part of the assumption is satisfied for any
distribution with bounded density. As to the second part, the
following lemma shows that it is satisfied for uniform and Gaussian
distributions (with an explicit $c_2$), and in fact for any non-trivial
distribution (with a distribution-dependent $c_2$ -- see also footnote
\ref{footnote:orthogonal}):
\begin{lemma}
The following hold:
\begin{itemize}
\item If $w$ is drawn from a zero-mean, unit-variance Gaussian, then
$\mathbb{E}[|w|]< 0.8$~.
\item If $w$ is drawn from a zero-mean, unit-variance uniform
distribution, then $\mathbb{E}[|w|]< 0.9$~.
\item If $w$ is drawn from any zero-mean, unit variance distribution
other than uniform on $\{-1,+1\}$, then $\mathbb{E}[|w|]< 1$.
\end{itemize}
\end{lemma}
\begin{proof}
The first two parts follow from standard results on Gaussian and uniform
distributions. As to the third part, it is easy to see that the support of
any distribution which satisfies the conditions cannot be a subset of
$\{-1,+1\}$, and therefore $w^2$ is not supported on a single value. By
Jensen's inequality and the fact that $\sqrt{\cdot}$ is a strictly concave
function, it follows that
$\mathbb{E}[|w|]=\mathbb{E}[\sqrt{w^2}]<\sqrt{\mathbb{E}[w^2]} = 1$.
\end{proof}
With such an initialization, we now show that gradient descent is
overwhelmingly likely to take at least exponential time to converge:
\begin{theorem}\label{thm:zero}
The following holds for some positive constants $c,c'$ independent of $k$:
Under
Assumptions
\ref{assump:f} and \ref{assump:initzero}, if gradient
descent is ran with any step size $\eta\leq \exp(ck)$, then with
probability at least
$1-\exp(-\Omega(k))$ over the initialization, the number of iterations
required to reach suboptimality less than $c'$ is at least
$\exp(\Omega(k))$.
\end{theorem}
In the above, $\Omega(\cdot)$ hides dependencies on the absolute constants in
the theorem statement and the assumptions. The proof (as well as
all other major proofs in this paper) is presented in \secref{sec:proofs}.
The intuition behind the theorem is quite simple: Under our assumptions, it is
easy to show that the product of any $\Omega(k)$ coordinates from
$w_1(1),\ldots,w_k(1)$ is overwhelmingly likely to be exponentially small in
$k$. Since the derivative
of our objective w.r.t. any $w_j$ has the form
$f'(\prod_i w_i)\prod_{i\neq j} w_i$,
it follows that the gradient is exponentially small in $k$. Moreover, we show
that the gradient is exponentially small at any point within a bounded distance
from the initialization (which is the main technical challenge of the proof,
since the gradient is by no means Lipschitz). As a
result, gradient descent will only make exponentially small steps. Assuming we
start from a point bounded away from a global minimum, it follows that the
number of required iterations must be exponentially large in $k$.
We note that the observation that Xavier initialization leads to highly skewed
values in deep enough networks is not new (see
\citet{saxe2013exact,pennington2017resurrecting}),
and has motivated alternative initializations such as orthogonal
initialization\footnote{It is interesting to note that in our setting,
orthogonal initialization amounts to choosing each $w_i$ in $\{-1,+1\}$, which
can easily cause non-convergence, e.g. for $F(w_1,\ldots,w_k) =
(\prod_i w_i-y)^2$ when $y\prod_i w_i(1)<0$ and small enough step
sizes.\label{footnote:orthogonal}} Our
contribution here is to rigorously analyze how this affects the optimization
process for our setting.
\subsection{Near-Identity Initialization}
We now turn to consider initializations where each $w_i$ is initialized close
to $1$. Here, it will be convenient to make deterministic rather than
stochastic assumptions on the initialization point (which are satisfied with
high probability for reasonable distributions):
\begin{assumption}\label{assump:initone}
For some absolute constants $c_1,c_2,c_3>0$ independent of
$k$, gradient descent is initialized at a point $\mathbf{w}(1)$ which satisfies
$\max_{j} |w_j(1)-1|~\leq~ k^{-c_1}$ and $c_2\leq\prod_i w_i(1)\leq c_3$.
\end{assumption}
To justify this assumption, note that if $w_1(1),\ldots, w_k(1)$ are chosen
i.i.d. and not in the range of $1\pm k^{-c_1}$ for some $c_1>0$, then their
product is likely to explode or vanish with $k$.
\begin{theorem}\label{thm:identity}
The following holds for some positive constants $c,c'$ independent of $k$:
Under
Assumptions \ref{assump:f} and \ref{assump:initone}, if gradient
descent is ran with any positive step size $\eta\leq c$, then the
number of
iterations required to reach suboptimality less than $c'$ is at least
$\exp(\Omega(k))$.
\end{theorem}
As before, $\Omega(\cdot)$ hides dependencies on the absolute constants in
the theorem statement, as well as those in the assumptions.
The formal proof appears in \secref{sec:proofs}. To help explain its intuition,
we provide in Figure
\ref{fig:evolution} the actual evolution of $w_j(t)$ for a typical run of
gradient descent, when $F(\mathbf{w})=F(w_1,\ldots,w_7) =
\frac{1}{2}(\prod_{i=1}^{7} w_i+1)^2$ and we initialize all coordinates
reasonably close to $1$.
Recall that for any $w_j(t)$, the gradient descent updates take the form
\[
\forall j,~ w_j(t+1)~=~ w_j(t)-\eta\left(\prod_i w_i(t)+1\right)\prod_{i\neq
j} w_i(t)~,
\]
where $\prod_i w_i(1)>0$.
Thus, initially, all parameters $w_j(t)$ decrease with $t$, as to be expected.
However, as their value fall to around or below $1$, their product decreases
rapidly to
$\exp(-\Omega(k))$. Since the gradient of each $w_j(t)$ scales as
$\prod_{i\neq j} w_i(t)$,
the magnitude of the gradients becomes very small, and the algorithm makes only
slow progress. Eventually, one of the parameters becomes negative, in which
case all other parameters start increasing, and the algorithm
converges. However, by a careful analysis, the length of the slow middle phase
can be shown to be exponential in the depth / number of parameters $k$.
\begin{figure}
\centering
\includegraphics[trim=2cm 0cm 0cm 0cm, clip=true, scale=1]{evolution2.pdf}
\caption{The left figure plots $F(\mathbf{w}(t))$ as a function of iteration $t$, for
$F(\mathbf{w})=(\prod_{i=1}^{7} w_i+1)^2$. The right figure plots
$w_1(t),w_2(t),\ldots,w_7(t)$ as a function of $t$. Best viewed in color.}
\label{fig:evolution}
\end{figure}
\subsection{A Positive Result}
Having established that the number of iterations is at least $\exp(\Omega(k))$,
we now show that this is nearly tight. Specifically, we prove that gradient
descent indeed converges in the settings studied so far, with a number of
iterations scaling as $\exp(\tilde{\mathcal{O}}(k))$ (this
can be interpreted as a constant for any constant $k$). For simplicity, we
prove this in the case where $f(\prod_i w_i) = \frac{1}{2}(\prod_i w_i-y)^2$,
but the technique can be easily generalized to other convex $f$ under mild
conditions. We note that the case of $y>0$ and each $w_i$ initialized to $1$ is
covered by the results in \citet{bartlett2018gradient}. However, here we show a
convergence result for other values of $y$, and even if $w_i$ are not all
initialized at $1$.
We will use the following assumptions on our objective and parameters:
\begin{assumption}\label{assump:pos}
The following hold for some absolute positive constants $c_1,c_2,c_3,c_4$
independent of $k$:
\begin{itemize}
\item $y=-c_1<0$
\item The initialization $w_1(1),\ldots,w_k(1)$ satisfies the following:
\begin{itemize}
\item $|w_i(1)|\leq c_2$ and $\prod_i w_i(1)>y$
\item $\min_{j\neq j'} \left||w_j(1)|-|w_{j'}(1)|\right|\geq k^{-c_4}$
\item $\max_{j,j'} \left|\prod_{i\notin \{j,j'\}}w_i(1)\right|\leq c_4$
\end{itemize}
\end{itemize}
\end{assumption}
The assumptions $y<0$ and $\prod_i w_i(1)>y$ ensure that the objective
satisfies the conditions of our negative results, for both
Xavier and near-identity initializations (the other cases can be studied using
similar techniques).
\begin{theorem}\label{thm:posid}
Consider the objective $F(\mathbf{w}) = \frac{1}{2}\left(\prod_i w_i-y\right)^2$.
Under Assumption \ref{assump:pos}, for any step size $\eta=k^{-c}$ for some
large enough constant $c>0$, and for any $\epsilon>0$, the number of
gradient descent iterations $t$ required for $F(\mathbf{w}_t)\leq \epsilon$ is at
most $\exp\left(\tilde{\mathcal{O}}(k)\right)\cdot\max\{1,\log(1/\epsilon)\}$.
\end{theorem}
\section{Multi-Dimensional Networks}\label{sec:experiments}
So far, we showed that for one-dimensional linear neural networks,
gradient descent can easily require exponentially many iterations (in the depth
of the network) to converge. However, these results are specific to the case
where the parameter matrix $W_i$ of each layer is one-dimensional, and do not
necessarily extend to higher dimensions. A possibly interesting exception is
when $F(W_1,\ldots,W_k)=\norm{\prod_i W_i - Y}_F^2$, and both $Y$ and the
initialization $W_1(1),\ldots,W_k(1)$ are diagonal matrices. In that case, it
is easy to show that the matrices produced by gradient descent remain diagonal,
and the objective can be rewritten as a sum of independent one-dimensional
problems for which our results would apply. However, this reasoning fails for
non-diagonal initializations and target matrices $Y$.
In this section, we study experimentally whether our theoretical results for
one-dimensional networks might also extend to multi-dimensional ones. In
particular, we consider the multi-dimensional generalization of the
objective function studied earlier:
\[
F(W_1,\ldots,W_k) = \frac{1}{2}\left\|\prod_{i=1}^{k}W_i-Y\right\|_{F}^2,
\]
where $W_1,\ldots,W_k$ are $d\times d$ square matrices (for $d=25$), $Y=-I$
($I$ being
the identity matrix), and $\|\cdot\|_{F}$ is the Frobenius norm. We
ran gradient descent on this objective using three initialization strategies:
\begin{enumerate}
\item \emph{Xavier initialization}: Each entry of each matrix $W_i$ was
initialized independently from a zero-mean Gaussian with variance
$\frac{1}{d}$.
\item \emph{Near-Identity initialization}: Each $W_i$ was initialized as $I+M$
where each entry of $M$ was sampled independently from a zero-mean Gaussian
with variance $\frac{1}{dk}$. Up to numerical constants, this is the
largest variance which ensures that
$\mathbb{E}[(\prod_{i=1}^{k}W_i)(\prod_{i=1}^{k}W_i)^\top]$
remains bounded independent of $d,k$. To see this, note that had we used
variance $\frac{c}{dk}$ for some constant $c$, then
$\mathbb{E}[W_iW_i^\top] = \left(1+\frac{c}{k}\right)I$ and thus
$\mathbb{E}[(\prod_{i=1}^{k}W_i)(\prod_{i=1}^{k}W_i)^\top]
= \left(1+\frac{c}{k}\right)^k I\approx \exp(c)I$.
\item \emph{Near-Identity initialization with smaller variance}: Each $W_i$ was
initialized as above, except that the variance of each entry in the matrix $M$
was $\frac{1}{(dk)^2}$.
\end{enumerate}
\begin{figure}
\includegraphics[scale=0.7,trim=3.4cm 0cm 2cm
0cm,clip=true]{graphs.pdf}\label{fig:graphs}
\caption{Mean and Standard Deviation of the log number of iterations required
for convergence, over $50$ trials, for each initialization strategy and depth
parameter $k\in \{2,3,\ldots,8\}$. `NC' refers to the percentage of runs (for a
given initialization and depth) which did not converge after $10^9$ iterations,
if any. Note that when some trials did not converge, the bars actually
under-estimate the mean convergence time if all trials were ran till
convergence (since they only represent runs which took a log number of
iterations less than $\log(10^9)=20.72...$). }
\end{figure}
For each random initialization strategy, and for depth parameter $k\in
\{2,3,\ldots,8\}$ , we ran 50 trials of gradient
descent, with a step size\footnote{Our results did not seem to change
significantly by taking other bounded step sizes.} of $10^{-3}$, until either
one
of the following two
stopping conditions occured:
\begin{itemize}
\item The objective value dropped
below $0.1$ (or equivalently, $\norm{\prod_{i=1}^{k}W_i-Y}_F\leq
\sqrt{1/5}$, a
rather mild requirement).
\item The number of iterations exceeded $10^9$ iterations, in which case the
algorithm was deemed to have failed to converge (note that from a practical
viewpoint, one billion
iterations is exceedingly large considering our problem size).
\end{itemize}
In Figure \ref{fig:graphs}, we plot the mean and standard deviation for the
\emph{logarithm} of the number of iterations required to make the objective
value less than $0.1$ (among the $50$ trials which converged). We also point
out the percentage of trials which did not converge, if any.
The figure strongly suggests that using both Xavier initialization and
near-identity initialization with small variance, the required runtime scales
exponentially with the depth (recall that the $y$-axis is in log scale). This
indicates that the phenomenon of exponential scaling with depth is not just an
artifact of one-dimensional networks, and can also occur in multi-dimensional
networks, even with reasonable random initializations. On the flip side, when
performing near-identity initialization with a large enough variance, we did
not observe such an exponential scaling (as evidenced in the middle plot in the
figure). Moreover, based on some additional experiments with other objective
functions, it appears that although gradient descent can sometime require
exponential time to converge, this phenomenon is not particularly common. A
possible explanation to this is that in one dimension, $\prod_i w_i$ had to
change sign, and hence pass through zero (see Figure \ref{fig:evolution}). This
brought the iterates to a ``flat'' region with exponentially small gradients.
In contrast, in multiple
dimensions, to continuously change $\prod_i W_i$ from a matrix to some other
matrix, it is always possible to go ``around'' any particular point. Our
experiments suggest that gradient descent indeed avoids problematic flat
regions in many cases, but not always. Overall, it seems quite possible that
for multi-dimensional networks, the exponential runtime dependence on the
depth can be avoided under reasonable assumptions -- however, some such
assumptions would be necessary, and would need to exclude either objectives of
the type we studied here, or some of the initializations. For example, such an
analysis might need to explicitly separate between one-dimensional and
multi-dimensional networks, or between near-identity initialization with
variance $1/dk$ and with variance $1/(dk)^2$ (which are both polynomially large
in $d,k$), and how to do so with existing analyses is currently unclear.
\section{Proofs}\label{sec:proofs}
\subsection{Proof of \thmref{thm:zero}}
The proof is based on the following two lemmas:
\begin{lemma}\label{lem:smallinit}
Suppose $w_1,\ldots,w_k$ are drawn i.i.d. from a distribution such that
$\mathbb{E}[|w_1|]\leq a$ for some $a>0$. Then
\[
\Pr\left(\max_j \left|\prod_{i\neq j}w_i\right|\geq
k a^{(k-1)/2}\right)~\leq~
a^{(k-1)/2}
\]
\end{lemma}
\begin{proof}
For any fixed $j$, by Markov's inequality and the i.i.d. assumption,
\[
\Pr\left(\left|\prod_{i\neq j}w_i\right|\geq ka^{(k-1)/2}\right)
~\leq~
\frac{\mathbb{E}\left[|\prod_{i\neq j}w_i|\right]}{ka^{(k-1)/2}}
~=~ \frac{\left(\mathbb{E}[|w_1|]\right)^{k-1}}{ka^{(k-1)/2}}
~\leq~ \frac{a^{k-1}}{ka^{(k-1)/2}} ~=~ \frac{1}{k}a^{(k-1)/2}~.
\]
Taking a union bound over all $j=1,2,\ldots,k$, the result follows.
\end{proof}
\begin{lemma}\label{lem:flatball}
Let $\alpha,\beta,\delta>0$ be fixed. Let $\mathbf{w}\in \mathbb{R}^k$ such that
$\max_j \left|\prod_{i\neq j}
w_i\right|\leq
\alpha$ and $\min_i |w_i|\geq \delta$. Then for any $\mathbf{v}$ such that
$\norm{\mathbf{v}-\mathbf{w}}\leq
\frac{\delta}{\sqrt{k-1}}
\log(\beta/\alpha)$, it holds that $|\prod_i v_i|\leq
\beta\norm{\mathbf{v}}_{\infty}$ as well as
$\norm{\nabla F(\mathbf{v})}~\leq~
\sup_{p:|p|\leq \beta\norm{\mathbf{v}}_{\infty}}|f'(p)|\cdot \sqrt{k}\beta$.
\end{lemma}
\begin{proof}
We claim that it is enough to prove the following:
\begin{align}
\forall \mathbf{w},\mathbf{v}\in\mathbb{R}^k~~&\text{s.t.}~~\max_j
\left|\prod_{i\neq j}
w_i\right|\leq
\alpha~~,~~\min_i |w_i|\geq \delta~~,~~
\max_j\left|\prod_{i\neq j}v_i\right|> \beta\notag\\
&\text{it holds that} ~~
\norm{\mathbf{v}-\mathbf{w}}~>~
\frac{\delta}{\sqrt{k-1}}\log(\beta/\alpha)~.\label{eq:toshow0}
\end{align}
Indeed, this would imply that for any $\mathbf{w}$ satisfying the conditions
above, and
any $\mathbf{v}$
s.t. $\norm{\mathbf{v}-\mathbf{w}}\leq \frac{\delta}{\sqrt{k-1}}\log(\beta/\alpha)$,
we must
have
$\max_j \left|\prod_{i\neq j}v_i\right|\leq\beta$,
and therefore $|\prod_i v_i|\leq \beta\norm{\mathbf{v}}_{\infty}$, as well as
$\norm{\nabla
F(\mathbf{v})}=\sup_{p:|p|\leq
\beta\norm{\mathbf{v}}_{\infty}}|f'(p)|\cdot\norm{(\prod_{i\neq
1}v_i,\ldots,\prod_{i\neq k}v_i)}\leq
\sup_{p:|p|\leq \beta\norm{\mathbf{v}}_{\infty}}|f'(p)|\sqrt{k}\beta$
by definition of $F$, as
required.
To prove \eqref{eq:toshow0}, we first state and prove the following
auxiliary result:
\begin{align}
\forall \mathbf{w},\mathbf{v}\in\mathbb{R}^{k-1}~~&\text{s.t.}~~
\forall i~v_i\geq w_i\geq 0~~,~~\prod_{i}
w_i\leq
\alpha~~,~~\min_i w_i\geq \delta~~,~~
\prod_{i}v_i> \beta\notag\\
&\text{it holds that} ~~
\norm{\mathbf{v}-\mathbf{w}}~>~
\frac{\delta}{\sqrt{k-1}}\log(\beta/\alpha)~.\label{eq:toshow}
\end{align}
This statement holds by the following calculation:
\begin{align*}
\norm{\mathbf{v}-\mathbf{w}}~&\geq~ \frac{1}{\sqrt{k-1}}
\norm{\mathbf{v}-\mathbf{w}}_1~=~\frac{1}{\sqrt{k-1}}\cdot\sum_i
(v_i-w_i)\\
&\stackrel{(*)}{\geq}~ \frac{1}{\sqrt{k-1}}\sum_i
w_i\left(\log(v_i)-\log(w_i)\right)~\geq~
\frac{\delta}{\sqrt{k-1}}\sum_i\left(\log(v_i)-\log(w_i)\right)\\
&=~
\frac{\delta}{\sqrt{k-1}}\log\left(\frac{\prod_i v_i}{\prod_i
w_i}\right)~>~
\frac{\delta}{\sqrt{k-1}}\log\left(\frac{\beta}{\alpha}\right)~,
\end{align*}
where $(*)$ is due to the fact that $\log(\cdot)$ is $1/z$-Lipschitz in
$[z,\infty)$, and the assumption that $v_i\geq w_i\geq 0$.
It remains to explain how \eqref{eq:toshow} implies \eqref{eq:toshow0}.
Indeed, let $\mathbf{w},\mathbf{v}$ be any two vectors in $\mathbb{R}^k$ which satisfy the
conditions of \eqref{eq:toshow0}. Now, suppose we transform them into
vectors $\mathbf{w}',\mathbf{v}'\in \mathbb{R}^{k-1}$ by the following procedure:
\begin{itemize}
\item Change the sign of every $w_i$ and $v_i$ to be positive
\item For any $i$ such that $v_i<w_i$, change $v_i$ to equal $w_i$.
\item Drop a coordinate $j$ which maximizes $|\prod_{i\neq j} v_i|$.
\end{itemize}
It is easy to verify that the resulting vectors $\mathbf{w}',\mathbf{v}'$ satisfy the
conditions
of \eqref{eq:toshow}, and $\norm{\mathbf{v}'-\mathbf{w}'}\leq \norm{\mathbf{v}-\mathbf{w}}$. Therefore,
by \eqref{eq:toshow}, $\norm{\mathbf{v}-\mathbf{w}}\geq \norm{\mathbf{v}'-\mathbf{w}'}\geq
\frac{\delta}{\sqrt{k-1}}\log(\beta/\alpha)$ as required.
\end{proof}
With these two lemmas in hand, we turn to prove the theorem. By
\lemref{lem:smallinit} and Assumption \ref{assump:initzero}, we have
\[
\Pr\left(\max_j \left|\prod_{i\neq j} w_i(1)\right|\geq
\exp(-2C k)\right)~\leq~
\exp(-C' k)~.
\]
for some fixed constants $C,C'>0$ and any large enough $k$.
Moreover, again by Assumption \ref{assump:initzero}, it holds for any $i$ that
$\Pr(|w_i(1)|\leq \exp(-C k))\leq \mathcal{O}(\exp(-C k))$, so by a
union bound,
\[
\Pr(\min_i |w_i| < \exp(-C k))~\leq~ \mathcal{O}(k\exp(-C k)).
\]
Finally, by Assumption \ref{assump:initzero}, Markov's inequality and a union
bound,
\[
\Pr(\norm{\mathbf{w}(1)}_{\infty}\geq \exp(C k))~\leq k\exp(-C k)
\]
Combining the last three displayed equations with a union bound, and
applying \lemref{lem:flatball} (with $\alpha=\exp(- 2C k))$,
$\beta=2\alpha$, and $\delta=\exp(-C k)$), we get the following: With
probability at least $1-\exp(-C' k)-\mathcal{O}(k\exp(-C
k))-k\exp(-C k)=1-\exp(-\Omega(k))$ over the choice of $\mathbf{w}(1)$,
\begin{itemize}
\item $\norm{\mathbf{w}(1)}_{\infty}\leq \exp(C k)$.
\item For any $\mathbf{v}$ at a distance at most $\exp(-C
k)\frac{\log(2)}{\sqrt{k-1}}$ from $\mathbf{w}(1)$, we have
\[
\norm{\mathbf{v}}_{\infty}~\leq~
\norm{\mathbf{w}(1)}_{\infty}+\exp(-C
k)\frac{\log(2)}{\sqrt{k-1}}~\leq~ \mathcal{O}(\exp(C k))~,
\]
\[
\left|\prod_i v_i\right|~\leq~ \beta\norm{\mathbf{v}}_{\infty} ~=~
2\exp(-2C k)\cdot \mathcal{O}\left(\exp(C k)\right)~=~
\mathcal{O}\left(\exp\left(-C k\right)\right)
\]
and
\begin{align*}
\norm{\nabla F(\mathbf{v})}~&\leq~
\sup_{p:|p|\leq \beta\norm{\mathbf{v}}_{\infty}}|f'(p)|\cdot\sqrt{k}\beta~\leq~
\sup_{p:|p|\leq \mathcal{O}\left(\exp\left(-C k\right)\right)}|f'(p)|\cdot2\sqrt{k}
\exp(-2C k)\\
&=~
\mathcal{O}\left(\sqrt{k}\exp(-2C k)\right)~.
\end{align*}
\end{itemize}
This has two implications:
\begin{enumerate}
\item Since the gradient descent updates are of the form
$\mathbf{w}(t+1)=\mathbf{w}(t)-\eta \nabla F(\mathbf{w}(t))$, and we can assume $\eta\leq
\exp(C k/2)$ by the theorem's conditions, the number of iterations
required to get to a distance
larger than $\exp(-C k)\frac{\log(2)}{\sqrt{k-1}}$ from $\mathbf{w}(1)$ is at
least
\[
\frac{\exp(-C k)\frac{\log(2)}{\sqrt{k-1}}}{\exp(C k/2)\cdot
\mathcal{O}(\sqrt{k}\exp(-2C k))}
~=~ \Omega\left(\frac{\exp(C k/2)}{k}\right)~,
\]
which is at least $\exp(\Omega(k))$ iterations.
\item As long as we are at a distance smaller than the above,
$
\left|\prod_i
v_i\right|~\leq~ \mathcal{O}(\exp(-Ck))\leq \exp(-\Omega(k))
$. In particular, $\prod_i v_i\geq
-1/2$ for large enough $k$, so by Assumption
\ref{assump:f} and definition of $F$, we have that
$F(\mathbf{v})-\inf_{\mathbf{v}}F(\mathbf{v})$ is lower bounded by a constant
independent of $k$.
\end{enumerate}
Overall, we get that with probability at least $1-\exp(-\Omega(k))$, we
initialize at some region in which all points are at least $\Omega(1)$
suboptimal, and at least $\exp(\Omega(k))$ iterations are required to escape
it. This immediately implies our theorem.
\subsection{Proof of \thmref{thm:identity}}
We begin with the following auxiliary lemma, and then turn to analyze the
dynamics of gradient descent in our setting.
\begin{lemma}\label{lem:gm}
For any positive scalars $\alpha,w_1,\ldots,w_k$ such that $\min_i w_i >
\alpha$,
\[
\prod_{i}(w_i-\alpha) ~\leq~
\left(\left(\prod_i w_i\right)^{1/k}-\alpha\right)^k~.
\]
\end{lemma}
\begin{proof}
Taking the $k$-th root and switching sides, the inequality in the lemma is
equivalent to proving
\[
\left(\prod_i (w_i-\alpha)\right)^{1/k}+\alpha~\leq~
\left(\prod_{i}w_i\right)^{1/k}.
\]
Letting $a_i = w_i-\alpha$, and $b_i=\alpha$ for all $i$, the above is
equivalent to proving that
\[
\left(\prod_i a_i\right)^{1/k}+\left(\prod_i b_i\right)^{1/k}~\leq~
\left(\prod_i (a_i+b_i)\right)^{1/k},
\]
namely that the sum of the geometric means of two positive sequences
$(a_i)$ and $(b_i)$ is at most the geometric mean of their sum $(a_i+b_i)$.
This follows from the superadditivity of the geometric mean (see
\citet[Exercise 2.11]{steele2004cauchy})
\end{proof}
\begin{lemma}\label{lem:gap}
If $\min_i w_i(t)\geq C$ and $\prod_i w_i(t)\leq C'$ for some positive
constants $C,C'$,
then for any $j,j'$,
\[
|w_j(t+1)^2-w_{j'}(t+1)^2|~\leq~ |w_j(t)^2-w_{j'}(t)^2|+C''\eta^2
\left(\prod_i w_i(t)\right)^2~,
\]
where $C''$ is some constant dependent only on $C,C'$ and the function $f$.
\end{lemma}
\begin{proof}
By definition,
\begin{align*}
w_j(t+1)^2&-w_{j'}(t+1)^2\notag\\
&=~\left(w_j(t)-\eta f'\left(\prod_i
w_i(t)\right)\prod_{i\neq
j}w_i(t)\right)^2 - \left(w_{j'}(t)-\eta f'\left(\prod_i
w_i(t)\right)\prod_{i\neq
j'}w_i(t)\right)^2 \notag\\
&=~ w_j(t)^2-w_{j'}(t)^2+\eta^2f'\left(\prod_i w_i(t)\right)^2\left(
\left(\prod_{i\neq j}w_i(t)\right)^2-\left(\prod_{i\neq
j'}w_i(t)\right)^2\right)\\
&=~ w_j(t)^2-w_{j'}(t)^2+\eta^2\left(\prod_i w_i(t)\right)^2\cdot
f'\left(\prod_i
w_i(t)\right)^2\left(\frac{1}{w_j(t)^2}-\frac{1}{w_{j'}(t)^2}\right)~.
\end{align*}
By assumption, $0\leq \prod_i w_i(t)\leq C'$ and $\max_j
\frac{1}{w_j(t)^2}\leq
\frac{1}{C^2}$. Therefore, by our assumptions on $f$, the displayed
equation above implies that
\[
|w_j(t+1)^2-w_{j'}(t+1)^2|~\leq~ |w_j(t)^2-w_{j'}(t)^2|+C''\eta^2
\left(\prod_i w_i(t)\right)^2
\]
for some constant $C''>0$ dependent on $C,C'$ and $f$ as required.
\end{proof}
\begin{lemma}\label{lem:decay}
Suppose that at some iteration $t$, for some constant $C$ independent of
$k$, it holds that $\max_i w_i(t)\leq C$ and $\prod_i w_i(t)\leq \beta$ for
some $\beta\in (0,C)$. Then
after at most
$
\tau\leq 1+\mathcal{O}(1)\cdot \frac{\beta^{1/k-1}}{\eta k}
$
iterations, if $\min_{j} w_j(r)\geq 1/2$ for all
$r=t,t+1,\ldots,t+\tau$, then
\begin{itemize}
\item Each $w_i(r)$ as well as $\prod_i w_i(r)$ monotonically decrease
in $r=t,t+1,\ldots,t+\tau$
\item For all $r=t,t+1,\ldots,t+\tau-1$, $\max_j
|w_j(r+1)-w_j(r)|\leq \mathcal{O}(1)\cdot\eta\beta$
\item $\prod_i w_i(t+\tau) \leq
\beta\cdot \exp(-1)$.
\end{itemize}
In the above, $\mathcal{O}(1)$ hides constants dependent only on $C$ and the
function $f$.
\end{lemma}
\begin{proof}
If $\prod_i w_i(t)\leq \beta\cdot\exp(-1)$, we can pick $\tau=0$, and the
lemma trivially holds. Otherwise, let $\tau$ be the smallest (positive)
index such that $\prod_i w_i(t)\leq \beta\cdot\exp(-1)$ (if no such index
exists, take $\tau=\infty$, although the arguments below imply that $\tau$
must be finite). Since we assume $w_i(r)$ for all $i$ are
positive, and $f$ is monotonically increasing,
\[
w_j(r+1)~=~ w_j(r)-\eta f'\left(\prod_i w_i(r)\right)\prod_{i\neq
j}w_i(r)~\leq~ w_j(r),
\]
so $w_j(r)$ monotonically decreases in $r$. Moreover, these are all
positive numbers by assumption, so $\prod_i w_i(r)$ monotonically decreases
in $r$ as well. This shows the first part of the lemma.
As to the second part, the displayed equation above, the fact that $w_j(r)$
and $\prod_i w_i(r)$ decrease in $r$, and our assumptions on $f$ imply that
for any $r<t+\tau$,
\begin{align*}
w_j(r+1)~&=~ w_j(r)-\eta f'\left(\prod_i w_i(r)\right)\prod_{i\neq
j}w_i(r)~=~w_j(r)-\frac{\eta}{w_j(r)} f'\left(\prod_i
w_i(r)\right)\prod_{i}w_i(r)\\
&=~ w_{j}(r)-\Theta(1)\cdot \eta\beta~.
\end{align*}
where $\Theta(1)$ hides constants dependent only on $f$ and $C$. As to the
third part of the lemma, fix some $s<\tau$, and repeatedly apply the
displayed equation above for
$r=t,t+1,\ldots,t+s$, to get that
that $w_j(t+s)\leq
w_j(t)-\Theta(1)\cdot\eta\beta s$ (which is still $\geq 1/2$ by the lemma
assumptions). In that case,
\begin{align*}
\prod_i w_i(t+s)~&\leq~ \prod_i
\left(w_i(t)-\Theta(1)\cdot\eta\beta s\right)
~\stackrel{(*)}{\leq}\left(\beta^{1/k}-\Theta(1)\cdot\eta\beta s\right)^k
~=~ \beta\left(1-\Theta(1)\cdot\eta\beta^{1-1/k}s\right)^k\\
&\leq~\beta\exp\left(-\Theta(1)\cdot
\eta\beta^{1-1/k}s k\right)
\end{align*}
where $(*)$ follows from \lemref{lem:gm} and the fact that $\prod_i
w_i(t)\leq \beta$. The right hand side in turn is at
most $\beta\cdot \exp(-1)$ for any $s\geq C'\beta^{1/k-1}/\eta
k$ for some constant $C'$. In particular, if $\tau>1+C'\beta^{1/k-1}/\eta
k$, then by choosing $s$
s.t. $\tau>s\geq C'\beta^{1/k-1}/\eta k$, we get that
$\prod_i w_i(t+s) \leq \beta\cdot \exp(-1)$ even though $s<\tau$, which
contradicts the definition of $\tau$. Hence $\tau\leq
1+C'\beta^{1/k-1}/\eta
k$ as stated in the lemma.
\end{proof}
Combining \lemref{lem:gap} and \lemref{lem:decay}, we have the following:
\begin{lemma}\label{lem:closealways}
For any constants $C>0$ and index $T$, if $\prod_i w_i(1)\leq C$ and
$w_i(t)\geq\frac{1}{2}$ for all $i=1,\ldots,k$ and $t=1,2,\ldots,T$, then
for all such $t$,
\begin{itemize}
\item Each $w_i(t)$ as well as $\prod_i w_i(t)$
monotonically decrease in $t$.
\item $\max_j |w_j(t+1)-w_j(t)|\leq \mathcal{O}(1)\cdot\eta$
\item $\max_{j,j'}|w_j(t)-w_{j'}(t)|~\leq~
k^{-\Omega(1)}+\mathcal{O}(1)\cdot \left(\eta^2+\frac{\eta}{k}\right)$~.
\end{itemize}
In the above, $\mathcal{O}(\cdot)$ hides constants dependent only on $C$ and the
constants in Assumptions \ref{assump:f} and \ref{assump:initone}.
\end{lemma}
\begin{proof}
The first two parts of the lemma follow from \lemref{lem:decay} and the
fact that by Assumption \ref{assump:initone}, $w_i(1)\leq 1+k^{-\Omega(1)}
\leq \mathcal{O}(1)$. As to the last part, define
$t_0\leq t_1\leq \ldots\leq t_s$ (where $t_0=1$) as the first indices $\leq
T$ such that for all
$r=0,\ldots,s$, $\prod_i
w_i(t_r)\leq
(\prod_i w_i(1))\exp(-r)$ (where $s$ is taken to be as large as possible).
By
\lemref{lem:decay}, we have the following:
\begin{itemize}
\item For all $r=0,\ldots,s-1$, $|t_{r+1}-t_{r}|\leq
1+\mathcal{O}(1)\cdot\frac{\exp(-r)^{1/k-1}}{\eta k}$~.
\item $|T-t_{s}|\leq 1+\mathcal{O}(1)\cdot\frac{\exp(-s)^{1/k-1}}{\eta k}$~.
\item For all $r=0,\ldots,s-1$ and any $t_{r} \leq t \leq t_{r+1}$,
we have $\prod_i w_i(t)\leq \mathcal{O}(1)\cdot\exp(-r)$.
\end{itemize}
Combining this with \lemref{lem:gap}, it follows that for any $j,j'$, and
any $r=0,\ldots,s-1$,
\begin{align*}
|w_j(t_{r+1})^2-w_{j'}(t_{r+1})^2|~&\leq~ |w_j(t_r)^2-w_{j'}(t_r)^2|
+ \mathcal{O}(1)\cdot\eta^2\exp(-2r)\cdot \left(1+\frac{\exp(-r)^{1/k-1}}{\eta
k}\right)\\
&\leq~ |w_j(t_r)^2-w_{j'}(t_r)^2|+\mathcal{O}(1)\cdot
\left(\eta^2\exp(-2r)+\frac{\eta
\exp(-r)}{k}\right)~,
\end{align*}
as well as
\[
|w_j(T)^2-w_{j'}(T)^2|~\leq~|w_j(t_s)^2-w_j(t_s)^2|+\mathcal{O}(1)\cdot
\left(\eta^2\exp(-2s)+\frac{\eta
\exp(-s)}{k}\right)~.
\]
Repeatedly applying the last two displayed equations, and using Assumption
\ref{assump:initone}, we get that
\begin{align*}
|w_j(T)^2-w_{j'}(T)^2|~&\leq~
|w_{j}(1)^2-w_{j'}(1)^2|+\mathcal{O}(1)\cdot\left(\eta^2\sum_{r=0}^{s}\exp(-2r)+\frac{\eta}{k}\sum_{r=0}^{s}\exp(-r)\right)\\
&\leq~
k^{-\Omega(1)}+\mathcal{O}(1)\cdot\left(\eta^2+\frac{\eta}{k}\right)~.
\end{align*}
Since $|w_j(T)^2-w_{j'}(T)^2|=|w_j(T)+w_{j'}(T)|\cdot
|w_j(T)-w_{j'}(T)|\geq |w_j(T)-w_{j'}(T)|$ (as we have $\min_i
w_i(T)\geq 1/2$ by assumption), we get that
$|w_j(T)-w_{j'}(T)|\leq
k^{-\Omega(1)}+\mathcal{O}(1)\cdot\left(\eta^2+\frac{\eta}{k}\right)$ as
required.
\end{proof}
With \lemref{lem:closealways} in hand, we can now prove the theorem. Let
$T$ be
the largest index such that $\min_i w_i(t)\geq 1/2$ for all
$t=1,2,\ldots,T$ (and $\infty$ if this holds for all $t$). It follows that
$\prod_i
w_i(t)\geq 0$,
and therefore, by Assumption \ref{assump:f}, $F(\mathbf{w}(t))-\inf_{\mathbf{w}} F(\mathbf{w})$ is
at least a constant independent of $k$ for all $t=1,2,\ldots,T$. Thus, to prove
the theorem, it is
enough to show that if $T<\infty$, then $T\geq \exp(\Omega(k))$.
By Assumption \ref{assump:initone} and \lemref{lem:closealways}, we have
that $w_1(1)\geq 1-k^{-\Omega(1)}$, $|w_1(t+1)-w_1(t)|\leq \mathcal{O}(1)\cdot\eta$,
and
$\max_j |w_j(t)-w_1(t)|\leq k^{-\Omega(1)}+\mathcal{O}(1)\cdot
\left(\eta^2+\frac{\eta}{k}\right)$.
On the
other hand, if
$T<\infty$, then $\min_i w_i(T+1)<1/2$. Therefore, if $k$ is large enough and
$\eta$ is small enough, there
exists some iteration $t\leq T$ such that $w_j(t)\in [2/3,3/4]$ for all $j$.
This means that $\prod_i w_i(t)\leq (3/4)^k = \exp(-\Omega(k))$. Thus, by
\lemref{lem:decay} (with $\beta=\exp(-\Omega(k))$, from iteration $t$ till
iteration $T$, each $w_j$ decreases
by at most $\mathcal{O}(1)\cdot \eta \beta\leq \exp(-\Omega(k))$ at each iteration.
By assumption, at iteration
$T+1$, there is some $w_j(T+1)< 1/2$, so we must have $T-t\geq
(2/3-1/2)/\exp(-\Omega(k)) =
\exp(\Omega(k))$ as required.
\subsection{Proof of \thmref{thm:posid}}
To prove the theorem, we first state and prove the following key lemma:
\begin{lemma}\label{lem:signswitch}
For any initialization $\mathbf{w}(1)$ and any $(\sigma_1,\ldots,\sigma_k)\in
\{-1,+1\}^k$, let $\mathbf{v}(1),\mathbf{v}(2),\ldots$ denote the iterates produced by
gradient descent starting from $\mathbf{v}(1):=(\sigma_1 w_1(1),\ldots,\sigma_k
w_k(1))$, w.r.t. the function
\[
F_{\sigma}(\mathbf{v}) := \frac{1}{2}\left(\prod_i v_i-\sigma y\right)^2~,
\]
where $\sigma:=\prod_i \sigma_i$.
Then for any $t\geq 1$,
\[
\mathbf{v}(t) = (\sigma_1 w_1(t),\ldots,\sigma_k w_k(t))~~~\text{and}~~~
F(\mathbf{w}(t)) = F_{\sigma}(\mathbf{v}(t))~.
\]
\end{lemma}
\begin{proof}
We prove the lemma by induction. The base case ($t=1$) is immediate from the
definitions and the fact that
\[
F_{\sigma}(\mathbf{v}(1)) = \frac{1}{2}\left(\prod_i \sigma_i w_i(1)-\sigma y\right)^2=
\frac{1}{2}\left(\sigma \prod_i w_i(1)-\sigma y\right)^2=
F(\mathbf{w}(1))~.
\]
Assuming that the induction hypothesis holds for $t$, and recalling that
$\sigma=\prod_i \sigma_i$, we have for any $j\in
\{1,\ldots,k\}$ that
\begin{align*}
v_j(t+1) &= v_j(t)-\left(\prod_i v_i(t)-\sigma y\right)\prod_{i\neq j}v_i(t) =
\sigma_j w_j(t)-\sigma\left(\prod_i w_i(t)-y\right)\prod_{i\neq j}\sigma_i
w_i(t)\\
&= \sigma_j\left(w_j(t)-\left(\prod_i w_i(t)-y\right)\prod_{i\neq
j}w_i(t)\right) =
\sigma_j
w_j(t+1)~.
\end{align*}
As a result,
\[
F_{\sigma}(\mathbf{v}(t+1)) = \frac{1}{2}\left(\prod_i v_i(t+1)-\sigma y\right)^2 =
\frac{1}{2}\left(\sigma \prod_i w_i(t+1)-\sigma y\right)^2 = F(\mathbf{w}(t+1))~.
\]
This establishes the inductive step for $t+1$, hence proving the lemma.
\end{proof}
The lemma implies that for
studying the dynamics of gradient descent starting from any initial point
$(w_1(1),\ldots,w_k(1))$, we can arbitrarily change the signs of its
coordinates, as long as the sign of $y$ is changed accordingly. In particular,
we will assume without loss of generality that all $w_1(1),\ldots,w_k(1)$ are
positive (again, as long as the sign of $y$ is fixed accordingly). The proof
then proceeds as follows:
\begin{itemize}
\item The simplest case is when after the sign transformations, $y>0$. By
our assumptions, this implies that both $y$ and $\prod_i w_i(1)$
switched from being negative (and satisfying $\prod_i w_i(1)>y$) to
positive, hence we now have $y>\prod_i w_i(1)>0$. In that case,
\lemref{lem:phase3} below implies that
$\exp(\tilde{\mathcal{O}}(k))\log(1/\epsilon)$ iterations suffice.
\item The case $y<0$ (which by our assumptions, implies $y<0<\prod_i
w_i(1)$) is more involved: First,
we show that after $t=\exp(\tilde{\mathcal{O}}(k))$ iterations, one (and only one) of
the coordinates of $\mathbf{w}(t)$ becomes non-positive (\lemref{lem:phase1}). Then,
we show that after at most one additional iteration, that non-positive
coordinate
becomes negative and bounded away from $0$ (\lemref{lem:phase2}), the other
coordinates remaining strictly positive. By \lemref{lem:signswitch}, we can
then argue that
at that time point, the dynamics become the same as the scenario where $y>0$,
and all coordinates of the iterate are strictly positive. Again applying
\lemref{lem:phase3}, we get that $\exp(\tilde{\mathcal{O}}(k))\log(1/\epsilon)$
additional iterations suffice for convergence.
\end{itemize}
\subsubsection{The case $y> \prod_i w_i(1)>0$}
We will need the following auxiliary lemma:
\begin{lemma}\label{lem:logab}
For any $a>0$, $b\geq 0$, $\log(a+b)\leq \log(a)+\frac{b}{a}$.
\end{lemma}
\begin{proof}
Since $\log(1+z)\leq z$ for all $z\geq 0$, we have
$\log(a+b)=\log(a(1+b/a))=\log(a)+\log(1+b/a)\leq \log(a)+b/a$.
\end{proof}
\begin{lemma}\label{lem:phase3}
Fix some $\gamma\geq \delta>0$. Suppose that $y>0$, and gradient descent on $F$
is initialized at some $\mathbf{w}(1)$ such that $\prod_i w_i(1)\in [0,y)$,
$w_{j^*}(1)\geq \delta$ for some $j^*\in \arg\min_i w_i(1)$, and
$w_j(1)\geq\gamma$ for all $j\neq j^*$. Assuming step size $\eta \leq
\delta^2/2ky^2$, we have that $F(\mathbf{w}(t))\leq \epsilon$ for any
$t ~\geq~
\frac{\log(y^2/2\epsilon)}{k\delta^2 \gamma^{2(k-2)}\eta}$.
\end{lemma}
\begin{proof}
Let
\[
\mathcal{W}~:= \left\{\mathbf{w}\in \mathbb{R}^k~:~ \prod_i w_i(1)\in [0,y)~,~\min_i
w_i(1)\geq \delta~,~\forall j\neq j^*~w_j(1)\geq\gamma\right\}
\]
denote the set of points in $\mathbb{R}^k$ which satisfy the
initialization conditions of the lemma.
First, we show that if the step size $\eta$ is small enough, then gradient
descent will remain in $\mathcal{W}$ forever. For that, it is enough to show that
for any $\mathbf{w}\in\mathcal{W}$, the update $\mathbf{w}':=\mathbf{w}-\eta\nabla F(\mathbf{w})$ produced by
gradient descent is in $\mathcal{W}$ as well. By definition of $\mathcal{W}$,
it is easily verified that $w'_i\geq
w_i>0$ for all $i$, so the only non-trivial condition to verify is that
$\prod_j
w'_j< y$. To
show this, we note that by \lemref{lem:logab},
\begin{align*}
\log\left(\prod_j w'_j\right) ~&=~ \sum_j \log(w'_j)~=~
\sum_j \log\left(w_j+\eta(y-\prod_i w_i)\prod_{i\neq j}w_i\right)\\
&\leq~
\sum_j \log(w_j)+\eta(y-\prod_i w_i)\sum_j\frac{\prod_{i\neq
j}w_i}{w_j}\\
&=~
\log\left(\prod_j w_j\right)+\eta(y-\prod_i w_i)\left(\prod_i
w_i\right)\sum_j\frac{1}{w_j^2}\\
&<~
\log\left(\prod_j w_j\right)+\eta (y-\prod_i w_i)y\sum_j \frac{1}{\delta^2}\\
&=~
\log\left(\prod_j w_j\right)+\eta \frac{yk}{\delta^2}\left(y-\prod_i
w_i\right)~.
\end{align*}
Thus, to ensure that $\prod_j w'_j< y$ (or equivalently, $\log(\prod_j
w'_j)< \log(y)$), it is enough to ensure that
\[
\log\left(\prod_j w_j\right)+\eta \frac{yk}{\delta^2}\left(y-\prod_i
w_i\right)~\leq~ \log(y)~.
\]
Rearranging the above, we require that
\[
\eta \frac{yk}{\delta^2}~\leq~ \frac{\log(y)-\log(\prod_j w_j)}{y-\prod_j
w_j}~.
\]
By the mean value theorem and the fact that $\prod_j w_j< y$, the right hand
side can be lower bounded by $\min_{z\in (0,y]} \log'(z) = 1/y$, so it is
enough to require
\[
\eta \frac{yk}{\delta^2}~\leq~ \frac{1}{y}~~~\Rightarrow~~~
\eta~\leq~ \frac{\delta^2}{ky^2}~,
\]
which indeed holds by assumption.
Having established that gradient descent will remain in $\mathcal{W}$ forever, we now
establish that the objective $F$ has a $\frac{2ky^2}{\delta^2}$-Lipschitz
gradient on $\mathcal{W}$: Indeed,
the Hessian of $F$ at any $\mathbf{w}\in\mathcal{W}$ can be easily verified to equal
\[
(\nabla^2 F(\mathbf{w}))_{r,s} = \begin{cases} \left(\prod_i
w_i-y\right)\frac{\prod_{i}w_i}{w_r w_s}+\frac{\prod_{i} w^2_i}{w_r w_s}&
r\neq s\\
\frac{\prod_{i}w_i^2}{w_r^2}& r=s\end{cases}~.
\]
Since $\mathbf{w}\in\mathcal{W}$, it follows that magnitude of each entry in the $k\times
k$ Hessian is
at most $y\cdot \frac{y}{\delta^2}+\frac{y^2}{\delta^2}=2y^2/\delta^2$,
and therefore its spectral norm (which is at most the Frobenius norm) can be
upper bounded by $2ky^2/\delta^2$.
The final ingredient we need is that $F$ satisfies
\[
\norm{\nabla F(\mathbf{w})}^2~\geq~ 2k\delta^{2}\gamma^{2(k-2)}F(\mathbf{w})
\]
for any $\mathbf{w}\in\mathcal{W}$ (this type of inequality is known as the
Polyak-\L{}ojasiewicz condition, which ensures linear convergence rates for
gradient descent on possibly non-convex functions -- see
\citet{polyak1963gradient,karimi2016linear}). This
follows
from $\norm{\nabla F(\mathbf{w})}^2$, by definition, being equal to
\[
(\prod_i w_i-y)^2\sum_j\left(\prod_{i\neq j}w_i\right)^2 =
2F(\mathbf{w})\sum_j\left(\prod_{i\neq j}w_i\right)^2 \geq
2F(\mathbf{w})k\left(\delta\gamma^{k-2}\right)^2~.
\]
Collecting these ingredients, we can now perform a standard analysis using the
Polyak-\L{}ojasiewicz condition: If we do a gradient step to get from
$\mathbf{w}\in\mathcal{W}$ to
$\mathbf{w}'\in \mathcal{W}$ (i.e. $\mathbf{w}':=\mathbf{w}-\eta\nabla F(\mathbf{w})$), and assuming $\eta\leq
\delta^2/2ky^2$, then
\begin{align*}
F(\mathbf{w}')~&\leq~ F(\mathbf{w})+\nabla
F(\mathbf{w})^{\top}(\mathbf{w}'-\mathbf{w})+\frac{ky^2}{\delta^2}\norm{\mathbf{w}'-\mathbf{w}}^2\\
&=~ F(\mathbf{w})-\eta\norm{\nabla F(\mathbf{w})}^2+\frac{ky^2}{\delta^2}\norm{\eta\cdot
\nabla
F(\mathbf{w})}^2\\
&=~ F(\mathbf{w})-\eta\left(1-\frac{ky^2}{\delta^2}\eta\right)\norm{\nabla
F(\mathbf{w})}^2\\
&\leq~ F(\mathbf{w})-\eta\cdot \frac{1}{2}\cdot
2k\delta^{2}\gamma^{2(k-2)}F(\mathbf{w})\\
&=~ \left(1-k\delta^{2}\gamma^{2(k-2)}\eta\right)F(\mathbf{w})~\leq~
\exp(-k\delta^{2}\gamma^{2(k-2)}\eta)F(\mathbf{w})~.
\end{align*}
Applying this inequality $t$ times, we
get that
\[
F(\mathbf{w}(t))~\leq~ \exp\left(-k\delta^2\gamma^{2(k-2)}\eta
t\right)F(\mathbf{w}(1))~\leq~\frac{y^2}{2} \exp\left(-k\delta^2\gamma^{2(k-2)}\eta
t\right)~.
\]
Equating the bound above to the target accuracy $\epsilon$ and solving for $t$,
the result follows.
\end{proof}
\subsubsection{The Case $y<0<\prod_i w_i(1)$}
We first state the following auxiliary lemma, which establishes that the gaps
between coordinates are monotonically increasing under suitable assumptions.
\begin{lemma}\label{lem:gapincrease}
Fix some coordinate indices $j,j'$ and iteration $t$, and suppose that
$w_j(t)\leq w_{j'}(t)$, $\min_i w_i(t)\geq 0$, and $y<0$. Then
$
w_{j'}(t)-w_j(t) ~\leq~ w_{j'}(t+1)-w_{j}(t+1)
$.
\end{lemma}
\begin{proof}
Dropping the $(t)$ index to simplify notation, we have by definition that
$w_{j'}(t+1)-w_{j}(t+1)$ equals
\begin{align*}
&\left(w_{j'}-\eta\left(\prod_i
w_i-y\right)\prod_{i\neq
j'}w_i\right)
-\left(w_j-\eta\left(\prod_i w_i-y\right)\prod_{i\neq j}w_i\right)\\
&= \left(w_{j'}-w_j\right)-\eta\left(\prod_i
w_i-y\right)\left(\prod_{i\neq j'}w_i-\prod_{i\neq j}w_i\right)\\
&= \left(w_{j'}-w_j\right)-\eta\left(\prod_i
w_i-y\right)\left(w_{j}-w_{j'}\right)\prod_{i\notin\{j,j'\}}w_i
~=~ \left(w_{j'}-w_j\right)\left(1+\eta\left(\prod_i
w_i-y\right)\prod_{i\notin\{j,j'\}}w_i\right)~.
\end{align*}
Since $y<0$ and $w_i\geq 0$ for all $i$, the above is at least $w_{j'}-w_j=
w_{j'}(t)-w_j(t)$ as required.
\end{proof}
\begin{lemma}\label{lem:phase1}
Suppose $y<0$, and that $\mathbf{w}(1)$ has positive entries which satisfy the theorem
assumptions. If $\eta\leq k^{-C}$ for some sufficiently large constant $C$,
then the
following hold for some
iteration
$t_0\leq
\exp(\tilde{\mathcal{O}}(k))/\eta$:
\begin{itemize}
\item There exists a unique $j^*=\arg\min_i w_i(t_0)$, and $-\mathcal{O}(1)\cdot
\eta\leq w_{j^*}(t_0)\leq 0$.
\item $\min_{j\neq j^*} w_j(t_0)\geq k^{-\mathcal{O}(1)}$, $\max_{j\neq
j^*}w_j(t_0)\leq \mathcal{O}(1)$, and
$\max_{j}\prod_{i\notin \{j,j^*\}}w_i(t_0)\leq \mathcal{O}(1)$.
\end{itemize}
\end{lemma}
It is important to note that the constants hidden in the $\mathcal{O}(\cdot)$ notation
do not depend on $\eta$ (although they may depend on $C$).
\begin{proof}
By \lemref{lem:gapincrease} and the theorem assumptions, the following holds
for all iterations
$t=1,2,\ldots,T$ where $\min_{i,t< T} w_i(t)\geq 0$: There exists a unique
$j^*=\arg\min_i w_i(1)$, $w_{j^*}(t)$ remains the unique smallest value
among $w_1(t),\ldots,w_k(t)$, and its distance from any other coordinate (which
was initially $k^{-\mathcal{O}(1)}$) is
monotonically increasing in $t$. In
particular, for any $t< T$,
$\min_{j\neq j^*}w_j(t)\geq k^{-\mathcal{O}(1)}$. As
a result, recalling that $y<0$, we have for all $t< T$ that
\begin{align*}
w_{j^*}(t+1)~&=~ w_{j^*}(t)-\eta\left(\prod_i
w_i(t)-y\right)\prod_{i\neq
j^*}w_i(t)~\leq~ w_{j^*}(t)-\eta\left(\prod_i
w_i(t)-y\right)\left(k^{-\mathcal{O}(1)}\right)^{k-1}\\
&\leq~ w_{j^*}(t)+\eta y\exp(-\tilde{\mathcal{O}}(k))~\leq~
w_{j^*}(t)-\eta\cdot\exp(-\tilde{\mathcal{O}}(k))~.
\end{align*}
We also assume that initially $w_{j^*}(1)\leq \mathcal{O}(1)$.
Therefore, after at most $t_0=\exp(\tilde{\mathcal{O}}(k))/\eta$
iterations, we will have $w_{j^*}(t_0)\leq 0$ for the first time.
It remains to show that $w_{j^*}(t_0)\geq -\mathcal{O}(1)\cdot\eta$, as well as
the second bullet in the lemma. To that end, we note
that up till iteration $t_0$, for any $j$, both $w_j(t)$ and $\prod_{i\neq
j} w_i(t)$ are monotonically decreasing in
$t$, and moreover, $t_0>1$ (since $w_{j^*}(t_0)\leq 0$ and we assume
$w_{j^*}(1)>0$). Thus, by Assumption \ref{assump:pos},
\begin{align*}
w_{j^*}(t_0) &= w_{j^*}(t_0-1)-\eta\left(\prod_i
w_i(t_0-1)-y\right)\prod_{i\neq
j^*}w_i(t_0-1)\\
&>
0-\eta\left(\prod_i w_i(1)-y\right)\prod_{i\neq j^*}w_i(1)
~\geq~ -\eta\cdot \mathcal{O}(1)~.
\end{align*}
Using this inequality, we have for any $j\neq j^*$
\[
w_j(t_0)~=~ w_{j^*}(t_0)+(w_j(t_0)-w_{j^*}(t_0)) ~\geq~
-\eta\cdot\mathcal{O}(1)+k^{-\mathcal{O}(1)}~,
\]
which is at least $k^{-\mathcal{O}(1)}$ if $\eta \leq k^{-C}$ for some
sufficiently large constant $C$. Finally, since $w_j(t)$ for any $j\neq
j^*$ is positive and monotonically decreasing up to iteration $t_0$, we have
$w_j(t_0)\leq w_j(1)\leq \mathcal{O}(1)$ and $\prod_{i\notin
\{j,j^*\}}w_i(t_0)\leq\prod_{i\notin \{j,j^*\}}w_i(1)\leq
\mathcal{O}(1)$ by Assumption \ref{assump:pos}.
\end{proof}
\begin{lemma}\label{lem:phase2}
Under the conditions of \lemref{lem:phase1},
\begin{itemize}
\item There exists a unique $j^*=\arg\min_i w_i(t_0+1)$, and
$w_{j^*}(t_0+1)~\leq~ -\eta\cdot
\exp(-\tilde{\mathcal{O}}(k))$.
\item $\min_{j\neq j^*} w_j(t_0+1)\geq k^{-\mathcal{O}(1)}$ and
$\prod_{i\neq j^*}w_i(t_0+1)\leq \mathcal{O}(1)$.
\end{itemize}
\end{lemma}
\begin{proof}
By \lemref{lem:phase1}, we have $\prod_i w_i(t_0)\leq 0$, as well as $\prod_i
w_i(t_0)= w_{j^*}(t_0)\cdot w_j(t_0)\cdot \prod_{i\notin \{j,j^*\}}w_i(t_0)\geq
-\mathcal{O}(1)\cdot \eta$ (where $j$ is arbitrary). This
implies
that for sufficiently small $\eta$, $\frac{y}{2}\leq \prod_i w_i(t_0)\leq 0$.
By definition of the gradient descent update, it follows that
$w_{j^*}(t_0+1)\leq
w_{j^*}(t_0)$ and for all $j\neq j^*$,
$w_j(t_0+1)\geq
w_j(t_0)$, which implies that $j^*$ remains the unique coordinate with smallest
value as we move from iteration $t_0$ to iteration $t_0+1$, as well as
$\min_{j\neq j^*} w_j(t_0+1)\geq \min_{j\neq
j^*}w_j(t_0)\geq
k^{-\mathcal{O}(1)}$.
We now turn to prove $w_{j^*}(t_0+1)~\leq~ -\eta\cdot
\exp(-\tilde{\mathcal{O}}(k))$. Using the fact that $\frac{y}{2}\leq \prod_i
w_i(t_0)\leq 0$ as noted earlier,
\[
w_{j^*}(t_0+1) = w_{j^*}(t_0)-\eta\left(\prod_i w_i(t_0)-y\right)\prod_{i\neq
j^*}w_i(t_0)~\leq~ 0+\eta\cdot\frac{y}{2}\cdot\left(k^{-\mathcal{O}(1)}\right)^{k-1}
~\leq~ -\eta\cdot \exp(-\tilde{\mathcal{O}}(k))~.
\]
Finally, to prove $\prod_{i\neq j^*} w_i(t_0+1)\leq \mathcal{O}(1)$, we have by
definition that for any $j\neq j^*$,
\[
w_{j}(t_0+1) ~=~ w_{j}(t_0)-\eta\left(\prod_i
w_i(t_0)-y\right)w_{j^*}(t_0)\cdot\prod_{i\notin \{j,j^*\}}w_i(t_0)~.
\]
Using the fact that
$\frac{y}{2}\leq \prod_i w_i(t_0)\leq 0$ as shown earlier, and noting that by
\lemref{lem:phase1}, $|w_{j^*}(t_0)|\leq \mathcal{O}(1)\cdot \eta$ and
$\left|\prod_{i\notin \{j,j^*\}}w_i(t_0)\right|\leq \mathcal{O}(1)$, it follows from
the displayed equation above that
$w_{j}(t_0+1)\leq w_{j}(t_0)+\mathcal{O}(1)\cdot \eta^2$. Therefore,
\[
\prod_{i\neq j^*} w_i(t_0+1)~\leq~ \prod_{i\neq
j^*}\left(w_i(t_0)+\mathcal{O}(1)\cdot \eta^2\right)~\leq~ \prod_{i\neq j^*}
\left(w_i(t_0)\left(1+\frac{\mathcal{O}(1)\eta^2}{w_i(t_0)}\right)\right)~.
\]
Since $\min_{j\neq j^*} w_j(t_0)\geq k^{-\mathcal{O}(1)}$ (where the $\mathcal{O}(1)$ does
not depend on $\eta$), then by picking $\eta\leq
k^{-C}$ for a sufficiently large $C$, the above is at most
$
\left(\prod_{i\neq
j^*}w_i(t_0)\right)\left(1+\frac{\mathcal{O}(1)}{k}\right)^{k-1}~\leq~
\left(w_j(t_0)\prod_{i\notin\{j,j^*\}} w_i(t_0)\right)\cdot \mathcal{O}(1)~\leq~
\mathcal{O}(1)$,
where we used \lemref{lem:phase1} and where $j$ is arbitrary.
\end{proof}
\subsubsection{Putting Everything Together}
As discussed at the beginning of the proof, we can assume w.l.o.g. that
$w_1(1),\ldots,w_k(1)$ are all positive (and in fact, $\min_i w_i(1)\geq
k^{-\mathcal{O}(1)}$ by our assumptions), and only consider the cases $y>\prod_i
w_i(1)>0$ and $y <0<\prod_i w_i(1)$.
\begin{itemize}
\item If $y>\prod_i
w_i(1)>0$, we can apply
\lemref{lem:phase3} with $\gamma = \delta = k^{-\mathcal{O}(1)}$ and any $\eta =
k^{-c}$ for some large enough constant $c$, to get a convergence
to an $\epsilon$-optimal solution in
$\exp(\tilde{\mathcal{O}}(k))\cdot\max\{1,\log\left(1/\epsilon\right)\}$
iterations.
\item If $y<0<\prod_i w_i(1)$, and assuming $\eta=k^{-c}$ for some large enough
constant
$c>0$,
then \lemref{lem:phase1} and \lemref{lem:phase2} together tell
us that after at most $\exp(\tilde{\mathcal{O}}(k))$ iterations, we get to an
iteration $t=t_0+1$ where $w_{j^*}(t)\leq
-\exp(-\tilde{\mathcal{O}}(k))$ for some
$j^*$,
$w_{j}(t)\geq k^{-\mathcal{O}(1)}$ for all $j\neq j^*$, and $0>
\prod_i w_i(t)\geq
-\mathcal{O}(1)\cdot \eta\geq -\mathcal{O}(1)\cdot k^{-\Omega(1)}>y$ for large enough $k$.
Therefore, by
\lemref{lem:signswitch}, the dynamics of gradient descent from this time point
is identical to case where we switch the signs of $y$ and $w_{j^*}$, so that
$y>0$, $w_{j^*}(t)\geq \exp(-\tilde{\mathcal{O}}(k))$,
$w_j(t)\geq k^{-\mathcal{O}(1)}$ for all $j\neq j^*$, and $y>\prod_i w_i(t)>0
$ for large enough $k$. Now applying
\lemref{lem:phase3} with
$\delta=\exp(-\tilde{\mathcal{O}}(k))$, $\gamma=k^{-\mathcal{O}(1)}$, and any step size
$\eta=k^{-c}$ for some large enough $c$, we get that
$\exp(\tilde{\mathcal{O}}(k))\cdot\max\{1,\log(1/\epsilon)\}$ additional iterations
suffice for
convergence. Overall, $\exp(\tilde{\mathcal{O}}(k))\cdot\max\{1,\log(1/\epsilon)\}$
iterations are
sufficient.
\end{itemize}
\bibliographystyle{plainnat}
|
\section{Introduction}
Automated Essay Scoring (AES) and Automated Short Answer Scoring (ASAS) has become more prevalent among testing agencies \cite{ETS, IEA, PEG,Intellimetric}. These systems are often designed to address one task and one task alone; to determine whether a written piece of text addresses a question or not. These engines were originally based on either hand-crafted features or term frequency–inverse document frequency (TF-IDF) approaches \cite{BOW}. More recently, these techniques have been superseded by the combination of word-embeddings and neural networks \cite{Embedding, NNAES1, NNAES2}. For semantically simple responses, the accuracy of these approaches can often be greater than accuracy of human raters, however, these systems are not trained to appropriately deal with the anomalous cases in which a student writes something that elicits concern for the writer or those around them, which we simply call an `alert'. Typically essay scoring systems do not handle alerts, but rather, separate systems must be designed to process these types of responses before they are sent to the essay scoring system. Our goal is not to produce a classification, but rather to use the same methods developed in AES, ASAS and sentiment analysis \cite{TangSentiment, WangSentiment} to identify some percentage of responses that fit patterns seen in known alerts and send them to be assessed by a team of reviewers.
Assessment organizations typically perform some sort of alert detection as part of doing business. In among hundreds of millions of long and short responses we find cases of alerts in which students have outlined cases of physical abuse, drug abuse, depression, anxiety, threats to others or plans to harm themselves \cite{EssayHelp}. Such cases are interesting from a linguistic, educational, statistical and psychological viewpoint \cite{Language}. While some of these responses require urgent attention, given the volume of responses many testing agencies deal with, it is not feasible to systematically review every single student response within a reasonable time-frame. The benefits of an automated system for alert detection is that we can prioritize a small percentage which can be reviewed quickly so that clients can receive alerts within some fixed time period, which is typically 24 hours. Given the prevalence of school shootings and similarly urgent situations, reducing the number of false positives can effectively speed up the review process and hence optimize our clients ability to intervene when necessary.
As a classification problem in data science, our problem has all the hallmarks of the most difficult problems in natural language processing (NLP) \cite{LillyNLP}; alerts are anomalous in nature making training difficult, the data is messy in that it contains misspellings (both misused real words and incorrectly spelled words) \cite{Spelling}, students often use student specific language or multi-word colloquialisms \cite{Multiword} and the semantics of alerts can be quite complex and subtle, especially when the disturbing content is implicit rather than explicit. The responses themselves are drawn from a wide range of free-form text responses to questions and student comments from a semantically diverse range of topics, including many that are emotive in nature. For example, the semantic differences between an essay on gun-control and a student talking about getting a gun can be very subtle. Sometimes our systems include essays on emotive topics because the difference in language between such essays and alerts can be very small. Students often use phrases like ``kill me now" as hyperbole out of frustration rather than a genuine desire to end ones life, e.g., "this test is so boring, kill me now". To minimize false positives, the engine should attempt to evaluate context, not just operate on key words or phrases.
When it comes to neural network design, there are two dominant types of neural networks in NLP; convolutional neural networks (CNN) and recurrent neural networks (RNN)\cite{RNNNLP}. Since responses may be of an arbitrary length different recurrent neural networks are more appropriate tools for classifying alerts \cite{CNNNLP}. The most common types of cells used in the design of recurrent neural networks are Gated Recurrent Units (GRU)s \cite{GRU2014} and Long-Short-Term-Memory (LSTM) units \cite{LSTM1997}. The latter were originally designed to overcome the vanishing gradient problem \cite{VanishingGrad}. The GRU has some interesting properties which simplify the LSTM unit and the two types of units can give very similar results \cite{GRUvsLSTM}. We also consider stacked versions, bidirectional variants \cite{Bidirectional1997} and the effect of an attention mechanism \cite{Bahdanau:Attention}. This study has been designed to guide the creation of our desired final production model, which may include higher stacking, dropouts (both regular and recurrent) and may be an ensemble of various networks tuned to different types of responses \cite{Ensemble}. Similar comparisons of architectures have appeared in the literature \cite{Tweets, NNAES2}, however, we were not able to find similar comparisons for detecting anomalous events.
In section \ref{Data} we outline the nature of the data we have collected, a precise definition of an alert and how we processed the data for the neural network. In section \ref{RNNs} we outline the definition of the models we evaluate and how they are defined. In section \ref{sec:Results} we outline our methodology in determining which models perform best given representative sensitivities of the engine. We attempt to give an approximation of the importance of each feature of the final model.
\section{Defining the Data}\label{Data}
The American Institutes for Research tests up to 1.8 million students a day during peak testing periods. Over the 2016--2017 period AIR delivered 48 million online tests across America. Each test could involve a number of comments, notes and long answer free-form text responses that are considered to be a possible alerts as well as equations or other interactive items that are not considered to be possible alerts. In a single year we evaluate approximately 90 million free-form text responses which range anywhere from a single word or number to ten thousand word essays. These responses are recorded in html and embedded within an xml file along with additional information that allows our clients to identify which student wrote the response. The first step in processing such a response is to remove tags, html code and any non-text using regular expressions.
To account for spelling mistakes, rather than attempt to correct to a vocabulary of correctly spelled words, we constructed an embedding with a vocabulary that contains both correct and incorrectly spelled words. We do this by using standard algorithms \cite{Word2Vec} on a large corpus of student responses (approximately 160 million responses). The embedding we created reflects the imperfect manner in which students use words \cite{mispellingEmbedding}. For example, while the words 'happems' and 'ocures' are both incorrectly spelled versions of 'happens' and 'occurs' respectively, our embedding exhibits a high cosine similarity between the word vectors of the correct and incorrect versions. The embedding we created was an embedding into 200 dimensional space with a vocabulary consisting of 1.12 million words. Using spelling dictionaries we approximate that the percentage of correctly spelled words in the vocabulary of this embedding is approximately 7\%, or roughly 80,000 words, while the remaining 93\% are either misspellings, made up words or words from other languages. Lastly, due to the prevalence of words that are concatenated (due to a missing space), we split up any word with a Levenstein distance that is greater than two from our vocabulary into smaller words that are in the vocabulary. This ensures that any sentence is tokenized into a list of elements, almost all of which have valid embeddings.
In our classification of alerts, with respect to how they are identified by the team of reviewers, we have two tiers of alerts, Tier A and Tier B. Tier A consists of true responses that are alarming and require urgent attention while Tier B consists of responses that are concerning in nature but require further review. For simplification, both types of responses are flagged as alerts are treated equivalently by the system. This means the classification we seek is binary. Table \ref{TierA} and Table \ref{TierB} outline certain subcategories of this classification in addition to some example responses.
\begin{table}[!ht]
\ra{1.3}
\begin{tabular}{p{3cm} p{5cm} p{8cm}}\\ \toprule
{\bf Category A} & {\bf Details} & {\bf Examples} \\ \toprule
Harm to self or another being. & Suicide, self-harm, or extreme depression; threats or reports of violence, rape, abuse, drugs, or neglect. & I wanna kill myself. Why does my dad beat me at night?\\ \midrule
Contains mention of a gun. & Doesn’t have to be threatening. & I want a sniper rifle.\\ \midrule
Specific and serious request for help.& Not test-related & I hate my life, please help.\\\bottomrule\\
\end{tabular}
\caption{Student response that meet the requirement to trigger an immediate alert notification to client. \label{TierA}}
\end{table}
\begin{table}[!ht]
\ra{1.3}
\begin{tabular}{p{3cm} p{5cm} p{8cm}}\toprule
{\bf Category B} & {\bf Details} & {\bf Examples} \\ \toprule
Non-specific request for help. & Not specific to test or harm.& I need help real bad help me please.\\ \midrule
Sexual Imagery. & Without threats or reports of abuse. & My uncle touches me.\\ \midrule
Violent words or phrases. & No explicit reports of being the perpetrator or victim of violence, but text seems suspect.& Death by suffocation.\\ \midrule
Signs of depression, self-loathing, or anxiety .
& Sad, lack of social support, dissatisfaction for life, grief, anxiety, negative attitude towards self. & No one loves me. I am stupid and ugly.\\\bottomrule \\
\end{tabular}
\caption{Student response contains watch words or phrases (these may be required to undergo review further). \label{TierB}}
\end{table}
The American Institutes for Research has a hand-scoring team specifically devoted to verifying whether a given response satisfies the requirements of being an alert. At the beginning of this program, we had very few examples of student responses that satisfied the above requirements, moreover, given the diverse nature of what constitutes an alert, the alerts we did have did not span all the types of responses we considered to be worthy of attention. As part of the initial data collection, we accumulated synthetic responses from the sites Reddit and Teen Line that were likely to be of interest. These were sent to the hand-scoring team and assessed as if they were student responses. The responses pulled consisted of posts from forums that we suspected of containing alerts as well as generic forums so that the engine produced did not simply classify forum posts from student responses. We observed that the manner in which the students engaged with the our essay platform in cases of alerts mimicked the way in which students used online forums in a sufficiently similar manner for the data to faithfully represent real alerts. This additional data also provided crucial examples of classes of alerts found too infrequently in student data for a valid classification. This initial data allowed us to build preliminary models and hence build better engines.
Since the programs inception, we have greatly expanded our collection of training data, which is summarized below in Table \ref{traindat}. While we have accumulated over 1.11 million essay responses, which include many types of essays over a range of essay topics, student age ranges, styles of writing as well as a multitude of types of alerts, we find that many of them are mapped to the same set of words after applying our preprocessing steps. When we disregard duplicate responses after preprocessing, our training sample consists of only 866,137 unique responses.
\begin{table}[!ht]
\begin{tabular}{l l | c c c c} \toprule
Category & & Alerts & Normal & Unclassified & Total \\ \toprule
Training & Synthetic & 5012 & 67025&0& 72037 \\
(with Duplicates)& Real &7448 & 1035530&0 & 1042978\\ \midrule
Total&& 12460 & 1102555 & 0 & 1115015 \\ \toprule
Training & Synthetic &4912&57988& 0& 62900 \\
(Unique) & Real &5615 &797622& 0 & 803237\\ \midrule
Total && 10527 & 855610 & 0 & 866137 \\ \toprule
Threshold data & & 14 & 186 &199814 & 200014 \\ \bottomrule \\
\end{tabular}
\caption{The table gives the precise number of examples used in training both before preprocessing (with possible duplicates) and after (unique responses) as well as an unclassified set we used for determining an approximation of the percentage of responses flagged by the engine at various levels of sensitivity.\label{traindat}}
\end{table}
Our training sample has vastly over-sampled alerts compared with a typical responses in order to make it easier to train an engine. This also means that a typical test train split would not necessarily be useful in determining the efficacy of our models. The metric we use to evaluate the efficacy of our model is an approximation of the probability that a held-out alert is flagged if a fixed percentage of a typical population were to be flagged as potential alerts.
This method also lends itself to a method of approximating the number of alerts in a typical population. we use any engine produced to score a set of responses, which we call the threshold data, which consisted of a representative sample of 200,014 responses. Using these scores and given a percentage of responses we wish to flag for review, we produce a threshold value in which scores above this threshold level are considered alerts and those below are normal responses. This threshold data was scored using our best engine and the 200 responses that looked most like alerts were sent to be evaluated by our hand-scorers and while only 14 were found to be true alerts. Using the effectiveness of the model used, this suggests between 15 and 17 alerts may be in the entire threshold data set. We aggregated the estimates at various levels of sensitivity in combination with the efficacy of our best model to estimate that the rate of alerts is approximately 77 to 90 alerts per million responses. Further study is required to approximate what percentage are Tier A and Tier B.
\section{Recurrent Structures Considered}\label{RNNs}
Since natural languages contain so many rules, it is inconceivable that we could simply list all possible combinations of words that would constitute an alert. This means that the only feasible models we create are statistical in nature. Just as mathematicians use elementary functions like polynomials or periodic functions to approximate smooth functions, recurrent neural networks are used to fit classes of sequences. Character-level language models are typically useful in predicting text \cite{charlevel}, speech recognition \cite{Speech} and correcting spelling, in contrast it is generally accepted that semantic details are encoded by word-embedding based language models \cite{embedNLP}.
Recurrent neural networks are behind many of the most recent advances in NLP. We have depicted the general structure of an unfolded recurrent unit in figure \ref{RNN}. A single unit takes a sequence of inputs, denoted $x$ below, which affects a set of internal states of the node, denoted $a$, to produce an output, $h_n$. A single unit either outputs a single variable, which is the output of the last node, or a sequence of the same length of the input sequence, $h = (h_1, \ldots, h_n)$, which may be used as the input into another recurrent unit.
A layer of these recurrent units is a collection of independent units, each of which may pick up a different aspect of the series. A recurrent layer, consisting of $t$ independent recurrent units, has the ability to take the most important/prevalent features and summarize those features in a vector of length $t$. When we feed the sequence of outputs of one recurrent layer into another recurrent layer, we call this a stacked recurrent layer. Analogous to the types of features observed in stacking convolutional and dense layers in convolutional neural networks \cite{DeepImages}, it is suspected that stacking recurrent layers allows a neural network to model more semantically complex features of a text \cite{DeepRNN, DeepOpinionRNN}.
\begin{figure}[!ht]
\begin{tikzpicture}[scale=1.3]
\node [draw, fill=blue!10,circle,minimum size=.9cm] (a) at (-3,1) {${}_a$};
\node [draw, fill=blue!10,circle,minimum size=.9cm] (0) at (0,1) {$0$};
\node [draw, fill=blue!10,circle,minimum size=.9cm] (a1) at (1,1) {${}_{a_1}$};
\node [draw, fill=blue!10,circle,minimum size=.9cm] (a2) at (2,1) {${}_{a_2}$};
\node [draw, fill=blue!10,circle,minimum size=.9cm] (anm1) at (4,1) {${}_{a_{n-1}}$};
\node [draw, fill=blue!10,circle,minimum size=.9cm] (an) at (5,1) {${}_{a_n}$};
\node at (3,1) {$\ldots$};
\node at (3,0) {$\ldots$};
\node at (3,2) {$\ldots$};
\node [draw, fill=red!10,circle,minimum size=.9cm] (x) at (-3,0) {${}_x$};
\node [draw, fill=red!10,circle,minimum size=.9cm] (x1) at (1,0) {${}_{x_1}$};
\node [draw, fill=red!10,circle,minimum size=.9cm] (x2)at (2,0) {${}_{x_2}$};
\node [draw, fill=red!10,circle,minimum size=.9cm] (xnm1) at (4,0) {${}_{x_{n-1}}$};
\node [draw, fill=red!10,circle,minimum size=.9cm] (xn) at (5,0) {${}_{x_n}$};
\node [draw, fill=green!10,circle,minimum size=.9cm] (h) at (-3,2) {${}_{h_n}$};
\node (h1) at (1,2) {${}_{h_1}$};
\node (h2)at (2,2) {${}_{h_2}$};
\node (hnm1) at (4,2) {${}_{h_{n-1}}$};
\node [draw, fill=green!10,circle,minimum size=.9cm] (hn) at (5,2) {${}_{h_n}$};
\draw[->,thick] (x)--(a);
\draw[->,thick] (x1)--(a1);
\draw[->,thick] (x2)--(a2);
\draw[->,thick] (xnm1)--(anm1);
\draw[->,thick] (xn)--(an);
\draw[->,thick] (0)--(a1);
\draw[->,thick] (a1)--(a2);
\draw[->,thick] (a2) -- (2.5,1);
\draw[->,thick] (3.5,1) -- (anm1);
\draw[->,thick] (anm1) -- (an);
\draw[->,thick] (a1) -- (h1);
\draw[->,thick] (a2) -- (h2);
\draw[->,thick] (anm1) -- (hnm1);
\draw[->,thick] (an) -- (hn);
\draw[->,thick] (a) -- (h);
\end{tikzpicture}
\caption{\label{RNN}When we unfold an RNN, we express it as a sequence of cell each accepting, as input, an element of the sequence. The output of the RNN is the output of the last state.}
\end{figure}
The collections of variables associated with the state of the recurrent units, which are denoted $a$ in figure \ref{RNN}, and their relations between the inputs, $x$, and the outputs are what distinguishes simple recurrent units, GRUs and LSTM units. In our case, $x =(x_1,x_2, \ldots, x_n)$ is a sequence of word-vectors. The underlying formulas for gated recurrent units are specified by the initial condition $h_0 = 0$ and
\begin{subequations}
\begin{align}
z_t &= \sigma_g (W_z x_t + U_z h_{t-1} + b_z),\\
r_t &= \sigma_g (W_r x_t + U_r h_{t-1} + b_r),\\
h_t & = \tilde{z}_t \circ h_{t-1} + z_t \circ y_t,\\
y_t &= \sigma_h (W_h x_t + U_h(r_t \circ h_{t-1}) + b_h),
\end{align}
\end{subequations}
where $\circ$ denotes the element-wise product (also known as the Hadamard product), $x_t$ is an input vector $h_t$ is an output vector, $z_t$ is and update gate, $\tilde{z}_t = 1-z_t$, $r_t$ is a reset gate, subscripted variables $W$, $U$ and $b$ are parameter matrices and a vector and $\sigma_h$ and $\sigma_h$ are the original sigmoid function and hyperbolic tangent functions respectively \cite{GRU2014}.
The second type of recurrent unit we consider is the LSTM, which appeared in the literature before the GRU and contains more parameters \cite{LSTM1997}. It was created to address the vanishing gradient problem and differs from the gated recurrent unit in that it has more parameters, hence, may be regarded as more powerful.
\begin{subequations}
\begin{align}
f_t &= \sigma_g (W_f x_t + U_f h_{t-1} + b_f),\\
i_t &= \sigma_g (W_i x_t + U_i h_{t-1} + b_i),\\
o_t &= \sigma_g (W_o x_t + U_o h_{t-1} + b_o),\\
c_t &= f_t \circ c_{t-1} + i_t \circ y_t,\\
h_t &= o_t \circ \sigma_h(c_t), \\
y_t & = \sigma_h (W_z x_t + U_z h_{t-1} + b_z),
\end{align}
\end{subequations}
where $x_t$ is the input, $c_t$ is the cell state vector, $f_t$ is the forget gate, $i_t$ is the input gate, $o_t$ is the output gate and $h_t$ is the output, $z_t$ is a function of the input and previous output while subscripted variables $W$, $U$ and $b$ are parameter matrices and a vector. Due to their power, LSTM layers are ubiquitous when dealing with NLP tasks and are being used in many more contexts than layers of GRUs \cite{NLPLSTM}.
\begin{figure}[!ht]
\begin{tikzpicture}[scale=.9]
\draw[fill=blue!10] (0,0) rectangle (6,3);
\node [draw, fill=green!10,circle,minimum size=1cm] (ct) at (4,1.5) {${}_{c_t}$};
\draw[very thick,->] (ct) -- (4,3.5);
\draw[very thick,->] (-.5,1.5) -- (ct);
\node [draw, fill=red!10,rectangle,minimum size=.5cm] (ft) at (.7,1.5) {${}_{f_t}$};
\node [draw, fill=red!10,rectangle,minimum size=.5cm] (ot) at (4,2.5) {${}_{o_t}$};
\node [draw, fill=red!10,rectangle,minimum size=.5cm] (it) at (3,.7) {${}_{i_t}$};
\node [draw, fill=yellow!10,circle] (yt) at (1.4,.7) {${}_{y_t}$};
\draw[very thick] (it) -- (yt);
\node[draw, fill=blue!10,circle,minimum size=.2cm] (plus) at (3,1.5) {};
\node at (3,1.5) {$+$};
\draw[very thick] (plus) -- (it);
\node at (-.8,1.5) {$c_{t-1}$};
\node at (4,3.8) {$h_{t}$};
\node at (6.5,2.5) {$h_{t}$};
\node at (1.4,-.5) (xt) {$x_{t}$};
\node at (-.8,.7) (htm1) {$h_{t-1}$};
\draw[very thick,->] (xt) -- (yt);
\draw[very thick,->] (ot) -- (6.2,2.5);
\draw[very thick,->] (htm1) -- (yt);
\draw[very thick,->] (ct) -- (6.2,1.5);
\node at (6.5,1.5) {$c_t$};
\end{tikzpicture}
\hspace{1cm}
\begin{tikzpicture}[scale=.9]
\draw[fill=blue!10] (0,0) rectangle (6,3);
\node at (3.8,3.8) (ht) {$h_{t}$};
\node at (4.6,-.5) (xt) {$x_{t}$};
\node at (-.8,.7) (htm1) {$h_{t-1}$};
\node [draw, fill=red!10,rectangle,minimum size=.5cm] (zt) at (4.6,1.8) {${}_{z_t}$};
\node [draw, fill=yellow!10,circle] (yt) at (4.6,.7) {${}_{y_t}$};
\draw[very thick,->] (-.4,.7) --(yt);
\draw[very thick,->] (xt) --(yt);
\draw[very thick] (yt) --(zt);
\draw[very thick] (1.5,.7) --(1.5,1.8) -- (zt);
\node[draw, fill=blue!10,circle,minimum size=.2cm] (plus) at (3.8,1.8) {};
\node at (3.8,1.8) {$+$};
\node [draw, fill=red!10,rectangle,minimum size=.5cm] (tzt) at (3,1.8) {${}_{\tilde{z}_t}$};
\node [draw, fill=red!10,rectangle,minimum size=.5cm] (ot) at (3,.7) {${}_{r_t}$};
\draw[very thick,->] (plus) --(ht);
\draw[very thick,->] (3.8,2.5) --(6.2,2.5);
\node at (6.5,2.5) {$h_{t}$};
\end{tikzpicture}
\caption{The left is an LSTM; the gates $f_t$, $i_t$ and $o_t$ are vectors of values between 0 and 1 that augment the input being passed through them by selecting which features to keep and which to discard. The input $x_t$, $h_{t-1}$, $c_{t-1}$ and $y_t$ are more generically valued vectors. The GRU is based on a similar concept with a simpler design where an update gate, $z_t$ decides which features of the previous output to keep as output of the new cell and which features need to contain input specific information, which is stored in $y_t$.}
\end{figure}
Given a recurrent unit, the sequence $x$ is fed into the recurrent unit cell by cell in the order it appears, however, it was found that some recurrent networks applied to translation benefited from reversing the ordering of the sequence, so that the recurrent units are fed the vectors from last to first as opposed to first to last. Indeed, it is possible to state the most important information at the beginning of a text or at the end. The idea behind bidirectional recurrent units is that we double the number of set units and have half the units fed the sequence in the right order, while the other half of the units are fed the sequence in reverse. Due to the lack of symmetry in the relations between states, we are potentially able to model new types of sequences in this way.
The last mechanism we wish to test is an attention mechanism \cite{Bahdanau:Attention}. The key to attention mechanisms is that we apply weights to the sequences, $h_t$, outputted by the recurrent layer, not just the final output. This means that the attention is a function of the intermediate states of the recurrent layer as well as the final output. This may be useful when identifying when key phrases are mentioned for example. This weighted sequence is sent to a soft-max layer to create a context vector. The attention vector is then multiplied by $h_t$ to produce resulting attention vector, $\tilde{h}_t$. We have implemented the following attention mechanism
\begin{subequations}
\begin{align}
\tilde{h}_t &= c_t h_t,\\
c_t &= \sum \alpha_{t,j} h_j,\\
\alpha_{i,j} &= \dfrac{\exp(e_{ij})}{\sum \exp(e_{ik})},
\end{align}
\end{subequations}
where $h_t$ was the output from the LSTM layer, the $e_{ij}$ are linear transformations of the $h_j$ and $\tilde{h}_t$ is the attended output, i.e., the output of the attention layer \footnote{An attention function, e.g., $f = \mathrm{tanh}$, is often used so that $\tilde{h}_t = f(c_t h_t)$. We tested this approach only to find similar overall results in our experiments \cite{MPM2015:Attention}.}. This mechanism has been wildly successful in machine translation \cite{MPM2015:Attention, GoogleTranslation} and other tasks \cite{Xu}.
\section{Methodology and Results}\label{sec:Results}
Unlike many tasks in NLP, our goal is not to explicitly maximize accuracy. The framework is that we may only review a certain percentage of documents, given this, we want to maximize the probability than an alert will be caught. I.e., the cost of a false-positive is negligible, while we consider false negatives to be more serious. Conversely, this same information could be used to set a percentage of documents required to be read in order to have have some degree of certainty that an alert is flagged. If we encode all alerts with the value 1 and all normal documents with a value of 0, any neural network model will serve as a statistical mechanism in which an alert that was not used in training will, a priori, be given a score by the engine from a distribution of numbers between 0 and 1 which is skewed towards 1 while normal documents will also have scores from another distribution skewed towards 0. The thresholds values where we set are values in which all scores given by the engine above the cut-off are considered possible alerts while all below are considered normal. We can adjust the number of documents read, or the percentage of alerts caught by increasing or decreasing this cut-off value.
To examine the efficacy of each model, our methodology consisted of constructing three sets of data:
\begin{enumerate}
\item{A small collection of alerts that was removed from training.}
\item{Our training set consisting of almost all our data except for a small collection of alerts.}
\item{A sufficiently large corpus of unclassified generic test responses, which was sampled as uniformly as possible over the various sources of responses.}
\end{enumerate}
The idea is that we use the generic test responses to determine how each model would score the types of responses the engine would typically see. While the number of alerts in any set can vary wildly, it is assumed that the set includes both normal and alert responses in the proportions we expect in production. Our baseline model is logistic regression applied to a TF-IDF model with latent semantic analysis used to reduce the representations of words to three hundred dimensions. This baseline model performs poorly at lower thresholds and fairly well at higher thresholds.
\begin{table}[!ht]
\begin{tabular}{l c c c c c c c }\toprule
{\bf Model} & {\bf Configuration} & {\bf 0.1 \%} & {\bf 0.3\%} & {\bf 0.5\%} & {\bf 1\%} & {\bf 2\%} & {\bf 4\%} \\ \toprule
BOW + LSA + Logistic Regression& & 64.8 & 82.2 & 86.9 & 91.3 & 94.6 & 96.3 \\\toprule
GRU & (512) & 51.4 & 62.3 & 67.4 & 73.9 & 80.3 & 85.4\\ \midrule
Stacked GRU & (256,256) & 73.2 & 80.6 & 83.5 & 87.1 & 90.6 & 93.2\\ \midrule
Bidirectional GRU & (256) & 55.7 & 64.7 & 68.8 & 74.9 & 81.6 & 86.6\\ \midrule
Bidirectional Stacked GRU & (128,128) & 80 & 86.8 & 89.2 & 92.2 & 94.4 & 95.2\\ \midrule
GRU with Attention & (512) & 55.4 & 67.5 & 72 & 77.6 & 84 & 91.9\\ \midrule
Stacked GRU with Attention & (256,256) & 69.1 & 78.7 & 81.9 & 86.2 & 89.5 & 92.6 \\ \midrule
Bidirectional GRU with Attention& (256) &59.8 & 70 & 75.6 & 80.6 & 87.4 & 93.3\\ \midrule
Bidirectional Stacked GRU with Attention & (128,128) & 76.6 & 84.2 & 85.9& 90.2 & 93.4 & 95.3\\ \midrule
LSTM &(512) & 66.6 & 72.8 & 75.8 & 78.6 & 86 & 92.1 \\ \midrule
Stacked LSTM &(256,256) & 80.8 & 87.5 &89.6 & 93 & 94.8 & 96.7 \\ \midrule
Bidirectional LSTM &(256) & 62 & 69.3 & 72.6 & 77.8 & 83 & 87.5\\ \midrule
Bidirectional Stacked LSTM & (128,128) & 83.5 & 87.8 & 90.6 & 93.2 & 94.2 & 96.2 \\ \midrule
LSTM with Attention & (512) & 62.2 & 79.4 & 85.6 & 89.5 & 92.8 & 95.6\\ \midrule
Stacked LSTM with Attention &(256,256) & 86 & 90.3 & 91.7 & 93.6 & 95.3 & 97.2 \\ \midrule
Bidirectional LSTM with Attention &(256) & 66.5 & 81.5 & 86.4 & 91.5 & 94.3 & 96.4 \\ \midrule
Bidirectional Stacked LSTM with Attention & (128,128) & {\bf 86.2} & {\bf 91} & {\bf 93.5} & {\bf 95.5} & {\bf 96.8} & {\bf 98.7}\\ \bottomrule \\
\end{tabular}
\caption{Approximations of the percentage of alerts caught by each model for each percentage allowed to be reviewed. \label{Results}}
\end{table}
To evaluate our models, we did a 5-fold validation on a withheld set of 1000 alerts. That is to say we split our set into 5 partitions of 200 alerts, each of which was used as a validation sample for a neural network trained on all remaining data. This produced five very similar models whose performance is given by the percentage of 1000 alerts that were flagged. The percentage of 1000 alerts flagged was computed for each level of sensitivity considered, as measured by the percentage of the total population flagged for potentially being an alert.
Each of the models had 512 recurrent units (the attention mechanisms were not recurrent), hence, in stacking and using bidirectional variants, the number of units were halved. We predominantly trained on using Keras with Tensorflow serving the back-end. The machines we used had NVIDIA Tesla K80s. Each epoch took approximately two to three hours, however, the rate of convergence was such that we could restrict our attention to the models formed in the first 20 epochs as it was clear that the metrics we assessed had converged fairly quickly given the volume of data we had. The total amount of GPU time spent on developing these models was in excess of 4000 hours.
To give an approximation of the effect of each of the attributes we endowed our models with, we can average over the effectiveness of each model with and without each attribute in question. It is clear that that stacking two layers of recurrent units, each with half as many cells, offers the greatest boost in effectiveness, followed by the difference in recurrent structures followed by the use of attention. Using bidirectional units seems to give the smallest increase, but given the circumstances, any positive increase could potentially save lives.
\begin{table}[!ht]
\centering
\begin{tabular}{l|c}\toprule
{\bf Attribute} & Effect \\\toprule
LSTM vs GRU & 8.05\% increase \\
Bidirectional vs Normal & 2.20\% increase \\
Attention vs No Attention & 4.15\% increase \\
Stacked vs Flat & 14.98 \% increase \\ \bottomrule
\end{tabular}
\caption{The effect of each of the attributes we endowed our networks.}
\label{attributes}
\end{table}
\section{Conclusions}\label{Conclusion}
The problem of depression and violence in our schools is one that has recently garnered high levels of media attention. This type of problem is not confined to the scope of educational research, but this type of anomaly detection is also applicable to social media platforms where there are posts that indicate potential cases of users alluding to suicide, depression, using hate-speech and engaging in cyberbullying. The program on which this study concerns is in place and has contributed to the detection an intervention of cases of depression and violence across America. This study itself has led to a dramatic increase in our ability to detect such cases.
We should also mention that the above results do not represent the state-of-the-art, since we were able to take simple aggregated results from the models to produce better statistics at each threshold level than our best model. This can be done in a similar manner to the work of \cite{Ensemble}, however, this is a topic we leave for a future paper. It is also unclear as to whether traditional sentiment analysis provides additional information from which better estimates may be possible.
\section{Acknowledgements}
I would like to thank Jon Cohen, Amy Burkhardt, Balaji Kodeswaran, Sue Lottridge and Paul van Wamelen for their support and discussions.
|
\section{Introduction}
\label{sec:Intro}
Humans prefer to form groups and act collectively. These groups have evolved from simple settlements in ancient times to huge nations in modern times; defined by multiple causes like languages, common heritage, geographical boundaries, and even ideologies. Often human cooperation has been the motivating force behind the rapid progress of man. This cooperation \cite{Perc_2017} has extended from blood relatives to totally unrelated individuals. Contrarily, evolution has been responsible for drawing distinctions among themselves in their bids for the ``survival of the fittest''. The segregation \cite{Schelling_1969,Schelling_1971} can be seen in various forms of race, caste, class, religion, political ideology, etc. The assortment of positive and negative aspects of human social behavior makes it extremely complex and convoluted with multiple parameters playing crucial roles. Thus, it is extremely difficult to assess and model the complexity of human social behavior, ranging from bonding, co-operation, support to greed, jealousy, conflict, aggression, coup, war, etc.
Entire world has seen time and again different forms of conflicts, aggression, war, and terrorism, which have plagued mankind from antiquity. Anti-social phenomenon notably possesses very different characteristics than normal social behavior. The interactions among anti-social agents are very low and the occurrences of events tend to be independent of each other.
A conflict is an activity which takes place between conscious (not necessarily rational) beings when their interests are mutually inconsistent with each other. A conflict is usually associated with violent activities. The human society has been riddled with conflicts. The first known conflict, a case of inter-group violence, was in eastern Africa around 10,000 years ago as an attempt to seize resources - territory, women, food stored in pots, which resulted in the killing of over two dozen prehistoric men, women, and children \cite{Lahr_2016}.
However, as there has been more progress in civilization, humans have become more materialistic and self-centered, and gone beyond competition for tangible resources; they have adopted causes like religion, racial superiority, etc., as pretexts for killing others.
Many people, including Karl Marx and Friedrich Engels, have proposed theories of social conflicts.
Apart from social scientists, physicists and data scientists have recently tried to perform in-depth studies and provide mathematical models, statistical
and time series analysis of the empirical data and tried to propose potential solutions to the menaces of terrorism, conflicts and other social phenomena, leading to the development of the field of sociophysics \cite{Castellano_2009,Chakrabarti_2006,Sen_2014,Abergel_2017}. Sociophysics is marked by the belief that large-scale statistical measurement of social variables reveals underlying relational patterns that can be explained by theories and laws found in natural sciences, and physics in particular.
In this chapter, we focus on the data dependent statistical analyses of three major anti-social phenomena, viz., ethnic conflicts (EC), human right violations (HR), and terrorism (GTD) \cite{Sharma_2017_a,Clauset_2007,Husain_2018}.
An ethnic conflict is a conflict between two or more contending ethnic groups where each group fights for its position within the society on the basis of ethnicity, derived from common descent, culture, language and sometimes, even a common identity.
Similarly, a human right violation is said to occur when the basic fundamental rights of a person or a group of persons are infringed upon. Both these anti-social phenomena are based on the conflicts between one or more contesting parties (two sets of actors). However, terrorism differs from these conflicts in the sense that the casualties occurring in a terrorist event are direct or indirect targets of the terrorist groups (sources). The aim of terrorism is not limited to eliminating the target group or destruction of resources, rather it is specifically carried out to send out a psychological message to the adversary \cite{Richardson_2013}. In other words, unlike in ethnic conflicts and human rights violations, the terrorist attacks are carried out to send across a message to the opponent \cite{Cutter_2014}.
Here, we use the publicly available data from: (a) GDELT database \cite{GDELT, GDELTcodebook}, which has news reports in media consisting of records of a wide range of socio-economic and political events, viz. ethnic conflicts and human rights violations, over a long period of time, and (b) GTD project \cite{GTDcloud, GTDcodebook}, which has recorded the terrorist attack incidents that occurred in the last half-century across the globe. We aggregate these anti-social events over time, and study the temporal evolution of these events. We present here the results of several analyses like recurrence intervals, Hurst R/S analysis, etc., that reveal the long memory of these time series \cite{Tilak_2012}. Further, we filter the data country-wise, and study the correlations of these anti-social events within the individual countries. Using the multi-dimensional scaling, we cluster the countries together in terms of the co-movements with respect to temporal growths of these anti-social events. The time series of these events reveal interesting statistical regularities and correlations.
The article is organized as follows. Section~\ref{sec_data_method_results} describes the data description, methodology and results in detail. Section~\ref{sec:conclusion} contains the concluding remarks.
\section{Data description, Methodology and Results}
\label{sec_data_method_results}
\subsection{Data description}
We have used the Global Database of Events, Language, and Tone (GDELT)~\cite{GDELT, GDELTcodebook} which is an open source database hosted and managed by GDELT project through \textit{Google Cloud}. GDELT monitors the world's news media from nearly every corner of every country in print, broadcast, and web formats, in over 100 languages, every moment of every day. The GDELT project is a real-time open database, where the human society is seen through the eyes of the world's news media, reaching deeply into local events, reaction, discourse, and emotions of the most remote corners of the world. The entire GDELT event database is available and can be extracted using \textit{Google BigQuery}. We filtered all events related to ethnic conflicts (EC) and human rights violations (HR) happening around the world spanning over a large time scale. We procured $45,942$ events for EC and $48,295$ for HR for a 15 year period, 2001-2015.
We have also analysed the data on terrorism events. For the analysis we have used the Global Terrorism Database (GTD) which is an open-source database provides a detailed account of terrorist events around the world from 1970-2017~\cite{GTDcloud, GTDcodebook}. The event database is hosted by the National Consortium for the Study of Terrorism and Responses to Terrorism (START), University of Maryland. We procured $72,521$ events for the same 15 year period, 2001-2015.
The list of all the countries analysed, containing names along with their three letter ISO codes, is given in Table \ref{table_country}.
\subsection{Methodology and Results}
\label{sec_methods}
The GDELT and GTD data sets contain detailed information about the anti-social events, viz. ethnic conflicts (EC), human rights violations (HR) and terrorist attacks (GTD), on the scale of a day. Our overarching aim is to observe nature of the memory of each of the time-series (EC, HR and GTD) and the cross-correlations among them, within a country. Further, we would like to group the countries together on the basis of their long-term evolution trends and correlations. First, we study simple statistics of auto-correlations, Hurst R/S analysis and recurrence intervals distribution, of the detrended time series. Later, we study the co-movements of the countries on the events spaces using the multidimensional scaling technique.
We have considered the data for the period 2001-2015, and generated daily time series of EC, HR and GTD, as shown in Fig.~\ref{fig:events} (a). The black curves show the long time trends, which imply that the time-series are not stationary. To see the spread of the events, we computed the complementary cumulative density function (CCDF) of the events: For the probability density function (PDF) $P(n)$ as $n$ reported events per day, the cumulative density function (CDF) is $F(n) = P(N\leq n)$; then, the CCDF is $Q(n) = 1-F(n)$, such that it estimates the probability of the events are above a particular level $n$, $P(N>n)$. As the empirical PDFs are often too noisy (specially toward the tails) to be relied upon for statistics, it is known that integrating a signal improves its ``Signal-to-Noise ratio''. So, we plot the CCDF as it reduces the noise content and makes the information contained by the signal clearer.
\begin{figure}[!h]
\includegraphics[width=0.49\textwidth]{Fig1a}
\llap{\parbox[b]{2.4in}{(\textbf{a})\\\rule{0ex}{1.6in}}}
\includegraphics[width=0.47\textwidth]{Fig1b}
\llap{\parbox[b]{2.3in}{(\textbf{b})\\\rule{0ex}{1.6in}}}
\caption{\textbf{(a)} Plots of the time evolution of the number of events $n$ reported daily for EC, HR and GTD during the period 2001-2015 with trends (black solid curves).
\textbf{(b)} The complementary cumulative density function (CCDF) $Q(n$) that $n$ or more events are reported on a particular day. The data seems to fit well to a stretched exponential of the form $(\exp[-an^b])$ for EC (red circles), HR (green diamonds) and GTD (blue squares), with exponents given in Table~\ref{table_CCDF}.}
\label{fig:events}
\end{figure}
As the number of news entries $n$ per day is a stochastic variable, we often see bursts of activities for all the three anti-social phenomena: EC, HR and GTD. Due to large inter-day fluctuations in the number of reports and the bursty nature, the CCDF shows a broad distribution. Fig.~\ref{fig:events} (b) shows the plots for the CCDF $Q(n$) that $n$ or more events are reported on a particular day, for the three time-series. Each of the curves is well-fitted by a stretched exponential of the form, $(\exp[-an^b])$ with exponents given in Table~\ref{table_CCDF}.
\begin{table}[!h]
\centering
\captionsetup{justification=centering}
\caption{Exponent values for events of EC, HR and GTD}
\begin{tabular}{|l|l|l|l|}
\hline
\backslashbox{Series}{Exponents} & \textbf{a} & \textbf{b} \\ \hline\hline
\textbf{EC} & 0.50$\pm$ 0.01 & 2.49 $\pm$ 0.01 \\ \hline
\textbf{HR} & 0.76 $\pm$ 0.01 & 2.91 $\pm$ 0.01 \\ \hline
\textbf{GTD} & 0.48 $\pm$ 0.01 & 3.17 $\pm$ 0.01 \\ \hline
\end{tabular}
\label{table_CCDF}
\end{table}
The auto-correlation is the correlation of a signal with a time-delayed copy of itself, as a function of delay or lag. In simple words, it is the similarity between observations as a function of the time lag between them, which can be used for finding repeating patterns or periodicity obscured by noise.
The Hurst exponent is a popular measure of long-term memory in a time series, which relates to the auto-correlations of the same and the rate at which these auto-correlations decrease as the time-lag between the pair of values increases.
Extreme events are rare in natural as well as social phenomena, but it is essential to study their properties as the consequences of extreme events are often enormous \cite{Santhanam_2008,Chicheportiche_2014,Chicheportiche_2017}. As researchers, we are often interested in the question that how long would we have to wait for extreme events of a certain magnitude to recur. We thus fix a threshold $X^{(q)}$ and consider only the events of magnitude higher than $X^{(q)}$, where $q$ denotes the quantile.
We define the \emph{recurrence interval} as the time interval between two consecutive extreme events:
\begin{equation}
R_t= \begin{cases}\text{NA}&,X(t)<X^{(q)}\\\inf\left\{\tau>0\mid X(t+\tau)\geqX^{(q)}\right\}&,X(t)\geqX^{(q)}\end{cases},
\label{eq_reccur}
\end{equation}
where $X(t)$ is an event occurring at time $t$, $X^{(q)}$ is the threshold, and $\tau$ is a time lag.
A real data series usually exhibit non-stationarity of various forms such as ``seasonal
effects", ``trends", etc. Though it is very difficult to completely eliminate non-stationarity, its effect can be reduced by introducing some corrective measures. Each type of non-stationarity requires a different type of correction. Here, none of the original time series of events is stationary, as can be seen in Fig.~\ref{fig:events}, where the black dashed line shows the inherent trend of the time series. We calculate the trend with a polynomial of degree 10, and then divide the original signal by the computed trend, resulting in a detrended series. Fig.~\ref{fig:Reccur} (a) shows the events time series after detrending it. The black dashed line, which shows trend of this detrended time series, is thus flat. The detrended events time series can be assumed to be weakly stationary.
\begin{figure}
\includegraphics[width=0.65\textwidth]{Fig2a}
\llap{\parbox[b]{3.1in}{(\textbf{a})\\\rule{0ex}{2.2in}}}
\hspace{-0.7cm}
\includegraphics[width=0.33\textwidth]{Fig2b}
\llap{\parbox[b]{1.6in}{(\textbf{b})\\\rule{0ex}{2.1in}}}
\caption{\textbf{(a)} Detrended time series for EC, HR and GTD with a flat trend (black solid line). \textbf{(b)} shows the plots of complementary cumulative density function $Q(r)$ that $r$ events recurred at quantiles $q=0.7$ and $q=0.9$. The data for quantiles $q=0.7$ and $q=0.9$ have been found to fit well to stretched exponentials, $(a\exp[-bn^c])$ with parameters given in Tables~\ref{table_recur07} and \ref{table_recur09}.}
\label{fig:Reccur}
\end{figure}
\begin{table}[!h]
\centering
\captionsetup{justification=centering}
\caption{Parameter values for recurrence CCDF at $q=0.7$}
\begin{tabular}{|l|l|l|l|}
\hline
\backslashbox{Series}{Parameters} & \textbf{a} & \textbf{b} & \textbf{c} \\ \hline\hline
\textbf{EC} & 3.53$\pm$ 0.37 & 1.27$\pm$ 0.10 & 0.49 $\pm$ 0.03 \\ \hline
\textbf{HR} & 3.69 $\pm$ 0.46 & 1.31 $\pm$ 0.12 & 0.49 $\pm$ 0.03 \\ \hline
\textbf{GTD} & 6.95 $\pm$ 0.71 & 1.94 $\pm$ 0.10 & 0.38 $\pm$ 0.01 \\ \hline
\end{tabular}
\label{table_recur07}
\end{table}
\begin{table}[!h]
\centering
\captionsetup{justification=centering}
\caption{Parameter values for recurrence CCDF at $q=0.9$}
\begin{tabular}{|l|l|l|l|}
\hline
\backslashbox{Series}{Parameters} & \textbf{a} & \textbf{b} & \textbf{c} \\ \hline\hline
\textbf{EC} & 1.67$\pm$ 0.09 & 0.54$\pm$ 0.04 & 0.50 $\pm$ 0.02 \\ \hline
\textbf{HR} & 1.69 $\pm$ 0.10 & 0.56 $\pm$ 0.05 & 0.51 $\pm$ 0.02 \\ \hline
\textbf{GTD} & 4.72 $\pm$ 0.72 & 1.58 $\pm$ 0.15 & 0.28 $\pm$ 0.02 \\ \hline
\end{tabular}
\label{table_recur09}
\end{table}
We hence analyse the CCDF of the recurrence time intervals on the detrended time series (see Fig.~\ref{fig:Reccur} (a)) to quantify the extreme events or duration of recurrences of an event.
Fig.~\ref{fig:Reccur} (b) shows the plots of CCDF $Q(r)$ that $r$ events recurred at quantiles $q=0.7$ and $q=0.9$. The data for quantiles $q=0.7$ and $q=0.9$ seem to fit well to stretched exponentials, $(a\exp[-bn^c])$, with parameters given in Tables~\ref{table_recur07} and \ref{table_recur09}. It should be mentioned that as the quantile $q$ increases, the distribution becomes fatter, i.e., lower recurrence time intervals occur less frequently. This observation is quite obviously explained by the fact that at higher
values of $q$ there are fewer extreme events and they are spread apart.
It is often not possible to comprehend certain effects using empirical data. Thus, the results obtained by analyses of empirical data generally need to be compared against standard benchmarks.
In such situations, artificial data can be simulated according to required specifications and the simulated data can then serve as reliable benchmarks. Therefore, we first use Gaussian noises (white and fractional) to understand certain effects and use them as benchmarks for comparing the empirical statistics.
Gaussian noise is a statistical noise having a probability density function equal to that of the Normal (or Gaussian) distribution; a special case is the white Gaussian noise (wGn) or Brownian motion, in which the increments (values at any pair of times) are identically distributed and statistically independent (and hence uncorrelated).
Thus, it has no auto-correlation for positive lags, and an exponentially decreasing recurrence interval distribution. We illustrate a white Gaussian noise in Fig.~\ref{fig:benchmark} (a).
A fractional Brownian motion is a generalization of Brownian motion. The main difference between fractional Brownian motion and regular Brownian motion is that the increments in Brownian motion are independent, whereas increments for fractional Brownian motion are not.
A fractional Gaussian noise (fGn) with Hurst exponent $0 \leq H \leq 1$, is defined as a continuous-time Gaussian process $B_H(t)$ on $[0, T]$, which starts at zero, has expectation zero for all $t$ in $[0, T]$, and has a co-variance function. Mathematically,
\begin{align}
&\forall (t,s)\in \mathbb{R}_+^2, \nonumber\\
&\mathbb{E}[B_H(t)]=0\\
&\mathbb{E}[B_H(t)B_H(s)]=\frac{|t|^{2H}+|s|^{2H}-|t-s|^{2H}}{2}. \label{eq:brownian}
\end{align}
The auto-correlation function (ACF) of a fractional Gaussian noise with Hurst exponent $H$ is given by:
\begin{equation}
ACF(\tau) \rightarrow \dfrac{\mid \tau + 1 \mid ^{2H} + \mid \tau - 1 \mid ^{2H} - 2\mid \tau \mid ^{2H}}{2} .
\label{autocorr_FGN}
\end{equation}
For a stationary process with auto-correlations decaying $ACF(\tau) \sim \tau^{-\gamma}$ (long-memory processes), it can be shown mathematically $\gamma = 2-2H$ \cite{Tarnopolski_2016}.
Fig.~\ref{fig:benchmark} (b) shows the ACF and PDF of recurrence time intervals for fractional Gaussian noise with Hurst exponent $H = 0.8$. If the underlying time series has auto-correlation, then the extreme events are auto-correlated as well. Evidently the presence of the auto-correlation renders the probability density function of the recurrence time intervals to be a stretched exponential, instead of a pure exponential as observed in the case of white Gaussian noise.
\begin{figure}
\includegraphics[width=0.95\linewidth]{Fig3a}
\llap{\parbox[b]{4.6in}{(\textbf{a})\\\rule{0ex}{2.2in}}}
\includegraphics[width=0.95\linewidth]{Fig3b}
\llap{\parbox[b]{4.4in}{(\textbf{b})\\\rule{0ex}{2.2in}}}
\caption{\textbf{(a)} Time series for white Gaussian noise, probability density function for white noise (independent and identically distributed variables), auto-correlation function of time series, probability density function of recurrence time intervals at $q = 0.75$. The time series was generated using the \textit{rnorm()} function in R-software for statistical computing. \textbf{(b)} Time series for fractional Gaussian noise with Hurst index $H=0.8$, probability density function of the fractional Gaussian noise (dependent and identically distributed variables), auto-correlation function of the time series, probability density function for recurrence intervals at $q = 0.75$. The time series was generated using the \textit{simFGN0()} function in R-software for statistical computing. }
\label{fig:benchmark}
\end{figure}
The Hurst exponent is a useful statistical method for inferring the properties of a time series.
There are various methods to calculate Hurst exponent, which measures the existence of trend or `persistence' or long-range memory present in the time series. We used the rescaled range (R/S) method to compute the Hurst exponent~\cite{Chakraborti_2005}.
The rescaled range (R/S) method is calculated for a time series, $X_{1},X_{2},\dots ,X_{T}$, as follows~\cite{torresexponente}:
We first break the long time series with $T$ data points, into shorter windows of $n$ data points, such that there are $m=T/n$ windows. For each of the $m$ windows of size $n$, we have the partial time series $X_{1},X_{2},\dots ,X_{n}$, for which we calculate the rescaled range:
\begin{enumerate}
\item Calculate the mean $\mu={\frac {1}{n}}\sum _{{i=1}}^{{n}}X_{i}$
\item Create a mean adjusted series $Y_{i}=X_{{i}}-\mu{\text{ for }}i=1,2,\dots ,n\,$
\item Calculate the cumulative deviate series $Z_{t}=\sum _{{i=1}}^{{t}}Y_{{i}}{\text{ for }}t=1,2,\dots ,n$
\item Compute the range $R(n)=\max \left(Z_{1},Z_{2},\dots ,Z_{n}\right)-\min \left(Z_{1},Z_{2},\dots ,Z_{n}\right)$
\item Compute the standard deviation $S(n)={\sqrt {{\frac {1}{n}}\sum _{{i=1}}^{{n}}\left(X_{{i}}-\mu \right)^{{2}}}}$, where $\mu$ is the mean for the partial time series $X_{1},X_{2},\dots ,X_{n}$.
\item Calculate the rescaled range $ R(n)/S(n) $ and average over all the partial time series of length $n$.
\end{enumerate}
The Hurst exponent is estimated by fitting the power law: $\mathbb{E} [R(n)/S(n)]=Cn^{H}$ to the empirical data, where $C$ is a constant. This can be done by plotting $\log[R(n)/S(n)]$ as a function of $ \log n$, and fitting a straight line. The value of the slope gives the Hurst exponent $H$, such that
\begin{itemize}
\item A value in the range $0\leq H<0.5$ indicates a time series with `anti-persistent' behavior,
\item a value in the range $0.5<H\leq 1$ indicates a time series with long-term positive auto-correlation (`persistent' behavior),
\item a value of $H=0$ indicates a pink noise,
\item a value of $H=0.5$ indicates a completely uncorrelated series (Brownian motion).
\end{itemize}
\begin{figure}[!h]
\includegraphics[width=0.95\textwidth]{Fig4a}
\llap{\parbox[b]{4.6in}{(\textbf{a})\\\rule{0ex}{1.6in}}}\\
\includegraphics[width=0.95\textwidth]{Fig4b}
\llap{\parbox[b]{4.6in}{(\textbf{b})\\\rule{0ex}{1.6in}}}\\
\includegraphics[width=0.95\textwidth]{Fig4c}
\llap{\parbox[b]{4.6in}{(\textbf{c})\\\rule{0ex}{1.6in}}}
\caption{Plot for the auto-correlation of detrended time series and Hurst exponent based on R/S analysis having exponent for \textbf{(a)} EC $0.75 \pm 0.01$, \textbf{(b)} HR $0.78 \pm 0.01$ and \textbf{(c)} GTD $0.82 \pm 0.01$. As the value of the exponent is greater than $0.5$ so the time series shows the persistence behavior for all EC, HR and GTD. }
\label{fig:acf_dfa}
\end{figure}
Fig.~\ref{fig:acf_dfa} shows the auto-correlation of the detrended time series and Hurst exponent based on R/S analysis having exponent for (a) EC $0.75 \pm 0.01$, (b) HR $0.78 \pm 0.01$ and (c) GTD $0.82 \pm 0.01$. The auto-correlation for GTD is decaying exponentially. As the value of the exponent is greater than $0.5$ so the time series shows the persistence behavior for all EC, HR and GTD.
Next, we study the co-movements among the different countries across the globe. Fig.~\ref{fig:Corr_MDS} (a) shows the time series of $n$ events (EC, HR and GTD) during the period of 2001-2015 for a few countries chosen arbitrarily. We take $N$ countries and aggregate the events over a year, producing $T=15$ data points for the period $2001-2015$.
To build the correlation matrices, we define the equal-time Pearson cross-correlation coefficient for the time series of the number of events per year $c_i$ as
\begin{equation}
\rho_{ij}(\tau) = \frac{\langle c_i c_j \rangle - \langle c_i \rangle \langle c_j \rangle}{\sigma_i\sigma_j}.
\label{Eq_corr}
\end{equation}
where $\sigma_i= \sqrt{\langle c_i^2 \rangle -\langle c_i\rangle^2}$ is the standard deviation of $c_i$, $i,j=1, \dots, N$, and $\langle \dots \rangle $ denotes average over the time period $\tau$. The elements $\rho_{ij}$ are restricted to the domain $ -1\leq \rho_{ij}\leq1$, where $\rho_{ij}=1$ signifies perfect correlations, $\rho_{ij}=-1$ perfect anti-correlations, and $\rho_{ij}=0$ corresponds to uncorrelated pairs.
It is difficult to estimate the exact correlation among $N$ time series, each of length $T$, as spurious correlations or `noise' are present in the finite time series (see Ref.~\cite{Pharasi_2018b}). The quality of the estimation of true correlation in a matrix strongly depends upon the ratio of the length of the time series $T$ and the number of time series $N$, $Q = T/N$. Correlation matrices are less noisy for higher value of $Q $.
As $N > T$, the corresponding cross-correlation matrices are also singular with $N-T+1$ zero eigenvalues, which leads to poor eigenvalue statistics. Thus, we use the power map technique \cite{Chakraborti_2018, Pharasi_2018a,Pharasi_2018b} to break the degeneracy of eigenvalues at zero and suppress the noise. In this method, a non-linear distortion is given to each cross-correlation coefficient $(\rho_{ij})$ of the correlation matrix $\boldsymbol \rho$ by: $\rho_{ij} \rightarrow (\mathrm{sign} ~~\rho_{ij}) |\rho_{ij}|^{1+\epsilon}$, where $\epsilon$ is the distortion parameter; here we used $\epsilon=0.6$ (see Refs.~\cite{Pharasi_2018b,Pharasi_2018a} for choice of the parameter).
Fig.~\ref{fig:Corr_MDS} (b) shows the correlation matrices (after using the power mapping method), computed over different time series across the countries by using Eq.~\ref{Eq_corr}. The correlation matrix for EC shows more correlations (colored red) among the countries as compared to anti-correlations (colored blue). The correlation matrices for HR and GTD look very different. In order
to visualize the correlations, we apply the multidimensional scaling (MDS) technique. First, we transform the correlation matrix $\boldsymbol \rho$ into distance matrix $\textbf{D}$, as
\begin{eqnarray}
d_{ij} = \sqrt{2 (1- \rho_{ij})},
\label{Eq:distance}
\end{eqnarray}
such that $2 \geq d_{ij} \geq 0$.
After transforming the correlation matrix into distance matrix, we generate the MDS map.
The MDS algorithm is used to display the structure of similarity in terms of distances, as a geometrical map where each country corresponds to a set of coordinates in
the multidimensional space. MDS arranges different countries in this space according to the strength of the pairwise distances between them. Two similarly behaving countries are represented by two points that are close to each other, and two dissimilarly behaving countries are placed far apart in the map. In general, we choose the embedding dimension to be 2, so that we are able to plot the coordinates in the form of a map. It may be noted that coordinates are not necessarily unique, as we can arbitrarily translate and rotate them, as long as such transformations leave the distances unaffected.
Fig.~\ref{fig:Corr_MDS} (c) shows the 2D MDS plots for EC, HR and GTD based on the similarities/ distances among them.
\begin{figure}[!h]
\includegraphics[width=0.95\linewidth]{Fig5a}
\llap{\parbox[b]{4.7in}{(\textbf{a})\\\rule{0ex}{0.7in}}}
\includegraphics[width=0.95\linewidth]{Fig5b}
\llap{\parbox[b]{4.7in}{(\textbf{b})\\\rule{0ex}{1.2in}}}
\includegraphics[width=0.98\linewidth]{Fig5c}
\llap{\parbox[b]{4.5in}{(\textbf{c})\\\rule{0ex}{1.2in}}}
\caption{\textbf{(a)} Time series plots for number of events $n$ of different countries for EC, HR and GTD during the period of 2001-15. The list of country names can be seen in Table.~\ref{table_country}. \textbf{(b)} Correlation matrices for EC, HR and GTD. As number of countries $N$ are more than length of time series $T$, i.e. $N>>T$, the power mapping technique with a distortion $\epsilon=0.6$ is applied to the correlation matrix to suppress the noise. \textbf{(c)} 2D MDS plots for EC, HR and GTD during the period 2001-2015. The MDS plots show the co-movement of the countries: similar countries are grouped together and dissimilar ones placed far apart. }
\label{fig:Corr_MDS}
\end{figure}
At the end, we also calcuted the correlation among the different time series for individual countries. The correlations are computed among EC-HR, EC-GTD and HR-GTD. Few countries like ESP, IDN, ITA and RUS have low correlations among EC-HR, whereas ESP, FRA, ITA, LKA and RUS show anti-correlations for EC-GTD; countries like ESP, FRA, GRC, IDN, ITA, LKA show anti-correlations for HR-GTD. For further details, see Table.~\ref{table_corr}. It must be noted that these are just linear correlations, and causal relations cannot be inferred.
\begin{table}[!h]
\caption{Cross-correlation among events across countries.}
\begin{tabular}{||l||l|l|l|l||l||l|l|l|l|}
\hline
\textbf{S.No.} & \textbf{Country} & \textbf{EC-HR} & \textbf{EC-GTD} & \textbf{HR-GTD} & \textbf{S.No.} & \textbf{Country} & \textbf{EC-HR} & \textbf{EC-GTD} & \textbf{HR-GTD} \\ \hline \hline
1 & \textbf{AFG } & 0.72 & 0.80 & 0.84 & 11 &\textbf{ IRQ } & 0.61 & 0.73 & 0.85 \\ \hline
2 & \textbf{CHN } & 0.81 & 0.64 & 0.89 & 12 & \textbf{ISR } & 0.90 & 0.43 & 0.33 \\ \hline
3 & \textbf{DEU} & 0.65 & 0.95 & 0.49 & 13 & \textbf{ITA } & 0.21 & -0.05 & -0.09 \\ \hline
4 & \textbf{ESP} & 0.47 & -0.52 & -0.28 & 14 & \textbf{LKA } & 0.57 & -0.22 & -0.39 \\ \hline
5 & \textbf{FRA} & 0.91 & -0.08 & -0.10 & 15 & \textbf{NGA} & 0.73 & 0.77 & 0.77 \\ \hline
6 & \textbf{GBR} & 0.81 & 0.72 & 0.84 & 16 & \textbf{PAK} & 0.83 & 0.65 & 0.82 \\ \hline
7 & \textbf{GRC} & 0.91 & 0.18 & -0.01 & 17 & \textbf{RUS} & 0.29 & -0.07 & 0.04 \\ \hline
8 & \textbf{IDN} & 0.42 & 0.69 & -0.10 & 18 & \textbf{TUR} & 0.97 & 0.95 & 0.90 \\ \hline
9 & \textbf{IND} & 0.94 & 0.85 & 0.89 & 19 & \textbf{USA} & 0.61 & 0.38 & 0.11 \\ \hline
10 & \textbf{IRN} & 0.80 & 0.11 & 0.45 & 20 & \textbf{ZAF} & 0.96 & 0.21 & 0.17 \\ \hline
\end{tabular}
\label{table_corr}
\end{table}
\begin{table}[]
\caption{List of countries and their 3-letter ISO codes.}
\begin{tabular}{||l||l|l||l||l|l||l||l|l|}
\hline
\textbf{S.No.} & \textbf{Code} & \textbf{Country} & \textbf{S.No.} & \textbf{Code} & \textbf{Country} & \textbf{S.No.} & \textbf{Code} & \textbf{Country} \\ \hline \hline
1 & \textbf{AFG} & Afghanistan & 36 & \textbf{GEO} & Georgia & 71 & \textbf{NLD} & Netherlands \\ \hline
2 & \textbf{AGO} & Angola & 37 & \textbf{GHA} & Ghana & 72 & \textbf{NOR} & Norway \\ \hline
3 & \textbf{ALB} & Albania & 38 & \textbf{GMB} & Gambia & 73 & \textbf{NPL} & Nepal \\ \hline
4 & \textbf{ARG} & Argentina & 39 & \textbf{GRC} & Greece & 74 & \textbf{PAK} & Pakistan \\ \hline
5 & \textbf{ARM} & Armenia & 40 & \textbf{HKG} & Hong Kong & 75 & \textbf{PER} & Peru \\ \hline
6 & \textbf{AUS} & Australia & 41 & \textbf{HRV} & Croatia & 76 & \textbf{PHL} & Philippines \\ \hline
7 & \textbf{AZE} & Azerbaijan & 42 & \textbf{HTI} & Haiti & 77 & \textbf{POL} & Poland \\ \hline
8 & \textbf{BDI} & Burundi & 43 & \textbf{HUN} & Hungary & 78 & \textbf{PRK} & North Korea \\ \hline
9 & \textbf{BEL} & Belgium & 44 & \textbf{IDN} & Indonesia & 79 & \textbf{RUS} & Russia \\ \hline
10 & \textbf{BGD} & Bangladesh & 45 & \textbf{IND} & India & 80 & \textbf{RWA} & Rwanda \\ \hline
11 & \textbf{BGR} & Bulgaria & 46 & \textbf{IRL} & Ireland & 81 & \textbf{SAU} & Saudi Arabia \\ \hline
12 & \textbf{BIH} & Bosnia-Herzegovina & 47 & \textbf{IRN} & Iran & 82 & \textbf{SDN} & Sudan \\ \hline
13 & \textbf{BLR} & Belarus & 48 & \textbf{IRQ} & Iraq & 83 & \textbf{SEN} & Senegal \\ \hline
14 & \textbf{BRA} & Brazil & 49 & \textbf{ISR} & Israel & 84 & \textbf{SLE} & Sierra Leone \\ \hline
15 & \textbf{BTN} & Bhutan & 50 & \textbf{ITA} & Italy & 85 & \textbf{SLV} & El Salvador \\ \hline
16 & \textbf{CAN} & Canada & 51 & \textbf{JOR} & Jordan & 86 & \textbf{SOM} & Somalia \\ \hline
17 & \textbf{CHE} & Switzerland & 52 & \textbf{JPN} & Japan & 87 & \textbf{SRB} & Serbia \\ \hline
18 & \textbf{CHL} & Chile & 53 & \textbf{KEN} & Kenya & 88 & \textbf{SWE} & Sweden \\ \hline
19 & \textbf{CHN} & China & 54 & \textbf{KGZ} & Kyrgyzstan & 89 & \textbf{SYR} & Syria \\ \hline
20 & \textbf{CIV} & Cote D'ivoire & 55 & \textbf{KHM} & Cambodia & 90 & \textbf{TCD} & Chad \\ \hline
21 & \textbf{COG} & Democratic Republic of the Congo & 56 & \textbf{KOR} & South Korea & 91 & \textbf{THA} & Thailand \\ \hline
22 & \textbf{COL} & Colombia & 57 & \textbf{KWT} & Kuwait & 92 & \textbf{TUN} & Tunisia \\ \hline
23 & \textbf{CUB} & Cuba & 58 & \textbf{LBN} & Lebanon & 93 & \textbf{TUR} & Turkey \\ \hline
24 & \textbf{CYP} & Cyprus & 59 & \textbf{LBR} & Liberia & 94 & \textbf{TWN} & Taiwan \\ \hline
25 & \textbf{CZE} & Czech Republic & 60 & \textbf{LBY} & Libya & 95 & \textbf{UGA} & Uganda \\ \hline
26 & \textbf{DEU} & Germany & 61 & \textbf{LKA} & Sri Lanka & 96 & \textbf{UKR} & Ukraine \\ \hline
27 & \textbf{DNK} & Denmark & 62 & \textbf{LVA} & Latvia & 97 & \textbf{USA} & United States \\ \hline
28 & \textbf{DZA} & Algeria & 63 & \textbf{MDA} & Moldova & 98 & \textbf{UZB} & Uzbekistan \\ \hline
29 & \textbf{ECU} & Ecuador & 64 & \textbf{MEX} & Mexico & 99 & \textbf{VEN} & Venezuela \\ \hline
30 & \textbf{EGY} & Egypt & 65 & \textbf{MKD} & Macedonia & 100 & \textbf{VNM} & Vietnam \\ \hline
31 & \textbf{ESP} & Spain & 66 & \textbf{MMR} & Myanmar & 101 & \textbf{YEM} & Yemen \\ \hline
32 & \textbf{ETH} & Ethiopia & 67 & \textbf{MNP} & Northern Mariana Islands & 102 & \textbf{ZAF} & South Africa \\ \hline
33 & \textbf{FJI} & Fiji & 68 & \textbf{MYS} & Malaysia & 103 & \textbf{ZMB} & Zambia \\ \hline
34 & \textbf{FRA} & France & 69 & \textbf{NAM} & Namibia & 104 & \textbf{ZWE} & Zimbabwe \\ \hline
35 & \textbf{GBR} & United Kingdom & 70 & \textbf{NGA} & Nigeria & & \textbf{} & \\ \hline
\end{tabular}
\label{table_country}
\end{table}
\section{Concluding remarks}
\label{sec:conclusion}
In this paper, our goal was to do the time series analysis and apply data science approaches to the study of the daily anti-social events like ethnic conflicts (EC), human right violations (HR) and terrosrist attacks (GTD). As the time series were non-stationary, so we made them stationary by detrending them. We computed the recurrence interval distribution of events and made attempts to relate it with its auto-correlation function. Then we computed the Hurst exponent using the rescaled range (R/S) analyses, which gives the information about whether long memory is present or not. Further, our interest was to study the co-movements of the countries in the respective events spaces. To visualize the co-movements, we computed the cross-correlations among different countries, transformed the correlations into distances and then projected the distances into 2D multidimensional scaling maps.
\begin{acknowledgement}
The authors would like to thank Anirban Chakraborti, Vishwas Kukreti, Arun S. Patel and Hirdesh K. Pharasi for critical discussions and inputs.
KS acknowledges the University Grants Commission (Ministry of Human Resource Development, Govt. of India) for her senior research fellowship.
SSH and KS acknowledge the support by University of Potential Excellence-II grant (Project ID-47) of JNU, New Delhi, and the DST-PURSE grant given to JNU by the Department of Science and Technology, Government of India.
\end{acknowledgement}
\bibliographystyle{spmpsci}
|
\section{Introduction}
\section{Introduction}
In relative homological algebra, the theory of covers and envelopes is fundamental and important.
Let $R$ be a ring and $\Mod R$ the category of left $R$-modules. Given a subcategory of $\Mod R$,
it is always worth studying whether or when it is (pre)covering or (pre)enveloping.
This problem has been studied extensively, see \cite{BR}--\cite{HJ09} and references therein.
Let $R$ be a commutative noetherian ring and $C$ a semidualizing $R$-module,
and let $\mathcal{A}_C(R)$ and $\mathcal{B}_C(R)$ be the Auslander and Bass classes respectively.
By proving that both $\mathcal{A}_C(R)$ and $\mathcal{B}_C(R)$ are Kaplansky classes, Enochs and Holm
got in \cite[Theorems 3.11 and 3.12]{EH} that the pair $(\mathcal{A}_C(R),(\mathcal{A}_C(R))^\bot)$
is a perfect cotorsion pair, $\mathcal{A}_C(R)$ is covering and preenveloping
and $\mathcal{B}_C(R)$ is preenveloping. Holm and J{\o}rgensen introduced the notion of duality pairs
and proved the following remarkable result. Let $R$ be an arbitrary ring, and let $\mathscr{X}$ and
$\mathscr{Y}$ be subcategories of $\Mod R$ and $\Mod R^{op}$ respectively.
When $(\mathscr{X},\mathscr{Y})$ is a duality pair, the following assertions hold true:
(1) If $\mathscr{X}$ is closed under coproducts, then $\mathscr{X}$ is covering;
(2) if $\mathscr{X}$ is closed under products, then $\mathscr{X}$ is preenveloping; and
(3) if $_RR\in\mathscr{X}$ and $\mathscr{X}$ is closed under coproducts and extensions,
then $(\mathscr{X},\mathscr{X}^{\perp})$ is a perfect cotorsion pair
(\cite[Theorem 3.1]{HJ09}). By using it, they generalized the above result of Enochs and Holm to
the category of complexes, and Enochs and Iacob investigated in \cite{EI} the existence of Gorenstein
injective envelopes over commutative noetherian rings.
Let $R$ and $S$ be arbitrary rings and $_RC_S$ a semidualizing bimodule, and let $\mathcal{A}_C(R^{op})$
be the Auslander class in $\Mod R^{op}$ and $\mathcal{B}_C(R)$ the Bass class in $\Mod R$.
Our first main result is the following
\begin{theorem}\label{1.1} {\rm (Theorem \ref{3.3})}
\begin{enumerate}
\item[(1)] Both the pairs
$$(\mathcal{A}_C(R^{op}),\mathcal{B}_C(R))\ and\ (\mathcal{B}_C(R),\mathcal{A}_C(R^{op}))$$
are coproduct-closed and product-closed duality pairs; and furthermore, the former one is perfect.
\item[(2)] $\mathcal{A}_C(R^{op})$ is covering and preenveloping in $\Mod R^{op}$
and $\mathcal{B}_C(R)$ is covering and preenveloping in $\Mod R$.
\end{enumerate}
\end{theorem}
As a consequence of Theorem \ref{1.1}, we get that the pair
$$(\mathcal{A}_C(R^{op}),\mathcal{A}_C(R^{op})^{\bot})$$
is a hereditary perfect cotorsion pair and $\mathcal{A}_C(R^{op})$ is covering and preenveloping in $\Mod R^{op}$,
where $\mathcal{A}_C(R^{op})^{\bot}$ is the right $\Ext$-orthogonal class of $\mathcal{A}_C(R^{op})$ (Corollary \ref{3.4}).
This result was proved in \cite[Theorem 3.11]{EH} when $R$ is a commutative noetherian ring and $_RC_S={_RC_R}$.
By Theorem \ref{1.1} and its symmetric result, we have that $\mathcal{B}_C(R)$ is preenveloping in $\Mod R$
and $\mathcal{A}_C(S)$ is preenveloping in $\Mod S$. Moreover, we prove the following
\begin{theorem}\label{1.2} {\rm (Theorem \ref{3.7}(2))}
If $\mathcal{B}_C(R)$ is enveloping in $\Mod R$, then $\mathcal{A}_C(S)$ is enveloping in $\Mod S$.
\end{theorem}
Then we apply these results and their symmetric results to study the Auslander projective dimension of modules.
We obtain some criteria for computing the Auslander projective dimension of modules in $\Mod S$ (Theorem \ref{4.4}).
Furthermore, we get the following
\begin{theorem}\label{1.3} {\rm (Theorem \ref{4.10})}
If $_RC$ has an ultimately closed projective resolution, then
$$\mathcal{A}_C(S)={{C_S}^{\top}}={^{\bot}\mathcal{I}_C(S)},$$
where ${{C_S}^{\top}}$ is the $\Tor$-orthogonal class of $C_S$ and ${^{\bot}\mathcal{I}_C(S)}$ is
the left $\Ext$-orthogonal class of the subcategory $\mathcal{I}_C(S)$ of $\Mod S$ consisting of $C$-injective modules.
\end{theorem}
As a consequence, we have that if $_RC$ has an ultimately closed projective resolution,
then the projective dimension of $C_S$ is at most $n$ if and only if the Auslander projective dimension of
any module in $\Mod S$ is at most $n$ (Corollary \ref{4.11}).
\section{Preliminaries}
In this paper, all rings are associative with identities. Let $R$ be a ring. We use $\Mod R$ to denote the category
of left $R$-modules and all subcategories of $\Mod R$ are full and closed under isomorphisms.
For a subcategory $\mathscr{X}$ of $\Mod R$, we write
$${^\perp{\mathscr{X}}}:=\{A\in\Mod R\mid\operatorname{Ext}^{\geq 1}_{R}(A,X)=0 \mbox{ for any}\ X\in \mathscr{X}\},$$
$${{\mathscr{X}}^\perp}:=\{A\in\Mod R\mid\operatorname{Ext}^{\geq 1}_{R}(X,A)=0 \mbox{ for any}\ X\in \mathscr{X}\},$$
$${^{\perp_1}{\mathscr{X}}}:=\{A\in\Mod R\mid\operatorname{Ext}^{1}_{R}(A,X)=0 \mbox{ for any}\ X\in \mathscr{X}\},$$
$${{\mathscr{X}}^{\perp_1}}:=\{A\in\Mod R\mid\operatorname{Ext}^{1}_{R}(X,A)=0 \mbox{ for any}\ X\in \mathscr{X}\}.$$
For subcategories $\mathscr{X},\mathscr{Y}$ of $\Mod R$, we write $\mathscr{X}\perp\mathscr{Y}$ if
$\operatorname{Ext}^{\geq 1}_{R}(X,Y)=0$ for any $X\in \mathscr{X}$ and $Y\in \mathscr{Y}$.
\begin{definition} \label{2.1}
{\rm (\cite{E1,EJ00})
Let $\mathscr{X}\subseteq\mathscr{Y}$ be
subcategories of $\Mod R$. A homomorphism $f: X\to Y$ in
$\Mod R$ with $X\in\mathscr{X}$ and $Y\in \mathscr{Y}$ is called an {\bf $\mathscr{X}$-precover} of $Y$
if $\Hom_{R}(X^{'},f)$ is epic for any $X^{'}\in\mathscr{X}$; and $f$ is called {\bf right minimal}
if an endomorphism $h:X\to X$ is an automorphism whenever $f=fh$.
An {\bf $\mathscr{X}$-precover} $f: X\to Y$ is called an {\bf $\mathscr{X}$-cover} of $Y$
if it is right minimal. The subcategory $\mathscr{X}$ is called {\bf (pre)covering}
in $\mathscr{Y}$ if any object in $\mathscr{Y}$ admits an $\mathscr{X}$-(pre)cover.
Dually, the notions of an {\bf $\mathscr{X}$-(pre)envelope}, a {\bf left minimal homomorphism}
and a {\bf (pre)enveloping subcategory} are defined.}
\end{definition}
\begin{definition}\label{2.2}
{\rm (\cite{EJ00, GT12})
Let $\mathscr{U},\mathscr{V}$ be subcategories of $\Mod R$.
\begin{enumerate}
\item[(1)] The pair $(\mathscr{U},\mathscr{V})$ is called a {\bf cotorsion pair} in $\Mod R$ if
$\mathscr{U}={^{\bot_1}\mathscr{V}}$ and $\mathscr{V}={\mathscr{U}^{\bot_1}}$.
\item[(2)] A cotorsion pair $(\mathscr{U},\mathscr{V})$ is called {\bf perfect} if $\mathscr{U}$ is covering and $\mathscr{V}$
is enveloping in $\Mod R$.
\item[(3)] A cotorsion pair $(\mathscr{U},\mathscr{V})$ is called {\bf hereditary} if one of the following equivalent
conditions is satisfied.
\begin{enumerate}
\item[(3.1)] $\mathscr{U}\perp \mathscr{V}$.
\item[(3.2)] $\mathscr{U}$ is projectively resolving in the sense that $\mathscr{U}$ contains all projective modules
in $\Mod R$, $\mathscr{U}$ is closed under extensions and kernels of epimorphisms.
\item[(3.3)] $\mathscr{V}$ is injectively coresolving in the sense that $\mathscr{V}$ contains all injective modules
in $\Mod R$, $\mathscr{V}$ is closed under extensions and cokernels of monomorphisms.
\end{enumerate}
\end{enumerate}}
\end{definition}
Set $(-)^+:=\Hom_{\mathbb{Z}}(-,\mathbb{Q}/\mathbb{Z})$, where $\mathbb{Z}$ is the additive group of integers and $\mathbb{Q}$
is the additive group of rational numbers. The following is the definition of duality pairs (cf. \cite{EI,HJ09}).
\begin{definition}\label{2.3}
{\rm Let $\mathscr{X}$ and $\mathscr{Y}$ be subcategories of $\Mod R$ and $\Mod R^{op}$ respectively.
\begin{enumerate}
\item[(1)] The pair ($\mathscr{X},\mathscr{Y}$) is called a {\bf duality pair} if the following conditions are satisfied.
\begin{enumerate}
\item[(1.1)] For a module $X\in\Mod R$, $X\in\mathscr{X}$ if and only if $X^{+}\in \mathscr{Y}$.
\item[(1.2)] $\mathscr{Y}$ is closed under direct summands and finite direct sums.
\end{enumerate}
\item[(2)] A duality pair ($\mathscr{X},\mathscr{Y}$) is called {\bf (co)product-closed} if $\mathscr{X}$ is closed under
(co)products.
\item[(3)] A duality pair ($\mathscr{X},\mathscr{Y}$) is called {\bf perfect} if it is coproduct-closed,
$_RR\in\mathscr{X}$ and $\mathscr{X}$ is closed under extensions.
\end{enumerate}}
\end{definition}
We also recall the following remarkable result.
\begin{lemma}\label{2.4}
{\rm (\cite[p.7, Theorem]{EI} and \cite[Theorem 3.1]{HJ09})}
Let $\mathscr{X}$ and $\mathscr{Y}$ be subcategories of $\Mod R$ and $\Mod R^{op}$ respectively.
If $(\mathscr{X},\mathscr{Y})$ is a duality pair, then the following assertions hold true.
\begin{enumerate}
\item[(1)] If $(\mathscr{X},\mathscr{Y})$ is coproduct-closed, then $\mathscr{X}$ is covering.
\item[(2)] If $(\mathscr{X},\mathscr{Y})$ is product-closed, then $\mathscr{X}$ is preenveloping.
\item[(3)] If $(\mathscr{X},\mathscr{Y})$ is perfect, then $(\mathscr{X},\mathscr{X}^{\perp})$ is a perfect cotorsion pair.
\end{enumerate}
\end{lemma}
\begin{definition} \label{2.5}
{\rm (\cite{HW07}).
Let $R$ and $S$ be rings. An ($R,S$)-bimodule $_RC_S$ is called
{\bf semidualizing} if the following conditions are satisfied.
\begin{enumerate}
\item[(a1)] $_RC$ admits a degreewise finite $R$-projective resolution.
\item[(a2)] $C_S$ admits a degreewise finite $S$-projective resolution.
\item[(b1)] The homothety map $_RR_R\stackrel{_R\gamma}{\rightarrow} \Hom_{S^{op}}(C,C)$ is an isomorphism.
\item[(b2)] The homothety map $_SS_S\stackrel{\gamma_S}{\rightarrow} \Hom_{R}(C,C)$ is an isomorphism.
\item[(c1)] $\Ext_{R}^{\geq 1}(C,C)=0$.
\item[(c2)] $\Ext_{S^{op}}^{\geq 1}(C,C)=0$.
\end{enumerate}}
\end{definition}
Wakamatsu in \cite{W1} introduced and studied the so-called {\bf generalized tilting modules},
which are usually called {\bf Wakamatsu tilting modules}, see \cite{BR, MR}. Note that
a bimodule $_RC_S$ is semidualizing if and only if it is Wakamatsu tilting (\cite[Corollary 3.2]{W3}).
Examples of semidualizing bimodules are referred to \cite{HW07,W2}.
\section{Duality pairs}
In this section, $R$ and $S$ are arbitrary rings and $_RC_S$ is a semidualizing bimodule.
We write $(-)_*:=\Hom(C,-)$ and
$${{_RC}^{\bot}}:=\{M\in\Mod R\mid\Ext^{\geq 1}_{R}(C,M)=0\}\ \text{and}\
{{C_S}^{\bot}}:=\{B\in\Mod S^{op}\mid\Ext^{\geq 1}_{S^{op}}(C,B)=0\},$$
$${^{\top}{_RC}}:=\{N\in\Mod R^{op}\mid\Tor_{\geq 1}^{R}(N,C)=0\}\ \text{and}\
{{C_S}^{\top}}:=\{A\in\Mod S\mid\Tor_{\geq 1}^{S}(C,A)=0\}.$$
\begin{definition} \label{3.1}
{\rm (\cite{HW07})
\begin{enumerate}
\item[(1)] The {\bf Auslander class} $\mathcal{A}_{C}(R^{op})$ with respect to $C$ consists of all modules $N$
in $\Mod R^{op}$ satisfying the following conditions.
\begin{enumerate}
\item[(a1)] $N\in{^{\top}{_RC}}$.
\item[(a2)] $N\otimes _{R}C\in{{C_S}^{\perp}}$.
\item[(a3)] The canonical valuation homomorphism
$$\mu_N:N\rightarrow (N\otimes_RC)_*$$
defined by $\mu_N(x)(c)=x\otimes c$ for any $x\in N$ and $c\in C$ is an isomorphism in $\Mod R^{op}$.
\end{enumerate}
\item[(2)] The {\bf Bass class} $\mathcal{B}_C(R)$ with respect to $C$ consists of all modules $M$
in $\Mod R$ satisfying the following conditions.
\begin{enumerate}
\item[(b1)] $M\in{_RC^{\perp}}$.
\item[(b2)] $M_*\in{{C_S}^{\top}}$.
\item[(b3)] The canonical valuation homomorphism
$$\theta_M:C\otimes_SM_*\rightarrow M$$
defined by $\theta_M(c\otimes f)=f(c)$ for any $c\in C$ and $f\in M_*$ is an isomorphism in $\Mod R$.
\end{enumerate}
\item[(3)] The {\bf Auslander class} $\mathcal{A}_C(S)$ in $\Mod S$ and the {\bf Bass class} $\mathcal{B}_C(S^{op})$
in $\Mod S^{op}$ are defined symmetrically.
\end{enumerate}}
\end{definition}
The following result is crucial. From its proof, it is known that the conditions in the definitions
of $\mathcal{A}_{C}(R^{op})$ and $\mathcal{B}_C(R)$ are dual item by item.
\begin{proposition}\label{3.2}
\begin{enumerate}
\item[]
\item[(1)] For a module $N\in\Mod R^{op}$, $N\in\mathcal{A}_{C}(R^{op})$ if and only if $N^+\in\mathcal{B}_C(R)$.
\item[(2)] For a module $M\in\Mod R$, $M\in\mathcal{B}_C(R)$ if and only if $M^+\in\mathcal{A}_{C}(R^{op})$.
\end{enumerate}
\end{proposition}
\begin{proof}
(1) Let $N\in \Mod R^{op}$. Then we have the following
(a) \begin{align*}
&\ \ \ \ \ \ \ \ \ \ N\in{^{\top}{_RC}}\\
&\ \ \ \ \ \Leftrightarrow \Tor_{\geq 1}^R(N,C)=0\\
&\ \ \ \ \ \Leftrightarrow [\Tor_{\geq 1}^R(N,C)]^+=0\\
&\ \ \ \ \ \Leftrightarrow \Ext^{\geq 1}_R(C,N^+)=0\ \text{(by \cite[Lemma 2.16(b)]{GT12})}\\
&\ \ \ \ \ \Leftrightarrow N^+\in{_RC^{\perp}}.
\end{align*}
(b) \begin{align*}
&\ \ \ \ \ \ \ \ \ \ N\otimes _{R}C\in{{C_S}^{\perp}}\\
&\ \ \ \ \ \Leftrightarrow \Ext_{S^{op}}^{\geq 1}(C,N\otimes_RC)=0\\
&\ \ \ \ \ \Leftrightarrow [\Ext_{S^{op}}^{\geq 1}(C,N\otimes_RC)]^+=0\\
&\ \ \ \ \ \Leftrightarrow \Tor_{\geq 1}^S(C,(N\otimes_RC)^+)=0\ \text{(by \cite[Lemma 2.16(d)]{GT12})}\\
&\ \ \ \ \ \Leftrightarrow \Tor_{\geq 1}^S(C,(N^+)_*)=0\ \text{(by \cite[Lemma 2.16(a)]{GT12})}\\
&\ \ \ \ \ \Leftrightarrow (N^+)_*\in{{C_S}^{\top}}.
\end{align*}
(c) By \cite[Lemma 2.16(c)]{GT12}, the canonical valuation homomorphism
$$\alpha:C\otimes_S(N\otimes_RC)^+ \to [\Hom_{S^{op}}(C,N\otimes_RC)]^+$$
defined by $\alpha(c\otimes g)(f)=gf(c)$ for any $c\in C$, $g\in (N\otimes_RC)^+$ and $f\in\Hom_{S^{op}}(C,N\otimes_RC)$
is an isomorphism in $\Mod R$. By \cite[Lemma 2.16(a)]{GT12}, the canonical valuation homomorphism
$$\beta:(N\otimes_RC)^+\to \Hom_R(C,N^+)$$
defined by $\beta(g)(c)(x)=g(x\otimes c)$ for any $g\in (N\otimes_RC)^+$, $c\in C$ and $x\in N$
is an isomorphism in $\Mod S$. So $$1_C\otimes\beta:C\otimes_S(N\otimes_RC)^+\to C\otimes_S\Hom_R(C,N^+)$$
via $(1_C\otimes\beta)(c\otimes g)=c\otimes\beta(g)$ for any $c\in C$ and $g\in (N\otimes_RC)^+$ is
an isomorphism in $\Mod R$.
Consider the following diagram
\begin{gather*}
\begin{split}
\xymatrix{
& C\otimes_S(N\otimes_RC)^+ \ar[rr]^{\alpha} \ar [d]^{1_C\otimes\beta} && [\Hom_{S^{op}}(C,N\otimes_RC)]^+ \ar [d]^{(\mu_N)^+} \\
& C\otimes_S\Hom_R(C,N^+) \ar[rr]^{\theta_{N^+}} &&N^+,}
\end{split}
\end{gather*}
where
$$(\mu_N)^+:[\Hom_{S^{op}}(C,N\otimes_RC)]^+\to N^+$$
via $(\mu_N)^+(f^{'})=f^{'}\mu_N$ for any $f^{'}\in [\Hom_{S^{op}}(C,N\otimes_RC)]^+$ is a natural homomorphism in $\Mod R$,
and
$$\theta_{N^+}:C\otimes_S\Hom_R(C,N^+)\to N^+$$
defined by $\theta_{N^+}(c\otimes f^{''})=f^{''}(c)$ for any $c\in C$ and $f^{''}\in\Hom_R(C,N^+)$
is a canonical valuation homomorphism in $\Mod R$. Then for any $c\in C$, $g\in(N\otimes_RC)^+$ and $x\in N$, we have
$$(\mu_N)^+\alpha(c\otimes g)(x)=\alpha(c\otimes g)\mu_N(x)=g\mu_N(x)(c)=g(x\otimes c)$$
$$\theta_{N^+}(1_C\otimes\beta)(c\otimes g)(x)=\theta_{N^+}(c\otimes \beta(g))(x)=\beta(g)(c)(x)=g(x\otimes c),$$
Thus $$(\mu_N)^+\alpha=\theta_{N^+}(1_C\otimes\beta),$$
and therefore $\mu_N$ is an isomorphism $\Leftrightarrow$ $(\mu_N)^+$ is an isomorphism
$\Leftrightarrow$ $\theta_{N^+}$ is an isomorphism.
We conclude that $N\in\mathcal{A}_{C}(R^{op})\Leftrightarrow N^+\in\mathcal{B}_C(R)$.
(2) Let $M\in \Mod R$. Then we have the following
(a) \begin{align*}
&\ \ \ \ \ \ \ \ \ \ M\in{_RC^{\perp}}\\
&\ \ \ \ \ \Leftrightarrow \Ext^{\geq 1}_R(C,M)=0\\
&\ \ \ \ \ \Leftrightarrow [\Ext^{\geq 1}_R(C,M)]^+=0\\
&\ \ \ \ \ \Leftrightarrow \Tor_{\geq 1}^R(M^+,C)=0\ \text{(by \cite[Lemma 2.16(d)]{GT12})}\\
&\ \ \ \ \ \Leftrightarrow M^+\in{^{\top}{_RC}}.
\end{align*}
(b) \begin{align*}
&\ \ \ \ \ \ \ \ \ \ M_*\in{{C_S}^{\top}}\\
&\ \ \ \ \ \Leftrightarrow \Tor_{\geq 1}^S(C,M_*)=0\\
&\ \ \ \ \ \Leftrightarrow [\Tor_{\geq 1}^S(C,M_*)]^+=0\\
&\ \ \ \ \ \Leftrightarrow \Ext_{S^{op}}^{\geq 1}(C,(M_*)^+)=0\ \text{(by \cite[Lemma 2.16(b)]{GT12})}\\
&\ \ \ \ \ \Leftrightarrow \Ext_{S^{op}}^{\geq 1}(C,M^+\otimes_RC)=0\ \text{(by \cite[Lemma 2.16(c)]{GT12})}\\
&\ \ \ \ \ \Leftrightarrow M^+\otimes_RC\in{{C_S}^{\perp}}.
\end{align*}
(c) By \cite[Lemma 2.16(a)]{GT12}, the canonical valuation homomorphism
$$\tau:[C\otimes_S\Hom_R(C,M)]^+\to \Hom_{S^{op}}(C,[\Hom_R(C,M)]^+)$$
defined by $\tau(g^{'})(c)(f)=g^{'}(c\otimes f)$ for any $g^{'}\in [C\otimes_S\Hom_R(C,M)]^+$, $c\in C$ and $f\in\Hom_R(C,M)$
is an isomorphism in $\Mod R^{op}$. By \cite[Lemma 2.16(c)]{GT12}, the canonical valuation homomorphism
$$\sigma:M^+\otimes_RC\to [\Hom_R(C,M)]^+$$
defined by $\sigma(g\otimes c)(f)=gf(c)$ for any $g\in M^+$, $c\in C$ and $f\in\Hom_R(C,M)$
is an isomorphism in $\Mod S^{op}$. So $$\Hom_{S^{op}}(C,\sigma):\Hom_{S^{op}}(C,M^+\otimes_RC)\to \Hom_{S^{op}}(C,[\Hom_R(C,M)]^+)$$
via $\Hom_{S^{op}}(C,\sigma)(g^{''})=\sigma g^{''}$ for any $g^{''}\in \Hom_{S^{op}}(C,M^+\otimes_RC)$ is
an isomorphism in $\Mod R^{op}$.
Consider the following diagram
\begin{gather*}
\begin{split}
\xymatrix{
& M^+ \ar[rr]^{(\theta_M)^+} \ar [d]^{\mu_{M^+}} && [C\otimes_S\Hom_R(C,M)]^+ \ar [d]^{\tau} \\
& \Hom_{S^{op}}(C,M^+\otimes_RC) \ar[rr]^{\Hom_{S^{op}}(C,\sigma)} &&\Hom_{S^{op}}(C,[\Hom_R(C,M)]^+),}
\end{split}
\end{gather*}
where
$$(\theta_M)^+:M^+\to [C\otimes_S\Hom_R(C,M)]^+$$
via $(\theta_M)^+(g)=g\theta_M$ for any $g\in M^+$ is a natural homomorphism in $\Mod R^{op}$, and
$$\mu_{M^+}:M^+\to \Hom_{S^{op}}(C,M^+\otimes_RC)$$ defined by $\mu_{M^+}(g)(c)=g\otimes c$ for any $g\in M^+$ and $c\in C$
is a canonical valuation homomorphism in $\Mod R^{op}$. Then for any $g\in M^+$, $c\in C$ and $f\in\Hom_R(C,M)$, we have
$$\tau(\theta_M)^+(g)(c)(f)=(\theta_M)^+(g)(c\otimes f)=g\theta_M(c\otimes f)=gf(c),$$
$$\Hom_{S^{op}}(C,\sigma)\mu_{M^+}(g)(c)(f)=\sigma\mu_{M^+}(g)(c)(f)=\sigma(g\otimes c)(f)=gf(c),$$
Thus $$\tau(\theta_M)^+=\Hom_{S^{op}}(C,\sigma)\mu_{M^+},$$
and therefore $\theta_M$ is an isomorphism $\Leftrightarrow$ $(\theta_M)^+$ is an isomorphism
$\Leftrightarrow$ $\mu_{M^+}$ is an isomorphism.
We conclude that $M\in\mathcal{B}_C(R)\Leftrightarrow M^+\in\mathcal{A}_{C}(R^{op})$.
\end{proof}
As a consequence, we get the following
\begin{theorem}\label{3.3}
\begin{enumerate}
\item[]
\item[(1)] The pair
$$(\mathcal{A}_C(R^{op}),\mathcal{B}_C(R))$$ is a perfect coproduct-closed and product-closed duality pair
and $\mathcal{A}_C(R^{op})$ is covering and preenveloping in $\Mod R^{op}$.
\item[(2)] The pair
$$(\mathcal{B}_C(R),\mathcal{A}_C(R^{op}))$$ is a coproduct-closed and product-closed duality pair
and $\mathcal{B}_C(R)$ is covering and preenveloping in $\Mod R$.
\end{enumerate}
\end{theorem}
\begin{proof}
It follows from \cite[Proposition 4.2(a)]{HW07} that both $\mathcal{A}_C(R^{op})$ and $\mathcal{B}_C(R)$ are closed under
direct summands, coproducts and products. So by Lemma \ref{2.4}(1)(2) and Proposition \ref{3.2}, we have that
both the pairs
$$(\mathcal{A}_C(R^{op}),\mathcal{B}_C(R))\ {\rm and}\ (\mathcal{B}_C(R),\mathcal{A}_C(R^{op}))$$
are coproduct-closed and product-closed duality pairs, $\mathcal{A}_C(R^{op})$ is covering and preenveloping in $\Mod R^{op}$
and $\mathcal{B}_C(R)$ is covering and preenveloping in $\Mod R$. Moreover, $\mathcal{A}_C(R^{op})$ is projectively resolving
by \cite[Theorem 6.2]{HW07}, so the duality pair $(\mathcal{A}_C(R^{op}),\mathcal{B}_C(R))$ is perfect.
\end{proof}
We write
$$\mathcal{A}_C(R^{op})^{\bot}:=\{Y\in\Mod R^{op}\mid \Ext_{R^{op}}^{\geq 1}(N,Y)=0\ \text{for any}\ N\in\mathcal{A}_C(R^{op})\}.$$
The following corollary was proved in \cite[Theorem 3.11]{EH} when $R$ is a commutative noetherian ring and $_RC_S={_RC_R}$.
\begin{corollary}\label{3.4}
The pair $$(\mathcal{A}_C(R^{op}),\mathcal{A}_C(R^{op})^{\bot})$$ is a hereditary perfect cotorsion pair
and $\mathcal{A}_C(R^{op})$ is covering and preenveloping in $\Mod R^{op}$.
\end{corollary}
\begin{proof}
It follows from Theorem \ref{3.3}(1) and Lemma \ref{2.4}(3).
\end{proof}
The following two results are the symmetric versions of Theorem \ref{3.3} and Corollary \ref{3.4} respectively.
\begin{theorem}\label{3.5}
\begin{enumerate}
\item[]
\item[(1)] The pair $$(\mathcal{A}_C(S),\mathcal{B}_C(S^{op}))$$ is a perfect coproduct-closed and product-closed duality pair
and $\mathcal{A}_C(S)$ is covering and preenveloping in $\Mod S$.
\item[(2)] The pair $$(\mathcal{B}_C(S^{op}),\mathcal{A}_C(S))$$ is a coproduct-closed and product-closed duality pair
and $\mathcal{B}_C(S^{op})$ is covering and preenveloping in $\Mod S^{op}$.
\end{enumerate}
\end{theorem}
We write
$$\mathcal{A}_C(S)^{\bot}:=\{X\in\Mod S\mid \Ext_{S}^{\geq 1}(N^{'},X)=0\ \text{for any}\ N^{'}\in\mathcal{A}_C(S)\}.$$
\begin{corollary}\label{3.6}
The pair $$(\mathcal{A}_C(S),\mathcal{A}_C(S)^{\bot})$$ is a hereditary perfect cotorsion pair
and $\mathcal{A}_C(S)$ is covering and preenveloping in $\Mod S$.
\end{corollary}
Holm and White proved in \cite[Proposition 4.1]{HW07} that there exist the following (Foxby) equivalences of categories
$$\xymatrix@C=16ex{{\mathcal{A}_C(S)}\ar@<0.8ex>[r]_-{\sim}^-{C\otimes_{S}-}&
{\mathcal{B}_{C}(R)}\ar@<0.8ex>[l]^-{\Hom_{R}(C,-)},&}$$
$$\xymatrix@C=16ex{{\mathcal{A}_C(R^{op})}\ar@<0.8ex>[r]_-{\sim}^-{-\otimes_{R}C}&
{\mathcal{B}_{C}(S^{op})}\ar@<0.8ex>[l]^-{\Hom_{S^{op}}(C,-)}.&}$$
Compare this result with Theorems \ref{3.3} and \ref{3.5}.
By Theorems \ref{3.3}(2) and \ref{3.5}(1), $\mathcal{B}_C(R)$ is preenveloping in $\Mod R$
and $\mathcal{A}_C(S)$ is preenveloping in $\Mod S$. In the following result, we construct
an $\mathcal{A}_C(S)$-preenvelope of a given module in $\Mod S$ from a $\mathcal{B}_C(R)$-preenvelope
of some module in $\Mod R$.
\begin{theorem}\label{3.7}
\begin{enumerate}
\item[]
\item[(1)] Let $N\in\Mod S$ and
$$f:C\otimes_SN\to B$$ be a $\mathcal{B}_C(R)$-preenvelope of $C\otimes_SN$ in $\Mod R$.
Then we have
\begin{enumerate}
\item[(1.1)] $$f_*\mu_N:N\to B_*$$ is an $\mathcal{A}_C(S)$-preenvelope of $N$ in $\Mod S$.
\item[(1.2)] If $f$ is a $\mathcal{B}_C(R)$-envelope of $C\otimes_SN$, then $f_*\mu_N$ is an $\mathcal{A}_C(S)$-envelope of $N$.
\end{enumerate}
\item[(2)] If $\mathcal{B}_C(R)$ is enveloping in $\Mod R$, then $\mathcal{A}_C(S)$ is enveloping in $\Mod S$.
\end{enumerate}
\end{theorem}
\begin{proof}
(1.1) Let $N\in\Mod S$ and
$$f:C\otimes_SN\to B$$ be a $\mathcal{B}_C(R)$-preenvelope in $\Mod R$.
By \cite[Proposition 4.1]{HW07}, we have $B_*\in\mathcal{A}_C(S)$.
Let $g\in\Hom_S(N,A)$ with $A\in\mathcal{A}_C(S)$. By \cite[Proposition 4.1]{HW07}
again, we have $C\otimes_SA\in\mathcal{B}_C(R)$. So there exists $h\in\Hom_R(B,C\otimes_SA)$
such that $1_C\otimes g=hf$, that is, the following diagram
$$\xymatrix@R=15pt@C=15pt{
&&C\otimes_SN\ar[d]_{1_C\otimes g}\ar[r]^{\ \ \ \ f}&B\ar@{-->}[ld]^{h}&\\
&&C\otimes_SA&&}$$
commutes. From the following commutative diagram
$$\xymatrix@R=20pt@C=20pt{
N\ar[r]^{g}\ar[d]_{\mu_N} &A\ar[d]^{\mu_A}\\
(C\otimes_SN)_*\ar[r]^{(1_C\otimes g)_*}&(C\otimes_SA)_*,}$$
we get $\mu_Ag=(1_C\otimes g)_*\mu_N$. Because $\mu_A$ is an isomorphism, we have
$$g={\mu_A}^{-1}(1_C\otimes g)_*\mu_N=({\mu_A}^{-1}h_*)(f_*\mu_N),$$
that is, the following diagram
$$\xymatrix@R=15pt@C=15pt{
&&N\ar[d]_{g}\ar[r]^{f_*\mu_N}&B\ar@{-->}[ld]^{{\mu_A}^{-1}h_*}&\\
&&A&&}$$
commutes. Thus $f_*\mu_N:N\to B_*$ is an $\mathcal{A}_C(S)$-preenvelope of $N$.
(1.2) By (1.1), it suffices to prove that if $f$ is left minimal, then so is $f_*\mu_N$.
Let $f$ be left minimal and $h\in\Hom_S(B_*,B_*)$ such that $f_*\mu_N=h(f_*\mu_N)$. Then we have
$$(1_C\otimes f_*)(1_C\otimes \mu_N)=1_C\otimes (f_*\mu_N)=1_C\otimes (h(f_*\mu_N))
=(1_C\otimes h)(1_C\otimes f_*)(1_C\otimes \mu_N).\eqno{(3.1)}$$
From the following commutative diagram
$$\xymatrix@R=20pt@C=20pt{
C\otimes_S(C\otimes_SN)_*\ar[r]^{1_C\otimes f_*}\ar[d]_{\theta_{C\otimes_SN}} &C\otimes_SB_*\ar[d]^{\theta_B}\\
C\otimes_SN\ar[r]^{f}&B,}$$
we get
$$f\theta_{C\otimes_SN}=\theta_B(1_C\otimes f_*).\eqno{(3.2)}$$
So we have
\begin{align*}
&\ \ \ \ \ \ \ \ \ \ f=f1_{C\otimes_SN}\\
&\ \ \ \ \ \ \ \ \ \ \ \ =f(\theta_{C\otimes_SN}(1_C\otimes \mu_N))\ \text{(by \cite[Proposition 2.2(1)]{Wi}})\\
&\ \ \ \ \ \ \ \ \ \ \ \ =\theta_B(1_C\otimes f_*)(1_C\otimes \mu_N)\ \text{(by (3.2))}\\
&\ \ \ \ \ \ \ \ \ \ \ \ =\theta_B(1_C\otimes h)(1_C\otimes f_*)(1_C\otimes \mu_N)\ \text{(by (3.1))}\\
&\ \ \ \ \ \ \ \ \ \ \ \ =\theta_B(1_C\otimes h)({\theta_B}^{-1}\theta_B)(1_C\otimes f_*)(1_C\otimes \mu_N)\ \text{(because $\theta_B$ is an isomorphism)}\\
&\ \ \ \ \ \ \ \ \ \ \ \ =\theta_B(1_C\otimes h){\theta_B}^{-1}f\theta_{C\otimes_SN}(1_C\otimes \mu_N)\ \text{(by (3.2))}\\
&\ \ \ \ \ \ \ \ \ \ \ \ =\theta_B(1_C\otimes h){\theta_B}^{-1}f1_{C\otimes_SN}\ \text{(by \cite[Proposition 2.2(1)]{Wi}})\\
&\ \ \ \ \ \ \ \ \ \ \ \ =\theta_B(1_C\otimes h){\theta_B}^{-1}f.
\end{align*}
Because $f$ is left minimal, $\theta_B(1_C\otimes h){\theta_B}^{-1}$ is an isomorphism, which implies that
$1_C\otimes h$ and $(1_C\otimes h)_*$ are also isomorphisms. From the following commutative diagram
$$\xymatrix@R=20pt@C=20pt{
B_*\ar[r]^{h}\ar[d]_{\mu_{B_*}} &B_*\ar[d]^{\mu_{B_*}}\\
(C\otimes_SB_*)_*\ar[r]^{(1_C\otimes h)_*}&(C\otimes_SB_*)_*,}$$
we get
$$(1_C\otimes h)_*\mu_{B_*}=\mu_{B_*}h.$$
Because $B_*\in\mathcal{A}_C(S)$ by \cite[Proposition 4.1]{HW07}, $\mu_{B_*}$ is an isomorphism.
It follows that $h$ is also an isomorphism and $f_*\mu_N$ is left minimal.
(2) It follows from the assertion (1.2) immediately.
\end{proof}
We do not know whether a $\mathcal{B}_C(R)$-preenvelope of given module in $\Mod R$
can be constructed from an $\mathcal{A}_C(S)$-preenvelope of some module in $\Mod S$,
and do not know whether the converse of Theorem \ref{3.7}(2) holds true.
By Theorems \ref{3.3}(2) and \ref{3.5}(1), $\mathcal{B}_C(R)$ is covering in $\Mod R$
and $\mathcal{A}_C(S)$ is covering in $\Mod S$. In the following result, we construct
a $\mathcal{B}_C(R)$-cover of a given module in $\Mod R$ from an $\mathcal{A}_C(S)$-cover
of some module in $\Mod S$.
\begin{proposition}\label{3.8}
Let $M\in\Mod R$ and $$g:A\to M_*$$ be an $\mathcal{A}_C(S)$-cover of $M_*$ in $\Mod S$.
Then
$$\theta_M(1_C\otimes g):C\otimes_SA\to M$$ is a $\mathcal{B}_C(R)$-cover of $M$ in $\Mod R$.
\end{proposition}
\begin{proof}
Let $M\in\Mod R$ and $$g:A\to M_*$$ be an $\mathcal{A}_C(S)$-cover of $M_*$ in $\Mod S$.
By \cite[Proposition 4.1]{HW07}, we have $C\otimes_SA\in\mathcal{B}_C(R)$.
Let $f\in\Hom_R(B,M)$ with $B\in\mathcal{B}_C(R)$. By \cite[Proposition 4.1]{HW07}
again, we have $B_*\in\mathcal{A}_C(S)$. So there exists $h\in\Hom_S(B_*,A)$
such that ${f}_*=gh$, that is, the following diagram
$$\xymatrix{ & B_* \ar[d]^{{f}_*} \ar@{-->}[ld]_{h}\\
A \ar[r]^{g} & M_*}$$
commutes. From the following commutative diagram
$$\xymatrix@R=20pt@C=20pt{
C\otimes_SB_*\ar[r]^{1_C\otimes {f}_*}\ar[d]_{\theta_B} &C\otimes_SM_*\ar[d]^{\theta_M}\\
B\ar[r]^{f}&M,}$$
we get $f\theta_B=\theta_M(1_C\otimes {f}_*)$. Because $\theta_B$ is an isomorphism, we have
$$f=\theta_M(1_C\otimes {f}_*){\theta_B}^{-1}=\theta_M(1_C\otimes (gh)){\theta_B}^{-1}
=(\theta_M(1_C\otimes g))((1_C\otimes h)){\theta_B}^{-1}),$$
that is, the following diagram
$$\xymatrix{ & B \ar[d]^{f} \ar@{-->}[ld]_{(1_C\otimes h)){\theta_B}^{-1}}\\
C\otimes_SA \ar[r]_{\theta_M(1_C\otimes g)} & M}$$
commutes. Thus $\theta_M(1_C\otimes g):C\otimes_SA\to M$ is a $\mathcal{B}_C(R)$-precover of $M$.
In the following, it suffices to prove that $\theta_M(1_C\otimes g)$ is right minimal.
Let $h\in\Hom_R(C\otimes_SA,C\otimes_SA)$ such that $\theta_M(1_C\otimes g)=(\theta_M(1_C\otimes g))h$. Then we have
$$(\theta_M)_*(1_C\otimes g)_*=(\theta_M(1_C\otimes g))_*=((\theta_M(1_C\otimes g))h)_*
=(\theta_M)_*(1_C\otimes g)_*h_*.\eqno{(3.3)}$$
From the following commutative diagram
$$\xymatrix@R=20pt@C=20pt{
A\ar[r]^{g}\ar[d]_{\mu_A} &M_*\ar[d]^{\mu_{M_*}}\\
(C\otimes_SA)_*\ar[r]^{(1_C\otimes g)_*}&(C\otimes_SM_*)_*,}$$
we get
$$\mu_{M_*}g=(1_C\otimes g)_*\mu_A.\eqno{(3.4)}$$
So we have
\begin{align*}
&\ \ \ \ \ \ \ \ \ \ g=1_{M_*}g\\
&\ \ \ \ \ \ \ \ \ \ \ \ =(\theta_M)_*\mu_{M_*}g\ \text{(by \cite[Proposition 2.2(1)]{Wi}})\\
&\ \ \ \ \ \ \ \ \ \ \ \ =(\theta_M)_*(1_C\otimes g)_*\mu_A\ \text{(by (3.4))}\\
&\ \ \ \ \ \ \ \ \ \ \ \ =(\theta_M)_*(1_C\otimes g)_*h_*\mu_A\ \text{(by (3.3))}\\
&\ \ \ \ \ \ \ \ \ \ \ \ =(\theta_M)_*(1_C\otimes g)_*\mu_A{\mu_A}^{-1}h_*\mu_A\ \text{(because $\mu_A$ is an isomorphism)}\\
&\ \ \ \ \ \ \ \ \ \ \ \ =(\theta_M)_*\mu_{M_*}g{\mu_A}^{-1}h_*\mu_A\ \text{(by (3.4))}\\
&\ \ \ \ \ \ \ \ \ \ \ \ =1_{M_*}g{\mu_A}^{-1}h_*\mu_A\ \text{(by \cite[Proposition 2.2(1)]{Wi}})\\
&\ \ \ \ \ \ \ \ \ \ \ \ =g{\mu_A}^{-1}h_*\mu_A.
\end{align*}
Because $g$ is right minimal, ${\mu_A}^{-1}h_*\mu_A$ is an isomorphism, which implies that
$h_*$ and $1_C\otimes h_*$ are also isomorphisms. From the following commutative diagram
$$\xymatrix@R=20pt@C=20pt{
C\otimes_S(C\otimes_SA)_*\ar[r]^{1_C\otimes h_*}\ar[d]_{\theta_{C\otimes_SA}} &C\otimes_S(C\otimes_SA)_*\ar[d]^{\theta_{C\otimes_SA}}\\
C\otimes_SA\ar[r]^{h}&C\otimes_SA,}$$
we get
$$h\theta_{C\otimes_SA}=\theta_{C\otimes_SA}(1_C\otimes h_*).$$
Because $C\otimes_SA\in\mathcal{B}_C(R)$ by \cite[Proposition 4.1]{HW07}, $\theta_{C\otimes_SA}$ is an isomorphism.
It follows that $h$ is also an isomorphism and $\theta_M(1_C\otimes g)$ is right minimal.
\end{proof}
We do not know whether an $\mathcal{A}_C(S)$-cover of a given module in $\Mod S$
can be constructed from a $\mathcal{B}_C(R)$-cover of some module in $\Mod R$.
\section{The Auslander projective dimension of modules}
For a subcategory $\mathscr{X}$ of $\Mod S$ and $N\in \Mod S$, the \textbf{$\mathscr{X}$-projective dimension}
$\mathscr{X}$-$\pd_SN$ of $N$ is defined as $\inf\{n\mid$ there exists an exact sequence
$$0 \to X_n \to \cdots \to X_1\to X_0 \to N\to 0$$
in $\Mod S$ with all $X_i\in\mathscr{X}\}$, and we set $\mathscr{X}$-$\pd_SN$ infinite
if no such integer exists. We call $\mathcal{A}_{C}(S)$-$\pd_SN$ the \textbf{Auslander projective dimension}
of $N$. For any $n\geq 0$, we use $\Omega^n(N)$ to denote the $n$-th syzygy of $N$ (note: $\Omega^0(N)=N$).
\begin{lemma}\label{4.1}
Let $N\in\Mod S$ and $n\geq 0$. If $\mathcal{A}_{C}(S)$-$\pd_SN\leq n$ and
$$0\to K_n \to A_{n-1} \to \cdots \to A_1\to A_0 \to N\to 0$$
be an exact sequence in $\Mod S$ with all $A_i$ in $\mathcal{A}_{C}(S)$,
then $K_n\in\mathcal{A}_{C}(S)$; in particular, $\Omega^n(N)\in\mathcal{A}_C(S)$.
\end{lemma}
\begin{proof}
Because $\mathcal{A}_{C}(S)$ is projectively resolving and is closed under direct summands and coproducts
by \cite[Theorem 6.2 and Proposition 4.2(a)]{HW07}, the assertion follows from \cite[Lemma 3.12]{AB}.
\end{proof}
We use $\mathcal{A}_{C}(S)$-$\pd^{<\infty}$ to denote the subcategory of $\Mod S$ consisting of modules
with finite Auslander projective dimension.
\begin{proposition}\label{4.2}
$\mathcal{A}_{C}(S)$-$\pd^{<\infty}$ is closed under extensions, kernels of epimorphisms and
cokernels of monomorphisms.
\end{proposition}
\begin{proof}
Let $$0\to N_1 \to N_2 \to N_3 \to 0$$
be an exact sequence in $\Mod S$ and $n\geq 0$. If
$\max\{\mathcal{A}_{C}(S)$-$\pd_SN_1,\mathcal{A}_{C}(S)$-$\pd_SN_3\}\leq n$,
then by Lemma \ref{4.1}, there exist exact sequences
$$0\to \Omega^n(N_1)\to P^{n-1}_1 \to \cdots\to P^{1}_1 \to P^{0}_1 \to N_1 \to 0,$$
$$0\to \Omega^n(N_3)\to P^{n-1}_3 \to \cdots\to P^{1}_3 \to P^{0}_3 \to N_3 \to 0$$
in $\Mod S$ with all $P_i^j$ projective and $\Omega^n(N_1),\Omega^n(N_3)\in\mathcal{A}_{C}(S)$.
Then we get exact sequences
$$0\to K_n\to P^{n-1}_1\oplus P^{n-1}_3 \to \cdots\to P^{1}_1 \oplus P^{1}_3\to P^{0}_1\oplus P^{0}_3 \to N_2 \to 0,$$
$$0\to \Omega^n(N_1)\to K_n \to \Omega^n(N_3)\to 0$$
in $\Mod S$. By \cite[Theorem 6.2]{HW07}, we have $K_n\in\mathcal{A}_{C}(S)$ and
$\mathcal{A}_{C}(S)$-$\pd_SN_2\leq n$.
If $\max\{\mathcal{A}_{C}(S)$-$\pd_SN_1,\mathcal{A}_{C}(S)$-$\pd_SN_2\}\leq n$, then by Corollary \ref{3.6}
and Lemma \ref{4.1}, there exist $\Hom_S(\mathcal{A}_{C}(S),-)$-exact exact sequences
$$0\to A^{n}_1\to A^{n-1}_1 \to \cdots\to A^{1}_1 \to A^{0}_1 \to N_1 \to 0,$$
$$0\to A^{n}_2\to A^{n-1}_2 \to \cdots\to A^{1}_2 \to A^{0}_2 \to N_2 \to 0$$
in $\Mod S$ with all $A_i^j$ in $\mathcal{A}_{C}(S)$. By \cite[Theorem 3.6]{Hu}, we get an exact sequence
$$0\to A^n_1\to A^{n-1}_1\oplus A^{n}_2 \to \cdots\to A^{0}_1 \oplus A^{1}_2\to A^{0}_2 \to N_3 \to 0$$
in $\Mod S$, and so $\mathcal{A}_{C}(S)$-$\pd_SN_3\leq n+1$.
If $\max\{\mathcal{A}_{C}(S)$-$\pd_SN_2,\mathcal{A}_{C}(S)$-$\pd_SN_3\}\leq n$, then by Corollary \ref{3.6}
and Lemma \ref{4.1}, there exist $\Hom_S(\mathcal{A}_{C}(S),-)$-exact exact sequences
$$0\to A^{n}_2\to A^{n-1}_2 \to \cdots\to A^{1}_2 \to A^{0}_2 \to N_2 \to 0,$$
$$0\to A^{n}_3\to A^{n-1}_3 \to \cdots\to A^{1}_3 \to A^{0}_3 \to N_3 \to 0$$
in $\Mod S$ with all $A_i^j$ in $\mathcal{A}_{C}(S)$. By \cite[Theorem 3.2]{Hu}, we get exact sequences
$$0\to A^n_2\to A^{n-1}_2\oplus A^{n}_3 \to \cdots\to A^{1}_2 \oplus A^{2}_3\to A \to N_1 \to 0,$$
$$0\to A \to A_{2}^0 \oplus A_{3}^1\to A^{0}_3 \to 0$$
in $\Mod S$. By \cite[Theorem 6.2]{HW07}, we have $A\in\mathcal{A}_{C}(S)$,
and so $\mathcal{A}_{C}(S)$-$\pd_SN_1\leq n$.
\end{proof}
We write
$$\mathcal{I}_C(S):=\{I_*\mid I\ {\rm \ is\ injective\ in}\ \Mod R\}.$$
The modules in $\mathcal{I}_C(S)$ is called {\bf $C$-injective} (\cite{HW07}).
Let $Q$ be an injective cogenerator for $\Mod R$. Then
$$\mathcal{I}_C(S)=\Prod_SQ_*$$ by \cite[Proposition 2.4(2)]{LHX},
where $\Prod_SQ_*$ is the subcategory of $\Mod S$ consisting of direct summands
of products of copies of $Q_*$. By \cite[Lemma 2.16(b)]{GT12}, we have the following
isomorphism of functors
$$\Hom_R(\Tor_i^S(C,-),Q)\cong\Ext_S^i(-,Q_*)$$
for any $i\geq 1$. This gives the following
\begin{lemma}\label{4.3}
${C_S}^{\top}={^{\bot}\mathcal{I}_C(S)}$.
\end{lemma}
For a subcategory $\mathscr{X}$ of $\Mod S$, a sequence in $\Mod S$ is called
{\bf $\Hom_{S}(-,\mathscr{X})$-exact} if it is exact after applying the functor
$\Hom_{S}(-,X)$ for any $X\in\mathscr{X}$.
Now we give some criteria for computing the Auslander projective dimension of modules.
\begin{theorem}\label{4.4}
Let $N\in\Mod S$ with $\mathcal{A}_{C}(S)$-$\pd_SN<\infty$ and $n\geq 0$.
Then the following statements are equivalent.
\begin{enumerate}
\item[(1)] $\mathcal{A}_{C}(S)$-$\pd_SN\leq n$.
\item[(2)] $\Omega^n(N)\in\mathcal{A}_{C}(S)$.
\item[(3)] $\Tor^S_{\geq n+1}(C,N)=0$.
\item[(4)] There exists an exact sequence
$$0\to H \to A \to N\to 0$$
in $\Mod S$ with $A\in \mathcal{A}_{C}(S)$ and $\mathcal{I}_C(S)$-$\pd_SH\leq n-1$.
\item[(5)] There exists a ($\Hom_{S}(-,\mathcal{I}_C(S))$-exact) exact sequence
$$0\to N \to H^{'} \to A^{'} \to 0$$
in $\Mod S$ with $A^{'}\in \mathcal{A}_{C}(S)$ and $\mathcal{I}_C(S)$-$\pd_SH^{'}\leq n$.
\end{enumerate}
\end{theorem}
\begin{proof}
By Lemma \ref{4.1} and the dimension shifting, we have $(1)\Leftrightarrow (2)\Rightarrow (3)$.
$(3)\Rightarrow (2)$ Because $\Tor^S_{\geq n+1}(C,N)=0$ by (3), we have $\Omega^n(N)\in{C_S}^\top$,
and so $\Omega^n(N)\in{^{\bot}\mathcal{I}_C(S)}$ by Lemma \ref{4.3}.
Note that all projective modules in $\Mod S$ are in $\mathcal{A}_C(S)$ by \cite[Theorem 6.2]{HW07}.
Because $\mathcal{A}_{C}(S)$-$\pd_SN<\infty$ by assumption,
we have $\mathcal{A}_{C}(S)$-$\pd_S\Omega^n(N)<\infty$ by Proposition 4.2.
Assume that $\mathcal{A}_{C}(S)$-$\pd_S\Omega^n(N)=m(<\infty)$ and
$$0\to A_m \to \cdots \to A_1 \to A_0 \to \Omega^n(N) \to 0 \eqno{(4.1)}$$
is an exact sequence in $\Mod S$ with all $A_j$ in $\mathcal{A}_C(S)$.
Because $\mathcal{A}_C(S)\subseteq {C_S}^{\top}={^{\bot}\mathcal{I}_C(S)}$ by
Lemma \ref{4.3}, the exact sequence (4.1) is $\Hom_{S}(-,\mathcal{I}_C(S))$-exact.
By \cite[Theorem 3.11(1)]{TH}, we have the following $\Hom_{S}(-,\mathcal{I}_C(S))$-exact exact sequence
$$0\to A_j \to U_j^0\to U_j^1 \to \cdots \to U_j^i \to \cdots$$
in $\Mod S$ with all $U_j^i$ in $\mathcal{I}_C(S)$ for any $0\leq j\leq m$ and $i\geq 0$.
It follows from \cite[Corollary 3.5]{Hu} that there exist
the following two exact sequences
$$0\to \Omega^n(N) \to U \to \oplus _{i=0}^mU_i^{i+1} \to \oplus_{i=0}^mU_i^{i+2}
\to \oplus_{i=0}^mU_i^{i+3} \to \cdots,$$
$$0\to U_m^0 \to U_m^1\oplus U_{m-1}^0 \to \cdots \to \oplus _{i=2}^mU_i^{i-2}
\to \oplus_{i=1}^mU_i^{i-1} \to \oplus _{i=0}^mU_i^i \to U \to 0,$$
and the former one is $\Hom_{S}(-,\mathcal{I}_C(S))$-exact. Because $\mathcal{I}_C(S)$ is closed
under finite direct sums and cokernels of monomorphisms by \cite[Proposition 5.1(c) and Corollary 6.4]{HW07},
we have $U\in\mathcal{I}_{C}(S)$. By \cite[Theorem 3.11(1)]{TH} again, we have
$\Omega^n(N)\in\mathcal{A}_{C}(S)$.
$(1)\Rightarrow (4)$ By \cite[Theorem 6.2]{HW07}, $\mathcal{A}_C(S)$ is closed under extensions.
By \cite[Theorem 3.11(1)]{TH}, we have that
$\mathcal{I}_C(S)$ is an $\mathcal{I}_C(S)$-coproper cogenerator
for $\mathcal{A}_C(S)$ in the sense of \cite{Hu2}. Then the assertion follows from \cite[Theorem 4.7]{Hu2}.
$(4)\Rightarrow (5)$ Let
$$0\to H \to A \to N\to 0$$
be an exact sequence in $\Mod S$ with $A\in \mathcal{A}_{C}(S)$ and $\mathcal{I}_C(S)$-$\pd_SH\leq n-1$.
By \cite[Theorem 3.11(1)]{TH}, there exists a $\Hom_{S}(-,\mathcal{I}_C(S))$-exact exact sequence
$$0\to A\to U\to A^{'}\to 0$$ in $\Mod S$ with $U\in\mathcal{I}_C(S)$
and $A^{'}\in \mathcal{A}_{C}(S)$. Consider the following push-out diagram
$$\xymatrix{
& & 0 \ar[d] &0 \ar@{-->}[d] & \\
0 \ar[r] & H \ar@{==}[d] \ar[r] & A \ar[d]\ar[r] &N \ar@{-->}[d] \ar[r] & 0 \\
0 \ar@{-->}[r] & H \ar@{-->}[r] & U \ar[d] \ar@{-->}[r] & H^{'} \ar@{-->}[d] \ar@{-->}[r] & 0 \\
& & A^{'} \ar[d]\ar@{==}[r] & A^{'}\ar@{-->}[d] & \\
& & 0 & 0. & } $$
By the middle row in this diagram, we have $\mathcal{I}_C(S)$-$\pd_SH^{'}\leq n$. Because the middle column
in the above diagram is $\Hom_{S}(-,\mathcal{I}_C(S))$-exact, the rightmost column is also
$\Hom_{S}(-,\mathcal{I}_C(S))$-exact by \cite[Lemma 2.4(2)]{Hu} and it is the desired exact sequence.
$(5)\Rightarrow (1)$ Let
$$0\to N \to H^{'} \to A^{'} \to 0$$
be an exact sequence in $\Mod S$ with $A^{'}\in \mathcal{A}_{C}(S)$ and $\mathcal{I}_C(S)$-$\pd_SH^{'}\leq n$.
Then there exists an exact sequence
$$0\ra U_n\ra \cdots\ra U_1\ra U_0\ra H^{'}\ra 0$$ in $\Mod S$ with all $U_i$ in $\mathcal{I}_C(S)$.
Set $H:=\mbox{Ker}(U_0\ra H^{'})$.
Then $\mathcal{I}_C(S)$-$\pd_SH\leq n-1$. Consider the following pull-back diagram
$$\xymatrix@R=20pt@C=20pt{& 0 \ar@{-->}[d] & 0 \ar[d]&& &\\
& H \ar@{==}[r] \ar@{-->}[d] & H \ar[d]& &&\\
0 \ar@{-->}[r] & A \ar@{-->}[d] \ar@{-->}[r] & U_0 \ar[d] \ar@{-->}[r] &A^{'} \ar@{==}[d] \ar@{-->}[r] & 0\\
0 \ar[r] & N \ar@{-->}[d]\ar[r] & H^{'} \ar[r] \ar[d] & A^{'} \ar[r] & 0 &\\
& 0 & 0. & && }$$
Applying \cite[Theorem 6.2]{HW07} to the middle row in this diagram yields $A\in \mathcal{A}_{C}(S)$.
Thus $\mathcal{A}_{C}(S)$-$\pd_SN\leq n$ by the leftmost column in the above diagram.
\end{proof}
The only place where the assumption $\mathcal{A}_{C}(S)$-$\pd_SN<\infty$ in Theorem \ref{4.4} is used is in showing
$(3)\Rightarrow(2)$. By Theorem \ref{4.4}, it is easy to get the following standard observation.
\begin{corollary}\label{4.5}
Let $$0\to L \to M \to K\to 0$$
be an exact sequence in $\Mod S$. Then we have
\begin{enumerate}
\item[(1)] $\mathcal{A}_{C}(S)$-$\pd_{S}K \leq \max\{\mathcal{A}_{C}(S)$-$\pd_{S}M,\mathcal{A}_{C}(S)$-$\pd_{S}L+1\}$,
and the equality holds true if $\mathcal{A}_{C}(S)$-$\pd_{S}M \neq \mathcal{A}_{C}(S)$-$\pd_{S}L$.
\item[(2)] $\mathcal{A}_{C}(S)$-$\pd_{S}L \leq \max\{\mathcal{A}_{C}(S)$-$\pd_{S}M,\mathcal{A}_{C}(S)$-$\pd_{S}K-1\}$,
and the equality holds true if $\mathcal{A}_{C}(S)$-$\pd_{S}M\neq \mathcal{A}_{C}(S)$-$\pd_{S}K$.
\item[(3)] $\mathcal{A}_{C}(S)$-$\pd_{S}M \leq \max\{\mathcal{A}_{C}(S)$-$\pd_{S}L,\mathcal{A}_{C}(S)$-$\pd_{S}K\}$,
and the equality holds true if $\mathcal{A}_{C}(S)$-$\pd_{S}K \neq \mathcal{A}_{C}(S)$-$\pd_{S}L+1$.
\end{enumerate}
\end{corollary}
The following corollary is an addendum to the implications $(1)\Rightarrow (4)$ and $(1)\Rightarrow (5)$ in Theorem \ref{4.4}.
\begin{corollary}\label{4.6}
Let $N\in\Mod S$ with $\mathcal{A}_{C}(S)$-$\pd_SN=n(<\infty)$.
Then there exist exact sequences
$$0\to H \to A \to N\to 0,$$
$$0\to N \to H^{'} \to A^{'} \to 0$$
in $\Mod S$ with $A,A^{'}\in \mathcal{A}_{C}(S)$ and $\mathcal{I}_C(S)$-$\pd_SH=\mathcal{I}_C(S)$-$\pd_SH^{'}=n$.
\end{corollary}
\begin{proof}
Let $N\in\Mod S$ with $\mathcal{A}_{C}(S)$-$\pd_SN=n(<\infty)$.
By Theorem \ref{4.4}, there exists an exact sequence
$$0\to H \to A \to N\to 0$$
in $\Mod S$ with $A\in \mathcal{A}_{C}(S)$ and $(\mathcal{A}_{C}(S)$-$\pd_SH\leq)\mathcal{I}_C(S)$-$\pd_SH\leq n-1$.
By Theorem \ref{4.4} again, we have $\sup\{i\geq 0\mid \Tor^S_{i}(C,N)\neq 0\}=n$.
So $\sup\{i\geq 0\mid \Tor^S_{i}(C,H)\neq 0\}=n-1$, and hence
$\mathcal{A}_{C}(S)$-$\pd_SH=n-1$ by Theorem \ref{4.4}.
It follows that $\mathcal{I}_C(S)$-$\pd_SH=n-1$.
By Theorem \ref{4.4}, there exists an exact sequence
$$0\to N \to H^{'} \to A^{'} \to 0$$
in $\Mod S$ with $A^{'}\in \mathcal{A}_{C}(S)$ and $(\mathcal{A}_{C}(S)$-$\pd_SH\leq)\mathcal{I}_C(S)$-$\pd_SH^{'}\leq n$.
By Corollary \ref{4.5}(3), we have $\mathcal{A}_{C}(S)$-$\pd_SH=\mathcal{A}_{C}(S)$-$\pd_SN=n$,
and so $\mathcal{I}_C(S)$-$\pd_SH^{'}=n$.
\end{proof}
Let $N\in \Mod S$. Bican, El Bashir and Enochs proved in \cite{BBE} that $N$ has a flat cover. We use
$$\cdots \buildrel {f_{n+1}} \over \longrightarrow F_n(N) \buildrel {f_n} \over \longrightarrow \cdots
\buildrel {f_2} \over \longrightarrow F_1(N) \buildrel {f_1} \over \longrightarrow F_0(N) \buildrel {f_0}
\over \longrightarrow N \to 0\eqno{(4.2)}$$
to denote a minimal flat resolution of $N$ in $\Mod S$, where each $F_i(N)\to\Im f_i$ is a flat cover of $\Im f_i$.
\begin{lemma}\label{4.7}
Let $N\in\Mod S$ and $n\geq 0$. If $\Tor_{1\leq i\leq n}^S(C,N)=0$, then we have
\begin{enumerate}
\item[(1)] There exists an exact sequence
$$0\rightarrow \Ext_R^{n+1}(C,\Ker(1_C\otimes f_{n+1}))\rightarrow N\stackrel{\mu_{N}}{\longrightarrow}
(C\otimes_SN)_*\rightarrow\Ext_R^{n+2}(C,\Ker(1_C\otimes f_{n+1}))\rightarrow 0$$
in $\Mod S$.
\item[(2)] $\Ext_R^{1\leq i \leq n}(C,\Ker(1_C\otimes f_{n+1}))=0$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) The case for $n=0$ follows from \cite[Proposition 3.2]{TH}. Now suppose $n\geq 1$. If $\Tor_{1\leq i\leq n}^S(C,N)=0$,
then the exact sequence (4.2) yields the following exact sequence
$$0\to \Ker(1_C\otimes f_{n+1}) \to C\otimes_SF_{n+1}(N) \buildrel {1_C\otimes f_{n+1}}\over \longrightarrow
C\otimes_SF_{n}(N) \buildrel {1_C\otimes f_{n}}\over \longrightarrow\cdots$$
$$\buildrel {1_C\otimes f_{2}}\over \longrightarrow C\otimes_SF_{1}(N) \buildrel {1_C\otimes f_{1}}
\over \longrightarrow C\otimes_SF_{0}(N) \buildrel {1_C\otimes f_{0}}\over \longrightarrow C\otimes_SN \to 0\eqno{(4.3)}$$
in $\Mod R$. Because all $C\otimes_SF_i(N)$ are in $_RC^{\bot}$ by \cite[Lemma 2.3(1)]{TH}, we have
$$\Ext_R^{1}(C,\Ker(1_C\otimes f_1))\cong \Ext_R^{n+1}(C,\Ker(1_C\otimes f_n)),$$
$$\Ext_R^{2}(C,\Ker(1_C\otimes f_1))\cong \Ext_R^{n+2}(C,\Ker(1_C\otimes f_n)).$$
Now the assertion follows from \cite[Proposition 3.2]{TH}.
(2) Applying the functor $(-)_*$ to the exact sequence (4.3) we get the following commutative diagram
$$\xymatrix{
&&F_{n+1}(N) \ar[d]^{\mu_{F_{n+1}(N)}}\ar[r]^{\ \ \ \ f_{n+1}}& F_n(N)\ar[d]^{\mu_{F_{n}(N)}}\ar[r]^{f_n}
& \cdots \ar[r]^{f_1} & F_0(N)\ar[d]^{\mu_{F_{0}(N)}}\\
0\ar[r]&(\Ker(1_C\otimes f_{n+1}))_*\ar[r]& (C\otimes_SF_{n+1}(N))_*\ar[r]^{\ \ \ \ (1_C\otimes f_{n+1})_*\ \ \ \ }
& (C\otimes_SF_{n}(N))_*\ar[r]^{\ \ \ \ \ \ (1_C\otimes f_{n})_*}\ar[r]& \cdots \ar[r]^{(1_C\otimes f_{1})_* \ \ \ \ \ \ \ }& (C\otimes_SF_{0}(N))_*.}$$
All columns are isomorphisms by \cite[Lemma 4.1]{HW07}. So the bottom row in this diagram is exact.
Because all $C\otimes_SF_{i}(N)$ are in $_RC^{\bot}$, we have
$\Ext_R^{1\leq i \leq n}(C,\Ker(1_C\otimes f_{n+1}))=0$.
\end{proof}
Let $X\in \Mod R$ and let $$\cdots \buildrel {g_{n+1}} \over \longrightarrow P_n \buildrel {g_n} \over \longrightarrow
\cdots \buildrel {g_2} \over \longrightarrow P_1 \buildrel {g_1} \over \longrightarrow P_0 \buildrel {g_0} \over \longrightarrow X \to 0$$
be a projective resolution of $X$ in $\Mod R$. If there exists $n\geq 1$ such that $\Im g_n\cong \oplus_jW_j$, where each $W_j$ is
isomorphic to a direct summand of some $\Im g_{i_j}$ with $i_j<n$, then we say that $X$ {\bf has an ultimately closed projective resolution at $n$};
and we say that $X$ {\bf has an ultimately closed projective resolution} if it has an ultimately closed projective resolution at some $n$ (\cite{J}).
It is trivial that if $\pd_{R}X$ (the projective dimension of $X$) $\leq n$, then $X$ has an ultimately closed projective resolution at $n+1$.
Let $R$ be an artin algebra. If either $R$ is of finite representation type or the square of the radical of $R$ is zero,
then any finitely generated left $R$-module has an ultimately closed projective resolution (\cite[p.341]{J}).
Following \cite{Wi}, a module $N\in\Mod S$ is called {\bf $C$-adstatic} if $\mu_N$ is an isomorphism.
\begin{proposition}\label{4.8}
Let $N\in\Mod S$ and $n\geq 1$. If $\Tor_{1\leq i\leq n}^S(C,N)=0$, then $N$ is $C$-adstatic
provided that one of the following conditions is satisfied.
\begin{enumerate}
\item[(1)] $\pd_{R}C\leq n$.
\item[(2)] $_RC$ has an ultimately closed projective resolution at $n$.
\end{enumerate}
\end{proposition}
\begin{proof} (1) It follows directly from Lemma \ref{4.7}(1).
(2) Let
$$\cdots \buildrel {g_{n+1}} \over \longrightarrow P_n \buildrel {g_n} \over \longrightarrow
\cdots \buildrel {g_2} \over \longrightarrow P_1 \buildrel {g_1} \over \longrightarrow
P_0 \buildrel {g_0} \over \longrightarrow C \to 0$$ be a projective resolution of $C$ in $\Mod R$
ultimately closed at $n$. Then $\Im g_n\cong \oplus_jW_j$ such that each $W_j$ is
isomorphic to a direct summand of some $\Im g_{i_j}$ with $i_j<n$. Let $N\in\Mod S$ with
$\Tor_{1\leq i\leq n}^S(C,N)=0$. By Lemma \ref{4.7}(2), we have
$$\Ext^{1}_R(\Im g_{i_j},\Ker(1_C\otimes f_{n+1}))\cong \Ext^{i_j+1}_R(C,\Ker(1_C\otimes f_{n+1}))=0.$$
Because $W_j$ is isomorphic to a direct summand of some $\Im g_{i_j}$, we have $\Ext^{1}_R(W_j,\Ker(1_C\otimes f_{n+1}))=0$
for any $j$, which implies
\begin{align*}
&\ \ \ \ \Ext^{n+1}_R(C,\Ker(1_C\otimes f_{n+1}))\\
& \cong \Ext^{1}_R(\Im g_{n},\Ker(1_C\otimes f_{n+1}))\\
& \cong \Ext^{1}_R(\oplus_jW_j,\Ker(1_C\otimes f_{n+1}))\\
& \cong \Pi_j\Ext^{1}_R(W_j,\Ker(1_C\otimes f_{n+1}))\\
& =0.
\end{align*}
Then by Lemma \ref{4.7}(2), we conclude that $\Ext_R^{1\leq i \leq n+1}(C,\Ker(1_C\otimes f_{n+1}))=0$.
Similar to the above argument we get $\Ext_R^{n+2}(C,\Ker(1_C\otimes f_{n+1}))=0$.
It follows from Lemma \ref{4.7}(1) that $\mu_N$ is an isomorphism and $N$ is $C$-adstatic.
\end{proof}
\begin{corollary}\label{4.9}
For any $n\geq 1$,
a module $N\in \Mod S$ satisfying $\Tor_{0\leq i\leq n}^S(C,N)=0$ implies $N=0$ provided that
one of the following conditions is satisfied.
\begin{enumerate}
\item[(1)] $\pd_{R}C\leq n$.
\item[(2)] $_RC$ has an ultimately closed projective resolution at $n$.
\end{enumerate}
\end{corollary}
\begin{proof}
Let $N\in \Mod S$ with $\Tor_{0\leq i\leq n}^S(C,N)=0$. By Proposition \ref{4.8},
we have that $N$ is $C$-adstatic and $N\cong (C\otimes_SN)_*=0$.
\end{proof}
We now are in a position to give the following
\begin{theorem}\label{4.10}
If $_RC$ has an ultimately closed projective resolution, then
$$\mathcal{A}_C(S)={{C_S}^{\top}}={^{\bot}\mathcal{I}_C(S)}.$$
\end{theorem}
\begin{proof}
By the definition of $\mathcal{A}_C(S)$ and Lemma \ref{4.3}, we have
$\mathcal{A}_C(S)\subseteq{{C_S}^{\top}}={^{\bot}\mathcal{I}_C(S)}$.
Now let $N\in{^{\bot}\mathcal{I}_C(S)}$ and let $f:C\otimes_SN\to B$
be a $\mathcal{B}_C(R)$-preenvelope of $C\otimes_SN$ in $\Mod R$ as in Theorem \ref{3.7}.
Because $\mathcal{B}_C(R)$ is injectively coresolving
in $\Mod R$ by \cite[Theorem 6.2]{HW07}, $f$ is monic. By Proposition \ref{4.8},
$\mu_N$ is an isomorphism. Then by Theorem \ref{3.7}(1), we have a monic $\mathcal{A}_C(S)$-preenvelope
$$f^0:N\rightarrowtail A^0$$ of $N$, where $f^0=f_*\mu_N$ and $A^0=B_*$. So we have a
$\Hom_S(-,\mathcal{A}_C(S))$-exact exact sequence
$$0\to N \buildrel {f^0} \over \longrightarrow A^0 \to N^1 \to 0$$
in $\Mod S$, where $N^1=\Coker f^0$. Because $A^0\in{^{\bot}\mathcal{I}_C(S)}$,
we have $N^1\in{^{\bot}\mathcal{I}_C(S)}$. Similar to the above argument, we get a
$\Hom_S(-,\mathcal{A}_C(S))$-exact exact sequence
$$0\to N^1 \buildrel {f^1} \over \longrightarrow A^1 \to N^2 \to 0$$
in $\Mod S$ with $A^1\in\mathcal{A}_C(S)$ and $N^2\in{^{\bot}\mathcal{I}_C(S)}$. Repeating this procedure,
we get a $\Hom_S(-,\mathcal{A}_C(S))$-exact exact sequence
$$0\to N \buildrel {f^0} \over \longrightarrow A^0 \buildrel {f^1} \over \longrightarrow
A^1\buildrel {f^2} \over \longrightarrow \cdots \buildrel {f^i}
\over \longrightarrow A^i\buildrel {f^{i+1}} \over \longrightarrow \cdots$$
in $\Mod S$ with all $A^i$ in $\mathcal{A}_C(S)$. Because $\mathcal{I}_C(S)\subseteq\mathcal{A}_C(S)$
by \cite[Corollary 6.1]{HW07}, this exact sequence is $\Hom_S(-,\mathcal{I}_C(S))$-exact.
By \cite[Theorem 3.11(1)]{TH}, there exists a $\Hom_S(-,\mathcal{A}_C(S))$-exact exact sequence
$$0\to A^i\to U^i_0 \to U^i_1 \to \cdots \to U^i_j\to \cdots$$
in $\Mod S$ with all $U^i_j$ in $\mathcal{I}_C(S)$ for any $i,j\geq 0$. Then by \cite[Corollary 3.9]{Hu},
we get the following $\Hom_S(-,\mathcal{A}_C(S))$-exact exact sequence
$$0\to N\to U^0_0 \to U^0_1\oplus U^1_0 \to \cdots \to \oplus_{i=0}^{n}U^i_{n-i}\to \cdots$$
in $\Mod S$ with all terms in $\mathcal{I}_C(S)$. It follows from \cite[Theorem 3.11(1)]{TH} that
$N\in\mathcal{A}_C(S)$. The proof is finished.
\end{proof}
We use $\pd_{S^{op}}C$ and $\fd_{S^{op}}C$ to denote the projective and flat dimensions of $C_S$ respectively.
\begin{corollary}\label{4.11}
If $_RC$ has an ultimately closed projective resolution,
then the following statements are equivalent for any $n\geq 0$.
\begin{enumerate}
\item[(1)] $\pd_{S^{op}}C\leq n$.
\item[(2)] $\mathcal{A}_{C}(S)$-$\pd_{S}N \leq n$ for any $N\in\Mod S$.
\end{enumerate}
\end{corollary}
\begin{proof}
Assume that $_RC$ has an ultimately closed projective resolution. By Theorem \ref{4.10},
we have $\mathcal{A}_C(S)={{C_S}^{\top}}$. Then
it is easy to see that $C_S$ is flat (equivalently, projective) if and only if $\mathcal{A}_{C}(S)=\Mod S$,
so the assertion for the case $n=0$ follows. Now let $N\in\Mod S$ and $n\geq 1$.
$(2)\Rightarrow (1)$ By (2) and Theorem \ref{4.4}, we have
$\Omega^n(N)\in\mathcal{A}_{C}(S)(\subseteq{C_S}^\top)$. Then by the dimension shifting,
we have $\Tor_{\geq n+1}^S(C,N)=0$, and so $\pd_{S^{op}}C=\fd_{S^{op}}C\leq n$.
$(1)\Rightarrow (2)$ If $\pd_{S^{op}}C\leq n$, then $\Omega^n(N)\in{{C_S}^{\top}}$ by the dimension shifting.
By Theorem \ref{4.10}, we have $\Omega^n(N)\in\mathcal{A}_{C}(S)$ and $\mathcal{A}_{C}(S)$-$\pd_{S}N \leq n$.
\end{proof}
\vspace{0.3cm}
{\bf Acknowledgements.}
This research was partially supported by NSFC (Grant No. 11571164).
|
\section{Introduction}
\label{intro}
Neutrino and antineutrino absorption reactions in nuclear matter play
a crucial role in core-collapse
supernovae~\cite{Reddy.Prakash.Lattimer:1998,Burrows.Reddy.Thompson:2006}.
In delayed neutrino-heating mechanism, the energy transport by
electron neutrinos and antineutrinos from a hot protoneutron star
(PNS) is responsible for reviving the stalled shock-wave and
generating an explosion~\cite{Bethe1985,Janka2012}. The
charged-current reactions $\nu_e + n \to p + e$ and
$\bar\nu_e + p \to n + e^+$ provide a dominant mechanism for heating
the matter behind the shock. The efficiency of neutrino-heating
mechanism and the success of the explosion depend critically on the
luminosity and energy spectra for $\nu_e$ and $\bar\nu_e$. In turn,
these characteristics are determined by the neutrino transport inside
the protoneutron star, and thus depend on the neutrino opacities in
the dense ($\rho\simeq10^{12}-10^{14}~\mathrm{g/cm^3}$) and hot
($T\simeq 5-10~\mathrm{MeV}$) nuclear matter. The main opacity sources
for $\nu_e$ and $\bar\nu_e$ are the same absorption reactions on
nucleons, which are responsible for the neutrino-heating mechanism.
For dense neutron-rich conditions relevant in the protoneutron star,
earlier calculations have found that many-body correlations due to
strong interactions between nucleons can modify the neutrino opacities
significantly. In particular, it was
found~\cite{MartinezPinedo_2012,Roberts_PRC2012} that the difference
in the neutron and proton mean-field potential energies enhances
(suppresses) the (anti)neutrino opacity considerably with respect to
that in the noninteracting gas of free nucleons. Meanwhile,
in~\cite{Roberts_PRC2012} it was shown that random-phase approximation
corrections, accounting for long-range $ph$ correlations, act in the
opposite direction, but their effect is less significant as compared
to that of mean-field potential difference.
It was emphasized in many works (see, for example, Refs.~\cite{Sawyer_PRD11,Sawyer_PRC40,Reddy_PRC59}) that neutrino opacities need to be consistent with the model
of the strong interaction used in the equation of state for protoneutron star matter. Specifically, a particular choice of the particle-hole interaction should be
consistent with the EOS employed to compute the composition of charge-neutral $\beta$-equilibrated matter~\cite{Reddy_PRC59}. In this paper we use the EOS based on the Skyrme
effective interaction to compute $\nu_e$ and $\bar\nu_e$ charged-current opacities in the outer region of PNS called the neutrino-sphere. We employ self-consistent RPA calculations,
i.e., both the Hartree-Fock mean-field and the residual $ph$ interaction are derived from the same Skyrme potential. In this work we will consider five Skyrme parameterizations:
KDE0v1, LNS, NRAPR, SKRA, SQMC700. It was shown in Ref.~\cite{Dutra_PRC85} that these Skyrme forces satisfy a series of experimentally extracted constraints for infinite nuclear matter properties.
The paper is organized as follows. In Sec.~\ref{formalism} we briefly present the Skyrme-RPA formalism to compute the Fermi and Gamow-Teller strength functions which
determine the absorption cross section for $\nu_e$ and $\bar\nu_e$. Our results for the Fermi and Gamow-Teller strength functions as well as for (anti)neutrino opacities are presented in Sec.~\ref{results}. Finally, we conclude in Sec.~\ref{conslusions}.
\section{Formalism}\label{formalism}
The charged-current opacity for $\nu_e$ and $\bar\nu_e$ can be calculated by numerical integration of the absorption differential cross section per unit volume.
In the nonrelativistic limit for nucleons, the double-differential cross-section for a (anti)neutrino with initial energy $E_\nu$ is given by~\cite{Roberts_PRC2012}
\begin{align}\label{CrSect}
& \frac{1}{V}\frac{d^2\sigma}{d\cos\theta\,dE_e}=\frac{G^2_F \cos^2\theta_C}{2\pi}p_e E_e (1-f_e(E_e))
\\ \notag
& \times\bigl[(1+\cos\theta)S^{(\pm)}_0(\omega, q) + g^2_A (3 - \cos\theta)S^{(\pm)}_1(\omega, q) \bigr].
\end{align}
The energy transfer to the nuclear medium is $\omega = E_\nu - E_e$,
while the momentum transfer is $q=|\bm{q}|=|\bm{p}_\nu - \bm{p}_e|$.
$S^{(\pm)}_0$ and
$S^{(\pm)}_1$ are the strength functions (also called dynamical
structure factors) associated with the Fermi and Gamow-Teller
operators, $\bm{\tau}^{(\pm)}$ and $\bm{\sigma\tau}^{(\pm)}$,
respectively. Note, that in Eq.~\eqref{CrSect} the plus sign refers to
a neutrino absorption, while the minus sign refers to an antineutrino
absorption. Thus, for (anti)neutrino absorption the cross section
includes Fermi and Gamow-Teller strength functions for $n\to p$
($p\to n$) transitions. Finally, $1-f_e(E_e)$ is the Pauli blocking
factor for the outgoing lepton. In the protoneutron star matter
positrons are nondegenerate and their spectrum approaches the
Boltzmann distribution. Therefore, we can neglect the final-state
blocking for antineutrino absorption.
The strength functions $S^{(\pm)}_{0,1}$ embody all spatial and
temporal correlations between nucleons arising from strong
interaction. At $T\ne 0$ they include summation of transition
probabilities from thermally excited states $|i\rangle$ to final
states $|f\rangle$ such that $\omega = E_f-E_i$:
\begin{align}
S^{(\pm)}_\alpha(q,\omega) = \sum_{i,f} P_i|\langle f |\bm{Q}^{(\pm)}_\alpha| i \rangle|^2\delta(E_f-E_i - \omega),
\end{align}
with $P_i=Z^{-1}\mathrm{e}^{-E_i/T}$ being the probability of finding
the system in the state $|i\rangle$, and $Z=\sum_ie^{-E_i/T}$
is the partition function. The one-body transition operators
$\bm{Q}^{(\pm)}_\alpha$ are defined as $\bm{Q}^{(\pm)}_0 = \sum_j e^{i\bm{q}\bm{r}_j}\bm{\tau}^{(\pm)}_j$, $ \bm{Q}^{(\pm)}_1 = \sum_j e^{i\bm{q}\bm{r}_j}\bm{\sigma}_j\bm{\tau}^{(\pm)}_j,$
where the index $j$ stands for the nucleon. The strength functions obey the detailed balance theorem~\cite{Hernandez_NPA658}
\begin{equation}\label{DB}
S^{(\pm)}_\alpha(q,\omega) = e^{(\omega \pm \Delta \mu_{np})/T} S^{(\mp)}_\alpha (q,-\omega),
\end{equation}
where $\Delta \mu_{np} = \mu_n - \mu_p$ is the difference between the neutron and proton chemical potentials,
as well as they satisfy the non-energy weighted sum rule~\cite{Hernandez_NPA658}
\begin{equation}\label{sum_rule}
\int^{+\infty}_{-\infty} d\omega\bigl[ S^{(+)}_\alpha(q,\omega) - S^{(-)}_\alpha(q,\omega)\bigr] = \rho_n - \rho_p,
\end{equation}
which depends only on the difference between neutron and proton densities.
We obtain the strength functions from the imaginary part of the response function $\Pi^{(\alpha)}(q,\omega) $ according to the fluctuation-dissipation theorem
\begin{equation}
S^{(\pm)}_\alpha(q,\omega) = -\frac{1}{\pi}\frac{\mathfrak{Im}
\Pi^{(\pm)}_\alpha(q,\omega)}{1-\mathrm{e}^{-(\omega \pm \Delta \mu_{np})}}.
\end{equation}
In its turn $\Pi^{(\pm)}_\alpha$ is related to the particle-hole propagator (or the retarded $ph$ Green's function)
\begin{equation}
\Pi^{(\pm)}_\alpha(q,\omega) = 2 \int\frac{d^3\bm{k}}{(2\pi)^3} G^{(\pm)}_\alpha(k,q,\omega).
\end{equation}
Here and below, for the momentum averages we adopt the notation
\begin{equation}
\int\frac{d^3\bm{k}}{(2\pi)^3} f(\bm{k})G^{(\pm)}_\alpha(k,q,\omega) \equiv\langle f G^{(\pm)}_\alpha\rangle.
\end{equation}
With this notation $ \Pi^{(\pm)}_\alpha(q,\omega) = 2\langle G^{(\pm)}_\alpha\rangle$.
Neglecting long-range $ph$ correlations $G^{(\pm)}_\alpha = G^{(\pm)}_{HF} $, and the Hartree-Fock particle-hole propagator is given by
\begin{equation}
G^{(\pm)}_{HF}(k, q, \omega) = \frac{ f_{n(p)}(\bm{k}) -f_{p(n)}(\bm{k} + \bm{q})}{\omega + E_{n(p)}(\bm{k}) - E_{p(n)}(\bm{k} + \bm{q}) + i\eta }.
\end{equation}
Here $f_\tau(\bm{k})=\bigl[1+\mathrm{e}^{(E_\tau(\bm{k})-\mu_\tau)/T}\bigr]^{-1}$ is the Fermi-Dirac distribution for nucleons and $E_\tau(\bm{k})$ is the single particle energy given
by
\begin{equation}\label{disp}
E_\tau(\bm{k}) = \frac{\bm{k}^2}{2m_\tau} + U_\tau,
\end{equation}
where $m_\tau$ is the effective mass and $U_\tau$ is the Hartree-Fock potential obtained with the Skyrme interaction.
Using the detailed balance~\eqref{DB} we can write the Hartree-Fock strength function for $n\to p$ ($p\to n$) transitions as
\begin{align}\label{S_HF}
S^{(\pm)}_{HF} (q,\omega)= 2\int \frac{ d^3\bm{k}}{(2\pi)^3}\,\delta(E_{p(n)} - E_{n(p)}-\omega)
\notag\\
\times f_{n(p)}( \bm{k}) (1-f_{p(n)}(\bm{k} + \bm{q})).
\end{align}
Assuming $U_\tau=0$ and considering the bare nucleon masses we get the strength function for a noninteracting Fermi gas.
To go beyond the HF approximation we take into the residual proton-neutron interaction by resumming a class of so-called ring diagrams and obtain the well known RPA~\cite{Fetter}.
The RPA correlated $ph$ propagator is the solution of the Bethe-Salpeter integral equation
\begin{align}\label{BS_eq}
G^{(\pm)}_{\alpha}(k_1, q, \omega) = G^{(\pm)}_{HF} (k_1, q, \omega) + G^{(\pm)}_{HF} (k_1, q, \omega)
\notag\\
\times\sum_{\alpha'}\int\frac{ d^3\bm{k_2}}{(2\pi)^3} V^{(\alpha,\alpha')}_{ph}(\bm{q},\bm{k}_1,\bm{k}_2) G^{(\pm)}_{\alpha'}(k_2, q, \omega),
\end{align}
where $V^{(\alpha,\alpha')}_{ph}$ is the residual interaction matrix element. For the Skyrme effective interaction containing a zero-range spin-orbit term this proton-neutron matrix element
takes the form~\cite{Davesne_PRC89}
\begin{multline}
V_{ph}^{(SM;S'M')}(\bm{q},\bm{k}_1,\bm{k}_2) \equiv
\\
\langle \bm{q}+\bm{k}_1,\bm{k}_1^{-1};SM|V_{ph}|\bm{q}+\bm{k}_2,\bm{k}_2^{-1};S'M'\rangle =
\\
\delta_{SS'}\delta_{MM'}\bigl(W^{(S)}_1 + W^{(S)}_2\bm{k}^2_{12} \bigl)
\notag \\
+ q W^{(1)}_{SO}\bigl(\delta_{S'0}\delta_{S1}M(k_{12})^{(1)}_{-M} + \delta_{S'1}\delta_{S0}M'(k_{12})^{(1)}_{M'}\bigr),
\end{multline}
where $\bm{k}_{12} = \bm{k}_1 - \bm{k}_2$ is the relative hole
momentum, while the rank-1 tensor $(k_{12})^{(1)}_{M}$ is defined
in~\cite{Davesne_PRC89}. The coefficients $W^{(S)}_{1,2}$ are the
combinations of the Skyrme parameters $(x_i, t_i)$, the transferred
momentum $q$ and the total density $\rho=\rho_n+\rho_p$. Their
detailed expressions are given in~\cite{Hernandez_NPA658}. It is
important to note that the spin-orbit term introduces a coupling
between $S=0$ and $S=1$ spin channels and removes the degeneracy on
the spin projection $M$.
Following the method suggested by Margueron~\textit{et
al}~\cite{Margueron_PRC74}, a closed algebraic system for
$\langle G^{(\pm)}_\alpha\rangle$ is obtained by multiplying the
Bethe-Salpeter equation~\eqref{BS_eq} with the functions $k^2_1$,
$k_1 Y_{1\mu}$ and integrating over the momentum. Finally, the system
of algebraic equations can be written in a compact form for each
channel ($\alpha=S,M$) as
\begin{align}\label{BS1}
&(1-\widetilde W_1^{(\alpha)}\beta_0^{(\pm)} - W_2^{(\alpha)} q^2\beta_2^{(\pm)})\langle G^{(\pm)}_\alpha\rangle
\notag\\
&- W_2^{(\alpha)}\beta_0^{(\pm)}\langle k^2 G^{(\pm)}_\alpha\rangle+ 2 W_2^{(\alpha)}q\beta_1^{(\pm)}
\notag\\
&\times\sqrt{\frac{4\pi}{3}}\langle k Y_{10} G^{(\pm)}_\alpha\rangle = \beta_0^{(\pm)},
\end{align}
\begin{align}\label{BS2}
-(&\widetilde W_1^{(\alpha)}q^2\beta_2^{(\pm)} + W_2^{(\alpha)} q^4\beta_5^{(\pm)})\langle G^{(\pm)}_\alpha\rangle
\notag\\
&+(1-W_2^{(\alpha)}q^2\beta_2^{(\pm)})\langle k^2 G^{(\pm)}_\alpha\rangle + 2 W_2^{(\alpha)}q^3\beta_4^{(\pm)}
\notag\\
&\times\sqrt{\frac{4\pi}{3}}\langle k Y_{10} G^{(\pm)}_\alpha\rangle = q^2\beta_2^{(\pm)},
\end{align}
\begin{align}\label{BS3}
-(&\widetilde W_1^{(\alpha)}q\beta_1^{(\pm)} + W_2^{(\alpha)} q^3\beta_4^{(\pm)})\langle G^{(\pm)}_\alpha\rangle
\notag\\
&- W_2^{(\alpha)}q\beta_1^{(\pm)}\langle k^2 G^{(\pm)}_\alpha\rangle + (1+2 W_2^{(\alpha)}q^2\beta_3^{(\pm)})
\notag\\
&\times\sqrt{\frac{4\pi}{3}}\langle k Y_{10} G^{(\pm)}_\alpha\rangle = q\beta_1^{(\pm)}.
\end{align}
The parameter $\widetilde W_1^{(\alpha)}$ describes the coupling between spin channels induced by the
spin-orbit interaction
\begin{equation}
\widetilde W_1^{(\alpha)} = W_1^{(\alpha)} + C^{(\alpha)}\frac{W^2_{SO}q^4(\beta_2^{(\pm)}-\beta_3^{(\pm)})}{1+W_2^{(\alpha')}q^2(\beta_2^{(\pm)}-\beta_3^{(\pm)})}.
\end{equation}
Here $\alpha'$ is defined with respect to $\alpha$ as $S'=1-S$, while $C^{(\alpha)}=1$ if $S=0$ and $C^{(\alpha)}=\frac12 M^2 $ if $S=1$.
The six functions $\beta^{(\pm)}_l(q,\omega),~(l=0,5)$ include the momentum averages of the HF propagator $G^{(\pm)}_{HF}$ and their
explicit expressions are given ~\cite{Hernandez_NPA658}.
If we replace $\widetilde W_1^{(\alpha)}$ in Eqs.~(\ref{BS1}-\ref{BS3}) by $W_1^{(\alpha)}$ we obtain the results of Ref.~\cite{Hernandez_NPA658}, related to the central part of the interaction,
as it should be. We also would like to note that for ambient conditions (density and temperature) considered in the present work, the effect of the spin-orbit interaction is negligible and
for Gamow-Teller strength function we have $S^{(\pm)}_{S=1,M=0}\approx S^{(\pm)}_{S=1,M=1}\equiv S^{(\pm)}_{S=1}$.
\section{Results}\label{Results}\label{results}
To illustrate the general features of the strength functions we have
chosen the LNS Skyrme parametrization. The temperature evolution of
Fermi and GT strength functions is shown in Fig.~\ref{fig-1} for both
$n\to p$ and $p\to n$ transition at a small momentum transfer
$q=0.2$~fm$^{-1}$. We have performed calculations at density
$\rho=0.02$~fm$^{-3}$ for a proton fraction
$Y_p=\rho_p/\rho = 0.1$. As a reference, the HF and noninteracting
Fermi gas (FG) strength functions -- which are independent of the spin
channel - are also shown.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{fig1}
\caption{The finite temperature ($T=2,\,4,8$~MeV) HF and RPA strength functions at $q=0.2\,\mathrm{fm}^{-3}$ as functions of energy $\omega$ for $n\to p$ (left panels) and $p\to n$ (right panels) transitions. The dashed curves give the strength function for the noninteracting Fermi gas (FG). Note, that the strength functions are scaled by the factor shown inside each panel.}
\label{fig-1}
\end{figure}
Comparing the noninteracting $n\to p$ FG strength functions with the HF ones we find that they have the same shape, but the interaction shifts the HF strength function to larger negative energy transfer. It means
that the energy of the nuclear system is decreased for $n\to p$ transitions. In contrast, the interaction pushes the HF $p\to n$ strength function to higher energy transfer,
thereby increasing the energy of the nuclear system. This is understood as follows. The dispersion relation for the nucleon is given by Eq.~\eqref{disp} and the energy shift $\Delta U_{np}=U_n-U_p$ is closely related to the symmetry energy, which describes how the energy of nuclear matter changes as one moves away
from the equal number of protons and neutrons. For the LNS Skyrme interaction the potential difference at considered neutron-rich conditions
is $\Delta U_{np}\approx 12$~MeV. Thus the proton is more strongly bound than the neutron and the difference in potentials serves as a threshold for the $p\to n$ reactions.
Allowing for repulsive $ph$ interaction at low temperatures
significantly suppresses the continuum part of the HF strength
functions and gives rise to well-defined Fermi and Gamow-Teller
collective modes either at positive or negative energy above the
particle-hole band. Moreover, due to the detailed balance~\eqref{DB},
each collective mode peak in the $n\to p$ strength function has its
image in the $p \to n$ strength function at $-\omega$. One important
point is that these $p\to n$ collective modes appear only due to
thermal effects and they are located below the particle-hole band.
The centroid of collective excitations is almost stable against
thermal effects, while their width grows with increasing temperature
due to Landau damping. At high temperatures collective modes become
absorbed into thermally extended particle-hole band.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{fig2}
\caption{Finite temperature ($T=8$~MeV) strength functions calculated with the LNS Skyrme potential at three values of momentum transfer. The density and the proton fraction are the same as in Fig.~\ref{fig-1} }
\label{fig-2}
\end{figure}
Figure~\ref{fig-2} shows a typical modification of the Fermi and
Gamow-Teller strength functions, $S^{(\pm)}_0$ and $S^{(\pm)}_{1}$,
with increasing momentum transfer. On the same figure the HF strength
functions are shown for reference. As we have mentioned before and as
it follows from Eq.~\eqref{S_HF}, the evolution of the HF strength
functions for $n\to p$ ($p\to n$) transitions follows that of the
noninteracting Fermi gas, but shifted by $\Delta U_{np}\approx 12$~MeV
to lower (higher) energies. In particular, we can observe the
enlarging the particle-hole band with increasing momentum transfer $q$
following the overall reduction of the strength. The sum
rule~\eqref{sum_rule} requires that for neutron-rich conditions the
momentum increase has a smaller effect on the total $n\to p$ strength
than on that for the $p\to n$ transitions.
The middle and lower panels of Fig.~\ref{fig-2} show that the effects of increasing momentum are remarkably similar for the Fermi and Gamow-Teller strength functions.
At low momentum transfer the $n\to p $ ($p\to n$) RPA strength functions in both the $S=0$ and $S=1$ channels have relatively narrow peaks above (below) the maximum
of the HF strength. When $q$ increases, the peaks become broader and finally the RPA strength evolves to a structureless shape. Furthermore, although at large $q$ the $ph$ correlations
have a relatively smaller effect than at small $q$, the $n\to p $ RPA strength functions remain shifted to larger energies with respect to the HF strength. In contrast, because of diminishing contribution from thermally excited collective modes, the maximums of the $p\to n $ Fermi and Gamow-Teller strength functions move from low energies to higher ones and finally they appear above the HF maximum.
\begin{figure*}
\centering
\includegraphics[width=\textwidth,clip]{fig3}
\caption{Differential cross sections for neutrino (upper panels) and antineutrino (lower panels) absorption, evaluated at beta equilibrium conditions
$\rho=0.02\,\mathrm{fm}^{-3}$, $T=8\,\mathrm{MeV}$, and $Y_p = 0.029$. The Skyrme force LNS is used in the HF and RPA calculations.
In each panel the dashed line shows the noninteracting Fermi gas cross sections. For $E_{\bar\nu_e}=10$ antineutrinos the HF cross section is scaled by the factor $10^4$. }
\label{fig-3}
\end{figure*}
The differential absorption cross sections for neutrinos and antineutrinos as a function of the outgoing lepton energy are shown in Fig.~\ref{fig-3}.
We consider incoming neutrino energies $E_{\nu,\bar\nu}=10,~20,~50$ and $100$~MeV and the conditions of the medium are $\rho=0.02\,\mathrm{fm}^{-3}$,
$T=8\,\mathrm{MeV}$. The proton fraction and the electron chemical potential are $Y_p = 0.029$ and $\mu_e=47.05\,\mathrm{MeV}$, respectively, which correspond to beta-equilibrium (i.e., $\mu_e=\mu_n-\mu_p$) for the given temperature, density and the assumed
nuclear Skyrme interaction LNS. The mean-field potential difference is $\Delta U_{np}=14.66\,\mathrm{MeV}$.
It is seen from the figure, the peak of the differential cross section
is shifted by about $\Delta U_{np}$ to higher (lower) energies for
(anti)neutrino absorption with respect to the noninteracting Fermi gas
model and, hence, the mean-field potential contribution increases
(reduces) the energy of the emitted electron (positrons). This
observation is in line with the elastic
approximation~\cite{Reddy.Prakash.Lattimer:1998} where the energy of
the outgoing electron (positron) is $E_e=E_\nu +\Delta U_{np}$
($E_{e^+}=E_{\bar\nu} -\Delta U_{np}$). For low energy neutrinos,
$E_\nu<\mu_e$, the additional energy $\Delta U_{np}$ is not enough to
overcome the Pauli blocking for the outgoing electron. However, this
energy is sufficient to put the outgoing electron in a less blocked
portion of phase space, thereby leading to an exponential enhancement
of the cross section. For high-energy neutrinos, $E_\nu>\mu_e$, the
cross section is approximately proportional to the phase space factor
$p_eE_e$ and the effects of the mean field contribution diminish with
increasing $E_\nu$.
For the antineutrino HF cross section the important difference from the neutrino case comes from the fact that the reaction threshold is increased by $\Delta U_{np}$. The increased reaction threshold leads to a significant suppression of the HF cross section for low-energy
antineutrinos. For $E_{\bar\nu} < \Delta U_{np} $ antineutrinos the HF cross section differs from zero due to thermal effects which smear the proton and neutron Fermi surfaces.
As was stated above, we can neglect the Pauli blocking for the outgoing positron and for $E_{\bar\nu} > \Delta U_{np}$ antineutrinos the
HF cross section is suppressed solely by the reduced available phase space.
Let us now consider the role of RPA correlations on the differential cross sections. It is seen from the upper panels of Fig.~\ref{fig-3} that RPA correlations suppress the neutrino absorption cross section. The degree of suppression depends on neutrino energy and it is most significant for low-energy neutrinos when the transferred momentum is also low and the RPA strength functions are dominated by the resonance contributions (see~Fig.~\ref{fig-2}). The Gamow-Teller and Fermi resonances are clearly visible in the differential cross section for low-energy neutrinos, however their contribution is strongly suppressed by the phase-space blocking for outgoing electrons. With increasing neutrino energy, the contribution of collective states to the $\nu_e$ cross section decreases and the RPA corrections become smaller. Concerning $\bar\nu_e$ absorption, for low-energy antineutrinos the HF cross section is almost zero due to reaction threshold. However, RPA correlations and thermal effects give rise to collective peaks which decrease the reaction threshold and enhance the cross section significantly. The contribution of collective states diminishes with increasing antineutrino energy and the RPA corrections tend to reduce the cross section as in the $\nu_e$ case.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{fig4}
\caption{Neutrino (left panel) and antineutrino (right panel)
opacities, evaluated at beta-equilibrium conditions
$\rho = 0.02\,\mathrm{fm}^{-3}$, $T=8\,\mathrm{MeV}$, and
$Y_p=0.029$ with the Skyrme force LNS.}
\label{fig-4}
\end{figure}
Figure~\ref{fig-4} shows neutrino and antineutrino opacities evaluated
at the same conditions considered in Fig.~\ref{fig-3}. The results
shown in Fig.~\ref{fig-4} follow the trends expected from the results
for the differential cross sections. Namely, due to presence of the
mean-field potentials the opacity for low-energy neutrinos increases
with respect to the noninteracting case, but at larger neutrino
energies the mean-field contribution becomes less important and the
opacities computed with and without mean-field potentials approach
each other asymptotically. In contrast, the antineutrino HF opacity is
reduced relative to the Fermi gas result and the reaction threshold
manifests itself at $\Delta U_{np}$. For incoming (anti)neutrino
energy $E_\nu=24$~MeV which is the mean thermal energy $E_\nu\sim 3T$
at temperature $T=8$~MeV the HF neutrino opacity is enhanced by about
an order of magnitude compared to the noninteracting Fermi gas, while
the antineutrino opacity is suppressed by almost a factor of
$5$. While mean-field effect tends to increase the neutrino opacity,
RPA correlations decrease it. For $E_\nu=24$~MeV neutrinos, the RPA
opacity is suppressed by almost a factor of 4 relative to the HF
one. For antineutrino absorption, the inclusion of RPA correlations
decreases the reaction threshold and increases opacity for low-energy
neutrinos. However, at higher neutrino energies RPA correlations
reduces the opacity.
We now compare the opacities evaluated with different Skyrme forces mentioned in the Introduction. Results for the parametrizations KDE0v1, LNS, NRAPR, SKRPA, and SQMC700 are shown in Fig.~~\ref{fig-5}.
The RPA opacities do not differ significantly between the last four Skyrme parametrizations. In contrast,
the HF calculations based on the KDE0v1 force predict
a mean-field potential difference $\Delta U_{np}=20.7\,\mathrm{MeV}$ which is substantially larger than that for other four parametrizations. As a result, low-energy neutrino (antineutrino) opacities evaluated with the KDE0v1 force turn out to be well above (below) those expected from other Skyrme forces.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{fig5}
\caption{Opacity for neutrino and antineutrino, evaluated at conditions $\rho = 0.02\,\mathrm{fm}^{-3}$, $T=8\,\mathrm{MeV}$, and $Y_p=0.029$ with the different Skyrme parameterizations.}
\label{fig-5}
\end{figure}
\section{Conclusions}\label{conslusions}
In this paper, we have studied the combined effects of the mean-field
potential difference and RPA correlations on the $\nu_e$ and
$\bar\nu_e$ opacity in the matter of protoneutron stars. By solving
the Bethe-Salpeter equation with the residual particle-hole
interaction derived from a Skyrme potential, we estimated the Fermi
and Gamow-Teller strength functions which are directly related to the
(anti)neutrino absorption cross section. Strength functions are
analyzed for varying conditions of momentum transfer and temperature
for a representative Skyrme potential LNS. It is shown that both
mean-field effects and RPA correlations play a significant role in the
strength function redistribution relative to the noninteracting Fermi
gas. In agreement with the previous
studies~\cite{MartinezPinedo_2012,Roberts_PRC2012}, the mean-field
corrections enhance the neutrino opacity and suppress the antineutrino
opacity. Our calculations, however, indicate that RPA correlations
may significantly suppress the HF neutrino opacity at thermal neutrino
energies in agreement with previous studies~\cite{Reddy_PRC59,Burrows.Sawyer:1999}. For antineutrino
absorption the relative contribution of RPA corrections is energy
dependent: for near threshold antineutrinos they enhance the opacity,
while for higher energies they act in the opposite direction. The
joint effect of mean-field energy shift and long-range RPA
correlations make the protoneutron star matter more opaque
(transparent) for (anti)neutrino radiation as compared to the
noninteracting Fermi gas model.
\begin{acknowledgement}
This work was supported by the Heisenberg-Landau Program. G.M.-P. is
partly supported by the Deutsche Forschungsgemeinschaft through
contract SFB~1245.
\end{acknowledgement}
|
\section{Introduction}
Producing sentences which are perceived as natural by a human
addressee---a property which we will denote as \textit{fluency}\footnote{Alternative names include \textit{naturalness},
\textit{grammaticality} or \textit{readability}.
Note that the exact definitions of all those terms vary slightly throughout the literature.}
throughout this
paper
---is a crucial goal of all natural language generation (NLG) systems:
it makes interactions more natural, avoids misunderstandings and,
overall, leads to higher user satisfaction and user trust \cite{usertrust2018}.
Thus,
fluency evaluation
is important, e.g., during system development,
or for filtering unacceptable generations at application time.
However, fluency evaluation of NLG systems constitutes a hard challenge:
systems
are often not limited to reusing words from the input, but can generate
in an \textit{abstractive} way. Hence, it is not guaranteed that a correct output will match any of a finite number of given references.
This results in difficulties for current reference-based evaluation,
especially of fluency,
causing word-overlap metrics like ROUGE \cite{lin-och:2004:ACL}
to correlate only weakly with human
judgments \cite{toutanova2016dataset}.
As a result, fluency evaluation of NLG is often
done manually, which is costly and time-consuming.
\begin{table}[t!]
\centering
\setlength{\tabcolsep}{4.pt}
\begin{tabular}{| l c |}\hline
If access to a synonym dictionary is & \\
likely to be of use, then this package may & 3 \\
be of service. & \\\vspace{-0.14cm}
& \\
Participants are invited to submit a set & \\
pair do domain name that is already & 1.6 \\
taken along with alternative. & \\\vspace{-0.14cm}
&\\
Even \$15 was The HSUS. & 1 \\ \hline
\end{tabular}
\caption{Example compressions from our dataset with their fluency scores; scores in $[1, 3]$, higher is better.\label{tab:example}}
\end{table}
Evaluating sentences on their fluency, on the other hand, is a linguistic ability of humans which has been the subject
of a decade-long debate in cognitive science.
In particular,
the question has been raised whether the grammatical knowledge that underlies
this ability is probabilistic
or categorical in nature \cite{chomsky2002syntactic,manning2003probabilistic,sprouse2007continuous}.
Within this context,
\newcite{lau2017grammaticality}
have recently shown that neural language models (LMs)
can be used for modeling human ratings of acceptability.
Namely, they found
SLOR \cite{pauls2012large}---sentence log-probability which is normalized by
unigram log-probability and sentence length---to correlate
well with acceptability judgments at the sentence level.
However, to the best of our knowledge, these insights have so far gone disregarded by the natural language processing (NLP) community.
In this paper, we investigate the practical implications of \newcite{lau2017grammaticality}'s findings for fluency evaluation of NLG, using the task of automatic compression \cite{Knight:2000:SSS:647288.721086,mcdonald2006discriminative}
as an example (cf. Table \ref{tab:example}).
Specifically, we test our hypothesis that
SLOR should be a suitable metric for
evaluation of compression fluency which
(i) does not rely on references;
(ii) can naturally be applied at the sentence level (in contrast to the system level); and
(iii) does not need human fluency annotations of any kind.
In particular the first aspect, i.e., SLOR not needing references, makes it a promising candidate for automatic evaluation. Getting rid of human references has practical importance in a variety of settings, e.g., if references are unavailable due to a lack of resources for annotation, or if obtaining references is impracticable. The latter would be the case, for instance, when filtering system outputs at application time.
We further introduce WPSLOR, a novel, WordPiece \cite{wu2016google}-based version of SLOR, which drastically reduces model size and training time.
Our experiments show that both approaches correlate better with human judgments
than traditional word-overlap metrics, even though the latter do rely on reference compressions.
Finally,
investigating the case of available references and how to incorporate them, we combine WPSLOR and ROUGE
to ROUGE-LM, a novel reference-based metric,
and increase the correlation with human fluency ratings even further.
\paragraph{Contributions. }
To summarize, we make the following contributions:
\begin{enumerate}
\item We empirically show that SLOR is a good referenceless metric for the evaluation of NLG fluency at the sentence level.
\item We introduce WPSLOR, a WordPiece-based version of SLOR, which disposes of a more compact LM without a significant loss of performance.
\item We propose ROUGE-LM, a reference-based metric, which achieves a significantly higher correlation with human fluency judgments than all other metrics in our experiments.
\end{enumerate}
\section{On Acceptability}
Acceptability judgments, i.e., speakers' judgments of the well-formedness of sentences, have been the basis of much linguistics research \cite{chomsky1964aspects,schutze2016empirical}: a speakers intuition about a sentence is used to draw conclusions about a language's rules.
Commonly, ``acceptability'' is used synonymously with ``grammaticality'', and speakers are in practice asked for grammaticality judgments or acceptability judgments interchangeably. Strictly speaking, however, a sentence can be unacceptable, even though it is grammatical -- a popular example is Chomsky's phrase ``Colorless green ideas sleep furiously.'' \cite{chomsky2002syntactic} In turn, acceptable sentences can be ungrammatical, e.g., in an informal context or in poems \cite{newmeyer1983grammatical}.
Scientists---linguists, cognitive scientists, psychologists, and NLP researcher alike---disagree about how to represent
human linguistic abilities. One subject of debates are
acceptability judgments: while, for many, acceptability is a binary condition on membership in a set of well-formed sentences \cite{chomsky2002syntactic}, others assume that it is gradient in nature \cite{heilman2014predicting,toutanova2016dataset}.
Tackling this research question, \newcite{lau2017grammaticality}
aimed at modeling human acceptability judgments automatically, with the goal to gain insight into the nature of
human perception of acceptability. In particular, they tried to answer the question: Do humans judge acceptability on a gradient scale?
Their experiments showed a strong correlation between human
judgments and
normalized sentence log-probabilities under a variety of LMs for artificial data they had created by translating and back-translating
sentences with neural models. While they tried different types of LMs, best results were obtained for neural models, namely recurrent neural networks (RNNs).
In this work, we investigate if approaches which have proven successful for modeling acceptability can
be applied to the NLP problem of automatic fluency evaluation.
\section{Method}
In this section, we first describe SLOR and the intuition behind this score. Then, we introduce WordPieces, before explaining how we combine the two.
\subsection{SLOR}
SLOR assigns to a sentence $S$ a score which consists of its log-probability under a given LM,
normalized by unigram log-probability and length:
\begin{align}
\text{SLOR}(S) = &\frac{1}{|S|} (\ln(p_M(S)) \\\nonumber
&- \ln(p_u(S)))
\end{align}
where $p_M(S)$ is the probability assigned to the sentence under the LM.
The unigram probability $p_u(S)$ of the sentence is calculated as
\begin{equation}
p_u(S) = \prod_{t \in S}p(t)
\end{equation}
with $p(t)$ being the unconditional probability of a token $t$, i.e., given no context.
The intuition behind subtracting unigram log-probabilities is that a token which is rare on its own
(in contrast to being rare at a given position in the sentence)
should not bring down the sentence's rating.
The normalization by sentence length is necessary in order to not prefer shorter sentences over equally fluent longer ones.\footnote{Note that the sentence log-probability which is normalized by sentence length corresponds to the negative cross-entropy.}
Consider, for instance, the following pair of sentences:
\begin{align}
\textrm{(i)} ~ ~ &\textrm{He is a citizen of France.}\nonumber\\
\textrm{(ii)} ~ ~ &\textrm{He is a citizen of Tuvalu.}\nonumber
\end{align}
Given that both sentences are of equal length and assuming that France appears more often in a given LM training set than Tuvalu, the length-normalized log-probability of sentence (i) under the LM would most likely be higher than
that of sentence (ii). However, since both sentences are equally fluent, we expect
taking each token's unigram probability into account to lead to a more suitable score for our purposes.
We calculate the probability of a sentence with a long-short term memory (LSTM, \mbox{\newcite{hochreiter1997long}}) LM, i.e., a special type of RNN LM, which has been trained on a large
corpus. More details on LSTM LMs can be found, e.g., in \newcite{sundermeyer2012lstm}.
The unigram probabilities for SLOR are estimated using the same corpus.
\subsection{WordPieces}
Sub-word units like WordPieces \cite{wu2016google} are getting increasingly important in NLP.
They
constitute a compromise between characters and words: On the one hand,
they yield a smaller vocabulary, which reduces model size and
training time, and improve handling of rare words, since those are partitioned into
more frequent segments.
On the other hand, they contain more information than characters.
WordPiece models are estimated
using a data-driven approach which maximizes the LM likelihood
of the training corpus as described in \newcite{wu2016google} and \newcite{6289079}.
\begin{table}[t!]
\centering
\setlength{\tabcolsep}{4.5pt}
\begin{tabular}{l | c c c c }
&\textbf{ILP} & \textbf{NAMAS} & \textbf{SEQ2SEQ} & \textbf{T3} \\\hline
fluency & 2.22 & 1.30 & 1.51 & 1.40\\
\end{tabular}
\caption{\label{tab:system_average} Average fluency ratings for each compression system in the dataset by \newcite{toutanova2016dataset}.}
\end{table}
\subsection{WPSLOR}
We propose a novel version of SLOR, by incorporating a LM which is trained on a corpus which has been split by a
WordPiece\footnote{https://github.com/google/sentencepiece} model. This leads to a smaller vocabulary, resulting in a LM with less parameters, which is faster to train (around 12h compared to roughly 5 days for the word-based version in our experiments).
We will refer to the word-based SLOR as WordSLOR and to our newly proposed WordPiece-based version as WPSLOR.
\section{Experiment}
Now, we present our main experiment, in which we assess the performances of WordSLOR and WPSLOR as fluency evaluation metrics.
\subsection{Dataset}
We experiment on the compression dataset by \newcite{toutanova2016dataset}.
It contains single sentences and two-sentence paragraphs from
the Open American National Corpus (OANC),
which belong to 4 genres: \textit{newswire}, \textit{letters}, \textit{journal}, and \textit{non-fiction}.
Gold references are manually created
and the outputs of 4
compression systems
(ILP (extractive), NAMAS (abstractive), SEQ2SEQ (extractive), and T3 (abstractive);
cf. \newcite{toutanova2016dataset} for details) for the test data
are provided. Each example has 3 to 5 independent human ratings for content and fluency.
We are interested in the latter, which is rated
on an ordinal scale from 1 (disfluent) through 3 (fluent).
We experiment on the $2955$ system outputs for the test split.
Average
fluency scores per system are shown in Table \ref{tab:system_average}.
As can be seen, ILP produces the best output.
In contrast, NAMAS is the worst system for fluency.
In order to be able to judge the reliability of the human annotations, we follow the procedure suggested by \newcite{TACL732} and used by \newcite{toutanova2016dataset}, and compute the quadratic weighted $\kappa$ \cite{cohen1968weighted} for the human fluency scores of the system-generated compressions as $0.337$.
\subsection{LM Hyperparameters and Training}
We train our LSTM
LMs on the English Gigaword corpus \mbox{\cite{parker2011english}}, which consists of news data.
The hyperparameters of all LMs are tuned using perplexity on a held-out part of Gigaword, since we expect LM perplexity and final evaluation performance of WordSLOR and, respectively, WPSLOR to correlate.
Our best networks consist of two layers with 512 hidden units each, and are trained for $2,000,000$
steps with a minibatch size of 128. For optimization, we employ ADAM \mbox{\cite{kingma2014adam}}.
\subsection{Baseline Metrics}
\paragraph{ROUGE-L. }
Our first baseline is ROUGE-L \cite{lin-och:2004:ACL},
since it is the most commonly used metric for compression tasks. ROUGE-L
measures the similarity of two sentences based on their longest common subsequence.
Generated and reference compressions are tokenized and lowercased.
For multiple references, we only make use of the one with the highest score for each example.
\paragraph{N-gram-overlap metrics. }
We compare to the best n-gram-overlap metrics from \newcite{toutanova2016dataset};
combinations of linguistic units (bi-grams (LR2) and tri-grams (LR3)) and scoring measures (recall (R) and F-score (F)).
With multiple references, we consider the union of the sets of n-grams.
Again, generated and reference compressions are tokenized and lowercased.
\paragraph{Negative cross-entropy. } We further compare to the negative LM cross-entropy, i.e., the log-probability which is only normalized by sentence length.
The score of a sentence $S$ is calculated as
\begin{equation}
\text{NCE}(S) = \tfrac{1}{|S|} \ln(p_M(S))
\end{equation}
with $p_M(S)$ being the probability assigned to the sentence by a LM.
We employ the same LMs as for SLOR, i.e., LMs trained on words (WordNCE)
and WordPieces (WPNCE)
\paragraph{Perplexity.} Our next baseline is perplexity,
which corresponds to the exponentiated cross-entropy:
\begin{equation}
\text{PPL}(S) = \exp(-\text{NCE}(S))
\end{equation}
\paragraph{About BLEU.} Due to its popularity, we also performed initial experiments with BLEU \cite{papineni-EtAl:2002:ACL}. Its correlation with human scores was so low that we do not consider it
in our final experiments.
\subsection{Correlation and Evaluation Scores}
\paragraph{Pearson correlation. }
Following earlier work \cite{toutanova2016dataset}, we evaluate our metrics using Pearson correlation with human judgments.
It is defined as the covariance divided by the product of the standard deviations:
\begin{equation}
\rho_{X,Y} = \frac{\text{cov}(X,Y)}{\sigma_X \sigma_Y}
\end{equation}
\paragraph{Mean squared error. }
Pearson
cannot accurately judge a metric's performance for sentences of very similar quality, i.e., in the extreme case of rating outputs of identical quality, the correlation is either not defined or $0$, caused by noise of the evaluation model.
Thus, we additionally evaluate using mean squared error (MSE), which is defined as
the squares of residuals after a linear transformation, divided by the sample size:
\begin{equation}
\text{MSE}_{X,Y} = \underset{f}{\min}\frac{1}{|X|}\sum\limits_{i = 1}^{|X|}{(f(x_i) - y_i)^2}
\end{equation}
with $f$ being a linear function.
Note that, since MSE is invariant to linear transformations of $X$ but not of $Y$, it is a non-symmetric quasi-metric.
We apply it with $Y$ being the human ratings.
An additional advantage as compared to Pearson is that it has an interpretable meaning: the expected error made by a given metric as compared to the human rating.
\begin{table}[t!]
\centering
\setlength{\tabcolsep}{6.2pt}
\begin{tabular}{l | c | c c }
\textbf{metric} & \textbf{refs} & \textbf{Pearson} & \textbf{MSE} \\ \hline
{WordSLOR} & \textbf{0} & \textbf{0.454\phantom{$^\ast$}} & \textbf{0.261}\phantom{$^\ast$}\\
{WPSLOR} & \textbf{0} & 0.437\phantom{$^\ast$} & 0.267\phantom{$^\ast$} \\ \hline
WordNCE & \textbf{0} & 0.403$^\ast$ & 0.276$^\ast$ \\
WPNCE & \textbf{0} & 0.413$^\ast$ & 0.273$^\ast$ \\ \hline
{WordPPL} & \textbf{0} & 0.325$^\ast$ & 0.295$^\ast$ \\
{WPPPL} & \textbf{0} & 0.344$^\ast$ & 0.290$^\ast$ \\ \hline
{ROUGE-L-mult} & $3-5$& 0.429$^\ast$ & 0.269\phantom{$^\ast$} \\
{LR3-F-mult} & $3-5$ & 0.405$^\ast$ & 0.275$^\ast$\\
{LR2-F-mult} & $3-5$ & 0.375$^\ast$ & 0.283$^\ast$ \\
{LR3-R-mult} & $3-5$ & 0.412$^\ast$ & 0.273$^\ast$ \\\hline
{ROUGE-L-single} & 1 & 0.406$^\ast$ & 0.275$^\ast$ \\\hline
\end{tabular}
\caption{\label{tab:overallresults} Pearson correlation (higher is better) and MSE (lower is better) for all metrics; best results in bold; \textit{refs}=number of references used to compute the metric.}
\end{table}
\begin{table*}[h!]
\centering
\setlength{\tabcolsep}{4.5pt}
\begin{tabular}{l | c | c c c c | c c c c }
& & \multicolumn{4}{|c|}{Pearson} & \multicolumn{4}{|c}{MSE} \\
& \textbf{refs} & {}{\textbf{ILP}} & {}{\textbf{NAMAS}} & {}{\textbf{S2S}} & {}{\textbf{T3}} & {}{\textbf{ILP}} & {}{\textbf{NAMAS}} & {}{\textbf{S2S}} & {}{\textbf{T3}}
\\
\hline\hline
\# samples & & {}{679} & {}{762} & {}{767} & {}{747} & {}{679} & {}{762} & {}{767} & {}{747}
\\
\hline\hline
WordSLOR & \textbf{0}& 0.363$^\ast$& 0.340$^\ast$& \textbf{0.257\phantom{$^\ast$}}& 0.343\phantom{$^\ast$}& 0.307$^\ast$& 0.104\phantom{$^\ast$}& \textbf{0.161\phantom{$^\ast$}}& 0.174\phantom{$^\ast$}\\
WPSLOR & \textbf{0}& 0.417$^\ast$& 0.312$^\ast$& 0.201$^\ast$& \textbf{0.360\phantom{$^\ast$}}& 0.292$^\ast$& 0.106$^\ast$& 0.166\phantom{$^\ast$}& \textbf{0.172\phantom{$^\ast$}} \\
\hline
WordNCE &\textbf{0} & 0.311$^\ast$& 0.270$^\ast$& 0.128$^\ast$& 0.342\phantom{$^\ast$} & 0.319$^\ast$& 0.109$^\ast$& 0.170$^\ast$& 0.174\phantom{$^\ast$} \\
WPNCE &\textbf{0} & 0.302$^\ast$& 0.258$^\ast$& 0.124$^\ast$& 0.357\phantom{$^\ast$}& 0.322$^\ast$& 0.110$^\ast$& 0.170$^\ast$& \textbf{0.172\phantom{$^\ast$}} \\
\hline
ROUGE-L-mult & $3-5$& 0.471\phantom{$^\ast$}& \textbf{0.392\phantom{$^\ast$}}& 0.013$^\ast$& 0.256$^\ast$& 0.275\phantom{$^\ast$}& \textbf{0.100\phantom{$^\ast$}}& 0.173$^\ast$& 0.184$^\ast$ \\
LR3-F-mult & $3-5$& \textbf{0.489\phantom{$^\ast$}}& 0.266$^\ast$& 0.007$^\ast$& 0.234$^\ast$& \textbf{0.269\phantom{$^\ast$}}& 0.109$^\ast$& 0.173$^\ast$& 0.187$^\ast$ \\
LR2-F-mult & $3-5$& 0.484\phantom{$^\ast$}& 0.213$^\ast$& -0.013$^\ast$& 0.236$^\ast$& 0.271\phantom{$^\ast$}& 0.112$^\ast$& 0.173$^\ast$& 0.186$^\ast$ \\
LR3-R-mult & $3-5$& 0.473\phantom{$^\ast$}& 0.246$^\ast$& -0.002$^\ast$& 0.232$^\ast$& 0.275$^\ast$& 0.111$^\ast$& 0.173$^\ast$& 0.187$^\ast$\\
\hline
ROUGE-L-single & 1 & 0.363$^\ast$& 0.308$^\ast$& 0.008$^\ast$& 0.263$^\ast$& 0.307$^\ast$& 0.107$^\ast$& 0.173$^\ast$& 0.184$^\ast$ \\
\hline
\end{tabular}
\caption{\label{tab:analysis} Pearson correlation (higher is better) and MSE (lower is better), reported by compression system; best results in bold;
\textit{refs}=number of references used to compute the metric.}
\end{table*}
\subsection{Results and Discussion}
As shown in Table \ref{tab:overallresults},
WordSLOR and WPSLOR correlate best with human judgments:
WordSLOR (respectively WPSLOR) has a $0.025$ (respectively $0.008$)
higher Pearson correlation
than the best word-overlap metric ROUGE-L-mult,
even though the latter requires multiple reference compressions.
Furthermore, if we consider with ROUGE-L-single a setting with a single given reference,
the distance to WordSLOR increases to $0.048$
for Pearson correlation. Note that, since having a single reference is very common, this result is highly relevant for practical applications.
Considering MSE, the top two metrics are still WordSLOR and WPSLOR, with a $0.008$ and, respectively,
$0.002$ lower error than the third best metric, ROUGE-L-mult.
\footnote[]{*Significantly worse than best (bold) result with $p < 0.05$; one-tailed; Fisher-Z-transformation for Pearson, two sample t-test for MSE.}
Comparing WordSLOR and WPSLOR, we find no significant differences:
$0.017$
for Pearson and $0.006$ for MSE.
However, WPSLOR uses a more compact LM and, hence,
has a shorter training time,
since the vocabulary is smaller
($16,000$ vs. $128,000$ tokens).
Next, we find that
WordNCE and WPNCE perform roughly on par with
word-overlap metrics. This is interesting, since they, in contrast to traditional metrics, do not require reference compressions.
However, their correlation with human fluency judgments is strictly lower than that of their respective SLOR counterparts.
The difference between WordSLOR and WordNCE is bigger than that between WPSLOR and
WPNCE.
This might be due to accounting for differences in frequencies being
more important for words than for WordPieces.
Both WordPPL and WPPPL clearly underperform as compared to all other metrics in our experiments.
The traditional word-overlap metrics all perform
similarly.
ROUGE-L-mult and {LR2-F-mult} are best and worst, respectively.
\subsection{Analysis I: Fluency Evaluation per Compression System}
\label{subsec:ana_system}
The results per compression system (cf. Table \ref{tab:analysis}) look different from the correlations in Table \ref{tab:overallresults}: Pearson
and MSE are both lower. This is due to the outputs of each given system being of comparable quality.
Therefore, the datapoints are similar and, thus, easier to fit for the linear function used for MSE.
Pearson, in contrast, is lower due to its invariance to linear transformations of both variables.
Note that this effect is smallest for ILP, which has
uniformly distributed targets ($\text{Var}(Y) = 0.35$ vs. $\text{Var}(Y) = 0.17$ for SEQ2SEQ).
Comparing the metrics, the two SLOR approaches perform best for SEQ2SEQ and T3. In particular, they outperform the best word-overlap metric baseline by $0.244$ and $0.097$ Pearson correlation as well as $0.012$ and $0.012$ MSE, respectively. Since T3 is an abstractive system,
we can conclude that WordSLOR and WPSLOR are applicable even for systems that are not limited to make use of a fixed repertoire of words.
For ILP and NAMAS, word-overlap metrics obtain best results. The differences in performance, however, are with a maximum difference of $0.072$ for Pearson and ILP much smaller than for SEQ2SEQ. Thus, while the differences are significant, word-overlap metrics do not outperform our SLOR approaches by a wide margin.
Recall, additionally, that word-overlap metrics rely on references being available, while our proposed approaches do not require this.
\subsection{Analysis II: Fluency Evaluation per Domain}
\label{subsec:ana_domain}
Looking next at the correlations for all models but different domains (cf. Table \ref{tab:analysis_domain}),
we first observe that
the results across domains are similar, i.e., we do not observe the same effect as in Subsection \ref{subsec:ana_system}.
This is due to the distributions of scores being uniform ($\text{Var}(Y) \in [0.28, 0.36]$).
Next, we focus on an important question: How much does the performance of our SLOR-based metrics depend on the domain, given that the respective LMs are trained on Gigaword, which consists of news data?
Comparing the evaluation performance for individual metrics, we observe that, except for \textit{letters}, WordSLOR and WPSLOR perform best across all domains: they outperform the best word-overlap metric by at least $0.019$ and at most $0.051$ Pearson correlation, and at least $0.004$ and at most $0.014$ MSE. The biggest difference in correlation is achieved for the \textit{journal} domain.
Thus, clearly even LMs which have been trained on out-of-domain data
obtain competitive performance for fluency evaluation.
However, a domain-specific LM might additionally improve the metrics' correlation with human judgments.
We leave a more detailed analysis of the importance of the training data's domain for future work.
\begin{table*}[h!]
\centering
\setlength{\tabcolsep}{4.5pt}
\begin{tabular}{l | c | c c c c | c c c c }
& & \multicolumn{4}{|c|}{Pearson} & \multicolumn{4}{|c}{MSE} \\
& \textbf{refs}
& {}{\textbf{letters}} & {}{\textbf{journal}} & {}{\textbf{news}} & {}{\textbf{non-fi}} & {}{\textbf{letters}} & {}{\textbf{journal}} &
{}{\textbf{news}} & {}{\textbf{non-fi}}
\\
\hline\hline
\# samples &
& {}{640} & {}{999} & {}{344} & {}{972} & {}{640} & {}{999} &
{}{344} & {}{972}
\\
\hline\hline
WordSLOR & \textbf{0}& 0.452\phantom{$^\ast$}& \textbf{0.453\phantom{$^\ast$}}& \textbf{0.403\phantom{$^\ast$}}& \textbf{0.484\phantom{$^\ast$}}& 0.258\phantom{$^\ast$}& \textbf{0.250\phantom{$^\ast$}}& \textbf{0.234\phantom{$^\ast$}}& \textbf{0.278\phantom{$^\ast$}} \\
WPSLOR & \textbf{0}& 0.435$^\ast$& 0.415$^\ast$& 0.389\phantom{$^\ast$}& 0.483\phantom{$^\ast$}& 0.263\phantom{$^\ast$}& 0.260\phantom{$^\ast$}& 0.237\phantom{$^\ast$}& 0.278\phantom{$^\ast$} \\
\hline
WordNCE &\textbf{0} & 0.395$^\ast$& 0.412$^\ast$& 0.342$^\ast$& 0.425$^\ast$& 0.273$^\ast$& 0.261$^\ast$& 0.247\phantom{$^\ast$}& 0.297$^\ast$ \\
WPNCE &\textbf{0} & 0.424$^\ast$& 0.398$^\ast$& 0.363\phantom{$^\ast$}& 0.460\phantom{$^\ast$}& 0.266$^\ast$& 0.265$^\ast$& 0.243\phantom{$^\ast$}& 0.286\phantom{$^\ast$} \\
\hline
ROUGE-L-mult & $3-5$& \textbf{0.487\phantom{$^\ast$}}& 0.382$^\ast$& 0.384\phantom{$^\ast$}& 0.451$^\ast$& \textbf{0.247\phantom{$^\ast$}}& 0.269$^\ast$& 0.238\phantom{$^\ast$}& 0.289\phantom{$^\ast$} \\
LR3-F-mult & $3-5$& 0.404$^\ast$& 0.402$^\ast$& 0.278$^\ast$& 0.439$^\ast$& 0.271$^\ast$& 0.264$^\ast$& 0.258$^\ast$& 0.293\phantom{$^\ast$} \\
LR2-F-mult & $3-5$& 0.390$^\ast$& 0.363$^\ast$& 0.292$^\ast$& 0.395$^\ast$& 0.275$^\ast$& 0.273$^\ast$& 0.256$^\ast$& 0.306$^\ast$ \\
LR3-R-mult & $3-5$& 0.420$^\ast$& 0.395$^\ast$& 0.272$^\ast$& 0.453\phantom{$^\ast$}& 0.267$^\ast$& 0.266$^\ast$& 0.259$^\ast$& 0.288\phantom{$^\ast$} \\
\hline
ROUGE-L-single & 1 & 0.453\phantom{$^\ast$}& 0.347$^\ast$& 0.335$^\ast$& 0.450$^\ast$& 0.258$^\ast$& 0.277$^\ast$& 0.248\phantom{$^\ast$}& 0.289\phantom{$^\ast$} \\
\hline
\end{tabular}
\caption{\label{tab:analysis_domain} Pearson correlation (higher is better) and MSE (lower is better), reported by domain of the original sentence or paragraph; best results in bold;
\textit{refs}=number of references used to compute the metric.}
\end{table*}
\section{Incorporation of Given References}
\begin{table*}[t!]
\centering
\setlength{\tabcolsep}{6.8pt}
\begin{tabular}{| c l |}
\hline
\textbf{model} & \textbf{generated compression} \\ \hline
ILP & Objectives designed to lead incarcerated youth to an understanding of grief and loss\\
& related influences on their behavior.\\
ILP & In Forster's A Passage to India is created.\\
SEQ2SEQ & Jogged my thoughts back to Muscat Ramble. \\
SEQ2SEQ & Between Sagres and Lagos, pleasant beach with fishing boats, and a market. \\
T3 & Your support of the Annual Fund maintaining the core values in GSAS the ethics. \\\hline
\end{tabular}
\caption{\label{tab:weird_ratings} Sentences for which raters were unsure if they were perceived as problematic due to fluency or content issues, together with
the model which generated them.}
\end{table*}
ROUGE was shown to correlate well with ratings of a generated text's content or meaning at the sentence level \cite{toutanova2016dataset}. We further expect
content and fluency ratings to be correlated. In fact, sometimes it is difficult to distinguish which one is problematic: to illustrate this, we show some extreme examples---compressions which got the highest fluency rating and
the lowest possible content rating by at least one rater, but the lowest fluency score and the highest content score by
another---in Table \ref{tab:weird_ratings}.
We, thus, hypothesize that ROUGE should contain information about fluency which
is complementary to SLOR,
and want to make use of references for fluency evaluation, if available.
In this section, we experiment with two \emph{reference-based} metrics -- one trainable one, and one that can be used without fluency annotations, i.e., in the same settings as pure word-overlap metrics.
\subsection{Experimental Setup}
First, we assume a setting in which we have the following available: (i) system outputs whose fluency is to be evaluated, (ii) reference generations for evaluating system outputs, (iii) a small set of system outputs with references, which has been annotated for fluency by human raters, and (iv) a large unlabeled corpus for training a LM.
Note that available fluency annotations are often uncommon in real-world scenarios; the reason we use them is that they allow for a proof of concept.
In this setting, we train scikit's \cite{scikit-learn} support vector regression
model (SVR)
with the default parameters on predicting fluency, given WPSLOR and ROUGE-L-mult.
We use $500$ of our total $2955$ examples for each of training and development,
and the remaining $1955$ for testing.
Second, we simulate a setting in which we have only access to (i) system outputs which should be evaluated on fluency, (ii) reference compressions, and (iii) large amounts of unlabeled text.
In particular, we assume to not have fluency ratings
for system outputs, which makes training a regression model impossible.
Note that this is the standard setting in which word-overlap metrics are applied.
Under these conditions, we propose to normalize both given scores by mean and variance, and to simply add them up. We call this new reference-based metric ROUGE-LM.
In order to make this second experiment comparable to the SVR-based one, we use the same $1955$ test examples.
\begin{table}[]
\centering
\setlength{\tabcolsep}{.92pt}
\begin{tabular}{l | c | c c | c c}
&\textbf{metric} & \textbf{refs} & \textbf{train?} & \textbf{Pearson} & \textbf{MSE} \\
\hline
1 &\textbf{SVR}: & $3-5$& yes & \textbf{0.594} & \textbf{0.217} \\
&ROUGE+WPSLOR & & & & \\
2 & \textbf{ROUGE-LM} &$3-5$ & no & 0.496 & 0.252 \\ \hline
3 & ROUGE-L-mult & $3-5$& no & 0.430 &0.273 \\
4 & WPSLOR & 0 & no & 0.439 & 0.270 \\ \hline
\end{tabular}
\caption{\label{results:content}Combinations; all differences except for 3 and 4 are statistically significant; \textit{refs}=number of references used to compute the metric; ROUGE=ROUGE-L-mult; best results in bold.}
\end{table}
\subsection{Results and Discussion}
Results are shown in Table \ref{results:content}.
First, we can see that using SVR (line 1) to combine ROUGE-L-mult and WPSLOR outperforms both individual scores (lines 3-4) by a large margin.
This serves as a proof of concept: the information contained in the two approaches is indeed complementary.
Next, we consider the setting where only references and no annotated examples are available.
In contrast to SVR (line 1), ROUGE-LM (line 2) has only the same requirements as conventional word-overlap metrics (besides a large corpus for training the LM, which is easy to obtain for most languages).
Thus, it can be used in the same settings as other word-overlap metrics.
Since ROUGE-LM---an uninformed combination---performs significantly better than both ROUGE-L-mult and WPSLOR on their own,
it should be the metric of choice for evaluating fluency with given references.
\section{Related Work}
\subsection{Fluency Evaluation}
Fluency evaluation is related to grammatical error
detection \cite{atwell1987detect,wagner2007comparative,
schmaltz2016sentence,liu2017exploiting}
and grammatical error correction \cite{islam2011correcting,
ng2013conll,ng2014conll,bryant2015far,
yuan2016grammatical}.
However, it differs from those in several aspects; most importantly,
it is concerned with the degree to which errors matter
to humans.
Work on automatic fluency evaluation in NLP has been rare.
\newcite{heilman2014predicting} predicted the fluency
(which they called \textit{grammaticality})
of sentences written by English language learners. In contrast to ours, their approach is supervised.
\newcite{stent2005evaluating}
and \newcite{cahill2009correlating} found only low correlation between automatic metrics and fluency ratings for
system-generated English paraphrases and the output of a German surface realiser, respectively.
Explicit fluency evaluation of NLG, including compression and the related task of summarization, has mostly been performed manually.
\newcite{vadlapudi-katragadda:2010:SRW}
used LMs for the evaluation of summarization fluency,
but their models were based on part-of-speech tags, which we do not require, and they were non-neural. Further,
they evaluated longer texts, not single sentences like we do.
\newcite{toutanova2016dataset}
compared 80 word-overlap metrics for evaluating the content and fluency of compressions, finding only low
correlation with the latter.
However, they did not propose an alternative evaluation.
We aim at closing this gap.
\subsection{Compression Evaluation}
Automatic compression evaluation has mostly had a strong focus on content.
Hence, word-overlap metrics like ROUGE \cite{lin-och:2004:ACL} have been widely used for compression evaluation.
However, they have certain shortcomings, e.g., they correlate best for extractive compression, while we, in contrast,
are interested in an approach which generalizes to abstractive systems.
Alternatives include success rate \cite{jing2000sentence},
simple accuracy \cite{bangalore2000evaluation}, which is based on the edit distance between the
generation and the reference,
or word accuracy \cite{hori2004speech}, the equivalent for multiple references.
\subsection{Criticism of Common Metrics for NLG}
In the sense that we promote an explicit
evaluation of fluency, our work is in line with previous criticism
of evaluating NLG tasks
with a single score produced by word-overlap metrics.
The need for better evaluation for machine translation (MT) was expressed, e.g., by \newcite{callison2006re}, who doubted the meaningfulness of BLEU,
and claimed that a higher BLEU score was neither a necessary precondition nor a proof of
improved translation quality.
Similarly, \newcite{song2013bleu} discussed BLEU being unreliable at the sentence or sub-sentence level
(in contrast to the system-level), or for only one single reference.
This was supported by \newcite{isabelle-cherry-foster:2017:EMNLP2017},
who proposed a so-called challenge set approach as an alternative.
\newcite{graham-EtAl:2016:COLING} performed a large-scale evaluation of human-targeted metrics for machine translation, which can be seen as a compromise between human evaluation and fully automatic metrics. They also found fully automatic metrics to correlate only weakly or moderately with human judgments.
\newcite{bojar2016ten} further confirmed that automatic MT evaluation methods do not perform well with a
single reference.
The need of better metrics for MT has been addressed since 2008 in the WMT metrics shared
task \cite{bojar-EtAl:2016:WMT2,W17-4755}.
For unsupervised dialogue generation, \newcite{liu-EtAl:2016:EMNLP20163} obtained close to no correlation with human judgements for BLEU, ROUGE and METEOR.
They contributed this in a large part to the unrestrictedness of dialogue answers, which makes it hard to match given references.
They emphasized that the community should move away from these metrics for dialogue generation tasks,
and develop metrics that correlate more strongly with human judgments.
\newcite{elliott-keller:2014:P14-2} reported the same for BLEU and image caption generation.
\newcite{duvsek2017referenceless} suggested an RNN
to evaluate NLG at the utterance level, given
only the input meaning representation.
\input{long.future_work}
\section{Conclusion}
We empirically confirmed the effectiveness of
SLOR,
a LM score which accounts for the effects of sentence length
and individual unigram probabilities,
as a metric for fluency evaluation of the NLG task of automatic compression at the sentence level.
We further introduced WPSLOR, an adaptation of SLOR to WordPieces,
which reduced both model size and training time at a similar evaluation performance.
Our experiments showed that our proposed referenceless metrics correlate significantly better with fluency ratings for the outputs of compression systems
than traditional word-overlap metrics on a benchmark dataset.
Additionally, they
can be applied even in settings where no references are available, or would be costly to obtain.
Finally, for given references, we proposed the
reference-based metric ROUGE-LM, which consists of a combination of WPSLOR and ROUGE.
Thus, we were able to obtain
an even more accurate fluency evaluation.
\section*{Acknowledgments}
We would like to thank Sebastian Ebert and Samuel Bowman for their detailed and helpful feedback.
\section{Future Work}
The work presented in this paper brings up multiple
interesting next steps for future research.
First, in Subsection \ref{subsec:ana_domain}, we investigated the performances of WordSLOR and WPSLOR in dependence of the domain of the considered text. We concluded that an application was possible even for unrelated domains. However, we did not experiment with alternative LMs, which leaves the following questions unresolved: (i) Would training LMs on specific domains improve WordSLOR's and WPSLOR's correlation with human fluency judgments, i.e., to what degree are the properties of the training data important? (ii) How does the size of the training corpus influence performance?
Ultimatly, this research
could lead to answering the following question: Is it better to train a LM on a small, in-domain corpus or
on a large corpus from another domain?
Second, we showed that, in certain settings, Pearson correlation does not
give reliable insight into a metric's performance. Since in general evaluation of \emph{evaluation metrics} is hard, an important topic for future research
would be the investigation of better ways to do so, or to study under which conditions a metric's performance can be assessed best.
Last but not least,
a straight-forward continuation of our research would encompass the investigation of SLOR as a fluency metric for other NLG tasks or languages. While the results for compression strongly suggest a general applicability to a wider range of NLP tasks, this has yet to be confirmed empirically. As far as other languages are concerned, the question what influence a given language's grammar has would be an interesting research topic. |
\section*{Review Topics}
%
\pagestyle{fancy}
\renewcommand{\headrulewidth}{0pt}
\IEEEpeerreviewmaketitle
\section{Introduction}
\let\thefootnote\relax\footnotetext{\noindent \textsuperscript{*}Parts of this paper are patent pending. \\ \textsuperscript{\textdagger}This work was done while the author worked at the ABB Robotics R\&D Center, 3055 Orchard Dr., San Jose, CA 95134.}
Robots have revolutionized the manufacturing process by performing tasks more efficiently and accurately at a lower operational cost. However, sudden breakdowns or malfunctions of robots can result in a steep decrease in production quality and quantity, which often entails substantial financial losses.
Robot failures and malfunctions can have various causes, e.g., moving heavy objects continuously leads to the deterioration of robots gears over time. As a result of this deterioration, gaps can appear between the gear teeth, which results in so-called \emph{backlash}. Backlash is a clearance or lost motion in a mechanism. It is noticeable in a gearbox when the direction of movement is reversed and the slack or lost motion is taken up before the reversal of motion is complete.
In order to prevent unscheduled maintenance due to sudden disruptions of normal operation, \emph{predictive maintenance} has gained a lot of attention recently~\cite{hornung2014model, liu2005model, mobley2002introduction, garcia2006simap}.
The goal of predictive maintenance is to prevent unexpected equipment failure by constantly monitoring the
performance and condition of equipment in operation and to extrapolate from the current condition when corrective maintenance will be required \cite{grall2002continuous, hashemian2010state}. Predictive maintenance can be performed on a software and hardware level. Software faults are, e.g., communication problems, controller software malfunctioning, etc. Hardware issues are mostly related to broken sensors, broken gears, backlash, etc. The focus of this paper is to identify hardware issues through analyzing the internal signals of the robots. An accurate system capable of detecting hardware problems requires additional internal sensors. However, using extra sensors to perform predictive maintenance is not feasible in many situations because it increases the cost, complexity, system weight, and requires extra space inside the robot system~\cite{hornung2014model}. For this reason, predictive maintenance systems based on machine learning algorithms have gained popularity recently and this approach is the focus of this paper.
\begin{figure*}[th]
\begin{minipage}[b]{\linewidth}
\centering
\centerline{\includegraphics[width=8cm]{training.jpg}}
\centerline{ (a)}\medskip
\vspace{-2mm}
\end{minipage}
\begin{minipage}[b]{\linewidth}
\centering
\centerline{\includegraphics[width=12cm]{test.jpg}}
\centerline{ (b)}\medskip
\vspace{-4mm}
\end{minipage}
\caption{Flowchart of the proposed work: (a) Training from the healthy robot with task $A$; (b) Evaluating (test) the operating robot in the customer facility with unknown task $B$ using domain adaptation.}
\label{fig:schematic}
\vspace{-5mm}
\end{figure*}
There have been several research activities recently in this domain~\cite{hornung2014model, park2016multimodal, borgi2017data, sjostrand2012dynamic, pinto2019robot, moisescu2018retracted, sathish2019data}. Hornung et al.\ propose to map data to a positive and a negative group, where the positive group includes all areas of data space accessed during normal operation while data in the negative group belongs to the unknown area of data space~\cite{hornung2014model}. In order to approximate the positive data space, radial basis function kernels are used. Anomalies are detected by training a support vector machine (SVM) on positive and negative data. In another work, Park et al.\ propose to use multimodal sensory data such as haptic, auditory, visual, and kinematic signals to train a hidden Markov model (HMM)\cite{park2016multimodal}. The trained HMM provides a probability of the test data belonging to class associated with normal behavior, which can be used as a criterion for anomaly detection.
Generally, these approaches consist of data pre-processing, feature extraction, dimension reduction, and a model-based classifier. A model is trained for each axis of the robot based on the training data collected while the robot is healthy and performing a specific task. Later, the health of the robot is evaluated by collecting data and comparing it with the trained model.
The problem with existing predictive maintenance algorithms is that they raise false alarms when the task for which data is collected differs from the task used during the training phase. When the task changes, features such as frequency components extracted from data change as well. This change in the features fools the systems to erroneously raise an alarm, although the robot is healthy and operating normally. In order to avoid false alarms due to an operational change, current systems require the model to be retrained when altering the task of the robot.
In this paper, we propose to use \emph{transfer learning}, also known as knowledge transfer and domain adaptation, as an intermediate step in order to avoid retraining the model every time the task of the robot changes. In the proposed method, the training data is collected while the robot is healthy and performing some task $A$. Subsequently, the short-time Fourier transform (STFT) of the data and their combinations are calculated to derive the features used to determine the condition of the robot, i.e., these features constitute the input to the predictive maintenance algorithm. These features are in a subspace called \emph{source domain}, ${\mathcal D}_s$, representing the healthy robot. The test data is collected while the robot is doing task $B$. Test data is in the \emph{target domain}, ${\mathcal D}_t$. The goal of the transfer learning algorithm is to transfer the model learned in the source domain to the target domain~\cite{pan2010survey}. We assume that labeled data of healthy and faulty robots are not available in the target domain. Therefore, unsupervised methods must be used to transfer the knowledge between the domains. In this paper, we propose to use \emph{manifold alignment}~\cite{wang2009general, boucher2015aligning}, which is a local-preserving algorithm that finds a common subspace of the source and the target domains. Figure~\ref{fig:schematic} shows the flowchart of the proposed system.
Furthermore, since it is not convenient to ask the users of the robot to run the transfer learning algorithm each time the task of the robot changes, we assume that the test data is from the target domain. Thus, the manifold alignment algorithm is continuously applied to the features extracted from the test data to make sure that the comparison between healthy data and the test data is performed in the same domain. The experiments on real-world datasets obtained from two types of industrial robots show a significant performance improvement in dealing with operational changes, as well as the ability of the algorithm to correctly identify anomalies.
The following section describes the real-world dataset, pre-processing, feature extraction, and detection methods. Subsequently, the proposed predictive maintenance method based on the manifold alignment algorithm is presented in detail, followed by the presentation and discussion of experimental results based on real-world data traces. The final section concludes the findings of this paper and proposes directions for future work.
\section{Scenario}\label{sec:scenario}
\subsection{Dataset}
The dataset is collected from two robot types. Robot ${\mathcal A}$ is from a series of single arm robots with 6 axis. Their reach is up to 3m (10ft) and their payload is around 100kg (220lb). Robot ${\mathcal B}$ is also from a series of 6-axis single arm robots and reach of up to 4m (13ft) and a maximum payload of 600kg (1320lb). Three signals, position, speed, and torque, are recorded from the controller of these robots for each axis separately. The sampling frequency of speed and torque signals are $2kHz$ while the position signal is sampled at a frequency of $250Hz$.
Since the dataset comprises real-world data from robots in operation, we cannot provide more details about the robots, nor their specific tasks due to confidentiality obligations. However, we believe that the generic nature of our proposed predictive maintenance method makes it possible to apply it successfully for a broad range of robot types and tasks.
\subsection{Pre-processing}
Since the sampling frequency of the position signal is lower than the other two signals, as mentioned above, the position signal is upsampled to get the same number of data points for each signal.
The position signal is upsampled by first inserting 7 data points with value zero between pairs of actual signals, which artificially increases the data rate to $2kHz$. Next, the zero values are overwritten using the
cubic interpolation method, which fits a third degree polynomial function to the samples of the position signal~\cite{fritsch1980monotone}. The fitted polynomial function is used to approximate the values of position signals at the newly inserted sample points. After pre-processing, each signal $x(t) = [speed(t), position(t), torque(t)]^T$ provides one sample at a time $t$, where $speed(t)$, $position(t)$, and $torque(t)$ are scalar values.
\subsection{Feature Extraction}
After the data is pre-processed, a certain set of features is required to be extracted from the given three signals. The features should not only accurately represent normal data but also discriminate normal from anomalous data well. In this paper, the short-time Fourier transform (STFT) of the three signals in $x(t)$ and their combinations is computed as the set of features. The reason for choosing STFT is that it adequately represents the time-frequency distribution of the signals, which is suitable to identify various anomalies. For example, gearbox malfunctioning and backlash appears as vibration during robot operation. STFT is able to capture these vibrations as different frequency components at different time steps. However, note that finding the appropriate features for the robot predictive maintenance is not the focus of this paper. Rather, we focus on using appropriate (given) features to detect anomalies without causing false positives due to operation changes. As we will show later, the standard STFT-based approach to identify features already yields promising results in experiments using real-world data.
\subsection{Detection}
The absolute values of STFT magnitude are used to build the subspace representing the healthy robot.
Naturally, the performance of the anomaly detection mechanism critically hinges upon the accurateness of this subspace.
Before delving into our proposed method to derive such a subspace in the subsequent section, we briefly summarize a common and straightforward approach: principal component analysis (PCA) can be applied to form the subspace, i.e., the principal components of the absolute STFT values constitute the subspace associated with a healthy robot.
Given this subspace, training and test datasets are projected to this subspace using the principal components. The $\ell_2$-norm distance between the training dataset (of a healthy robot) and the test dataset in this subspace is used as a criterion to identify whether the test data represents a healthy or faulty robot. If the test dataset comes from a healthy robot, the subspace should be able to represent the features of the test data well and the distance to the training data in this subspace should be small. On the other hand, a large distance is an indication of a faulty robot.
Formally, let ${\bf X}_{s}$ and ${\bf X}_{t}$ represent absolute STFT values of training (healthy) and test data, respectively, where rows are samples and columns are the flattened absolute STFT values of a particular sample. Our distance criterion is defined as $d = \|{\bf X}_{s}{\bf P}-{\bf X}_{t}{\bf P} \|_2^2$, where ${\bf P}$ is the matrix whose columns are the principal components. We formulate the detection problem as:
\begin{equation}
\begin{array}{cc}
H_0: & d \ge \varepsilon \\
H_1: & d < \varepsilon
\end{array}
\end{equation}
\noindent where $\varepsilon$ is the threshold determining whether the test data represents a healthy or faulty condition. The parameter $\varepsilon$ can be set based on the mean distance and the variance for samples associated with a healthy robot. In other words, the threshold is determined by the significance level, $\alpha$, or the percentile of the distribution.
\section{Proposed Method}
\label{sec:domain}
The problem with existing methods is that the subspace built from the features of the training data, collected from a healthy robot, of one specific operation is not able to represent the data coming from the same healthy robot but executing another operation. As a result, the $\ell_2$-norm distance increases and the algorithm raises a false alarm. In this section, transfer learning is used to find a common subspace of the training and test data. Then, the $\ell_2$-norm distance is calculated in this common subspace.
Let ${\mathcal X}_s$ represent the feature space in the source domain, and ${\bf X}_s \in {\mathbb R}^{N \times K}$ are $K$ features (absolute STFT values) of $N$ training samples drawn from this space. Let ${\bf X}_t \in {\mathbb R}^{N \times K}$ be $K$ features of $N$ test samples (absolute STFT values) drawn from the target feature space ${\mathcal X}_t$ and collected while the robot is in operation. Since the application of the robot during training and test are different, their corresponding feature spaces are different, i.e., ${\mathcal X}_s \neq {\mathcal X}_t$. Therefore, there is a need for a transfer learning algorithm to reduce the difference between these two spaces while preserving the geometric properties~\cite{pan2011domain}. This reduction can be achieved by finding a common subspace between source and target spaces through minimizing a certain cost function. Manifold alignment is an unsupervised tranfer learning (or domain adaptation) algorithm that provides a closed-form solution~\cite{boucher2015aligning, wang2009general}. Having a closed-form solution makes it possible to implement it in a computationally efficient manner, which can be a requirement for applications with real-time constraints. For this reason, we chose manifold alignment to perform domain adaptation. The manifold alignment algorithms replaces the application of PCA as discussed in the previous section. Hence, the input to the manifold alignment algorithm is the STFT values, ${\bf X}_s$ and ${\bf X}_t$, and the output is the computed distance $d$.
More precisely, the manifold alignment algorithm computes low-rank embeddings (LREs) of ${\bf X}_s$ and ${\bf X}_t$ in a joint subspace. The LREs are then used to calculate the distance.
%
%
We will now discuss how the LREs of the source and target features, ${\bf X}_s $ and ${\bf X}_t $, are calculated.
The objective is to minimize the following loss function:
\begin{equation}
\begin{array}{c}
\min_{R_s} {1 \over 2} {\|{\bf X}_s-{\bf X}_s{\bf R}_s \|}_F^2 + \lambda {\|{\bf R}_s\|}_* \\
\min_{R_t} {1 \over 2} {\|{\bf X}_t-{\bf X}_t{\bf R}_t \|}_F^2 + \lambda {\|{\bf R}_t\|}_*,
\end{array}
\end{equation}
\noindent where $\lambda>0$, ${\|.\|}_F$ and ${\|.\|}_*$ are Frobenius and spectral norms, respectively. In this equation, ${\bf X}_s{\bf R}_s$ and ${\bf X}_t{\bf R}_t$ are the low rank embeddings of ${\bf X}_s $ and ${\bf X}_t $, respectively, and ${\bf R}_s$ and ${\bf R}_t$ are their reconstruction coefficient matrices.
\begin{align}
{\bf R}_s &= {\hat {\bf V}}_s({\bf I}-{\bf \hat S}^{-2}_s){\bf \hat V}^T_s \;\;\text{and}\\
{\bf R}_t &= {\bf \hat V}_t({\bf I}-{\bf \hat S}^{-2}_t){\bf \hat V}^T_t,
\end{align}
where ${\hat {\bf S}}_s$ is the diagonal matrix of all singular values greater than one and matrix ${\hat {\bf V}}_s.$ is comprised of the corresponding right-singular vectors. The block reconstruction coefficient matrix is given by:
\begin{equation}
{\bf R}=\left [
\begin{array}{cc}
{\bf R}_s & {\bf 0} \\
{\bf 0} & {\bf R}_t
\end{array}
\right ]
\end{equation}
The inter-set correspondence between the samples of the training and test datasets is represented by
$${\bf C} = \left [
\begin{array}{cc}
{\bf 0} & {\bf I} \\
{\bf I} & {\bf 0}
\end{array}
\right ],$$
where ${\bf I}$ is the identity matrix.
After finding the LRE of source and target samples, the projection matrices from the source and the target space into the common subspace and the embedding of the source and target samples are calculated by minimizing the following cost function:
\begin{equation}
(1-\mu) \| {\bf F} - {\bf R}{\bf F} \|_F^2+\mu \sum_{i,j=1}^N{ \|{\bf F}_i-{\bf F}_j\|^2 C(i,j)},
\end{equation}
\noindent where $\mu \in [0,1]$ determines the importance of the local geometry (first term) vs.\ the inter-set correspondence (second term). The cost function can be simplified to the following expression~\cite{boucher2015aligning}:
\begin{equation}\label{eq:cost2}
(1-\mu)({\bf I}-{\bf R})^T({\bf I}-{\bf R}) + 2\mu {\bf L},
\end{equation}
\noindent where ${\bf L}$ is the Laplacian matrix of ${\bf C}$. This cost function is minimized by replacing
$$
{\bf F}=\left [ \begin{array}{c} {\bf F}_s \\ {\bf F}_t \end{array} \right ]
$$
with the $d$ smallest eigenvectors of Equation~\ref{eq:cost2}. ${\bf F}$ is the $d$-dimensional embedding of $N$ training and $N$ test features in the common subspace.
Since ${\bf X}_s$ is assumed to be the training features collected when the robot is healthy, the $d$ dimensional embedding of the test dataset is compared to the embedding of the test dataset using the Euclidean distance, i.e.,
$$\delta(t) = {\sqrt{\|{\bf F}_s - {\bf F}_t \|^2_2}}.$$
If ${\bf X}_t$
is anomalous, its distance is larger than for ${\bf X}_s$. In order to perform hypothesis testing on the test data, the metric $\delta$ is calculated for several normal datasets to build a probability distribution function. The empirical distribution of the metric is the positive half of the Laplace distribution with $\mu=0$, and any ${\bf X}_t$ outside of the confidence interval is marked as anomalous.
\section{Experimental Results}
\label{sec:exp}
In this section, the proposed algorithm is evaluated on a lab-generated dataset and two separate real-world datasets from robot ${\mathcal A}$ and robot ${\mathcal B}$.\footnote{The actual robot names cannot be provided for confidentiality reasons as the robots are used in production.} In order to show the efficacy of the proposed method in robot predictive maintenance, the standard PCA-based approach is used as well to build a subspace of the features (STFT of signals and their combinations) of the healthy robot from the training data. The extracted features of the test dataset are projected to this subspace and compared with the training dataset.
The results obtained using our proposed approach based on transfer learning are then compared against this PCA-based solution.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{roc.png}
\caption{Receiver operating characteristic curve calculated for the $\{50, 75, 90, 95, 97.5, 99\}th$ percentiles. The experiment is repeated 10 times for each percentile and the results are averaged.}
\label{fig:roc}
\end{figure}
\subsection{Lab Generated Dataset}
In this section, we generate a simulated dataset that resembles speed, position, and torque signals of real robots to evaluate
the effectiveness of the proposed approach. The length of each signal on each day is $N=300$ samples. The experiment spans $10$ consecutive days. The operation of the robot during the first three days does not change and the robot is healthy. The set of generated healthy signals are
\begin{flalign*}
& Position=2 \sin(2 \pi 9 n)+\mathcal{N}(0,10^{-4})& \\
& Speed=10 \sin(2 \pi 10 n)+\mathcal{N}(0,10^{-4}) & \\
& Torque= \sin(2 \pi 10 n) + 0.05 \sin(2 \pi 300 n)+\mathcal{N}(0,10^{-4}),&
\end{flalign*}
\noindent where $\mathcal{N}(0,10^{-4})$ is an added noise randomly generated from a Gaussian distribution with zero mean and standard deviation of $10^{-4}$. The robot arm, in a typical operation, moves mostly between two locations (pick and place). Therefore, the position signal looks like a combination of sinus signals oscillating over time. Since the speed and torque have a direct relationship with the position signal, they also resemble combinations of sinus signals. However, the torque signal is typically noisy and has a saw-edged shape. For this reason, we added the term $0.05 \sin(2 \pi 300 n)$, which results in torque signals that are similar in shape to those observed in real-world applications.
On the fourth day, the operation does not change---which means the three signals will be the same as before. However, the torque signal shows an anomaly, which starts at the $150^\mathit{th}$ sample and lasts for $60$ samples. For these $60$ samples, the torque signal is
\begin{alignat*}{1}
Torque &= \sin(2 \pi 10 n) + 0.3 \sin(2 \pi 100 n) \\
& + 0.3 \sin(2 \pi n)+\mathcal{N}(0,10^{-4}).
\end{alignat*}
The robot continues its operation one more day with normal behavior and the same operation. On the sixth day, the operation of the robot changes to
\begin{flalign*}
& Position=2 \sin(2 \pi 4 n)+\mathcal{N}(0,1e^{-4})& \\
& Speed=10 \sin(2 \pi 5 n)+\mathcal{N}(0,1e^{-4}) & \\
& Torque= \sin(2 \pi 5 n) + 0.05 \sin(2 \pi 300 n)+\mathcal{N}(0,1e^{-4}),&
\end{flalign*}
\noindent and lasts for three days. Then, the operation goes back to the first operation. Thus, the total $10$ days of operation can be summarized as $\{O_1, O_1, O_1, A_{O_1}, O_1, O_2, O_2, O_2, O_1 \}$, where $O_1$ and $O_2$ denote the first and second operations, respectively, and $A_{O_1}$ is the anomaly happening while the robot is performing operation $O_1$. A correct algorithm detects only this single anomaly. Detecting any operation change is considered as a false alarm.
We have applied the proposed algorithm to this generated dataset and calculated the true positive rate (TPR) and false positive rate (FPR) for the $\{50, 75, 90, 95, 97.5, 99\}th$ percentiles. This experiment was repeated $10$ times for both the proposed algorithm and the PCA-based algorithm at each percentile level and the results are averaged. Figure~\ref{fig:roc} shows the receiver operating characteristic (ROC) curve. As the figure shows, the ROC curve of the PCA-based method falls below random guess line while the proposed algorithm has the best possible shape, which proves the efficacy of the proposed method.
\begin{figure*}[t]
\centering
\begin{minipage}[b]{.49\textwidth}
\centering
\centerline{\includegraphics[width=8cm,height=4.5cm]{operation_DA.png}}
\centerline{ (a)}\medskip
\vspace{-2mm}
\end{minipage} %
\begin{minipage}[b]{.49\textwidth}
\centering
\centerline{\includegraphics[width=8cm,height=4.4cm]{operation_PCA.png}}
\centerline{ (b)}\medskip
\end{minipage}
\caption{{Robot ${\mathcal A}$}: first experiment using a 190-day data trace without any anomalies and three changes in the task of the robot. (a) Distance values over 190 days calculated using the proposed method. (b) Distance values over 190 days calculated using the PCA-based method. Three changes in the task are apparent on the $30^{\mathit{th}}$, $87^{\mathit{th}}$, and $125^{\mathit{th}}$ day in (b), whereas there is no discernible change in the distance in (a).}
\vspace{-2mm}
\label{fig:exp1}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[width=8cm,height=4.5cm]{anomaly_DA.png}}
\centerline{(a)}\medskip
\vspace{-2mm}
\end{minipage}%
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[width=8cm,height=4.5cm]{anomaly_PCA.png}}
\centerline{(b)}\medskip
\vspace{-2mm}
\end{minipage}
\caption{{Robot ${\mathcal B}$}: second experiment using a 148-day data trace with an anomaly on the last day and no change in the task of the robot. (a) Distance values over 148 days calculated using the proposed method. (b) Distance values over 148 days calculated using the PCA-based method. The anomaly is apparent on the last day in both (a) and (b) but the relative change is more pronounced using the proposed method.}
\label{fig:exp2}
\end{figure*}
\subsection{Real-World Dataset}
\subsubsection{Robot ${\mathcal A}$}
Three signals (speed, position, torque) of the fourth axis of robot ${\mathcal A}$ were collected for 190 consecutive days. The dataset on each day is 3 seconds long. No break down or anomaly was reported for this axis of the robot; however, the task of the robot changed three times during these 190 days. After pre-processing, the dataset on day 1 is used as the training data (healthy robot) and the respective STFT values constitute the extracted features ${\bf X}_s$. The datasets of the following days are used as the test dataset to form ${\bf X}_t$. $\delta(t)$ is calculated for each day with respect to the first day to identify any changes with respect to the first day. The assumption is that the robot is healthy on the first day, which can be the time of inspection or commissioning at the robot manufacturing facility. Figure~\ref{fig:exp1}(a) shows $\delta(t), \hspace{2mm} t \in \{1,2,\ldots,190\}$. It is apparent in this figure that none of the changes in the task of the robot is identified as anomalous behavior when using our proposed method based on transfer learning (domain adaption). On the other hand, Figure~\ref{fig:exp1}(b) shows a different picture when using PCA projection. All three changes in the task of the robot are obvious in this figure. The resulting changes in the computed distances indicate that the use of conventional methods such as PCA fails to distinguish between a change in the mechanical condition of the robot from a change in its task. A second observation when comparing these two figures is the magnitude of the distance function $\delta(t)$. The $\delta(t)$ values in Figure~\ref{fig:exp1}(b) are significantly larger than those of Figure~\ref{fig:exp1}(a). The larger distances provide further evidence that ${\bf X}_s$ and ${\bf X}_t$s are not in the same subspace. On the other hand, the small values of $\delta(t)$ in Figure~\ref{fig:exp1}(a) indicate that their projection onto the common subspace renders them comparable.
\subsubsection{Robot ${\mathcal B}$}
In the second experiment, three signals of the fifth axis of robot ${\mathcal B}$ were collected for 148 days. The dataset on each day is again 3 seconds long as in the first experiment. The axis of the robot broke down on the $149^{\mathit{th}}$ day. However, the task of the robot never changed during the entire 148 days. The training and test data sets are constructed as before:
The first day is considered the training data and the features are extracted from this dataset as ${\bf X}_s$. The extracted features of the other $147$ days form ${\bf X}_t$. The test data is compared to the training data using the $\ell_2$-norm to obtain the distance $\delta(t)$ at time $t$. Figure~\ref{fig:exp2}(a) shows the computed distances of the proposed algorithm. As in the first experiment, PCA is also used to build the subspace of the healthy robot, and the resulting distance function is depicted in Figure~\ref{fig:exp2}(b). According to the two plots in Figure~\ref{fig:exp2}, both methods perform equally well in identifying the trend of $\delta(t)$ leading to the breakdown of the axis on the $149^{\mathit{th}}$ day. However, comparing the $\delta(t)$ values of both plots demonstrates the training and test data are being compared in the same subspace in the proposed algorithm in contrast to the PCA-based method in Figure~\ref{fig:exp2}(b). As a result, there is a significantly sharper relative increase in the distance around and on the day the anomaly occurred when using our proposed method, implying that it can more distinctly identify anomalies.
\section{Conclusion}
Predictive maintenance systems are of great interest to manufacturers of various kinds of mechanical devices and apparatuses such as drives and robots. The goal of such systems is to build a model of the healthy apparatus and continuously compare the current condition of the device, captured in the form of measured signals, with the model to identify any possible errors or malfunction. The challenge with the application of predictive maintenance systems in robotics is that the model requires to be retrained every time the task of the robot changes using conventional methods. We showed that transfer learning (domain adaptation) addresses this challenge. While we used manifold alignment algorithm to project the features extracted from the training and test data onto a common subspace, it is worth investigating other unsupervised transfer learning algorithms. We used distance-based methods to compare the test data with the training data in the common subspace to identify any anomaly. The future work will consider time-series analysis methods in the common subspace to capture the dynamics of systems.
%
%
\bibliographystyle{IEEEbib}
|
\section{Introduction}
Community ecology has been deeply shaked by the book of Hubbell (2001) \cite{hubbel} that elaborated on the idea that the dynamics of ecological communities might be mainly shaped by random processes. A number of predictions made by this neutral theory of biodiversity have been indeed corroborated by empirical evidence (Hubbell \cite{hubbel}, Condit et al.\cite{Conditall}, Jabot and Chave \cite{JabotChave2009}). This good performance of neutral models for reproducing empirical patterns has stimulated the mathematical study of neutral ecological models (Etienne \cite{etienne2005}, Fuk et al \cite{FOC2017}), in connection with the rich body of work on evolutionary neutral models (Volkov et al. \cite{VBHM2003}, Ewens \cite{ewens2004}, Muirhead and Wakeley \cite{mw2009}).\\
More recent empirical evaluations of neutral predictions on tropical forest data have focused on the temporal dynamics of individual populations and have shown that the temporal variance of population sizes was actually larger than the one typically predicted by neutral models (Chisholm et al. \cite{Chisholmall}). These authors have suggested that this may be due to species-specific responses to the temporal variability of the environment. Subsequent modelling studies have elaborated on this idea (Kalyuzhny et al. \cite{kalyuzhny}, Jabot and Lohier \cite{jabotlohier}, Krone and Neuhauser \cite{KNeuhauser} and more mathematically-oriented contributions on this topic have since been made (\cite{GriffthsTavare,expansionmultiallelicWF,daninoshnerb,danino2016effect,danino2018stability,danino2018fixation,fung2017species,Viotprocees,Grieshammer,cmv}). All these studies share the same idea that community dynamics is influenced by a species-specific selection coefficient and that this selection coefficient is temporally varying, so that good and bad periods are subsequently experienced by all species within the community.\\
The aim of this contribution is to provide a comprehensive study of a simple model encapsulating this kind of ideas within a weak selection framework. Subsequent works within a strong selection framework will complement the present study.
Indeed, the simplest model for the evolution of population is surely the Moran model (or its cousin the discrete Wright-Fisher model), in which in a given population an individual is chosen to die (uniformly in the population) and then a child chooses his parent proportionally to the abundance in the previous population. One may also add immigration, i.e. a probability that the child comes from another community, and selection so that some species (or traits) have a selective advantage. Kalyuzhny et al \cite{kalyuzhny}, to bypass the neutrality of the model, chosed to consider immigration and selection as random processes (independent of the Moran system), but which preserves neutrality "in mean". One of the main goal is of course to study the effect of these models on biodiversity and its evolution. There are many ways to measure biodiversity. We will focus here on the Simpson index \cite{Simpson}, usually considered in neutral model \cite{etienneOlff2}: it measures the probability that two individuals uniformly chosen may be of the same species. More precisely denoting $X^{i}$ the number of individuals of species $i$, $S$ the number of species and $J$ the toal size of the population, the Simpson index is given by
$${\mathcal S}=\sum\limits_{i=1}^{S}\frac{X^{i}(X^{i}-1)}{J(J-1)}$$
thus varying (roughly) from $0$ to $1$, from maximal to minimal diversity. Using backward approach, and Kingman's coalescent, an explicit formula may be given for the (asymptotic in time) Simpson index in the neutral case with immigration as $\frac{1-m}{1+J-2m}$, as otherwise it is 1 as a particular specie will almost surely invade all the population. Note that a closed formula (even for the expectation) of the Simpson index at a given time, is usually not reachable.\\
We will consider in this paper this Moran model with immigration and selection as general random process. As said previously such models were recently considered for example by Griffiths \cite{Griffiths,GriffthsTavare}, Kalyuznhy et al \cite{kalyuzhny} for a simulation study, but no theoretical framework towards the Simpson index. The backward approach constitutes the works of Krone and Neuhauser \cite{KNeuhauser,NKrone} leading to new coalescent type processes which are however quite difficult to study and may not give a closed formula for the Simpson index. A recent work by Grieshammer \cite{Grieshammer} considers a forward in time approach but he does not focus on the Simpson index. Our approach is only forward here. As is often done in population genetics, we will consider the large population approximation. Our first task is then to justify this asymptotic to a Wright-Fisher diffusion process in random environment in the weak immigration and selection case. As a flavour, with only two species, the evolution of the proportion of one species is given by
$$dX_t=m_t(p_t-X_t)dt+s_tX_t(1-X_t)dt+\sqrt{2X_t(1-X_t)}dB_t$$
where $m_t$ is the immigration process, $p_t$ the probability that this species is chosen, and $s_t$ the selection advantage. It will be done by the usual martingale method. Another quantitative approach will be considered in \cite{ggp2018}. The Simpson index is then a quadratic form involving the proportion of each species, and by It\^o's formula it involves higher order term. The equation for the expectation of the Simpson index is therefore not closed. We will then introduce a quantitative approximation procedure for the expectation of the Simpson index, in the quenched case (corresponding to a given random environment) and in some particular case in the annealed case for two or more species. We will also study in some simple case (constant parameter) the long time behaviour of the Simpson index. It is reminiscent with the very recent work of Coron, M\'el\'eard and Villemonais \cite{cmv} (discovered while finishing this work). Very schematically our approach is the following
\begin{enumerate}
\item approximate the true discrete process by a SDE, i.e. Wright-Fisher process;
\item approximate the expectation of the Simpson index for the Wright-Fisher process by a deterministic ODE;
\item approximate the infinite time expectation of the Simpson index by the finite time, through evaluation of the speed of convergence towards equilibrium of the Wright-Fisher process.
\end{enumerate}
In Section 2, we introduce the Moran model in random environment and prove its convergence in the large population limit. In Section 3, we focus on the two species case where the approximation of the Simpson index is studied in the quenched case as well as its long time behaviour. Section 4 generalizes to a large number of species and also considers the annealed case when the selection parameter has a particular form (a variant of a Wright-Fisher process). The last section contains technical proofs or recall some known results for the Wright-Fisher process.
\section{The Moran model in random environment and its approximation in large population }
\subsection{Discrete model with selection and immigration}
\quad\\
In this section we describe in detail the discrete model, i.e. the Moran process, which is the basic of our study. One may also consider here the Wright-Fisher discrete process with adequat change.
The Moran process is an evolution of population model, in which a single event occurs at each time step. More precisely each event corresponds to the death of an individual and the birth of another who replaces it.\\
We consider a population, whose size is constant over time equal to $J$, composed of $S+1$ species . The proportion of the $i$ species at the $n^{th}$ event is denoted $X^{i}_{n}$, $i \in \mathbb{S}=\{1,...,S+1\}$, $n\in \mathbb{N}$.\\As usual one we know $(X_n^{i})_{i=1,..,S}$, we deduce the proportion for the last species, $X_{n}^{S+1}=1-\sum\limits_{i=1}^{S}X^{i}_{n}$\\
We note $X_{n}$ the species vector or abundance vector having for coordinate $i$, $X_{n}^{i}$.
The dynamics of evolution follows the following pattern at the step $n$:
\begin{enumerate}
\item The individual designated to die is chosen uniformly among the community.
\item The one which replaces it, chooses his parent in the community with probability $1-m_{n}$ (filiation) or a parent from the immigration process with probability $m_{n}$ (immigration). The quantity $m_{n}$ varies between $0$ and $1$, it can be random and time dependent.
\item If there is immigration the chosen parent is from the species $i$, with probability $p^{i}_{n}$ $i \in \mathbb{S}$. The $p_{n}^{i}$ verify $\sum\limits_{i \in \mathbb{S}}p^{i}_{n}=1$ and can be time dependent and random. We note $p$ for the vector having for coordinate $i$, $p_{n}^{i}$.
\item
In a filiation, the chosen parent is of the species $i$ with probability $\frac{X^{i}_{n}(1+s^{i}_{n})}{1+\sum\limits_{k=1}^{S+1}X^{k}s^{k}_{n}}$. \\
The $s_{n}^{i}, i \in \mathbb{S}$ are the selection parameters , they may be time dependent and random.
Furthermore, we assume $s_{n}^{S+1}=0$. Indeed, we can obtain it from any configuration by changing all the coefficients by $s_{n}^{i}=\frac{\tilde{s}^{i}_{n}-\tilde{s}^{S+1}_{n}}{1+\tilde{s}_{n}^{S+1}} $.
\end{enumerate}
We will assume throughout this work that $m_n,p_n,s_n$ are autonomous, in the sense that their evolution do not depend on $(X_n)_{n\ge0}$. We will further assume that $(m_n,p_n,s_n)_{n\ge0}$ is a Markov chain. Note also that $m_n,p_n,s_n$ may also depend in some sense of the size of the population $J$, but we do not add another superscript to get lighter notations.\\
This model therefore describes a Markovian dynamic in which selection and immigration play an important role. Immigration already introduced by Hubbell \cite{hubbel} avoids the definitive invasion of the community by a species. Selection changes the dynamics of a species related to the neutral model( \cite{kalyuzhny}).
The temporal evolution of the population could be simulated numerically from the transition matrix of the Markov system. Let us describe precisely these transition probabilities for the evolution of proportions. The assumption for the dynamics for the immigration and selection will be given later on.\\
Let $x$ be the vector having for coordinate $i$, $x^{i}$ and suppose $m_{n}$ ,$p_{n}$ known.
Denote $\Delta=\frac{1}{J}$, so for the $i$ species:
\begin{eqnarray*}
P_{x^{i}+}&=& \mathbb{P}(X^{i}_{n+1}=x+\Delta|X_{n}=x)\\ & =&(1-x^{i})\left(m_{n}p^{i}_{n}+(1-m_{n})\frac{x^{i}(1+s^{i}_{n})}{1+\sum\limits_{k=1}^{S+1}x^{k}s^{k}_{n}}\right),
\end{eqnarray*}
\begin{eqnarray*}
P_{x^{i}-} &=&\mathbb{P}(X^{i}_{n+1}=x-\Delta|X_{n}=x))\\ & =&x^{i}\left(m_{n}(1-p^{i}_{n})+(1-m_{n})\left(1-\frac{x^{i}(1+s^{i}_{n})}{1+\sum\limits_{k=1}^{S+1}x^{k}s^{k}_{n}}\right)\right),
\end{eqnarray*}
\begin{eqnarray*}
P_{x^{i+}x^{j-}} &=&\mathbb{P}(\{X^{i}_{n+1}=x^{i}+\Delta\} \cap\{ X^{j}_{n+1}=x^{j}-\Delta\}|X_{n}=x)\\
&=&x^{j}\left(m_{n}(1-p_{n}^{i})+(1-m_{n})\frac{x^{i}(1+s_{n}^{i})}{1+\sum\limits_{k=1}^{s+1}x^{k}s_{n}^{k}}\right).
\end{eqnarray*}
With the dynamics of $(m_n,p_n,s_n)$ given, one may of course simulate exactly the vector of proportion $X_n$ and thus evaluate the expectation of the Simpson index, which is what we will do to validate our approximation procedure, but when the population size $J$ is very large, it may be computationally too costly (and even impossible).
Thus we will approach the dynamics of this model by a stochastic differential system continuous in time.
\subsection{To a limit in large population}
\quad\\
In this section, we explain how approaching the dynamics of the preceding model by a diffusion, and associated process for the immigration and selection processes, when $J$ goes to infinity. \\
We need to define a $J$ dependent time scale. Indeed, when $J$ goes to infinity, the time scale has to change, expectation and variance are about $\frac{1}{J}$ and $\frac{1}{J^2}$, and goes to $0$ when $J $ goes to infinity. It corresponds to considering a large number of event for the Markov chain, to obtain a non-trivial convergence of our discrete process towards a limit process, i.e. not look at the event-by-event evolution as we did before but in packets of several events. \\
Several choices for scales are possible, each one leads to study a different process. We choose to study a continuous multidimensional diffusion in time.\\
\subsubsection{Diffusion approximation}
\quad\\
In a general framework, the limiting process we obtain is characterized by the first moment and the covariance matrix of the infinitesimal variation of abundance. \\
More precisely if we note $\Delta_{t}$ the infinitesimal variation in time (which depends on the scale ) and $\Delta X_{t} =X_{t+\Delta_{t}}-X_{t}$ the infinitesimal variation of abundance, the diffusion process is characterized by the quantities:
$$b(x)=\lim\limits_{\Delta_{t}\Rightarrow 0}\frac{E[\Delta X_{t}|X_{t}=x]}{\Delta_{t}}$$
$$\sigma_{i,j}(x)=\lim\limits_{\Delta_{t}\Rightarrow 0}\frac{\mbox{Cov}[\Delta X_{t+\Delta_{t}}(i),\Delta X_{t+\Delta_{t}}(j)|X_{t}=x]}{\Delta_{t}} \quad i,j \in \mathbb{S}$$
\vspace{3mm}
The following property characterizes the order (relative to J) of the expectation, variance, and covariance of the abundance variation of a species during an event:
\begin{prop}
\begin{enumerate}
\item $\mathbb{E}[X^{i}_{n+1}-X^{i}_{n}|X_{n}=x]=\Delta (P_{x^{i}+}-P_{x^{i}-})$
\item $\mbox{Var}[X^{i}_{n+1}-X^{i}_{n}|X_{n}=x]=\Delta^{2}((P_{x^{i}+}+P_{x^{i}-})-(P_{x^{i}+}-P_{x^{i}-})^{2})$
\end{enumerate}
\end{prop}
The proof is standard calculus and thus omitted. The last property shows that the expectation is of the order of $\frac{1}{J}$ whereas the variance is of the order of $\frac{1}{J^2}$. The choice we make to preserve a stochastic part in our limit equation is to consider the infinitesimal time variation is of the order of $\frac{1}{J^2}$ .Other choices would have led to a Piecewise Deterministic Markov process in which only the parameters $s$, $m$, $p$ brings randomness. It will be left for further study.\\
\textbf{A scale in $\frac{1}{J^2}$ and the weak selection and immigration}.\\
Let $\Delta_{t}=\frac{1}{J^2}$ et $n=tJ^2$.\\
In this case, $\mathbb{E}[X_{t+\Delta_{t}}-x|X_{t}=x]=O(J)$ and the limit would be infinite. To hope for a finite term and thus to observe the influence of $ s $ and $ m $ in our limit DSE, we must assume that $ s $ and $m$ are inversely proportional to $ J $.
We now assume that migration and speciation are weak.
\begin{prop}\label{exp-var}
Let $\Delta_{t}=\frac{1}{J^2}$ and $n=tJ^2$, note
$m'(t)=m(t)\times J$ and $s'(t)=s(t)\times J $ .
So when $J$ goes to infinity:\\
\begin{enumerate}
\item $\mathbb{E}[X^{i}_{t+\Delta_{t}}-x^{i}|X_{t}=x]=\left(m_{t}'(p_{t}^{i}-x^{i})+x^{i}(s'^{i}_{t}-\sum\limits_{k\in \mathbb{S}}x^{k}s'^{k}_{t})\right)\times \Delta_{t}+o( \Delta_{t})$
\item $\mbox{\rm Var}[X^{i}_{t+\Delta_{t}}-x^{i}|X_{t}=x]=2x^{i}(1-x^{i})\Delta_{t} +o( \Delta_{t})$
\item $\mbox{\rm Cov}[X^{i}_{t+\Delta_{t}},X^{l}_{t+\Delta_{t}}|X_{t}=x]=-2x^{i}x^{l}\Delta_{t}+o( \Delta_{t}) \quad \forall i \neq l.$
\end{enumerate}
\end{prop}
We will now introduce notations and assumptions. Remark that a very recent work by Bansaye et al \cite{scallim} considered a very general framework for population model convergence in random environment, but in their work the environment is usually i.i.d. whereas we are in a Markovian setting. To get shorter statement and proofs, we will make considerable simplifications for our main assumption.
{\bf Assumption (A)}.\begin{itemize}
\item the process $(p_n)$ is assumed to be constant, which corresponds to a non evolving pool of immigration;
\item the process $(m^J)$ is an autonomous Markov chain, and consider its rescaled piecewise linear extension $\tilde m_t^J=J\,m^{J}_{\lfloor tJ^2\rfloor}$, which is assumed to take values in a finite set $E_m$ and uniformly bounded (in $J$). Let denote $P_{m^J}(m,m',t)$ its transition probabilities and assume for all $m\not=m'$,$$\lim\limits_{J\rightarrow\infty}P_{m^J}\left(m,m',\frac{1}{J^{2}}\right)\times J^{2} =Q(m,m')$$ is well defined;
\item the process $(s^J)$ is an autonomous Markov chain, and consider its rescaled piecewise linear extension $\tilde s_t^J=J\,s^{J}_{\lfloor tJ^2\rfloor}$, which is assumed to take values in a finite set $E_s$ and uniformly bounded (in $J$). Let denote $P_{s^J}(s,s',t)$ its transition probabilities and assume for all $s\not=s'$,$$\lim\limits_{J\rightarrow\infty}P_{s^{J}}\left(s,s',\frac{1}{J^{2}}\right)\times J^{2} =Q^{s}(s,s')$$ is well defined.
\item The processes $(m_n)$ and $(s_n)$ are supposed to be independent.
\end{itemize}
These assumptions about the limits of transitions probabilities are intended to ensure the convergence in law of $s$ and $m$ towards Markovian jump processes when $J$ goes to infinity.
Denote again $U^{J}_{t}=\begin{pmatrix}
X^{J}_{t} \\ s^{J}_{t}\\m_{t}^{J}
\end{pmatrix}$ taking values in $E=E_{x}\times E_{s}\times E_{m}$ a compact set of $\mathbb{R}^{2S+1}$ and for $\Gamma \in E$, we define for $k\in \mathbb{N}$, $t\in [\frac{k}{J²},\frac{k+1}{J²}[$ , $\pi_{J}(U_{t},\Gamma)=\pi_{J}(U_{\frac{k}{J^{2}}},\Gamma)=\mathbb{P}(U^{J}_{\frac{k+1}{J^{2}}}\in \Gamma| U^{J}_{\frac{k}{J^2}})$.
We are now in position to state our diffusion approximation result.
\begin{theorem}
\label{th:diffapprox}
Assume {\bf (A)}
then when $J$ goes to infinity the sequence of processes $(X_{t}^{J})$ converges in law to the process $(X_{t})$ whose coordinates are solutions of the following stochastic differential equation
\begin{equation}
\begin{pmatrix}
dX^{1}_{t}\\
\vdots\\
dX^{S}_{t}
\end{pmatrix}
=
\begin{pmatrix}
m_{t}\times (p_{t}^{1}-X^{1}_{t})+X^{1}_{t}\left(s_{t}^{1}-\sum\limits_{k\in \mathbb{S}}X^{k}s^{k}_{t}\right)\\
\vdots\\
m_{t}\times (p^{S}_{t}-X^{S}_{t})+X^{S}_{t}\left(s^{S}_{t}-\sum\limits_{k\in \mathbb{S}}X^{k}s^{k}_{t}\right)
\end{pmatrix}dt
+ \sigma(X_{t}) dB_{t} \label{eqprincipale2}
\end{equation}
where $\sigma$ is such that $\sigma.\sigma^{*}=a$ with $a_{i,j}=-2x^{i}x^{j}$ if $i\neq j$ and $a_{i,i}=2x^{i}(1-x^{i})$ and where
$s_{t}=\lim s_{t}^{J}J$ and $m_{t}=\lim m_{t}^{J}J$ are the Markovian jump processes with generators $Q^{s}$ and $Q^{m}$ and for initial conditions $s_{0}=\lim Js_{0}^{J}$ and $ m_{0}=\lim J m_{0}^{J}$. \\
\end{theorem}
\begin{proof}
The proof is in Section \ref{diffapprox} and relies on the usual Martingale Problem Method.
\end{proof}
Let us give some remarks
\begin{enumerate}
\item We may also consider the proportions $p_{i}$ random but their law would be $J$ dependent (through the change of time) and it is in disharmony with the biological model which assumes the pool independent of the community size.
\item Other types of processes for $s$ would have led to similar results, for example we could take a diffusion, with obvious modifications.
\item It is possible to give an upper bound of the error made by performing the diffusion approximation, by a direct approach not relying on the martingale problem method. It will be the purpose of \cite{ggp2018}.
\end{enumerate}
\subsubsection{Simpson index}
\quad\\
Our main object to quantify the biodiversity will be the Simpson index :
\begin{equation}
\label{simpson}
\mathscr{S}_t=\sum\limits_{i=1}^{S+1}(x_{t}^{i})^{2}
\end{equation}
with $x_{t}^{S+1}=1-\sum\limits_{i=1}^{S}x_{t}^{i}$. \\
Notice that this expression is the limits of the discrete Simpson index when $J$ goes to the infinity. Its dynamic is given by a non autonomous stochastic differential equation.
\begin{prop}
Denote as usual $p_{t}^{S+1}=1-\sum\limits_{i=1}^{S}p_{t}^{i},\quad \textnormal{and} \quad X^{S+1}_{t}=1-\sum\limits_{i=1}^{S}X^{i}_{t}$.\\
So the Simpson index is solution of the following equation :\\
\vspace{2mm}
$$d\mathscr{S}_{t}=2(1-\mathscr{S}_{t})-2\sum\limits_{i=1}^{S}s_{t}^{i}X_{t}^{i}(\mathscr{S}_{t}-X_{t}^{i})+2m_{t}\left(\sum\limits_{i=1}^{S+1}p^{i}X^{i}_{t}-\mathscr{S}_{t}\right)dt+dM_{t} \label{eqSn}$$\\
where $M_{t}$ is a martingale.\\
The drift is composed of three terms, the first is the drift in the neutral case without immigration (autonomous equation), the second is a term due to the presence of selection only and the third to the presence of immigration.
\end{prop}
This proposition follows from It\^o's calculus and details may be found in Section 5. As we will consider mainly the evolution of the expectation of the Simpson index, we do not describe the martingale term.
Now that the large limit SDE is established, we may consider the approximation of the Simpson index. We will first consider a simplified case, but which contains all the main difficulties: the two species case.
\section{Approximation of the Simpson index in the quenched or deterministic case: the two species case.}
In this part, we study a population of only two species. The equations obtained are in dimension one and the quantities are easier to calculate. It is a basic example to understand the dynamic in greater dimension.\\
In all this section we will suppose that the {\it selection and immigration processes are deterministic}, which also amounts to consider the quenched case, i.e. we fix the random environment (immigration and selection), as our goal will be to give a numerical method to approximate $\mathbb{E}[\mathscr{S}_{t}]$ at a lower cost. We will see in the next section how to consider the annealed case for a very particular case of selection and immigration.
In a second part, we will look for constant selection and immigration the behaviour of the process in long time.\\
The one dimensional Simpson index is $$\mathscr{S}_{t}=X_{t}^{2}+(1-X_{t})^{2}$$
Following the result of the previous section we are thus interested in the case where $X_{t}$ and $\mathscr{S}_{t}$ dynamics are given by
\begin{equation}
dX_{t}=m_{t}(p_{t}-X_{t})dt+s_{t}X_{t}(1-X_{t})dt+\sqrt{2X_{t}(1-X_{t}}dB_{t} \label{eq2}
\end{equation}
\begin{align}
d\mathscr{S}_{t}=&4X_{t}(1-X_{t})\times\left(1+s_{t}(X_{t}-\frac{1}{2})\right)+2m_{t}(p_{t}-X_{t})(2X_{t}-1)dt\nonumber\\
&+2(2X_{t}-1)\, \sqrt{2X_{t}(1-X_{t})}dB_{t} \label{eqS2}
\end{align}
where $m_t$ and $s_t$ are the rescaled limit immigration and selection processes.
\begin{remarque}
There is a first interesting feature when analysing the instantaneous behaviour of the dynamic of $\mathscr{S}_t$ in the case where there is no immigration (and thus the Simpson index will tend to 1): when $|s_t|<2$, the drift of $\mathscr{S}_t$ is always positive so that a even variability if the selection is small does not change the trend to non diversity. At the opposite, if the selection is sufficiently strong it may change locally the behaviour of the Simpson index, and we may thus imagine that change of fitness may lead to oscillation of the Simpson index. We will illustrate this phenomenon numerically.
\end{remarque}
We will now concentrate on a method to approximate all the moments of $X_t$, and thus an approximation of $\mathbb{E}[\mathscr{S}_t]$.
\subsection{Approximation of the moments of $X_{t}$}
\subsubsection{The approximation theorem}
\quad\\
As remarked earlier the momentum of $ \mathscr{S} _ {t} $ can not be calculated directly, as the equation of $\mathbb{E}[ \mathscr {S} _ {t} ]$ is not autonomous\eqref{eq2}. However we only need to evaluate the first two moments of $X_{t}$. We will see that it will be more difficult than planned. Indeed, taking expectation in \eqref{eqS2} (recalling that $m_t$ and $s_t$ are considered deterministic), we get:
\begin{equation}\label{eqmmt}
d\mathbb{E}[X_{t}]=m_{t}p_{t}-m_{t}\mathbb{E}[X_{t}]
+s_{t}\big(\mathbb{E}[X_{t}]-\mathbb{E}[X_{t}^{2}]\big)dt
\end{equation}
By analyzing \eqref{eqmmt}, we cannot express the first momentum of $X_{t}$ without the second moment and when considering the second moment, the third will appear and so on. It is thus impossible to express the momentum of $X_{t}$ as the solution of an autonomous equation (except for the trivial case where $s_t=0$)\\
Nevertheless, the following theorem gives a way of obtaining an approximation of the first moments of $X_{t}$ by solving a differential system whose size is all the greater that one wishes to be precise.
\begin{theorem}
Denote
$\tilde{A}^n_t$ the tridiagonal matrix whose coefficients are given by $\tilde{a}_{i+1,i}=(i+p_tm_t)(i+1)$, $\tilde{a}_{i,i}=i(s_t-i+1-m_t)$, $\tilde{a}_{i,i+1}=-is_t$ for i in $\{1,...,n-1\}$ and $\tilde{a}_{n,n}=-n(n-1+m_t)$
Let consider the following system of ordinary differential equations
$$d\tilde M_{t}=\tilde{A}^n_{t}\times \tilde M_{t}dt+\mathscr{C}_tdt$$
where $\mathscr{C}_t=(m_tp_t,0,...,0)^t$.\\
So for $j$ fixed, the $j^{th} $ coordinate of the solution $\tilde M_{t}$ converges when $ n $ (the size of the differential system) tends to infinity towards $j^{th}$ momentum of $X_{t}$. The error committed is at most $$\frac{\sqrt[]{n}\|s\|_{\infty}^{n-1}}{(n-1)!}.$$
\end{theorem}
As seen from the upper bound, the convergence is quite fast and even enable to approximate the Laplace transform of $X_t$ quite efficiently. It is mainly due to the fact that we have only two species here. We will see in the next section what happens for three species and will explain how it deteriorates with the number of species.
\subsubsection{Proof of the theorem}
We will begin by considering the error between the solution of our approaching system and the real solution. For practical reasons, each coordinate of the error is multiplied by a coefficient independent of the system size. Let us begin by giving the (non autonomous) system of ordinary differential equations verified by the moments of $X_t$
\begin{prop} Let $M_{t}$ being the vector having for coordinate $i$ the $\mathbb{E}[X_t^i]$ (up to $n^{th}$ moment). $M_{t}$ is solution of
$$dM_{t}=\tilde{A}_{t}^{n}\times M_{t}dt+\mathscr{C}_tdt+B_{t}dt$$
where $B_{t}=(0,...,0, ns_{t}\mathbb{E}[X_{t}^{n}(1-X_{t})])^T$
\end{prop}
\begin{proof}
\quad\\
It is of course a simple consequence of It\^o's Formula $X^{i}_{t}$:
\begin{align*}
dX_{t}^{i}=&iX_{t}^{i-1}\left(\left(m_{t}(p_{t}-X_{t})+s_{t}X_{t}(1-X_{t})\right)\times (i-1)(1-X_{t})\right)dt+d\mathscr{M}_{t}\\=&-is_{t}X_{t}^{i+1}+i\left(s_{t}-(i-1)-m_{t}\right)X_{t}^{i}+i(i-1+m_{t}p_{t})X_{t}^{i-1} dt+d\mathscr{M}t_{t}
\end{align*}
where $\mathscr{M}$ is a martingale and by taking the expectation
$$d\mathbb{E}[X_{t}^{i}]=-is_{t}\mathbb{E}[X_{t}^{i+1}]+i(s_{t}-(i-1)-m_{t})\mathbb{E}[X_{t}^{i}]+i(i-1+m_{t}p_{t})\mathbb{E}[X_{t}^{i-1}] dt$$
to recover the coefficients of the matrix $\tilde{A}_{t}^{n}$.
\end{proof}
We define the vector of errors $\Delta_{t}^{n}$ by $$\Delta_{t}^{n}(i)=(s_{t})^{i-1}\times\frac{M_{t}^{i}-\tilde{M_{t}}^{i}}{(i-1)!},$$
and we introduce also $\tilde\mu^n_t=(0,...,0,\mu_t^n)^T$ with $$\mu_{t}^{n}=\frac{(s_{t})^{n-1}}{(n-1)!}\times ns_{t}E[X_{t}^{n}(1-X_{t})].$$
Note that when we multiply, for $i$ fixed, the difference $M_{t}^{i}-\tilde{M_{t}^{i}}$ by a coefficient independent of $n$, the speed of convergence of the $i^{th}$ coordinate of the error change but not the fact that this quantity tends towards $0$. On the other hand, it forces the dependent coordinates of $n$ to tend to $0$. So we just need to prove that $\|\Delta_{t}^{n}\|$ goes to $0$, which is not the case for $M_{t}^{i}-\tilde{M_{t}^{i}}$.
\begin{prop}
\quad
$\Delta_{t}^{n}$ is the solution of the following differential system:
\begin{equation}
\left\{
\begin{aligned}
&d\Delta_{t}^{n}=A_{t}^{n}\times \Delta_{t}^{n}dt+\tilde{\mu}_{t}^{n}dt\\
&\Delta^{n}_{0}=0
\end{aligned}
\right.
\label{Em}
\end{equation}
and $$A_{t}^{n}=\begin{pmatrix}
a_{t}^{1}&b_{t}^{1}&&&(0)\\
c_{t}^{1}&a_{t}^{2}&b_{t}^{2}\\
&\ddots&\ddots&\ddots&\\
&&c_{t}^{n-2}&a_{t}^{n-1}&b_{t}^{n-1}\\
(0)&&&c_{t}^{n-1}&a_{t}^{n}
\end{pmatrix}$$
with
\begin{align*}
a_{t}^{i}&=i\big(s_{t}-(i-1)-m_{t}\big), \quad i< n\\
a_{t}^{n}&=-n\big((n-1)+m_{t}\big)\\
b_{t}^{i}&=-i^{2} \quad i< n\\
c_{t}^{i}&=s_{t}i(1-\frac{m_{t}p_{t})}{i-1})\quad i< n\\
\end{align*}
\end{prop}
\begin{proof}
\quad\\
Indeed, by substraction, $$d(M_{t}^{n}-\tilde{M}_{t}^{n})=\tilde{A_{t}^{n}}\times (M_{t}^{n}-\tilde{M}_{t}^{n})+\begin{pmatrix}
0\\ \vdots\\ 0\\s_{t}nE[(X_{t})^{n}(1-X_{t})]\end{pmatrix}.$$
This involves for all $ k<n$\\
\begin{align*}
d(M_{t}^{k}-\tilde{M}_{t}^{k})&=k(k-1+p_{t}m_{t})(M_{t}^{k-1}-\tilde{M}_{t}^{k-1})\\&+k(s_{t}-(k-1)-m_{t})(M_{t}^{k}-\tilde{M}_{t}^{k})-ks_{t}(M_{t}^{k+1}-\tilde{M}_{t}^{k+1}).\\
\end{align*}
Next, we multiply by$\frac{(s_{t})^{k-1}}{(k-1)!}$,
\begin{align*}
d(\Delta_{t}^{k,n})&=\frac{(M_{t}^{k-1}-\tilde{M}_{t}^{k-1})s_{t}^{k-2}}{(k-2)!}\frac{s_{t}k(k-1+p_{t}m_{t})}{k-1}\\&\qquad+\frac{ks_{t}^{k-1}}{(k-1)!}\left(s_{t}-(k-1)-m_{t})(M_{t}^{k}-\tilde{M}_{t}^{k}\right)\\&\qquad
-\frac{s_{t}^{k}(M_{t}^{k+1}-\tilde{M}_{t}^{k+1})}{(k)!}k^{2}\\
&=ks_{t}\big(1-\frac{m_{t}p_{t})}{k-1}\big)\Delta_{t}^{k-1,n}+k(s_{t}-(k-1)-m_{t})\Delta_{t}^{k,n}-k^{2}\Delta_{t}^{k+1,n}.
\end{align*}
Now if k=n,
\begin{eqnarray*}
d\Delta_{t}^{n,n}&=&n\left[s_{t}\left(1-\frac{m_{t}p_{t}}{n-1}\right)\Delta_{t}^{n-1,n}-(m_{t}+n-1)\Delta_{t}^{n,n}\right.\\
&&\qquad\left.+\frac{s_{t}^{i-1}}{(n-1)!} s_{t}\mathbb{E}[X_{t}^{n}(1-X_{t})]\right]
\end{eqnarray*}
We thus find the coefficients of the previous equation system.
\end{proof}
We may now provide the solution to the system of equation for the error.
\begin{prop}
The solution of $\eqref{Em}$ can be written $$\Delta_{t}^{n}=\int_{0}^{t}\exp\left(\int_{q}^{t}A_{u}^{n}\right)\tilde{\mu}_{q}^{n}dq$$
\end{prop}
\begin{proof}
First, the solution of $\eqref{Em}$ can be written as the solution of the homogeneous system and a particular solution.
$$\Delta_{t}^{n}=\mathcal{K}\exp\left(\int_{q}^{t}A_{u}^{n}du\right)+\exp\left(\int_{0}^{t}A_{u}^{n}du\right)\int_{0}^{t}\exp\left(-\int_{0}^{q}A_{u}^{n}du\right)\tilde{\mu}_{t}^{n}(q)dq
$$ As $\Delta_{0}^{n}=0$ necessarily $\mathcal{K}=0$ and $\Delta_{t}^{n}=\int_{0}^{t}\exp\left(\int_{q}^{t}A_{u}^{n}du\right)\tilde{\mu}_{t}^{n}(q)dq.$
\end{proof}
We will now show that, for a fixed time interval, $\|\Delta_{t}^{n}\|_{2}$ is uniformly bounded by a quantity which goes to 0 when n goes to infinity. In the following, $\|.\|_{F}$ stands for the Frobenius norm. Thanks to the formula of the previous proposition:
\begin{align*}
\|\Delta_{t}^{n}(t)\|_{2}^{2}&=\left\|\int_{0}^{t}\exp\left(\int_{q}^{t}A_{u}^{n}du\right)\tilde{\mu}_{q}^{n}dq\right\|_{2}^{2}\\
&=\sum\limits_{k=1}^{n}\left(\int_{0}^{t}(\exp\left(\int_{q}^{t}A_{u}^{n}du\right)\tilde{\mu}_{q}^{n})dq\right)_k^{2}\\
&\leqslant t\times \sum\limits_{k=1}^{n}\int_{0}^{t}\left(\exp\left(\int_{q}^{t}A_{u}^{n}du\right)\tilde{\mu}_{q}^{n}\right)_{k}^{2}dq\\
&\leqslant t \int_{0}^{t}\left\|\exp\left(\int_{q}^{t}A_{u}^{n}du\right)\tilde{\mu}_{q}^{n}\right\|_{2}^{2}dq \\
&\leqslant t \int_{0}^{t}\left\|\exp\left(\int_{q}^{t}A_{u}^{n}du\right)\right\|_{F}^{2}\,\|\tilde{\mu}_{q}^{n}\|_{2}^{2}
dq\\
&\leqslant t^{2} \times \sup\limits_{x\in[0,t]}\left\|\exp\left(\int_{x}^{t}A_{u}^{n}du\right)\right\|_{F}^{2}\times \sup\limits_{x\in[0,t]}|\mu_{x}^{n}|^{2}
\end{align*}
The following easy lemma gives us an upper bound for the two previous norms:
\begin{lemma}
\quad\\
If $X \in [0,1]$ then $sX^{n}(1-X)=O(\frac{1}{n})$ and there is a constant $ c_ {1} $ independent of n such that$\sup\limits_{x\in[0,t]}|\mu_{x}^{n}|\leqslant \sup\limits_x\in[0,t]s_{x}^{n-1}\frac{c_{1}}
{(n-1)!}$
\end{lemma}
\begin{proof}
The maximum of $X^{n}(1-X)$ on $[0,1]$ is achieved in $\frac{n}{n+1}$ and it's worth $(1-\frac{1}{n+1})^{n}\times \frac{1}{n+1}$.\\
This quantity is of the order of $\frac{1}{n}$ when $n$ goes to infinity.\\
The upper bound of $\mu_{x}^{n}$ follow.
\end{proof}
Next, \begin{align*}
\left\|\exp\left(\int_{x}^{t}A_{u}^{n}du\right)\right\|_{F}^{2}=&Tr\left(\exp\left(\int_{x}^{t}A_{u}^{n}du\right)^{T}\times \exp\left(\int_{x}^{t}A_{u}^{n}du\right)\right)\\
=&Tr\left(\exp\left(\int_{x}^{t}A_{u}^{n}+(A_{u}^{n})^{T}du\right)\right)\\
=&\sum\limits_{\lambda_{i} } \exp\left(\lambda_{i}^{x}\right)\\
\leqslant & n\times \max\limits_{i \in \{1...n\}}(\
\exp(\lambda_{i}^{x}))
\end{align*}
where the $\lambda_{i}^{x}$ are the eigenvalues of $\int_{x}^{t}A_{u}^{n}+(A_{u}^{n})^{T}du$ .\\
Then if $\max\limits_{i \in \{1...n\}}\exp(\lambda_{i}^{x})$ is independent of $n$, there is a constant $c_{2}$ independent of $n$ verifying $$ \sup\limits_{x\in[0,t]}\left\|\exp(\int_{x}^{t}A_{u}^{n}du)\right\|_{F}\leqslant \sqrt[]{n}\times c_{2}$$
And then there exist a constant $C$ such as $$\|\Delta_{t}^{n}(t)\|_{2}\leqslant\sqrt[]{n} c_{2}\times \frac{c_{1}\|s\|_{\infty}^{n-1}}{(n-1)!}\leqslant\frac{C\sqrt[]{n}\|s\|_{\infty}^{n-1}}{(n-1)!} $$
It remains to show that the eigenvalues of $\int_{x}^{t}A^{n}_{u}+(A_{u}^{n})^{T}du$ (which are real) have an upper bound independent of n.
\begin{prop}
Let $A_{t}^{n}+(A_{t}^{n})^{T}=
\begin{pmatrix}
a_{t}^{1}&b_{t}^{1}&&&\\
b_{t}^{1}&a_{t}^{2}&b_{t}^{2}\\
&\ddots&\ddots&\ddots&\\
&&b_{t}^{n-2}&a_{t}^{n-1}&b_{t}^{n-1}\\
&&&b_{t}^{n-1}&a_{t}^{n}
\end{pmatrix}
$
with
\begin{align*}
a_{t}^{i}&=2i\big(s_{t}-(i-1)-m_{t}\big), \quad i< n\\
a_{t}^{n}&=-n\big((n-1)+m_{t}\big)\\
b_{t}^{i}&=s_{t}i(1-\frac{m_{t}p_{t})}{i-1})-(i-1)^{2}
\end{align*}
then the eigenvalues of $\int_{x}^{t}(A_{u}^{n})^{T}+A_{u}^{n}du$ are uniformly bounded in $n$ .
\end{prop}
\begin{proof}
To prove this result we need to use Gershgorin's disk. The eigenvalues of $\int_{x}^{t}(A_{u}^{n})^{T}+A_{u}^{n}du$ are included in the union of the disk $D_{i}$ whose centers are the $i^{th}$ term on the diagonal ($\int_{x}^{t}a_{u}^{i}du$) and for radius the sums of the coefficients norms on the $i^{th}$ line except the diagonal term$(|\int_{x}^{t}b_{u}^{i-1}du|+|\int_{x}^{t}b_{u}^{i}du|)$. It is important to consider the forms of the discs in our case.
As we have to show that the eigenvalues have a upper bound independently of n. We just need to look at the shape of the discs for $i $ and $n$ big enough. From the matrix if $i<n$, their centers are
$$2i\left(\int_{x}^{t}s_{u}du-(i-1)(t-x)-\int_{x}^{t}m_{u}du\right)$$ and their radius are equal to
\begin{align*}
\left|\int_{x}^{t}b_{u}^{i-1}du\right|+\left|\int_{x}^{t}b_{u}^{i}\right|&=\left|\int_{x}^{t}s_{u}i\left(1-\frac{m_{u}p_{u}}{i-1}\right)-(i-1)^{2}du\right|\\&\qquad+\left|\int_{x}^{t}s_{u}(i+1)\left(1-\frac{m_{u}p_{u}}{i}\right)-i^{2}du\right|\\
&=\left|i\int_{x}^{t}s_{u}du-(i^2-2i)(t-x)+i\int_{x}^{t}m_{u}du\right|\\&\qquad+\left|i\int_{x}^{t}s_{u}du-i^2(t-x)-i\int_{x}^{t}m_{u}du\right|+O(1)\\
&=2i\left((i-1)(t-x)-\int_{x}^{t}s_{u}du+\int_{x}^{t}m_{u}du\right)+O(1).
\end{align*}
So the maximum value for an eigenvalue of $\int_{x}^{t}(A_{u}^{n})^{T}+A_{u}^{n}du$ belonging to $D_{i}$ is
\begin{align*}
&\int_{x}^{t}a_{u}^{i}du+\left|\int_{x}^{t}b_{u}^{i-1}du\right|+\left|\int_{x}^{t}b_{u}^{i}du\right|\\&\qquad\qquad=2i\left(\int_{x}^{t}s_{u}du-(i-1)(t-x)-\int_{x}^{t}m_{u}du\right)\\&\qquad\qquad\qquad+2i\left((i-1)(t-x)-\int_{x}^{t}s_{u}du+\int_{x}^{t}m_{u}du\right)+O(1)\\ &\qquad\qquad=O(1)
\end{align*} as soon as $i$ is big enough. If $i=n$ the same reasoning is still working. So the eigenvalues of $(A^{n})^{T}+A^{n}$ have an upper bound independent of $n$.
\end{proof}
This property concludes the proof and we get the following upper bound$$\|\Delta_{t}^{n}(t)\|_{\infty}\leqslant\|\Delta_{t}^{n}(t)\|_{2}\leqslant\frac{C\sqrt[]{n}\|s\|_{\infty}^{n-1}}{(n-1)!} $$
The constant $ C $ depends on time (exponentially) and therefore this algorithm will be less accurate if we look at the behavior of the process in long time. This result gives a satisfactory approximation of the $ \mathscr {S} _ {t} $'s moment. Convergence is very fast, and the algorithm boils down to solving a linear differential system. To ensure the interest of this method we can compare the expectation of Simpson's index obtained by a Monte Carlo method from the discrete model to that obtained with this approximation. The Figure 1 presents such an approximation.
\begin{figure}[htbp]
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{MMCvsAPPROX_XT.png}
\end{center}
\end{minipage}
\hfill
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{MMCvsAPPROX.png}
\end{center}
\end{minipage}
\label{fig:image1}
\caption{Are plotted the approximate values of $E[\mathscr{S}_{t}]$ and $E[X_{t}]$ by the precedent method from the approximation in large population and by MC method from the discreet model. The number of simulated trajectories for MC mean is $ 500$ (red and blue), $J=1000$,$m=2$,$p=0.5$,$X_{0}=0.2$, $s$ switches between $2$ and $-2 $ at regular time intervals, the size of the approaching linear system is $100$.}
\end{figure}
\subsection{Numerical applications}
\quad\\
The simulations presented in this part are obtained from the previous theorem.
The values of $ s, m, p $ are those of the large population approximation and not that of the discrete model.
The size of the approaching system will be usually between 80 and 144 depending on the needed precision.
\subsubsection{Influence of $s$ on Simpson Index.}
\quad\\
In this part, $p=0.5$. Now we know how to approximate the expectation of $ \mathscr {S} _ {t} $, so we can check the influence of $ s $ on this quantity. Let us make precise a statement enounced when deriving $\mathscr{S}_t$.
\begin{prop}
\quad\\
If $\|s\|_{\infty}$ is smaller than $2$ and if $m=0$, $E[ \mathscr {S} _ {t} ]$ is increasing.
\end{prop}
\begin{proof}
\quad\\
If we refer to the equation of $ dE [\mathscr {S} _ {t}] $ (cf \eqref {eqS2}), we see that if $\|s\|_{\infty}<2$, the quantity $ 4X_{t}(1-X_{t})\times\left(1+s_{t}(X_{t}-\frac{1}{2})\right)$ is positive whatever the initial condition, so Simpson's index mean is always growing .
\end{proof}
In other words, selection alone can not bring about a renewal of biodiversity.
On the other hand if for some $t$, $s_{t}> 2$, $ E [\mathscr {S} _ {t}] $ can decrease.
In this case the more the selection is important, more the decay is pronounced.
We will see in the last part that this phenomenon can be generalized to a larger number of species. Figure 2 present these different behaviours with respect to $s$ and Figure 3 the combination of initial parameters and selection under which the Simpson's index is decreasing.
\begin{figure}[htbp]
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{ES_for_several_s.png}
\label{fig:image2}
\caption{Several trajectories of $E [\mathscr {S} _ {t}]$ are drawn for different s. $X_{0}=0.1$, m=0. }
\end{center}
\end{minipage}
\hfill
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{Figure_1.png}
\label{fig:image3}
\caption{The colored area represents the pairs $(X, s)$ for which the slope of the Simpson's index will be negative.}
\end{center}
\end{minipage}
\end{figure}
\subsubsection{Approximation of $T_{1}$, $T_{0}$, $T_{1,0}$.}
\quad\\
In the special case where $m_{t}=0$, a species inevitably invaded the community in a finite time. We define by $T_{1}$, (respectively $T_{0}$), the smallest time from which the process $X_{t}$ reaches $1$ (respectively $0$) and $T_{1,0}=min( T_{1}, T_{0}) $.\\
Thanks to the approximation of moments we can obtain an approximation of the $ T_ {1} $ distribution function .
In fact we know $X_{t}^{n}$ for $n$ big enough and so since $\lim\limits_{n\rightarrow \infty}E[(X_{t})^{n}]=1\times \mathbb{P}(T_{1}<t)$ we obtain $P(T_{1}<t)$.
The same way with $1-X_{t}$ gives $\mathbb{P}(T_{0}<t)$. Figure 4 gives an approximation of the distribution function of $T_1$ when there is no immigration.
\begin{figure}[htbp]
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{distributionfonctionT.png}
\end{center}
\end{minipage}
\caption{Distribution function of $T_{1}$ for m=0,p=0.5,$X_{0}=0.2$, $s=2$. The size of the approaching linear system is 100}
\label{fig:image4}
\end{figure}
We can use the same method to obtain $T_{\mathscr{S}_{t}}$ the probability that $\mathscr{S}_{t}$ is equal to $1$ at time $t$.
\subsection{Long time behavior}
\quad\\
Two cases are distinguished in this part, the case $ m = 0 $ and the case $ m \neq 0 $. In this first case there is no immigration and necessarily a species invade the community. Invasion times and the probability that the species with a selective advantage will invade the community are calculated. In the second case the system admits an invariant measure, we explain it and we specify the speeds of convergence towards this measure.
More details about the behavior of stochastic processes in long time can be found in \cite {evo} and \cite {SMB}.
\subsubsection{The case without immigration ($m=0$): absorption}
\quad\\
The results in this section are partially well known and we include them only to get a full picture of the behavior of the Simpson index.\\ Recall that $\mathscr{S}_{t}$ satisfies the equation
\begin{equation}
d\mathscr{S}_{t}=4X_{t}(1-X_{t})\times\big(1+s_{t}(X_{t}-\frac{1}{2})\big)dt+dM_{t}.
\label{eqS2b}
\end{equation}
If there is no more immigration the states $0$ and $1$ are absorbing, and it is then well known that the process reaches them in finite time almost surely. For more details about this refer to \cite{evo}. Let $T_{1}$ et $T_{0}$ the hit times of 1 and 0 for the random variable$X_{t}$ et $T_{1}\wedge T_{0}=T_{1, 0}$.
\begin{prop}\label{prop:absorb}
Suppose $s_{t}=s \in \mathbb{R}$,
\begin{enumerate}
\item $T_{1, 0}<\infty$ almost surely, for all initial condition $X_{0}$.\\
\item Let g be the solution of
\begin{equation}
x(1-x)g"(x)+sx(1-x)g'(x)=-1 \quad et \quad
g(0)=g(1)=0 \label{g}
\end{equation}
then $ E_{X_{0}}[T_{1, 0}]=g(X_{0})$. \\
\item $P(T_{1}<T_{0})=\frac{e^{-sX_{0}}-1}{e^{-s}-1}.$
\end{enumerate}
\end{prop}
The proof is given in section \ref{absorb}
Let us consider some particular case which illustrates that the same behavior may be obtained with varying selection. Suppose $T>0$ and $s_{t}$ is a constant function on intervals $[kT,(k+1)T]$, $k\in \mathbb{N}$ which can take the values $ s_ {0} $ or $ -s_ {0} $ for $ s_ {0}> 0 $, randomly. We will establish a result similar to the constant case. Let us begin by the following lemma which asserts that without selection one may reach the boundary at any time.
\begin{lemma}
Consider the following process
\begin{equation}
dX_{t}=\sqrt[]{2X_{t}(1-X_{t})}dB_{t} \label{eq2neutre}
\end{equation}
Note $T_{1}=\inf\{t,X_{t}=1\} $, an initial condition $x \in ]0,1]$ and a time $t$ .\\
Then $\mathbb{P}_{x}(T_{1}<t)>0$, in other words, $ 1 $ is accessible for $ X_ {t} $ from any non-zero initial condition and in a as little time as one wants.
\end{lemma}
\begin{proof}
Assume that $t_{0}=\inf\{t,\mathbb{P}_{x}(T_{1}<t)>0\}>0$.\\
$t_{0}$ is well defined. Remark that $ \mathbb{E} [X_ {t}] $ is constant in time because $ X_ {t} $ is a bounded martingale. Then $0<x=\lim\limits_{t\rightarrow\infty}E[X_{t}]=1\times \mathbb{P}_{x}(T_{1}<T_{0})+0$ as $\mathbb{E}_{x}[T_{1,0}]<\infty$ and so $\mathbb{P}_{x}(T_{1}<\infty)>0$. Then, let $\Delta_{t}>0$ be such that $t_{0}-2\Delta_{t}>0$. We will show then that there is $ y $ such that$\mathbb{P}_{y}(T_{1}<\Delta_{t})>0$.\\
If that was not the case then $\forall y \in [0,1],\mathbb{P}_{y}(T_{1}<\Delta_{t})=0 $ and
\begin{align*}
\mathbb{P}_{x}(T_{1}<t_{0})=&\mathbb{E}_{x}[\mathds{1}_{T_{1}<t_{0}}]=\mathbb{E}_{x}[\mathds{1}_{T_{1}<t_{0}}\times(\mathds{1}_{T_{1}<\Delta_{t}}+\mathds{1}_{T_{1}>\Delta_{t}})]\\
=&\mathbb{E}_{x}[\mathds{1}_{T_{1}<\Delta_{t}}]+\mathbb{E}_{x}[\mathds{1}_{T_{1}<t_{0}}\times\mathds{1}_{T_{1}>\Delta_{t}}]\\
=&\mathbb{E}_{x}[\mathds{1}_{T_{1}<t_{0}}\times\mathds{1}_{T_{1}>\Delta_{t}}]=\mathbb{E}_{x}[\mathds{1}_{T_{1}>\Delta_{t}}E[\mathds{1}_{T_{1}<t_{0}}|X_{\Delta_{t}}]]\\
=&\mathbb{E}_{x}[\mathds{1}_{T_{1}>\Delta_{t}}\times \mathbb{P}_{X_{\Delta_{t}}}(T_{1}<t_{0}-\Delta_{t})]\\
=&\mathbb{E}_{x}[\mathds{1}_{T_{1}>\Delta_{t}} \mathbb{E}_{X_{\Delta_{t}}}[\mathds{1}_{T_{1}>\Delta_{t}} \cdots \mathbb{E}_{X_{n\Delta_{t}}}[\mathds{1}_{T_{1}>\Delta_{t}}\\&\qquad\qquad\qquad\times \mathbb{P}_{(n+1)\Delta_{t}}(T_{1}<t_{0}-(n+1)\Delta_{t})]]].
\end{align*}
Let us choose $n$ such that $t_{0}-(n+1)<\Delta_{t}$. Then $\mathds{1}_{T_{1}>\Delta_{t}}\times \mathbb{P}_{(n+1)\Delta_{t}}(T_{1}<t_{0}-(n+1)\Delta_{t})\leqslant \mathds{1}_{T_{1}>\Delta_{t}}\mathbb{P}_{(n+1)\Delta_{t}}(T_{1}<\Delta_{t})=0$ and $\mathbb{P}_{x}(T_{1}<t_{0}) =0$ which is contrary to the assumptions. Now, we show that $\mathbb{P}_{x}(T_{1}<t_{0}-\Delta_{t})>0$:
\begin{align*}
\mathbb{P}_{x}(T_{1}<t_{0}-\Delta_{t})&=\mathbb{E}_{x}[\mathds{1}_{T_{1}<t_{0}-\Delta_{t}}\times(\mathds{1}_{T_{y}<t_{0}-2\Delta_{t}}+\mathds{1}_{T_{y}>t_{0}-2\Delta_{t}})]\\
& \geqslant \mathbb{E}_{x}[\mathds{1}_{T_{1}<t_{0}-\Delta_{t}}\times\mathds{1}_{T_{y}<t_{0}-2\Delta_{t}}]\\
&\geqslant \mathbb{E}_{x}[\mathds{1}_{T_{y}<t_{0}-2\Delta_{t}}E[\mathds{1}_{T_{1}<t_{0}-\Delta_{t}}|T_{y}]]\\
&\geqslant \mathbb{E}_{x}[\mathds{1}_{T_{y}<t_{0}-2\Delta_{t}}\mathbb{P}_{y}(T_{1}<t_{0}-\Delta_{t}-T_{y})]\\
&\geqslant \mathbb{E}_{x}[\mathds{1}_{T_{y}<t_{0}-2\Delta_{t}}\mathbb{P}_{y}(T_{1}<\Delta_{t})]\\
&\geqslant \mathbb{P}_{x}(T_{y}<t_{0}-2\Delta_{t})\times \mathbb{P}_{y}(T_{1}<\Delta_{t}).
\end{align*}
But we know that $ \mathbb{P}_{y}(T_{1}<\Delta_{t})>0$ by the previous calculation and the local uniform ellipticity of diffusion also ensures us that $\mathbb{P}_{x}(T_{y}<t_{0}-2\Delta_{t})>0$. So we obtain $ \mathbb{P}_{x}(T_{1}<t_{0}-\Delta_{t})>0$ which is contrary to the fact that $t_{0}>0$.
So $\forall t>0, \mathbb{P}_{x}(T_{1}<t)>0$ which concludes the proof of the lemma.
\end{proof}
Of course, this result may be adapted for the process $dX_{t}=sX_{t}(1-X_{t})dt+\sqrt[]{2X_{t}(1-X_{t})}dB_{t}.$ Indeed if $s>0$ the drift goes in the right direction. Else, if $s<0$, we obtain a symmetric result by replacing in the previous reasoning $ T_ {1} $ by $ T_ {0} $.
\begin{prop}
\quad\\
Let $s_{0}>0$, $T>0$ and $s_{t}$ a constant function on the intervals $[kT,(k+1)T]$, $k\in \mathbb{N}$ which can take the values $ s_ {0} $ or $-s_ {0} $ values randomly.\\
Let consider the process
\begin{equation}
dX_{t}=s_{t}X_{t}(1-X_{t})dt+\sqrt[]{2X_{t}(1-X_{t})}dB_{t} \label{eq2b}
\end{equation}
Finally $T_{1,0}=\inf\{t,X_{t}=0 \quad ou \quad X_{t}=1\}$ then $\mathbb{P}_{x}(T_{1,0}<\infty)=1$.
\end{prop}
\begin{proof}
The idea is to show that for each time interval of size $ T $, the probability of reaching $ 1 $ or $ 0 $ is non-zero and independent of the position where the process is located. So we compare the probability that our process reaches $ 1 $ or $ 0 $ to a geometric law. So let us first show that $\forall t \in ]0,T],\exists \alpha>0$, such that $ \forall x \in [0,1], \mathbb{P}_{x}(T_{1,0}<t)>\alpha$. Suppose $s>0$ on $[0,T]$ and denote $g_{i}(x)= \mathbb{P}_{x}(T_{i}<t)$, $i \in \{0,1\}$. Both functions are continuous, $g_{0}$ is decreasing and $g_{0}(0)=1$, $g_{0}(1)=0$, whereas $g_{1}$ is increasing and $g_{1}(1)=1$, $g_{1}(0)=0$.
Then there exist a $x_{0} \in ]0,1[$ such as $g_{0}(x_{0})=g_{1}(x_{0})=\alpha_{+}$. And by the previous lemma since $ s $ does not vary on $] 0, T] $ we have that $\alpha_{+}>0$. A symmetric reasoning for $ s <$ 0 guarantees us the existence of a $\alpha_{-}>0$.\\
Let $\alpha =2\min(\alpha_{-},\alpha_{+})$, $\alpha$ is then strictly positive and $ \forall x \in [0,1]$, $\mathbb{P}_{x}(T_{1,0}<t)= \mathbb{P}_{x}(T_{1}<t)+ \mathbb{P}_{x}(T_{0}<t)\geqslant \alpha$. Then, for $t\in [nT,(n+1)T]$, using previous inequality:
\begin{align*}
\mathbb{P}_{x}(T_{1,0}<t)\geqslant& \mathbb{P}_{x}(T_{1,0}<nT)\\
\geqslant &\mathbb{E}_{x}[\mathds{1}_{T_{1,0}<(n-1)T}]+\mathbb{E}_{x}[\mathds{1}_{T_{1,0}>(n-1)T}\mathbb{P}_{X_{(n-1)T}}(T_{1,0}<T)]\\
\geqslant&\mathbb{P}_{x}(T_{1,0}<(n-1)T)+\alpha \times \mathbb{E}_{x}[1-\mathds{1}_{T_{0,1}<(n-1)T}]\\
\geqslant& (1-\alpha)\times \mathbb{P}_{X_{(n-1)T}}(T_{1,0}<(n-1)T)+\alpha\\
\geqslant& \alpha+\alpha(1-\alpha)+....+\alpha(1-\alpha)^{n-1}
\geqslant 1-(1-\alpha)^{n}
\end{align*}
We obtain when $n$ goes to infinity $\mathbb{P}_{x}(T_{1,0}<\infty)=1$.
\end{proof}
Even with frequent changes of fitness, a species always ends up invading the community if $m$ is zero. We may then consider the process with immigration. Note that while preparing this paper, comparable (and even more general) results were obtained (in the multi-allelic case) by Coron et al \cite{cmv}.
\subsubsection{The case with immigration ($m\neq 0$): invariant measure}
\quad\\
Assume now $s$, $m$ and $p$ are constants. The long time behavior for varying selection and immigration is far more complicated and may lead to interesting behavior that will be considered in another paper. We thus consider the following process:
$$dX_{t}=m(p-X_{t})+sX_{t}(1-X_{t})dt+\sqrt[]{2X_{t}(1-X_{t})}dB_{t}$$
and will consider the long time behavior in a quantitative way, i.e. not using Meyn-Tweedie's theory, but rather via a Poincar\'e inequality.\\
Our process $ X_ {t} $ is Markovian and evolves in a range bounded by 0 and 1. At first we will ask ourselves what is the behaviour of our process in the neighbourhood of 0 and 1, by considering the criterion given by Feller cf\cite{Feller1,Feller2}. According to the values of $ m $, $ p $ and $ s $, our process will have different behaviours in the neighbourhood of $1$ and $0$.
We have already seen that if $m=0$ then $0$ and $1$ are absorbing states reached by the process in a finite time almost surely.(The same hold if $ m $ is non-zero and if $ p $ is $0$ or $1$.) Now if $ m $ and $ p $ are not trivial, $0$ and $1$ are no longer absorbing. In other words, immigration prevents the invasion of the community by a species. Moreover, for some values of $ m $ and $ p $ these two states are not accessible, i.e the process can not reach them in a finite time.
\begin{prop}\label{prop:access}
The state $ 1 $ (respectively $ 0 $) is accessible by the process $ X_ {t} $ if and only if $ m (1-p) <1 $ (respectively $ mp <1 $) and regular otherwise.
\end{prop}
The proof will be given in section \ref{access}. In the case of inaccessible or reflective boundaries (which is our the case), the law of the process $ X_ {t} $ admit a density and converge in long time to an invariant measure. This measure has a density, denoted $ \pi $. In addition $ \pi $ is a solution of the Fokker-Planck equation:
$$0=\frac{\partial }{\partial t}\pi=-\frac{\partial\pi }{\partial y}\big(m(p-y)+sy(1-y)\big)+\frac{\partial² \pi}{\partial^{2} y}y(1-y)$$
The solution of this equation is:
\begin{equation}
\pi(y)=c\times y^{mp-1}\times (1-y)^{m(1-p)-1}\times \exp(sy) \label{mesinv2}
\end{equation}
The constant $c$ is chosen so that $\int_{0}^{1}\pi(y)dy=1$.
The following Figures 5 and 6 show the influence of the parameters on the expectation and the variance of Simpson's equilibrium index.
\begin{figure}[htbp]
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{Savecm.png}
\label{fig:image2b}
\end{center}
\end{minipage}
\hfill
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{VarSavecm.png}
\label{fig:image2c}
\end{center}
\end{minipage}
\caption{Expectation and variance of the Simpson index at equilibrium against m for several values of s,
p=0.5
\end{figure}
\begin{figure}
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{Savecs.png}
\end{center}
\end{minipage}
\hfill
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{VarSavecs.png}
\end{center}
\end{minipage}
\label{fig:image5}
\caption{Here are plotted the variance and the expectation of the simpson index at equilibrium against s for several values of m.
p=0.5, size of linear approaching system is 100. }
\end{figure}
\begin{figure}
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{Savecp.png}
\end{center}
\end{minipage}
\hfill
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{VarSavecp.png}
\end{center}
\end{minipage}
\label{fig:image6}
\caption{Here are plotted the variance and the expectation of the Simpson index at equilibrium against p for several values of s.
m=2, size of linear approaching system is 100. }
\end{figure}
\vspace{30mm}
Let us now quantify the convergence to equilibrium. Recall at first that the process $ X_ {t} $ has for generator $ \mathscr {L} $ and for invariant measure $ \pi $. Denote $P_{t}f(x)=\mathbb{E}[f(X_{t})|f(0)=x]$ the associated semigroup. In fact, when $s=0$, the full spectrum is known, see for example Shimakura \cite{shimakura1977} which provides a spectral gap value $m$. It will imply an exponential convergence to equilibrium in $ L_{\pi}^{2} $.
\begin{prop}
Let us suppose that $s>0$. The following Poincar\'e inequality is valid, i.e. for every smooth function $f$
$$\mbox{Var}_{\pi}(f)\leqslant \min\left(\frac{e^s}{m},\frac{8e^{(1-M)s}}{m}\right)\int_{0}^{1}f'^{2}(x)\times x(1-x)d\pi(x)$$
where $M$ is a median of $\pi$. As a consequence $\pi_{t}$ converge to $\pi$ in $L^{2}_{\pi}$ exponentially:
$$\mbox{Var}_{\pi}(P_{t}f)\leqslant e^{-2\max\left(\frac{e^s}{m},\frac{8e^{(1-M)s}}{m}\right)}\mbox{Var}_{\pi}(f).$$
\end{prop}
\begin{proof}
Using usual Holley-Stroock's perturbation argument we easily deduce, that the Poincar\'e constant is at most $e^s/m$. We will now use Hardy's type condition (see for example \cite{ABC}) for Poincar\'e condition, that we recall now
\begin{lemma}
Let $\pi$ and $\nu$ be two measures and $M$ the median of $\pi$.\\
Let $$B^{+}_{M}=\sup\limits_{x>M}\int_{x}^{1}d\pi\int_{M}^{x}\frac{1}{\nu(t)}dt,\qquad B^{-}_{M}=\sup\limits_{x<M}\int_{0}^{x}d\pi\int_{x}^{M}\frac{1}{\nu(t)}dt.$$
If $B^{-}_{M}$ et $B^{+}_{M}$ are bounded, then the following Poincar\'e inequality holds
$$\mbox{Var}_\pi(f)\le c_P\int f'^2d\nu.$$
In addition, the optimal constant $c_P$ verifies $$\frac{1}{2}\max(B^{+}_{M},B^{-}_{M})\leqslant c_P\leqslant 4\max(B^{+}_{M},B^{-}_{M})$$.
\end{lemma}
We apply the lemma to $\pi$ and $\nu=x(1-x)\pi$ using both sides of the estimates. Denote $\pi_0$ and $\nu_0$ the case where $s=0$ and the Poincar\'e constant is $m$. Then
\begin{eqnarray*}
B^{+,s}_M&=&\sup_{x>M}\int_x^1d\pi\int_M^x\frac{1}{\nu(t)}dt\\
&=&\sup_{x>M}\int_x^1 e^{st}d\pi_0(t)\int_M^x\frac{e^{-st}}{\nu_0(t)}dt\\
&\le& e^{(1-M)s}B^{+,0}_M
\end{eqnarray*}
The same reasoning shows that $B^{-,s}_M\le B^{-,0}_M$.
\end{proof}
Of course, one can do easily the same for $s<0$ using a symmetric reasoning. If the order is good with respect to the immigration parameter, as the case $s=0$ is optimal, it is an open question to look at the dependence with respect to the selection parameter. We may also consider a convergence in entropy, via the logarithmic Sobolev inequality without selection established by Stannat \cite{stannat2000} or Miclo \cite{miclo2003} and the same line of proof using Holley-Stroock perturbation argument or the Hardy type condition for logarithmic Sobolev inequality (see again \cite{ABC}). Note that the convergence in entropy entails a convergence in total variation via Cszisar-Pinsker-Kullback inequality, but the constant involved are less explicit so we omit the details.\\\medskip
This quantitative long time behaviour enables us to give an error while approximating the asymptotic Simpson index (being a smooth function of the species). As usual, an $L^2$ decay will enable us to consider long time behaviour for initial measures whose density with respect to the invariant measure is bounded, which in could prevent starting from a Dirac measure. However due to regularization, and so waiting a time $t_0$, enables (loosing on the constants in the decay) to start from a Dirac measure. See for example \cite{BCG}.
\section{Generalization to a larger number of species or in random environment}
In this section we provide extensions of the two species case to 1) finite number of species, 2) two species case in a particular random environment, namely Wright-Fisher diffusion environment.
\subsection{Expectation approximation for three species}
\quad\\
In fact we will give the main ideas for $S=2$. Extension to a larger number of species is only technically involved and requires no further arguments. Denote $X_{t}$ and $Y_{t}$ the proportions of the two main species, $s^{x}_{t}$ and $s_{t}^{y}$ their selection parameters and $p_{t}^{x}$ and $p_{t}^{y}$ their proportions in the pool. The immigration parameter will still be denoted $m_t$. The method presented for $ S = 1 $ in the previous section can be generalized to a larger number of species. It will have of course some limitations: greater the number of species is, larger will be the size of the approaching linear system. In fact, the derivative of the expectation of order $ n $ involves only the expectation of lower and higher order in the case of two species. Now with 3 species we also need to know the expectation of the form $ \mathbb{E} [X_{t}^{n}Y_{t}^{k}] $ for $ k, n$ in $\mathbb{N} $. We will thus need a system of size $ N^{2} $.
We present here the extension of our approximation for 3 species.
\begin{equation}
\begin{pmatrix}
dX_{t}\\
dY_{t}
\end{pmatrix}
=
\begin{pmatrix}
m_{t}(p_{t}^{x}-X_{t})+X_{t}(s_{t}^{x}-X_{t}s_{t}^{x}-Y_{t}s_{t}^{y})\\
m_{t}(p_{t}^{y}-Y_{t})+Y_{t}(s_{t}^{y}-X_{t}s_{t}^{x}-Y_{t}s_{t}^{y})\\
\end{pmatrix}dt
+ \sigma (X_{t},Y_{t})dB_{t} \label{eqprincipalemulti}
\end{equation}
where $\sigma$ verifies $\sigma.\sigma^{*}(x,y)=a(x,y)$ with \\
$$a(x,y)=2\begin{pmatrix} x(1-x)&-xy\\ -xy&y(1-y)\\ \end{pmatrix} $$
We have to calculate $d(X_{t} ^{n}Y_{t}^{k})$ with It\^o's formula :
\begin{align}
d(X_{t} ^{n}Y_{t}^{k})= \nonumber &\big(m_{t}(p_{t}^{x}-X_{t})+X_{t}(s_{t}{x}-X_{t}s_{t}^{x}-Y_{t}s_{t}^{y})\big)nX_{t} ^{n-1}Y_{t}^{k}\\ \nonumber
&+\big(m_{t}(p^{y}_{t}-Y_{t})+Y_{t}(s^{y}_{t}-X_{t}s_{t}^{x}-Y_{t}s^{y}_{t})\big)kX_{t} ^{n}Y_{t}^{k-1}\\ \nonumber
&+n(n-1)(1-X_{t})X_{t} ^{n-1}Y_{t}^{k} +k(k-1)(1-Y_{t})X_{t} ^{n}Y_{t}^{k-1}\\ \nonumber
&-2nkX_{t} ^{n}Y_{t}^{k}+ d\mathscr{M}_{t}\\ \nonumber
=&X_{t} ^{n-1}Y_{t}^{k}n(m_{t}p_{t}^{x}+n-1)\\ \nonumber
&+X_{t} ^{n}Y_{t}^{k-1}k(m_{t}p^{y}_{t}+k-1)\\ \nonumber
&+X_{t} ^{n}Y_{t}^{k}\big(-m_{t}(n+k)-2kn-k(k-1)-n(n-1)+ns_{t}^{x}+ks_{t}^{y}\big)\\ \nonumber
&-X_{t} ^{n+1}Y_{t}^{k}s_{t}^{x}(n+k)\\ \nonumber
&-X_{t} ^{n}Y_{t}^{k+1}s_{t}^{y}(n+k)\\
&+d\mathscr{M}_{t}\label{coeff3speces}
\end{align}
here $\mathscr{M}_{t}$ is a martingale. Then $ d\mathbb{E} [X_ {t} ^ {n} Y_ {t} ^ {k}] $ is expressed in terms of 4 other quantities which complicates the one dimensional calculations. Moreover we must define what are the neglected expectations on which we will make an approximation, that is how to close the system. We can decide we make an approximation to the order $ N $ then that we neglect all the terms of higher order in the expression of $dE[X_{t} ^{n}Y_{t}^{k}]$ where $max(n,k)=N$.
Suppose we want to get the expectation up to order $N$ we need exactly $\sum_{k=1}^{N}2k+1$ quantities. And so the size of the approaching differential linear system will be of order $ (N + 1) ^ {2} $. \\
The following figure represents the complexity of the problem, for example $ d\mathbb{E} [X_{t}Y_{t}] $ is expressed as a function of the expectations of the quantities to which the blue arrows point.
$$\begin{matrix}
\quad X\textcolor{orange}{\rightarrow} \qquad \textcolor{blue}{\leftarrow}XY \textcolor{blue}{\rightarrow} \qquad Y\\
\quad \textcolor{orange}{\swarrow} \quad \qquad \quad \textcolor{blue}{\swarrow}\quad \quad \quad \textcolor{blue}{\searrow}\qquad \qquad\\
\qquad X^2\qquad X^{2}Y \qquad X^{2}Y^{2}\qquad Y^{2}X\qquad Y^{2}
\end{matrix}$$
Algorithmically it is not very difficult to build the matrix approaching the expectations of the diffusion. We must begin by giving a vector composed of the different expectation of size $ (N + 1)^2 $. For that, let us define an application that transforms the expectation of order $ (n, k) $ that is to say $ \mathbb{E} [X_{t} ^ {n} Y_{t} ^{k}] $ into an integer which corresponds to its coordinate in the expectation vector.
Next we build the matrix $A_{N²}$ as in the case of two species from the coefficients calculated in \ref{coeff3speces}.
We consider as error each expectation $(n,k)$ with $max(n,k)>N$ in the It\^o formula. So that the error is composed of $2N+1$ terms. And the approximation boils down to solving numerically a linear system. As an example consider the case $N=1$, we therefore involve four expectations which are $\mathbb{E}[X^{n}Y^{k}], k,n \in \{0,1\} $. The order imposed by $ \phi $ is therefore $(1,\mathbb{E}[X],\mathbb{E}[Y],\mathbb{E}[XY])$. Three terms compose the error: $\mathbb{E}[s_{t}^{x}X^{2}],\mathbb{E}[s_{t}^{y}Y^{2}],2\mathbb{E}[(s_{t}^{x}+s_{t}^{y})XY^{2}]$. We can also, as in the case of two species, prove the convergence of this algorithm by following exactly the same pattern as in the one species case.
The renormalizing coefficients of the expectation $(n,k)$ then become
$\frac{(s_{t}^{x})^{n}(s_{t}^{y})^{k}}{(n-1)!(k-1)!}$. It can similarly be shown that the error is at most of the order of $\frac{N^{2}\max(s_{t}^{x},s_{t}^{y})^{N}}{(N-1)!}$.\\
The extension to a larger number is straightforward and will entail an error of the order $\frac{N^{\frac{S}{2}+2}\|s_t\|_\infty^N}{N!}$, and it will still be reasonable but requires computations of a system of size $N^{S}$ which may be prohibitive for large $S$.
\subsubsection*{Numerical applications}
We can easily program such an algorithm and check that the results obtained are in agreement with quantities obtained by Monte Carlo method. See following figures:
\underline{Basic example}
\begin{figure}[htbp]
\label{basic}
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{approx3espS.png}
\end{center}
\end{minipage}
\hfill
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{approx3espX.png}
\end{center}
\end{minipage}
\label{fig:image7}
\caption{On the left hand side expectation of the Simpson index and Monte Carlo mean, and on the right hand side expectations of each species and their Monte Carlo means}
\end{figure}
We consider here a case with no immigration and constant selection parameter. The number of simulated trajectories for MC mean is 1000, $J=1000$, $m=0$, $X_{0}=0.5$, $Y_{0}=0.3$, $s_{y}=2$, $s_{x}=1$, the size of the approaching linear system is 144. Figure 8 plots approximate values of $\mathbb{E}[\mathscr{S}_{t}]$ and $\mathbb{E}[X_{t}]$ by the precedent method from the approximation in large population and by MC method from the discrete model.
\quad\\
\underline{Time dependent parameter case}
\begin{figure}[htbp]
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{approx3espS_svar.png}
\end{center}
\end{minipage}
\hfill
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{approx3espX_svar.png}
\end{center}
\end{minipage}
\label{fig:image11}
\caption{Left hand side Simpson index, right hand with the expectations of the three species.}
\end{figure}
In this second example, we consider once again a case without immigration and time dependent selection parameter. The number of simulated trajectories for MC mean is 1000, $J=1000$, $m=0$, $X_{0}=0.5$, $Y_{0}=0.3$, $s_{y}=2$, $s_{x}$ is piecewise constant taking two values $4$ and $-4$ at regular time intervals, the size of the approaching linear system is 144. Figure 9 plot the approximate values of $E[\mathscr{S}_{t}]$ and $E[X_{t}]$ by the precedent method from the approximation in large population and by MC method from the discrete model.
\subsection{When the selection is a diffusion.}
In the third Section we gave a method to get the moments of $X_{t}$, and thus $\mathbb{E}(\mathscr{S}_{t})$ for a time dependent immigration/selection parameter. If these parameters are random but autonomous, it gives a way to approximate the expectation of the Simpson index by doing a Monte Carlo mean with respect to the environment, passing from quenched to annealed. It would be however more interesting to evaluate directly the expectation of the Simpson index without further Monte Carlo simulations. It seems quite impossible to give a general algorithm for every environment but we will give in this section an efficient approximation method in a particular case. We consider for the selection parameter $s_t$ a rescaled Wright-Fisher diffusion, whose leading Brownian motion is independent of the one leading the SDE for the species evolution. This choice assures us that $s_t$ is a diffusion evolving in a bounded set and the choice of the different parameters leads to a wide choice of a Moran process with immigration.
\subsubsection{The expectation approximation}
\quad\\
Let us just give first the diffusion approximation result for this particular case, whose proof is even simpler as it relies on usual approximation diffusion for Markov chains.
\begin{theorem}
Assume that $(v^{J}_{n})_{n \in \mathbb{N}}$ is a Moran process without selection with size $J$ and the parameters $m^{s}$ et $p^{s}$.
Let $c$ and $b$ two constants such as $s^{J}_{n}= cv^{J}_{n}-b \quad \forall n \in \mathbb{N}$, and assume that $X_{n}$ follow a Moran process with size $J$ and parameters $(m_{n})_{n \in \mathbb{N}}$, $(p_{n})_{n \in \mathbb{N}}$ et $(s^{J}_{n})_{n \in \mathbb{N}}$ describe in the first part.
Let $U^{J}_{n}$ be the process having for coordinates $X^{J}_{n}$ et $s^{J}_{n}$.
Then when $J$ goes to infinity, the process $U^{J}_{tJ²}$ converge in law to the process $U_{t}$ which coordinates are solutions of the following stochastic differential equation:
\begin{equation}
\begin{pmatrix}
dX_{t}\\
dv_{t}
\end{pmatrix}
=
\begin{pmatrix}
m'_{t}(p_{t}-X_{t})+s'_{t}X_{t}(1-X_{t})\\
m'^{s}_{t} (p^{s}_{t}-v_{t})
\end{pmatrix}dt
+\begin{pmatrix}
\sqrt[]{2X_{t}(1-X_{t})}\\
\sqrt[]{2v_{t}(1-v_{t})}
\end{pmatrix} dB_{t} \label{eqprincipale}
\end{equation}
where $s_{t}'=\frac{s_{t}}{J}=cv_{t}-b$, $m_{t}'=\frac{m_{t}}{J}$, $m_{t}'^{s}=\frac{m_{t}^{s}}{J}$.
\end{theorem}
To approach the expectation of $X_{t}$ we use the method describe previously for three species, here $v$ play the same role as a third species.
However the dynamics is not exactly the same, the It\^o formula gives us:
\begin{align}
d(X_{t} ^{n}v_{t}^{k})= \nonumber &X_{t} ^{n-1}v_{t}^{k}n(m'_{t}p_{t}+n-1)\\ \nonumber
&+X_{t} ^{n}v_{t}^{k-1}k(m'^{s}_{t}p^{s}_{t}+k-1)\\ \nonumber
&+X_{t} ^{n}v_{t}^{k}\big(-(m'_{t}+b)n-km_{t}'^{s}-k(k-1)-n(n-1)\big)\\ \nonumber
&+X_{t} ^{n+1}v_{t}^{k}nb\\ \nonumber
&+X_{t} ^{n}v_{t}^{k+1}cn\\
&-X_{t} ^{n+1}v_{t}^{k+1}nc\\
&+d\mathscr{M}_{t}\label{coeff3speces2}
\end{align}
with $\mathscr{M}_{t}$ a martingale. Then as previously we close our system, for a given $N$, and to do so to neglect all the terms of higher order in the expression of $d\mathbb{E}[X_{t} ^{n}v_{t}^{k}]$ where $max(n,k)=N$. And now the algorithm is able to calculate all the expectations of the form $ \mathbb{E}[X_{t} ^{n}v_{t}^{k}]$ and so obtain the expectation of the Simpson index. The proof follows the same pattern. The renormalizing coefficients of the expectation $(n,k)$ in the proof allow to control the eigenvalues of the matrix thanks to the Gershgorin disks as before.
Many choices are possible and we take here the coefficient $\frac{1}{\sqrt[4]{n!k!}}$. This choice leads to a convergence speed at most of the order of $\frac{N^2}{\sqrt[4]{(N)!}}$.
\begin{figure}[htbp]
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{test_approximationXetSal_atoires__2_.png}
\end{center}
\end{minipage}
\label{fig:image12}
\caption{Are plotted the approximate values of $\mathbb{E}[v_{t}]$ and $\mathbb{E}[X_{t}]$ by the precedent method from the approximation in large population and by MC method from the discrete model. The number of simulated trajectories for MC mean is $5000 $, $J=1000$, $X_{0}=0.2$, $v_{0}=0.7$, $m^{s}=4$, $m=2$, $p=p^{s}=0.5$, $c=3$, $b=0.5$ the size of the approaching linear system is 144}
\end{figure}
\subsubsection{Comparison with the neutral model.}
\quad\\
In this part we compare the case where $s$ is "neutral on average", to the neutral case with $s=0$.
Thanks to the previous method one can for example calculate the average Simpson index in the case where the selection expectation is 0. For it let's take $p^{s}=1/2$, $v_{0}=1/2$ (this enforces $c=-2b$). The following figures show the results:
\begin{figure}[htbp]
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.3]{EsperS.png}
\end{center}
\end{minipage}
\hfill
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.3]{VarS.png}
\end{center}
\end{minipage}
\label{fig:image13}
\caption{Comparison with neutral case. Approximate values of $\mathbb{E}[\mathscr{S}_{t}]$ and $\mathbb{E}[X_{t}]$ by the precedent method from the approximation in large population $X_{0}=0.5$, $m^{s}=1$, $m=2$, $p=0.5$, $c=3$, size of the approaching linear system is 144}
\end{figure}
\begin{figure}[htbp]
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.3]{compre_neutre2.png}
\end{center}
\end{minipage}
\hfill
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.3]{variance01.png}
\end{center}
\end{minipage}
\label{fig:image14}
\caption{Comparison with the neutral case. Approximate values of $\mathbb{E}[\mathscr{S}_{t}]$ and $\mathbb{E}[X_{t}]$ by the precedent method from the approximation in large population $X_{0}=0.1$, $m^{s}=0.5$, $m=2$, $p=0.5$, $c=5$, size of the approaching linear system is 144.}
\end{figure}
We thus see that a selection even if neutral in mean, involves deeper mechanism which lead to a different behaviour than the neutral one. Of course the Simpson index involves not only the expectation of one species but also the moment of order two.
\subsection{Effect of selection on increase of biodiversity}
\quad\\
We have already seen in the case of two species that selection alone could contribute to the decrease of the average Simpson index in the absence of immigration. There was however a threshold for $ s $ under which such a phenomenon could not occur. We sort of generalize it here to any number of species.
\begin{prop}
Note as previously $ S + 1 $ the number of species in the community with $s_i$ the selection parameter for species $i$. Then if all $ \|s_ {i} \|_{\infty} $ are less than $ \frac {1}{2} $ the Simpson index is increasing in the absence of immigration.
In other words, the selection can not be the source of the diversity decreasing.
\end{prop}
\begin{proof}
Assume all the $s_{t}^{i}, \forall t$ are between $\alpha$ and $-\alpha$ .
First write:
\begin{align*}
1-\mathscr{S}_{t}=2&\sum\limits_{i=1}^{S}X^{i}_{t}(1-X^{i}_{t})-\sum\limits_{j\neq i}X^{i}_{t}X^{j}_{t}\\
=&\sum\limits_{i=1}^{S}X^{i}_{t}(1-X^{i}_{t})+\sum\limits_{i=1}^{S}X^{i}_{t}(1-\sum\limits_{i=1}^{S}X^{i}_{t})\\
=&\sum\limits_{i=1}^{S+1}X^{i}_{t}(1-X^{i}_{t})\\
\end{align*}
Then,
\begin{align*}
d\mathbb{E}[\mathscr{S}_{t}]=&2\mathbb{E}[1-\mathscr{S}_{t}]-2\mathbb{E}[\sum\limits_{i=1}^{S}s_{t}^{i}X_{t}^{i}(\mathscr{S}_{t}-X_{t}^{i})]dt\\
= & 2\mathbb{E}[1-\mathscr{S}_{t}-\sum\limits_{i=1}^{S}s_{t}^{i}X_{t}^{i}(\mathscr{S}_{t}-1)-\sum\limits_{i=1}^{S}s_{t}^{i}X_{t}^{i}(1-X_{t}^{i})]dt\\
\geqslant & 2\mathbb{E}[1-\mathscr{S}_{t}-\sum\limits_{i=1}^{S}s_{t}^{i}X_{t}^{i}(\mathscr{S}_{t}-1)-\alpha\sum\limits_{i=1}^{S}X_{t}^{i}(1-X_{t}^{i})]dt\\
\geqslant & 2\mathbb{E}[(1-\mathscr{S}_{t})(1+\sum\limits_{i=1}^{S}s_{t}^{i}X_{t}^{i})-\alpha\sum\limits_{i=1}^{S}X_{t}^{i}(1-X_{t}^{i})]dt\\
\geqslant & 2\mathbb{E}[(1-\mathscr{S}_{t})(1-\alpha+\sum\limits_{i=1}^{S}s_{t}^{i}X_{t}^{i})+\alpha\sum\limits_{i=1}^{S}X_{t}^{i}(1-\sum\limits_{i=1}^{S}X_{t}^{i})]dt\\
\geqslant & 2\mathbb{E}[(1-\mathscr{S}_{t})(1-\alpha+\sum\limits_{i=1}^{S}s_{t}^{i}X_{t}^{i})]dt\\
\geqslant & 2\mathbb{E}[(1-\mathscr{S}_{t})(1-\alpha(1+\sum\limits_{i=1}^{S}X_{t}^{i}))]dt
\end{align*}
and so if $\forall i, \alpha\leqslant \frac{1}{2}$ , $ d\mathbb{E}[\mathscr{S}_{t}]\geqslant 0$ and $\mathbb{E}[\mathscr{S}_{t}]$ is increasing.
\end{proof}
Remark that this bound is certainly not optimal, as the two species case indicates but true for each $S$.
\subsection{Long time behaviour.}
\quad\\
We will once again assume in this part $s,m,p$ are constants. If $m=0$ then a species will still invade the community definitively. On the other hand, if $ m \neq 0 $, the law of the vector of abundance, converges in a long time to a unique invariant measure. Consider the generator of the diffusion \eqref{eqprincipale} which is the generator of the Wright-Fisher diffusion with selection and mutation:
$$\mathscr{L}f(x)=\sum\limits_{i,j=1}^{S}x^{i}(\delta_{i,j}-x^{j})\frac{\partial²f(x)}{\partial x^{i}\partial x^{j}j}+\sum\limits_{i=1}^{S}\left(m(p^{i}-x^{i})+x^{i}\left(s^{i}-\sum\limits_{i=1}^{S}x^{i}s^{i}\right)\right)\frac{\partial f(x)}{\partial x^{i}}$$
A reversible and stationary measure for the diffusion \eqref{eqprincipale} is given by (see for example \cite{theinfinitneutralalleles,wright1977,Griffiths}:
$$\pi_{S}(dx)=C\times \exp\left(\sum\limits_{i,j=1}^{S+1}s^{i}x^{i}x^{j}\right)\times (x^{1})^{mp^{1}-1}\times...\times (x^{S+1})^{mp^{S+1}-1}dx_{1}...dx_{S}$$
Where $x^{S+1}=1-\sum\limits_{i=1}^{S}x^{i}$ and $p^{S+1}=1-\sum\limits_{i=1}^{S}p^{i}$, $s^{S+1}=0$. C is a constant just like $\int \pi_{S}(dx)=1$.
Of course, when $s$ and $m$ are time dependent, periodic for example, an invariant measure will not exist. The next figure presents the approximate values of $\mathbb{E}[\mathscr{S}_{t}]$ and $\mathbb{E}[X^{i}_{t}]$ for $i$ in $\{1,2,3\}$ by the precedent method from the approximation in large population and by Monte Carlo method from thr discret model. The number of simulated trajectories for Monte Carlo mean is $5000$, $J=500$, $m$ is a time dependant piecewise process, it takes alternatively the values of $3$ and $0$ at regular time intervals. $X_{0}=0.5$, $p_{x}=0.33$, $Y_{0}=0.3$, $p_{y}=0.33$, $s_{y}$ et $s_{x}$ are Markovian jump processes, the size of the approaching linear system is 144.
\begin{figure}[htbp]
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{maleatoire3espsconst.png}
\label{malea}
\end{center}
\caption{$\mathbb{E}[\mathscr{S}_{t}]$}
\end{minipage}
\hfill
\begin{minipage}[c]{.45\linewidth}
\begin{center}
\includegraphics[scale=0.35]{maleatoire3espsconstXY.png}
\label{}
\end{center}
\caption{$\mathbb{E}[X_{t}], \mathbb{E}[Y_{t}]$}
\label{Tux}
\end{minipage}
\end{figure}
Concerning the long time behaviour, we may once again refer to \cite{shimakura1977} for the spectral gap which is $e^{(S+1)\sum_1^{S+1}s^i}/m$ by Holley-Stroock's perturbation argument. Unfortunately, it is not possible to refine this argument as there is no Hardy's type inequalities in this case. Once again it is also possible to derive a logarithmic Sobolev inequality, and thus convergence in entropy (and total variation) but constants are less explicit.
\section{Proofs}
In this section we gather the proofs, technical or more or less well known.
\subsection{Proof of the diffusion approximation, Theorem \ref{th:diffapprox}}
\label{diffapprox}
\quad\\
In the following proof we'll get back to a martingale problems. All the results used in this section can be funded in \cite{diffmultidim} p267-272.
For the sake of clarity, assume that $ m = 0 $, and that $ S = 1 $. \\The multidimensional case is treated exactly the same way. \\
We can put $h=\frac{1}{J²}$ and $U$ means here $(x,s)$ where $x \in E_{x}$ and $s \in E_{s}$.\\
Let $a(U)=x(1-x)$, $b(U)=sx(1-x)$ and $L_{x}f(U)=b(U)\times \frac{\partial f}{\partial x}(U)+a(U)\times \frac{\partial² f}{\partial x²}(U)$ the generator of the SDFE \eqref{eqprincipale} and $L_{s}f(U)=\sum\limits_{y\in E_{s}}
Q_{s,y}f(x,y)$ the generator of a Markovian jump process applied to a function depending of the population variable.
Let's start with the following lemma
\begin{lemma}(\cite{diffmultidim} p268)
\quad\\
Let $f$ be a $C^{\infty}$ function, note $A_{J}f(U)=\int_{E}f(y)-f(u)d\pi_{J}(U,dy)$ then $J²A_{J}f$ converge uniformly to $L_{x}f+L_{s}f$
\end{lemma}
\begin{proof}
\begin{align*}
A_{J}f(U)=&\int_{E}f(y)-f(u)d\pi_{J}(U,dy)=\int_{E}f(z,w)-f(x,s)d\pi_{J}(U=(x,s),dy)\\
=&\int_{E}f(z,w)-f(x,s)d\pi_{J}(U,dy)\\=&\int_{E}f(z,w)-f(x,w)+f(x,w)-f(x,s)d\pi_{J}(U,dy)\\
=&\int_{E}f(z,w)-f(x,w)d\pi_{J}(U,dy)+\int_{E}f(x,w)-f(x,s)d\pi_{J}(U,dy)\\
=&\int_{E} f(z,w)-f(x,w)d\pi_{J}(U,dy) + \sum\limits_{w\in E_{s}}f(x,w)-f(x,s)P_{J}^{i}(x,y,h).
\end{align*}
Via Taylor's formula, we obtain
\begin{align*}
&\int_{E} f(z,w)-f(x,w)d\pi_{J}(U,dy)=\mathbb{E}[f(X_{t+h},s_{t+h})-f(x,s_{t+h})|U_{t}=(x,s)]\\
&=\frac{\partial f}{\partial x}(x,s)\mathbb{E}[X_{t+h}-x|U_{t}]+\frac{\partial² f}{2\partial x²}(x,s)\mathbb{E}[(X_{t+h}-x)²|U_{t}]\\
&+ \frac{\partial² f}{\partial x\partial s}(x,s)\mathbb{E}[X_{t+h}-x|U_{t}]E[s_{t+h}-s|U_{t}]+O(\mathbb{E}[\| U_{t+h}-U_{t}\|³|U_{t}]).
\end{align*}
we give the limits of the previous quantities,
\begin{itemize}
\item $\lim\limits_{h\rightarrow 0} \sup\limits_{E}\|\frac{1}{h}\mathbb{E}[X^{J}_{t+h}-x|U^{J}_{t}]\|=b(x,s)$ by the second propriety (\ref{exp-var}),
\item $\lim\limits_{h\rightarrow 0} \sup\limits_{E}\|\frac{1}{h}\mathbb{E}[(X^{J}_{t+h}-x)²|U^{J}_{t}]\|=a(x,s)$ by the second propriety (\ref{exp-var}),
\item $\lim\limits_{h\rightarrow 0} \sup\limits_{E}\|\frac{1}{h}\mathbb{E}[X^{J}_{t+h}-x|U^{J}_{t}]\mathbb{E}[s^{J}_{t+h}-s|U^{J}_{t}]\|=0$ because
\begin{eqnarray*}
|\mathbb{E}[s^{J}_{t+h}-s|U_{t}]|&=&|\sum\limits_{w\in E_{s}}(w-s_{t})P_{J}^{i}(s_{t},w,h)|\\& \leqslant& \max\limits_{x,y\in E_{s}}|x-y|\sum\limits_{w\in E_{s}/s_{t}}P_{J}^{i}(s_{t},w,h),
\end{eqnarray*}
\item $\lim\limits_{h\rightarrow 0} \sup\limits_{E}\mathbb{E}[\| U^{J}_{t+h}-U\|³|U^{J}_{t}]=0$ because $\int_{E}\sup\limits_{i\leqslant S}|y_{i}-u_{i}|^{3}\pi_{J}(U,dy) \rightarrow 0 $ and by the previous point.
\end{itemize}
Then, going to the limit in the previous expression,
\begin{align*}
\lim\limits_{h\rightarrow 0} &\sup\limits_{E}|A_{J}f(U)\frac{1}{h}-L_{x}f(U)-L_{s}f(U)|\\
\le&\lim\limits_{h\rightarrow 0} \sup\limits_{E}|\frac{1}{h}\int_{E} f(z,w)-f(x,w)d\pi_{J}(U,dy)- L_{x}f(U)|\\
&\qquad+\lim\limits_{h\rightarrow 0} \sup\limits_{E}|\frac{1}{h}\sum\limits_{w\in E_{s}}\left(f(x,w)-f(x,s)\right)P_{J}^{i}(x,y,h)-L_{s}f(U)|\\ =&0
\end{align*}
And this expression conclude the proof.
\end{proof}
Now let $f$ be $\mathscr{C}^{\infty}$, then
\begin{align*}
\mathbb{E}_{U}[f(U^{J}_{t})]&=f(U)+\mathbb{E}_{U}\left[ \sum\limits_{k=1}^{|tJ²|-1}\mathbb{E}[f(U^{J}_{(k+1)h})-f(U^{J}_{kh})|U^{J}_{kh}]\right]\\
&=f(U)+\mathbb{E}_{U}\left[ \sum\limits_{k=1}^{|tJ²|-1}A_{J}f(U_{kh})\right]\\
&=f(U)+\mathbb{E}_{U}\left[ \sum\limits_{k=1}^{|tJ²|-1}\frac{h}{h}A_{J}f(U_{kh})\right]\\
\end{align*}
and so $$\mathbb{E}_{U}[f(U^{J}_{t+h})-f(U)-\sum\limits_{k=1}^{|tJ²|-1}\frac{h}{h}A_{J}f(U_{kh})]=0$$
i.e $f(U^{J}_{t})-f(U)-\sum\limits_{k=1}^{|tJ²|-1}\frac{h}{h}A_{J}f(U_{kh})$ is a martingale for $ \pi^{J}$.\\
Moreover, note the sum is a Riemann sum and the previous lemma ensures when $ J $ tends to infinity the convergence of $$f(U^{J}_{t})-f(U)-\sum\limits_{k=1}^{|tJ²|-1}\frac{h}{h}A_{J}f(U_{kh})$$ towards $$f(U_{t})-f(U)-\int_{0}^{t}L_{x}f(U_{s})+L_{s}f(U_{s})ds.$$
We need now to find a probability measure on the Borel sets of the canonical space $\mathscr{C}([0,1],R)$ verifying the martingale problem for $ L_ {x} + L_ {s} $.
Let us show now that $ \pi ^ {J} $ admits an adherent value in the space of probability measure on the Borel of $\mathscr{C}([0,1],R)$ with the norm $$\|\pi^{J}\|=\sup\limits_{f\in C
}\frac{|\int fd\pi^{J}|}{|f|_{\infty}}$$ (which is a norm since the $ \pi^{J} $ are supported in $ [0,1] $).\\
Let note $ \pi^{J} f=\int_{E} fd\pi^{J}$.
Let $ (f_{n}) _ {\mathbb{N}} $ be a dense sequence in the space of continuous functions then $ (\pi^{J} f_ {n}) _ {J} $ is a sequence of $ \mathbb{R} $ having an adherence value in $ \mathbb {R} $ because it is uniformly bounded by $ |f_{n}|_{\infty} $. Then by diagonal extraction, eventually for a subsequence, $ \pi ^ {J}f_{n} $ converges to a certain $ \phi_ {fn} $ in $ \mathbb{R} $ for all $ n $. And by the uniformly continuous extension theorem, we define $ \phi_ {f} $ for all $ f $ of $ \mathscr{C} $. And since $ \phi_ {f} $ is a linear form, the Riesz-Markov theorem ensures the existence of a unique measure $ \mu $ such that $ \int fd \mu = \phi_ {f} $.
Since this is true for all $ f \in \mathscr{C} $, by considering the constant function equal to $ 1 $, we find $ \mu (\Omega) = 1 $ and $ \mu $ is a probability. The convergence of $ \| \pi^{J} - \mu \| $ to $ 0 $ is then immediate in view of the chosen norm.\\
Thus our sequence $ \pi^{J} $ admits an accumulation point. So, there is at least one $ \pi $ and one $ X $ process that satisfy the martingale problem associated with $ L_ {x} + L_ {s} $. \\
And so $\pi$ verifies $\int f(U_{t})-f(U)-\int_{0}^{t}L_{x}f(U_{s})+L_{s}f(U_{s})ds d\pi(U,U_{t})=0$.
So it exist at least a solution to the martingale problem for $ L_ {x} + L_ {s} $ .
\quad\\
If the uniqueness of this martingale problem is verified then the process converges in law to the diffusion process (our $X_t$) defined by $ a $, $ b $ and thus and $s$ the jump process of generator $ Q $, since they are both solutions of the same problem of martingale. The proof of uniqueness is quite standard, following Ethier \cite{ETHIER} when $ s $, $ m $, $ p $ constant. A straightforward modification allows to obtain the result for $ s $, $ m $, $ p $ random.
\subsection{Proof of Proposition \ref{prop:absorb}}
\label{absorb}
\quad\\
Let $g$ be the solution of the differential equation \eqref{g}. Let us first verify that $ g$ is well defined on $[0,1]$. It must be ensured that the solutions do not diverge in 0 and 1, in which case the second member of the equation is not defined. For that we can write the solution of this equation. So $$g(x)=\int_{0}^{x}e^{-su}\left(K+\int_{\frac{1}{2}}^{u}\frac{e^{st}}{t(1-t)}dt\right)du+C$$ where $C,K$are constant. As $e^{-su}$ is bounded on$[0,1]$, there are two positive constants $B$ and $D$ such that $$\lim\limits_{x\rightarrow 0}|g(x)|\leqslant \lim\limits_{x\rightarrow 0}\int_{a}^{x}\int_{a}^{u} \frac{D}{t}dtdu + B\leqslant \infty$$
as $\ln$ is integrable on a neighbourhood of 0. Thus $g$ is well defined on [0,1] and bounded (because continue).\\
So we have $g(X_{t\wedge T_{1, 0}})=g(X_{0})+\int_{0}^{t\wedge T_{1, 0}}g'(X_{u})\sqrt[]{2X_{u}(1-X_{u})}dB_{u}-t\wedge T_{1, 0}$.\\
But the process $\int_{0}^{t\wedge T_{1, 0}}g'(X_{u})\sqrt[]{2X_{u}(1-X_{u})}dB_{u}$ is a stopped martingale because $t\wedge T_{1, 0}$
is a stopping time and $g'(X_{u})\sqrt[]{2X_{u}(1-X_{u})}$ is adapted to the considered filtration.
We deduce that $\mathbb{E}_{X_{0}}[g(X_{t\wedge T_{1, 0}})]=g(X_{0})-\mathbb{E}_{X_{0}}[t\wedge T_{1, 0}]$ and the first property,
i.e $\mathbb{E}_{X_{0}}[t\wedge T_{1, 0}]\leqslant 2 \sup\limits_{[0,1]}(g)\leqslant \infty$, and thus the second point is shown.Now if $t\rightarrow \infty$, $\mathbb{E}_{X_{0}}[g(X_{t\wedge T_{1, 0}})]\rightarrow0$ because $g(0)=g(1)=0$ and we find again $g(X_{0})=E_{X_{0}}[ T_{1, 0}]$.\\
To prove the third point, consider $f(x)=e^{-sx}-1$. Then $f$ is solution of
$f''(x)+sf'(x)=0$ and
$f(0)=0$. By It\^o's formula, we obtain $d(f(X_{t})=f(X_{0})+dMt $. As $T_{1}$ and $T_{0}$ are stopping times $T_{1}\wedge T_{0}=T_{1, 0}$ is also a stopping time. So$ f(X_{ T_{0,1} })=f(X_{0})+dM'_{t}$ où $dM'_{t}$ is still a martingale. By taking expectation we have $\mathbb{E}_{X_{0}}[f(X_{ T_{0,1}})]=f(X_{0})=f(1)\mathbb{P}(T_{1}<T_{0})$ and we deduce $ \mathbb{P}(T_{1}<T_{0})=\frac{e^{-sX_{0}}-1}{e^{-s}-1}$.
\subsection{Proof of Proposition \ref{prop:access}}
\label{access}
\quad\\
Let us consider the speed measure and the scale function as in Feller \cite{Feller2}.
\begin{align*}
m(y)&=\frac{1}{2y(1-y)}exp(\int_{a}^{y}s+\frac{m(p-x)}{x(1-x)}dx) \quad \textnormal{(speed measure)}\\
&=c\times y^{mp-1}\times (1-y)^{m(1-p)-1}\times \exp(sy),\quad c \in \mathbb{R}
\end{align*}
\begin{align*}
\mu(t)&=\int_{a}^{t} exp(-\int_{a}^{y}s+\frac{m(p-x)}{x(1-x)}dx)dy \quad \textnormal{(scale function)}\\
&=c'\int_{a}^{t} y^{-mp}\times (1-y)^{-m(1-p)}\times \exp(-sy)dy,\quad c' \in \mathbb{R}.
\end{align*}
Then $1$ is reachable if and only if $\mu(1)<\infty$ and $\int_{\frac{1}{2}}^{1}\mu(1)-\mu(y)m(y)dy<\infty$. It is easily seen that $\mu(1)<\infty$ if and only if $m(1-p)<1$.\\
Next, $\int_{\frac{1}{2}}^{1}(\mu(1)-\mu(y)m(y))dy<\infty$ if and only if $\mu(1)-\int\mu(y)m(y)$ is integrable on a neighborhood of 1. But
\begin{align*}
\mu(1)-\mu(y)m(y)&\leqslant \int_{y}^{1}x^{-mp}(1-x)^{-m(1-p)}y^{mp-1}(1-y)^{m(1-p)-1}e^{sy-sx}dx\\
&\leqslant \int_{y}^{1}(1-x)^{-m(1-p)}dx\times(1-y)^{m(1-p)-1}\frac{1}{y}\\
&\leqslant a\times \frac{1}{y},
\end{align*}
for some constant $a$. This quantity is well defined and integrable on a neighborhood of 1. So 1 is reachable if and only if $m(1-p)<1$.\\
Now if $m(1-p) \leqslant 1$, $1$ is not reachable, it is regular (reflective barriers) if and only if $m(y)$ is integrable. It is indeed the case here, $-21<m(1-p)-1\leqslant 0$. Of course the same holds for $0$.
\bibliographystyle{plain}
|
\section{Introduction}
Since its discovery in 1964 by Penzias and Wilson \cite{Penzias}, the cosmic microwave background (CMB) has served as the main source of information about the current and past Universe. Originating in the process of recombination at about 380$\,$000 years after the Big Bang at redshift of about 1100, the CMB temperature distribution on the celestial sphere displays the energy density distribution on the Last Scattering Surface when the Universe became transparent to electromagnetic radiation. The CMB intensity follows a nearly perfect Planckian distribution and its average temperature has been measured to be $T_0=2.72548 \pm 0.00057 \, \mathrm{K}$\cite{Fixsen} with minor fluctuations of order $10^ {-5}$--$10^{-4} \, \mathrm{K}$ and one larger dipole modulation of order $10^{-3}\, \mathrm{K}$, which is assumed to be of pure kinematic origin. The anisotropy of the CMB temperature has been measured first by the \textit{Cosmic Background Explorer} satellite from 1989 to 1993, followed up by the Wilkinson Microwave Anisotropy Probe (\textit{\textit{WMAP}}) mission from 2001 to 2010. The most recent CMB investigation satellite, \textit{Planck}, was launched in 2009 and shutdown in 2013. Its 2015 results from the second data release provide the most precise values of cosmological parameters \cite{Planck2015I,Planck2015XIII} measured up to this point. Recently, they have been refined in the 2018 data release \cite{Planck2018I,Planck2018VI}.
It is commonly assumed that the tiny temperature fluctuations follow a Gaussian and statistically isotropic distribution. This assumption has been confirmed by the \textit{Planck} mission to a large extent\cite{Planck2015XVI,Planck2015XVII}, but nevertheless the search for possible non-Gaussianities\cite{KomatsuNonGaussianity} and statistical anisotropies \cite{Schwarz2004,schupp1,Abramo,Schwarz2007,Aurich,Schwarz2009,Gruppuso,Schwarz2010,Zunckel,Sung} has been rich and certain anomalies have been found, as, for example, unusual (anti-)correlation of the lowest multipoles with the Cosmic Dipole as well as with each other, a sign of parity asymmetry and a lack of large-angle correlation (see e.g. the review\cite{Schwarz2015}).
Common tools in these analyses are multipole vectors (MPVs) which were introduced for cosmological data analysis in \cite{MultiCopi} and whose properties have been elaborated in \cite{schupp1,KatzWeeks,Dennis2004,Dennis2005,Dennis2007}. For the most recent results on possible CMB anomalies using multipole vectors and an overview over the mathematical approaches see \cite{Pinkwart,PinkwartSchwarz,Oliveira}.
MPVs are closely related to Bloch coherent states (see \cite{schupp1}) which were also used in the past to prove special cases of Lieb's conjecture\cite{Schupp} for the Wehrl entropy.
In this work, we develop and compare several rotationally invariant measures of randomness on functions on the two-sphere, namely, the angular, projection, and quadratic pseudoentropies. We show that for $l=2$ the Wehrl and angular entropy can be expressed as a function of the squared chordal distance of MPVs. We find that all these measures except the quadratic one show the same features, making the quadratic pseudoentropy the least preferred measure. Because of the shared features, we then restrict ourselves to the numerically fastest method, and use it to analyze \textit{Planck} 2015 and 2018 full sky as well as \textit{WMAP} 7-year Internal Linear Combination (ILC) maps. The angular pseudoentropy allows for comparing the data to many ensembles of Gaussian and isotropic random maps up to $l=1000$ in short computing time. With a better theoretical understanding of confidence levels, also the Wehrl entropy could be used easily since the computing time for a single map is still reasonable. In general, it is especially nice to have a single number for each multipole even in the case that the data would not be Gaussian and isotropic. In this case, the CMB would be described by more than one degree of freedom (d.o.f.) per multipole. Non-Gaussian distributions need higher correlation functions and anisotropic distributions yield an $m$-dependent two-point function. In the tradition of thermodynamics, with these pseudoentropies one can approximately reduce a possibly large set of data again to one number for each multipole. Since all considered types of entropies show a similar behavior the information does not depend on the definition of the entropy. Eventually there exists also an extension of the angular entropy to ranges and collections of multipoles, which we call range angular entropy.
This paper is organized as follows: In Sec.~\ref{CMB statistics}, we briefly recapitulate the basic ingredients of CMB spherical harmonic statistics. Afterwards, in Sec.~\ref{math}, we introduce our methods mathematically, clarify their properties, and show the connection to multipole vectors. We also provide a simple proof of the convergence of the projection entropy to the Wehrl entropy up to a term which is independent of the input density matrix. Section \ref{Results} is dedicated to the application of our methods to real data. We compare the different pseudoentropy methods, then we apply the angular pseudoentropy to 2015 \textit{Planck} and 7-year \textit{WMAP} full sky foreground-cleaned maps before comparing the 2015 results to those obtained with 2018 data and also applying the range entropy and comparing it to the statistics used before. Eventually, in Sec.~\ref{sum} we summarize and discuss our findings.
\section{CMB statistics}
\label{CMB statistics}
As a function on $S^2$ the CMB temperature fluctuations $\Delta T := \delta T/T_0$ can be decomposed uniquely according to irreducible representations of $SO(3)$, i.e.,, into spherical harmonics
\begin{equation}
\Delta T(\theta,\phi) = \sum_{l=1}^{\infty}\sum_{m=-l}^l a_{lm}Y_{lm}(\theta,\phi) \, \in \mathbb{R},
\end{equation}
where the $l=0$-summand is omitted because the fluctuations average to zero, and $\theta$, $\phi$ denote the usual spherical coordinates. The fact that $\Delta T$ is real together with the property $Y_{l,-m} = (-1)^m Y_{lm}^* $ impose the constraints
\begin{equation}
a_{l,-m} = (-1)^m a_{lm}^*
\end{equation}
on the spherical harmonic coefficients, leaving for each multipole number $l$ exactly $2l+1$ real d.o.f. The multipole number $l$ corresponds to angular scales of $\approx 180/l\, \deg$. The orthonormality of $\{Y_{lm}\}$ allows to compute the coefficients from the temperature map via
\begin{equation}
a_{lm} = \int_{S^2}\! \mathrm{d}\Omega \, \Delta T(\Omega)Y_{lm}^*(\Omega).
\end{equation}
Simple inflationary models together with linear perturbation theory predict nearly Gaussian temperature fluctuations and a further common assumption is statistical isotropy. The spherical harmonic coefficients inherit both properties from the temperature map, meaning that
\begin{equation}
p(\vec{a}_l) = \frac{1}{\mathcal{N}}e^{-\frac{1}{2}\vec{a}_l^{\dagger} \mathbf{D}_l \vec{a}_l} \quad \text{(Gaussianity),}
\end{equation}
where $p$ denotes the joint probability distribution, $\vec{a}_l = (a_{l0},\ldots,a_{ll})^T$, $C^l_{mn} = \left( \mathbf{D}_l^{-1} \right) _{mn} = \langle a_{lm}a^*_{ln}\rangle$ and $\mathcal{N}$ denotes a normalization constant, and
\begin{align}
\forall \mathbf{R} \in SO(3), \vec{e}_1,\ldots,\vec{e}_n \in S^2, n \in \mathbb{N}: \nonumber \\
G_n(\{\mathbf{R}\vec{e}_i\}) = G_n(\{\vec{e}_i\}) \quad \text{(isotropy),}
\end{align}
where $G_n(\{\vec{e}_i\}) = \langle \prod_{i=1}^n \Delta T (\vec{e}_i) \rangle$ denotes the $n$-point function of temperature fluctuations. The isotropy condition is equivalent to rotationally invariance of the joint $a_{lm}$-probability distribution. The averaging $\langle . \rangle$ is meant to be performed over all possible universes, which of course is not possible, wherefore we are left with a natural and inevitable variance in all quantities, called cosmic variance. If we impose both Gaussianity and isotropy then the two-point correlation of spherical harmonic coefficients is diagonal
\begin{equation}
\label{cl_eq}
C^l_{mn} = C_l \delta_{mn}.
\end{equation}
In practice, the power spectrum $C_l$ is calculated using the unbiased estimator
\begin{equation}
\label{cl_estimator}
\hat{C}_l = \frac{1}{2l+1}\sum_m |a_{lm}^2| \, , \, \langle \hat{C}_l \rangle = C_l,
\end{equation}
with cosmological variance
\begin{equation}
\mathrm{var}(\hat{C}_l) = \frac{2}{2l+1}C_l^2.
\end{equation}
\section{pseudoentropies and their properties}
\label{math}
In this section, we shall discuss the mathematical properties and interrelation of various macroscopic entropy measures that can be used as powerful tools to analyze Gaussianity and isotropy of the CMB and can also be useful in other contexts. This section contains a review of the mathematical background as well as new definitions, results and insights. The motivation to look for macroscopic entropy measures is the same as in statistical physics: A microscopic description of a physical system, e.g., the positions and momenta of a fluid, is useful for simulation purposes, but not when comparing to a real fluid. Instead, one would resort to the study of macroscopic quantities and parameters like internal energy, temperature, entropy, pressure etc.~that are well-defined because of symmetries. In the analysis of the CMB, the $a_{lm}$ coefficients are an analog of the microscopic quantities. For low $l$ these and derived quantities like multipole vectors can be studied individually, but for high~$l$ this quickly becomes impractical: The \textit{Planck} mission data easily comprises several million reliable data points. For the CMB the obvious underlying spacetime symmetry is rotation invariance. The representations of the rotation group decompose into irreducible components labeled by the angular momentum quantum number $l$ (multipole expansion) and we can focus on fixed-$l$ subspaces. A loose analog of internal energy is the angular power spectrum, i.e., the $C_l$ coefficients [see Eqs.~\eqref{cl_eq} and \eqref{cl_estimator}]. They have proven immensely useful in the analysis of the CMB and its cosmological implications, but when it comes to questions of isotropy and preferred directions, individual $m$ matter and an analog of entropy would be useful. A natural idea is to consider the abstract quantum state $|\psi\rangle := \sum a_{lm} |l,m\rangle$ that can be formally computed from the $a_{lm}$ and associate an entropy $S$ to it. We will usually focus on one $l$ at a time and normalize the states by rescaling the~$a_{lm}$ appropriately. Since the states are by construction pure, the von Neumann entropy will be trivially zero, but there are also non-trivial pseudoentropies that can distinguish pure states and turn out to be sensitive to non-Gaussianity and anisotropy. The general strategy is as follows:
\begin{equation}
\begin{split}
T(\theta,\phi) &\rightarrow a_{lm} \rightarrow |\psi\rangle := \sum a_{lm} |l,m\rangle \\
& \rightarrow \rho = |\psi\rangle\!\langle \psi| \rightarrow \rho_\text{mixed} \rightarrow S \ ,
\end{split}
\end{equation}
where $\rho_\text{mixed}$ is obtained from the pure state $\rho$ by applying a rotationally symmetric quantum channel $\Phi$, i.e., a completely positive map between Hilbert spaces with possibly different dimensions, or by computing its lower symbol, i.e., its expectation value in spin coherent states. The latter choice leads to the Wehrl entropy \cite{Wehrl2,Lieb78,Schupp}
\begin{equation}
S_W = - (2l+1) \int \frac{d\Omega}{4\pi} |\langle \Omega|\psi\rangle|^2 \ln |\langle \Omega|\psi\rangle|^2 \ , \label{Wehrlentropie}
\end{equation}
where $|\Omega\rangle$ is a spin-$l$ coherent state. The Wehrl entropy was first proposed in \cite{Helling} as a useful tool for CMB analysis. See Fig.~\ref{wehrl_30} for a showing the Wehrl entropy for CMB data. Closely related is the ``quadratic entropy'' that is obtained by replacing $-x \ln x$ in the formula for the Wehrl entropy by the concave function $x(1-x)$,
\begin{equation}
S_{\text{quad}} = 1 - \frac{2l+1}{4l+1} \left|P_{2l} |\psi \otimes \psi \rangle\right|^2 \ , \label{quadraticentropie}
\end{equation}
where $P_{2l}$ is the projector onto the spin-$2l$ part, i.e., the highest spin component of the tensor product.
Other examples using the choice $-x \ln x$ are what we call angular entropy
\begin{align}
S_{\text{ang}} =& \, \mathrm{Tr} \left[ \phi\Big(\sum_{i=1}^3 \frac{L_i |\psi\rangle\!\langle \psi| L_i}{l(l+1)} \Big) \right] \\
\label{phi_def}
\text{with} \quad \phi(x) :=& -x \ln x \ ,
\end{align}
where the $L_i$ are angular momentum generators in the spin-$l$ representation, and $j$-projection entropy
\begin{equation}
S_{\text{proj}}^{(j)} = \mathrm{Tr} \left[\phi\Big(\frac {2l+1}{2(l+j)+1} P_{l+j}\big(|\psi\rangle\!\langle \psi| \otimes \mathbf 1\big)P_{l+j}\Big) \right]
\end{equation}
where $\mathbf 1$ is the unit operator on a spin-$j$ ancilla $[j]$ and $\phi$ is as in Eq.~\eqref{phi_def}.
An overview of these entropies applied to CMB data can be found in Fig.~\ref{Ent_comp}. In the following section we shall explain the mathematics in detail, derive relations between the various entropies and point out interesting side results including a fast way to compute multipole vectors. Readers that are mostly interested in results and numerics can skip to the algorithm \eqref{ang_alt}-\eqref{ang_range} at the end of Sec.~\ref{Math_a}.
\subsection{Coherent states, multipole vectors and entropy}
\label{Math_a}
Coherent states were originally introduced by Schr\"{o}dinger \cite{Schrodinger} and are well known in the context of the quantum harmonic oscillator, where they can be defined either as eigenstates of the lowering operator or, equivalently, as elements of the orbit of the ground state under the Heisenberg group. Perelomov \cite{Perelomov} has generalized the latter notion to orbits of a fiducial vector in some representation of a Lie group under the action of that group. The choice of the fiducial vector is essential for the properties of the resulting coherent states.
Spin coherent states -- also called Bloch coherent states -- in a spin-$l$ irreducible representation $[l] \equiv {\mathbb C}^{2l+1}$ of $SU(2)$ with $2l+1 \in \mathbb N$
are defined as orbits of the highest weight vector $|l,l\rangle$. The stability group of that vector is $U(1)$ and spin coherent states can thus be labeled by points $\Omega =(\theta, \phi)$ on the sphere $S_2 \cong SU(2)/U(1)$,
\begin{equation}
|\Omega_l\rangle = \mathcal R(\Omega) |l,l\rangle \ ,
\end{equation}
where $\mathcal R(\Omega)$ denotes a rotation that takes the north pole to the point $\Omega$ and $l$ labels the representation of $SU(2)$. For the CMB data $l$ will be an integer, but everything we discuss here is also valid for half-integer $l$.
For $l= \frac 12$ this gives for example
\begin{equation}
|\Omega_{\frac 12}\rangle = e^{-i\phi/2} \cos \tfrac\theta 2 |\tfrac 12, \tfrac 12\rangle
+ e^{+i\phi/2} \sin \tfrac\theta 2 |\tfrac 12, -\tfrac 12\rangle \ . \label{spin12}
\end{equation}
Coherent states inherit nice properties from the underlying fiducial vector. A particular important one is that the tensor product of coherent states is again a coherent state and lies in the highest spin component:
\begin{equation}
|\Omega_l\rangle \otimes |\Omega_{j}\rangle
= |\Omega_{l+j}\rangle \label{productproperty}
\end{equation}
Using this property repeatedly yields an explicit formula for any spin from \eqref{spin12}:
\begin{equation}
\label{factorcoherent}
\begin{split}
|\Omega_l\rangle &= |\Omega_{\frac 12}\rangle \otimes \ldots \otimes |\Omega_{\frac 12} \rangle \\
&= \sum_{m=-l}^l \binom{2l}{l+m}^{\!\frac 12} e^{-im \phi /2}\,\cos^{l+m}(\tfrac\theta 2) \, \sin^{l-m}(\tfrac\theta 2) \, |l,m\rangle
\end{split}
\end{equation}
Interestingly, such a product representation in terms of spin-$\tfrac 12$ states exists for any state $|\psi\rangle \in [l]$, but except for coherent states, a projection onto the highest spin component and a renormalization are required \cite{Schupp}:
\begin{equation}
|\psi_l\rangle = \sum_{m=-l}^l a_{lm} |l,m\rangle = c P_l\Big( |\Omega^{(1)}_{\frac 12}\rangle \otimes \ldots \otimes |\Omega^{(2l)}_{\frac 12} \rangle\Big) \ , \label{factorpsi}
\end{equation}
where $P_l$ is the projector onto $[l]$ and $c$ is a normalization constant.
The~$\Omega^{(i)}$ point into the direction of the $2l$ multipole vectors that characterize the state~$|\psi\rangle$. Contracting \eqref{factorcoherent} with \eqref{factorpsi} and using the stereographic projection to express points on the sphere in terms of complex numbers $z = e^{i\phi}\cot (\tfrac \theta 2)$, leads to a polynomial
\begin{equation}
\sum_{m=-l}^l \binom{2l}{l+m}^{\!\frac 12} z^{l+m} a_{lm} \ ,
\end{equation}
whose $n \leq 2l$ zeros (roots) correspond to points on the sphere that are antipodal to $n$ of the $2l$ multipole vectors. The remaining $2l -n$ multipole vectors point to the south pole of the sphere. For the CMB data $l$ is an integer, $\Delta T(\Omega)$ is real and consequently $a^*_{lm} = (-)^m a_{l,-m}$. This implies that the zeroes of the polynomial are located at pairs of antipodal points on the sphere and the multipole vectors come in anti-aligned pairs (for details see \cite{Helling}). In \cite{Schupp} this method was introduced to determine explicit formulas for the Wehrl entropy and to prove Lieb's conjecture. Applying those explicit formulas to the case $l = 2$ with two pairs of anti-aligned multipole vectors of length $1/2$ gives the following formula for the Wehrl entropy as a function of the squared chordal distance $\epsilon = \sin^2 (\frac{\alpha}2)$ between the vectors, where $\alpha$ is the angle between them:
\begin{equation}
S_{\mathrm{W}}(\epsilon ) = c - \ln c + \frac{32}{15} - \ln 6 \ ,
\end{equation}
where
\begin{equation}
c = c(\epsilon) := \frac1 {1 -\epsilon (1-\epsilon )} \ .
\end{equation}
For the angular entropy a similar computation gives
\begin{equation}
\label{Sang(eps)}
S_\text{ang}(\epsilon ) =- \frac c2 \left((1-\epsilon )^2 \ln (1-\epsilon )^2 + \epsilon ^2 \ln \epsilon ^2\right) - \ln \frac c2 \
\end{equation}
with $c$ as above.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{Both_S.png}
\caption{Dependence of the angular pseudoentropy $S_{\mathrm{ang}}(\epsilon)$ (solid) and the Wehrl entropy $S_{\mathrm{W}}(\epsilon)$ (dashed) on the squared chordal distance $\epsilon$ between multipole vectors on the sphere with radius $r=\frac{1}{2}$ for $l=2$. The maximum is obtained when the
multipole vectors are orthogonal to each other and the minimum
is obtained when both multipole vectors are the same.
}
\label{BothS}
\end{center}
\end{figure}
Plots of the two functions look very similar (see Fig.~\ref{BothS}), confirming the observed similarities in behavior of the two entropy measures in the CMB analysis (see Fig.~\ref{Ent_comp}).
The polynomial method provides a very fast and convenient way to determine multipole vectors and has been used in \cite{Helling,Dennis2007} and many other publications to analyze the CMB.
Spin coherent states are complete via Schur's lemma
\begin{equation}
(2l +1) \int \frac{d \Omega}{4\pi}\, |\Omega_l\rangle\!\langle\Omega_l| = P_l \ , \label{complete}
\end{equation}
where $P_l$ is the projector onto $[l]$. They are normalized $\langle\Omega_l|\Omega_l\rangle = 1$ but not orthogonal
\begin{equation}
|\langle\Omega_l | \Omega'_l\rangle|^2 = \cos^{4l} (\mbox{\large$\sphericalangle$}{(\Omega, \Omega')}) \ , \label{notorthogonal}
\end{equation}
i.e.,\ they form an overcomplete basis of $[l]$.
In the $l \rightarrow \infty$ limit, $(2l+1) |\langle \Omega_l | \Omega'_l\rangle|^2$ becomes a delta function $\delta(\Omega, \Omega')$ and in this limit the coherent states form an infinite-dimensional orthonormal basis labeled by points on the sphere.
A striking property of coherent states is that the diagonal matrix elements
\begin{equation}
A(\Omega) = \langle\Omega_l|A|\Omega_l\rangle \quad \text{ (lower symbol)}
\end{equation}
of an operator $A$ on $[l]$ already determine that operator uniquely: Let $C = A - B$ with an arbitrary operator $B$, then $C(\Omega) = 0$ for all $\Omega$ implies $C = 0$, i.e., $A=B$. The proof uses analytic properties of the lower symbol. The lower symbol is thus a faithful representation of an operator.
Using Eq.~\eqref{complete}, the trace of an operator $A$ on $[l]$ can be computed as an integral over its lower symbol
\begin{equation}
\mathrm{Tr}_{[l]} (A) = (2l+1) \int \frac{d \Omega}{4\pi} \, \langle\Omega_l|A|\Omega_l\rangle = (2l+1) \int \frac{d \Omega}{4\pi} \, A(\Omega) \ . \label{traceformula}
\end{equation}
Another interesting property is that any operator $A$ on $[l]$ can be expanded diagonally in coherent states
\begin{equation}
A = (2l+1) \int \frac{d \Omega}{4\pi} \, h_A(\Omega) |\Omega_l\rangle\!\langle\Omega_l| \ , \label{upper}
\end{equation}
where $h_A(\Omega)$ is called an upper symbol of $A$. These two properties are in fact closely related: Contracting Eq.~\eqref{upper} with an operator $C$ gives
$\mathrm{Tr} (C^\dagger A) \propto \int d \Omega \, h_A(\Omega)\, \Omega^*(C)$, i.e., the operators that can be represented by an upper symbol as in Eq.~\eqref{upper}, are orthogonal to the operators that are in the kernel of the lower symbol map. Hermitean operators have real lower and upper symbols. Positive semi definite operators and density matrices have unique non-negative lower symbols, but the same is in general not true for upper symbols. Following Wehrl, these properties suggest to interpret the lower symbol of a density matrix $\rho$, which is by definition positive semi definite and normalized, as a probability density and compute
\begin{equation}
(2 l +1) \int \frac{d \Omega}{4\pi} \, \phi( \rho(\Omega)) = (2 l +1) \int \frac{d \Omega}{4\pi} \, \phi(c) \ .
\end{equation}
With $\phi(x) = x$ we can verify the normalization and get $\mathrm{Tr}( \rho) = 1$. With $\phi = - x\ln x$ we compute the Shannon entropy of the probability density $\rho(\Omega)$, which is precisely the Wehrl entropy
\begin{equation}
S_W(\rho) = -(2l +1) \int \frac{d \Omega}{4\pi} \, \rho(\Omega) \ln \rho(\Omega) \ , \label{Wehrl}
\end{equation}
with the special case \eqref{Wehrlentropie} for a pure state $\rho = |\psi\rangle\!\langle\psi|$. The Wehrl entropy was introduced as a semi-classical entropy, which is mathematically better behaved than the Boltzmann entropy in classical statistical mechanics. The Wehrl entropy is always larger than the von Neumann entropy and it is positive even for pure states. Furthermore its definition is rotationally symmetric and it is hence perfectly suited for our purposes. The minimum of the Wehrl entropy is attained for coherent states. This fact is surprisingly difficult to prove. It was first shown for low spin in \cite{Schupp} and then finally in general in \cite{Lieb14,Lieb16}. Computational evidence suggested in fact an analogous but much stronger conjecture for any concave function $\phi(x)$ \cite{Schupp,Schupp_unpub}, which has also been settled affirmatively in \cite{Lieb16}. For our application, the maximal value of the Wehrl entropy is more interesting. The exact value is not known, in fact finding the maximizing pure state is another hard problem, but a reasonable upper limit can be obtained very simply from a totally mixed state: $S_W \leq \ln(2l +1)$. To summarize: The Wehrl entropy is the Shannon entropy of the probability density obtained from the (faithful) lower symbol representation of a density matrix. It has all the right properties for our purposes. The only drawback is that its computation with suitable precision has a high computational complexity. We will now introduce and discuss several alternatives with similar properties, but better computability.
Instead of $- x\ln x$, one can consider other concave functions defined on the interval~$[0,1]$. For $\phi(x) = x(1-x)$ we obtain the quadratic entropy
\begin{equation}
S_\text{quad}(\rho) = 1 - \frac{2l+1}{4l+1} \mathrm{Tr}\left(P_{2l} (\rho \otimes \rho )\right) \ ,
\end{equation}
where $P_{2l}$ is the projector onto spin $2l$ and we have used the following trick \cite{Schupp}:
\begin{align*}
\left(\rho(\Omega)\right)^2 &= \langle\Omega_l|\rho|\Omega_l\rangle^2 \\
&= \langle\Omega_l\otimes \Omega_l|\rho\otimes \rho|\Omega_l\otimes \Omega_l\rangle
= \langle\Omega_{2l}|\rho\otimes \rho|\Omega_{2l}\rangle
\end{align*}
and the trace formula \eqref{traceformula} adapted to spin-$2l$. For a pure state $\rho = |\psi\rangle\!\langle\psi|$ we obtain formula \eqref{quadraticentropie}. The quadratic entropy has a large computational advantage as compared to the Wehrl entropy (see Fig.~\ref{Time}), but its qualitative behavior is a bit different, as can be seen in Fig.~\ref{quad_30}. This is not surprising, since many important properties of entropy like additivity and (strong) sub-additivity depend crucially on the choice of the function $\phi(x) =- x\ln x$. We shall hence not pursue quadratic entropy any further and will try to identify and construct other alternatives of the Wehrl entropy that share its characteristic features but are computationally more accessible. Let us start by reconsidering the main ingredient of Wehrl entropy.
Let $\rho$ be a density matrix on $[l] = \mathbb C^{2l+1}$ and introduce an ancilla Hilbert space $[j] = \mathbb C^{2j+1}$. Using the product property \eqref{productproperty} and normalization of coherent states, we can rewrite the lower symbol $\rho(\Omega)$ that enters the formula for the Wehrl entropy as follows:
\begin{equation}
\begin{split}
\langle\Omega_l|\rho|\Omega_l\rangle &= \langle\Omega_l|\rho|\Omega_l\rangle \langle\Omega_j|\Omega_j\rangle
= \langle\Omega_l \otimes \Omega_j|\rho \otimes \mathbf 1|\Omega_l \otimes \Omega_j\rangle \\
&= \langle\Omega_{l+j}|\rho \otimes \mathbf 1|\Omega_{l+j}\rangle \ ,
\end{split}
\end{equation}
where $\mathbf 1$ is the unit operator on $[j]$. The values of the lower symbol are thus the diagonal elements of a family of infinite-dimensional matrices
\begin{equation}
\rho_j(\Omega, \Omega') = \langle\Omega_{l+j}|\rho \otimes \mathbf 1|\Omega'_{l+j}\rangle \ . \label{infdimmatrix}
\end{equation}
By an infinite-dimensional compact analog of the Schur-Horn theorem the diagonal elements $\rho(\Omega)$ are majorized by the eigenvalues of the $\rho_j(\Omega, \Omega')$ matrices. This implies that any concave function of the values $\rho(\Omega)$ will be larger or equal to the respective function of the eigenvalues of $\rho_j(\Omega, \Omega')$. The Wehrl entropy is therefore larger than or equal to the von Neumann entropy of $\rho_j(\Omega, \Omega')$. For convex functions the inequalities are reversed. See e.g.~\cite{Bengtsson} for an overview of the mathematical background. In the limit $j \rightarrow \infty$ and in view of Eq.~\eqref{notorthogonal} the off-diagonal matrix elements of $\rho_j(\Omega, \Omega')$ become zero and the inequalities become equalities.
Using the property \eqref{complete} on both sides of Eq.~\eqref{infdimmatrix} we can recover a finite-dimensional matrix
\begin{equation}
P_{l+j}\left(\rho \otimes \mathbf 1\right) P_{l+j} \label{projmatrix}
\end{equation}
from $\rho_j(\Omega, \Omega')$, where $P_{l+j}$ is the projector onto the highest spin component $[l+j]$ of the tensor product. The matrix \eqref{projmatrix} has the same eigenvalues as $\rho_j(\Omega, \Omega')$. In fact, if
\begin{equation}
P_{l+j}\left(\rho \otimes \mathbf 1\right) P_{l+j} |V_\lambda\rangle = \lambda |V_\lambda\rangle \label{eigenvalproj}
\end{equation}
then $V_\lambda (\Omega) := \langle \Omega_{l+j} | V_\lambda\rangle$ satisfies
\begin{equation}
(2(l+j) + 1) \int \frac{d \Omega'}{4\pi} \, \langle\Omega_{l+j}|\rho \otimes \mathbf 1|\Omega'_{l+j}\rangle V_\lambda(\Omega') = \lambda V_\lambda(\Omega) \label{eigenvalint}
\end{equation}
and vice versa if $V_\lambda(\Omega)$ is a solution of Eq.~\eqref{eigenvalint}, then $|V_\lambda\rangle = (2(l+j) +1) \int \frac{d \Omega}{4\pi} \, |\Omega_{l+j}\rangle V_\lambda (\Omega)$ satisfies Eq.~\eqref{eigenvalproj}. We have shown that the eigenvalues of the matrix \eqref{projmatrix} majorize the values of the lower symbol of $\rho$ in the sense explained above, namely that inequalities are implied for concave (or convex) functions of these values. It can furthermore be shown that pure states majorize mixed ones and that among the pure states, projectors $|\Omega\rangle\!\langle \Omega|$ onto coherent states will lead to matrices \eqref{eigenvalproj} that majorize all other choices.
Among the concave functionals we are in particular interested in entropy and define an appropriately normalized mixed density matrix
\begin{equation}
\rho_{\text{proj}}^{(j)} = \frac{2l+1}{2(l+j)+1} P_{l+j}\left(\rho \otimes \mathbf 1\right) P_{l+j} \ , \label{mixproj}
\end{equation}
whose von Neumann entropy is what we call the ``projection entropy''
\begin{equation}
S_{\text{proj}}^{(j)}(\rho) = \mathrm{Tr}\left[ \phi\Big(\frac {2l+1}{2(l+j)+1} P_{l+j}\big(\rho\otimes \mathbf 1\big)P_{l+j}\Big) \right],
\end{equation}
with $\phi(x) = x \ln(x)$.
From the fact that the mixed density matrix \eqref{mixproj} has at most $2j+1$ non-zero eigenvalues, we get an upper bound for the projection entropy \cite{Buciumas} $S_{\text{proj}}^{(j)}(\rho) \leq \ln(2j+1)$.
From the $[l+j]$ perspective the Wehrl entropy should also be computed from Eq.~\eqref{mixproj} and we get the aforementioned inequalities. The only differences from the original definition of Wehrl entropy \eqref{Wehrl} is a rescaling of the density matrix and related renormalization of the integral, which leads to a shift in entropy and the following inequality:
\begin{equation}
\label{proj_wehrl_conv}
S_W(\rho) \geq S_{\text{proj}}^{(j)}(\rho) + \ln\left( \frac {2l+1}{2(l+j)+1}\right) \ .
\end{equation}
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{Proj_Wehrl.png}
\caption{Comparison of the Wehrl entropy $S_{\mathrm{W}}(l)$ with the $j=1$-,$10$-,$100$-projection pseudoentropies minus a $j$- and $l$-dependent term, $S_{\mathrm{proj}}^{(j)}(l) + \ln\left(\frac{2l+1}{2(l+j)+1}\right)$, for the NILC 2015 map on the range $[1,30]$. For $j \rightarrow \infty$ the latter converges to the former, but not uniformly.}
\label{ProjWehrl}
\end{center}
\end{figure}
In the limit $j \rightarrow \infty$ this inequality becomes an equality, see Fig.~\ref{ProjWehrl} for the converge of $S_{\text{proj}}^{(j)}$ to $S_{\mathrm{W}}$ for Needlet Internal Linear Combination (NILC) \textit{Planck} data and \cite{Lieb91} for an alternative proof. The projector $P_{l+j}: [l] \otimes [j] \rightarrow [l+j]$ can be expressed in terms of Clebsch-Gordan coefficients:
\begin{equation}
P_{l+j} \big( |l,m\rangle \otimes |j,M\rangle\big) = \sqrt{\frac{\binom{2l}{l+m}\binom{2j}{j+M}}{\binom{2(l+j)}{l+j+m+M}}} \ .
\end{equation}
For large $j$ the projection method provides a good way to compute the Wehrl entropy with high precision. For small $j$ we get an entropy measure with all the nice properties of Wehrl entropy, but a pretty large computational advantage. We will now focus on the case where $\rho$ is a pure state, i.e., $\rho = |\psi_l\rangle\!\langle\psi_l|$ with $|\psi_l\rangle$ computed from the $a_{lm}$ of the CMB data with $l$ fixed. For a pure state $\rho$ the matrix \eqref{projmatrix} can be rewritten as the Gram matrix of a set of vectors $\vec V_M \in [l+j]$ that are labeled by a basis of $[j]$:
\begin{align}
P_{l+j}\left(|\psi_l\rangle\!\langle\psi_l| \otimes \mathbf 1\right) P_{l+j} & = \sum_{M=-j}^j \vec V_M \vec V_M{}^\dagger \\
\text{with} \quad \vec V_M &= P_{l+j}\big(|\psi_l\rangle\otimes|j,M\rangle\big) \ .
\end{align}
The dual Gram matrix
\begin{equation}
\begin{split}
\mathrm{Tr}_{[l+j]} \big( \vec V_M \vec V_{M'}{}^\dagger \big) &= \vec V_{M'}{}^\dagger \cdot \vec V_M \\
&= \big(\langle\psi_l|\otimes\langle j,M'|\big) P_{l+j} \big(|\psi_l\rangle\otimes|j,M\rangle\big)
\end{split}
\end{equation}
has the same non-zero eigenvalues as the original matrix, because for any matrix $C$, $C C^\dagger$ and $C^\dagger C$ have the same non-zero singular values.
We can therefore also use the dual Gram matrix for the computation of the projection entropy. Appropriately normalized and written in basis-independent notation we have
\begin{align}
\tilde \rho_{\text{proj}}^{(j)} &= \frac{2l+1}{2(l+j)+1} \langle \psi_l| P_{l+j}|\psi_l\rangle \\
S_{\text{proj}}^{(j)}(\rho) &= - \mathrm{Tr} \left( \tilde \rho_{\text{proj}}^{(j)} \ln \left(\tilde \rho_{\text{proj}}^{(j)}\right) \right) \ ,
\end{align}
where the expectation value is taken in the first tensor slot of $P_{l+j}$.
Unlike $\rho_{\text{proj}}^{(j)}$ the new density matrix $\tilde \rho_{\text{proj}}^{(j)}$ is in general not a faithful representation of the underlying $\rho$ for $j < l$, but the entropy is precisely the same, while its computation involves smaller matrices and is faster. The computational advantage is particularly large for small $j$. The projection entropy computed in this way is an excellent tool for the analysis of the CMB and other spherically distributed data.
Expanding the unit operator on $[j]$ in Eq.~\eqref{mixproj} in terms of basis states, it can be seen that the map $\rho \rightarrow \rho_{\text{proj}}^{(j)}$ is in fact a trace preserving completely positive map (quantum channel) $[l] \rightarrow [l+j]$ in Kraus form:
\begin{align}
\rho_{\text{proj}}^{(j)} &= \sum_M A_M \rho A_M{}^\dagger \ , \quad \sum A_M{}^\dagger A_M = 1 \\
\quad A_M &= \sqrt{\frac{2l+1}{2(l+j)+1}}P_{l+j} |j,M\rangle \ ,
\end{align}
where the last terms can also be written $P_{l+j} |j,M\rangle = \sum_m |j+l,m+M\rangle\!\langle l, m|$. There is a similar formula for the transformation of the density matrix in the the dual Gram matrix formulation. In view of the $j\rightarrow \infty$ limit, the lower symbol of a density matrix can also be interpreted as resulting from a completely positive map.
We shall now introduce yet another natural choice of a rotationally invariant quantum channel, leading to what we call ``angular entropy'', which shares the nice properties of the aforementioned entropies with the additional advantage of being even faster to compute. Let $L_1$, $L_2$, $L_3$ be the standard angular momentum generators in the spin-$l$ representation and define a mixed density matrix and entropy via
\begin{align}
\rho_\text{ang} &= \frac{1}{l(l+1)} \sum_{i=1}^3 L_i \rho L_i{}^\dagger \\
\quad S_\text{ang} &= - \mathrm{Tr}\left( \rho_\text{ang} \ln \left( \rho_\text{ang} \right) \right) \ . \label{angentropy}
\end{align}
The transformation is obviously of Kraus form and therefore completely positive. It is trace-preserving because $C = \sum_i L_i{}^\dagger L_i$ is the quadratic casimir and has value $l(l+1)$ in the spin $l$ representation. The formula for angular entropy can be written in a basis-independent way by replacing
$\sum L_i \otimes L_i$ by $\tfrac 12 (\Delta C - C\otimes 1 - 1 \otimes C)$, where $\Delta C$ the coproduct of the casimir. In practice the formula is usually rewritten in terms of $\tfrac 1{\sqrt 2} L_\pm$ instead of $L_1$ and $L_2$. Therefore we have included the dagger $\dagger$ in Eq.~\eqref{angentropy}, which is of course not necessary for Hermitian $L_i$.
For a pure state $\rho = |\psi\rangle\!\langle \psi|$, there is also a dual Gram matrix formulation of the angular entropy:
\begin{equation}
\label{ang_alt}
G_{ij} = \langle \psi | C^{-1} L_i{}^\dagger L_j|\psi\rangle \ , \quad S_\text{ang} = - \mathrm{Tr} ( G \ln (G) )\
\end{equation}
(see also \cite{PowerEntropyI,PowerEntropyII} for application to CMB data).
In the way we have written this formula, it is now in fact no longer restricted to individual angular momentum (multipole) numbers $l$. It can also be applied to a range or even a selection of $l$: One simply needs to insert an appropriately normalized state
\begin{equation}
\label{psi_range}
|\psi\rangle = \sum_{l \in \text{select}} \sum_m a_{lm} |l,m\rangle \ ,
\end{equation}
leading to the following explicit algorithm: First determine the Hermitian $3\times 3$ dual Gram matrix $G$
\begin{equation}
G = \begin{pmatrix} G_{11} & G_{12} & G_{13}\\ G^*_{12} & G_{22} & G_{23} \\ G^*_{13} & G^*_{23} & G_{33} \end{pmatrix}
\end{equation}
with matrix elements
\begin{equation}
G_{ij} = \sum_{l \in \text{select}} \frac{1}{2l(l+1)}\mathfrak{g}_{ij}(l,\{a_{lm}\}) \, ,
\end{equation}
with
\begin{align*}
\mathfrak{g}_{11} &= \sum_{m=-l}^l (l+m+1)(l-m) \cdot |a_{lm}|^2 \\
\mathfrak{g}_{12} &= \sum_{m=-l+1}^{l-1}\sqrt{(l^2-m^2)((l+1)^2-m^2)} \cdot a_{l,m-1} a^*_{l,m+1} \\
\mathfrak{g}_{13} &= \sqrt{2}\sum_{m=-l}^{l-1} (m+1)\sqrt{(l+m+1)(l-m)} \cdot a_{lm}a^*_{l,m+1} \\
\mathfrak{g}_{22} &= \sum_{m=-l}^l (l-m+1)(l+m) \cdot |a_{lm}|^2 \\
\mathfrak{g}_{23} &= \sqrt{2}\sum_{m=-l+1}^l (m-1) \sqrt{(l-m+1)(l+m)} \cdot a_{lm}a^*_{l,m-1} \\
\mathfrak{g}_{33} &= 2 \sum_{m=-l}^l m^2 \cdot |a_{lm}|^2 \ ,
\end{align*}
where ``select'' is a chosen selection of multipole angular momentum quantum numbers~$l$. In this paper we typically select a single value at a time, but this can also be a range of values or an even more complex selection. The angular entropy is then computed in terms of the three eigenvalues $\lambda_i \in [0,1)$ of
the normalized mixed angular density matrix $\rho_\text{ang} = G/{\mathrm{Tr} (G)}$, with $\mathrm{Tr}(G) = \sum_{l \in \mathrm{select}} (2l+1) \hat{C}_l$,
\begin{equation}
\label{ang_range}
S_\text{ang} = - \mathrm{Tr}( \rho_\text{ang} \ln \left(\rho_\text{ang}\right) )= - \sum \lambda_i \ln (\lambda_i) \
\end{equation}
(see \cite{Kopp} for an overview of algorithms for the fast and precise computation of eigenvalues of Hermitian $3 \times 3$ matrices). There exist also two range entropy measures using the Wehrl entropy which were identified in \cite{Fintzen}.
The computation of the angular entropy involves only $3 \times 3$ matrices and their eigenvalues. It is by far the fastest method and numerical experiments with actual and simulated data show that it has similar behavior as the Wehrl entropy.
From a theoretical point of view there are some similarities between the angular and projection entropy and hence also the Wehrl entropy: As we have mentioned, the projector $P_{l+j}$ is related to the Clebsch-Gordon decomposition of $[l] \otimes [j]$. Likewise, the angular momentum operators can be interpreted as Clebsch-Gordon coefficients of the decomposition of $[l] \otimes [l]$, but while the projector is onto the highest spin component, the angular momentum generators pick out the adjoint spin-$1$ representation. For the angular entropy there are similar conjectures as for the Wehrl and projection entropies, which are still open and under current consideration. There are several further generalized pseudoentropies -- for example one could choose a (convex) function of the casimir in the definition of angular entropy, e.g.
\begin{equation}
\frac{1}{l(l+1)}\sum_{i_1,\ldots,i_k}L_{i_1}\cdots L_{i_k} \rho L_{i_k} \cdots L_{i_1} \ .
\end{equation}
Below, we focus on the projection and angular entropies that we have defined in this section.
\subsection{Probability distribution of the angular pseudoentropy}
In this section, we want to compare the behavior of the angular pseudoentropy for isotropic and Gaussian maps to its behavior for maps that are constructed from multipole vectors which are distributed uniformly on the sphere according to the surface measure. The latter are statistically isotropic but not Gaussian, hence we investigate deviations from Gaussianity without violating statistical isotropy. For later convenience we often use the logarithmic reciprocal distance of the angular pseudoentropy from its theoretical maximum $X:=\ln\left( \frac{1}{\ln(3)-S_{\mathrm{ang}}} \right)$ as the quantity in investigation. Whenever $X$ is plotted as the independent variable, the dependent probability and cumulative densities are meant to be $p_X$ and $F_X$, and not $p_S$ and $F_S$. On the other hand, if we plot $S$ as the independent variable, the dependent densities are $p_S$ and $F_S$. The probability densities $p_S$ and $p_X$ are related via Eq.\,\eqref{analytical_l2_a}. Note that $X(S)$ is a monotonic function and hence minimal/maximal $S$ corresponds to minimal/maximal $X$. Since we do not take into account the other entropies in this section, we drop the subscript 'ang' in the text.
In the future one should also consider small deviations from isotropy and Gaussianity and investigate the behavior of the entropy distribution in dependence on the small deviation parameters. In this work we leave it at the most simple deviation from Gaussianity in the form of uniform MPVs but consider maps which are constructed from partly Gaussian $a_{lm}$ and partly uniform MPVs as well. Our main aim is to show, that there is a distinction between Gaussian and non-Gaussian maps in the entropy statistics. It should be noted that Gaussianity and statistical independence of the $a_{lm}$ for given $l$ are quite closely related and that a major cause of a deviation in the entropy could result from statistical independence of the $a_{lm}$.
\subsubsection{Semi-analytical distribution for uniform MPVs at $l=2$}
\begin{figure}
\includegraphics[width = 0.49\textwidth]{l2_theo.pdf}
\caption{Probability distribution for uniformly distributed multipole Vectors of the logarithmic reciprocal of the entropy at $l=2$; calculated by solving $S(\epsilon) = s$ for $\epsilon$ numerically with $S(\epsilon)$ given in \eqref{Sang(eps)} and using the analytical formulas \eqref{analytical_l2_a} and \eqref{analytical_l2_b}. The dotted vertical lines display the asymptotics.}
\label{l2_theo}
\end{figure}
For $l=2$, we have the analytical formula \eqref{Sang(eps)} which expresses the angular pseudoentropy as a function of the squared chordal distance $\epsilon$ between multipole vectors. This can be used to obtain an expression for the probability distribution of $X=\ln\left( \frac{1}{\ln(3)-S} \right)$ if the probability distribution of $\epsilon$ is known. If we consider uniformly distributed multipole vectors on the sphere, which yield an isotropic but non-Gaussian map, the ($l=2$)-case is particularly simple. One can fix the first MPV to be $(0,0,1)^T$ and the second to be an arbitrary vector with length $1/2$ and $z\geq 0$. Then for the angle $\Theta$ between both we have $p^{\mathrm{uni}}_{\Theta}(\theta) = \sin(\theta)$ and since $\epsilon(\Theta) = \sin^2(\Theta/2)$, the probability distribution for $\epsilon$ is $p^{(\mathrm{uni},2)}_{\epsilon}(\epsilon) \equiv 2$. This induces the following probability distributions for $S$ and $X$:
\begin{align}
\label{analytical_l2_a}
p^{(\mathrm{uni},2)}_X(x) &= e^{-x} p^{(\mathrm{uni},2)}_S\left(\ln(3)-e^{-x}\right) \\
p^{(\mathrm{uni},2)}_S(s) &= \frac{2}{\frac{\mathrm{d}S}{\mathrm{d}\epsilon}|_{\epsilon(s)}}.
\label{analytical_l2_b}
\end{align}
Unfortunately, $S(\epsilon) = s$ is a transcendental equation and therefore has to be solved for $\epsilon(s)$ numerically. Fig.~\ref{l2_theo} shows $p^{(\mathrm{uni},2)}_X(x)$. Large and small values of the pseudo angular entropy are preferred in this case, because the slope of $S(\epsilon)$ approaches zero in these regimes. Since $S(\epsilon)$ is compactly supported, so is $p_X^{(\mathrm{uni},2)}$. As is shown below, the case $l=2$ is special among all multipoles.
\subsubsection{Numerical distributions at $l>2$}
For $l>2$ the analytical result for $S$ is a complicated expression and therefore we resort to Monte Carlo simulations. We computed probability and cumulative distributions for $l=2$ (Fig.~\ref{l2Gauss} in Appendix \ref{app_plots}) as well as $l=3$--$7$ with $10^5$ random ensembles (Figs.~\ref{mpv_pdf_cdf},\ref{pdf_cdf_ang},\ref{prob_fit},\ref{mpv_vs_gaussian}) and for $l=20,40,60,80,100$ with only $10^2$ random ensembles (see Figs.~\ref{prob_highl},\ref{l100} in Appendix \ref{app_plots}). The angular pseudoentropy is capable of distinguishing clearly between isotropic Gaussian maps and isotropic non-Gaussian maps connected with uniformly distributed MPVs, especially at high $l$, but not at $l=2,3$.
\begin{figure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width = \textwidth]{pdf_mpv_log.pdf}
\end{subfigure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width = \textwidth]{cdf_mpv_log.pdf}
\end{subfigure}
\caption{Probability (top) and cumulative (bottom) distribution of the logarithmic reciprocal of the angular pseudoentropy for uniformly distributed multipoles vectors at multipoles $l=2,3,4,5,6,7$; calculated with $10^5$ random ensembles and smoothed.}
\label{mpv_pdf_cdf}
\end{figure}
Figure \ref{mpv_pdf_cdf} shows the distributions for uniform MPVs at large angular scales. For increasing $l$, the distribution gradually moves to smaller entropy values. This behavior carries on to larger multipole numbers $l\leq 100$ (see Fig.~\ref{prob_highl}). The large entropy behavior shows up to be universal on the range $l \in [3,7]$. Due to the low number of ensembles, we cannot confirm this property for larger $l$, but we observe that the right tail does not stretch further out and hence is bounded from above by the right tail at lower multipoles. This means that from $X = 2.5$ on the probability distribution for uniform MPVs is effectively zero.
\begin{figure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width = \textwidth]{pdf_Gaussian_log.pdf}
\end{subfigure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width = \textwidth]{cdf_Gaussian_log.pdf}
\end{subfigure}
\caption{Probability (top) and cumulative (bottom) distribution of the logarithmic reciprocal of the angular pseudoentropy for Gaussian and isotropic $a_{lm}$ at multipoles $l=2,3,4,5,6,7$; calculated with $10^5$ random ensembles and smoothed.}
\label{pdf_cdf_ang}
\end{figure}
The distributions for isotropic and Gaussian maps at large angular scales (without $l=2$) are shown in Fig.~\ref{pdf_cdf_ang} and the distribution at $l=2$ in Fig.~\ref{l2Gauss} in Appendix \ref{app_plots}. For $l=2$ the distribution peaks at $X \approx 1.5$, decreases towards smaller entropy values and becomes zero at $X \approx 0.9$. The reason for this behavior is that $S$ is in general tightly bounded at the dipole and that MPVs from Gaussian and isotropic maps tend to repel each other. If only two MPVs are present, the most likely configuration is that of orthogonal MPVs, which results in a maximal $S$. For larger $l$, i.e., a higher number of MPVs, the number of configurations that admit a maximal distance increases and hence the distribution is smoothed. $l=3$ is a transition multipole between the smooth and stretched higher multipoles and the sharp and restricted $l=2$. From $l=3$ on, the distribution moves to larger entropy values, which is confirmed at higher multipoles in Fig.~\ref{prob_highl}. The general shape and the width of the probability distribution is approximately conserved when changing $l$ (except for $l=2,3$), only the expectation value is shifted. Hence also confidence levels in $X$ are approximately constant (see also Fig.~\ref{Fit_plot}).
One can try to fit the cumulative distributions of $S$ for isotropic, Gaussian data with a simple function. The comparison between an $e^{-a(x-\ln(3))^2}$-fit and the cumulative distribution as well as between the derivative of the fit function and the probability distribution (see Fig.~\ref{prob_fit}) shows good agreement at the right tail and moderate agreement at the left tail. We conclude that a Gaussian form of the cumulative distribution provides a good first guess also for the probability distribution but should be refined to arrive at a better agreement at the left tail. It should be noted that no analytical result for the left tail is known. Already the calculation of the general lower bound of the entropies is a difficult mathematical problem whose solution for the Wehrl entropy took several decades. Nevertheless, since the whole distribution moves in shape to the right, also the left tail moves to the right when increasing $l$.
\begin{figure}[h]
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \columnwidth]{fit_cdf_Gaussian.pdf}
\end{subfigure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \columnwidth]{fit_pdf_Gaussian.pdf}
\end{subfigure}
\caption{Cumulative (top) and probability (bottom) distribution of the angular pseudoentropy for isotropic, Gaussian $a_{lm}$ at $l=4,5,6,7$; calculated with $10^5$ random ensembles and smoothed. The cumulative distribution was fitted with $f(x;a) = e^{-a(x-\log(3))^2}$ and is shown together with the fit functions. The probability distribution is shown together with the derivatives of the fit functions $f'(x;a)$.}
\label{prob_fit}
\end{figure}
\begin{figure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{l6_log.pdf}
\end{subfigure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{l6_cdf_log.pdf}
\end{subfigure}
\caption{Comparison between isotropic Gaussian $a_{lm}$ and uniformly distributed multipole vectors of probability (top) and cumulative (bottom) distribution of the logarithmic reciprocal of the angular pseudoentropy at $l=6$; calculated with $10^5$ random ensembles and smoothed.}
\label{mpv_vs_gaussian}
\end{figure}
Comparing the distributions, one observes that for increasing $l$ the entropy decreases for uniform MPVs and increases for isotropic, Gaussian maps. While the overlap at $l=6$ is already small but could still have an effect (see Fig.~\ref{mpv_vs_gaussian}), the distributions are clearly distinguished at $l=100$ (see Fig.~\ref{l100}). The distinction between Gaussian and non-Gaussian maps improves for increasing multipole number.
For a quantitative estimate of the behavior of the angular pseudoentropy when only small deviations from Gaussianity are considered, one can investigate the probability distribution using maps that are constructed partly from uniform MPVs and partly from MPVs that are extracted from an isotropic and Gaussian map (see Fig.~\ref{GaussAndUni} in Appendix \ref{app_plots} for $l=6$). It is shown that already a small deviation from Gaussianity in the form of a Gaussian map with one single MPV replaced by a uniformly distributed MPV yields a sizable deviation in the probability distribution and that the distribution converges to the distribution for uniform MPVs rapidly when the number of uniform MPVs is increased. Hence, the entropy measure is highly sensitive to non-Gaussianity. A different but numerically more complicated approach would be to consider the convex combination of the isotropic, Gaussian joint probability distribution of spherical harmonic coefficients and an non-Gaussian distribution. This would have the advantage that the convex deviation parameter could be arbitrarily tuned but it would have the disadvantage of arbitrariness in the choice of the added non-Gaussian contribution. We postpone such an investigation to later works.
It should be noted that an equivalent expression to the angular entropy has already been introduced under the name of power entropy in \cite{PowerEntropyI} but without reference to the Wehrl entropy and completely positive maps. Furthermore, that work made the wrong assumption that the maximal entropy value $\ln(3)$ would be obtained for isotropic maps. The method was applied to \textit{Planck} and \textit{WMAP} in \cite{PowerEntropyII} but with the main focus on the correlation of multipoles with the quadrupole. There, no large-scale anomalies were observed, but correlations with the quadrupole were found on a wider range of scales.
\section{Application to CMB data}
\label{Results}
We use \textit{Planck} 2015 second release data, in particular, the four cleaned full sky maps COMMANDER, NILC, Spectral Estimation via Expectation Maximisation (SEVEM), and Spectral Matching Independent Component Analysis
(SMICA), together with the \textit{WMAP} 7-year ILC cleaned full sky map. The names stand for different cleaning algorithms applied to the original data. COMMANDER uses astrophysical models in order to fill in masked regions that contain foreground contamination, NILC stands for "Needlet Internal Linear Combination" and represents a refinement of the ILC algorithm using needlets in harmonic space, SEVEM uses template fitting and SMICA fills masked regions by a Metropolis Monte Carlo random process. Later in this section we also compare the 2015 results we obtain with results obtained from recently published 2018 \textit{Planck} data. We process the data using the Healpy\cite{Healpix} and Numpy packages for Python 2.7. In order to compute confidence levels, a number of ensembles of Gaussian and isotropic random $a_{lm}$ are treated as input
data for the various entropies. Depending on the entropy the number of ensembles ranges from $30$ to $10^4$.
\subsection{Comparison of pseudoentropies}
The considered pseudoentropies differ in computational expense (see Fig.~\ref{Time}). Computing the angular entropy up to $l=1000$ takes about $90$ seconds per run, while the quadratic entropy is slightly more slow. The quadratic entropy is also fast to compute, but it should be used with care since, due to its $-x^2$ instead of the usual $x \log(x)$ behavior it lacks some of the usual entropy properties. Because the projection entropy converges to the Wehrl entropy for $l \rightarrow \infty$ up to a term which does not depend on the data [see Eq.~\eqref{proj_wehrl_conv} in Sec.\ref{math}] its running time converges as well. Clearly, the Wehrl entropy is the quantity that needs the largest computation time, namely about $3000$ seconds up to $l_{max} = 30$ (for system resources, see Appendix \ref{system}).
\begin{figure}
\begin{center}
\includegraphics[width = \columnwidth]{Time_1000.png}
\caption{Comparison of running times on a standard home computer for one set of $a_{lm}$ when computing the pseudoentropies from $l=1$ up to $l=l_{max}$. For system resources see Appendix \ref{system}.}
\label{Time}
\end{center}
\end{figure}
Figure \ref{Ent_comp} shows that all measures except the quadratic pseudo-entropy exhibit very similar features in the data analysis, which has also been noticed in \cite{Minkov}. In particular, we observe unusually large values at $l=5$ and $28$ and conspicuously small values at $l = 6, 16, 17$ and $30$. On the other hand, the quadratic pseudoentropy singles out other unlikely multipoles, e.g. $l = 14$. This shows again that this measure should be used with care and that the other measures suit our purposes better. It is interesting to see that the most unusual multipoles $l = 5$ and $l = 28$ have an entropy that is far above the expectation value. For non-Gaussian or non-isotropic maps
one would in general expect the entropy to be lower than the expectation, as will be shown later.
In every plot the \textit{Planck} SEVEM map clearly deviates from the other maps from $l=10$ on, showing values of each entropy which are too small, hence indicating a preferred direction in SEVEM. On the other hand, from the comparison of \textit{WMAP} to the \textit{Planck} maps it becomes clear that \textit{WMAP} has already been fairly accurate on large angular scales, since in all of the entropies the \textit{WMAP} line sticks closely to the \textit{Planck} lines. An analysis of unusual multipoles and the differences of the various maps will be given in Sec.~\ref{results_ang}.
The Gaussian expectation values of the angular and quadratic entropies -- in both cases we plot the logarithm of the reciprocal distance to the theoretical maximum -- as well as the Wehrl entropy are monotonously increasing functions of $l$, approaching the maximal values for $l \rightarrow \infty$, while the projection entropies -- logarithmic reciprocal distance plotted as well -- decrease in the low-$l$-regime and increase for larger values of $l$. The logarithmic reciprocal plotting turns out to be especially useful because the $\sigma$-regions do not decrease for large multipoles in this measure. In fact, as will be shown later in more detail, the confidence levels are constant from intermediate $l$ on, while the Gaussian expectation for the angular entropy shows a very simple functional dependence on $l$ as well. Furthermore, in contrast to unlogarithmic plotting both the upper and lower confidence levels have approximately the same width allowing for a better identification of unusual multipoles. For comparison, see Fig.~\ref{unlog_ang} in Appendix \ref{app_plots}, which shows the pure angular pseudoentropy.
It should be noted that smoothed confidence levels appear only in the plots. When calculating p-Values in this work, they are calculated directly numerically with the data and no smoothing takes place.
Concluding, the agreement of features in the different pseudoentropies suggests considering only the numerically cheapest entropy apart from the quadratic one. Hence, in the following only the angular pseudoentropy will be considered.
\begin{figure*}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width = \textwidth]{1-30_proj_j1_10000Ens_log_smooth.png}
\subcaption{($j=1$)-projection entropy, 10000 ensembles of random $a_{lm}$.}
\end{subfigure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width = \textwidth]{1-30_ang_10000Ens_log_smooth.png}
\subcaption{Angular entropy, 10000 ensembles of random $a_{lm}$.}
\label{ang30}
\end{subfigure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width = \textwidth]{1-30_proj_j10_1000Ens_log_smooth.png}
\subcaption{($j=10$)-projection entropy, 1000 ensembles of random $a_{lm}$.}
\end{subfigure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width = \textwidth]{1-30_wehrl_30Ens_smooth.png}
\subcaption{Wehrl entropy, 30 ensembles of random $a_{lm}$.}
\label{wehrl_30}
\end{subfigure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width = \textwidth]{1-30_proj_j100_100Ens_log_smooth.png}
\subcaption{($j=100$)-projection entropy, 100 ensembles of random $a_{lm}$.}
\end{subfigure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width = \textwidth]{1-30_quadratic_1000Ens_log_smooth.png}
\subcaption{Quadratic entropy, 1000 ensembles of random $a_{lm}$.}
\label{quad_30}
\end{subfigure}
\caption{Comparison of pseudoentropies from $l=1$ to $l=30$. The sigma boundaries were determined by a certain number of random maps and have been smoothed with a Gaussian filter in all plots. Note, that concerning the angular entropy the smoothing broadens the confidence levels for $l=2,4$ and straightens them for $l=3$. For higher multipole numbers the smoothing does not add or remove any features.}
\label{Ent_comp}
\end{figure*}
\subsection{Results for angular pseudoentropy with 2015 data}
\label{results_ang}
From fig.~\ref{ang30} we read off that for the angular entropy in the range $1\leq l \leq 30$ five NILC data points lie at $2\sigma$ or even outside of it ($l=5,16,17,28,30$), two of which are even close to $3\sigma$ ($l=5,28$). One could now argue, that it is expected that some data points lie at low confidence levels, but a quick estimation shows that the deviations observed here are still unlikely. The probability for five out of 30 data points to lie outside $2\sigma$ approximately equals the Poisson distribution for five events with a mean rate $\lambda = 1.2 = 30\cdot 0.04$, i.e.,
\begin{equation}
P_{\lambda}(5) = \frac{\lambda^{5}}{5!}e^{-\lambda} \approx 0.6 \%,
\end{equation}
implying that the significance of these unlikely data points is above $2\sigma$.
Turning to higher multipole numbers it would be beneficial to find a method of calculating confidence levels even faster. In this regard we observe that in the logarithmic reciprocal depiction the Gaussian expectation value and confidence levels of the angular entropy behave in a simple fashion, namely the expectation can be fitted with $f(x) =a \, \mathrm{log}(bx+c) $ and the confidence levels with $g(x) = a(1-e^{-b(x-1)})$ (see Fig.~\ref{Fit_plot}). In particular, it turns out that the confidence levels are constant from about $l=30$ up to $l=100$. In the following we assume that this holds true for $l>100$. This assumption is justified by continuity of the angular pseudoentropy and the isotropic, Gaussian probability distribution of spherical harmonic coefficients, i.e.,~no sudden jumps should be expected. Tab.~\ref{Fit_params_ang} in Appendix \ref{app_plots} contains all optimal parameters.
\begin{figure}
\includegraphics[width = \columnwidth]{Angular_Fit_1-100-10000.png}
\caption{Expectation, upper and lower confidence levels of $\log\left(\frac{1}{\log(3)-S_{\mathrm{ang}}(l)}\right)$ with $10000$ ensembles of isotropic and Gaussian random $a_{lm}$ up to $l=100$. The dashed lines represent fits.}
\label{Fit_plot}
\end{figure}
While the fit of the expectation value coincides well with the numerical graph on the whole considered range, the lower confidence fits are not suited for $l=2$ and $l=3$ and the upper confidence fits suit the numerical results from $l=3$ ($1\sigma$), $l=4$ ($2\sigma$) and $l=20$ ($3\sigma$). Hence, using the fits in analysis slightly underestimates the most conspicuous multipoles with values above the expectation value in the range $3 \leq l \leq 20$, but the fits allow for a comparison of the entropy to the expectation from $l=1$ to $l=1000$.
\begin{figure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width = \textwidth]{1-1000_ang_with_fit_log.png}
\end{subfigure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{1-1000_ang_with_fit_log_masked.png}
\end{subfigure}
\caption{Angular pseudoentropy up to $l=1000$ with fitted expectation and confidence levels. Unmasked map (top) and map with SEVEM mask applied and the masked region not filled (bottom).}
\label{ang_1000}
\end{figure}
In Fig.~\ref{ang_1000}, we applied the fits for the angular entropy up to $l=1000$, once for the pure cleaned full sky maps and once masked with the SEVEM mask and without refilling the masked region. In the second case \textit{WMAP} was taken out because of dissimilar NSIDE number of this map and the SEVEM mask. The unmasked shows no obvious deviation of COMMANDER, NILC and SMICA from the expected behavior of a Gaussian map on the whole observed range of multipoles, while single multipoles stick out, as for example NILC at $l=896$, but the data does not exhibit unusual global deviations from the expectation, i.e., deviations on a large range of angular scales. On the other hand \textit{WMAP} and SEVEM clearly fall off from $l=200$ on. While the \textit{WMAP} data is commonly accepted to be inaccurate on very small angular scales, the large drop of SEVEM surprises at first glance. However, the masked plot shows that the deviation of SEVEM from the other \textit{Planck} maps can be explained largely by the strong influence of residual foreground pollution in the SEVEM map. Indeed, the masked \textit{Planck} maps all coincide very well on the whole range, leaving only minor deviations. It can be seen that the pure masking process lowers the entropy for large values of $l$ indicating, as expected, that masking singles out certain directions by removing the galactic plane. This can nicely be seen by taking into account the dashed red line which shows the entropy of a coherent state which represents a map that is confined to a single direction. That masking lowers angular pseudoentropy is not \textit{a priori} clear since we normalize the $a_{lm}$ before computing pseudoentropies and hence there is no lack of absolute power due to masking.
\begin{figure}
\includegraphics[width = \columnwidth]{890-910_angular_fit.png}
\caption{Angular pseudoentropy between $l=890$ and $l=910$, confidence levels and expectation calculated with 1000 random ensembles and smoothed with a Gaussian filter.}
\label{ang_high_l}
\end{figure}
\begin{table}
\begin{tabular}{|c||c||c|c|}
\hline
$l$ & pipeline & $p$[\%] & $\bar{p}_{\text{geom}}$[\%] \\
\hline \hline
\multirow{3}{*}{895} & NILC & 18.8 & \\
& COMMANDER & 36.6 & 17.7 \\
& SMICA & 8.1 & \\
\hline
\multirow{3}{*}{896} & NILC & 0.1 & \\
& COMMANDER & 29.5 & 1.2 \\
& SMICA & 0.6 & \\
\hline
\multirow{3}{*}{897} & NILC & 46.3 & \\
& COMMANDER & 45.5 & 23.4 \\
& SMICA & 6.1 & \\
\hline
\multirow{3}{*}{898} & NILC & 2.6 & \\
& COMMANDER & 13.1 & 6.9 \\
& SMICA & 9.7 & \\
\hline
\multirow{3}{*}{899} & NILC & 3.1 & \\
& COMMANDER & 10.5 & 5.9 \\
& SMICA & 6.4 & \\
\hline
\multirow{3}{*}{900} & NILC & 2.2 & \\
& COMMANDER & 1.1 & 1.7 \\
& SMICA & 2.1 & \\
\hline
\multirow{3}{*}{901} & NILC & 4.1 & \\
& COMMANDER & 5.4 & 4.3 \\
& SMICA & 3.7 & \\
\hline
\multirow{3}{*}{902} & NILC & 0.4 & \\
& COMMANDER & 44.0 & 7.7 \\
& SMICA & 25.9 & \\
\hline
\multirow{3}{*}{903} & NILC & 36.4 & \\
& COMMANDER & 33.2 & 37.9 \\
& SMICA & 45.1 & \\
\hline
\multirow{3}{*}{904} & NILC & 3.0 & \\
& COMMANDER & 31.4 & 13.9 \\
& SMICA & 28.3 & \\
\hline
\multirow{3}{*}{905} & NILC & 45.7 & \\
& COMMANDER & 17.3 & 32.3 \\
& SMICA & 42.8 & \\
\hline
\end{tabular}
\caption{P-values of angular pseudoentropy for $895 \leq l \leq 905$, calculated with 1000 ensembles of Gaussian $a_{lm}$, rounded to one decimal place. In the third column the geometric mean of the p-values of the three maps was chosen because of the multiplicative behavior of probabilities.}
\label{pvals_high_l}
\end{table}
Taking a closer look to angular scales around $l=900$ (see Fig.~\ref{ang_high_l} and Table \ref{pvals_high_l}) for the full sky maps reveals that NILC behaves unusually between $l=895$ and $l=905$. We measure unusualness of multipoles with the p-value, using the convention
\begin{align}
p(l) & := \int_{S_{\mathrm{ang}}(l)}^{\mathrm{log}(3)}\! \mathrm{d}s\, p_S(s) \quad \text{if } S_{\mathrm{ang}}(l) > \langle S_{\mathrm{ang}} \rangle \\
p(l) & := \int_{S_{\mathrm{ang}}^{\mathrm{min}}}^{S_{\mathrm{ang}}(l)}\! \mathrm{d}s\, p_S(s) \quad \text{if } S_{\mathrm{ang}}(l) < \langle S_{\mathrm{ang}} \rangle,
\end{align}
where $p_S(s)$ denotes the probability distribution of $S_{\mathrm{ang}}$ for Gaussian and isotropic $a_{lm}$ and $\langle S_{\mathrm{ang}} \rangle$ the respective expectation value. The entropy value at $l=896$ lies outside the $3\sigma$-region with p-value $\lessapprox 0.1$, where the reason for the inequality is the low number (1000) of random ensembles that have been used to calculate this value, which yields a resolution of $0.1$. This means that on average at most one out of 1000 realizations is expected to be larger than the expectation and to be as unusual as or more unusual than the data point. On the other hand one interpretation is that one out of 1000 multipoles is expected to be at least as unusual as the data point. Since $l=896$ is the only NILC data point on $1 \leq l \leq 1000$ that is outside of $3\sigma$, this multipole is still allowed by statistics. Nevertheless, the NILC values of the angular entropy exhibit small p-values at $l=896,898,899,900,901,902,904$. SMICA behaves a bit less extreme than NILC on the considered range and COMMANDER stays inside or close to the $1\sigma$-region with on average large p-values. While for the multipoles $l=898$ to $l=901$ all the pipelines behave similarly, they deviate from each other at the other multipoles between $l=890$ and $l=910$.
In order to estimate the significance of this multipole range, we calculate the geometric mean over p-values,
\begin{equation}
\langle p \rangle_{\mathrm{geom}}(\mathrm{map}) = \left(\prod_{l=895}^{905} p^{(\mathrm{map})}(l)\right)^{1/11},
\end{equation}
and compare it to the distribution of p-values for Gaussian and isotropic random maps, see Fig.~\ref{p_geo}. For NILC the geometric mean is $\langle p \rangle_{\mathrm{geom}}(\mathrm{NILC}) = 4.4\%$. From 1000 ensembles of random Gaussian and isotropic maps not a single map attains such a small mean p-value, hence we can give an estimate on the upper bound of the likelihood of the NILC data in the given multipole range assuming Gaussianity and isotropy as a null hypothesis
\begin{equation}
\mathbb{L}(\mathrm{NILC};895 \leq l \leq 905) \lessapprox 0.1\%,
\end{equation}
i.e., an about $3\sigma$-significance. It should be noted that by averaging over a range of $l$-modes one does not take into account correlations of these modes induced by inhomogeneous noise. This effect could cause large upper uncertainties in the likelihood and should be considered seriously in more detailed studies. For judging all three full sky maps together, we use the geometric mean of the p-value over the three maps and proceed with these mean p-values as with NILC, resulting in a mean p-value of $8.6\%$ on $[895,905]$ with a likelihood of
\begin{align}
\begin{split}
&\mathbb{L}(\mathrm{COMMANDER}\cdot\mathrm{NILC}\cdot\mathrm{SMICA}; 895 \leq l \leq 905) \\
&\approx 0.8\%,
\end{split}
\end{align}
i.e., a more than $2\sigma$-significance. One should keep in mind, that the significance might be lowered when taking into account correlations between the different maps, which are caused by the simple fact, that all of them are derived from the same physical data. Using the geometric mean implicitly assumes that the ingredients are statistically independent. Hence this significance should be seen rather as a first approximation. The intention in taking the geometric mean over different maps is to obtain a p-value which is to some extent independent of the specifics of the different cleaning algorithms.
\begin{figure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width=\columnwidth]{2015Geomean_2_30_1.png}
\subcaption{Range $l\in [2,30]$ for 10000 ensembles of Gaussian and isotropic random maps.}
\label{p_geo_30}
\end{subfigure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width=\columnwidth]{2015Geomean_895_905_1.png}
\subcaption{Range $l\in [895,905]$ for 1000 ensembles of Gaussian and isotropic random maps.}
\label{p_geo_905}
\end{subfigure}
\caption{Probability distribution (unnormalized) of the geometric mean of p-values plotted with 50 bins. The blue and black vertical lines show the NILC p-value and the geometric mean of p-values from COMMANDER, NILC and SMICA for 2015 data.}
\label{p_geo}
\end{figure}
We conclude that either NILC, and to a lesser extent SMICA, might induce bad characteristics to the data on the mentioned scales or that the COMMANDER algorithm might induce arbitrary isotropy and/or Gaussianity on these scales and therefore distorts the real data. Furthermore, even if one considers all three maps at once by multiplying the p-values and comparing to the expectation, the data is inconsistent with the assumption of isotropy and Gaussianity at a $2\sigma$-level.
It should be noted that this might well be a selection effect due to the particular chosen, non-physically motivated range of scales that was considered. One should ask how likely it is to find a range of multipoles of the given size that which yields an average p-value as low as the considered range. Moreover we do not perform a fully developed and precise statistical analysis, hence the estimated significance might need corrections. The essence here is that the entropy method is capable of highlighting unusual behavior at high multipole numbers.
\begin{table}
\begin{tabular}{|c||c||c|c|}
\hline
$l$ & pipeline & $p$[\%] & $\bar{p}_{\mathrm{geom}}$[\%] (excluding SEVEM) \\
\hline \hline
\multirow{5}{*}{5} & NILC & 1.03 & \\
& COMMANDER & 1.35 & \\
& SMICA & 0.99 & 0.92 \\
& SEVEM & 16.80 & \\
& \textit{WMAP} & 0.53 & \\
\hline
\multirow{5}{*}{16} & NILC & 1.52 & \\
& COMMANDER & 2.01 & \\
& SMICA & 3.58 & 1.89 \\
& SEVEM & 32.10 & \\
& \textit{WMAP} & 1.16 & \\
\hline
\multirow{5}{*}{17} & NILC & 1.66 & \\
& COMMANDER & 1.06 & \\
& SMICA & 3.82 & 2.35 \\
& SEVEM & 19.50 & \\
& \textit{WMAP} & 4.56 & \\
\hline
\multirow{5}{*}{28} & NILC & 0.07 & \\
& COMMANDER & 0.65 & \\
& SMICA & 1.13 & 0.26 \\
& SEVEM & 4.47 & \\
& \textit{WMAP} & 0.09 & \\
\hline
\multirow{5}{*}{30} & NILC & 1.29 & \\
& COMMANDER & 0.32 & \\
& SMICA & 0.45 & 0.63 \\
& SEVEM & 1.93 & \\
& \textit{WMAP} & 0.84 & \\
\hline
\end{tabular}
\caption{P-values of angular pseudoentropy for most conspicuous large angle multipoles, calculated with 10000 ensembles of Gaussian $a_{lm}$, rounded to two decimal places.}
\label{pvals_tab}
\end{table}
Now, we return to the large angular range $l \leq 30$, where we computed p-values with 10000 sets of random $a_{lm}$, see Fig.~\ref{pvals_plot} in Appendix \ref{app_plots} for a plot of the p-values. Tab.~\ref{pvals_tab} shows the six most pronounced large scale multipoles revealing again that unmasked SEVEM does not exhibit the same behavior as the other maps, yielding large p-values at these multipoles while the other maps show small p-values. It turns out that $l=28$ with an average (excluding SEVEM) p-value of about $0.26$\% sticks out most, followed by $l=30$ and $l=5$. Although two of the most conspicuous multipoles display a too large value of the entropy, neither too large nor too small values can directly be identified to be preferred. At these angular scales, the three non-SEVEM \textit{Planck} maps behave quite similarly and the large discrepancy between COMMANDER and NILC is not yet present. Another feature that can be observed is the slight improvement of \textit{Planck} compared to \textit{WMAP}, even for $l < 250$ because on average \textit{WMAP} yields the smallest p-values at these conspicuous multipoles
\begin{figure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width = \textwidth]{1-30_ang_random_not_masked.png}
\subcaption{Unmasked}
\end{subfigure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width = \textwidth]{1-30_ang_random_masked.png}
\subcaption{Masked with \textit{WMAP} intensity mask. The Cosmic Dipole was included before masking and after masking a dipole was removed with Healpy remove\_{}dipole, then the masked region was filled with the original data, see Fig.~\ref{maps} in App.~\ref{app_plots}.}
\end{subfigure}
\caption{Angular pseudoentropy up to $l=30$ from $10$ isotropic and Gaussian random maps with power spectrum (up to $l=200$) of \textit{WMAP} 7-year ILC.}
\label{10_randoms}
\end{figure}
It has been conjectured in \cite{schupp1} that some of the large scale features could be produced by parts of the data processing, namely by a non-linearity in the masking process which mixes the large dipole moment to higher moments when subtracting the dipole. We try to answer this question in a very simplified approach. In Fig.~\ref{10_randoms} we plot the angular entropy for ten isotropic and Gaussian random full sky maps and in a second step we add the non-relativistic contribution of a dipole to the map
\begin{equation}
T(\vec{e}) \rightarrow T'(\vec{e}) = T(\vec{e}) + A\, \vec{d} \cdot \vec{e},
\end{equation}
where $A = 3364.5 \, \mu\mathrm{K}$ denotes the Cosmic Dipole amplitude and $\vec{d} = (x(l,b),y(l,b),z(l,b))^T$ with $(l,b) = (264.00\deg , 48.24\deg)$ \cite{Planck2015I} denotes its direction in the galactic coordinate system, then we mask the map and remove the dipole afterwards again, using this time the build-in Healpy function \textit{remove\_{}dipole}, which returns a map $\widetilde{T}$ that is the closest -- in the meaning of a least square fit -- map to the original $T$ among those maps obeying $\sum_{p\in\mathcal{P}} \vec{e}_p \widetilde{T}_p = 0$, where $\mathcal{P}$ denotes the set of all unmasked pixels. Finally, we refill the masked region with the original data $T$ in order to receive a full sky map, see Fig.~\ref{maps} in App.~\ref{app_plots} for a depiction of this process by maps in Mollweide view. Since the \textit{WMAP} and \textit{Planck} maps behave similarly on large angular scales and working with \textit{WMAP} is computationally cheaper than working with \textit{Planck} maps -- \textit{WMAP} has $\mathrm{NSide} = 512$ and \textit{Planck} $\mathrm{NSide} = 2048$ -- we use the \textit{WMAP} intensity mask for masking as well as the \textit{WMAP} power spectrum up to $l=200$ as the variance of the isotropic and Gaussian $a_{lm}$.
Although a sizable residual effect of the dipole can be seen in the maps, the entropies get modified only slightly. Clearly, at $l=1$ the entropies show the residual part of the dipole, and also at higher $l$ the curves are distorted a little, but the described procedure does not impose any large anomalies and especially it does not result in conspicuous values at $l=5,16,17,28,30$. Thus, we conclude that with our simplified approach no sizable mixing of dipole power to higher multipoles via the masking process can be observed. Finally note that other masking processes with Fourier methods were also applied to the angular and projection pseudoentropies in \cite{Mittelstaedt} and more extensively in \cite{OtgonbaatarMaster}\cite{Otgonbaatar}.
\subsection{Comparison of 2015 and 2018 data with angular pseudoentropy}
In the following we compare the angular pseudoentropy of 2015 \textit{Planck} data to the newest 2018 data release.
Since the 2018 component separation process has been optimized for polarization data, it is expected to come equipped with a few drawbacks in temperature maps, especially for COMMANDER \cite{Planck2018IV}, which carries more residual foreground contamination in the 2018 than in the 2015 temperature map. One should expect to see this feature in the angular pseudoentropy and indeed Fig.~\ref{15vs18_COMMANDER}, which shows the comparison of 2015 and 2018 COMMANDER angular entropy on the two ranges considered in Sect.~\ref{results_ang}, as well as Fig.~\ref{All15vs18}, which shows the relative deviation of 2018 to 2015 data for all \textit{Planck} foreground cleaned full sky maps, confirm this expectation. While even for large angular scales the deviation of COMMANDER is larger than that of SMICA and NILC, for small angular scales COMMANDER drops even below SEVEM. Since in our work we do not want to mask the maps, but need to work with full sky data, it becomes obvious that for our purposes the 2015 COMMANDER temperature data should be preferred to the 2018 data.
\begin{figure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{15vs18_COMMANDER_2_30.png}
\subcaption{Range $[2,30]$}
\end{subfigure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{15vs18_COMMANDER_890_910.png}
\subcaption{Range $[890,910]$}
\end{subfigure}
\caption{Comparison of angular entropy with 2015 and 2018 data for COMMANDER.}
\label{15vs18_COMMANDER}
\end{figure}
Both NILC and SMICA show only a slight deviation in 2018 compared to 2015, both on small and large angular scales with NILC 2018 entropy being identical to the 2015 entropy with at most $5\%$ deviation, which is reached only at $l=28$, see Figs.~\ref{15vs18_NILC}, \ref{15vs18_SMICA} and \ref{All15vs18}. For all other multipoles the NILC deviation is nearly negligible. Since the NILC component separation process has been left nearly unaltered from 2015 to 2018, NILC is most useful for observing the effects of the improved Cosmic Dipole calculation and the removed AD non-linearity. The influence of the former is restricted mainly to a very slight reduction of significance of the both most unlikely multipoles we considered, namely $l=28,896$.
SEVEM has been clearly enhanced in 2018, as shown in Figs.~\ref{15vs18_SEVEM} and \ref{All15vs18}. In 2015 data the angular entropy of SEVEM was far too low from $l=13$. That behavior came particularly clear at small angular scales. In the preceding sections we argued that this effect stems from the residual contamination of SEVEM data by the galactic plane. In 2018 the entropy is constantly shifted to higher values from $l=13$ on, approaching a nearly constant relative improvement of about $18\%$ at small angular scales. Nevertheless for our purposes the SEVEM map still lacks quality at small angular scales and visible residual foreground pollution is left.
\begin{figure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{15vs18_NILC_2_30.png}
\subcaption{Range $[2,30]$}
\end{subfigure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{15vs18_NILC_890_910.png}
\subcaption{Range $[890,910]$}
\end{subfigure}
\caption{Comparison of angular entropy with 2015 and 2018 data for NILC.}
\label{15vs18_NILC}
\end{figure}
\begin{figure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{15vs18_SEVEM_2_30.png}
\subcaption{Range $[2,30]$}
\end{subfigure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{15vs18_SEVEM_890_910.png}
\subcaption{Range $[890,910]$}
\end{subfigure}
\caption{Comparison of angular entropy with 2015 and 2018 data for SEVEM.}
\label{15vs18_SEVEM}
\end{figure}
\begin{figure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{15vs18_SMICA_2_30.png}
\subcaption{Range $[2,30]$}
\end{subfigure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{15vs18_SMICA_890_910.png}
\subcaption{Range $[890,910]$}
\end{subfigure}
\caption{Comparison of angular entropy with 2015 and 2018 data for SMICA.}
\label{15vs18_SMICA}
\end{figure}
\begin{figure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{All_15vs18_2_30.png}
\subcaption{Range $[2,30]$}
\end{subfigure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{All_15vs18_890_910.png}
\subcaption{Range $[890,910]$}
\end{subfigure}
\caption{Relative deviation of (logarithmic reciprocal deviation from maximal value of the) angular entropy with 2018 data compared to 2015 data for all maps.}
\label{All15vs18}
\end{figure}
\begin{figure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{2018Geomean_2_30_1.png}
\subcaption{Range $[2,30]$ calculated with 10000 ensembles.}
\end{subfigure}
\begin{subfigure}[c]{\columnwidth}
\includegraphics[width = \textwidth]{2018Geomean_895_905_1.png}
\subcaption{Range $[895,905]$ calculated with 1000 ensembles.}
\end{subfigure}
\caption{Probability density of geometric mean of p-values and respective values for NILC and the geometric mean of maps for 2018 data.}
\label{Geommean2018}
\end{figure}
At large angular scales COMMANDER and SMICA exhibit a joint deviation behavior at the most unlikely multipoles. Both large entropy values at $l=5$ and $l=28$ are slightly suppressed, but still outside of $2\sigma$, and the small entropy value at $l=30$ is enlarged. While for SMICA all three multipoles still lie outside of $2\sigma$, COMMANDER shifts them towards smaller confidence. In contrast to that, the two multipoles $l=16,17$ get shifted to more unlikely values in both maps. These considerations show that the most conspicuous multipoles at large angular scales could partly be caused by unoptimized component separation, but that AD non-linearity and the Cosmic Dipole identification show only a minor effect, since NILC is nearly unaltered.
At small angular scales we can only use NILC and SMICA since SEVEM is still off and COMMANDER has been degraded due to polarization optimization. The main observation concerning single multipoles, that can be made at this point, is that for SMICA the most unlikely multipole $l=896$ on the considered range is improved but the previously normal multipole $l=910$ is shifted towards $2\sigma$.
In Tab.~\ref{tab_geommeans} in App.~\ref{app_plots} we gather p-values and likelihoods for 2015 and 2018 data. There we also include the range $[890,910]$ in order to compare it to $[895,905]$. Since COMMANDER is off at small angular scales in 2018, we also consider the geometric mean of NILC and SMICA alone. The range $[2,30]$ is normal in both data releases with 2015 being slightly better than 2018 both for NILC alone and for the geometric mean of COMMANDER, NILC and SMICA. Fig.~\ref{Geommean2018} shows that even though the geometric mean of p-values is smaller than the expectation, the significance for that is too low and hence we can conclude that using the angular entropy method the whole range $[2,30]$ is compatible with the assumption of isotropic and Gaussian temperature fluctuations, which was also pointed out in \cite{PinkwartSchwarz} for the range $[2,50]$.
In contrast the range $[895,905]$ displays unlikely behavior in both releases. We observe a slight enhancement from $4.4 \%$ to $6.2 \%$ in the geometric mean of p-values for NILC and from $5.3\%$ to $8.1\%$ in the geometric mean of p-values for the geometric mean of NILC and SMICA. These enhancements correspond to changes in likelihood from $0.1\%$ in 2015 to $0.5\%$ in 2018 for the latter and no change of likelihood for the former. None of thousand random isotropic and Gaussian maps admits such a low geometric mean of p-values as NILC in both 2015 and 2018, hence the Likelihood is bounded from above by $\approx 0.1\%$ in both releases. Enlarging the range a bit from $[895,905]$ to $[890,910]$ increases the likelihoods about a factor of $1$ to $15$, but keeps them under $2.5\%$. We can conclude that the improvements made in the 2018 data processing improve also the small angular scales, but the features are still at nearly $2\sigma$ for the mean of NILC and SMICA at $[890,910]$ and at or outside of $3\sigma$ for NILC at $[895,905]$. At this point we should clarify again that the question, how likely it is to find a range of such width outside of $3\sigma$ is postponed to the future and that here we might fall for the selection effect.
\subsection{Results for range angular pseudoentropy with 2015 data}
The range angular pseudoentropy provides an additional measure for quantifying unlikeliness of multipole ranges and also collections of different multipoles which are not necessarily in a row, see Eqs.~\eqref{psi_range}-\eqref{ang_range} in Sect.~\ref{math}. Unfortunately it is a mix of a correlation and an averaging measure, hence one needs to consider both the range angular and the single multipole angular entropies in order to identify effects of correlation of different multipoles, which are usually expressed by small range entropies. On the other hand, if one is solely interested in the mean likelihood of a given range, the geometric mean of p-values of the single multipole angular entropy is surely the better measure. Aside the partial correlation interpretation, the big advantage of the range angular entropy is its pseudoentropy nature and henceforth its interpretation as an entanglement measure. In Tab.~\ref{pvaltab} in Appendix \ref{app_plots} we gather p-values for the range angular entropy for various ranges on large angular scales and for the small scale range $[895,905]$, as well as the signed deviation of the entropy for 2015 NILC from the isotropic and Gaussian expectation.
One directly observes that the range entropy of the range $[2,3]$ is too small at $2\sigma$. Having in mind Fig.~\ref{ang30} this low p-value should mainly stem from correlation of $l=2$ and $l=3$. Since the single value angular entropy is related to multipole vectors via their chordal distances, we propose that this feature is the same as the correlation of the quadrupole and octupole multipole vectors (see \cite{Schwarz2015} for more details).
The next observation concerns the range $[2,6]$. Even though $l=5$ is too high at $\approx 1\%$ p-value and $l=6$ too low at $\approx 7\%$ p-value, the only unlikely collection of multipoles on this range containing $l=5$ or $l=6$ is the collection $\{2,5\}$ which gives a range entropy that is too small at nearly $2\sigma$-level. Since the average of $\{2,5\}$ should not differ significantly from $\{3,5\}$, which in turn has a p-value of above $50\%$, the low p-value of $\{2,5\}$ stems from anti-correlation or entanglement of $l=2$ with $l=5$. The multipole vectors of $l=2$ and $l=5$ are unusually widespread over the sphere. Hence, we draw the conclusion that the unusually large value of $l=5$ concerning the single value angular entropy is induced by the CMB quadrupole, which itself is mainly influenced by the Cosmic Dipole, that is assumed to constitute the main ingredient of large scale multipole vector anomalies (see \cite{PinkwartSchwarz}). We furthermore point out that to the authors knowledge this (anti-)correlation of the CMB quadrupole with $l=5$ has not been observed so far.
The range $[2,30]$ yields a p-value of around $18\%$ which is compatible with the likelihood of about $15\%$ calculated with the geometric mean of p-values for the single multipole angular entropy. The range entropy lies slightly below the expectation indicating a mixture of a slight averaged preference for a direction on the sky and a slight correlation of multipoles, though being within $1\sigma$.
For the collection $\{5,28\}$ the p-value is smaller than one would expect from the average of both multipoles indicating correlation of these two multipoles.
On small angular scales the p-value $0.5\%$ for the range $[895,905]$ is compatible with the likelihood from the geometric mean, which we gave the approximate upper bound $0.1\%$. For both ranges $[2,30]$ and $[895,905]$ we obtain slightly larger p-values with the range entropy than likelihoods with the geometric mean, which could be caused by reduction of significance due to averaging of large and small values of the angular entropy. The fact that the range entropy lies below the expectation again indicates, that a direction might be preferred in the data and/or different multipoles might be correlated.
\section{Summary and Discussion}
\label{sum}
Building upon the Wehrl entropy we introduced three types of pseudoentropies which approximate the Wehrl entropy but allow for much faster computation and hence analyzing CMB data up to $l=1000$. Those entropies are the $j$-projection, the quadratic and the angular entropy. While the quadratic is simple in fashion it should be disregarded, because of its $x^2$ instead of $x\log(x)$ behavior and the numerical problems one runs into with it for high $l$ if not approximating the expression. All pseudoentropies are rotationally invariant measure of quantum randomness or entanglement on spin-$l$ states and hence on each multipole of CMB temperature fluctuations on the sphere. Contrary to the usual von Neumann entropy these pseudoentropies do not vanish for pure states. In the spirit of thermodynamics, the entropies are useful for reducing $2l+1$ d.o.f. per multipole to a single number per multipole just as one usually does with $C_l$ but which complements it in the case of anisotropies or non-Gaussianities.
We showed that for $l=2$ both the Wehrl and the angular entropy depend only on the squared chordal distance of multipole vectors, yielding another view on this method and a connection to many previous studies of CMB analysis.
Although our focus was on introducing the methods and clarifying their properties, in order to demonstrate the usage of these methods we applied the introduced types of pseudoentropies for analysis of CMB temperature full sky maps and it turned out that they all show similar behavior and the same characteristic features of the maps, except for the quadratic pseudoentropy. Since the angular entropy is the computationally cheapest measure, the rest of the analysis was devoted solely to the angular entropy, which reaches its maximum $\log(3)$ for maximally mixed states, which cannot be reached by pure temperature maps, and its minimum probably for coherent states; it is mathematically known for sure for $l=1/2,1$ only. The physical data from the \textit{Planck} 2015 maps and \textit{WMAP} ILC was compared to isotropic and Gaussian maps, and some multipoles with particularly small p-values were found, in particular $l=5,16,17,28,30$ on large angular scales, and the range $895 \leq l \leq 905$ for NILC, the likelihood of which we approximately bounded from above by $0.1\%$, but there could be as well further unusual angular scales and we did not take into account the selection effect statistically. On average three out of four \textit{Planck} maps do not show abnormal global behavior for $l \leq 1000$, that means deviations from isotropy and Gaussianity on a large range of scales, and the abnormality of the fourth map -- SEVEM -- can be removed by masking the galactic plane. As expected, the \textit{Planck} maps can be considered as a clear improvement compared to \textit{WMAP} on small angular scales $l>200$.
A comparison of isotropic, Gaussian random maps to maps constructed from uniform multipole vectors showed that our method is sensitive to deviations from Gaussianity, resp.~statistical dependence of spherical harmonic coefficients, and, due to its rotationally invariant nature, also isotropy. One should note that uniform multipole vectors are clearly distinguishable from isotropic and Gaussian maps for which the multipole vectors exhibit repulsion.
We were not able to identify the masking process as a reason for all or some of the mentioned conspicuous multipoles with a simple masking approach. The fact that these multipoles have low p-values in all of the maps except for SEVEM indicates a different reason behind them. Nevertheless, one should not withhold that a statistical fluke cannot be excluded, even if the p-value for $l=28$ lies below half a percent.
Comparing \textit{Planck} 2015 to 2018 data confirmed the expectation that the COMMANDER full sky 2018 map cannot be used for our methods at large angular scales without masking. The SEVEM full sky map has been enhanced from 2015 to 2018 but still carries too much foreground pollution in it when not masked. SMICA only deviates slightly and NILC is nearly unaltered. The unlikely features we observe are present in both data releases but with slightly less significance in 2018 than in 2015. The fact that NILC is left nearly unchanged suggests that AD non-linearity and unoptimized Cosmic Dipole removal do not account for the observed features. Nevertheless the component separation still might do.
Eventually we considered the angular range entropy as a mixture of a measure of range-or-collection-averaged angular entropy and of correlation between different multipoles. It turned out that the results for the ranges $[2,30]$ and $[895,905]$ are consistent with the likelihoods obtained from the geometric mean of p-values of the standard angular entropy. We found the anti-correlation of the quadrupole with $l=5$ and the correlation of $l=2,3$ especially interesting. While the latter supports the previously observed quadrupole-octupole correlation, the former hints towards a connection between the high angular entropy value at $l=5$ and the CMB dipole, which itself has been proposed to be influenced mainly by the Cosmic Dipole in the past.
There are several tasks left for the future. First of all a deeper statistical analysis needs to be done, especially considering wider and smaller angular scale ranges. Our analysis has shown, that there might be something hidden at small angular scales and it would be interesting to see more results on this. For identifying possible foreground effects it would be useful to apply our methods to the foreground maps from \textit{Planck} data and investigate to what extent certain features are foreground residuals of the component separation. Furthermore the influence of noise, especially inhomogeneous noise, on the results needs to be evaluated. So far we have introduced the methods, explained some of their mathematical behavior and performed a perfunctory analysis in order to illustrate their usage. Furthermore one could try to apply some of these methods to polarization data. For the Wehrl entropy it is clear that a direct generalization is possible, but for the angular entropy further analysis is needed. Angular and Wehrl entropy are also useful in quantum information theory as real entanglement measures. Finally, our work is also interesting from a mathematical perspective. The fact that the Wehrl entropy is minimized in general by SU(N)-coherent states is known as the Lieb conjecture and has been proven \cite{Lieb16}. The same question is, however, still open for the angular entropy.
\begin{acknowledgments}
Based on observations obtained with \textit{Planck} (http://www.esa.int/\textit{Planck}), an ESA science mission with instruments and contributions directly funded by ESA Member States, NASA, and Canada; and \textit{WMAP} (https://map.gsfc.nasa.gov/).
We acknowledge financial support by the Deutsche Forschungs Gemeinschaft Research Training Group 1620 \emph{Models of Gravity}. We thank Valentin Buciumas, Jessica Fintzen, Otgonbaatar Myagmar, and Gregor Mittelstaedt for collaboration at earlier stages of the research leading to this publication. Useful discussions with Elliott Lieb are gratefully acknowledged. We also thank the referee for useful comments and suggestions.
\end{acknowledgments}
\bibliographystyle{cj}
|
\subsection{Methods}
Our experiment consists of polystyrene sulfate latex particles (Invitrogen), with diameters 3.7~$\mu$m (Lot 1839598) and 5.4~$\mu$m (Lot 1818113) in roughly equal numbers (Fig.~\ref{fig:readexp}a), adsorbed at the interface between decane (``99\%+,'' ACROS Organics) and deionized water in a 60~mm-diameter glass dish~\citep{Keim:2014hu}. The particle suspension is handled using pipette tips and Eppendorf tubes that are free of surface treatments (Axygen ``Maxymum Recovery'') and it includes 50\% ethanol as a spreading agent.
These particles exhibit long-range electrostatic repulsion~\cite{Masschaele:2010da}, and so at the concentrations used here (area fraction $0.36 \pm 0.04$ \cite{suppmat}) each particle is mechanically over-constrained by its neighbors but does not touch them---forming a soft, frictionless jammed 2D solid with a typical spacing $a=8.2$~$\mu$m between particle centers, as measured at the first peak of the pair correlation function $g(r)$~\cite{trackpyv04}.
\begin{figure}
\begin{center}
\includegraphics[width=2in]{apparatus.pdf}
\end{center}
\caption{Schematic of the interfacial stress rheometer apparatus. A magnetized needle is adsorbed at the interface, along with the particle monolayer, in a channel formed by two 18~mm square glass coverslips that are held 3.2~mm apart by a nylon clamp (not shown). Dashed lines indicate that the meniscus rises slightly at the walls. Arrow indicates the direction of magnetically-driven needle motion, parallel to the channel.
\label{fig:apparatus}
}
\end{figure}
We use an interfacial shear rheometer~\citep{Brooks:1999ky,Reynaert:2008dm,Keim:2014hu} to shear the material. As shown in Fig.~\ref{fig:apparatus}, a steel needle is adsorbed at the interface and is positioned by a magnetic field between two glass walls; a computer-controlled perturbative field then drives the needle sinusoidally. Shear is nearly uniform, due to the no-slip boundary conditions at the needle and walls, and the high ratio of interfacial shear stress to bulk liquid (oil and water) shear stress---the Boussinesq number $\text{Bq} \sim 100$~\cite{Reynaert:2008dm}. The needle is 230~$\mu$m in diameter and 32~mm long; it protrudes from the ends of the channel to keep this yield-stress material from forming solid ``plugs'' there. This means that the working sample is approximately 18~mm long and 1.5~mm wide on each side of the needle.
Synchronously with shearing, we image and track $\sim$40,000 particles in a $1.9 \times 1.4$~mm area~\citep{Crocker:1996wp,trackpyv04}. We use a long-distance microscope (Infinity K2/SC) and 4-megapixel machine-vision camera (Mikrotron 4CXP) at a magnification of 0.82~$\mu$m/pixel and a frame rate of 30 frames/s. To find particle locations, we first apply a binary threshold to each image and estimate the locations of large particles. This allows us to compute the precise centroids of particles in two passes, optimized separately for large and small particles. High-throughput tracking is performed with the open-source ``trackpy'' software~\cite{trackpyv04,Crocker:1996wp} using the channel-flow prediction and adaptive search features, with the help of an image-registration algorithm to compensate for occasional motions of the microscope due to external vibrations. To reduce the effect of spurious rearrangements caused by particle-tracking errors, we discard any particle that is not tracked continuously over an entire set of samples, \emph{e.g.}\ the entire readout process.
Analysis involves measuring the differences between particle positions at two different times. For each particle, we subtract the average motion of a region of nearby material (radius $R_\text{disp} = 8.5a$), to avoid spurious signals due to small motions of the camera or variation of the needle position, yielding $\Delta \vec r_\text{local}$~\cite{Keim:2014hu,philatracksv02}. Choosing $R_\text{disp} = 4.5a$ or $16.5a$ does not change our qualitative results~\cite{suppmat}.
\subsection{Training and Readout}
All of the experiments reported here follow the protocol: (1) a ``reset'' phase where we apply 6 cycles with strain amplitude $\sim 70\%$ at 0.1 Hz; (2) a ``training'' phase where we apply oscillatory shear at 0.05 Hz with a repeating pattern of strain amplitudes for 176 cycles, recording video for the last 24; (3) a ``readout'' phase. Figure~\ref{fig:readexp}a shows strain \emph{vs.}\ time at the end of one experiment. Training involves the pattern of amplitudes $\gamma_2$, $\gamma_2$, $\gamma_2$, $\gamma_2$, $\gamma_1$, $\gamma_1$, $\gamma_1$, $\gamma_1$ (176 cycles = 22 repetitions). We use $\gamma_1 = \gamma_2$ (Fig.~\ref{fig:readexp}), as well as $\gamma_1 < \gamma_2$ and $\gamma_1 > \gamma_2$ (Fig.~\ref{fig:read2}). $\gamma_1$ is always applied \emph{last} before readout. Amplitudes are repeated within the pattern to reduce the possibility that the material would ``learn'' a 2-cycle trajectory, in which the amplitude of one cycle always predicts the amplitude of the next. The duration of training is much longer than the $\sim$15 cycles typically required to reach an apparent steady state~\cite{Keim:2013je,Keim:2014hu}, so that by the end of training virtually all particles return to the same positions after a complete 8-cycle pattern, despite many rearrangements (Fig.~\ref{fig:readexp}a). The median normalized MSD after 8 cycles in the steady state is $0.0010$, which is a scale for the noise floor in measurements like Fig.~\ref{fig:readexp}c.
\begin{figure}
\includegraphics[width=3.43in]{read2.pdf}
\caption{
\label{fig:read2}
Two memories.
\textbf{(a)} Experimental protocol as in Fig.~\ref{fig:readexp}b, but with $\gamma_2 = 4\%, \gamma_1 = 3\%$.
\textbf{(b)} Experimental readout for all particles as in Fig.~\ref{fig:readexp}c. ``$4, 3$'' is the protocol in panel (a) and shows evidence of both memories; ``$3,4$'' indicates $\gamma_2 = 4\%$ was applied last before readout, erasing the memory at 3\%.
\textbf{(c)} Simulations by Adhikari and Sastry~\cite{Adhikari:2018il}, after training with 6\% and 4\%.
}
\end{figure}
Consistent with other studies~\cite{Fiocco:2014bz,Fiocco:2015kr,Adhikari:2018il,Mukherji:2019hp}, we see evidence for both single and multiple memories. When we train with both 3\% and 4\% strain, applying $\gamma_1 = 3\%$ last before readout (Fig.~\ref{fig:read2}a, and ``4, 3'' curve in Fig.~\ref{fig:read2}b), we observe a memory at $\gamma_\text{read} = 3\%$, but we also see evidence for a memory above 3\%: MSD in that region is distinct from the ``3'' curve. The result is very different when we exchange $\gamma_1, \gamma_2$ and apply the larger amplitude last (``3, 4''): the signature of the smaller value is gone, which differs from the expected behavior of a dilute suspension~\cite{Keim:2011dv,Paulsen:2019ki,Keim:2019aa}. These results bear a resemblance to return-point memory (RPM). In the present context, RPM means that a cycle with amplitude $\gamma_1$ restores the system to the state it had after the previous cycle with $\gamma_1$ (i.e., minimizes MSD), so long as strain did not exceed $\gamma_1$ in the interim (the difference between ``4, 3'' and ``3, 4'' training)~\cite{barker83,Sethna:1993ts,Paulsen:2019ki,Keim:2019aa}.
In the rest of this paper, we explore the possibility that RPM could at least partially explain memory in amorphous solids.
\subsection{Model and Mechanism}
\begin{figure}
\includegraphics[width=3.43in]{preisach.pdf}
\caption{
\label{fig:preisach}
The Preisach model illustrates how specific memories can emerge from hysteresis.
\textbf{(a)} Applied field $H$ is varied to store and read memories of 4\% (trained state TS$_4$) and 3\% (TS$_3$). Labels correspond to panels \textbf{(b)--(e)}, which show the model's many hysterons, plotted according to $H^+$, $H^-$. Increasing $H$ converts hysterons to the $+1$ state, growing the lighter-shaded region rightward; decreasing $H$ grows the $-1$ region downward. For example, to go from (b) to (c) we decrease $H$ from 4\% to $-4\%$, then increase to 0.
In (e), a cycle with $\gamma_\text{read} = 1\%$ flipped hysterons in the outlined region to $+1$, but did not flip them back,causing a difference with the trained state (d).
\textbf{(f)} A simulated Preisach model is read out by measuring the fraction of hysterons $f_\text{diff}$ that differ from a trained state. Measuring relative to TS$_3$ yields both memories (the ``4,3'' curve). Measuring relative to TS$_4$ (``4,3 from TS$_4$''), or applying the 4\% amplitude last before readout (``3,4''), shows a 4\% memory only.
}
\end{figure}
We illustrate the mechanics of RPM with the Preisach model, originally used to study hysteresis in ferromagnets~\cite{barker83, Preisach:1935jm}. It considers many hysteretic subsystems, or ``hysterons," that are coupled to an external field $H$. The $i$th hysteron will ``flip" from its --1 state to +1 when $H$ is increased past $H_i^+$; it will flip back to --1 when $H$ is decreased past $H_i^-$. $H_i^+$ and $H_i^-$ are distributed uniformly from $-0.1$ to $0.1$ to represent disorder, and $H_i^+ > H_i^-$ to represent dissipative dynamics. We apply the training and readout protocol in Fig.~\ref{fig:preisach}a, and monitor hysterons' states in Fig.~\ref{fig:preisach}(b--e). The figure shows that our earlier definition of RPM is recursive: when $H_\text{read} = 4\%$, we recover the same state (b) as when amplitude 4\% was last applied, regardless of the intervening storage and recovery of a 3\% memory. In effect, there are \emph{two} trained states, which we denote TS$_3$ and TS$_4$.
Figure~\ref{fig:preisach}e highlights hysterons that because of their $H^+$, are placed in the +1 state by applying $H \ge 1\%$; but because of their $H^-$, require $H \le -3\%$ to be fully reversed. During a readout cycle with amplitude $H_\text{read} \ge 3\%$, these hysterons would each flip to $+1$ and back to $-1$, but with $H_\text{read} = 1\%$, they are stuck in their +1 states. Extended to all hysterons, this basic mechanism of RPM means that reducing the driving amplitude leaves the entire system in a different state, but it also means that previous states TS$_3$ and TS$_4$ can be restored by increasing the amplitude to previous values.
Figure~\ref{fig:preisach}f shows that the readout protocol in Figs.~\ref{fig:readexp}c and \ref{fig:read2}a can also read RPM in a simulated Preisach model with 25,000 hysterons, with non-monotonic curves as in the experimental results. Instead of MSD, at the end of each cycle we measure the fraction of hysterons $f_\text{diff}$ that do not match a trained state. In the ``4, 3'' curve, the change in slope as $H_\text{read}$ passes 4\% comes from the many hysterons with $H^+ > 4\%$ or $H^- < -4\%$ that were heretofore inactive. Figure~\ref{fig:preisach}f also verifies the recursive nature of RPM with respect to TS$_3$ and TS$_4$.
These curves roughly match our experiments, and are strikingly similar to results from molecular dynamics simulations (Fig.~\ref{fig:read2}b)~\cite{Fiocco:2014bz, Fiocco:2015kr, Adhikari:2018il}, despite key differences in these systems' physics that we discuss below. (In recent bubble raft experiments~\cite{Mukherji:2019hp} the larger memory was instead observed as a second minimum, but the annular geometry in that work makes direct comparisons difficult.)
\begin{figure}
\includegraphics[width=3.3in]{activity.pdf}
\caption{
\label{fig:activity}
Training and readout at the level of individual rearrangements of particles, from one of the ``4, 3'' movies averaged in Fig.~\ref{fig:read2}b.
\textbf{(a)} Strain at end of training, and readout, as in Fig.~\ref{fig:read2}a. Pairs of thick blue and red vertical lines indicate trained states TS$_4$ and TS$_3$ that follow 4\% and 3\% cycles, respectively, and the corresponding times during readout (see axis at bottom of figure).
\textbf{(b)} MSD of all particles in one movie, measured from 3\% (upper) and 4\% (lower) trained states. An open diamond or circle marks the value at the end of each cycle, as plotted in Fig.~\ref{fig:read2}. Closed symbols show values computed from only the 5 labeled clusters of rearranging particles in Fig.~\ref{fig:readexp}a; values for $\gamma_\text{read} > 4$ are too large to be plotted here. Both TS$_4$ and TS$_3$ are recovered approximately during readout.
\textbf{(c)} MSD of each labeled cluster in Fig.~\ref{fig:readexp}a, calculated relative to TS$_4$ (thick blue curve) and TS$_3$ (thin red curve). A blue circle and red diamond mark the end of each cycle. Each group's state at the end of a cycle is hysteretic, and depends on the strain amplitude in a different way. Collectively they give rise to the memory readouts in panel (b).
}
\end{figure}
We now consider whether the mechanism illustrated by the Preisach model---hysteretic subsystems that can get stuck in one state when the driving amplitude is decreased---is relevant for our amorphous solid. We examine a movie with ``4, 3'' training in Fig.~\ref{fig:activity}a. To identify candidate subsystems of rearranging particles, we focus on the region in Fig.~\ref{fig:readexp}a (13\% of the recorded area), and compare each particle's position at the beginning of readout with its position at all other times that $\gamma = 0$ (twice per cycle), during the interval in Fig.~\ref{fig:activity}a. We mark a particle as rearranging (Fig.~\ref{fig:readexp}a) if its $\|\Delta \vec r_\text{local} / a\|^2 \ge 0.025$ in any sample~\cite{suppmat}. To identify discrete subsystems, we use neighbor relationships (separation $< 1.5a$) to group these particles into contiguous rearranging clusters. Five clusters of interest are labeled A--E in Fig.~\ref{fig:readexp}a.
Figure~\ref{fig:activity}b shows global MSD of all particles (as in Fig.\ \ref{fig:read2}b) during training and readout, now computed 30 times per cycle, relative to two trained states: after the last application of $\gamma_1 = 3\%$ (TS$_3$, upper plot), and after the last application of $\gamma_2 = 4\%$ (TS$_4$, lower plot). However, we now also plot the MSD of the 5 labeled clusters only (closed symbols). This small fraction of the material is enough to qualitatively reproduce the global behavior. In Fig.~\ref{fig:activity}c we plot the MSD for each labeled cluster separately, relative to TS$_3$ (thin curves, diamonds) and TS$_4$ (thick curves, circles).
Figure~\ref{fig:activity}c shows that the global memory arises from local hysteresis and disorder. Because of disorder, each cluster rearranges at a different value of global strain, and so each cluster plays a different role in storing and reading memories. For instance, cluster ``B'' contributes strongly to the memory of TS$_3$: when driving amplitude is reduced at the start of readout, ``B'' gets stuck in a rearranged state relative to TS$_3$, and does not switch fully back to its original state until $\gamma_\text{read} \gtrsim 2.5\%$. In this way, ``B'' plays the same role that the highlighted hysterons of Fig.~\ref{fig:preisach}e did in the Preisach model.
Likewise, cluster ``C'' lets the material discriminate between $\gamma_\text{read} = 3\%$ and $\gamma_\text{read} = 4\%$. When the strain amplitude is reduced from 4\% to 3\%, cluster ``C'' stops switching states, and doesn't resume until $\gamma_\text{read} \ge 3.5\%$. Similarly, cluster ``A'' distinguishes $\gamma_\text{read} \le 4\%$, contributing to the readout of the 4\% memory. Cluster ``D'' distinguishes among values of $\gamma_\text{read}$, but it ends every cycle in the same state --- it is unused by our readout method. Finally, cluster ``E'' is nearly latent until $\gamma_\text{read} > 4\%$, and so reports the largest amplitude during training.
\subsection{Discussion}
By observing the motions of particles, and considering a simple example of RPM, we have shown how our material's memory arises from the hysteresis of individual rearranging clusters, each of which responds differently to global deformations. Hysteresis is responsible for the non-monotonic readout curves in Figs.~\ref{fig:readexp}c and \ref{fig:read2}b, and explains why this behavior is different from dilute suspensions (Fig.~\ref{fig:readexp}d), in which the steady state exhibits kinematic reversibility, not hysteresis.
Our results raise the question of how this behavior is connected with the physics of amorphous solids. A single rearranging cluster has hysteresis and is coupled to external shear stress, analogous to a hysteron in the Preisach model. However, it is also coupled to elastic deformations of the surrounding material~\cite{Eshelby:1957dg,Keim:2014hu}, so it may interact with nearby clusters, violating an assumption of the Preisach model. Indeed, when we measure the $\gamma^+_i$ and $\gamma^-_i$ (analogous to $H^+_i, H^-_i$), we find they depend on strain amplitude, presumably due to other, nearby rearrangements becoming active or inactive as the amplitude is varied (Fig.~\ref{fig:activity}b). RPM is proven to hold exactly only when interactions are ``ferromagnetic'' (each rearrangement encourages others) \cite{Sethna:1993ts}, but here we can also have ``antiferromagnetic'' interactions, depending on the relative positions of rearranging clusters~\cite{Eshelby:1957dg,Mungan2019}. Instead of the Preisach model, we can look to studies of disordered magnetic systems more generally, where despite complex, frustrated interactions, RPM may still hold at least approximately~\cite{deutsch04, Pierce:2005dd, Hovorka:2008ks}, especially in a steady state under cyclic driving~\cite{Hovorka:2008ks, Gilbert:2015db, Mungan:2019fh,Mungan2019}.
In the magnetic systems just discussed, disorder is quenched---the Hamiltonian prescribes couplings of a fixed population of subsystems to each other and an external field---facilitating the return to previous states. But disorder in deformed solids is generally \emph{not} quenched~\cite{Priezjev:2013hp, Keim:2014hu, Nagamanasa:2014jx, Regev:2015hs}. Instead, the transient at the beginning of each experiment remodels the material irreversibly, until we are left with a stable population of repeating rearrangements~\cite{Regev:2013es,Keim:2014hu,Regev:2015hs}. Remarkably, even as we subsequently reduce the strain amplitude and change the state of the system, this population largely persists~\cite{Mungan2019}. The few outlier trials we discard from our analysis~\cite{suppmat} may be exceptions.
While it exactly describes the behavior of only a few kinds of systems~\cite{Keim:2019aa}, here return-point memory is a generic prototype of how a rich global memory behavior can arise from disorder and local hysteresis (i.e. metastability). This suggests that the kind of memory discussed here might not only be present in the many kinds of amorphous solids, but could also be found or even engineered in many other types of systems~\cite{Keim:2019aa, Laurson:2012bc,Slotterback:2012fa, Ren:2013cp, Royer:2015dj,Dobroka:2017fc} given appropriate driving. Finally, by illuminating the mechanism for this behavior, our work points to a more precise question: why our material's self-organized steady states, despite frustration and marginal stability~\cite{Regev:2015hs,Mungan:2019aa,Mungan2019}, are so amenable as we vary driving to retrieve memories.
\begin{acknowledgments}
For illuminating discussions we thank Muhittin Mungan, Karin Dahmen, Sidney Nagel, Srikanth Sastry, Ajay Sood, and Joseph Paulsen. We also thank Paulo Arratia, David Gagnon, Larry Galloway, Luke Horowitz, Dani Medina, and Peter Nelson for their help. NCK thanks the Kavli Institute for Theoretical Physics for its hospitality, supported in part by NSF grant PHY-1748958. Minus K Technology donated the 100BM-1 vibration isolation platform used in experiments. This work was supported by NSF grant DMR-1708870, by an RSCA grant from California Polytechnic State University, San Luis Obispo, and by the William and Linda Frost Fund.
\end{acknowledgments}
|
\section{Introduction}
Fully homomorphic encryption is one of the great advances of modern cryptography. First discovered by Gentry in 2009 \cite{Gentry:2009}, it allows one to delegate the processing of \emph{encrypted} information to a party without access to the secret key. In classical computing, an enormous body of work has gone into developing and optimizing this protocol (see \cite{Peikert:2015} for a summary).
As the development of large scale quantum computers progresses, we must consider the cryptographic consequences of their arrival. While quantum computers can bolster the security of some cryptographic protocols \cite{Bennett:1984}, they can also obviate the security of others \cite{Shor:1995}. Fortunately, the security of existing homomorphic encryption schemes is derived from hard problems on lattices \cite{Brakerski:2013, Gentry:2013}, which are expected to be computationally infeasible for quantum computers to solve. Nonetheless, it is natural to ask: \vspace{.5 cm} \\ \emph{Can quantum computers allow for fully homomorphic encryption schemes which exhibit information-theoretic, rather than computational, security?} \vspace{.5cm}
For the strongest security definitions, we review why the answer is no while expanding on the limitations appearing in \cite{Newman:2017, Lai:2017, Yu:2014, Baumeler:2013}.
\section{Homomorphic Encryption}
\subsection{Classical homomorphic encryption}
A (classical) homomorphic encryption scheme is typically defined as an asymmetric key encryption scheme with an additional functionality, called evaluation. This functionality allows a third party, in possession of a ciphertext, to meaningfully manipulate the underlying plaintext without possessing the secret key. Formally, a homomorphic encryption scheme HE is a four-tuple of (randomized) algorithms.
\begin{itemize}
\item[] {\bf HE.KeyGen$(1^\kappa, 1^L)= (pk, sk, evk)$}. A key generation algorithm that accepts security parameter $\kappa$ and evaluation parameter $L$. It outputs the public key $pk$ and secret key $sk$ which depend on $\kappa$. Additionally, it outputs a third key known as the evaluation key $evk$, which depends on $L$. This key assists with the additional evaluation functionality.
\item[] {\bf HE.Enc$(m,pk) = c$}. An encryption algorithm that accepts a public key $pk$ and a single bit plaintext $m$, and then outputs a ciphertext $c$. By slight abuse of notation, we also allow $m$ to be a bit string, and assume the algorithm performs encryption bit-by-bit.
\item[] {\bf HE.Dec$(c,sk) = m$}. A decryption algorithm that accepts a single ciphertext $c$ and secret key $sk$ and outputs a single bit plaintext $m$. Again, we allow multi-ciphertext inputs and assume decryption occurs bitwise.
\item[] {\bf HE.Eval$(C, (c_1, \ldots, c_n), evk) = c'$}. An evaluation algorithm that accepts a circuit $C$ with $n$ input wires. It further accepts $n$ ciphertexts $(c_1, \ldots, c_n)$ and an evaluation key $evk$. It outputs a single new ciphertext $c'$.
\end{itemize}
The encryption scheme HE should satisfy the usual properties of any encryption scheme, but should further satisfy the following homomorphic property. For some circuits $C$ which we call the \emph{permissible functions} of the scheme, we have the following commutative diagram.
\begin{center}
\hspace{2cm}\begin{tikzcd}
\mathcal{M} \arrow[r, "\textbf{HE.Enc}{(\cdot,pk)}"] \arrow{d}[swap]{C} &[1.5cm] \mathcal{C} \arrow[d, "\textbf{HE.Eval}{(C, \cdot, evk)}"] \\[1.5cm]
\mathcal{M} & \mathcal{C} \arrow[l, "\textbf{HE.Dec}{(\cdot, sk)}"]
\end{tikzcd}
\end{center}
Here, $\mathcal{M}$ is the space of valid plaintexts (i.e. all binary strings), and $\mathcal{C}$ is the space of valid ciphertexts. One can think of the data processing as occurring from top-to-bottom, and the encryption as occurring from left-to-right. Plainly, a party Alice can perform the computation of $C$ herself, or outsource the computation by sending an encrypted message to a third party Bob.
We further call a homomorphic encryption scheme \emph{compact} if the complexity of HE.Dec is independent of the function being evaluated. This precludes trivial schemes in which Bob simply sends back a description of the circuit to be evaluated, and the true evaluation function is embedded into the decryption itself. Here and throughout, we will only consider compact schemes.
We call a homomorphic encryption scheme \emph{leveled fully homomorphic} if the set of permissible functions of the scheme is the set of all circuits up to some size specified by the evaluation parameter $L$. We simply call a scheme \emph{fully homomorphic} if the set of permissible functions is the set of all circuits, independent of $L$.
Typically, the ciphertexts in homomorphic encryption schemes experience an accumulation of noise that scales with the evaluated circuit depth. Eventually, this noise will prevent accurate decryption, and this motivates the definition of leveled fully homomorphic schemes. It is important to note that Alice's work may scale with the circuit to be evaluated, but this is realized implicitly as preprocessing in the key generation phase. A bootstrapping procedure introduced in \cite{Gentry:2009} allows for the indefinite refreshing of noisy ciphertexts, but requires a stronger circular security assumption.
Additionally, homomorphic encryption schemes sometimes allow for some small probability of failure. To simplify the discussion, we demand that schemes are perfectly correct, but note that we can extend all our arguments to the imperfect case with some extra notational baggage \cite{Baumeler:2013}.
\subsection{IT-secure quantum homomorphic encryption}
In \cite{Broadbent:2014}, the problem of extending homomorphic encryption to the \emph{quantum} setting was considered. Quantum homomorphic encryption accomplishes a similar task to classical homomorphic encryption, but with some key differences. Formally, we can model QHE as three families of quantum channels parametrized by the input size $n$ acting on four Hilbert spaces: $\mathcal{H}_M$ the message space, $\mathcal{H}_K$ the key space, $\mathcal{H}_C$ the ciphertext space, and $\mathcal{H}_E$ the evaluated ciphertext space (see Figure \ref{HE_diagram}).
\begin{figure}[htb!]
\includegraphics[width=\linewidth]{HE_diagram.pdf}
\caption{A diagram of QHE matching the generality of the model in \cite{Yu:2014}. Usually, we take $\mathcal{H}_C = \mathcal{H}_E$.}
\label{HE_diagram}
\end{figure}
Focusing on IT-secure encryption, we define a symmetric key QHE scheme. Formally, it consists of the following families of channels.
\begin{itemize}
\item[] {\bf QHE.Enc$_n:L(\mathcal{H}_M) \longrightarrow L(\mathcal{H}_K \otimes \mathcal{H}_C)$}. An encryption isometry that accepts an input state $\ket{\psi}$ and then outputs a key $\rho_k \in L(\mathcal{H}_K)$ and a ciphertext $\rho_{c} \in L(\mathcal{H}_C)$ which includes an appended evaluation key \footnote{Note that a scheme defined with a randomly generated classical key can be transformed into a QHE scheme of the prescribed form by purifying the key.}.
\item[] {\bf QHE.Dec$_n:L(\mathcal{H}_K \otimes \mathcal{H}_C) \longrightarrow L(\mathcal{H}_M)$}. A decryption channel that accepts a key $\rho_k \in L(\mathcal{H}_K)$ and evaluated ciphertext $\rho_{e} \in L(\mathcal{H}_E)$, and returns a plaintext state $\rho_m \in L(\mathcal{H}_M)$.
\item[] {\bf QHE.Eval$^U_n: L(\mathcal{H}_C) \longrightarrow L(\mathcal{H}_E)$}. An evaluation channel for a unitary circuit $U$ with $n$ wires that accepts a ciphertext state $\rho_c \in \mathcal{H}_C$ and outputs an evaluated ciphertext state $\rho_{e} \in \mathcal{H}_E$. Usually $\mathcal{H}_C = \mathcal{H}_E$, but this is not required.
\end{itemize}
Notice that the security parameter $\kappa$, which determines the length of secret key and encoding size, does not appear explicitly. Since quantum IT-secure encryption must obey similar key length bounds as classical IT-secure encryption \cite{lai2018generalized}, the key length must grow as a function of the input for fixed security guarantees. Consequently, for a particular QHE scheme, we define $\kappa$ implicitly as some fixed polynomial of $n$. Furthermore, encryption is performed all at once, rather than bit-by-bit.
For leveled schemes, we similarly define the evaluation parameter $L$ as another fixed polynomial of $n$, and assume that the evaluation key is appended to the ciphertext. This differs only slightly from the usual computationally-secure QHE setup \cite{Broadbent:2014}, but removes the need for explicit key generation.
The homomorphic property of the scheme is defined analogously to the classical setting. There, homomorphic evaluation is usually defined piece-by-piece for the gates comprising a circuit. For \emph{universal} gate sets, such as $\{$AND,OR$\}$ or $\{$NAND$\}$, homomorphic evaluation of these constituent gates can be built into (leveled) FHE schemes. For QHE schemes, the set of permissible functions is augmented by a richer set of constituent gates. The definitions for compact and homomorphicity carry over analogously to the quantum setting.
\section{QHE proposals}
\subsection{Computationally secure proposals}
Computationally secure QHE was first considered in \cite{Broadbent:2014}, along with an appropriate generalization of CPA security. The authors proposed three QHE schemes. The first was a scheme that could homomorphically implement the set of Clifford circuits, and followed directly from the quantum one-time pad. The second was a \emph{quasicompact} scheme: its decryption function scaled quadratically \emph{only} in the number of $T$-gates of the circuit. The final scheme required an evaluation key size scaling superexponentially in the $T$-depth of the circuit, and so was restricted to efficiently evaluating circuits of constant $T$-depth.
More recently, \cite{Dulek:2016} proposed a scheme built off of previous work on instantaneous nonlocal computation \cite{Speelman:2015}. The scheme is centered around an evaluation key consisting of $T$-gadgets, which allows for the homomorphic evaluation of one $T$-gate per $T$-gadget. This provides a leveled quantum fully homomorphic encryption scheme for polynomial-sized circuits, as these are the circuits for which the evaluation key can be generated efficiently. This was further extended to a verifiable scheme in \cite{Dulek:2017}.
Very recently, a scheme was proposed for performing leveled QFHE with a purely classical client using entirely different means \cite{Mahadev:2017}. This scheme is built on an encrypted controlled-not function, which allows a server to apply the control-not gate blindly, i.e. without knowing whether they had or not. The function requires a classical HE scheme with a particular list of desiderata. Whether or not this constitutes a non-leveled scheme depends on whether one accepts the circular security of existing schemes satisfying these requirements.
One common thread throughout all of these proposals is that each is built on a classical FHE scheme, and so inherits its underlying computational security. It is natural to ask if quantum mechanics might allow for an \emph{information-theoretically} secure delegation of computation on encrypted information. To this end, we might be encouraged by the construction of universal blind computation \cite{Broadbent:2008}, which allows delegated BQP quantum computation that guarantees information-theoretically secure hiding of \emph{both} the plaintext \emph{and} the computation, at the expense of interaction between Alice and Bob.
\subsection{Information-theoretically secure proposals}
There have been several works aimed towards homomorphic encryption with information-theoretic security guarantees. In \cite{Tan:2014}, a homomorphic encryption scheme based on bosonic encodings was proposed. This scheme used a weaker version of information-theoretic security by bounding the information accessible by the adversary. This allowed them to realize a fully unitary group of permissible functions, although these were not universal as the dimension of the group scaled \emph{polynomially} in the input size.
Later, \cite{Ouyang:2015} proposed a homomorphic encryption scheme based off of randomized quantum codes and transversal gates. The permissible functions for this scheme included all Clifford circuits augmented by a constant number of $T$-gates. The security guarantees for this scheme were stronger, providing exponential suppression on the trace distance between any two ciphertexts.
More recently, \cite{Lai:2017} detailed a scheme based on the quantum one-time pad to implement the class of IQP circuits homomorphically. Their scheme offers similarly strong information-theoretic security guarantees.
Finally, \cite{Tan:2017} proposed a homomorphic encryption scheme with a limited class of operations and modest information-theoretic security claims, but which may be implemented on current optical technologies. For a broader survey of securely delegated quantum computing, see \cite{Fitzsimons:2017}.
\section{Restrictions on information-theoretically secure proposals}
We now elaborate on certain no-go theorems which limit the capacity of homomorphic encryption schemes to exhibit meaningful information-theoretic security. It is well-known that in the classical setting, perfect information-theoretically secure homomorphic encryption is impossible \cite{Fillinger:2012}. This follows from communication bounds established for perfectly secure single-server private information retrieval \cite{Kushilevitz:1997}, and their relaxations \cite{Shubina:2007}.
The first restriction on QHE information-theoretic security was proven in \cite{Yu:2014}. There, they use a data localization argument via the no-programming theorem \cite{Nielsen:1997} to show the following.
\begin{Theorem}[Yu, Perez-Delgado, Fitzsimons] \label{first}
Suppose that a QHE scheme implements a set of unitary permissible functions $\mathcal{S}$, with precisely zero mutual information between the plaintext and ciphertext. Then, the size of the evaluated ciphertext must be of size at least $\log_2(|\mathcal{S}|)$.
\end{Theorem}
When $S$ is the set of all classical reversible functions, this becomes $$\log_2((2^n)!) = (n - \log_2(e))2^n + O(n).$$ Thus, in the case of \emph{perfect} information-theoretic security, any QHE scheme implementing even just \emph{classical reversible} unitaries must be highly inefficient. It is important to note that this data localization technique bounds the efficiency of perfect IT-secure QHE for \emph{any} set of unitary permissible functions.
In spite of this limitation, we have mentioned several QHE schemes \cite{Tan:2014,Ouyang:2015,Lai:2017} that implement large sets of permissible functions with strong, but imperfect, information-theoretic security guarantees. As the information localization argument of \cite{Yu:2014} relies integrally on perfect information-theoretic security, several authors \cite{Yu:2014,Ouyang:2015,Aaronson:2017} have asked whether an $\epsilon$-relaxation of the IT-security guarantee may allow for much larger sets of permissible functions. Theorem \ref{earlier_main} resolves this question in the negative.
\begin{Theorem} \label{earlier_main}
Suppose a QFHE scheme satisfies, for all ciphertexts $\rho,\rho'$, $$\|\rho - \rho'\|_{\text{Tr}} \leq \epsilon$$ for some $\epsilon < 1$. Then the communication complexity of the scheme must be exponential in $n$.
\end{Theorem}
This theorem appeared concurrently in the appendix of \cite{Newman:2017}, where it was used to prove bounds on transversal gates for quantum codes, and in \cite{Lai:2017}, which used similar generalizations of single-server private information retrieval bounds to the quantum setting \cite{Baumeler:2013}. At the heart of these arguments is an application of Nayak's bound \cite{Nayak:1999}, which places limitations on the compression of classical information into quantum information in the context of quantum random access codes. We now extend Theorem~\ref{earlier_main} to more closely generalize Theorem~\ref{first} to the imperfect security setting by considering a straightforward modification of Nayak's original bound.
\begin{Definition} \label{QRAC}
An $(n,m,p)$\emph{-quantum random access code (QRAC)} is a mapping of $n$ classical bits into $m$ qubits, \emph{[$b \mapsto \rho_b$]}, along with a set of POVM's $\{M_i^0, M_i^1\}_{i=1}^n$ satisfying, for all $b\in \{0,1\}^n$ and $i\in[n]$, $$ Tr(M_i^{b_i}\rho_b) \geq p.$$
\end{Definition}
Informally, this is simply a compression of classical information into quantum information that allows for a local recovery with some probability of success. Nayak's bound then places a fundamental limitation on the recoverability of this compression \cite{Nayak:1999}. For Theorem \ref{main}, we will require a small generalization of the bound for QRACs encoding subsets of strings of length $n$.
\begin{Lemma} \label{main_lemma}
Let $F \subseteq \{0,1\}^n$. For each $x \in F$, define a classical-to-quantum encoding $x \mapsto \rho_x$, where each $\rho_x$ is a state on $m$ qubits. Suppose there exists a set of POVM's $\{M_i^0, M_i^1\}_{i=1}^n$ satisfying, for all $x \in F$ and $i\in[n]$, $$ Tr(M_i^{x_i}\rho_x) \geq p.$$
Let $H(\cdot)$ denote the binary entropy function. Then, $$m \geq \log|F| - n\cdot H(p).$$
\end{Lemma}
\begin{proof}
The proof follows from the original proof of Nayak's bound, but with the uniform distribution on $F$ rather than $\{0,1\}^n$.
For any $x \in \{0,1\}^*$, define $$F_x = \{y \in F| y \text{ is prefixed by } x\},$$ and let $\sigma_x = \frac{1}{|F_x|}\sum_{y \in F_x} \rho_y$. For $0 \leq k < n$ and for any $x \in \{0,1\}^k$ such that $|F_x| > 0$, we then have $$\sigma_x = \frac{|F_{x0}|}{|F_x|}\sigma_{x0} + \frac{|F_{x1}|}{|F_{x}|} \sigma_{x1}.$$ Let $X^x$ be a random variable drawn from the uniform distribution on $F$ conditioned on the first $k$ bits being $x$, and let $Y^x$ be a binary random variable obtained from a measurement of $\sigma_x$. Then by Holevo's theorem \cite{holevo1973bounds}, $$S(\rho_x) \geq \frac{|F_{x0}|}{|F_x|}S(\sigma_{x0}) + \frac{|F_{x1}|}{|F_{x}|} S(\sigma_{x1}) + I(X^x:Y^x).$$ If the probability of failure of correctly identifying a sample $\sigma_{x0}$ or $\sigma_{x1}$ conditioned on knowing the prefix $x$ is at most $1-p$, then we can apply Fano's inequality \cite{cover2012elements}, $$I(X^x:Y^x) \geq H\left(\frac{|F_{x0}|}{|F_{x}|}\right) - H(p).$$ We can then expand the bound for $S(\sigma)$ out into singletons $\rho_y$ so that, \begin{align*} S(\sigma) &\geq \frac{1}{|F|}\sum\limits_{y \in F}S(\rho_y) + \sum\limits_{k=0}^{n-1} \sum\limits_{x \in \{0,1\}^k} \frac{|F_x|}{|F|} \left(H\left(\frac{|F_{x0}|}{|F_x|}\right) - H(p) \right) \\ &\geq \sum\limits_{k=0}^{n-1} \sum\limits_{x \in \{0,1\}^k} \frac{|F_x|}{|F|}H\left(\frac{|F_{x0}|}{|F_x|}\right) - n\cdot H(p) \\ &\geq \sum\limits_{k=0}^{n-1} H(X_{k+1}|X_{1,\ldots,k}) - n\cdot H(p) \\ &\geq H(X) - n \cdot H(p) \end{align*} where the final line follows from the chain rule on conditional entropy. Since $X$ is the uniform distribution on $|F|$ elements and $S(\sigma)$ is at most $m$, the result follows.
\end{proof}
With this lemma in mind, we can extend the definition of a QRAC to include encodings of subsets of strings, and refer to such encodings as $(F,n,m,p)$-QRACs, where $F$ is a subset of $\{0,1\}^n$. Applying the lemma with $F = \{0,1\}^n$, we obtain Nayak's bound \cite{Nayak:1999}.
\begin{Theorem}[Nayak] \label{Nayak}
Any $(n,m,p)$-quantum random access code must satisfy $$m \geq n (1-H(p)).$$
\end{Theorem}
We now have the tools required to prove the extended imperfect IT-secure QFHE no-go theorem. The central idea is to extract a quantum random access code from an IT-secure QHE scheme, and then apply Lemma \ref{main_lemma} to lower bound the communication complexity.
\begin{Theorem} \label{main}
Suppose we have a QHE scheme whose permissible functions can be used to evaluate a set $F_n$ of Boolean functions on $n$ bits. Suppose further that, for any two ciphertexts $\rho, \rho'$ encrypting messages of $n$ bits, we have the $\epsilon$-IT security guarantee $$\| \rho - \rho' \|_{\text{Tr}} \leq \epsilon(n).$$ Then, the communication complexity of the protocol is at least $\log|F_n| - 2^n\cdot H(\epsilon(n)).$
\end{Theorem}
\begin{proof}
Throughout, we will use subscripts to denote the subsystem(s) on which an operator acts. Fix any $n \geq 1$.
For any $x \in \{0,1\}^n$, define the state $\rho_{KM}^x = \text{QHE.Enc}_n(\ket{x})$. This is pure as QHE.Enc is an isometry. Furthermore, because $\rho_M^x$ purifies to a state of dimension at most $2\dim(\mathcal{H}_M)$, we may assume that $\dim(\mathcal{H}_K) \leq \dim(\mathcal{H}_M)$.
By our IT-security guarantee, we have that for all $x,x'$, $$\|\rho_M^x - \rho_M^{x'}\|_{\text{\it Tr}} \leq \epsilon(n).$$ By the Schmidt decomposition, there must exist a unitary $U_K$ so that $$\|U_K \rho_{KM}^x U_K^\dag- \rho_{KM}^{x'}\|_{\text{\it Tr}} \leq \epsilon(n).$$ Fix any base point $x'$ and define $U_K^x$ to be an operator that satisfies the above equation.
Let $f:\{0,1\}^n \rightarrow \{0,1\}$ be any Boolean function in $F_n$. To be explicit, we may assume $f$ is a classical reversible permissible function restricted to the first output wire, without loss of generality. Consider $f$ as a length $2^n$ bit string with bit $i = f(i_{\text{binary}})$. Then we can define the classical-to-quantum encoding $f \mapsto \rho_f$ according to $$\rho_f = \text{QHE.Eval}_n^{U_f}(\rho^{x'}_{KM})$$ where $U_f$ is the unitary representing the full reversible circuit for $f$.
Furthermore, for any $x \in \{0,1\}^n$, we may define the measurement channel $$M_x(\cdot) = M_{Z_1} \circ \text{QHE.Dec}_n \circ U_K^x(\cdot)U_K^{x\dag}$$ where $M_{Z_1}$ is a $Z$-basis measurement on the first qubit.
We claim that this forms a $(S,2^n,\dim(\mathcal{H}_C) + \dim(\mathcal{H}_E), 1 - \epsilon(n))$-QRAC. To see this, note that \begin{align*} M_x(\rho_f) &= M_{Z_1} \circ \text{QHE.Dec}_n \circ U_K^x(\text{QHE.Eval}_n^{U_f}(\rho^{x'}_{KM}))U_K^{x\dag} \\ &= M_{Z_1} \circ \text{QHE.Dec}_n \circ (\text{QHE.Eval}_n^{U_f}(U_K^x\rho^{x'}_{KM}U_K^{x\dag}) \\ &\approx_{\epsilon(n)} M_{Z_1} \circ \text{QHE.Dec}_n \circ (\text{QHE.Eval}_n^{U_f}(\rho^{x}_{KM})) \\ &\approx_{\epsilon(n)} f(x),\end{align*} where $\approx_{\epsilon(n)}$ denotes $\epsilon(n)$ proximity in terms of trace distance.
In the above, the second line follows from the fact that evaluation is performed on the ciphertext alone, and so commutes with the keyspace unitary. The third line follows from the IT-security guarantee and contractivity of trace distance. The final line is exactly the homomorphic property of the scheme for $f \in F_n$.
Thus, the probability of failure of the resulting QRAC is at most $\epsilon(n)$. Recalling the assumption $\dim(\mathcal{H}_K) \leq \dim(\mathcal{H}_C)$, this gives the parameters of the claimed QRAC. Applying Lemma \ref{main_lemma}, we see that $\dim(\mathcal{H}_C) + \dim(\mathcal{H}_E) \geq \log|S| - 2^n\cdot H(\epsilon(n))$. Since $\dim(\mathcal{H}_C) + \dim(\mathcal{H}_E)$ is precisely the communication complexity of the protocol as a whole, the Theorem follows.
\end{proof}
It is worth noting that in the theorem, we only require encryptions of classical bits to be secure, since we are only concerned with classical functions. Even in the case of hiding a \emph{constant} fraction of information, setting $F_n$ to be the set of all Boolean functions on $n$ bits, we obtain Theorem \ref{earlier_main} as a corollary.
\begin{proof}[Proof of Theorem~\ref{earlier_main}]
Because a QFHE scheme can implement the full set of Boolean functions, Theorem \ref{main} gives us that the communication complexity is lower bounded by $2^n(1 - H(\epsilon))$, which is exponential in $n$ whenever $\epsilon < 1$.
\end{proof}
More generally, one might expect that for most information-theoretically secure schemes, realizing security on the order of $2^{-poly(n)}$ could be augmented to security on the order of $O(2^{-1.01n})$ up to a polynomial change in the function defining the security parameter. For such a scheme, $2^nH(\epsilon(n)) = o(1)$, giving us the following corollary.
\begin{Corollary}
Suppose we have a QHE scheme whose permissible functions can be used to evaluate a set $F_n$ of $n$-bit Boolean functions. Further suppose that for any pair of $n$-qubit ciphertexts $\rho,\rho'$, it satisfies $$\|\rho - \rho'\|_{\text{Tr}} \leq \epsilon(n)$$ for some $\epsilon(n) = O(2^{-1.01n})$. Then, its communication complexity is lower bounded by $\log|S| - o(1)$.
\end{Corollary}
Note that an identical argument will still give a $\theta(\log|F_n|)$ lower bound even if $\epsilon(n) = \theta(2^{-n})$, assuming $|F_n|$ grows super-exponentially. However, we reiterate that we would expect such schemes to be equivalent up to polynomially related security parameters.
In comparing Theorem \ref{main} to Theorem \ref{first}, we see that Theorem \ref{main} gives a nice generalization that matches the dependency on the number of classical permissible functions. Most importantly, by reducing to lower bounds in quantum random access codes, \emph{it is robust to imperfect IT-security}. The mild tradeoff for this reduction is that the lower bound is restricted to Boolean functions, rather than general unitaries.
\section{Discussion}
One thing that is of vital importance to notice: \emph{neither of these bounds rule out the possibility of IT-secure \emph{leveled} QFHE}. Because the evaluation parameter is chosen as a fixed polynomial of the input, leveled schemes only carry the promise of implementing circuits whose size is some polynomial in the input. A coarse counting argument shows that the number of such circuits can grow at most exponentially in the input size. These information-theoretic approaches are ill-equipped to give a super-polynomial lower bound, simply because the class of permissible functions is too small.
To this end, we might turn to complexity theory as a means of proving no-go results. Aaronson et al. have given complexity-theoretic evidence that perfectly-secure QFHE is impossible if it takes the natural but restricted form of an offline one-round generalized quantum encryption scheme \cite{Aaronson:2017}. Even so, we re-emphasize that this relaxation to imperfect security is non-trivial: for example \cite{Aaronson:2017} shows that the delegation of blind quantum computation by classical clients is highly unlikely, but there already exist such schemes that leak a portion of the data \cite{mantri2017flow}.
Although we have required perfect correctness throughout for simplicity, the proofs go through using generalizations of Nayak's bound in the QPIR setting that account for imperfect retrievals \cite{Baumeler:2013}. Furthermore, it would be nice to generalize the lower bound to account for unitary functions to fully generalize the result of \cite{Yu:2014}.
As with any no-go result, the most interesting question is how one might plausibly circumvent it. In this direction, there are at least two avenues to consider.
The first is to lessen the stringency of the security even further. We've seen that, when using a weaker security guarantee in terms of accessible information, we can implement a full \emph{continuum} of permissible functions \cite{Tan:2014}. Could something similar be made universal, and if not, how strong can this computational model be?
The second avenue to consider is to limit the number of implementable functions. One could certainly restrict oneself to exponentially large circuit classes, as several existing IT-secure QHE schemes do. Perhaps more interestingly, could one construct an efficient leveled QHE scheme with meaningful IT-security guarantees, and if so, what would such a scheme look like? In existing IT-secure QHE schemes, for a fixed key size, information about the underlying plaintext is leaked as you encrypt more bits, but not as you apply more gates. It seems that a leveled scheme would need to apply gates in such a way that the gates themselves leak information about the underlying plaintext; otherwise, the scheme might as well be non-leveled. This idea is not without precedence, since evaluation key gadgets that both carry information about the secret key and apply gates homomorphically are a cornerstone of computationally-secure QFHE \cite{Dulek:2016}.
Still, we reiterate that there is evidence that this cannot be done in the perfect-security setting. In the direction of proving impossibility, perhaps further complexity-theoretic bounds on the set of permissible functions can be applied when restricting to a single-round of interaction \cite{Aaronson:2017}.
\section{Acknowledgements}
The author would like to thank Florian Speelman and Eric Sabo for comments on an earlier draft, as well as Cupjin Huang for pointing out Lemma~\ref{main_lemma}. This research was supported in part by NSF grant 1717523. \\ {\it $^*$Email address:} <EMAIL>.
|
\section{Introduction}
\emph{Introduction.}---Recent technological advances in manipulating the radiation-matter interaction, especially at the level of a few atoms and photons, have allowed great strides in the implementation of quantum information processing protocols. The mastery of such interactions on various platforms \cite{Haroche2013,Reiserer2015,Wineland2013,Hennessy2007,Devoret2013} has made it possible to dispel mistrust regarding the possibility of a quantum computer becoming real \cite{Haroche1996}. Thus, one often sees new (or improved) quantum computing protocols being implemented on diverse setups \cite{Palmero2017,Nguyen2018,Welte2018,Zajac2017,Axline2018,Boixo2018,wendin2017}.
The most elementary interaction between atom and radiation was first described by Rabi in 1936 \cite{Rabi1936}, considering the radiation as a classical field and a dipolar interaction.
In its quantum version, it is possible to classify such a model into different regimes \cite{Rossatto2017,FDiaz2018,Kocku2018}. When the interaction energy is a small perturbation on the free energies, the quantum Rabi model can be reduced to the well-known Jaynes-Cummings (JC) model \cite{Jaynes1963} through the rotating-wave approximation, which describes the usual coherent exchange of a quantum of energy between a two-level atom and a single-mode bosonic field.
The JC model can be extended to nonlinear versions (nondipolar light-matter interaction) \cite{Shore1993}, in particular the multiphoton JC model \cite{Sukumar1981,Singh1982} that considers multiphoton exchange. Such a nonlinear Hamiltonian appears, at least for two-photon interaction, in trapped-ion domain \cite{Vogel1995,Meekhof1996}, optical \cite{Gau1992} and microwave \cite{brune1997} cavities, or even in superconducting circuits \cite{neil2015,Garziano2015,Bertet2005,Felicetti2018,Feli2018}. In Refs.~\cite{Feli2018,Felicetti2018}, a much more interesting scenario is presented since two-photon interactions can be implemented with substantial coupling strengths, differently from perturbative higher-order effects of a dipolar interaction that allow small effective coupling strengths \cite{Vogel1995,Meekhof1996,Garziano2015,Piazza2012,Hamsen2017}. Furthermore, new physical consequences emerge when the two-photon interaction becomes stronger and stronger \cite{Duan2016,Garbe2017,Chen2018}. The feasibility of achieving nonperturbative two-photon interactions and the possibility to broaden it for general multiphoton processes stimulate a pursuit of novel physical phenomena and applications to such nondipolar interactions, such as new ways to manipulate quantum information.
In this Letter we investigate the multiphoton JC Hamiltonian ($N$-photon interaction) and how it can be employed in quantum information science. Given the system in the strong coupling regime, we show how to perform arbitrary rotations in Fock space involving vacuum and $M$-photon states ($M<N$), and how to implement a quantum device that allows the transmission of a field only in a finite superposition (or mixture) of Fock states (quantum scissor \cite{Leoski2011,Hoi2012,Peropadre2013}). In particular, the two-photon JC interaction can be used for single-photon generation in a different fashion as compared to protocols involving standard JC interactions \cite{Lindkvist2014,Peng2016}. Additionally, we show that the system can also provide a rich physics in the weak coupling regime, in particular multiphoton electromagnetically induced transparencylike phenomena; i.e., the multiphoton absorption of the bosonic mode can be canceled out due to a quantum destructive interference, inducing a narrow (controllable) reflectivity window for nonlinear probe fields.
\emph{Model.}---We consider the $N$-photon JC model \cite{Sukumar1981,Singh1982} ($\hbar=1$),
%
\begin{equation}
H_{0}=\omega a^{\dagger}a+\omega_{0}\frac{\sigma_{z}}{2}+g(\sigma_{+}a^{N}+ {\rm{H.c.}}),\label{hamiltonian}
\end{equation}
%
with $\omega$ and $a$ ($a^{\dagger}$) being the frequency and the annihilation (creation) operator of the single-mode bosonic field. The atomic frequency transition between the ground $\ket{g}$ and excited $\ket{e}$ states is $\omega_{0}$, while $\sigma_{+}=\left(\sigma_{-}\right)^{\dagger}=\left|e\right\rangle \left\langle g \right|$ and $\sigma_{z}=\sigma_{+}\sigma_{-}-\sigma_{-}\sigma_{+}$. The atom-field coupling is $g$ and ${\rm{H.c.}}$ stands for Hermitian conjugate. This Hamiltonian describes a coherent exchange of $N$ excitations of the bosonic mode with a two-level atom. In Fig.~\ref{fig1}(a) we illustrate the energy-level diagram for a resonant interaction ($\omega_{0}=N\omega$). The lowest eigenstates are uncorrelated, involving the atomic ground state and up to $N-1$ excitations in the bosonic mode,
%
\begin{equation}
\left|\Psi_{g,n}\right\rangle =\left|g,n\right\rangle \,\,\text{for}\,\,\, 0 \le n < N,\label{groundstates}
\end{equation}
%
with eigenenergies $E_{g,n}=(n - \tfrac{N}{2} )\omega$. The remaining (dressed) eigenstates are
%
\begin{eqnarray}
\left|\pm,n\right\rangle =\frac{1}{\sqrt{2}}\left(\left|g,n\right\rangle \pm\left|e,n-N\right\rangle \right) \,\,\text{for}\,\,\, n \ge N,
\label{excitedstates}
\end{eqnarray}
%
with eigenenergies $E_{\pm,n}=(n - \tfrac{N}{2} )\omega\pm g\sqrt{\tfrac{n!}{(n-N)!}}$. Therefore, the system exhibits a finite-dimensional harmonic-oscillator spectrum \cite{Miranowicz2001} for the lowest eigenstates and then followed by JC-like doublets, as depicted in Fig.~\ref{fig1}(a). For the sake of illustration, we hereafter consider the single-mode bosonic field as a mode of an optical cavity, as sketched in Fig.~\ref{fig1}(b).
\begin{figure}[b]
\includegraphics[trim = 0mm 0mm 0mm 0mm, clip, width=0.48\textwidth]{Fig1.pdf}
\caption{(a) Energy spectrum of the $N$-photon Jaynes-Cummings model. The system has $N$ equidistant uncorrelated eigenstates, $\left|g,n\right\rangle $ for $0\le n \le N-1$, and dressed eigenstates, $\left|\pm,n\right\rangle =\frac{1}{\sqrt{2}}\left(\left|g,n\right\rangle \pm\left|e,n-N\right\rangle \right)$ for $n \ge N$ (Jaynes-Cummings-like doublets). The transition $\ket{g,0} \to \ket{\pm,N}$ can be induced by an $N$-photon driving field on the cavity mode, while the adjacent uncorrelated states can be addressed via linear driving fields (single-photon transitions). (b) Pictorial representation of the driven atom-cavity system. Giving an input field, the output field will be limited to the ($N-1$)-photon state.}
\label{fig1}
\end{figure}
The dressed states can be directly excited from the ground state via a nonlinear driving field \cite{brune1997,neil2015,Andersen2016,Mallet2011} on the cavity mode. To this end, we assume a driving field of strength $\varepsilon(t)$, single-photon frequency $\omega_{p}$, phase $\chi$, and nonlinearity $M$, which is described by $H_{P}=\varepsilon(t)[a^{M}e^{i(M\omega_{p}t-\chi)}+{\rm{H.c.}}]$. Thus, the total Hamiltonian, written in a quasi-time-independent frame, reads
%
\begin{equation}
H_{I}=\Delta_{p}\left(a^{\dagger}a+N\frac{\sigma_{z}}{2}\right)+[\varepsilon(t)a^{M}e^{-i\chi}+ga^{N}\sigma_{+}+{\rm{H.c.}}],\label{eq:interactionhamiltonian}
\end{equation}
%
with $\Delta_{p}=\omega-\omega_{p}$. Finally, considering the cavity and the two-level atom coupled to their reservoir in the white-noise limit (Born, rotating-wave and Markov approximations in the system-reservoir interactions) \cite{gardiner2004,Petruccione2007}, the system dynamics at $T=0$K is governed by the master equation
%
\begin{equation}
\begin{split}\frac{d\rho}{dt}=& -i[H_{I},\rho]+\gamma(2\sigma_{-}\rho\sigma_{+}-\sigma_{+}\sigma_{-}\rho-\rho\sigma_{+}\sigma_{-})\\
& +\kappa(2a\rho a^{\dagger}-a^{\dagger}a\rho-\rho a^{\dagger}a)+\gamma_{\phi}(\sigma_{z}\rho\sigma_{z}-\rho),
\end{split}
\label{eq:masterequation}
\end{equation}
%
in which $\gamma$ and $\gamma_{\phi}$ are the atomic polarization decay and dephasing rates, respectively, while $\kappa$ is the decay rate of the cavity field amplitude. Moreover, this master equation is valid whenever we are out of the ultrastrong and deep-strong atom-cavity coupling regimes ($g\ll\omega,\omega_{0}$), and for small excitation numbers \cite{Blais2011}. We can numerically solve Eq.~\eqref{eq:masterequation} by truncating the Fock space of the cavity mode according to $\varepsilon$ and $M$ \cite{Johansson2012}.
\emph{Strong coupling regime: Arbitrary rotation in Fock space and quantum scissor.}---First we analyze the system dynamics in the strong coupling regime; i.e., when the coherent atom-field interaction exceeds all relaxation processes ($g>\kappa,\gamma,\gamma_{\phi}$).
Adjusting the driving field resonantly to the cavity mode frequency ($\omega_{p}=\omega$) and assuming the system in its ground state initially, two cases appear if the driving field is considered as a weak probe field [$\varepsilon_m \equiv \max{(| \varepsilon(t) |)} \ll g$]. (i) When $M>N$, the dressed states of the system are not populated since the probe field is out of resonance with the transitions $\ket{g,0} \to \ket{\pm,M}$, also for $\ket{\pm,M} \to \ket{\pm,2M}$ and so on. Consequently, the system remains in the ground state while the driving field is completely reflected by the cavity mirror. (ii) When $M<N$, a more interesting situation takes place, in which the probe field is able to induce the system to the uncorrelated states only. For instance, for $N=2$, the most excited uncorrelated state is $|g,1\rangle$, which can be populated with a linear probe field ($M=1$). Nevertheless, the two-photon state (and consequently the higher ones) is not populated since the transitions $|g,1\rangle \to |\pm,2\rangle$ are not resonant with this linear probe field, thus prohibiting the injection of more than one photon into the cavity mode (single-photon blockade \cite{[{For a recent review, see Section 6.6 of }][{, and references therein.}]Gu2017}). Hence, the system dynamics is restricted to the Fock subspace $\{|0\rangle,|1\rangle\}$, even if we take into account stronger driving fields, but ensuring that $g$ be strong accordingly. For different $N$ and $M$ ($<N$) similar situations appear, in which the system exhibits multiphoton blockade phenomena \cite{Gu2017}.
Considering the last case, let us derive an effective Hamiltonian in the limit of atom-field coupling much stronger than the driving field strength, i.e., when $g \gg \varepsilon_m$ (weak probe field). To this end, we rewrite $H_{P}$ (for $M<N$) in the eigenbasis of $H_{0}$. Then, adjusting $\omega_{p}=\omega$ and performing a rotating-wave approximation \cite{SM}, we have
%
\begin{equation}
H_{\rm{eff}} \simeq \varepsilon(t) e^{i\chi} \sum_{\substack{n=0 \\ (M<N)}}^{N-M-1}\sqrt{\frac{(n+M)!}{n !}}\left|g,n+M\right\rangle \langle g,n|+{\rm H.c.},\label{effective_hamiltonian}
\end{equation}
%
which is valid for $\varepsilon_m \ll g \sqrt{2\,(N-M)!}$ \cite{SM}. When $\frac{N}{2} \le M <N$, $H_{\rm{eff}}$ promotes an arbitrary rotation between the vacuum and the $M$-Fock states; i.e., $\ket{\Psi(t)} = e^{-i\int_{0}^{t}H_{\rm{eff}}(t^{\prime})dt^{\prime}}\ket{g,0}$ reads
%
\begin{equation}
\ket{\Psi(t)} =\ket{g} \left[\cos{\left(\frac{\theta_M(t)}{2}\right)\ket{0}}+e^{i\varphi}\sin{\left(\frac{\theta_M(t)}{2}\right)\ket{M}}\right],
\label{psi_rotating}
\end{equation}
%
with $\theta_M(t)=2\sqrt{M!}\int_{0}^{t} \varepsilon(t^\prime)dt^\prime$ and $\varphi=\chi - \frac{\pi}{2}$ (polar and azimuthal angles of the Bloch sphere, respectively). As the atom remains in its ground state throughout the dynamics, this rotation gate is immune to atomic relaxation processes. On the other hand, the cavity dissipation introduces errors to the gate because it destroys the quantum superposition of Eq.~\eqref{psi_rotating} and also leads the cavity to a final state outside the desired Hilbert space \{$\ket{0},\ket{M}$\}. In this sense, the greater the ratio $\varepsilon_m/\kappa$ the higher the gate fidelity, since in this case the rotation tends to occur before the cavity dissipation appreciably disturbs the unitary evolution. Such detrimental influence on the rotation gate can be seen in Fig.~\ref{Figure_rotations}, in which we illustrate the time evolution of the Fock state probabilities for $N=2$ and $M=1$ (upper panels) and for $N=3$ and $M=2$ (lower panels).
Therefore, we have shown a straightforward scheme to perform arbitrary rotations in Fock space (involving vacuum and $M$-Fock states), which could be implemented at least for $N=2$ with the current technology \cite{Feli2018,Felicetti2018}, moving toward a suitable route to unitary gates for Fock state qubits. We must remember that controlled qubit operations in Fock space are not a simple task \cite{Nielsen,Santos2005,Miled2014,Prado2014,Rosado2015,Krastanov2015,Heeres2015,Hofheinz2008,Liu2004,Law1996}, often demanding multistep processes and being sensitive to the dissipation of the atomic ancilla.
A linear driving field is able to inject up to $N-1$ excitations into the cavity mode. For instance, for $N=3$, $M=1$, and $\varepsilon_m \ll 2g$, we have $H_{\rm eff}\simeq\varepsilon(t)e^{i\chi} \left|g\right\rangle \langle g|(\left|1\right\rangle \langle 0|+\sqrt{2} \left|2\right\rangle \langle 1|)+{\rm H.c.}$, which indicates that the cavity mode can only be populated in the vacuum, one-, and two-photon states. Thus, our system also works out as a quantum scissor, generating a finite-dimensional quantum-optical state inside the cavity, but without requiring a nonlinear Kerr medium inside it as usually adopted in the literature for this purpose \cite{Leoski2011}. Our case has the advantage that the multiphoton interaction can achieve strong strengths as shown in Refs.~\cite{Feli2018,Felicetti2018}.
\begin{figure}[t]
\includegraphics[trim = 6mm 2mm 1mm 1mm, clip, width=0.48\textwidth]{rotation.pdf}
\caption{Rotations in Fock space involving the vacuum and $M$-Fock states. We plot the Fock state probabilities $P_{k}(t)={\rm{Tr}}[|k\rangle\langle k|\rho(t)]$ as a function of $\kappa t$ for $\gamma=\gamma_\phi = \kappa/2$ and $\chi=0$. In the upper panels $N=2$ and $M=1$, assuming (a) $g=1000\kappa$ and $\varepsilon(t)=\varepsilon_{0}=100\kappa$, and (b) $g=100\kappa$ and $\varepsilon_{0}=10\kappa$. We notice that the rotation fidelity can be high when $\varepsilon_{0} \gg \kappa$, but it decreases when $\varepsilon_{0} \gtrsim \kappa$ (cavity dissipation destroys the quantum superposition). In the lower panels $N=3$ and $M=2$, with (c) and (d) given in terms of the parameters of (a) and (b), respectively. For this configuration, we can see that the cavity dissipation can also lead the cavity to a final state outside the desired Hilbert space \{$\ket{0},\ket{2}$\}, since the Fock state $\ket{1}$ is populated through the decay channel $\ket{2} \to \ket{1}$. The symbols correspond to the approximate solution [Eq.~\eqref{psi_rotating}].}
\label{Figure_rotations}
\end{figure}
From the input-output theory, the field transmitted by a cavity mirror, opposite to the driven one, is exactly the intracavity field multiplied by the square root of the decay rate relative to this mirror \cite{walls2007quantum}. Then, from this perspective, our system also acts as a quantum-optical intensity limiter or Fock state filter; i.e., driving the cavity with an input field comprising any superposition (or mixture) of Fock states, the output field will be a finite superposition (or mixture) limited to the ($N-1$)-photon state, restricting the maximum intensity of the transmitted field [see Fig.~\ref{fig1}(b)]. If one wants to ensure that the field be almost entirely transmitted by a specific mirror, it is convenient to employ an asymmetric cavity \cite{Mcke2010}.
\begin{figure}[t]
\includegraphics[trim = 6mm 8mm 1mm 4mm, clip, width=0.48\textwidth]{filter.pdf}
\caption{Fock state filter. In the left-hand panels we plot the Fock state probabilities $P_{k}(t)={\rm Tr}[|k\rangle\langle k|\rho(t)]$ as a function of $\kappa (t-t_0)$ assuming a Gaussian probe pulse ($M=1$, $\varepsilon_{m}=4\kappa$ and $\eta=\sqrt{2}\kappa^{-1}$) and $\gamma=\gamma_{\phi}=\kappa/2$ for (a) $g=0$ and $g=20\kappa$ [(b) $N=2$ and (c) $N = 3$]. In the right-hand panels we plot $P_{k}$ in the stationary regime as a function of $g/\kappa$ assuming a continuous driving field with $\varepsilon(t)=\varepsilon_{0}=2\kappa$ for (d) $N=2$, (e) $N=3$, and (f) $N=4$.}
\label{Figure_filter}
\end{figure}
As mentioned above, the determination of the density matrix of the intracavity field is enough to derive the properties of the output field. Thus, let us first consider the driving field as a Gaussian pulse with amplitude $\varepsilon(t)=(\varepsilon_{m}/\sqrt{2\pi\eta^{2}})e^{-(t-t_{0})^{2}/2\eta^{2}}$, in which $\varepsilon_{m}$ and $\eta$ are the maximum amplitude and duration of the driving pulse, respectively, and $t_{0}$ is the time when its maximum arrives at the cavity mirror. For $g=0$ (driven empty cavity), we have non-null probabilities ($P_k$) of finding $k$ excitations in the cavity mode, as we can see in Fig.~\ref{Figure_filter}(a) considering $\varepsilon_{m}=4\kappa$ and $\eta=\sqrt{2}\kappa^{-1}$. However, depending on $N$, only some Fock states are populated when $g \neq 0$. For $N=2$ ($3$), the highest cavity-mode state populated is $|1\rangle$ ($|2 \rangle$) [Figs.~\ref{Figure_filter}(b) and \ref{Figure_filter}(c)], such that we can have a single-photon source for $N=2$. For weak atom-field coupling and intense driving fields, this quantum filter does not work out since the driving field is able to introduce more excitations in the cavity mode, as observed in Figs.~\ref{Figure_filter}(d) and \ref{Figure_filter}(f) for $g \lesssim \kappa , \varepsilon_{m}[2(N-M)!]^{-1/2}$. These figures show $P_k$ as a function of $g/\kappa$ in the steady state assuming a continuous driving field. We observe that $P_{k \ge N} \to 0$ as the system reaches stronger couplings ($g \gtrsim 10\kappa$), which elucidates the truncation of the Fock space accessible to the cavity mode. It is worth remarking that the Fock state filter yields similar results for nonlinear driving fields ($1<M<N$ with $N \ge 2$).
The two-photon JC Hamiltonian could be implemented via dispersive dipolar interaction involving a three-level superconducting artificial atom, similar to the scheme performed in Ref.~\cite{Wang2016}, where an effective coupling of the order of $1$ MHz ($\sim 10^3 \kappa$) was achieved, which would be sufficient to observe single-photon filter and rotations involving $\ket{0}$ and $\ket{1}$, as seen in Fig.~\ref{Figure_rotations}(a) [a complete rotation is achieved in a time ($\sim 30$ $\mu$s) 2 orders of magnitude shorter than the single-photon cavity lifetime, $1/\kappa \sim 1$ ms].
\emph{Weak coupling regime: Multiphoton-induced-reflectivity phenomenon.}---A weak driving field, of constant strength $\varepsilon(t) = \varepsilon_{0} \ll g$, frequency $\omega_{p}$, and the same nonlinearity of the JC Hamiltonian ($M=N)$, will be able to introduce $N$ excitations to the system if it is close to resonance with the transitions $\ket{g,0} \leftrightarrow \left|\pm,N\right\rangle $, i.e., if $N\omega_{p}\approx\omega N\pm g\sqrt{N!}$ (for $\omega_{0}=N\omega$). Nonetheless, if the decay rates of the first dressed states are large enough when compared to the effective Rabi frequency $g\sqrt{N!}$, an interference effect due to the different absorption paths can take place, canceling out the system absorption. This phenomenon is analogous to the electromagnetically induced transparency (EIT) that happens in three-level atoms driven by probe and control fields \cite{Fleischhauer2005}. In our case, however, the system does not become transparent to the probe field, it becomes highly reflective instead, as briefly discussed by the authors for $M=N=1$ in Ref.~\cite{Rossatto2013}, which gives rise to a multiphoton-induced-reflectivity phenomenon.
To illustrate this effect, in Fig.~\ref{Figure_EIT} we show the normalized mean number of intracavity excitations, i.e., the normalized absorption ($\langle a^{\dagger}a\rangle/\langle a^{\dagger}a\rangle_{\rm max}$), as a function of the detuning between the probe and cavity-mode frequencies $\Delta_{p}$ for different nonlinearities, with $\moy{a^{\dagger}a}_{\rm max}$ calculated with $g=\Delta_p=0$. All curves were obtained by taking the steady-state solution of the master equation [Eq.~\eqref{eq:masterequation}].
We notice that the stronger the atomic dissipation rates, the less the cancellation of the absorption of the probe field on the resonance ($\Delta_{p}=0$). This happens because the cavity decay rate ($\kappa$) plays the same role as the excited-state decay rate in usual EIT experiments, while both our $ \gamma$ and $\gamma_\phi$ play the same role as the dephasing rate \cite{Souza2015}, which is an agent that obstructs the destructive interference responsible for the cancellation of the absorption of the probe field in the transparency window (reflectivity window in our case). We can also observe that the width of the reflectivity window increases when we increase $N$. This is due to fact that the width depends not only on $g$, but rather on the effective vacuum Rabi frequencies $g \sqrt{N!}$, similar to the case in which there are many atoms in cavity-EIT experiments \cite{Mcke2010}. Thus, here we identify that the effective atom-mode coupling in our case corresponds to the Rabi frequency of the control field in usual EIT experiments \cite{Souza2015}. It is worth noting that in our case the destructive interference is not due to an interference between two single-photon absorption paths as happens in usual EIT phenomenon \cite{Fleischhauer2005} and analogues \cite{Rossatto2013,Souza2015}, but due to an interference between two cascade decay channels.
\begin{figure}[b]
\includegraphics[trim = 6mm 1mm 1mm 1mm, clip, width=0.48\textwidth]{EIT.pdf}
\caption{Normalized absorption spectrum of the cavity mode. The black dash-dotted lines are for $g=0$, while $g=\kappa/4$ with $\gamma=\gamma_\phi=10^{-4}\kappa$ (blue full lines) and $\gamma=\gamma_\phi=\kappa/10$ (red dashed lines). We have adjusted the strength of the continuous probe field $\varepsilon_0 = g/10\sqrt{N!}$ to keep the maximum average number of photons of the order of $10^{-3}$ for all cases: (a) nonlinearity $N=M=1$, (b) $2$, (c) $3$, and (d) $4$.}
\label{Figure_EIT}
\end{figure}
Finally, a multiphoton JC system in this configuration can be seen as a quantum-optical frequency blocker, i.e., a quantum device (filter) that inhibits the passage of multiphoton external probe fields within a narrow (controllable) reflectivity window in frequency domain. In addition, our study paves the way to further investigation (and possible applications) of processes involving more complex multiphoton quantum interference, e.g., the extension of the cavity-EIT phenomenon, which considers three-level atoms, to the case in which the atom-cavity interaction is nonlinear.
\emph{Conclusion.}---We have investigated a multiphoton Jaynes-Cummings system driven by a nonlinear driving field. Working in the strong atom-cavity coupling regime, we have shown that the cavity dynamics can be restricted to an upper-limited Fock subspace with the atom kept in its ground state. Thus, by driving the cavity mode with a coherent field (superposition of many Fock states), only the cavity states within the aforementioned Fock subspace can be populated; i.e., we have shown how to implement a quantum scissor without requiring a nonlinear Kerr medium inside the cavity. From the point of view of the cavity transmission, this Fock state filter can also be seen as a quantum-optical intensity limiter. Additionally, the nonlinearities of the system and driving field can be chosen in order to allow arbitrary rotations in the Fock space of a bosonic mode. On the other hand, working in the weak atom-cavity coupling regime, we have shown that the system exhibits a multiphoton-induced-reflectivity phenomenon, i.e., the multiphoton Jaynes-Cummings interaction induces a narrow (controllable) reflectivity window in frequency domain for multiphoton driving fields, such that this system can be used as a quantum-optical frequency blocker. Our results are useful for applications in quantum information protocols that require arbitrary rotations in Fock space \cite{Santos2005}, the generation of finite-dimensional quantum-optical states \cite{Leoski2001} and the manipulation of the optical response of atom-field systems \cite{Souza2013b}. As discussed, our results are general and could be implemented at least for two-photon interactions and driving fields with current technology \cite{Vogel1995,Meekhof1996,Gau1992,brune1997,neil2015,Garziano2015,Bertet2005,Felicetti2018,Feli2018}. However, the experimental verification for higher nonlinearities remains open, which should motivate further research on this topic, also motivated by some possible applications, e.g., the use of an $N$-photon Jaynes-Cummings system as a qudit with harmonic spectrum, a deeper investigation of the multiphoton blockade in the system or even how to use it as a single-photon source, and an exploration of multiphoton quantum interference for more-complex processes.
\begin{acknowledgments}
We thank M. H. Oliveira for the cavity drawing. This work was supported by the S\~{a}o Paulo Research Foundation (FAPESP) Grants No.~2013/04162-5 and No.~2013/23512-7, National Council for Scientific and Technological Development (CNPq) Grant No.~308860/2015-2, and Brazilian National Institute of Science and Technology for Quantum Information (INCT-IQ) Grant No.~465469/2014-0.
\end{acknowledgments}
|
\section{Introduction}
In this paper, we consider the Alfv\'en waves governed by the two-dimensional incompressible MHD equations in a finite channel $\Omega=\big\{(x,y)|x\in\mathbb{T}, y\in[-1,1]\big\}$:
\begin{equation}\label{eq: ideal MHD}
\left\{\begin{array}{l}
\partial_tU+U\cdot\nabla U-H\cdot\nabla H+\nabla P=0,\\
\partial_tH+U\cdot\nabla H-H\cdot\nabla U=0,\\
\nabla\cdot U=0,\ \ \nabla\cdot H=0,\\
U_2(t,x,y)|_{y=-1,1}=0, \ \ H_2(t,x,y)|_{y=-1,1}=0.\\
\end{array}\right.
\end{equation}
with initial data $U(0,x,y)$ and $H(0,x,y)$. Here $U=(U_1,U_2)$, $H=(H_1,H_2)$
and $P$ denote the velocity field, magnetic field, and the total pressure(kinetic plus magnetic) of the magnetic fluid, respectively.
This system has an equilibrium $U_s=(u(y),0)$, $H_s=(b(y),0)$, $P_s=\mathrm{const}$.
We focus on the secular behavior of the 2D linearized MHD equations around this equilibrium, which take the form
\begin{equation}\label{eq: linearized MHD}
\left\{\begin{array}{l}
\partial_tV_1+u\partial_xV_1+\partial_xp+u'V_2-b\partial_xB_1-b'B_2=0,\\
\partial_tV_2+u\partial_xV_2+\partial_yp-b\partial_xB_2=0,\\
\partial_tB_1+u\partial_xB_1+b'V_2-b\partial_xV_1-u'B_2=0,\\
\partial_tB_2+u\partial_xB_2-b\partial_xV_2=0,\\
\nabla\cdot V=0,\ \ \nabla\cdot B=0,\\
V_2(t,x,y)|_{y=-1,1}=0, \ \ B_2(t,x,y)|_{y=-1,1}=0.\\
\end{array}\right.
\end{equation}
with initial data $V(0,x,y)=(v_1(x,y), v_2(x,y))$
and $B(0,x,y)=(b_1(x,y), b_2(x,y))$. Let $w_0=\partial_xv_2-\partial_yv_1$
and $j_0=\partial_xb_2-\partial_yb_1$ be the initial vorticity and current density.
It is easy to deduce form \eqref{eq: linearized MHD} that the vorticity $w=\partial_xV_2-\partial_yV_1$
and the current density $j=\partial_xB_2-\partial_yB_1$ satisfy the following equations:
\begin{align}\label{eq: vorticity and current density}
\left\{\begin{array}{l}
\partial_tw+u\partial_xw-b\partial_xj=u''V_2-b''B_2,\\
\partial_tj+u\partial_xj-b\partial_xw=b''V_2-u''B_2+u'\partial_xB_1
-u'\partial_yB_2+b'\partial_yV_2-b'\partial_xV_1,
\end{array}\right.
\end{align}
here and in what follows, we use the notions $u, u',u''$ and $b, b',b''$ in stead by $u(y), \pa_yu(y), \pa_{yy}u(y)$ and $b(y), \pa_yb(y), \pa_{yy}b(y)$ for brevity.
In terms of the stream functions $\psi$: $V=(\partial_y\psi, -\partial_x\psi)$,
then we have $w=-\Delta\psi$. Similarly, there is a scalar function $\phi$
such that $B=(\partial_y\phi, -\partial_x\phi)$ and $j=-\Delta\phi$.
Then we can deduce the following system on $(\psi, \phi)$:
\begin{equation}\label{eq: psi phi}
\left\{\begin{array}{l}
\partial_t(\Delta\psi)+u\partial_x(\Delta\psi)-b\partial_x(\Delta\phi)=u''\partial_x\psi
-b''\partial_x\phi,\\
\partial_t(\Delta\phi)+u\partial_x(\Delta\phi)-b\partial_x(\Delta\psi)=b''\partial_x\psi
-u''\partial_x\phi-2u'\partial_x\partial_y\phi
+2b'\partial_x\partial_y\psi,
\end{array}\right.
\end{equation}
with boundary condition $\psi(t,x,\pm1)=\phi(t,x,\pm1)=0$.
And then taking the Fourier transform in $x$ of the above equations and inverting the operator $(\partial_y^2-\alpha^2)$
we get for $\alpha\neq0$,
\begin{equation}\label{eq: Fourier equations2}
\left\{\begin{array}{l}
\partial_t\widehat{\psi}+i\alpha u\widehat{\psi}-i\alpha b\widehat{\phi}=2i\alpha(\partial_y^2-\alpha^2)^{-1}\big(u''\widehat{\psi}
-b''\widehat{\phi}
+u'\partial_y\widehat{\psi}-b'\partial_y\widehat{\phi}\big),\\
\partial_t\widehat{\phi}+i\alpha u\widehat{\phi}-i\alpha b\widehat{\psi}=0.
\end{array}\right.
\end{equation}
Let
\begin{equation}\label{eq: matrix}
M_{\al}=-\Delta_{\al}^{-1}
\left[
\begin{matrix}
& u''-u\Delta_{\al} & -b''+b\Delta_{\al}\\
& b\Delta_{\al}+b''+2b'\pa_y & -u\Delta_{\al}-u''-2u'\pa_y
\end{matrix}
\right]
\end{equation}
where $\Delta_{\al}=\pa_y^2-\al^2$ and its inverse $\Delta_{\al}^{-1}$ satisfies $(\pa_y^2-\al^2)\Delta_{\al}^{-1}\psi(\al,y)=\psi(\al,y)$ with the boundary value
$\Delta_{\al}^{-1}\psi(\al,y)|_{y=\pm 1}=0$. Then
\begin{equation}\label{eq: matrix-form}
\partial_t\Big(
\begin{array}{ccc}
\widehat{\psi}\\
\widehat{\phi}\\
\end{array}
\Big)(t,\al,y)=-i\alpha M_{\al}\Big(
\begin{array}{ccc}
\widehat{\psi}\\
\widehat{\phi}\\
\end{array}
\Big)(t,\al,y).
\end{equation}
The study of the (in)stability of Alfv\'{e}n waves for MHD equations is a very active field in physics and mathematics and there is a number of works in this field \cite{Biskamp, PAD01, HP83, AL89, CU72, HZ15}.
If the equilibrium is not flowing ($U_s\equiv 0$), then for the Alfv\'{e}n waves in homogeneous magnetic fields (the case $b(y)=1$), it is easy to obtain the linear stability by using the fact that the current density $j$ and the vorticity $w$ satisfy the wave equation
\begin{equation*}
\pa_{tt}j-\pa_{xx}j=0,\ \ \pa_{tt}w-\pa_{xx}w=0,
\end{equation*}
with initial data $(j,w)(0,x,y)=(j_0,w_0)$ and $(\pa_tj, \pa_tw)(0,x,y)=(\pa_xw_0, \pa_xj_0)$. The above fact also implies that there is no hope of obtaining a decay estimate of the velocity. For nonlinear global stability in the homogeneous case, we refer the reader to \cite{BSS88}. For the nonresistive MHD equation in which the termm $-\mu\Delta U$ appears in (\ref{eq: ideal MHD}), we refer the reader to \cite{CMRR16, FMRR14, RW16, WZ17} for the local well-posedness results. The global well-posedness and stability results may be found in \cite{RWXZ14, TZ16} for the 2D case and \cite{AZ17,DZ17} for the 3D case. For the fully diffusive MHD equation in which both the terms $-\mu\Delta U$ and $-\nu\Delta H$ appear in (\ref{eq: ideal MHD}), we refer the reader to \cite{CL18, HXY18, WZ17}. For Alfv\'{e}n waves in inhomogeneous magnetic fields, there are few rigorous mathematical results. Grossmann and Tataronis \cite{GT72, TG72, TG73} predicted that the decay rate of the velocity is $O(t^{-1})$. Recently Ren and Zhao \cite{RZ17} gave a rigorous proof for the strict monotone positive magnetic field case in a finite channel. The mechanism leading to the damping is the phase mixing phenomenon, which is similar to the well-known Landau damping found by Landau in 1946 \cite{LL46} and proved by Mouhot and Villani in their remarkable work \cite{MV11}. This common phenomenon also appears in the Euler equation, which is called the inviscid damping. One may refer to \cite{BM10,KMC60, WO07, RS66, SAS95, WZZ18, WZZ1, WZZ2, CZ17} for the linear inviscid damping results and to \cite{ BM15, LZ11} for the nonlinear inviscid damping results.
If the plasma is flowing, the long-time behavior of the solution to linearized equation is not easy as before. One of the reasons is that $M_{\al}$ is not a self-adjoint operator when $u\neq 0$. Another reason is from the physical observation: the reconnection phenomenon. Reconnection of field lines is the process by which the topology of a flux surface structure in a plasma can change. For more details about the stability of Alfv\'en waves in a flowing plasma, we refer to \cite{Hameiri1983, ShiTok2010}. By using Fourier-Laplace analysis, Hirota, Tatsuno and Yoshida \cite{HTY05} studied a special case ($(u(y),b(y))=(k_1y,k_2y)$) and gave a formal analysis about the asymptotic behavior and predicted the existence of the magnetic island by assuming linear profiles of the ambient magnetic field and flow $(u(y),b(y))=(k_1y,k_2y)$ with $k_1,k_2=const$. The linear system is spectral stable, when $0\leq k_1<k_2$, however because of the existence of the magnetic island, the topology of the magnetic field lines in the final state may be different from the initial magnetic field lines. This instability, which tears and reconnects field lines, is called a tearing mode \cite{HZ15}.
In this paper, we study a more general case and provide a justification for Hirota, Tatsuno and Yoshida's prediction about the generation of magnetic island. We also carefully study the behavior of the magnetic island, which may be used as a modifying factor, when we linearized the MHD system and study the nonlinear instability of the Alfv\'en waves in a flowing plasma in the future.
Now we introduce some conditions on the background magnetic and velocity fields.
\begin{itemize}
\item[1.] Regularity, $(\bf R)$ : $u(y), b(y)\in C^5\left([-1,1]\right)$,
\item[2.] Stern stability, $(\bf SS)$: $|u(y)|\leq |b(y)|$,
\item[3.] Island, $(\bf I)$: $b(0)=u(0)=0$,
\item[4.] Monotone, $(\bf M)$: $b'(y)-|u'(y)|\geq c_0>0$
for some positive constant $c_0$.
\end{itemize}
One may regard $u(y)=ky, b(y)=k_0y$ with $k_0>|k|>0$ as an example. It is easy to check that $(\bf I)$
and $(\bf M)$ imply $(\bf SS)$.
\begin{theorem}\label{main thm}
Assume that $u(y), b(y)$ satisfy $(\bf R), (\bf I)$ and $(\bf M)$ and let $\big(\psi(t,x,y),\phi(t,x,y)\big)$ be the solution of (\ref{eq: psi phi}) with initial data $(\psi_0,\phi_0)\in H^{3}(-1,1)\times H^{4}(-1,1)$.
There holds that, \\
1. for $y=0$, as $t\rightarrow+\infty$ and $\al\neq0$
\begin{align*}
&\widehat{\psi}(t,\al,0)\rightarrow\frac{u'(0)}
{b'(0)}\widehat{\phi}_0(\al,0),\\
&\widehat{\phi}(t,\al,0)\equiv\widehat{\phi}_0(\al,0);
\end{align*}
2. for $0<y\leq 1$, as $t\rightarrow+\infty$, there exists $\Gamma^{+}(\al,y)$ such that
\begin{align*}
\widehat{\psi}(t,\al, y)&\rightarrow -\frac{u(y)}{b(y)}\big(b(y)\Gamma^{+}(\al,y)\big)\widehat{\phi}_0(\al,0),
\\
\widehat{\phi}(t,\al, y)&\rightarrow -\big(b(y)\Gamma^{+}(\al,y)\big)\widehat{\phi}_0(\al,0);
\end{align*}
3. for $-1\leq y<0$, as $t\rightarrow+\infty$, there exists $\Gamma^{-}(\al,y)$ such that
\begin{align*}
\widehat{\psi}(t,\al, y)&\rightarrow -\frac{u(y)}{b(y)}\big(b(y)\Gamma^{-}(\al,y)\big)\widehat{\phi}_0(\al,0),\\
\widehat{\phi}(t,\al, y)&\rightarrow -\big(b(y)\Gamma^{-}(\al,y)\big)\widehat{\phi}_0(\al,0).
\end{align*}
Moreover we have
\begin{eqnarray*}
\Gamma^{\pm}(\al,y)=\f{\va_{\pm}(\al,y)(u'(0)^2-b'(0)^2)}{b'(0)}
\int_{\pm1}^y\frac{1}{\big(u(y')^2-b(y')^2\big)\va_{\pm}(\al,y')^2}dy',
\end{eqnarray*}
where $\va_{\pm}$ solves $\pa_y\left((u^2-b^2)\pa_y\va_{\pm}\right)-\al^2(u^2-b^2)\va_{\pm}=0$ with boundary conditions $\va_{\pm}(\al,0)=1$ and $\pa_y\va_{\pm}(\al,0)=0$.
\end{theorem}
\begin{remark}\label{rmk:t revers}
Of course, by time-reversibility, Theorem \ref{main thm} is also true $t\to -\infty$. The limiting profile is independent of the initial stream function. As $V_2=-\pa_x\psi$ and $B_2=-\pa_x\phi$, the limiting profile of $\widehat{V_2}$ and $\widehat{B_2}$ are the same as $\widehat{\psi}$ and $\widehat{\phi}$.
\end{remark}
\begin{remark}\label{rmk: ident phi(t,0)}
The result $\widehat{\phi}(t,\al,0)\equiv \widehat{\phi}_0(\al,0)$ can be obtained by the equation \eqref{eq: Fourier equations2} and assumption ({\textbf{I}}).
\end{remark}
\begin{remark}\label{rmk: discontinuous}
By the fact that $\lim\limits_{y\to 0}b(y)\Gamma^{\pm}(\al,y)=-1$, we get that the limiting profile is continuous at $y=0$, precisely $\lim\limits_{y\to 0}\lim\limits_{t\to +\infty}\widehat{\psi}(t,\al,y)=\lim\limits_{t\to +\infty}\widehat{\psi}(t,\al,0)$. As for the its derivative, in \cite{HTY05}, the authors predict that the limiting profile has a derivative jump at $y=0$. Actually, we can get that if
\begin{eqnarray*}
-5u'(0)u''(0)+u''(0)b'(0)-u'(0)b''(0)+5b'(0)b''(0)\neq0,
\end{eqnarray*}
then there exists a positive constant $C$ such that
\begin{eqnarray*}
\big|\pa_y\big(b(y)\Gamma^{\pm}(\al,y)\big)\big|\geq C^{-1}\big(1+\big|\ln|y|\big|\big).
\end{eqnarray*}
That means in this case the final state of $\phi$ and $\psi$ does not decay to $W^{1,\infty}$, which may be useful in the study of the nonlinear instability.
\end{remark}
\begin{remark}\label{rmk: special case}
If $u(y)=ky$, $b(y)=k_0y$ for some constant $k_0>|k|\geq 0$, then $b(y)\Gamma^{\pm}(\al,y)$ are harmonic functions on $\mathbb{T}\times[0,\pm1]$ with boundary condition $b(0)\Gamma^{\pm}(\al,0)=-1$ and $b(\pm1)\Gamma^{\pm}(\al,\pm1)=0$. Then the final state is in $W^{1,\infty}$ and its profile is as follows.\\
\begin{tikzpicture}[thick, scale=1.6]
\draw[->](-1.5,0)--(1.5,0) node[below]{$y$};
\draw[->](0,-0.5)--(0,1.5);
\draw[color=red] plot[domain=0:1](\x, {-exp(-2)/(1-exp(-2))*exp(\x)+1/(1-exp(-2))*exp(-\x)});
\draw[color=blue] plot[domain=-1:0](\x, {-exp(2)/(1-exp(2))*exp(\x)+1/(1-exp(2))*exp(-\x)});
\draw (1,0.5) node{$-b\Gamma^+$};
\draw (-1,0.5) node {$-b\Gamma^-$};
\draw (-1,-0.15) node {$-1$};
\draw (1,-0.15) node {$1$};
\draw (0.25,1) node {$1$};
\end{tikzpicture}
\end{remark}
\begin{remark}\label{rmk: open problem}
The necessary and sufficient condition of the generation of the magnetic island is an open problem. Roughly speaking, according to \eqref{eq: matrix-form}, assuming there is a magnetic island at the finial state, then there is non-trivial solution of $M_{\al}\left(\begin{aligned}\widehat{\psi}\\\widehat{\phi}\end{aligned}\right)=0$. Formally we can get that the necessary condition of the generation of the magnetic island may be $0\in \sigma(M_{\al})$.
\end{remark}
\begin{remark}\label{rmk: add}
The convergence rate will be discussed in a separated work.
\end{remark}
\section{reduce the problem and the Sturmian equations}
Let $\Om$ be the domain that contains $\sigma(M_{\al})$.
Then we have the following representation formula of the solution to (\ref{eq: matrix-form}):
\begin{equation}\label{eq: psi and phi}
\Big(
\begin{array}{ccc}
\widehat{\psi}\\
\widehat{\phi}\\
\end{array}
\Big)(t,\alpha,y)=\frac{1}{2\pi i}\int_{\partial\Omega}e^{-i\alpha tc}(cI-M_{\al})^{-1}\Big(
\begin{array}{ccc}
\widehat{\psi}\\
\widehat{\phi}\\
\end{array}
\Big)(0,\alpha,y)dc.
\end{equation}
Then the large time behavior of the solution $\Big(
\begin{array}{ccc}
\widehat{\psi}\\
\widehat{\phi}\\
\end{array}
\Big)(t,\alpha,y)$ is reduced to the study of the resolvent $(cI-M_{\al})^{-1}$.
Suppose $\big(cI-M_{\al}\big)^{-1}\Big(\begin{array}{l} \widehat{\psi}_0\\ \widehat{\phi}_0\end{array}\Big)(\al,y)=\Big(\begin{array}{l} \Psi_1\\ \Phi_1\end{array}\Big)(\al,y,c)$, then
\begin{align*}
&\big(cI-M_{\al}\big)\Big(\begin{array}{l} \Psi_1\\ \Phi_1\end{array}\Big)(\al,y,c)=\Big(\begin{array}{l} \widehat{\psi}_0\\ \widehat{\phi}_0\end{array}\Big)(\al,y)\\
&\Leftrightarrow \left\{
\begin{array}{l}
-c\Delta_{\al}\Psi_1+u\Delta_{\al}\Psi_1
-u''\Psi_1-b\Delta_{\al}\Phi_1+b''\Phi_1
=\widehat{\om}_0,\\
-c\Delta_{\al}\Phi_1-b\Delta_{\al}\Psi_1-b''\Psi_1
+2u'\pa_y\Phi_1
+u\Delta_{\al}\Phi_1+u''\Phi_1-2b'\pa_y\Psi_1
=\widehat{j}_0.
\end{array}
\right.\\
&\Leftrightarrow \left\{
\begin{array}{l}
(u-c)\Delta_{\al}\Psi_1-u''\Psi_1-b\Delta_{\al}\Phi_1
+b''\Phi_1
=\widehat{\om}_0,\\
(u-c)\Phi_1-b\Psi_1=-\widehat{\phi}_0.
\end{array}
\right.
\end{align*}
Here $\Delta_{\al}=\partial_y^2-\al^2$.
Let $\Phi_1(\al,y,c)=b(y)\Phi_2(\al,y,c)$ then $\Psi_1(\al,y,c)=(u(y)-c)\Phi_2(\al,y,c)+\widehat{\phi}_0(\al,y)/b(y)$, then we get
\begin{equation}\label{eq:ODE1}
\begin{split}
&\pa_y\Big[\Big(\big(u(y)-c\big)^2-b(y)^2\Big)\pa_y\Phi_2(\al, y,c)\Big]
-\al^2\Big(\big(u(y)-c\big)^2-b(y)^2\Big)\Phi_2(\al,y,c)\\
&
=\widehat{\om}_0(\al,y)-\big(u(y)-c\big)\Delta_{\al}\Big(\frac{\widehat{\phi}_0(\al,y)}{b(y)}\Big)
+u''(y)\frac{\widehat{\phi}_0(\al,y)}{b(y)}
\end{split}
\end{equation}
For this equation, by denoting $\Phi_2(\al,y,c)=\Phi(\al,y,c)+\widehat{\phi}_0(0)\chi(y)/(cb(y))$
whereas $0 \leq\chi(y)\in C_0^{\infty}(\mathbf R):
\chi(y)=1$ for $|y|\leq\frac{1}{2}$ and $\chi(y)=0$ for $|y|\geq 3/4$, we can get that
\begin{align*}
&\pa_y\Big[\Big(\big(u(y)-c\big)^2-b(y)^2\Big)\pa_y\Phi(\al,y,c)\Big]
-\al^2\Big(\big(u(y)-c\big)^2-b(y)^2\Big)\Phi(\al,y,c)\\ &=\widehat{w}_0(\al,y)-\big(u(y)-c\big)\Delta_{\al}\Big(\frac{\widehat{\phi}_0(\al,y)
-\widehat{\phi}_0(\al,0)}{b(y)}\Big)
+u''(y)\frac{\widehat{\phi}_0(y)-\widehat{\phi}_0(\al,0)}{b(y)}\\
&\ \
-\frac{\widehat{\phi}_0(\al,0)}{c}\frac{1}{b(y)^3}
\Big\{2\big(u(y)-c\big)\big[c+\big(u(y)-c\big)\chi(y)\big]b'(y)^2\\
&\ \
-b(y)\big\{\big(u(y)-c\big)\big[c+\big(u(y)-c\big)\chi(y)\big]b''(y)
+2\big(u(y)-c\big)\big[\big(u(y)-c\big)\chi'(y)+u'(y)\chi(y)\big]b'(y)\big\}\\
&\ \ -b(y)^2\big\{\al^2\big(u(y)-c\big)\big[c+\big(u(y)-c\big)\chi(y)\big]
+cu''(y)-\big(u(y)-c\big)^2\chi''(y)\\
&\ \ -2\big(u(y)-c\big)u'(y)\chi'(y)\big\}
+b(y)^3b''(y)\chi(y)+b(y)^4(\al^2\chi(y)-\chi''(y))\Big\}\\
&\stackrel{def}{=}G(\al,y,c)
=G_1(\al,y,c)-\frac{\widehat{\phi}_0(\al,0)}{c}\frac{f(\al,y,c)}{b(y)^3},
\end{align*}
whereas
\begin{equation}\label{eq: G_1}
G_1(\al,y,c)=\widehat{\om}_0(\al,y)-(u(y)-c)\Delta_{\al}
\Big(\frac{\widehat{\phi}_0(y)-\widehat{\phi}_0(0)}{b(y)}\Big)
+u''(y)\frac{\widehat{\phi}_0(y)-\widehat{\phi}_0(0)}{b(y)}
\end{equation}
and
\begin{equation}\label{eq: f}
\begin{split}
f(\al,y,c)&=2\big(u(y)-c\big)\big[c+\big(u(y)-c\big)\chi(y)\big]b'(y)^2\\
& \ \
-b(y)\big\{\big(u(y)-c\big)\big[c+\big(u(y)-c\big)\chi(y)\big]b''(y)\\
& \ \
+2(u(y)-c)\big[\big(u(y)-c\big)\chi'(y)+u'(y)\chi(y)\big]b'(y)\big\}\\
&\ \
-b(y)^2\big\{\al^2\big(u(y)-c\big)\big[c+\big(u(y)-c\big)\chi(y)\big]
-\big(u(y)-c\big)^2\chi''(y)+cu''(y)\\
&\ \ -2\big(u(y)-c\big)u'(y)\chi'(y)\big\}
+b(y)^3b''(y)\chi(y)+b(y)^4\big(\al^2\chi(y)-\chi''(y)\big).
\end{split}
\end{equation}
On the one hand, it is easy to show that $f(\al,0,c)=0, (\pa_yf)(\al,0,c)=0$ and $(\pa_y^2f)(\al,0,c)=0$, Thus we can deduce that $f(\al,y,c)$ is $C^2$ function when $u(y), b(y)$ is $C^5$ function. Then we obtain that, due to the fact that $b(0)=0$,
\begin{equation*}
\frac{f(\al,y,c)}{b(y)^3}=\frac{y^3}{b(y)^3}\frac{f(\al,y,c)}{y^3}
=\frac{1}{\big(\int_0^1b'(sy)ds\big)^3}
\int_0^1t\int_0^t\int_0^s\pa_y^3f(\al,\tau y,c)d\tau ds dt
\end{equation*}
and
\begin{equation*}
\Big\|\frac{f(\al,y,c)}{b(y)^3}\Big\|_{L^{\infty}}\leq C.
\end{equation*}
On the other hand, thanks to the fact that
$\frac{\widehat{\phi}_0(y)-\widehat{\phi}_0(0)}{b(y)}
=\frac{\int_0^1(\pa_y\widehat{\phi}_0)(sy)ds}{\int_0^1b'(sy)ds},$ we have,
\begin{equation*}
\|G_1(y,c)\|_{H_y^1}\leq C\|\widehat{\om}_0\|_{H_y^{1}}+C\|\widehat{j}_0\|_{H_y^{3}}.
\end{equation*}
\section{the homogeneous Sturmian equations}
To solve the inhomogeneous Sturmian equation, we first construct two regular solutions of the homogeneous Sturmian equation:
\begin{align*}
& \pa_y\big[\big(u(y)+b(y)-c\big)\big(u(y)-b(y)-c\big)\pa_y\va(\al, y, c)\big]\nonumber\\
&\quad\quad
-\al^2\big(u(y)+b(y)-c\big)\big(u(y)-b(y)-c\big)\va(\al, y, c)=0.
\end{align*}
We will suppress the variable $\al$ for simplicity and write the function $\va(y,c)=\va(\al,y,c)$.
Let $W_+(y)=u(y)+b(y)$, $W_-(y)=u(y)-b(y)$, then
from $(\bf M)$, we get $W'_+(y)>0$, $W'_-(y)<0$. And we define $\mathcal{H}(y,c)=(W_+(y)-c)(W_-(y)-c)$,
where the constant coefficient $c$ will be taken in domain:
$c\in \Omega_{\ep_0}=\{z\in \mathbb{C}, dist(z,Ran\, W_+ \cup Ran\, W_-)\leq \ep_0\}$(the $\ep_0$ neighborhood of $Ran\, W_+ \cup Ran\, W_-$).
According to the relationship between $W_+(1), W_-(1), W_+(-1)$ and $W_-(-1)$, the domain have the following nine cases:
\begin{itemize}
\item[1.] $W_+(1)> W_-(-1)>0> W_-(1)>W_+(-1)$,
\item[2.] $W_+(1)= W_-(-1)>0> W_-(1)>W_+(-1)$,
\item[3.] $W_-(-1)> W_+(1)>0> W_-(1)>W_+(-1)$,
\item[4.] $W_+(1)> W_-(-1)>0> W_-(1)=W_+(-1)$,
\item[5.] $W_+(1)> W_-(-1)>0> W_+(-1)>W_-(1)$,
\item[6.] $W_+(1)=W_-(-1)>0> W_-(1)=W_+(-1)$,
\item[7.] $W_+(1)= W_-(-1)>0> W_+(-1)>W_-(1)$,
\item[8.] $W_-(-1)> W_+(1)>0> W_-(1)=W_+(-1)$,
\item[9.] $W_-(-1)> W_+(1)>0> W_+(-1)>W_-(1)$.
\end{itemize}
Before we introduce all these cases, we first introduce the extension lemma.
\begin{lemma}\label{lem: extend}
Let $f$ be a $C^k$ function defined on $[a,b]$ with $f'(b)>0$. Assume $M>f(b)$. Then for any $d>b$, there exists $F\in C^{k}([a,d])$ such that $F'(x)>0$ for $x\in [b,d]$ and $F(x)=f(x)$ for $x\in [a,b]$ and $F^{(m)}(b)=f^{(m)}(b)$ for $m=1,...,k$. Moreover, it holds that $\|F\|_{C^k}\leq C\|f\|_{C^k}$.
\end{lemma}
\begin{proof}
Let $\delta_1<\f{d-b}{4}$ be small enough which will be determined later and $0\leq \chi(x)\leq 1$ be a smooth non-negative function with compact support satisfying $|\chi'(x)|\leq C\delta_1^{-1}$, $\chi(x)=1$ for $0\leq x\leq \f14\delta_1$ and $\chi(x)=0$ for $x\geq \f34\delta_1$. Define for $x\in [b, b+\delta_1]$,
\begin{eqnarray*}
F(x)=f(b)+f'(b)(x-b)+\sum_{n=2}^{k}\f{f^{(n)}(b)}{n!}(x-b)^{n}\chi(x)
\end{eqnarray*}
Then $F'(x)=f'(b)+\sum_{n=2}^{k}\f{f^{(n)}(b)}{(n-1)!}(x-b)^{n-1}\chi(x)+\sum_{n=2}^{k}\f{f^{(n)}(b)}{n!}(x-b)^{n}\chi'(x-b)$, by the fact that $|\chi'(x)|\leq 1/\delta_1$, we have for $|x-b|\leq \delta_1$
\begin{eqnarray*}
F'(x)\geq f'(b)-C\max_{n=2,...,k}|f^{(n)}(b)|\delta_1.
\end{eqnarray*}
and $F(b+\delta_1)=f(b)+f'(b)\delta_1$, $F'(b+\delta_1)=f'(b)$ and $F^{(n)}(b+\delta_1)=0$ for $n\geq 2$.
By taking $\delta_1$ small enough such that $C\max_{n=2,...,k}|f^{(n)}(b)|\delta_1\leq \f12f'(b)$ and $5f'(b)\delta_1<M-f(b)$, then we have for $|x-b|\leq \delta_1$, $F'(x)\geq \f12f'(b)$ and $F(b+\delta_1)<M$.
Let $g(x)>0$ be a smooth function, such that $g(x)=f'(b)$ for $x\in [b+\delta_1,b+2\delta_1]$ and
\begin{eqnarray*}
\int_{b+\delta_1}^dg(x)dx=M-F(b+\delta_1).
\end{eqnarray*}
Let $F(x)=F(b+\delta_1)+\int_{b+\delta_1}^xg(x')dx'$ for $x\in [b+\delta_1,d]$, then $F(x)$ is the extension function.
\end{proof}
For {\bf Case 1}, we let
\begin{eqnarray*}
&\ & D_0\stackrel{def}{=}\big\{c\in[W_+(-1), W_+(1)]\big\},\\
&\ & D_{\ep_0}\stackrel{def}{=}\big\{c=c_r+i\ep, \ c_r\in[W_+(-1), W_+(1)], \ 0<|\ep|<\ep_0\big\},\\
&\ & B_{\ep_0}^{l}\stackrel{def}{=}\big\{c=W_+(-1)+\ep e^{i\theta},\ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},\\
&\ & B_{\ep_0}^{r}\stackrel{def}{=}\big\{c=W_+(1)-\ep e^{i\theta}, \ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},
\end{eqnarray*}
for some $\ep_0\in(0,1).$ We denote
$
\Omega_{\ep_0}\stackrel{def}{=}D_0\cup D_{\ep_0}\cup B_{\ep_0}^{l}\cup B_{\ep_0}^{r}.
$
We also define
\begin{equation*}
c_r=Re\, c\ for\ c\in D_0\cup D_{\ep_0}, \ c_r=W_+(-1)
\ for \ c\in B_{\ep_0}^{l}, \ c_r=W_+(1) \ for \ c\in B_{\ep_0}^{r}.
\end{equation*}
By Lemma \ref{lem: extend}, we can take a $C^5$ extension of $W_-$ to be $\wt{W}_-$ for $y\in[\rm{a}_-,\rm{a}_+]$ such that $\wt{W}_-(\rm{a}_-)=W_+(1)$, $\wt{W}_-(\rm{a}_+)=W_+(-1)$ and $\wt{W}_-'(y)<0$.
\begin{itemize}
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\geq 0$, we denote $y_{c_+}\in [0,1]$ with $y_{c_+}=(W_+)^{-1}(c_r)$, so that $W_+(y_{c_+})-c_r=0$ and $y_{c_-}\in [\rm{a}_-,0]$ with $y_{c_-}=(\wt{W}_-)^{-1}(c_r)$, so that $\wt{W}_-(y_{c_-})-c_r=0$.
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\leq 0$, we denote $y_{c_+}\in [0,\rm{a}_+]$ with $y_{c_+}=(\wt{W}_-)^{-1}(c_r)$, so that $\wt{W}_-(y_{c_+})=c_r$ and $y_{c_-}\in [-1,0]$ with $y_{c_-}=(W_+)^{-1}(c_r)$, so that $W_+(y_{c_-})-c_r=0$.
\item For $c\in B_{\ep_0}^{l}$ then $c_r=W_+(-1)=\wt{W}_-(\rm{a}_+)$, we denote $y_{c_+}=\rm{a}_+$ and $y_{c_-}=-1$.
\item For $c\in B_{\ep_0}^r$ then $c_r=W_+(1)=\wt{W}_-(a_-)$, we denote $y_{c_+}=1$ and $y_{c_-}=\rm{a}_-$.
\end{itemize}
We show the relationship between $y_{c_+}$, $y_{c_-}$ and $c_r$ by the following picture for the above case.
\begin{tikzpicture}[thick, scale=0.45]
\draw[very thin,color=gray];
\draw[->] (-7,0) -- (7,0) node[right] {$y$};
\draw (0.1,-0.01) node[anchor=north] {0}
(5,0) node[anchor=north] {1}
(-4.8,0) node[anchor=north] {-1};
\draw[dotted] (-6,4) -- (5,4);
\draw[dotted] (-5,-4) -- (6,-4);
\draw[dotted] (-5,-4) -- (-5,3);
\draw[dotted] (5,-3) -- (5,4);
\draw[red,thick] (-5,1.8) -- (5,1.8);
\draw[red] (0.3,2.2) node {$c_r$};
\draw[->] (0,-5) -- (0,5) node[above] {$c_r$};
\draw[thick] (0,0) parabola[bend at end] (1,1.2);
\draw[thick] (1.8,2) parabola[bend at end] (1,1.2);
\draw[thick] (1.8,2) parabola[bend at end] (2.6,2.8);
\draw[thick] (3.8,3.2) parabola[bend at end] (2.6,2.8);
\draw[thick] (3.8,3.2) parabola[bend at end] (5,4);
\draw[thick] (0,0) parabola[bend at end] (-1,-1.2);
\draw[thick] (-1.8,-2) parabola[bend at end] (-1,-1.2);
\draw[thick] (-1.8,-2) parabola[bend at end] (-2.6,-2.8);
\draw[thick] (-3.8,-3.2) parabola[bend at end] (-2.6,-2.8);
\draw[thick] (-3.8,-3.2) parabola[bend at end] (-5,-4);
\draw[thick] (0,0) parabola[bend at end] (-1.5,0.7);
\draw[thick] (-2.3,1.2) parabola[bend at end] (-1.5,0.7);
\draw[thick] (-2.3,1.2) parabola[bend at end] (-3.1,2.0);
\draw[thick] (-4,2.5) parabola[bend at end] (-3.1,2.0);
\draw[thick] (-4,2.5) parabola[bend at end] (-5,3);
\draw[red,dotted] (-6,4) parabola[bend at end] (-5,3);
\draw[thick] (0,0) parabola[bend at end] (1.5,-1.0);
\draw[thick] (2.5,-1.8) parabola[bend at end] (1.5,-1.0);
\draw[thick] (2.5,-1.8) parabola[bend at end] (3.1,-2.2);
\draw[thick] (4,-2.7) parabola[bend at end] (3.1,-2.2);
\draw[thick] (4,-2.7) parabola[bend at end] (5,-3);
\draw[red, dotted] (6,-4) parabola[bend at end] (5,-3);
\draw (5.1,4.4) node {\small $W_+(y)$};
\draw (4.1,-3.3) node {\small $W_-(y)$};
\draw[dotted] (-2.7,0) -- (-2.7,1.8);
\draw (-2.55,-0.5) node {\small $y_{c_-}$};
\draw[dotted] (1.7,0) -- (1.7,1.8);
\draw (1.8,-0.5) node {\small $y_{c_+}$};
\draw[dotted] (6,-4)-- (6,0); \draw (6,0.5) node {$\rm{a}_+$};
\draw[dotted] (-6,4)-- (-6,0); \draw (-6,-0.5) node {$\rm{a}_-$};
\draw (-3.5,4.6) node {\small $\wt{W}_-({\rm{a}}_-)=W_+(1)$};
\draw (3.5,-4.6) node {\small $\wt{W}_-({\rm{a}}_+)=W_+(-1)$};
\end{tikzpicture}
We also take a $C^5$ extension of $W_+$ to be $\wt{W}_+$ for $y\in [\rm{a}_-,{\rm{a}}_+]$, so that $\wt{W}_+'(y)>0$.
For {\bf{Case 2}}, we let
\begin{eqnarray*}
&\ & D_0\stackrel{def}{=}\big\{c\in[W_+(-1), W_+(1)]\big\},\\
&\ & D_{\ep_0}\stackrel{def}{=}\big\{c=c_r+i\ep, \ c_r\in[W_+(-1), W_+(1)], \ 0<|\ep|<\ep_0\big\},\\
&\ & B_{\ep_0}^{l}\stackrel{def}{=}\big\{c=W_+(-1)+\ep e^{i\theta},\ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},\\
&\ & B_{\ep_0}^{r}\stackrel{def}{=}\big\{c=W_+(1)-\ep e^{i\theta}, \ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},
\end{eqnarray*}
for some $\ep_0\in(0,1).$ We denote
$
\Omega_{\ep_0}\stackrel{def}{=}D_0\cup D_{\ep_0}\cup B_{\ep_0}^{l}\cup B_{\ep_0}^{r}.
$
We also define
\begin{equation*}
c_r=Re c\ for\ c\in D_0\cup D_{\ep_0}, \ c_r=W_+(-1)
\ for \ c\in B_{\ep_0}^{l}, \ c_r=W_+(1) \ for \ c\in B_{\ep_0}^{r}.
\end{equation*}
By Lemma \ref{lem: extend}, we can take a $C^5$ extension of $W_-$ to be $\wt{W}_-$ for $y\in[-1,{\rm{a}}_+]$ such that $\wt{W}_-({\rm{a}}_+)=W_+(-1)$ and $\wt{W}'_-(y)<0$.
\begin{itemize}
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\geq0$, we denote $y_{c_+}\in[0,1]$ with $y_{c_+}=(W_+)^{-1}(c_r)$, so that $W_+(y_{c_+})-c_r=0$ and $y_{c_-}\in[-1,0]$ with $y_{c_-}=(W_-)^{-1}(c_r)$, so that $W_-(y_{c_-})-c_r=0$.
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\leq0$, we denote $y_{c_+}\in[0,{\rm{a}}_+]$ with $y_{c_+}=(\wt{W}_-)^{-1}(c_r)$, so that $\wt{W}_-(y_{c_+})-c_r=0$ and $y_{c_-}\in[-1,0]$ with $y_{c_-}=(W_+)^{-1}(c_r)$, so that $W_+(y_{c_-})-c_r=0$.
\item For $c\in B_{\ep_0}^l$, then $c_r=W_+(-1)=\wt{W}_-({\rm{a}}_+)$, we denote $y_{c_+}={\rm{a}}_+$ and $y_{c_-}=-1$.
\item For $c\in B_{\ep_0}^r$, then $c_r=W_+(1)=W_-(-1)$, we denote $y_{c_+}=1$ and $y_{c_-}=-1$.
\end{itemize}
We show the relationship between $y_{c_+}, y_{c_-}$ and $c_r$ by the following picture for the above case.
\begin{tikzpicture}[thick, scale=0.5]
\draw[very thin,color=gray];
\draw[->] (-7,0) -- (7,0) node[right] {$y$};
\draw (0.1,-0.01) node[anchor=north] {0}
(5,0) node[anchor=north] {1}
(-4.8,0) node[anchor=north] {-1};
\draw[dotted] (-5,4) -- (5,4);
\draw[dotted] (-5,-4) -- (6,-4);
\draw[dotted] (-5,-4) -- (-5,4);
\draw[dotted] (5,-3) -- (5,4);
\draw[red,thick] (-5,1.8) -- (5,1.8);
\draw[red] (0.3,2.2) node {$c_r$};
\draw[->] (0,-5) -- (0,5) node[above] {$c_r$};
\draw[thick] (0,0) parabola[bend at end] (1,1.2);
\draw[thick] (1.8,2) parabola[bend at end] (1,1.2);
\draw[thick] (1.8,2) parabola[bend at end] (2.6,2.8);
\draw[thick] (3.8,3.2) parabola[bend at end] (2.6,2.8);
\draw[thick] (3.8,3.2) parabola[bend at end] (5,4);
\draw[thick] (0,0) parabola[bend at end] (-1,-1.2);
\draw[thick] (-1.8,-2) parabola[bend at end] (-1,-1.2);
\draw[thick] (-1.8,-2) parabola[bend at end] (-2.6,-2.8);
\draw[thick] (-3.8,-3.2) parabola[bend at end] (-2.6,-2.8);
\draw[thick] (-3.8,-3.2) parabola[bend at end] (-5,-4);
\draw[thick] (0,0) parabola[bend at end] (-1.5,1.1);
\draw[thick] (-2.5,1.9) parabola[bend at end] (-1.5,1.1);
\draw[thick] (-2.5,1.9) parabola[bend at end] (-3.1,2.6);
\draw[thick] (-4,3.1) parabola[bend at end] (-3.1,2.6);
\draw[thick] (-4,3.1) parabola[bend at end] (-5,4);
\draw[thick] (0,0) parabola[bend at end] (1.5,-1.0);
\draw[thick] (2.5,-1.8) parabola[bend at end] (1.5,-1.0);
\draw[thick] (2.5,-1.8) parabola[bend at end] (3.1,-2.2);
\draw[thick] (4,-2.7) parabola[bend at end] (3.1,-2.2);
\draw[thick] (4,-2.7) parabola[bend at end] (5,-3);
\draw[red, dotted] (6,-4) parabola[bend at end] (5,-3);
\draw (5.1,4.4) node {\small $W_+(y)$};
\draw (4.1,-3.3) node {\small $W_-(y)$};
\draw[dotted] (-2.5,0) -- (-2.5,1.8);
\draw (-2.3,-0.5) node {$y_{c_-}$};
\draw[dotted] (1.7,0) -- (1.7,1.8);
\draw (1.8,-0.5) node {$y_{c_+}$};
\draw[dotted] (6,-4)-- (6,0); \draw (6,0.5) node {${\rm{a}}_+$};
\draw (-3.5,4.5) node {\small $W_-(-1)=W_+(1)$};
\draw (3.5,-4.6) node {\small $\wt{W}_-({\rm{a}}_+)=W_+(-1)$};
\end{tikzpicture}
We also take a $C^5$ extension of $W_+$ to be $\wt{W}_+$ for $y\in [-1,{\rm{a}}_+]$, so that $\wt{W}_+'(y)>0$.
For {\bf{Case 3}}, we let
\begin{eqnarray*}
&\ & D_0\stackrel{def}{=}\big\{c\in[W_+(-1), W_-(-1)]\big\},\\
&\ & D_{\ep_0}\stackrel{def}{=}\big\{c=c_r+i\ep, \ c_r\in[W_+(-1), W_-(-1)], \ 0<|\ep|<\ep_0\big\},\\
&\ & B_{\ep_0}^{l}\stackrel{def}{=}\big\{c=W_+(-1)+\ep e^{i\theta},\ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},\\
&\ & B_{\ep_0}^{r}\stackrel{def}{=}\big\{c=W_-(-1)-\ep e^{i\theta}, \ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},
\end{eqnarray*}
for some $\ep_0\in(0,1).$ We denote
$
\Omega_{\ep_0}\stackrel{def}{=}D_0\cup D_{\ep_0}\cup B_{\ep_0}^{l}\cup B_{\ep_0}^{r}.
$
We also define
\begin{equation*}
c_r=Re c\ for\ c\in D_0\cup D_{\ep_0}, \ c_r=W_+(-1)
\ for \ c\in B_{\ep_0}^{l}, \ c_r=W_-(-1) \ for \ c\in B_{\ep_0}^{r}.
\end{equation*}
By Lemma \ref{lem: extend}, we can take a $C^5$ extension of $W_-$ to be $\wt{W}_-$ and $W_+$ to be $\wt{W}_+$ for $y\in[-1,{\rm{a}}_+]$ such that $\wt{W}_-({\rm{a}}_+)=W_+(-1)$, $\wt{W}'_-(y)<0$ and $\wt{W}_+({\rm{a}}_+)=W_-(-1)$, $\wt{W}'_+(y)>0$.
\begin{itemize}
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\geq0$, we denote $y_{c_+}\in[0,{\rm{a}}_+]$ with $y_{c_+}=(\wt{W}_+)^{-1}(c_r)$, so that $\wt{W}_+(y_{c_+})-c_r=0$ and $y_{c_-}\in[-1,0]$ with $y_{c_-}=(W_-)^{-1}(c_r)$, so that $W_-(y_{c_-})-c_r=0$.
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\leq0$, we denote $y_{c_+}\in[0,{\rm{a}}_+]$ with $y_{c_+}=(\wt{W}_-)^{-1}(c_r)$, so that $\wt{W}_-(y_{c_+})-c_r=0$ and $y_{c_-}\in[-1,0]$ with $y_{c_-}=(W_+)^{-1}(c_r)$, so that $W_+(y_{c_-})-c_r=0$.
\item For $c\in B_{\ep_0}^l$, then $c_r=W_+(-1)=\wt{W}_-({\rm{a}}_+)$, we denote $y_{c_+}={\rm{a}}_+$ and $y_{c_-}=-1$.
\item For $c\in B_{\ep_0}^r$, then $c_r=W_-(-1)=\wt{W}_+({\rm{a}}_+)$, we denote $y_{c_+}={\rm{a}}_+$ and $y_{c_-}=-1$.
\end{itemize}
We show the relationship between $y_{c_+}, y_{c_-}$ and $c_r$ by the following picture for the above case.
\begin{tikzpicture}[thick, scale=0.5]
\draw[very thin,color=gray];
\draw[->] (-7,0) -- (7,0) node[right] {$y$};
\draw (0.1,-0.01) node[anchor=north] {0}
(5,0) node[anchor=north] {1}
(-4.8,0) node[anchor=north] {-1};
\draw[dotted] (-5,4) -- (5.9,4);
\draw[dotted] (-5,-4) -- (5.9,-4);
\draw[dotted] (-5,-4) -- (-5,4);
\draw[dotted] (5,-3) -- (5,3);
\draw[red,thick] (-5,2) -- (5,2);
\draw[red] (0.3,2.2) node {$c_r$};
\draw[->] (0,-5) -- (0,5) node[above] {$c_r$};
\draw[thick] (0,0) parabola[bend at end] (1.5,1.0);
\draw[thick] (2.5,1.8) parabola[bend at end] (1.5,1.0);
\draw[thick] (2.5,1.8) parabola[bend at end] (3.9,2.3);
\draw[thick] (5,3) parabola[bend at end] (3.9,2.3);
\draw[red,dotted] (5,3) parabola[bend at end] (5.9,4);
\draw[thick] (0,0) parabola[bend at end] (-1,-1.2);
\draw[thick] (-1.8,-2) parabola[bend at end] (-1,-1.2);
\draw[thick] (-1.8,-2) parabola[bend at end] (-2.6,-2.8);
\draw[thick] (-3.8,-3.2) parabola[bend at end] (-2.6,-2.8);
\draw[thick] (-3.8,-3.2) parabola[bend at end] (-5,-4);
\draw[thick] (0,0) parabola[bend at end] (-1.5,1.1);
\draw[thick] (-2.5,1.9) parabola[bend at end] (-1.5,1.1);
\draw[thick] (-2.5,1.9) parabola[bend at end] (-3.1,2.6);
\draw[thick] (-4,3.1) parabola[bend at end] (-3.1,2.6);
\draw[thick] (-4,3.1) parabola[bend at end] (-5,4);
\draw[thick] (0,0) parabola[bend at end] (1.5,-1.0);
\draw[thick] (2.5,-1.8) parabola[bend at end] (1.5,-1.0);
\draw[thick] (2.5,-1.8) parabola[bend at end] (3.1,-2.2);
\draw[thick] (4,-2.7) parabola[bend at end] (3.1,-2.2);
\draw[thick] (4,-2.7) parabola[bend at end] (5,-3);
\draw[red, dotted] (5.9,-4) parabola[bend at end] (5,-3);
\draw (4.0,3.0) node {\small $W_+(y)$};
\draw (4.2,-3.2) node {\small $W_-(y)$};
\draw[dotted] (-2.55,0) -- (-2.55,2);
\draw (-2.55,-0.2) node {$y_{c_-}$};
\draw[dotted] (2.8,0) -- (2.8,2);
\draw (2.75,-0.2) node {$y_{c_+}$};
\draw[dotted] (5.9,-4)-- (5.9,4); \draw (5.9,0.5) node {$a_+$};
\draw (3.5,4.5) node {\small $\wt{W}_+({\rm{a}}_+)=W_-(-1)$};
\draw (3.5,-4.6) node {\small $\wt{W}_-({\rm{a}}_+)=W_+(-1)$};
\end{tikzpicture}
For {\bf{Case 4}}, we let
\begin{eqnarray*}
&\ & D_0\stackrel{def}{=}\big\{c\in[W_+(-1), W_+(1)]\big\},\\
&\ & D_{\ep_0}\stackrel{def}{=}\big\{c=c_r+i\ep, \ c_r\in[W_+(-1), W_+(1)], \ 0<|\ep|<\ep_0\big\},\\
&\ & B_{\ep_0}^{l}\stackrel{def}{=}\big\{c=W_+(-1)+\ep e^{i\theta},\ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},\\
&\ & B_{\ep_0}^{r}\stackrel{def}{=}\big\{c=W_+(1)-\ep e^{i\theta}, \ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},
\end{eqnarray*}
for some $\ep_0\in(0,1).$ We denote
$
\Omega_{\ep_0}\stackrel{def}{=}D_0\cup D_{\ep_0}\cup B_{\ep_0}^{l}\cup B_{\ep_0}^{r}.
$
We also define
\begin{equation*}
c_r=Re c\ for\ c\in D_0\cup D_{\ep_0}, \ c_r=W_+(-1)
\ for \ c\in B_{\ep_0}^{l}, \ c_r=W_+(1) \ for \ c\in B_{\ep_0}^{r}.
\end{equation*}
By Lemma \ref{lem: extend}, we can take a $C^5$ extension of $W_-$ to be $\wt{W}_-$ for $y\in[{\rm{a}}_-,1]$ such that $\wt{W}_-({\rm{a}}_-)=W_+(1)$ and $\wt{W}'_-(y)<0$.
\begin{itemize}
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\geq0$, we denote $y_{c_+}\in[0,1]$ with $y_{c_+}=(W_+)^{-1}(c_r)$, so that $W_+(y_{c_+})-c_r=0$ and $y_{c_-}\in[{\rm{a}}_-,0]$ with $y_{c_-}=(\wt{W}_-)^{-1}(c_r)$, so that $\wt{W}_-(y_{c_-})-c_r=0$.
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\leq0$, we denote $y_{c_+}\in[0,1]$ with $y_{c_+}=(W_-)^{-1}(c_r)$, so that $W_-(y_{c_+})-c_r=0$ and $y_{c_-}\in[-1,0]$ with $y_{c_-}=(W_+)^{-1}(c_r)$, so that $W_+(y_{c_-})-c_r=0$.
\item For $c\in B_{\ep_0}^l$, then $c_r=W_+(-1)=W_-(1)$, we denote $y_{c_+}=1$ and $y_{c_-}=-1$.
\item For $c\in B_{\ep_0}^r$, then $c_r=W_+(1)=\wt{W}_-({\rm{a}}_-)$, we denote $y_{c_+}=1$ and $y_{c_-}={\rm{a}}_-$.
\end{itemize}
We show the relationship between $y_{c_+}, y_{c_-}$ and $c_r$ by the following picture for the above case.
\begin{tikzpicture}[thick, scale=0.5]
\draw[very thin,color=gray];
\draw[->] (-7,0) -- (7,0) node[right] {$y$};
\draw (0.1,-0.01) node[anchor=north] {0}
(5,0) node[anchor=north] {1}
(-4.8,0) node[anchor=north] {-1};
\draw[dotted] (-6,4) -- (5,4);
\draw[dotted] (-5,-4) -- (5,-4);
\draw[dotted] (-5,-4) -- (-5,3);
\draw[dotted] (5,-4) -- (5,4);
\draw[red,thick] (-5,1.8) -- (5,1.8);
\draw[red] (0.3,2.2) node {$c_r$};
\draw[->] (0,-5) -- (0,5) node[above] {$c_r$};
\draw[thick] (0,0) parabola[bend at end] (1,1.2);
\draw[thick] (1.8,2) parabola[bend at end] (1,1.2);
\draw[thick] (1.8,2) parabola[bend at end] (2.6,2.8);
\draw[thick] (3.8,3.2) parabola[bend at end] (2.6,2.8);
\draw[thick] (3.8,3.2) parabola[bend at end] (5,4);
\draw[thick] (0,0) parabola[bend at end] (-1,-1.2);
\draw[thick] (-1.8,-2) parabola[bend at end] (-1,-1.2);
\draw[thick] (-1.8,-2) parabola[bend at end] (-2.6,-2.8);
\draw[thick] (-3.8,-3.2) parabola[bend at end] (-2.6,-2.8);
\draw[thick] (-3.8,-3.2) parabola[bend at end] (-5,-4);
\draw[thick] (0,0) parabola[bend at end] (-1.5,0.7);
\draw[thick] (-2.3,1.2) parabola[bend at end] (-1.5,0.7);
\draw[thick] (-2.3,1.2) parabola[bend at end] (-3.1,2.0);
\draw[thick] (-4,2.5) parabola[bend at end] (-3.1,2.0);
\draw[thick] (-4,2.5) parabola[bend at end] (-5,3);
\draw[red,dotted] (-6,4) parabola[bend at end] (-5,3);
\draw[thick] (0,0) parabola[bend at end] (1.5,-1.1);
\draw[thick] (2.5,-1.9) parabola[bend at end] (1.5,-1.1);
\draw[thick] (2.5,-1.9) parabola[bend at end] (3.1,-2.6);
\draw[thick] (4,-3.1) parabola[bend at end] (3.1,-2.6);
\draw[thick] (4,-3.1) parabola[bend at end] (5,-4);
\draw (5.1,4.4) node {\small $W_+(y)$};
\draw (3.1,-3.3) node {\small $W_-(y)$};
\draw[dotted] (-2.7,0) -- (-2.7,1.8);
\draw (-2.55,-0.5) node {$y_{c_-}$};
\draw[dotted] (1.7,0) -- (1.7,1.8);
\draw (1.8,-0.5) node {$y_{c_+}$};
\draw[dotted] (-6,4)-- (-6,0); \draw (-6,-0.5) node {$a_-$};
\draw (-3.5,4.5) node {\small $\wt{W}_-({\rm{a}}_-)=W_+(1)$};
\draw (3.5,-4.6) node {\small $W_-(1)=W_+(-1)$};
\end{tikzpicture}
We also take a $C^5$ extension of $W_+$ to be $\wt{W}_+$ for $y\in [{\rm{a}}_-,1]$, so that $\wt{W}_+'(y)>0$.
For {\bf{Case 5}}, we let
\begin{eqnarray*}
&\ & D_0\stackrel{def}{=}\big\{c\in[W_-(1), W_+(1)]\big\},\\
&\ & D_{\ep_0}\stackrel{def}{=}\big\{c=c_r+i\ep, \ c_r\in[W_-(1), W_+(1)], \ 0<|\ep|<\ep_0\big\},\\
&\ & B_{\ep_0}^{l}\stackrel{def}{=}\big\{c=W_-(1)+\ep e^{i\theta},\ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},\\
&\ & B_{\ep_0}^{r}\stackrel{def}{=}\big\{c=W_+(1)-\ep e^{i\theta}, \ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},
\end{eqnarray*}
for some $\ep_0\in(0,1).$ We denote
$
\Omega_{\ep_0}\stackrel{def}{=}D_0\cup D_{\ep_0}\cup B_{\ep_0}^{l}\cup B_{\ep_0}^{r}.
$
We also define
\begin{equation*}
c_r=Re\, c\ for\ c\in D_0\cup D_{\ep_0}, \ c_r=W_-(1)
\ for \ c\in B_{\ep_0}^{l}, \ c_r=W_+(1) \ for \ c\in B_{\ep_0}^{r}.
\end{equation*}
By Lemma \ref{lem: extend}, we can take a $C^5$ extension of $W_-$ to be $\wt{W}_-$ and $W_+$ to be $\wt{W}_+$ for $y\in[{\rm{a}}_-,1]$ such that $\wt{W}_-({\rm{a}}_-)=W_+(1)$, $\wt{W}'_-(y)<0$ and $\wt{W}_+({\rm{a}}_-)=W_-(1)$, $\wt{W}'_+(y)>0$.
\begin{itemize}
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\geq0$, we denote $y_{c_+}\in[0,1]$ with $y_{c_+}=(W_+)^{-1}(c_r)$, so that $W_+(y_{c_+})-c_r=0$ and $y_{c_-}\in[{\rm{a}}_-,0]$ with $y_{c_-}=(\wt{W}_-)^{-1}(c_r)$, so that $\wt{W}_-(y_{c_-})-c_r=0$.
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\leq0$, we denote $y_{c_+}\in[0,1]$ with $y_{c_+}=(W_-)^{-1}(c_r)$, so that $W_-(y_{c_+})-c_r=0$ and $y_{c_-}\in[{\rm{a}}_-,0]$ with $y_{c_-}=(\wt{W}_+)^{-1}(c_r)$, so that $\wt{W}_+(y_{c_-})-c_r=0$.
\item For $c\in B_{\ep_0}^l$, then $c_r=W_-(1)=\wt{W}_+({\rm{a}}_-)$, we denote $y_{c_+}=1$ and $y_{c_-}={\rm{a}}_-$.
\item For $c\in B_{\ep_0}^r$, then $c_r=W_+(1)=\wt{W}_-({\rm{a}}_-)$, we denote $y_{c_+}=1$ and $y_{c_-}={\rm{a}}_-$.
\end{itemize}
We show the relationship between $y_{c_+}, y_{c_-}$ and $c_r$ by the following picture for the above case.
\begin{tikzpicture}[thick, scale=0.5]
\draw[very thin,color=gray];
\draw[->] (-7,0) -- (7,0) node[right] {$y$};
\draw(0.1,-0.01) node[anchor=north] {0}
(5,0) node[anchor=north] {1}
(-4.8,0) node[anchor=north] {-1};
\draw[dotted] (-5.8,4) -- (5,4);
\draw[dotted] (-5.8,-4) -- (5,-4);
\draw[dotted] (-5,-3) -- (-5,3);
\draw[dotted] (5,-4) -- (5,4);
\draw[red,thick] (-5,1.8) -- (5,1.8);
\draw[red] (0.3,2.2) node {$c_r$};
\draw[->] (0,-5) -- (0,5) node[above] {$c_r$};
\draw[thick] (0,0) parabola[bend at end] (1,1.2);
\draw[thick] (1.8,2) parabola[bend at end] (1,1.2);
\draw[thick] (1.8,2) parabola[bend at end] (2.6,2.8);
\draw[thick] (3.8,3.2) parabola[bend at end] (2.6,2.8);
\draw[thick] (3.8,3.2) parabola[bend at end] (5,4);
\draw[thick] (0,0) parabola[bend at end] (-1.5,-1.0);
\draw[thick] (-2.5,-1.8) parabola[bend at end] (-1.5,-1.0);
\draw[thick] (-2.5,-1.8) parabola[bend at end] (-3.9,-2.3);
\draw[thick] (-5,-3) parabola[bend at end] (-3.9,-2.3);
\draw[red,dotted] (-5,-3) parabola[bend at end] (-5.8,-4);
\draw[thick] (0,0) parabola[bend at end] (-1.5,0.7);
\draw[thick] (-2.3,1.2) parabola[bend at end] (-1.5,0.7);
\draw[thick] (-2.3,1.2) parabola[bend at end] (-3.1,2.0);
\draw[thick] (-4,2.5) parabola[bend at end] (-3.1,2.0);
\draw[thick] (-4,2.5) parabola[bend at end] (-5,3);
\draw[red,dotted] (-5.8,4) parabola[bend at end] (-5,3);
\draw[thick] (0,0) parabola[bend at end] (1.5,-1.1);
\draw[thick] (2.5,-1.9) parabola[bend at end] (1.5,-1.1);
\draw[thick] (2.5,-1.9) parabola[bend at end] (3.1,-2.6);
\draw[thick] (4,-3.1) parabola[bend at end] (3.1,-2.6);
\draw[thick] (4,-3.1) parabola[bend at end] (5,-4);
\draw (5.1,4.4) node {\small $W_+(y)$};
\draw (6,-4) node {\small $W_-(y)$};
\draw[dotted] (-2.7,0) -- (-2.7,1.8);
\draw (-2.7,-0.5) node {$y_{c_-}$};
\draw[dotted] (1.7,0) -- (1.7,1.9);
\draw (1.8,-0.5) node {$y_{c_+}$};
\draw[dotted] (-5.8,-4)-- (-5.8,4); \draw (-6.2,-0.5) node {${\rm{a}}_-$};
\draw (-3.5,4.5) node {\small $\wt{W}_-({\rm{a}}_-)=W_+(1)$};
\draw (-3.5,-4.6) node {\small $\wt{W}_+({\rm{a}}_-)=W_-(1)$};
\end{tikzpicture}
For {\bf{Case 6}}, we let
\begin{eqnarray*}
&\ & D_0\stackrel{def}{=}\big\{c\in[W_-(1), W_+(1)]\big\},\\
&\ & D_{\ep_0}\stackrel{def}{=}\big\{c=c_r+i\ep, \ c_r\in[W_-(1), W_+(1)], \ 0<|\ep|<\ep_0\big\},\\
&\ & B_{\ep_0}^l\stackrel{def}{=}\big\{c=W_-(1)+\ep e^{i\theta},\ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},\\
&\ & B_{\ep_0}^r\stackrel{def}{=}\big\{c=W_+(1)-\ep e^{i\theta}, \ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},
\end{eqnarray*}
for some $\ep_0\in(0,1).$ We denote
$
\Omega_{\ep_0}\stackrel{def}{=}D_0\cup D_{\ep_0}\cup B_{\ep_0}^l\cup B_{\ep_0}^r.
$
We also define
\begin{equation*}
c_r=Re c\ for\ c\in D_0\cup D_{\ep_0}, \ c_r=W_-(1)
\ for \ c\in B_{\ep_0}^l, \ c_r=W_+(1) \ for \ c\in B_{\ep_0}^r.
\end{equation*}
\begin{itemize}
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\geq0$, we denote $y_{c_+}\in[0,1]$ with $y_{c_+}=(W_+)^{-1}(c_r)$, so that $W_+(y_{c_+})-c_r=0$ and $y_{c_-}\in[-1,0]$ with $y_{c_-}=(W_-)^{-1}(c_r)$, so that $W_-(y_{c_-})-c_r=0$.
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\leq0$, we denote $y_{c_+}\in[0,1]$ with $y_{c_+}=(W_-)^{-1}(c_r)$, so that $W_-(y_{c_+})-c_r=0$ and $y_{c_-}\in[-1,0]$ with $y_{c_-}=(W_+)^{-1}(c_r)$, so that $W_+(y_{c_-})-c_r=0$.
\item For $c\in B_{\ep_0}^l$, then $c_r=W_-(1)=W_+(-1)$, we denote $y_{c_+}=1$ and $y_{c_-}=-1$.
\item For $c\in B_{\ep_0}^r$, then $c_r=W_+(1)=W_-(-1)$, we denote $y_{c_+}=1$ and $y_{c_-}=-1$.
\end{itemize}
We show the relationship between $y_{c_+}$, $y_{c_-}$ and $c_r$ by the following picture.
\begin{tikzpicture}[thick, scale=0.5]
\draw[very thin,color=gray];
\draw[->] (-6,0) -- (6,0) node[right] {$y$};
\draw (0.1,-0.01) node[anchor=north] {0}
(5,0) node[anchor=north] {1}
(-5,0) node[anchor=north] {-1};
\draw[dotted] (-5,4) -- (5,4);
\draw[red,thick] (-5,2) -- (5,2);
\draw[red] (0.3,2.2) node {$c_r$};
\draw[dotted] (-5,-4) -- (5,-4);
\draw[dotted] (-5,-4) -- (-5,4);
\draw[dotted] (5,-4) -- (5,4);
\draw[->] (0,-5) -- (0,5) node[above] {$c_r$};
\draw[thick] (0,0) parabola[bend at end] (1,1.2);
\draw[thick] (1.8,2) parabola[bend at end] (1,1.2);
\draw[thick] (1.8,2) parabola[bend at end] (2.6,2.8);
\draw[thick] (3.8,3.2) parabola[bend at end] (2.6,2.8);
\draw[thick] (3.8,3.2) parabola[bend at end] (5,4);
\draw[thick] (0,0) parabola[bend at end] (-1,-1.2);
\draw[thick] (-1.8,-2) parabola[bend at end] (-1,-1.2);
\draw[thick] (-1.8,-2) parabola[bend at end] (-2.6,-2.8);
\draw[thick] (-3.8,-3.2) parabola[bend at end] (-2.6,-2.8);
\draw[thick] (-3.8,-3.2) parabola[bend at end] (-5,-4);
\draw[thick] (0,0) parabola[bend at end] (-1.5,1.1);
\draw[thick] (-2.5,1.9) parabola[bend at end] (-1.5,1.1);
\draw[thick] (-2.5,1.9) parabola[bend at end] (-3.1,2.6);
\draw[thick] (-4,3.1) parabola[bend at end] (-3.1,2.6);
\draw[thick] (-4,3.1) parabola[bend at end] (-5,4);
\draw[thick] (0,0) parabola[bend at end] (1.5,-1.1);
\draw[thick] (2.5,-1.9) parabola[bend at end] (1.5,-1.1);
\draw[thick] (2.5,-1.9) parabola[bend at end] (3.1,-2.6);
\draw[thick] (4,-3.1) parabola[bend at end] (3.1,-2.6);
\draw[thick] (4,-3.1) parabola[bend at end] (5,-4);
\draw (5.1,4.4) node {\small $W_+(y)$};
\draw (5.1,-5) node {\small $W_-(y)$};
\draw (-2.8,4.4) node {\small $W_+(1)=W_-(-1)$};
\draw (-2.8,-4.6) node {\small $W_-(1)=W_+(-1)$};
\draw[dotted] (-2.55,0) -- (-2.55,2);
\draw (-2.55,-0.5) node {$y_{c_-}$};
\draw[dotted] (1.8,0) -- (1.8,2);
\draw (1.8,-0.5) node {$y_{c_+}$};
\end{tikzpicture}
For {\bf{Case 7}}, we let
\begin{eqnarray*}
&\ & D_0\stackrel{def}{=}\big\{c\in[W_-(1), W_+(1)]\big\},\\
&\ & D_{\ep_0}\stackrel{def}{=}\big\{c=c_r+i\ep, \ c_r\in[W_-(1), W_+(1)], \ 0<|\ep|<\ep_0\big\},\\
&\ & B_{\ep_0}^{l}\stackrel{def}{=}\big\{c=W_-(1)+\ep e^{i\theta},\ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},\\
&\ & B_{\ep_0}^{r}\stackrel{def}{=}\big\{c=W_+(1)-\ep e^{i\theta}, \ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},
\end{eqnarray*}
for some $\ep_0\in(0,1).$ We denote
$
\Omega_{\ep_0}\stackrel{def}{=}D_0\cup D_{\ep_0}\cup B_{\ep_0}^{l}\cup B_{\ep_0}^{r}.
$
We also define
\begin{equation*}
c_r=Re c\ for\ c\in D_0\cup D_{\ep_0}, \ c_r=W_-(1)
\ for \ c\in B_{\ep_0}^{l}, \ c_r=W_+(1) \ for \ c\in B_{\ep_0}^{r}.
\end{equation*}
By Lemma \ref{lem: extend}, we can take a $C^5$ extension of $W_+$ to be $\wt{W}_+$ for $y\in[{\rm{a}}_-,1]$ such that $\wt{W}_+({\rm{a}}_-)=W_-(1)$, $\wt{W}'_+(y)>0$.
\begin{itemize}
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\geq0$, we denote $y_{c_+}\in[0,1]$ with $y_{c_+}=(W_+)^{-1}(c_r)$, so that $W_+(y_{c_+})-c_r=0$ and $y_{c_-}\in[-1,0]$ with $y_{c_-}=(W_-)^{-1}(c_r)$, so that $W_-(y_{c_-})-c_r=0$.
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\leq0$, we denote $y_{c_+}\in[0,1]$ with $y_{c_+}=(W_-)^{-1}(c_r)$, so that $W_-(y_{c_+})-c_r=0$ and $y_{c_-}\in[{\rm{a}}_-,0]$ with $y_{c_-}=(\wt{W}_+)^{-1}(c_r)$, so that $\wt{W}_+(y_{c_-})-c_r=0$.
\item For $c\in B_{\ep_0}^l$, then $c_r=W_-(1)=\wt{W}_+({\rm{a}}_-)$, we denote $y_{c_+}=1$ and $y_{c_-}={\rm{a}}_-$.
\item For $c\in B_{\ep_0}^r$, then $c_r=W_+(1)=W_-(-1)$, we denote $y_{c_+}=1$ and $y_{c_-}=-1$.
\end{itemize}
We show the relationship between $y_{c_+}, y_{c_-}$ and $c_r$ by the following picture for the above case.
\begin{tikzpicture}[thick, scale=0.5]
\draw[very thin,color=gray];
\draw[->] (-7,0) -- (7,0) node[right] {$y$};
\draw(0.1,-0.01) node[anchor=north] {0}
(5,0) node[anchor=north] {1}
(-4.8,0) node[anchor=north] {-1};
\draw[dotted] (-5,4) -- (5,4);
\draw[dotted] (-5.8,-4) -- (5,-4);
\draw[dotted] (-5,-3) -- (-5,4);
\draw[dotted] (5,-4) -- (5,4);
\draw[red,thick] (-5,1.8) -- (5,1.8);
\draw[red] (0.3,2.2) node {$c_r$};
\draw[->] (0,-5) -- (0,5) node[above] {$c_r$};
\draw[thick] (0,0) parabola[bend at end] (1,1.2);
\draw[thick] (1.8,2) parabola[bend at end] (1,1.2);
\draw[thick] (1.8,2) parabola[bend at end] (2.6,2.8);
\draw[thick] (3.8,3.2) parabola[bend at end] (2.6,2.8);
\draw[thick] (3.8,3.2) parabola[bend at end] (5,4);
\draw[thick] (0,0) parabola[bend at end] (-1.5,-1.0);
\draw[thick] (-2.5,-1.8) parabola[bend at end] (-1.5,-1.0);
\draw[thick] (-2.5,-1.8) parabola[bend at end] (-3.9,-2.3);
\draw[thick] (-5,-3) parabola[bend at end] (-3.9,-2.3);
\draw[red,dotted] (-5,-3) parabola[bend at end] (-5.8,-4);
\draw[thick] (0,0) parabola[bend at end] (-1.5,1.1);
\draw[thick] (-2.5,1.9) parabola[bend at end] (-1.5,1.1);
\draw[thick] (-2.5,1.9) parabola[bend at end] (-3.1,2.6);
\draw[thick] (-4,3.1) parabola[bend at end] (-3.1,2.6);
\draw[thick] (-4,3.1) parabola[bend at end] (-5,4);
\draw[thick] (0,0) parabola[bend at end] (1.5,-1.1);
\draw[thick] (2.5,-1.9) parabola[bend at end] (1.5,-1.1);
\draw[thick] (2.5,-1.9) parabola[bend at end] (3.1,-2.6);
\draw[thick] (4,-3.1) parabola[bend at end] (3.1,-2.6);
\draw[thick] (4,-3.1) parabola[bend at end] (5,-4);
\draw (5.1,4.4) node {\small $W_+(y)$};
\draw (6,-4) node {\small $W_-(y)$};
\draw[dotted] (-2.7,0) -- (-2.7,1.8);
\draw (-2.7,-0.5) node {$y_{c_-}$};
\draw[dotted] (1.7,0) -- (1.7,1.9);
\draw (1.8,-0.5) node {$y_{c_+}$};
\draw[dotted] (-5.8,-4)-- (-5.8,0); \draw (-6.2,-0.5) node {${\rm{a}}_-$};
\draw (-3.5,4.5) node {\small $W_-(-1)=W_+(1)$};
\draw (-3.5,-4.6) node {\small $\wt{W}_+({\rm{a}}_-)=W_-(1)$};
\end{tikzpicture}
We also take a $C^5$ extension of $W_-$ to be $\wt{W}_-$ for $y\in [{\rm{a}}_-,1]$, so that $\wt{W}_-'(y)<0$.
For {\bf{Case 8}}, we let
\begin{eqnarray*}
&\ & D_0\stackrel{def}{=}\big\{c\in[W_+(-1), W_-(-1)]\big\},\\
&\ & D_{\ep_0}\stackrel{def}{=}\big\{c=c_r+i\ep, \ c_r\in[W_+(-1), W_-(-1)], \ 0<|\ep|<\ep_0\big\},\\
&\ & B_{\ep_0}^{l}\stackrel{def}{=}\big\{c=W_+(-1)+\ep e^{i\theta},\ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},\\
&\ & B_{\ep_0}^{r}\stackrel{def}{=}\big\{c=W_-(-1)-\ep e^{i\theta}, \ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},
\end{eqnarray*}
for some $\ep_0\in(0,1).$ We denote
$
\Omega_{\ep_0}\stackrel{def}{=}D_0\cup D_{\ep_0}\cup B_{\ep_0}^{l}\cup B_{\ep_0}^{r}.
$
We also define
\begin{equation*}
c_r=Re c\ for\ c\in D_0\cup D_{\ep_0}, \ c_r=W_+(-1)
\ for \ c\in B_{\ep_0}^{l}, \ c_r=W_-(-1) \ for \ c\in B_{\ep_0}^{r}.
\end{equation*}
By Lemma \ref{lem: extend}, we can take a $C^5$ extension of $W_+$ to be $\wt{W}_+$ for $y\in[-1,{\rm{a}}_+]$ such that $\wt{W}_+({\rm{a}}_+)=W_-(-1)$, $\wt{W}'_+(y)>0$.
\begin{itemize}
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\geq0$, we denote $y_{c_+}\in[0,{\rm{a}}_+]$ with $y_{c_+}=(\wt{W}_+)^{-1}(c_r)$, so that $\wt{W}_+(y_{c_+})-c_r=0$ and $y_{c_-}\in[-1,0]$ with $y_{c_-}=(W_-)^{-1}(c_r)$, so that $W_-(y_{c_-})-c_r=0$.
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\leq0$, we denote $y_{c_+}\in[0,1]$ with $y_{c_+}=(W_-)^{-1}(c_r)$, so that $W_-(y_{c_+})-c_r=0$ and $y_{c_-}\in[-1,0]$ with $y_{c_-}=(W_+)^{-1}(c_r)$, so that $W_+(y_{c_-})-c_r=0$.
\item For $c\in B_{\ep_0}^l$, then $c_r=W_+(-1)=W_-(-1)$, we denote $y_{c_+}=1$ and $y_{c_-}=-1$.
\item For $c\in B_{\ep_0}^r$, then $c_r=W_-(-1)=\wt{W}_+({\rm{a}}_+)$, we denote $y_{c_+}={\rm{a}}_+$ and $y_{c_-}=-1$.
\end{itemize}
We show the relationship between $y_{c_+}, y_{c_-}$ and $c_r$ by the following picture for the above case.
\begin{tikzpicture}[thick, scale=0.5]
\draw[very thin,color=gray];
\draw[->] (-7,0) -- (7,0) node[right] {$y$};
\draw (0.1,-0.01) node[anchor=north] {0}
(5,0) node[anchor=north] {1}
(-4.8,0) node[anchor=north] {-1};
\draw[dotted] (-5,4) -- (5.9,4);
\draw[dotted] (-5,-4) -- (5,-4);
\draw[dotted] (-5,-4) -- (-5,4);
\draw[dotted] (5,-4) -- (5,3);
\draw[red,thick] (-5,2) -- (5,2);
\draw[red] (0.3,2.2) node {$c_r$};
\draw[->] (0,-5) -- (0,5) node[above] {$c_r$};
\draw[thick] (0,0) parabola[bend at end] (1.5,1.0);
\draw[thick] (2.5,1.8) parabola[bend at end] (1.5,1.0);
\draw[thick] (2.5,1.8) parabola[bend at end] (3.9,2.3);
\draw[thick] (5,3) parabola[bend at end] (3.9,2.3);
\draw[red,dotted] (5,3) parabola[bend at end] (5.9,4);
\draw[thick] (0,0) parabola[bend at end] (-1,-1.2);
\draw[thick] (-1.8,-2) parabola[bend at end] (-1,-1.2);
\draw[thick] (-1.8,-2) parabola[bend at end] (-2.6,-2.8);
\draw[thick] (-3.8,-3.2) parabola[bend at end] (-2.6,-2.8);
\draw[thick] (-3.8,-3.2) parabola[bend at end] (-5,-4);
\draw[thick] (0,0) parabola[bend at end] (-1.5,1.1);
\draw[thick] (-2.5,1.9) parabola[bend at end] (-1.5,1.1);
\draw[thick] (-2.5,1.9) parabola[bend at end] (-3.1,2.6);
\draw[thick] (-4,3.1) parabola[bend at end] (-3.1,2.6);
\draw[thick] (-4,3.1) parabola[bend at end] (-5,4);
\draw[thick] (0,0) parabola[bend at end] (1.5,-1.1);
\draw[thick] (2.5,-1.9) parabola[bend at end] (1.5,-1.1);
\draw[thick] (2.5,-1.9) parabola[bend at end] (3.1,-2.6);
\draw[thick] (4,-3.1) parabola[bend at end] (3.1,-2.6);
\draw[thick] (4,-3.1) parabola[bend at end] (5,-4);
\draw (4.0,3.0) node {\small $W_+(y)$};
\draw (3.0,-3.5) node {\small $W_-(y)$};
\draw[dotted] (-2.55,0) -- (-2.55,2);
\draw (-2.55,-0.2) node {$y_{c_-}$};
\draw[dotted] (2.8,0) -- (2.8,2);
\draw (2.75,-0.2) node {$y_{c_+}$};
\draw[dotted] (5.9,0)-- (5.9,4); \draw (5.9,-0.5) node {${\rm{a}}_+$};
\draw (3.5,4.5) node {\small $\wt{W}_+({\rm{a}}_+)=W_-(-1)$};
\draw (3.5,-4.6) node {\small $W_-(1)=W_+(-1)$};
\end{tikzpicture}
We also take a $C^5$ extension of $W_-$ to be $\wt{W}_-$ for $y\in [-1, {\rm{a}}_+]$, so that $\wt{W}_-'(y)<0$.
For {\bf{Case 9}}, we let
\begin{eqnarray*}
&\ & D_0\stackrel{def}{=}\big\{c\in[W_-(1), W_-(-1)]\big\},\\
&\ & D_{\ep_0}\stackrel{def}{=}\big\{c=c_r+i\ep, \ c_r\in[W_-(1), W_-(-1)], \ 0<|\ep|<\ep_0\big\},\\
&\ & B_{\ep_0}^{l}\stackrel{def}{=}\big\{c=W_-(1)+\ep e^{i\theta},\ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},\\
&\ & B_{\ep_0}^{r}\stackrel{def}{=}\big\{c=W_-(-1)-\ep e^{i\theta}, \ 0<\ep<\ep_0, \ \frac{\pi}{2}\leq\theta\leq\frac{3\pi}{2}\big\},
\end{eqnarray*}
for some $\ep_0\in(0,1).$ We denote
$
\Omega_{\ep_0}\stackrel{def}{=}D_0\cup D_{\ep_0}\cup B_{\ep_0}^{l}\cup B_{\ep_0}^{r}.
$
We also define
\begin{equation*}
c_r=Re c\ for\ c\in D_0\cup D_{\ep_0}, \ c_r=W_-(1)
\ for \ c\in B_{\ep_0}^{l}, \ c_r=W_-(-1) \ for \ c\in B_{\ep_0}^{r}.
\end{equation*}
By Lemma \ref{lem: extend}, we can take a $C^5$ extension of $W_+$ to be $\wt{W}_+$ for $y\in[{\rm{a}}_-,{\rm{a}}_+]$ such that $\wt{W}_+({\rm{a}}_-)=W_-(1)$, $\wt{W}_+({\rm{a}}_+)=W_-(-1)$ and $\wt{W}'_+(y)>0$.
\begin{itemize}
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\geq0$, we denote $y_{c_+}\in[0,{\rm{a}}_+]$ with $y_{c_+}=(\wt{W}_+)^{-1}(c_r)$, so that $\wt{W}_+(y_{c_+})-c_r=0$ and $y_{c_-}\in[-1,0]$ with $y_{c_-}=(W_-)^{-1}(c_r)$, so that $W_-(y_{c_-})-c_r=0$.
\item For $c\in D_0\cup D_{\ep_0}$ and $c_r\leq0$, we denote $y_{c_+}\in[0,1]$ with $y_{c_+}=(W_-)^{-1}(c_r)$, so that $W_-(y_{c_+})-c_r=0$ and $y_{c_-}\in[{\rm{a}}_-,0]$ with $y_{c_-}=(\wt{W}_+)^{-1}(c_r)$, so that $\wt{W}_+(y_{c_-})-c_r=0$.
\item For $c\in B_{\ep_0}^l$, then $c_r=W_-(1)=\wt{W}_+(a_-)$, we denote $y_{c_+}=1$ and $y_{c_-}={\rm{a}}_-$.
\item For $c\in B_{\ep_0}^r$, then $c_r=W_-(-1)=\wt{W}_+({\rm{a}}_+)$, we denote $y_{c_+}={\rm{a}}_+$ and $y_{c_-}=-1$.
\end{itemize}
We show the relationship between $y_{c_+}, y_{c_-}$ and $c_r$ by the following picture for the above case.
\begin{tikzpicture}[thick, scale=0.45]
\draw[very thin,color=gray];
\draw[->] (-7,0) -- (7,0) node[right] {$y$};
\draw (0.1,-0.01) node[anchor=north] {0}
(5,0) node[anchor=north] {1}
(-4.8,0) node[anchor=north] {-1};
\draw[dotted] (-5,4) -- (6,4);
\draw[dotted] (-6,-4) -- (5,-4);
\draw[dotted] (-5,-3) -- (-5,4);
\draw[dotted] (5,-4) -- (5,3);
\draw[red,thick] (-5,2) -- (5,2);
\draw[red] (0.3,2.2) node {$c_r$};
\draw[->] (0,-5) -- (0,5) node[above] {$c_r$};
\draw[thick] (0,0) parabola[bend at end] (1.5,1.0);
\draw[thick] (2.5,1.8) parabola[bend at end] (1.5,1.0);
\draw[thick] (2.5,1.8) parabola[bend at end] (3.9,2.3);
\draw[thick] (5,3) parabola[bend at end] (3.9,2.3);
\draw[red,dotted] (5,3) parabola[bend at end] (6,4);
\draw[thick] (0,0) parabola[bend at end] (-1.5,-1.0);
\draw[thick] (-2.5,-1.8) parabola[bend at end] (-1.5,-1.0);
\draw[thick] (-2.5,-1.8) parabola[bend at end] (-3.9,-2.3);
\draw[thick] (-5,-3) parabola[bend at end] (-3.9,-2.3);
\draw[red,dotted] (-5,-3) parabola[bend at end] (-6,-4);
\draw[thick] (0,0) parabola[bend at end] (-1.5,1.1);
\draw[thick] (-2.5,1.9) parabola[bend at end] (-1.5,1.1);
\draw[thick] (-2.5,1.9) parabola[bend at end] (-3.1,2.6);
\draw[thick] (-4,3.1) parabola[bend at end] (-3.1,2.6);
\draw[thick] (-4,3.1) parabola[bend at end] (-5,4);
\draw[thick] (0,0) parabola[bend at end] (1.5,-1.1);
\draw[thick] (2.5,-1.9) parabola[bend at end] (1.5,-1.1);
\draw[thick] (2.5,-1.9) parabola[bend at end] (3.1,-2.6);
\draw[thick] (4,-3.1) parabola[bend at end] (3.1,-2.6);
\draw[thick] (4,-3.1) parabola[bend at end] (5,-4);
\draw (4.0,3.2) node {\small $W_+(y)$};
\draw (5.5,-3.5) node {\small $W_-(y)$};
\draw[dotted] (-2.55,0) -- (-2.55,2);
\draw (-2.55,-0.5) node {$y_{c_-}$};
\draw[dotted] (2.8,0) -- (2.8,2);
\draw (2.75,-0.5) node {$y_{c_+}$};
\draw[dotted] (6,4)-- (6,0); \draw (6,-0.5) node {${\rm{a}}_+$};
\draw[dotted] (-6,-4)-- (-6,0); \draw (-6,-0.5) node {${\rm{a}}_-$};
\draw (3.5,4.5) node {\small $\wt{W}_+({\rm{a}}_+)=W_-(-1)$};
\draw (-3.5,-4.6) node {\small $\wt{W}_+({\rm{a}}_-)=W_-(1)$};
\end{tikzpicture}
We also take a $C^5$ extension of $W_-$ to be $\wt{W}_-$ for $y\in [{\rm{a}_-}, {\rm{a}}_+]$, so that $\wt{W}_-'(y)<0$.
In the last step of Case 1, 2, 3, 7, 8, 9, we only restrict the regularity and monotonicity of the extension.
\noindent{\bf Notations:}
We summarize the above nine cases and make the following notations.
We denote $a_+={\rm{a}}_+$ if we need to extend the definition of $W_{+}(y)$ or $W_{-}(y)$ for $y\geq 1$ and $a_+=1$ if we do not need to extend the definition of $W_{+}(y)$ nor $W_{-}(y)$ for $y\geq 1$. Similarly we also denote $a_-={\rm{a}}_-$ if we need to extend the definition of $W_{+}(y)$ or $W_{-}(y)$ for $y\leq -1$ and $a_-=-1$ if we do not need to extend the definition of $W_{+}(y)$ nor $W_{-}(y)$ for $y\leq -1$.
By letting $u(y)=\f{\wt{W}_+(y)+\wt{W}_-(y)}{2}$ and $b(y)=\f{\wt{W}_+(y)-\wt{W}_-(y)}{2}$, we extend $u$ and $b$. We also take a $C^3$ extension of $\widehat{\phi}_0(\al,y)$ for $y\in (1,a_+]\cup [a_-,-1)$ and extend $\widehat{\om}_0(\al,y)=0$ for $y\in (1,a_+] \cup [a_-,-1)$ then $cG(\al, y,c)\in L^{\infty}([a_-,a_+]\times \Om_{\ep_0})$ and $\|cG\|_{L^{\infty}}\leq C(\|\widehat{\om}_0\|_{H^1_y}+C\|\widehat{j}_0\|_{H^3_y})$.
We also use $W_{\pm}$ to represent $\wt{W}_{\pm}$, whether they are extended or not. For $\ep_0>0$, we let
\begin{eqnarray*}
&&D_0=[\min\{W_-(1),W_+(-1)\}, \max\{W_+(1),W_-(-1)\}], \\
&&D_{\ep_0}=\{z=c+i\ep:~c\in D_0, 0<|\ep|<\ep_0\},\\
&&B_{\ep_0}^l=\{z=\min\{W_-(1),W_+(-1)\}+\ep e^{i\th},0<\ep<\ep_0, \f{\pi}{2}\leq \th\leq \f{3\pi}{2}\},\\
&&B_{\ep_0}^r=\{z=\max\{W_+(1),W_-(-1)\}-\ep e^{i\th}, 0<\ep<\ep_0, \f{\pi}{2}\leq \th\leq \f{3\pi}{2}\},
\end{eqnarray*}
and $\Omega_{\ep_0}\stackrel{def}{=}D_0\cup D_{\ep_0}\cup B_{\ep_0}^{l}\cup B_{\ep_0}^{r}$.
We let $\mathcal{H}(y,c)=({W}_+(y)-c)({W}_-(y)-c)$ be well defined on $[a_-, a_+]\times \Om_{\ep_0}$. Let $d_+=[0,a_+]$ and $d_-=[a_-,0]$.
\subsection{Sturmian integral operator.}
Given $|\al|\geq 1$, let $A$ be a constant larger
than $C|\al|$ with $C\geq1$ independent of $\al$.
\begin{definition}\label{def: X+ norm}
For a function $f(y,c)$ defined on $d_{+}\times \Omega_{\ep_0}$ or $d_{-}\times \Omega_{\ep_0}$, we define
\begin{equation*}
\begin{split}
&\|f\|_{X_0^{\pm}}\stackrel{def}{=}\sup_{(y,c)\in d_{\pm}
\times D_0}\Big|\frac{f(y,c)}{\cosh(A(y-y_{c_{\pm}}))}\Big|,\\
&\|f\|_{X^{\pm}}\stackrel{def}{=}\sup_{(y,c)\in d_{\pm}
\times D_{\ep_0}\cup D_0}\Big|\frac{f(y,c)}{\cosh(A(y-y_{c_+}))}\Big|,\\
&\|f\|_{X_l^{\pm}}\stackrel{def}{=}\sup_{(y,c)\in d_{\pm}
\times B_{\ep_0}^{l}}\Big|\frac{f(y,c)}{\cosh(A(y-y_{c_{\pm}}))}\Big|,\\
&\|f\|_{X_r^{\pm}}\stackrel{def}{=}\sup_{(y,c)\in d_{\pm}\times B_{\ep_0}^{r}}\Big|\frac{f(y,c)}{\cosh(A(y-y_{c_{\pm}}))}\Big|,
\end{split}
\end{equation*}
where $c_r$ and $y_{c_{\pm}}$ were defined in the previous section satisfies $\mathcal{H}(y,c_r)=0$.
\end{definition}
\begin{definition}
For a function $f(y,c)$ defined on $d_+\times \Omega_{\ep_0}$ or $d_-\times \Omega_{\ep_0}$, we define
\begin{equation*}
\begin{split}
&\|f\|_{Y^{\pm}}\stackrel{def}{=}\|f\|_{X^{\pm}}
+\frac{1}{A}\big(\|\pa_yf\|_{X^{\pm}}+\|\pa_{c_r}f\|_{X^{\pm}}+\|\pa_{\ep}f\|_{X^{\pm}}\big),\\
&\|f\|_{Y_l^{\pm}}\stackrel{def}{=}\|f\|_{X_l^{\pm}}
+\frac{1}{A}\big(\|\pa_yf\|_{X_l^{\pm}}+\|\pa_{\ep}f\|_{X_l^{\pm}}+\|\pa_{\theta}f\|_{X_l^{\pm}}\big),\\
&\|f\|_{Y_r^{\pm}}\stackrel{def}{=}\|f\|_{X_r^{\pm}}
+\frac{1}{A}\big(\|\pa_yf\|_{X_r^{\pm}}+\|\pa_{\ep}f\|_{X_r^{\pm}}+\|\pa_{\theta}f\|_{X_r^{\pm}}\big),\\
&\|f\|_{Y_0^{\pm}}\stackrel{def}{=}\|f\|_{X_0^{\pm}}+\frac{1}{A}\big(\|\pa_yf\|_{X_0^{\pm}}
+\|\pa_cf\|_{X_0^{\pm}}\big)+\frac{1}{A^2}\|\pa_y\pa_cf\|_{X_0^{\pm}}.
\end{split}
\end{equation*}
\end{definition}
Now, we introduce the Sturmian integral operator, which will be used to give the solution formula of the homogeneous Sturmian equation.
\begin{definition}
Let $y\in d_+$ or $y\in d_-$, the Sturmian integral operator $S^{\pm}$ is defined by
\begin{equation*}
S^{\pm}f(y,c)\stackrel{def}{=}S_0^{\pm}\circ S_1^{\pm}f(y,c)=\int_{y_{c_{\pm}}}^y
\frac{\int_{y_{c_{\pm}}}^{y'}\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy',
\end{equation*}
where
\begin{equation*}
\begin{split}
&S_0^{\pm}f(y,c)\stackrel{def}{=}\int_{y_{c_{\pm}}}^yf(y',c)dy',\\
&S_1^{\pm}f(y,c)\stackrel{def}{=}
\frac{\int_{y_{c_{\pm}}}^{y}\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y)-c\big)\big(W_-(y)-c\big)}.
\end{split}
\end{equation*}
\end{definition}
\begin{proposition}\label{prop: S+ estimate}
For $y\in d_+$, there exists a constant $C_1$ independent of $A$ so that
\begin{equation*}
\begin{split}
&\|S^+f\|_{Y_0^+}\leq \frac{C_1}{A^2}\|f\|_{Y_0^+}, \
\|S^+f\|_{Y^+}\leq \frac{C_1}{A^2}\|f\|_{Y^+},\\
&\|S^+f\|_{Y_l^+}\leq \frac{C_1}{A^2}\|f\|_{Y_l^+},\
\|S^+f\|_{Y_r^+}\leq \frac{C_1}{A^2}\|f\|_{Y_r^+}.
\end{split}
\end{equation*}
Moreover, if $f\in C\big(d_+\times\Om_{\ep_0}\big)$, then
\begin{equation*}
S_0^+f,\ S_1^+f,\ S^+f\in C\big(d_+\times \Om_{\ep_0}\big).
\end{equation*}
\end{proposition}
\begin{proposition}\label{prop: S- estimate}
For $y\in d_-$, there exists a constant $C_1$ independent of $A$ so that
\begin{equation*}
\begin{split}
&\|S^-f\|_{Y_0^-}\leq \frac{C_1}{A^2}\|f\|_{Y_0^-},\
\|S^-f\|_{Y^-}\leq \frac{C_1}{A^2}\|f\|_{Y^-},\\
&
\|S^-f\|_{Y_l^-}\leq \frac{C_1}{A^2}\|f\|_{Y_l^-},\
\|S^-f\|_{Y_r^-}\leq \frac{C_1}{A^2}\|f\|_{Y_r^-}.
\end{split}
\end{equation*}
Moreover, if $f\in C\big(d_-\times\Om_{\ep_0}\big)$, then
\begin{equation*}
S_0^-f, \ S_1^-f, \ S^-f\in C\big(d_-\times \Om_{\ep_0}\big).
\end{equation*}
\end{proposition}
In the following, we only give the proof of the Proposition \ref{prop: S+ estimate} and Proposition \ref{prop: S- estimate} can be obtained by the same method.
\begin{proof}
By the fact that $W'_+(y)>0$, $W'_-(y)<0$ for $0\leq y\leq a_+$ and $W_+(0)=W_-(0)=0$, we have for $c_r\geq 0$, then
$W_+(y_{c_+})=c_r$, we can conclude that for $c\in \Omega_{\ep_0}$,
\begin{equation*}
\textrm{if} \ 0<y_{c_+}<y'<y<a_+ \ \textrm{or}\ 0<y<y'<y_{c_+}<a_+, \ \textrm{then} \
\Big|\frac{W_+(y')-c}{W_+(y)-c}\Big|\leq 1,
\end{equation*}
and
\begin{align*}
\textrm{if} \ 0<y_{c_+}<y'<y<a_+, \ \textrm{then}\
\Big|\frac{W_-(y')-c}{W_-(y)-c}\Big|\leq 1,\\
\textrm{if} \ 0<y<y'<y_{c_+}<a_+, \ \textrm{then}\
\Big|\frac{W_-(y')-c}{W_-(y)-c}\Big|
\leq C.
\end{align*}
And for the case $c_r\leq 0$, then $W_-(y_{c_+})=c_r$, we can conclude that for $c\in \Omega_{\ep_0}$,
\begin{equation*}
\textrm{if} \ 0<y_{c_+}<y'<y<a_+ \ or\ 0<y<y'<y_{c_+}<a_+, \ \textrm{then} \
\Big|\frac{W_-(y')-c}{W_-(y)-c}\Big|\leq 1,
\end{equation*}
and
\begin{align*}
\textrm{if} \ 0<y_{c_+}<y'<y<a_+, \ \textrm{then} \
\Big|\frac{W_+(y')-c}{W_+(y)-c}\Big|\leq 1,\\
\textrm{if} \ 0<y<y'<y_{c_+}<a_+, \ \textrm{then}\
\Big|\frac{W_+(y')-c}{W_+(y)-c}\Big|
\leq C.
\end{align*}
Thus we have that for $(y,c)\in[0,a_+]\times \Om_{\ep_0}$ and $ 0<y_{c_+}<y'<y<a_+$ or $0<y<y'<y_{c_+}<a_+$,
\begin{equation}\label{eq: W mono}
\Big|\frac{W_+(y')-c}{W_+(y)-c}\Big|\leq C,\quad
\Big|\frac{W_-(y')-c}{W_-(y)-c}\Big|\leq C.
\end{equation}
A direct calculation shows that for $(y,c)\in[0,a_+]\times D_0$
\begin{equation}\label{eq: S_0f}
\begin{split}
\|S_0^+f\|_{X_0^+}
&=\sup_{(y,c)\in[0,a_+]\times D_0}
\Big|\frac{1}{\cosh A(y-y_{c_+})}\int_{y_{c_+}}^y
\frac{f(z,c)}{\cosh A(y-y_{c_+})}\cosh A(z-y_{c_+})dz\Big|\\
&\leq \sup_{(y,c)\in[0,a_+]\times D_0}\Big|\frac{1}{\cosh A(y-y_{c_+})}\int_{y_{c_+}}^y\cosh A(z-y_{c_+})dz\Big|\|f\|_{X_0^+}\leq \frac{1}{A}\|f\|_{X_0^+}.
\end{split}
\end{equation}
Then we deduce
\begin{equation}\label{eq: S_11f}
\begin{split}
\Big\|S_1^+f(y,c)\Big\|_{X_0^+}
&\leq C\sup_{(y,c)\in[0,a_+]\times D_0}
\Big|\frac{y-y_{c_+}}{\cosh A(y-y_{c_+})}\int_0^1\cosh tA(y-y_{c_+})dt
\Big|\|f\|_{X_0^+}\\
&\leq \frac{C}{A}\|f\|_{X_0^+},
\end{split}
\end{equation}
which along with (\ref{eq: S_0f}) shows that
\begin{equation}\label{eq: Sf}
\|S^+f\|_{X_0^+}\leq \frac{C}{A^2}\|f\|_{X_0^+}.
\end{equation}
By using (\ref{eq: W mono}), we obtain
\begin{eqnarray}\label{0}
&\ & \left\|\frac{\int_{y_{c_+}}^{y}(W_-(y')-c)f(y',c)dy'}
{(W_+(y)-c)(W_-(y)-c)}\right\|_{X_0^+}
+\left\|\frac{\int_{y_{c_+}}^{y}(W_+(y')-c)(W_-(y')-c)f(y',c)dy'}
{(W_+(y)-c)^2(W_-(y)-c)}\right\|_{X_0^+}\nonumber\\
&\ &\leq C\sup_{(y,c)\in[0,1]\times D_0}\Big|\frac{\int_0^1\cosh tA(y-y_{c_+})dt}
{\cosh A(y-y_{c_+})\int_0^1W_+^{'}(y_{c_+}+t(y-y_{c_+}))dt}
\Big|\|f\|_{X_0^+}\nonumber\\
&\ & \leq C\|f\|_{X_0^+},
\end{eqnarray}
and
\begin{eqnarray}\label{00}
&\ & \left\|\frac{\int_{y_{c_+}}^{y}(W_+(y)-c)f(y',c)dy'}
{(W_+(y)-c)(W_-(y)-c)}\right\|_{X_0^+}
+\left\|\frac{\int_{y_{c_+}}^{y}(W_+(y')-c)(W_-(y')-c)f(y',c)dy'}
{(W_+(y)-c)(W_-(y)-c)^2}\right\|_{X_0^+}\nonumber\\
&\ & \leq C\sup_{(y,c)\in[0,a_+]\times D_0}\Big|\frac{1}{\cosh A(y-y_{c_+})}
\int_0^1\cosh tA(y-y_{c_+})dt
\Big|\|f\|_{X_0^+}
\nonumber\\
&\ & \leq C\|f\|_{X_0^+}.
\end{eqnarray}
Similarly, we also obtain that
\begin{eqnarray}\label{8}
\|S_0^+f\|_{Z^+}\leq \frac{C}{A}\|f\|_{Z^+}, \ \ \|S_1^+f\|_{Z^+}\leq \frac{C}{A}\|f\|_{Z^+},
\end{eqnarray}
\begin{equation}\label{88}
\begin{split}
&\left\|\frac{\int_{y_{c_+}}^{y}(W_-(y')-c)f(y',c)dy'}
{(W_+(y)-c)(W_-(y)-c)}\right\|_{Z^+}
+\left\|\frac{\int_{y_{c_+}}^{y}(W_+(y')-c)(W_-(y')-c)f(y',c)dy'}
{(W_+(y)-c)^2(W_-(y)-c)}\right\|_{Z^+}\\
&\leq C\|f\|_{Z^+}
\end{split}
\end{equation}
and
\begin{equation}\label{888}
\begin{split}
&\left\|\frac{\int_{y_{c_+}}^{y}(W_+(y)-c)f(y',c)dy'}
{(W_+(y)-c)(W_-(y)-c)}\right\|_{Z^+}
+\left\|\frac{\int_{y_{c_+}}^{y}(W_+(y')-c)(W_-(y')-c)f(y',c)dy'}
{(W_+(y)-c)(W_-(y)-c)^2}\right\|_{Z^+}\\
&\leq C\|f\|_{Z^+},
\end{split}
\end{equation}
here $Z^+$ can be taken as $X^+, X_l^+, X_r^+$. \\
A direct calculation shows that for $c\in D_0$,
\begin{equation*}
\partial_yS^+f(y,c)=S_1^+f(y,c),
\end{equation*}
and
\begin{equation}\label{Sc+}
\begin{split}
\partial_cS^+f(y,c)&=-\int_{y_{c_+}}^y
\frac{\int_{y_{c_+}}^{y'}\big(W_-(z)-c)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'
-\int_{y_{c_+}}^y
\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'\nonumber\\
&\ \
+\int_{y_{c_+}}^y
\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)^2\big(W_-(y')-c\big)}dy'\nonumber\\
&\ \
+\int_{y_{c_+}}^y
\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)^2}dy'\nonumber\\
&\ \
+\int_{y_{c_+}}^y
\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)\big(W_-(z)-c\big)\pa_cf(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy',
\end{split}
\end{equation}
and
\begin{equation*}
\begin{split}
\partial_{cy}S^+f(y,c)
&=-\frac{\int_{y_{c_+}}^{y}\big(W_-(y')-c\big)f(y',c)dy'}
{\big(W_+(y)-c\big)\big(W_-(y)-c\big)}
-\frac{\int_{y_{c_+}}^{y}\big(W_+(y')-c\big)f(y',c)dy'}
{\big(W_+(y)-c\big)\big(W_-(y)-c\big)}\\
&\ \
+\frac{\int_{y_{c_+}}^{y}\big(W_+(y')-c\big)\big(W_-(y')-c\big)f(y',c)dy'}
{\big(W_+(y)-c\big)^2\big(W_-(y)-c\big)}\\
&\ \
+\frac{\int_{y_{c_+}}^{y}\big(W_+(y')-c\big)\big(W_-(y')-c\big)f(y',c)dy'}
{\big(W_+(y)-c\big)\big(W_-(y)-c\big)^2}\\
&\ \
+\frac{\int_{y_{c_+}}^{y}\big(W_+(y')-c\big)\big(W_-(y')-c\big)\pa_cf(y',c)dy'}
{\big(W_+(y)-c\big)\big(W_-(y)-c\big)}.
\end{split}
\end{equation*}
Thus, from (\ref{eq: Sf}), (\ref{0}) and (\ref{00}), we obtain
\begin{align*}
\|S^+f\|_{Y_0^+}&=\|S^+f\|_{X_0^+}+\f{1}{A}\|\pa_yS^+f\|_{X_0^+}
+\f{1}{A}\|\pa_cS^+f\|_{X_0^+}+\f{1}{A^2}
\|\pa_{yc}S^+f\|_{X_0^+}\\
&\leq C\f{1}{A^2}\|f\|_{X_0^+}+\frac{C}{A^3}\|\pa_cf\|_{X_0^+}
\leq \f{C_1}{A^2}\|f\|_{Y_0^+}.
\end{align*}
For $c\in D_{\ep_0}$, we have
\begin{eqnarray*}
\pa_yS^+f(y,c)=S_1^+f(y,c),
\end{eqnarray*}
and
\begin{equation*}
\begin{split}
\pa_{\ep}S^+f(y,c)&
=S^+\pa_{\ep}f(y,c)+
i\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)^2\big(W_-(y')-c\big)}dy'\\
&\quad
+i\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)^2}dy'\\
&\quad
-i\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'
-i\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'.
\end{split}
\end{equation*}
If $c_r\geq 0$, then $W_+(y_{c_+})=c_r$, we obtain
\begin{equation*}
\begin{split}
\pa_{c_r}S^+f(y,c)
&=S^+\pa_{c_r}f(y,c)+
\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)^2\big(W_-(y')-c\big)}dy'\\
&\quad
+\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)^2}dy'\\
&\quad
-\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'
-\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'\\
&\quad
+i(W_+^{-1})'(c_r)f(y_{c_+},c)\big(W_-(y_{c_+})-c\big)
\int_{y_{c_+}}^y\frac{\ep}{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy',
\end{split}
\end{equation*}
and if $c_r\leq 0$, then $W_-(y_{c_+})=c_r$ and we have
\begin{equation*}
\begin{split}
\pa_{c_r}S^+f(y,c)
&=S^+\pa_{c_r}f(y,c)+
\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)^2\big(W_-(y')-c\big)}dy'\\
& \quad
+\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)^2}dy'\\
& \quad
-\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'
-\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'\\
& \quad
+i(W_-^{-1})'(c_r)f(y_{c_+},c)\big(W_+(y_{c_+})-c\big)
\int_{y_{c_+}}^y\frac{\ep}{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'.
\end{split}
\end{equation*}
Due to (\ref{eq: W mono}), we have
\begin{equation*}
\left|\frac{\big(W_+(y_{c_+})-c\big)}{\cosh A(y-y_{c_+})}
\int_{y_{c_+}}^y\frac{\ep}{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'\right|
\leq \frac{C_1|y-y_{c_+}|}{\cosh A(y-y_{c_+})}\leq \frac{C_1}{A},
\end{equation*}
\begin{equation*}
\left|\frac{\big(W_-(y_{c_+})-c\big)}{\cosh A(y-y_{c_+})}
\int_{y_{c_+}}^y\frac{\ep}{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'\right|
\leq \frac{C_1|y-y_{c_+}|}{\cosh A(y-y_{c_+})}\leq \frac{C_1}{A},
\end{equation*}
Then from which and (\ref{8})-(\ref{888}), we get
\begin{equation}
\|S^+f\|_{Y^+}\leq \frac{C_1}{A^2}\|f\|_{Y^+}.
\end{equation}
For $c\in B_{\ep_0}^l$, we have
\begin{equation*}
\pa_yS^+f(y,c)=S_1^+f(y,c),
\end{equation*}
and
\begin{equation*}
\begin{split}
\pa_{\ep}S^+f(y,c)&=S^+\pa_{\ep}f(y,c)+
e^{i\theta}\Big\{\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}
\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)^2\big(W_-(y')-c\big)}dy'\\
& \ \
+\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)^2}dy'\\
& \ \
-\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'
-\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'\Big\},
\end{split}
\end{equation*}
and
\begin{equation*}
\begin{split}
\pa_{\theta}S^+f(y,c)&=S^+\pa_{\theta}f(y,c)+
i\ep e^{i\theta}\Big\{\int_{y_{c_+}}^y
\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)^2\big(W_-(y')-c\big)}dy'\\
& \ \
+\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)^2}dy'\\
& \ \
-\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_+(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'
-\int_{y_{c_+}}^y\frac{\int_{y_{c_+}}^{y'}\big(W_-(z)-c\big)f(z,c)dz}
{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}dy'\Big\}.
\end{split}
\end{equation*}
Then we can get that by (\ref{8})-(\ref{888})
\begin{equation}
\|S^+f\|_{Y_l^+}\leq \frac{C_1}{A^2}\|f\|_{Y_l^+}.
\end{equation}
For $c\in B_{\ep_0}^r$, the proof is similar, we omit it for the sake of brevity.
Now we check the continuity. We rewrite $S_1^+f$ as
\begin{eqnarray*}
S_1^+f=\int_0^1K(t,y,c)f(y_{c_+}+t(y-y_{c_+}),c)dt,
\end{eqnarray*}
with $K(t,y,c)=\frac{(y-y_{c_+})\big(W_+(y_{c_+}+t(y-y_{c_+}))-c\big)\big(W_-(y_{c_+}+t(y-y_{c_+}))-c\big)}{(W_+(y)-c)(W_-(y)-c)}$.
Using the fact that for $(y,c)\in d_+\times\Om_{\ep_0}$ with $0<y<y'<y_{c_+}$ or $y_{c_+}<y'<y<a_+$, $\Big|\frac{\big(W_+(y')-c\big)\big(W_-(y')-c\big)}{\big(W_+(y)-c\big)
\big(W_-(y)-c\big)}\Big|\leq C$,
the continuity of $K(t,y,c)$ and the Lebesgue's dominated convergence theorem, we conclude the continuity of $S_1^+f$. The continuity of $S^+f$ follows from $S^+f=S_0^+\circ S_1^+f$.
\end{proof}
\subsection{Existence of the solution}
In the following, the constant $C$ may depend on $\al$.
The homogeneous Sturmian equation on $[0,a_+]$ is
\begin{equation}\label{eq: homo eq}
\Big\{\begin{array}{l}
\pa_y\Big(\mathcal{H}(y,c)\pa_y\va_+(y,c)\Big)=\al^2\mathcal{H}(y,c)\va_+(y,c),\\
\va_+(y_{c_+},c)=1, \ \pa_y\va_+(y_{c_+},c)=0.
\end{array}\Big.
\end{equation}
\begin{proposition}\label{prop: sol. hom. [0,1]}
1. For $c\in \Omega_{\ep_0}$, there exists a solution
$\va_+(y,c)\in C([0,a_+]\times \Omega_{\ep_0})$ of the
Sturmian equation \eqref{eq: homo eq} and $\pa_y\va_+(y,c)\in C([0,a_+]\times \Om_{\ep_0})$.
Moreover, there exists $\ep_1>0$ such that for any $\ep_0\in[0,\ep_1)$
and $(y,c)\in [0,a_+]\times \Omega_{\ep_0}$,
\begin{eqnarray*}
|\va_+(y,c)|\geq \frac{1}{2}, \ \ |\va_+(y,c)-1|\leq C|y-y_{c_+}|^2,
\end{eqnarray*}
where the constants $\ep_1, C$ may depend on $\al$.\\
2. For $c\in D_0$, for any $y\in[0,a_+]$, there is a constant $C$(depends on $\al$) such that,
\begin{eqnarray*}
\va_+(y,c)\geq \va_+(y',c)\geq 1, \ \ for \ \ 0\leq y_{c_+}\leq y'\leq y\leq 1
\ \ or\ \ 0\leq y\leq y'\leq y_{c_+}\leq 1;
\end{eqnarray*}
\begin{eqnarray*}
&\ &0\leq \va_+(y,c)-1\leq C\min\big\{\al^2(y-y_{c_+})^2, 1\big\}\va_+(y,c),
\end{eqnarray*}
\begin{eqnarray*}
C^{-1} |y-y_{c_+}|\leq |\pa_y\va_+(y,c)|\leq C |y-y_{c_+}|,
\end{eqnarray*}
and
\begin{eqnarray*}
|\pa_c\va_+(y,c)|\leq C |y-y_{c_+}|, \ \
|\pa_y\pa_c\va_+(y,c)|\leq C .
\end{eqnarray*}
\end{proposition}
The proof is based on the following lemmas.
\begin{lemma}\label{lem: Y+}
Let $c\in D_{\ep_0}$.
Then there exists a solution $\va_+(y,c)\in Y^+$ to the Sturmian equation \eqref{eq: homo eq}. Moreover, it holds
\begin{eqnarray*}
\|\va_+\|_{Y^+}\leq C.
\end{eqnarray*}
\end{lemma}
\begin{proof}
$\va_+$ satisfies
\begin{equation*}
\pa_y\Big(\mathcal{H}(y,c)\pa_y\va_+(y,c)\Big)=\al^2\mathcal{H}(y,c)\va_+(y,c),
\end{equation*}
from which, we infer that
\begin{equation*}
\va_+(y,c)=1+\int_{y_{c_+}}^y\frac{\al^2}{\mathcal{H}(y',c)}
\int_{y_{c_+}}^{y'}\mathcal{H}(z,c)\va_+(z,c)dzdy'.
\end{equation*}
This means that $\va_+$ satisfies
\begin{equation*}
\va_+(y,c)=1+\al^2S^+\va_+(y,c).
\end{equation*}
In follows from Proposition \ref{prop: S+ estimate} that the
operator $I-\al^2 S^+$ is invertible in the space $Y^+$, if
\begin{equation*}
\frac{\al^2C_1}{A^2}\leq \frac{1}{2}<1,
\end{equation*}
where $C_1$ is the constant in Proposition \ref{prop: S+ estimate}. Thus
\begin{equation*}
\va_+(y,c)=(I-\al^2S^+)^{-1}1.
\end{equation*}
Hence, $\|\va_+\|_{Y^+}\leq \|1\|_{Y^+}+\|\al^2S^+\va_+\|_{Y^+}
\leq C+\frac{1}{2}\|\va_+\|_{Y^+}$ implies $\|\va_+\|_{Y^+}\leq C$,
here $C$ is a constant independent of $A$ and $\al$.
\end{proof}
In a similar way as in Lemma \ref{lem: Y+}, we can show that
\begin{lemma}\label{lem: Y+l}
Let $c\in B_{\ep_0}^l$. Then there exists a solution
$\va_+(y,c)\in Y^+_l$ to the Sturmian equation \eqref{eq: homo eq}.
Moreover, it holds
\begin{equation*}
\|\va_+\|_{Y^+_l}\leq C.
\end{equation*}
\end{lemma}
\begin{lemma}\label{lem: Y+r}
Let $c\in B_{\ep_0}^r$. Then there exists a solution $\va_+(y,c)\in Y^+_r$
to the Sturmian equation \eqref{eq: homo eq}. Moreover, it holds
\begin{equation*}
\|\va_+\|_{Y^+_r}\leq C.
\end{equation*}
\end{lemma}
\begin{lemma}\label{lem: Y+0}
Let $c\in D_0$.
Then there exists a solution $\va_+(y,c)\in Y^+_0$ to the Sturmian equation \eqref{eq: homo eq}. Moreover, it holds
\begin{equation*}
\|\va_+\|_{Y^+_0}\leq C.
\end{equation*}
\end{lemma}
Now we are in a position to prove the Proposition \ref{prop: sol. hom. [0,1]}.
\begin{proof}
{\bf Proof of 1. }
Let us define
\begin{equation*}
\va_+(y,c)\stackrel{def}{=}\left\{\begin{array}{l}
\va_+^0(y,c) \ \ for \ \ c\in D_0,\\
\va_+^{\pm}(y,c) \ \ for \ \ c\in D_{\ep_0},\\
\va_+^l(y,c) \ \ for \ \ c\in B_{\ep_0}^l,\\
\va_+^r(y,c) \ \ for \ \ c\in B_{\ep_0}^r,\\
\end{array}\right.
\end{equation*}
where $\va_+^{\pm}, \va_+^l, \va_+^r, \va_+^0$ are given
by Lemma \ref{lem: Y+}, Lemma \ref{lem: Y+l}, Lemma \ref{lem: Y+r} and
Lemma \ref{lem: Y+0} respectively. Then $\va_+(y,c)$ is our desired solution.
By Proposition \ref{prop: S+ estimate} and using the formula
$\va_+(y,c)=\sum\limits_{k=0}^{+\infty}\al^{2k}(S^+)^k1$ for $\Omega_{\ep_0}$, we can conclude
that $\va_+(y,c)\in C([0,a_+]\times \Omega_{\ep_0})$.
Moreover, for $c\in D_0$, we have
\begin{eqnarray*}
(S^+)^k1(y,c)\geq 0, \ \ \va_+(y,c)\geq 1,
\end{eqnarray*}
which ensures that there exists $\ep_1>0$ so that for
any $\ep_0\in[0,\ep_1)$ and $(y,c)\in d_+\times \Om_{\ep_0}$,
\begin{eqnarray*}
|\va_+(y,c)|\geq \frac{1}{2}, \ \ |\va_+(y,c)|\leq C.
\end{eqnarray*}
Thanks to $\va_+=1+\al^2S^+\va_+$, we have
\begin{eqnarray}\label{va+}
\va_+(y,c)=1+\int_{y_{c_+}}^y
\frac{\al^2}{\mathcal{H}(y',c)}\int_{y_{c_+}}^{y'}\mathcal{H}(z,c)\va_+(z,c)dzdy',
\end{eqnarray}
from which, it follows that
\begin{eqnarray*}
\pa_y\Big(\mathcal{H}(y,c)\pa_y\va_+(y,c)\Big)=\al^2\mathcal{H}(y,c)\va_+(y,c),\ \ \va_+(y_{c_+},c)=1.
\end{eqnarray*}
Then $\va_+(y,c)$ satisfies the Sturmian equation \eqref{eq: homo eq}.
By the fact that $\pa_y\va_+(y,c)=\al^2S_1^+\va_+(y,c)$ and Proposition \ref{prop: S+ estimate}, we have $\pa_y\va_+(y,c)\in C(d_+\times \Om_{\ep_0})$.
From (\ref{va+}), we have
\begin{equation*}
|\va_+(y,c)-1|\leq \al^2\int_{y_{c_+}}^y\int_{y_{c_+}}^{y'}|\va_+(z,c)|
\Big|\frac{(W_+(z)-c)(W_-(z)-c)}{(W_+(y')-c)(W_-(y')-c)}\Big|dzdy'\leq C|y-y_{c_+}|^2.
\end{equation*}
\noindent{\bf Proof of 2. } Since we have $\pa_y\va_+(y,c)=\al^2S_1^+\va_+(y,c)$, thus for $y\geq z\geq y_{c_+}$ or $y\leq z\leq y_{c_+}$, $\frac{\mathcal{H}(z,c)}
{\mathcal{H}(y,c)}\geq 0$ and then $\pa_y\va_+(y,c)\geq 0$,
for $y\geq y_{c_+}$ and $\pa_y\va_+(y,c)\leq 0$, for $y\leq y_{c_+}$.
Since $\va_+(y,c)-1=\al^2S^+\va_+(y,c)$, using (\ref{eq: W mono}), we have
\begin{equation*}
0\leq S^+\va_+(y,c)\leq \Big(\int_{y_{c_+}}^y\int_{y_{c_+}}^{y'}\Big|\frac{\mathcal{H}(z,c)}
{\mathcal{H}(y',c)}\Big|dzdy'\Big)\va_+(y,c)\leq C|y-y_{c_+}|^2\va_+(y,c),
\end{equation*}
then we obtain
\begin{eqnarray*}
0\leq \va_+(y,c)-1\leq C|y-y_{c_+}|^2.
\end{eqnarray*}
By the fact that $|S_1^+\va_+(y,c)|\leq C |y-y_{c_+}|\va_+(y,c)$, we have
\begin{eqnarray*}
|\pa_y\va_+(y,c)|\leq C|y-y_{c_+}|\va_+(y,c).
\end{eqnarray*}
On the other hand, $\va_+\geq 1$ and $\mathcal{H}(y,c)\geq C^{-1}|y-y_{c_+}||y+y_{c_+}|$, thus
\begin{align*}
|S_1^+\va_+(y,c)|\geq \left|\int_{y_{c_+}}^y\Big|\frac{\mathcal{H}(z,c)}{\mathcal{H}(y,c)}\Big|dz\right|\geq C^{-1}|y-y_{c_+}|.
\end{align*}
By Lemma \ref{lem: Y+0}, we get $\va_+(y,c)\in Y_0^+$, which implies
\begin{eqnarray*}
|\va_+(y,c)|+|\pa_c\va_+(y,c)|+|\pa_c\pa_y\va_+(y,c)|\leq C,
\end{eqnarray*}
with $C$ depending on $\al$. Then by the fact
that $\pa_c\va_+(y_{c_+},c)=0$, we have
\begin{eqnarray*}
|\partial_c\va_+(y,c)|=\left|\int_{y_{c_+}}^y\pa_y\pa_c\va_+(y',c)dy'\right|\leq C|y-y_{c_+}|,
\end{eqnarray*}
with $C$ depending on $\al$.
This completes the proof of the proposition.
\end{proof}
Similarly, for the case $y\in [a_-,0]$, we solve the Sturmian equation:
\begin{equation}\label{eq: homo eq1}
\left\{\begin{array}{l}
\pa_y\Big(\mathcal{H}(y,c)\pa_y\va_-(y,c)\Big)=\al^2\mathcal{H}(y,c)\va_-(y,c),\\
\va_-(y_{c_-},c)=1, \va'_-(y_{c_-},c)=0.\\
\end{array}\right.
\end{equation}
Here we only give the conclusion, and omit the details of the proof for brevity.
\begin{proposition}\label{prop: sol. hom. [-1,0]}
1. For $c\in \Omega_{\ep_0}$, there exists a solution
$\va_-(y,c)\in C(d_-\times \Omega_{\ep_0})$ of the Sturmian
equation \eqref{eq: homo eq1} and $\pa_y\va_-(y,c)\in C(d_-\times \Om_{\ep_0})$.
Moreover, there exists $\ep_1>0$ such that for any $\ep_0\in[0,\ep_1)$ and $(y,c)\in d_-\times \Om_{\ep_0}$,
\begin{equation*}
|\va_-(y,c)|\geq \frac{1}{2}, \ \ |\va_-(y,c)-1|\leq C|y-y_{c_-}|^2,
\end{equation*}
where the constants $\ep_1, C$ may depend on $\al$.\\
2. For $c\in D_0$, we have that for any $y\in d_-$, there is a constant $C$(depends on $\al$) such that,
\begin{eqnarray*}
\va_-(y,c)\geq \va_-(y',c)\geq1, \ for \ -1\leq y_{c_-}
\leq y'\leq y\leq 0 \ or\ -1\leq y\leq y'\leq y_{c_-}\leq 0;
\end{eqnarray*}
and
\begin{eqnarray*}
&\ &0\leq \va_-(y,c)-1\leq C\min\big\{\al^2(y-y_{c_-})^2, 1\big\}\va_-(y,c)
\end{eqnarray*}
\begin{eqnarray*}
C^{-1}|y-y_{c_-}|\leq |\pa_y\va_-(y,c)|\leq C|y-y_{c_-}|,
\end{eqnarray*}
and
\begin{eqnarray*}
|\pa_c\va_-(y,c)|\leq C|y-y_{c_-}|,
\ \ |\pa_y\pa_c\va_-(y,c)|\leq C.
\end{eqnarray*}
\end{proposition}
The proposition can be proved mainly by the following lemmas.
\begin{lemma}
Let $c\in D_{\ep_0}$.
Then there exists a solution $\va_-(y,c)\in Y^-$ to the Sturmian
equation \eqref{eq: homo eq1}. Moreover, it holds
\begin{eqnarray*}
\|\va_-\|_{Y^-}\leq C.
\end{eqnarray*}
\end{lemma}
\begin{lemma}
Let $c\in B_{\ep_0}^l$. Then there exists a solution
$\va_-(y,c)\in Y^-_l$ to the Sturmian equation \eqref{eq: homo eq1}).
Moreover, it holds
\begin{equation*}
\|\va_-\|_{Y^-_l}\leq C.
\end{equation*}
\end{lemma}
\begin{lemma}
Let $c\in B_{\ep_0}^r$. Then there exists a solution $\va_-(y,c)\in Y^-_r$
to the Sturmian equation \eqref{eq: homo eq1}. Moreover, it holds
\begin{equation*}
\|\va_-\|_{Y^-_r}\leq C.
\end{equation*}
\end{lemma}
\begin{lemma}
Let $c\in D_0$.
Then there exists a solution $\va_-(y,c)\in Y^-_0$ to the Sturmian equation \eqref{eq: homo eq1}. Moreover, there holds
\begin{equation*}
\|\va_-\|_{Y^-_0}\leq C.
\end{equation*}
\end{lemma}
\begin{remark}
From the above construction, we note that $\va_+(y,c)$, $\va_-(y,c)$ may be not equal at the point $y=0$.
\end{remark}
\begin{remark}\label{rmk: epsilon}
By the definition of $Y^{\pm}$ and $Y_r^{\pm}$, $Y_l^{\pm}$ and Proposition \ref{prop: sol. hom. [0,1]}, Proposition \ref{prop: sol. hom. [-1,0]}, we have for $c\in D_0$ and $c_{\ep}=c+i\ep\in D_{\ep_0}\cup D_0$ with $0\leq |\ep|\leq \ep_0$,
\begin{eqnarray*}
|\va_{\pm}(y,c_{\ep})-\va_{\pm}(y,c)|\leq C|\ep|,
\end{eqnarray*}
and for $c_{\ep}=c+\ep e^{i\th}\in B_{\ep_0}^l$ or $c_{\ep}=c+\ep e^{i\th}\in B_{\ep_0}^r$ with $0\leq |\ep|\leq \ep_0$,
\begin{eqnarray*}
|\va_{\pm}(y,c_{\ep})-\va_{\pm}(y,c)|\leq C|\ep|.
\end{eqnarray*}
\end{remark}
\section{The inhomogeneous Sturmian equations}
\subsection{The Wronskian and its estimate}
In the following, we introduce for $c\in \Omega_{\ep_0}\setminus D_0$,
\begin{equation}
I_+(c)=\int_0^1\frac{1}{\mathcal{H}(y,c)\va_+(y,c)^2}dy,\ \
I_-(c)=\int_{-1}^0\frac{1}{\mathcal{H}(y,c)\va_-(y,c)^2}dy,
\end{equation}
\begin{equation}
P(c)=\va_-(0,c)^2(\va_+\pa_y\va_+)(0,c)-\va_+(0,c)^2(\va_-\pa_y\va_-)(0,c),
\end{equation}
\begin{eqnarray}\label{wronskian}
\mathcal{D}(c)=c^2P(c)I_+(c)I_-(c)-\va_+(0,c)^2I_+(c)-\va_-(0,c)^2I_-(c).
\end{eqnarray}
Here and as what follows, $\va_+(y,c),\va_-(y,c)$ are the solutions of the homogeneous Sturmian
equation constructed in Proposition \ref{prop: sol. hom. [0,1]} and
Proposition \ref{prop: sol. hom. [-1,0]}.
The Stern stability condition ({\bf{SS}}) was proved in \cite{Stern}. Here we recall the lemma and show the relationship between the Stern stability condition ({\bf{SS}}) and the Wronskian $\mathcal{D}(c)$.
\begin{lemma}\label{stern lemma}
If $|u(y)|\leq|b(y)|$ for $y\in[-1,1]$, $M_{\al}$ has no $H^1$ eigenvalue.
Thus for any $c\notin D_0$, the Sturmian equation
\begin{equation}\label{homo Sturmian}
\left\{\begin{array}{l}
\pa_y\Big(\mathcal{H}(y,c)\pa_y\Psi(y,c)\Big)-\al^2\mathcal{H}(y,c)\Psi(y,c)=0\\
\Psi(-1,c)=\Psi(1,c)=0.\\
\end{array}\right.
\end{equation}
has no $H^1(-1,1)$ solution. And then we have $\mathcal{D}(c)\neq0$ for $c\in \Om_{\ep_0}\setminus D_0$.
\end{lemma}
\begin{proof}
For $c=c_r+ic_i\notin D_0$, let $\Psi(y,c)\in H^1_0(-1,1)$ be a nontrivial solution of the Sturmian equation
\begin{equation}\label{sturmian eq}
\pa_y\Big(\mathcal{H}(y,c)\pa_y\Psi(y,c)\Big)-\al^2\mathcal{H}(y,c)\Psi(y,c)=0.
\end{equation}
Taking the inner product with $\overline{\Psi}(y,c)$ on both sides of (\ref{sturmian eq})
and by integration by parts, we obtain
\begin{eqnarray}\label{integral}
\int_{-1}^1\mathcal{H}(y,c)\Big(|\pa_y\Psi(y,c)|^2+\al^2|\Psi(y,c)|^2\Big)dy=0.
\end{eqnarray}
Due to
\begin{equation*}
\begin{split}
\mathcal{H}(y,c)&=\big(u(y)+b(y)-c\big)\big(u(y)-b(y)-c\big)\\
&=\big(u(y)+b(y)-c_r\big)\big(u(y)-b(y)-c_r\big)
-c_i^2-2ic_i\big(u(y)-c_r\big),
\end{split}
\end{equation*}
taking the real part of (\ref{integral}) gives
\begin{equation}\label{re}
\int_{-1}^1\Big[\big(u(y)+b(y)-c_r\big)\big(u(y)-b(y)-c_r\big)-c_i^2\Big]
\Big(|\pa_y\Psi(y,c)|^2+\al^2|\Psi(y,c)|^2\Big)dy=0,
\end{equation}
and taking the imagine part of (\ref{integral}) gives
\begin{eqnarray}\label{im}
\int_{-1}^1(u(y)-c_r)\Big(|\pa_y\Psi(y,c)|^2+\al^2|\Psi(y,c)|^2\Big)dy=0.
\end{eqnarray}
Then multiplying $2c_r$ on both sides of (\ref{im}) and adding (\ref{re}), we get
\begin{eqnarray*}
\int_{-1}^1\Big[u(y)^2-b(y)^2-c_r^2-c_i^2\Big]
\Big(|\pa_y\Psi(y,c)|^2+\al^2|\Psi(y,c)|^2\Big)dy=0.
\end{eqnarray*}
Thus if $|u(y)|\leq |b(y)|$ for $y\in[-1,1]$, we have $\Psi(y,c)\equiv0$, which leads to a contradiction.\\
Let $\Psi(y,c)\in H^1(-1,1)$ is a solution of (\ref{homo Sturmian}) and assume $\Psi(y,c)=\left\{\begin{array}{l}
\Psi_+(y,c), \ \ y\in[0,1], \\
\Psi_-(y,c), \ \ y\in[-1,0],\\
\end{array}\right.$ then we have for $y\in[0,1]$, $\Psi_+$ satisfies
\begin{equation}\label{Psi+}
\left\{\begin{array}{l}
\pa_y\Big(\mathcal{H}(y,c)\pa_y\Psi_+(y,c)\Big)-\al^2\mathcal{H}(y,c)\Psi_+(y,c)=0\\
\Psi_+(1,c)=0.\\
\end{array}\right.
\end{equation}
and for $y\in[-1,0]$, $\Psi_-$ satisfies
\begin{equation}\label{Psi-}
\left\{\begin{array}{l}
\pa_y\Big(\mathcal{H}(y,c)\pa_y\Psi_-(y,c)\Big)-\al^2\mathcal{H}(y,c)\Psi_-(y,c)=0\\
\Psi_-(-1,c)=0.\\
\end{array}\right.
\end{equation}
By Proposition \ref{prop: sol. hom. [0,1]} and Proposition \ref{prop: sol. hom. [-1,0]}, $\va_+,\va_-\neq 0$, then it holds that the equations (\ref{Psi+}) and (\ref{Psi-}) are equivalent to
\begin{equation*}
\left\{\begin{array}{l}
\pa_y\Big(\mathcal{H}\va_+^2\pa_y\Big(\frac{\Psi_+}{\va_+}\Big)\Big)=0\\
\Psi_+(1,c)=0,\\
\end{array}\right.
\quad\textrm{and}\quad
\left\{\begin{array}{l}
\pa_y\Big(\mathcal{H}\va_-^2\pa_y\Big(\frac{\Psi_-}{\va_-}\Big)\Big)=0\\
\Psi_-(1,c)=0.\\
\end{array}\right.
\end{equation*}
As $\mathcal{H}\neq0$ for $c\notin D_0$, by integration twice, we obtain that $\va_+(y,c)$ and
$\va_+(y,c)\int_1^y\frac{1}{\mathcal{H}(y,c)\va_+(y,c)^2}dy'$ are two independent solutions
of the homogeneous Sturmian equation for $y\in[0,1]$, and $\va_-(y,c)$
and $\va_-(y,c)\int_{-1}^y\frac{1}{\mathcal{H}(y,c)\va_-(y,c)^2}dy'$ are two independent
solutions of the homogeneous Sturmian equation for $y\in[-1,0]$.
Thus for $y\in[0,1]$ and $c\notin D_0$, we have
\begin{equation*}
\begin{split}
\Psi_+(y,c)&=\wt{\mu}_+(c) \va_+(y,c)\int_0^y\frac{1}{\mathcal{H}(y',c)\va_+(y',c)^2}dy'
+\nu_+(c)\va_+(y,c):= \Psi_+^0(y,c)\\
&=\mu_+(c) \va_+(y,c)\int_1^y\frac{1}{\mathcal{H}(y',c)\va_+(y',c)^2}dy':= \Psi_+^1(y,c),
\end{split}
\end{equation*}
and for $y\in[-1,0]$ and $c\notin D_0$,
\begin{equation*}
\begin{split}
\Psi_-(y,c)&=\mu_-(c) \va_-(y,c)\int_{-1}^y\frac{1}{\mathcal{H}(y',c)\va_-(y',c)^2}dy'
:=\Psi_-^{-1}(y,c)\\
&=\wt{\mu}_-(c) \va_-(y,c)\int_0^y\frac{1}{\mathcal{H}(y',c)\va_-(y',c)^2}dy'
+\nu_-(c)\va_-(y,c):=\Psi_-^0(y,c).
\end{split}
\end{equation*}
By the boundary conditions and the fact that $\Psi(y,c)\in H_0^{1}(-1,1)$, we get,
\begin{eqnarray*}
&\ & \Psi_+^0(1,c)=0, \ \Psi_-^0(-1,c)=0, \
\Psi_+^1(0,c)=\Psi_+^0(0,c), \\
&\ & \Psi_-^{-1}(0,c)=\Psi_-^0(0,c),\
\Psi_+^0(0,c)=\Psi_-^0(0,c), \ \pa_y\Psi_+^0(0,c)=\pa_y\Psi_-^0(0,c),
\end{eqnarray*}
which gives
\begin{equation*}
\left\{\begin{array}{l}
\mu_+(c)=\wt{\mu}_+(c), \ \
\mu_-(c)=\wt{\mu}_-(c),\\
I_+(c)\mu_+(c)+\nu_+(c)=0,\\
I_-(c)\mu_-(c)-\nu_-(c)=0,\\
\va_+(0,c)\nu_+(c)-\va_-(0,c)\nu_-(c)=0,\\
\va_-(0,c)\mu_+(c)-\va_+(0,c)\mu_-(c)
+c^2\big(\va_-\va_+\pa_y\va_+\big)(0,c)\nu_+(c)\\
\ \ -c^2\big(\va_+\va_-\pa_y\va_-\big)(0,c)\nu_-(c)=0.
\end{array}\right.
\end{equation*}
Thus we have
\begin{equation}\label{matrix}
W
\begin{bmatrix}
\mu_+(c)\\
\mu_-(c)\\
\nu_+(c)\\
\nu_-(c)
\end{bmatrix}
=0,
\end{equation}
where
$
W=\begin{bmatrix}
I_+(c)& 0 & 1& 0\\
0& I_-(c)& 0 & -1\\
0& 0& \va_+(0,c)& -\va_-(0,c)\\
\va_-(0,c)& -\va_+(0,c)& c^2\big(\va_-\va_+\pa_y\va_+\big)(0,c)
& -c^2\big(\va_+\va_-\pa_y\va_-\big)(0,c)
\end{bmatrix},
$
and the fact that \eqref{homo Sturmian} has no nontrivial $H^1$ solution, we obtain the Wronskian $\det(W)\neq 0$ holds for $c\in \Om_{\ep_0}\setminus D_0$, which gives
\begin{eqnarray*}
\det(W)=\mathcal{D}(c)=c^2P(c)I_+(c)I_-(c)-\va_+(0,c)^2I_+(c)-\va_-(0,c)^2I_-(c)\neq0.
\end{eqnarray*}
Thus we proved the lemma.
\end{proof}
\begin{remark}
Thanks to the continuity of $\va_+$ and $\va_-$, we have
that $\mathcal{D}(c)$ is continuous for $c\in \Om_{\ep_0}\setminus D_0$.
\end{remark}
Next we show that $\mathcal{D}(c)$ is continuous to the boundary.
For $c\in D_{\ep_0}\cup D_0$, let
\begin{equation}\label{sigma+}
\sigma_+(c)=W'_+(y_{c_+})\big(W_-(y_{c_+})-c\big)
-W'_-(y_{c_+})\big(W_+(y_{c_+})-c\big),
\end{equation}
\begin{equation}\label{sigma-}
\sigma_-(c)=W'_+(y_{c_-})\big(W_-(y_{c_-})-c\big)
-W'_-(y_{c_-})\big(W_+(y_{c_-})-c\big),
\end{equation}
\begin{equation}\label{Pi}
\Pi_+(c)=\int_0^1\frac{1}{\mathcal{H}(y,c)}\Big(\frac{1}{\va_+(y,c)^2}-1\Big)dy, \ \
\Pi_-(c)=\int_{-1}^0\frac{1}{\mathcal{H}(y,c)}\Big(\frac{1}{\va_-(y,c)^2}-1\Big)dy,
\end{equation}
For $c\in D_0$, let
\begin{equation}\label{R+1}
R_+^1(c)=\int_0^1\frac{W'_+(y_{c_+})\big[W_-(y_{c_+})-W_-(y)\big]
-W'_-(y_{c_+})\big[W_+(y_{c_+})-W_+(y)\big]}
{\big(W_+(y)-c\big)\big(W_-(y)-c\big)}dy,
\end{equation}
\begin{equation}\label{R+2}
R_+^2(c)=\int_0^1\frac{W'_+(y_{c_+})-W'_+(y)}{W_+(y)-c}
-\frac{W'_-(y_{c_+})-W'_-(y)}{W_-(y)-c}dy,
\end{equation}
\begin{equation}\label{R-1}
R_-^1(c)=\int_{-1}^0\frac{W'_+(y_{c_-})\big[W_-(y_{c_-})-W_-(y)\big]
-W'_-(y_{c_-})\big[W_+(y_{c_-})-W_+(y)\big]}
{\big(W_+(y)-c\big)\big(W_-(y)-c\big)}dy,
\end{equation}
\begin{equation}\label{R-2}
R_-^2(c)=\int_{-1}^0\frac{W'_+(y_{c_-})-W'_+(y)}{W_+(y)-c}
-\frac{W'_-(y_{c_-})-W'_-(y)}{W_-(y)-c}dy,
\end{equation}
For $c\in D_0\setminus \big\{0,W_+(1),W_-(1),W_+(-1),W_-(-1)\big\}$, let
\begin{eqnarray*}
\chi_+(c)=\left\{\begin{aligned}
1,\ c\in (W_-(1),W_+(1)),\\
0,\ c\notin [W_-(1),W_+(1)];
\end{aligned}
\right.
\quad
\chi_-(c)=\left\{\begin{aligned}
1,\ c\in (W_+(-1),W_-(-1)),\\
0,\ c\notin [W_+(-1),W_-(-1)].
\end{aligned}
\right.
\end{eqnarray*}
Then $\chi_+(c)^2+\chi_-(c)^2\geq 1$ for $c\in D_0\setminus \big\{0,W_+(1),W_-(1),W_+(-1),W_-(-1)\big\}$.
We denote for $c\in D_0\setminus \big\{0,W_+(1),W_-(1),W_+(-1),W_-(-1)\big\}$,
\begin{equation}\label{I+ re def}
I^{re}_+(c)=\Pi_+(c)+\frac{1}{\sigma_+(c)}
\left(R_+^1(c)+R_+^2(c)+\f12\ln \Big(\frac{W_+(1)-c}{c-W_-(1)}\Big)^2\right),
\end{equation}
\begin{equation}\label{I- re def}
I^{re}_-(c)=\Pi_-(c)+\frac{1}{\sigma_-(c)}
\left(R_-^1(c)+R_-^2(c)+\f12\ln\Big(\frac{W_-(-1)-c}{c-W_+(-1)}\Big)^2\right),
\end{equation}
\begin{equation}
\mathcal{D}^{re}(c)=c^2P(c)\Big(I^{re}_+(c)I^{re}_-(c)
-\frac{\pi^2\chi_+(c)\chi_-(c)}{\sigma_+(c)\sigma_-(c)}\Big)
-\va_+(0,c)^2I^{re}_+(c)-\va_-(0,c)^2I^{re}_-(c),
\end{equation}
\begin{equation}
\mathcal{D}^{im}(c)=c^2P(c)\Big(\frac{\pi I^{re}_+(c)\chi_-(c)}{\sigma_-(c)}
+\frac{\pi I^{re}_-(c)\chi_+(c)}{\sigma_+(c)}\Big)
-\frac{\pi\va_+(0,c)^2\chi_+(c)}{\sigma_+(c)}
-\frac{\pi\va_-(0,c)^2\chi_-(c)}{\sigma_-(c)}.
\end{equation}
We also define $l(x)=\ln(e+|x|^{-1})$ for $x\in \mathbb{C}$, so that $C(M)^{-1}(1+|\ln|x||)\leq l(x)\leq C(M)(1+|\ln|x||)$ for $|x|\leq M$.
\begin{remark}\label{rmk: sigma bdd}
By the definition of $\sigma_+(c)$ and $\sigma_-(c)$, for $c\in D_{\ep_0}\cup D_0$, we can easily get that there exists a positive constant $C$ such that
\begin{eqnarray*}
C^{-1} |c|\leq|\sigma_+(c)|\leq C |c|, \ C^{-1} |c|\leq|\sigma_-(c)|\leq C |c|.
\end{eqnarray*}
\end{remark}
\begin{proposition}\label{prop: D est}
There exists $\ep_0>0$, such that for $c_{\ep}\in \Om_{\ep_0}$, the following properties hold. \\
1. For $c\in D_0\setminus \big\{0,W_+(1),W_-(1),W_+(-1),W_-(-1)\big\}$, $c_{\ep}=c+i\ep$. It holds
\begin{eqnarray*}
\lim_{\ep\rightarrow0^{\pm}} \mathcal{D}(c_{\ep})=\mathcal{D}^{re}(c)\pm i\mathcal{D}^{im}(c).
\end{eqnarray*}
Moreover, there exists a constant $C\geq 1$ such that
\begin{eqnarray*}
\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2\geq C^{-1} >0.
\end{eqnarray*}
2. For $c_{\ep}=c+i\ep\in \Om_{\ep_0}\setminus D_0$, $0<\ep<\ep_0$, there exists a constant $\delta_0>0$ such that for $|c|<\delta_0$,
\begin{eqnarray*}
|\mathcal{D}(c_{\ep})|\geq \frac{C^{-1}}{|c_{\ep}|};
\end{eqnarray*}
3. For $c_{\ep}\in \Omega_{\ep_0}\setminus D_0$, it holds that
\begin{eqnarray*}
|\mathcal{D}(c_{\ep})|\geq \frac{
l\big(c_{\ep}-W_-(-1)\big)l\big(W_+(-1)-c_{\ep}\big)l\big(c_{\ep}-W_-(1)\big)l\big(W_+(1)-c_{\ep}\big)}{C|c_{\ep}|}.
\end{eqnarray*}
Here we recall $l(x)=\ln(e+|x|^{-1})$.
\end{proposition}
The proof of the proposition is mainly dependent on the following lemmas.
\begin{lemma}\label{lem: I limit+}
For $c_{\ep}= c+i\ep \in D_{\ep_0}$, $c\in D_0\setminus \big\{W_+(1),W_-(1)\big\}$, $\ep\in(0,\ep_0)$. It holds that
\begin{equation}\label{I+ lim}
\lim_{\ep\rightarrow 0{\pm}}\sigma_+(c_{\ep})I_+(c_{\ep})
=\sigma_+(c)I_+^{re}(c)\pm i\pi\chi_+(c),
\end{equation}
where
$\sigma_+(c_{\ep})$ is defined as \eqref{sigma+}.
\end{lemma}
\begin{proof}
Let $c_{\ep}=c+i\ep$.
Due to the fact that
\begin{equation}\label{cI+}
\sigma_+(c_{\ep})I_+(c_{\ep})
=\sigma_+(c_{\ep})\int_0^1\frac{1}{\mathcal{H}(y,c_{\ep})}
\Big(\frac{1}{\va_+(y,c_{\ep})^2}-1\Big)dy
+\sigma_+(c_{\ep})\int_0^1\frac{1}{\mathcal{H}(y,c_{\ep})}dy,
\end{equation}
and by Proposition \ref{prop: sol. hom. [0,1]}, $\va_+(y,c_{\ep})$ is continuous for $(y,c_{\ep})\in [0,a_+]\times \Omega_{\ep_0}$ and for $\ep_0$ small enough,
\begin{equation*}
|\va_+(y,c_{\ep})|\geq \frac{1}{2}, \ \ |\va_+(y,c_{\ep})-1|\leq C|y-y_{c_+}|^2.
\end{equation*}
By using the fact that $y_{c_-}\leq 0\leq y_{c_+}$, we have
\begin{equation}\label{mono}
|y-y_{c_+}|\leq |y-y_{c_-}|,
\end{equation}
and for $c_{\ep}\in\Omega_{\ep_0}$,
\begin{equation}\label{1}
\Big|\frac{1}{\mathcal{H}(y,c_{\ep})}
\Big(\frac{1}{\va_+(y,c_{\ep})^2}-1\Big)\Big|\leq C\frac{|y-y_{c_+}|^2}{|y-y_{c_+}||y-y_{c_-}|}\leq C.
\end{equation}
Thus, by the Lebesgue's dominated convergence theorem, we have that as $\ep\rightarrow0$,
\begin{equation*}
\sigma_+(c_{\ep})\int_0^1\frac{1}{\mathcal{H}(y,c_{\ep})}
\Big(\frac{1}{\va_+(y,c_{\ep})^2}-1\Big)dy\rightarrow\sigma_+(c)
\int_0^1\frac{1}{\mathcal{H}(y,c)}
\Big(\frac{1}{\va_+(y,c)^2}-1\Big)dy.
\end{equation*}
On the other hand, we have
\begin{equation}\label{sigma I+}
\begin{split} &\sigma_+(c_{\ep})\int_0^1\frac{1}{\big(W_+(y)-c_{\ep}\big)\big(W_-(y)-c_{\ep}\big)}dy\\
&=
\int_0^1\frac{W'_+(y_{c_+})\big(W_-(y_{c_+})-W_-(y)\big)
-W'_-(y_{c_+})\big(W_+(y_{c_+})-W_+(y)\big)}
{\big(W_+(y)-c_{\ep}\big)\big(W_-(y)-c_{\ep}\big)}dy\\
&\quad+
\int_0^1\frac{W'_+(y_{c_+})-W'_+(y)}{W_+(y)-c_{\ep}}
-\frac{W'_-(y_{c_+})-W'_-(y)}{W_-(y)-c_{\ep}}dy\\
&\quad
+
\int_0^1\frac{W'_+(y)}{W_+(y)-c_{\ep}}-\frac{W'_-(y)}{W_-(y)-c_{\ep}}dy
=I_1(c_{\ep})+I_2(c_{\ep})+I_3(c_{\ep}).
\end{split}
\end{equation}
By denoting that $h(y,y_{c_+})=W'_+(y_{c_+})\big(W_-(y_{c_+})-W_-(y)\big)-W'_-(y_{c_+})\big(W_+(y_{c_+})-W_+(y)\big)$, we have $h(y_{c_+},y_{c_+})=0$
and $(\pa_yh)(y_{c_+},y_{c_+})=0$. Thus we obtain that
\begin{equation*}
h(y,y_{c_+})=(y-y_{c_+})^2\int_0^1\int_0^1\pa_{yy}h\big(y_{c_+}+ts(y-y_{c_+})\big)dtds,
\end{equation*}
and then
\begin{equation}\label{h bdd}
|h(y,y_{c_+})|\leq C|y-y_{c_+}|^2.
\end{equation}
Due to $|(W_+(y)-c_{\ep})(W_-(y)-c_{\ep})|\geq C|y-y_{c_+}||y-y_{c_-}|$, we get by (\ref{mono}) and (\ref{h bdd}),
\begin{equation}\label{A+1}
\Big|\frac{h(y,y_c)}
{(W_+(y)-c_{\ep})(W_-(y)-c_{\ep})}\Big|\leq \frac{C|y-y_{c_+}|^2}{|y-y_{c_+}||y-y_{c_-}|}\leq C,
\end{equation}
and
\begin{equation}\label{A+2}
\Big|\frac{W'_+(y_{c_+})-W'_+(y)}{W_+(y)-c_{\ep}}
-\frac{W'_-(y_{c_+})-W'_-(y)}{W_-(y)-c_{\ep}}\Big|
\leq \frac{C|y-y_{c_+}|}{|y-y_{c_+}|}
+\frac{C|y-y_{c_+}|}{|y-y_{c_-}|}\leq C.
\end{equation}
Therefore $|I_1(c_{\ep})|+|I_2(c_{\ep})|\leq C$ and by the Lebesgue's dominated convergence theorem, we have as $\ep$ tends to $0$,
\begin{equation*}
\begin{split}
I_1(c_{\ep})&\rightarrow
\int_0^1\frac{h(y,y_{c_+})}
{\big(W_+(y)-c\big)\big(W_-(y)-c\big)}dy,\\
I_2(c_{\ep})&\rightarrow
\int_0^1\frac{W'_+(y_{c_+})-W'_+(y)}{W_+(y)-c}
-\frac{W'_-(y_{c_+})-W'_-(y)}{W_-(y)-c}dy.
\end{split}
\end{equation*}
Due to $W_+(0)=W_-(0)=0$, we also have for $c_{\ep}\in \Omega_{\ep_0}\setminus D_0$,
\begin{equation}\label{eq: I3}
I_3(c_{\ep})=\ln\frac{W_+(1)-c_{\ep}}{W_-(1)-c_{\ep}}.
\end{equation}
And for $c_{\ep}=c+i\ep, c\in D_{\ep_0}\setminus D_0$, we have
\begin{equation}
I_3(c_{\ep})=
\frac{1}{2}\ln\frac{\big(W_+(1)-c\big)^2+\ep^2}{\big(W_-(1)-c\big)^2+\ep^2}
+i\arctan\frac{W_+(1)-c}{\ep}
-i\arctan\frac{W_-(1)-c}{\ep},
\end{equation}
and then
\begin{align*}
\lim_{\ep\rightarrow0+}I_3(c_{\ep})=\f12\ln\Big(\frac{W_+(1)-c}{c-W_-(1)}\Big)^2+i\pi\chi_+(c), \\
\lim_{\ep\rightarrow0-}I_3(c_{\ep})=
\f12\ln\Big(\frac{W_+(1)-c}{c-W_-(1)}\Big)^2-i\pi\chi_+(c).
\end{align*}
Thus, from (\ref{cI+}), we get
\begin{equation*}
\lim_{\ep\rightarrow 0^{\pm}}\sigma_+(c_{\ep})I_+(c_{\ep})
=\sigma_+(c)\Pi_+(c)+R_+^1(c)+R_+^2(c)+\f12\ln \Big(\frac{W_+(1)-c}{c-W_-(1)}\Big)^2\pm i\pi\chi_+(c).
\end{equation*}
This complete the proof of the lemma.
\end{proof}
\begin{lemma}\label{lem: I limit-}
For $c_{\ep}= c+i\ep \in D_{\ep_0}$ and $c\in D_0\setminus\big\{W_-(-1), W_+(-1)\big\}$. It holds that
\begin{equation}\label{I- lim}
\lim_{\ep\rightarrow 0{\pm}}\sigma_-(c_{\ep})I_-(c_{\ep})
=\sigma_-(c)I_-^{re}(c)\pm i\pi\chi_-(c),
\end{equation}
where $\sigma_-(c_{\ep})$ is defined by \eqref{sigma-}.
\end{lemma}
\begin{proof}
The proof of the lemma is same as the proof of Lemma \ref{lem: I limit+}. A direct calculation gives
\begin{align*}
&\sigma_-(c_{\ep})I_-(c_{\ep})\\
&=\sigma_-(c_{\ep})\int_{-1}^0\frac{1}{\mathcal{H}(y,c_{\ep})}
\left(\frac{1}{\va_-(y,c_{\ep})^2}-1\right)dy
+\sigma_-(c_{\ep})\int_{-1}^0\frac{1}{\mathcal{H}(y,c_{\ep})}dy\\
&=\sigma_-(c_{\ep})\int_{-1}^0\frac{1}{\mathcal{H}(y,c_{\ep})}
\left(\frac{1}{\va_-(y,c_{\ep})^2}-1\right)dy
+\int_{-1}^0\frac{g(y,y_{c_+})}
{\big(W_+(y)-c_{\ep}\big)\big(W_-(y)-c_{\ep}\big)}dy\\
&\quad+
\int_{-1}^0\frac{W'_+(y_{c_-})-W'_+(y)}{W_+(y)-c_{\ep}}
-\frac{W'_-(y_{c_-})-W_-{'}(y)}{W_-(y)-c_{\ep}}dy\\
&\quad
+\int_{-1}^0\frac{W'_+(y)}{W_+(y)-c_{\ep}}-\frac{W'_-(y)}{W_-(y)-c_{\ep}}dy\buildrel\hbox{\footnotesize def}\over =\sigma(c_{\ep})\Pi_-(c_{\ep})+J_1(c_{\ep})+J_2(c_{\ep})+J_3(c_{\ep}),
\end{align*}
with $g(y,y_{c_-})=W'_+(y_{c_-})\big[W_-(y_{c_-})-W_-(y)\big]
-W'_-(y_{c_-})\big[W_+(y_{c_-})-W_+(y)\big]$.
Proposition \ref{prop: sol. hom. [-1,0]} implies
\begin{eqnarray*}
\Big|\frac{1}{\mathcal{H}(y,c_{\ep})}
\Big(\frac{1}{\va_-(y,c_{\ep})^2}-1\Big)\Big|\leq C,
\end{eqnarray*}
and thus $|\Pi_-(c_{\ep})|\leq C$ and as $\ep\rightarrow0$,
\begin{equation*}
\sigma_-(c_{\ep})\Pi_-(c_{\ep})\rightarrow\sigma_-(c)
\int_{-1}^0\frac{1}{\mathcal{H}(y,c)}
\Big(\frac{1}{\va_-(y,c)^2}-1\Big)dy.
\end{equation*}
By the fact that
\begin{equation}\label{eq: g est}
|g(y,y_{c_+})|\leq C|y-y_{c_+}|^2,
\end{equation}
and then
\begin{eqnarray*}
\left|\frac{W'_+(y_{c_-})\big[W_-(y_{c_-})-W_-(y)\big]
-W'_-(y_{c_-})\big[W_+(y_{c_-})-W_+(y)\big]}
{\big(W_+(y)-c_{\ep}\big)\big(W_-(y)-c_{\ep}\big)}\right|\leq C,
\end{eqnarray*}
and
\begin{eqnarray*}
\Big|\frac{W'_+(y_{c_-})-W'_+(y)}{W_+(y)-c_{\ep}}
-\frac{W'_-(y_{c_-})-W'_-(y)}{W_-(y)-c_{\ep}}\Big|
\leq C,
\end{eqnarray*}
and the Lebesgue's dominated convergence theorem, we have $|J_1(c_{\ep})|+|J_2(c_{\ep})|\leq C$ and as $\ep$ tends to $0$,
\begin{equation*}
\begin{split}
J_1(c_{\ep})&\rightarrow
\int_{-1}^0\frac{W'_+(y_{c_-})\big[W_-(y_{c_-})-W_-(y)\big]
-W'_-(y_{c_-})\big[W_+(y_{c_-})-W_+(y)\big]}
{\big(W_+(y)-c\big)\big(W_-(y)-c\big)}dy,\\
J_2(c_{\ep})&\rightarrow
\int_{-1}^0\frac{W'_+(y_{c_-})-W'_+(y)}{W_+(y)-c}
-\frac{W'_-(y_{c_-})-W'_-(y)}{W_-(y)-c}dy.
\end{split}
\end{equation*}
Due to $W_+(0)=W_-(0)=0$, we also have for $c_{\ep}\in \Omega_{\ep_0}\setminus D_0$,
\begin{equation}\label{eq: J3}
J_3(c_{\ep})=\ln\frac{W_-(-1)-c_{\ep}}{W_+(-1)-c_{\ep}}.
\end{equation}
And for $c_{\ep}=c+i\ep, c\in D_{\ep_0}\setminus D_0$, we have
\begin{equation*}
J_3(c_{\ep})=
\frac{1}{2}\ln\frac{\big(W_-(-1)-c\big)^2+\ep^2}{\big(W_+(-1)-c\big)^2+\ep^2}
-i\arctan\frac{W_+(-1)-c}{\ep}
+i\arctan\frac{W_-(-1)-c}{\ep}.
\end{equation*}
And then, we get
\begin{eqnarray*}
&&\lim_{\ep\rightarrow0+}J_3(c_{\ep})=\f12\ln\Big(\frac{W_-(-1)-c}{c-W_+(-1)}\Big)^2+i\pi\chi_-(c),\\
&&\lim_{\ep\rightarrow0-}J_3(c_{\ep})=\f12
\ln\Big(\frac{W_-(-1)-c}{c-W_+(-1)}\Big)^2-i\pi\chi_-(c).
\end{eqnarray*}
Thus we complete the proof of lemma.
\end{proof}
\begin{lemma}\label{lem: I bdd}
There is $\ep_0>0$ such that for $c_{\ep}\in \Omega_{\ep_0}\setminus D_0$, there exists a constant $C\geq1$ such that
\begin{eqnarray*}
\frac{l\big(W_+(\pm1)-c_{\ep}\big)l
\big(W_-(\pm1)-c_{\ep}\big)}{C|c_{\ep}|}\leq |I_{\pm}(c_{\ep})|
\leq \frac{Cl\big(W_+(\pm1)-c_{\ep}\big)l
\big(W_-(\pm1)-c_{\ep}\big)}{|c_{\ep}|}.
\end{eqnarray*}
\end{lemma}
\begin{proof}
We only give the estimate of $I_+(c_{\ep})$, the estimate of $I_-(c_{\ep})$ can be obtained by the same way and we omit the details here.
Due to (\ref{cI+}), we have
\begin{eqnarray*}
I_+(c_{\ep})=\Pi_+(c_{\ep})
+\frac{I_1(c_{\ep})+I_2(c_{\ep})+I_3(c_{\ep})}{\sigma_+(c_{\ep})}.
\end{eqnarray*}
The upper bound is directly follows from the fact that $|\Pi_+(c_{\ep})|+|I_1(c_{\ep})|+|I_2(c_{\ep})|\leq C$ and
\begin{align*}
|I_3(c_{\ep})|\leq
\left|\frac{1}{2}\ln\frac{\big(W_+(1)-c\big)^2+\ep^2}{\big(W_-(1)-c\big)^2+\ep^2}\right|+C
\leq Cl\big(W_+(1)-c_{\ep}\big)l
\big(W_-(1)-c_{\ep}\big).
\end{align*}
For the lower bounded. At first, we consider the case of $|c|<\delta_0$ for some $\delta_0>0$ small enough.
By Remark \ref{rmk: sigma bdd} and the fact that $|\Pi_+(c_{\ep})|\leq C$, we have
\begin{align*}
|I_+(c_{\ep})|&\geq \frac{\big|I_1(c_{\ep})+I_2(c_{\ep})+I_3(c_{\ep})\big|}{|\sigma_+(c_{\ep})|}
-|\Pi_+(c_{\ep})|\\
&\geq \frac{\big|Im(I_1(c_{\ep})+I_2(c_{\ep})+I_3(c_{\ep}))\big|}{C|c_{\ep}|}
-C\\
&\geq \frac{\big|Im(I_3(c_{\ep}))\big|}{C|c_{\ep}|}
-\frac{\big|Im(I_1(c_{\ep})+I_2(c_{\ep}))\big|}{C|c_{\ep}|}
-C.
\end{align*}
We have for $|c|<\delta_0$.
\begin{equation*}
\begin{split}
Im(I_1(c_{\ep}))&=\int_0^1\frac{\ep h(y,y_{c_+})(W_+(y)+W_-(y)-2c)}
{\big((W_+(y)-c)^2+\ep^2\big)\big((W_-(y)-c)^2+\ep^2\big)}dy,\\
Im(I_2(c_{\ep}))&=\int_0^1\frac{\ep\big(W'_+(y_{c_+})-W'_+(y)\big)}
{(W_+(y)-c)^2+\ep^2}dy-\int_0^1\frac{\ep\big(W'_-(y_{c_+})-W'_-(y)\big)}
{(W_-(y)-c)^2+\ep^2}dy.
\end{split}
\end{equation*}
By (\ref{h bdd}) and $|W_+(y)+W_-(y)-2c|\leq C|y-y_{c_+}|+C|y-y_{c_-}|\leq C|y-y_{c_-}|$, we have
\begin{equation}\label{I1 im}
|Im(I_1(c_{\ep}))|\leq C\ep\int_0^1\frac{y-y_{c_-}}{(y-y_{c_-})^2+\ep^2}dy\leq C\ep|\ln\ep|,
\end{equation}
and
\begin{equation}\label{I2 im}
\begin{split}
|Im(I_2(c_{\ep}))|&\leq C\ep\int_0^1\frac{|y-y_{c_+}|}{(y-y_{c_+})^2+\ep^2}dy
+C\ep\int_0^1\frac{|y-y_{c_+}|}{(y-y_{c_-})^2+\ep^2}dy\\
&\leq C\ep\int_0^1\frac{|y-y_{c_+}|}{(y-y_{c_+})^2+\ep^2}dy
+C\ep\int_0^1\frac{|y-y_{c_-}|}{(y-y_{c_-})^2+\ep^2}dy\\
&\leq C\ep|\ln\ep|.
\end{split}
\end{equation}
On the other hand, under the assumption $|c|<\delta_0$,
\begin{equation}\label{I3 im}
|Im(I_3(c_{\ep}))|=\Big|\arctan\frac{W_+(1)-c}{\ep}-\arctan\frac{W_-(1)-c}{\ep}\Big|
\geq\frac{3\pi}{4}.
\end{equation}
By choosing $\delta_0\leq \f{W_+(1)-W_-(1)}{1000}\ep_0$ small enough with $\ep_0$ in Proposition \ref{prop: sol. hom. [0,1]} also small enough, then (\ref{I1 im}), (\ref{I2 im}) and (\ref{I3 im}) imply for $|c|<\delta_0$,
\begin{equation}\label{eq: I+0 low}
|I_+(c_{\ep})|\geq \frac{C^{-1}}{|c_{\ep}|}-\frac{C\ep|\ln\ep|}{|c_{\ep}|}-C\geq\frac{C^{-1}}{|c_{\ep}|}.
\end{equation}
For the case of $|c_{\ep}-W_+(1)|<\delta_0$, we have
\begin{equation}
|\Pi_+(c_{\ep})|+\Big|\frac{I_1(c_{\ep})+I_2(c_{\ep})}{\sigma_+(c_{\ep})}\Big|\leq C.
\end{equation}
Thus for $|c_{\ep}-W_+(1)|<\delta_0$, by \eqref{eq: I3},
\begin{eqnarray*}
\begin{split}
|I_+(c_{\ep})|&\geq \frac{|I_3(c_{\ep})|}{|\sigma_+(c_{\ep})|}
-\frac{|I_1(c_{\ep})+I_2(c_{\ep})|}{|\sigma_+(c_{\ep})|}
-|\Pi_+(c_{\ep})|\\
&\geq \frac{\Big|\ln\big|W_+(1)-c_{\ep}\big|\Big|}{C}-C\\
&\geq C^{-1}\big|\ln|W_+(1)-c_{\ep}|\big|
\end{split}
\end{eqnarray*}
Similarly, we have for the case of $\big|c_{\ep}-W_-(1)\big|<\delta_0$,
\begin{eqnarray*}
|I_+(c_{\ep})|\geq C^{-1}|\ln\big|W_-(1)-c_{\ep}|\big|
\end{eqnarray*}
For $c_{\ep}\in \Om_{\ep_0}\setminus D_0$ with $|c_{\ep}|\geq \d_0$, $\big|c_{\ep}-W_-(1)\big|\geq \delta_0$ and $\big|c_{\ep}-W_-(1)\big|\geq \delta_0$, Lemma \ref{lem: I limit+} implies $\sigma_+(c_{\ep})I_+(c_{\ep})$ is continuous to the boundary with its boundary value $\sigma_+(c)I_+^{re}\pm i\pi\chi_+(c)$. Let $c_{\ep}=c+i\ep$ with $c\in D_0$ and $|c-W_+(1)|\geq \delta_0$ and $|c-W_-(1)|\geq \delta_0$, then if $\chi_+(c)=0$($c<W_-(1)$ in Case 1, 2, 3, or $c>W_+(1)$ in Case 3, 8, 9,) then $\mathcal{H}(y,c)>0$ which implies $I_+^{re}(c)=\int_0^1\f{1}{\mathcal{H}(y,c)\va_+(y,c)^2}dy>0$ and if $\chi_+(c)\neq 0$, then $|\sigma_+(c)I_+^{re}\pm i\pi\chi_+(c)|\geq \pi$.
Therefore there is $\ep_0$ such that for any $0\leq |\ep|\leq \ep_0$ and $c_{\ep}=c+i\ep$
\begin{eqnarray*}
|\sigma_+(c_{\ep})I_+(c_{\ep})|\geq \f{\pi}{2}.
\end{eqnarray*}
Thus we conclude that
\begin{eqnarray*}
|I_+(c_{\ep})|\geq \frac{l\big(W_+(1)-c_{\ep}\big)l
\big(W_-(1)-c_{\ep}\big)}{C|c_{\ep}|}.
\end{eqnarray*}
This completes the proof of the lemma.
\end{proof}
\begin{lemma}\label{lem: D>0}
Let $c\in D_0\setminus \big\{0,W_+(1),W_-(1),W_+(-1),W_-(-1)\big\}$, $c_{\ep}=c+i\ep$. It holds
\begin{equation*}
\lim_{\ep\rightarrow0^{\pm}} \mathcal{D}(c_{\ep})=\mathcal{D}^{re}(c)\pm i\mathcal{D}^{im}(c).
\end{equation*}
Moreover, we have that there exists a constant $C\geq 1$ such that
\begin{equation*}
\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2\geq C^{-1} >0.
\end{equation*}
\end{lemma}
\begin{proof}
By Lemma \ref{lem: I limit+} and Lemma \ref{lem: I limit-}, we get for $D_0\setminus \big\{0,W_+(1),W_-(1),W_+(-1),W_-(-1)\big\}$,
\begin{eqnarray*}
\lim_{\ep\rightarrow 0{\pm}}I_+(c)=I^{re}_+(c)\pm \frac{i\pi\chi_+(c)}{\sigma_+(c)},\quad
\lim_{\ep\rightarrow 0{\pm}}I_-(c)
=I^{re}_-(c)\pm \frac{i\pi\chi_-(c)}{\sigma_-(c)}.
\end{eqnarray*}
Thus for $c_{\ep}=c+i\ep$ and $c\in D_0\setminus \big\{0,W_+(1),W_-(1),W_+(-1),W_-(-1)\big\}$, we get $P(c_{\ep})\to P(c)$ as $\ep\to 0$ and
\begin{align*}
&\lim_{\ep\rightarrow0\pm}\mathcal{D}(c_{\ep})\\
&=c^2P(c)\Big(I^{re}_+(c)\pm\frac{i\pi\chi_+(c)}{\sigma_+(c)}\Big)
\Big(I^{re}_-(c)\pm\frac{i\pi\chi_-(c)}{\sigma_-(c)}\Big)\\
&\quad-\va_+(0,c)^2\Big(I^{re}_+(c)\pm\frac{i\pi\chi_+(c)}{\sigma_+(c)}\Big)
-\va_-(0,c)^2\Big(I^{re}_-(c)\pm\frac{i\pi\chi_-(c)}{\sigma_-(c)}\Big)\\
&=c^2P(c)\Big(I^{re}_+(c)I^{re}_-(c)
-\frac{\pi^2\chi_+(c)\chi_-(c)}{\sigma_+(c)\sigma_-(c)}\Big)-\va_+(0,c)^2I^{re}_+(c)
-\va_-(0,c)^2I^{re}_-(c)\\
&\quad
\pm i\left(c^2P(c)\Big(\frac{\pi I^{re}_+(c)\chi_-(c)}{\sigma_-(c)}+\frac{\pi\ I^{re}_-(c)\chi_+(c)}
{\sigma_+(c)}\Big)-\frac{\pi\va_+(0,c)^2\chi_+(c)}{\sigma_+(c)}
-\frac{\pi\va_-(0,c)^2\chi_-(c)}{\sigma_-(c)}\right)\\
&=\mathcal{D}^{re}(c)\pm i\mathcal{D}^{im}(c),
\end{align*}
And then we have, for $c\in D_0\setminus \big\{0,W_+(1),W_-(1),W_+(-1),W_-(-1)\big\}$ with $\chi_+(c)\chi_-(c)=1$,
\begin{align*}
&\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2\\
&=c^4P(c)^2I^{re}_+(c)^2I^{re}_-(c)^2
+\frac{\pi^4c^4P(c)^2}{\sigma_+(c)^2\sigma_-(c)^2}
+\va_+(0,c)^4I^{re}_+(c)^2+\va_-(0,c)^4I^{re}_-(c)^2\\
&\quad
-2c^2P(c)\va_+(0,c)^2I^{re}_+(c)^2I^{re}_-(c)
-2c^2P(c)\va_-(0,c)^2I^{re}_+(c)I^{re}_-(c)^2\\
&\quad
+2\va_+(0,c)^2\va_-(0,c)^2I^{re}_+(c)I^{re}_-(c)
+\frac{\pi^2}{\sigma_-(c)^2}c^4P(c)^2I^{re}_+(c)^2\\
&\quad
+\frac{\pi^2}{\sigma_+(c)^2}c^4P(c)^2I^{re}_-(c)^2
+\frac{\pi^2}{\sigma_-(c)^2}\va_-(0,c)^4
+\frac{\pi^2}{\sigma_+(c)^2}\va_+(0,c)^4\\
&\quad
+\frac{2\pi^2\va_+(0,c)^2\va_-(0,c)^2}{\sigma_+(c)\sigma_-(c)}
-\frac{2\pi^2}{\sigma_-(c)^2}c^2P(c)\va_-(0,c)^2I^{re}_+(c)\\
&\quad
-\frac{2\pi^2}{\sigma_+(c)^2}c^2P(c)\va_+(0,c)^2I^{re}_-(c)\\
&=\Big[c^2P(c)I^{re}_+(c)I^{re}_-(c)
-\va_+(0,c)^2I^{re}_+(c)-\va_-(0,c)^2I^{re}_-(c)\Big]^2\\
&\quad +
\frac{\pi^2}{\sigma_-(c)^2}\Big[c^2P(c)I^{re}_+(c)-\va_-(0,c)^2\Big]^2
+\frac{\pi^2}{\sigma_+(c)^2}\Big[c^2P(c)I^{re}_-(c)-\va_+(0,c)^2\Big]^2\\
&\quad
+\frac{\pi^4c^4P(c)^2}{\sigma_+(c)^2\sigma_-(c)^2}
+\frac{2\pi^2\va_+(0,c)^2\va_-(0,c)^2}{\sigma_+(c)\sigma_-(c)}.
\end{align*}
On one hand, for $c\in D_0\setminus \big\{0,W_+(1),W_-(1),W_+(-1),W_-(-1)\big\}$, we have
$$\sigma_+(c)\sigma_-(c)\geq C^{-1}|c|^2>0.$$
Indeed, by the fact that $W'_+(y)\geq C^{-1}>0, W'_-(y)\leq -C^{-1}<0$, we have
\begin{eqnarray*}
&&0>\sigma_+(c)=W'_+(y_{c_+})\big(W_-(y_{c_+})-W_-(y_{c_-})\big)\geq -C^{-1}(y_{c_+}-y_{c_-})\geq -C^{-1}|c|, \\
&&0>\sigma_-(c)=-W'_-(y_{c_+})\big(W_+(y_{c_-})-W_+(y_{c_+})\big)\geq -C^{-1}(y_{c_+}-y_{c_-})\geq -C^{-1}|c|,
\end{eqnarray*}
On the other hand, by Proposition \ref{prop: sol. hom. [0,1]} and Proposition \ref{prop: sol. hom. [-1,0]}, we have
\begin{equation}\label{eq: P est}
\begin{split}
C|c|\geq |P(c)|&=|\va_-(0,c)^2(\va_+\pa_y\va_+)(0,c)-\va_+(0,c)^2(\va_-\pa_y\va_-)(0,c)|\\
&=|\va_-(0,c)^2(\va_+\pa_y\va_+)(0,c)|+|\va_+(0,c)^2(\va_-\pa_y\va_-)(0,c)|\\
&\geq C^{-1}(|y_{c_+}|+|y_{c_-}|)\geq C^{-1}|c|,
\end{split}
\end{equation}
and then
\begin{equation}
\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2\geq \frac{\pi^4c^4P(c)^2}{\sigma_+(c)^2\sigma_-(c)^2}
+\frac{2\pi^2\va_+(0,c)^2\va_-(0,c)^2}{\sigma_+(c)\sigma_-(c)}\geq C^{-1}\big(c^2+\frac{1}{c^2}\big)>C^{-1}.
\end{equation}
For $c\in D_0\setminus \big\{0,W_+(1),W_-(1),W_+(-1),W_-(-1)\big\}$ with $\chi_+(c)=0$ then $\chi_-(c)=1$, we have
\begin{align*}
\mathcal{D}^{im}(c)=\f{\pi}{\sigma_-(c)}(I_+^{re}(c)c^2P(c)-\va_-(0,c)^2)
\end{align*}
In this case $c>W_+(1)$(in Case 3, 8, 9,) or $c<W_-(1)$(in Case 1, 2, 3), then $\mathcal{H}(y,c)>0$ for $y\in [0,1]$, thus $I_+^{re}(c)>0$ and by the fact that $P(c)<-C^{-1}|c|\leq 0$, we obtain that $|\mathcal{D}^{im}(c)|>C^{-1}$ for $c\in (W_+(1), W_-(-1)]$ and $c\in [W_+(-1), W_-(1))$.
For $c\in D_0\setminus \big\{0,W_+(1),W_-(1),W_+(-1),W_-(-1)\big\}$ with $\chi_-(c)=0$ then $\chi_+(c)=1$(in Case 1, 4, 5, 7, 9), we also have for $c\in (W_-(-1),W_+(1)]$ and $c\in [W_-(1),W_+(-1))$
\begin{align*}
|\mathcal{D}^{im}(c)|=\Big|\f{\pi}{\sigma_+(c)}(I_-^{re}(c)c^2P(c)-\va_+(0,c)^2)\Big|\geq C^{-1}.
\end{align*}
Thus we complete the proof of the lemma.
\end{proof}
\begin{lemma}\label{lem: Re sigma}
Let $c_{\ep}\in \Omega_{\ep_0}\setminus\{0\}$. It holds that
\begin{eqnarray*}
\Big|\frac{Re(\sigma_+(c_{\ep}))}{Re(\sigma_-(c_{\ep}))}-1\Big|\leq C|c|.
\end{eqnarray*}
\end{lemma}
\begin{proof}
We have
\begin{equation}\label{Re sigma}
\begin{split}
&Re(\sigma_+(c_{\ep}))=W'_+(y_{c_+})\big(W_-(y_{c_+})-c\big)
-W'_-(y_{c_+})\big(W_+(y_{c_+})-c\big),\\
&Re(\sigma_-(c_{\ep}))=W'_+(y_{c_-})\big(W_-(y_{c_-})-c\big)
-W'_-(y_{c_-})\big(W_+(y_{c_-})-c\big),
\end{split}
\end{equation}
For the case $c\geq 0$, then $W_+(y_{c_+})=c=W_-(y_{c_-})$, we have
\begin{align*}
Re(\sigma_+(c_{\ep}))=W'_+(y_{c_+})\big(W_-(y_{c_+})-c\big),\ Re(\sigma_-(c_{\ep}))=-W'_-(y_{c_-})\big(W_+(y_{c_-})-c\big),
\end{align*}
and then due to $W_+(0)=W_-(0)=0$,
\begin{align*}
Re(\sigma_+(c_{\ep}))&=W'_+(y_{c_+})\big(W_-(y_{c_+})-W_+(y_{c_+})\big)\\
&=-W'_+(y_{c_+})y_{c_+}\lambda_+(y_{c_+}),\\
Re(\sigma_-(c_{\ep}))&=-W'_-(y_{c_-})\big(W_+(y_{c_-})-W_-(y_{c_-})\big)\\
&=-W'_-(y_{c_-})y_{c_-}\lambda_-(y_{c_-}),
\end{align*}
where
\begin{align*}
&\lambda_+(y_{c_+})=-\int_0^1\big[W'_-(sy_{c_+})-W'_+(sy_{c_+})\big]ds, \\
&\lambda_-(y_{c_-})=\int_0^1\big[W'_+(sy_{c_-})-W'_-(sy_{c_-})\big]ds,
\end{align*}
Thus we obtain $\lambda_+(y_{c_+})\geq C^{-1}$, $\lambda_-(y_{c_-})\geq C^{-1}$ and $|Re(\sigma_{\pm})(c_{\ep})|\geq C^{-1}|c|$.
We also have
\begin{align*}
\frac{Re(\sigma_+(c_{\ep}))}{Re(\sigma_-(c_{\ep}))}-1
&
=\frac{\lambda_+(y_{c_+})\big(W'_+(y_{c_+})y_{c_+}-W'_-(y_{c_-})y_{c_-}\big)}
{\lambda_-(y_{c_-})W'_-(y_{c_-})y_{c_-}}+\frac{\lambda_+(y_{c_+})-\lambda_-(y_{c_-})}{\lambda_-(y_{c_-})}.
\end{align*}
Due to $y_{c_-}=W_-^{-1}(c)=W_-^{-1}\big(W_+(y_{c_+})\big)$ and
\begin{align*}
&|W'_+(y_{c_+})y_{c_+}-W'_-(y_{c_-})y_{c_-}|\\
&=\left|\int_0^{y_{c_+}} sW''_+(s)
-(W_-^{-1}\circ W_+)(s)(W''_-\circ W_-^{-1}\circ W_+)(s)
\big((W_-^{-1})'\circ W_+\big)(s)W'_+(s)ds\right|\\
&\leq C \int_0^{y_{c_+}} \left|s\right|
+\left|(W_-^{-1}\circ W_+)(s)\right|ds\leq C|c|^2,
\end{align*}
then we get
\begin{align*}
\Big|\frac{\lambda_+(y_{c_+})\big(W'_+(y_{c_+})y_{c_+}-W'_-(y_{c_-})y_{c_-}\big)}
{\lambda_-(y_{c_-})W'_-(y_{c_-})y_{c_-}}\Big|\leq C|c|.
\end{align*}
On the other hand,
\begin{align*}
&|\lambda_+(y_{c_+})-\lambda_-(y_{c_-})|\\
&=\left|\int_0^1\big[W'_-(sy_{c_+})-W'_-(sy_{c_-})
-W'_+(sy_{c_+})+W'_+(sy_{c_-})\big]ds\right|\\
&=|(y_{c_+}-y_{c_-})|\left|\int_0^1s\int_0^1\big[\big(W''_--W''_+\big)
\big(sy_{c_-}+ts(y_{c_+}-y_{c_-})\big)\big]dtds\right|\leq C|c|.
\end{align*}
For $c\leq 0$, then $W_-(y_{c_+})=c=W_+(y_{c_-})$, we have
\begin{align*}
Re(\sigma_+(c_{\ep}))&=-W'_-(y_{c_+})\big(W_+(y_{c_+})-W_-(y_{c_+})\big)
=-W'_-(y_{c_+})y_{c_+}\wt{\lambda}_+(y_{c_+}),\\
Re(\sigma_-(c_{\ep}))&=W'_+(y_{c_-})\big(W_-(y_{c_-})-W_+(y_{c_-})\big)
=-W'_+(y_{c_-})y_{c_-}\wt{\lambda}_-(y_{c_-}),
\end{align*}
where
\begin{align*}
&\wt{\lambda}_+(y_{c_+})=\int_0^1\big[W'_+(sy_{c_+})-W'_-(sy_{c_+})\big]ds, \\
&\wt{\lambda}_-(y_{c_-})=-\int_0^1\big[W'_-(sy_{c_-})-W'_+(sy_{c_-})\big]ds,
\end{align*}
Thus we obtain $\wt{\lambda}_+(y_{c_+})\geq C^{-1}$, $\wt{\lambda}_-(y_{c_-})\geq C^{-1}$ and $|Re(\sigma_{\pm}(c_{\ep}))|\geq C^{-1}|c|$.
We also have
\begin{align*}
\frac{Re(\sigma_+(c_{\ep}))}{Re(\sigma_-(c_{\ep}))}-1
&
=\frac{\wt{\lambda}_+(y_{c_+})\big(W'_-(y_{c_+})y_{c_+}-W'_+(y_{c_-})y_{c_-}\big)}
{\wt{\lambda}_-(y_{c_-})W'_+(y_{c_-})y_{c_-}}
+\frac{\wt{\lambda}_+(y_{c_+})-\wt{\lambda}_-(y_{c_-})}{\wt{\lambda}_-(y_{c_-})}.
\end{align*}
Due to $y_{c_-}=W_+^{-1}(c)=W_+^{-1}\big(W_-(y_{c_+})\big)$ and
\begin{equation*}
|W'_-(y_{c_+})y_{c_+}-W'_+(y_{c_-})y_{c_-}|\leq C|c|^2,
\end{equation*}
then we get
\begin{equation*}
\Big|\frac{\wt{\lambda}_+(y_{c_+})\big(W'_-(y_{c_+})y_{c_+}-W'_+(y_{c_-})y_{c_-}\big)}
{\wt{\lambda}_-(y_{c_-})W'_+(y_{c_-})y_{c_-}}\Big|\leq C|c|
\end{equation*}
and
\begin{equation*}
|\wt{\lambda}_+(y_{c_+})-\wt{\lambda}_-(y_{c_-})|\leq C|c|.
\end{equation*}
Therefore,
\begin{align*}
\Big|\frac{Re(\sigma_+(c_{\ep}))}{Re(\sigma_-(c_{\ep}))}-1\Big|\leq C|c|.
\end{align*}
This completes the proof of the Lemma.
\end{proof}
Now, we present the proof of Proposition \ref{prop: D est}.
\begin{proof}
The first part of the Proposition follows directly from Lemma \ref{lem: D>0}.
For the case of $|c|<\delta_0$, we have
\begin{align}
\mathcal{D}(c_{\ep})&=c_{\ep}^2P(c_{\ep})I_+(c_{\ep})I_-(c_{\ep})
-\big(\va_+(0,c_{\ep})^2-1\big)I_+(c_{\ep})-\Pi_+(c_{\ep})\nonumber\\
&\ \
-\big(\va_-(0,c_{\ep})^2-1\big)I_-(c_{\ep})
-\Pi_-(c_{\ep})\nonumber\\
&\ \
-\frac{I_1(c_{\ep})+I_2(c_{\ep})+I_3(c_{\ep})}{\sigma_+(c_{\ep})}
-\frac{J_1(c_{\ep})+J_2(c_{\ep})+J_3(c_{\ep})}{\sigma_-(c_{\ep})}\nonumber\\
&=c_{\ep}^2P(c_{\ep})I_+(c_{\ep})I_-(c_{\ep})
-\big(\va_+(0,c_{\ep})^2-1\big)I_+(c_{\ep})-\Pi_+(c_{\ep})\nonumber\\
&\ \
-\big(\va_-(0,c_{\ep})^2-1\big)I_-(c_{\ep})
-\Pi_-(c_{\ep})\nonumber\\
&\ \ -\frac{I_1(c_{\ep})+I_2(c_{\ep})+I_3(c_{\ep})}{\sigma_+(c_{\ep})}
-\frac{J_1(c_{\ep})+J_2(c_{\ep})+J_3(c_{\ep})}{\sigma_+(c_{\ep})}\nonumber\\
&\ \
-\frac{\frac{\sigma_+(c_{\ep})}{\sigma_-(c_{\ep})}-1}{\sigma_+(c_{\ep})}
\big(J_1(c_{\ep})+J_2(c_{\ep})+J_3(c_{\ep})\big),
\end{align}
where we recall in the proof of Lemma \ref{lem: I limit+} and Lemma \ref{lem: I limit-},
\begin{eqnarray*}
\sigma_+(c_{\ep})I_+(c_{\ep})=\sigma_+(c_{\ep})\Pi_+(c_{\ep})+I_1(c_{\ep})+I_2(c_{\ep})+I_3(c_{\ep}),\\
\sigma_-(c_{\ep})I_-(c_{\ep})=\sigma_-(c_{\ep})\Pi_-(c_{\ep})+J_1(c_{\ep})+J_2(c_{\ep})+J_3(c_{\ep}).
\end{eqnarray*}
By Proposition \ref{prop: sol. hom. [0,1]}, Proposition \ref{prop: sol. hom. [-1,0]}, Lemma \ref{lem: I bdd} and \eqref{eq: P est}, we obtain
\begin{align}\label{eq: up bdd1}
&\Big|c_{\ep}^2P(c_{\ep})I_+(c_{\ep})I_-(c_{\ep})-\big(\va_+(0,c_{\ep})^2-1\big)I_+(c_{\ep})
-\big(\va_-(0,c_{\ep})^2-1\big)I_-(c_{\ep})\Big|\nonumber\\
&\leq |c_{\ep}|^2|P(c_{\ep})||I_+(c_{\ep})||I_-(c_{\ep})|+C\big|\va_+(0,c_{\ep})-1\big|
|I_+(c_{\ep})|+C\big|\va_-(0,c_{\ep})-1\big||I_-(c_{\ep})|\nonumber\\
& \leq C|c_{\ep}|.
\end{align}
We also have
\begin{equation*}
Im(I_1(c_{\ep}))=\int_0^1\frac{\ep h(y,y_{c_+})\big(W_+(y)+W_-(y)-2c\big)}
{\big((W_+(y)-c)^2+\ep^2\big)\big((W_-(y)-c)^2+\ep^2\big)}dy,
\end{equation*}
by (\ref{h bdd}) and $|W_+(y)+W_-(y)-2c|\leq C|y-y_{c_-}|$, we have
\begin{equation*}
\Big|\frac{h(y,y_{c_+})\big(W_+(y)+W_-(y)-2c\big)}
{\big((W_+(y)-c)^2+\ep^2\big)\big((W_-(y)-c)^2+\ep^2\big)}\Big|\leq C\frac{|y-y_{c_-}|}{|(y-y_{c_-})^2+\ep^2|}.
\end{equation*}
Thus we can get that for $|c|<\delta_0$ with $\delta_0$ small enough,
\begin{equation*}
|Im(I_1(c_{\ep}))|\leq C\ep|\ln \ep|.
\end{equation*}
And similarly, we have
\begin{equation*}
Im(I_2(c_{\ep}))=\ep\int_0^1\frac{W'_+(y_{c_+})-W'_+(y)}{(W_+(y)-c)^2+\ep^2}dy
-\ep\int_0^1\frac{W'_-(y_{c_+})-W'_-(y)}{(W_-(y)-c)^2+\ep^2}dy,
\end{equation*}
which implies
\begin{equation*}
\big|Im(I_2(c_{\ep}))\big|\leq C\ep|\ln\ep|.
\end{equation*}
By the same argument, we can deduce that for $|c|<\delta_0$,
\begin{equation*}
\big|Im(J_1(c_{\ep})+J_2(c_{\ep}))\big|\leq C\ep|\ln\ep|.
\end{equation*}
Thus from the above, we have
\begin{equation}\label{eq: IJ12 bdd}
\big|Im(I_1(c_{\ep})+I_2(c_{\ep})+J_1(c_{\ep})+J_2(c_{\ep}))\big|\leq C\ep|\ln\ep|.
\end{equation}
Due to $
Im(I_3(c_{\ep})+J_3(c_{\ep}))=2\arctan\frac{W_+(1)-c}{\ep}-2\arctan\frac{W_-(1)-c}{\ep},
$
we get for $|c|\leq \delta_0$,
\begin{equation}\label{eq: IJ3 low bdd}
\big|Im(I_3(c_{\ep})+J_3(c_{\ep}))\big|\geq\frac{3\pi}{2}.
\end{equation}
On the other hand, we have $C^{-1}|c|\leq \big|Re(\sigma_-(c_{\ep}))\big|\leq C|c|$ and
\begin{eqnarray*}
Im(\sigma_+(c_{\ep}))=-\ep \big(W'_+(y_{c_+})-W'_-(y_{c_+})\big),\\
Im(\sigma_-(c_{\ep}))=-\ep \big(W'_+(y_{c_-})-W'_-(y_{c_-})\big),
\end{eqnarray*}
and then
\begin{equation*}
C^{-1}\ep\leq \big|Im(\sigma_-(c_{\ep}))\big|\leq C\ep,
\end{equation*}
and
\begin{align*}
\big|Im(\sigma_+(c_{\ep}))-Im(\sigma_-(c_{\ep}))\big|
&=\ep\left|W'_+(y_{c_+})-W'_+(y_{c_-})-W'_-(y_{c_+})+W'_-(y_{c_-})\right|
\leq C\ep|c|.
\end{align*}
Then from which and by Lemma \ref{lem: Re sigma}, we get
\begin{align}\label{eq: sig bdd}
\Big|\frac{\sigma_+(c_{\ep})}{\sigma_-(c_{\ep})}-1\Big|
&=\left|\frac{Re(\sigma_+(c_{\ep}))-Re(\sigma_-(c_{\ep}))
+i\big(Im(\sigma_+(c_{\ep}))-Im(\sigma_-(c_{\ep}))\big)}
{Re(\sigma_-(c_{\ep}))+iIm(\sigma_-(c_{\ep}))}\right|\nonumber\\
&
\leq\Big|\frac{Re(\sigma_+(c_{\ep}))}
{Re(\sigma_-(c_{\ep}))}-1\Big|
+\Big|\frac{Im(\sigma_+(c_{\ep}))-Im(\sigma_-(c_{\ep}))}
{Re(\sigma_-(c_{\ep}))+iIm(\sigma_-(c_{\ep}))}\Big|\nonumber\\
&\leq C|c|+\frac{C\ep |c|}{C^{-1}(\ep+|c|)}
\leq C|c|.
\end{align}
Therefore by the fact that $|\Pi_{\pm}(c_{\ep})|+|J_1(c_{\ep})|+|J_2(c_{\ep})|\leq C$ and that for $|c|\leq \delta_0$, $|J_3(c_{\ep})|\leq C$ and \eqref{eq: P est}, \eqref{eq: up bdd1}, \eqref{eq: IJ12 bdd}, \eqref{eq: IJ3 low bdd} and \eqref{eq: sig bdd} we obtain for $|c|<\delta_0$
\begin{align}\label{D 0}
|\mathcal{D}(c_{\ep})|&\geq \frac{|I_1(c_{\ep})+I_2(c_{\ep})+I_3(c_{\ep})+J_1(c_{\ep})+J_2(c_{\ep})+J_3(c_{\ep})|}{|\sigma_+(c_{\ep})|}
-\big|\Pi_+(c_{\ep})\big|-\big|\Pi_-(c_{\ep})\big|\nonumber\\
&\ \
-\Big|c_{\ep}^2P(c_{\ep})I_+(c_{\ep})I_-(c_{\ep})
-\big(\va_+(0,c_{\ep})^2-1\big)I_+(c_{\ep})
-\big(\va_-(0,c_{\ep})^2-1\big)I_-(c_{\ep})\Big|\nonumber\\
&\ \ -\Big|\frac{\frac{\sigma_+(c_{\ep})}{\sigma_-(c_{\ep})}-1}{\sigma_+(c_{\ep})}
\big(J_1(c_{\ep})+J_2(c_{\ep})+J_3(c_{\ep})\big)\Big|\nonumber\\
&\geq \frac{\big|Im(I_3(c_{\ep})+J_3(c_{\ep}))\big|}{\big|\sigma_+(c_{\ep})\big|}
-\frac{\big|Im(I_1(c_{\ep})+I_2(c_{\ep})+J_1(c_{\ep})+J_2(c_{\ep}))\big|}
{\big|\sigma_+(c_{\ep})\big|}-C-C|c_{\ep}|\nonumber\\
&\ \ -\frac{\big|\frac{\sigma_1(c_{\ep})}{\sigma_2(c_{\ep})}-1\big|}
{\big|\sigma_+(c_{\ep})\big|}\Big|J_1(c_{\ep})+J_2(c_{\ep})+J_3(c_{\ep})\Big|\nonumber\\
&\geq \frac{C^{-1}}{|c_{\ep}|}-\frac{C\ep|\ln\ep|}{|c_{\ep}|}-C-C|c_{\ep}|
\geq\frac{C^{-1}}{|c_{\ep}|}.
\end{align}
\noindent{\bf Case 1. $W_+(1)\neq W_-(-1)$ and $W_+(-1)\neq W_-(1)$. }\\
For $|c_{\ep}-W_{\pm}(1)|\leq \delta_0$, by Lemma \ref{lem: I limit+}, we get
\begin{align*}
\Big|\frac{c_{\ep}^2P(c_{\ep})I_-(c_{\ep})\big(I_1(c_{\ep})+I_2(c_{\ep})\big)}{\sigma_+(c_{\ep})}\Big|+\Big|\frac{\va_+(0,c_{\ep})^2\big(I_1(c_{\ep})+I_2(c_{\ep})\big)}
{\sigma_+(c_{\ep})}\Big|+\Big|\va_-(0,c_{\ep})^2I_-(c_{\ep})\Big|\leq C,
\end{align*}
and thus, we have
\begin{align*}
|\mathcal{D}(c_{\ep})|&\geq \left|Im\, \left(\frac{c_{\ep}^2P(c_{\ep})I_-(c_{\ep})I_3(c_{\ep})}
{\sigma_+(c_{\ep})}-\frac{\va_+(0,c_{\ep})^2I_3(c_{\ep})}{\sigma_+(c_{\ep})}\right)\right|\\
&\quad
-\Big|\frac{c_{\ep}^2P(c_{\ep})I_-(c_{\ep})\big(I_1(c_{\ep})+I_2(c_{\ep})\big)}{\sigma_+(c_{\ep})}\Big|-\Big|\frac{\va_+(0,c_{\ep})^2\big(I_1(c_{\ep})+I_2(c_{\ep})\big)}
{\sigma_+(c_{\ep})}\Big|-\Big|\va_-(0,c_{\ep})^2I_-(c_{\ep})\Big|\\
&\geq \left|Im\,\Big(\frac{c_{\ep}^2P(c_{\ep})I_-(c_{\ep})I_3(c_{\ep})}
{\sigma_+(c_{\ep})}\Big)
-Im\,\Big(\frac{\va_+(0,c_{\ep})^2I_3(c_{\ep})}{\sigma_+(c_{\ep})}\Big)\right|-C.
\end{align*}
Due to
\begin{align*}
&Im\Big(\frac{c_{\ep}^2P(c_{\ep})I_-(c_{\ep})I_3(c_{\ep})}
{\sigma_+(c_{\ep})}\Big)
-Im\Big(\frac{\va_+(0,c_{\ep})^2I_3(c_{\ep})}{\sigma_+(c_{\ep})}\Big)\\
&=Im\Big(\frac{c_{\ep}^2P(c_{\ep})I_-(c_{\ep})}
{\sigma_+(c_{\ep})}\Big)Re(I_3(c_{\ep}))+Re\Big(\frac{c_{\ep}^2P(c_{\ep})I_-(c_{\ep})}
{\sigma_+(c_{\ep})}\Big)Im(I_3(c_{\ep}))\\
&\ \
-Im\Big(\frac{\va_+(0,c_{\ep})^2}{\sigma_+(c_{\ep})}\Big)Re(I_3(c_{\ep}))
-Re\Big(\frac{\va_+(0,c_{\ep})^2}{\sigma_+(c_{\ep})}\Big)Im(I_3(c_{\ep}))\\
&=Re\Big(\frac{c_{\ep}^2P(c_{\ep})}{\sigma_+(c_{\ep})}\Big)Im(I_-(c_{\ep}))Re(I_3(c_{\ep}))
+Im\Big(\frac{c_{\ep}^2P(c_{\ep})}{\sigma_+(c_{\ep})}\Big)Re(I_-(c_{\ep}))Re(I_3(c_{\ep}))\\
&\ \
-Im\Big(\frac{\va_+(0,c_{\ep})^2}{\sigma_+(c_{\ep})}\Big)Re(I_3(c_{\ep}))
+Re\Big(\frac{c_{\ep}^2P(c_{\ep})I_-(c_{\ep})}
{\sigma_+(c_{\ep})}\Big)Im(I_3(c_{\ep}))\\
&\ \
-Re\Big(\frac{\va_+(0,c_{\ep})^2}{\sigma_+(c_{\ep})}\Big)Im(I_3(c_{\ep}))\\
&=\frac{c^2P(c)}{\sigma_+(c)}Im(I_-(c_{\ep}))Re(I_3(c_{\ep}))
-\frac{\va_+(0,c)^2}{\sigma_+(c)}Im(I_3(c_{\ep}))\\
&\ \
+Re\Big(\frac{c_{\ep}^2P(c_{\ep})}{\sigma_+(c_{\ep})}
-\frac{c^2P(c)}{\sigma_+(c)}\Big)Im(I_-(c_{\ep}))Re(I_3(c_{\ep}))
+Im\Big(\frac{c_{\ep}^2P(c_{\ep})}{\sigma_+(c_{\ep})}\Big)Re(I_-(c_{\ep}))Re(I_3(c_{\ep}))\\
&\ \
-Im\Big(\frac{\va_+(0,c_{\ep})^2}{\sigma_+(c_{\ep})}\Big)Re(I_3(c_{\ep}))
+Re\Big(\frac{c_{\ep}^2P(c_{\ep})I_-(c_{\ep})}
{\sigma_+(c_{\ep})}\Big)Im(I_3(c_{\ep}))\\
&\ \
-Re\Big(\frac{\va_+(0,c_{\ep})^2}{\sigma_+(c_{\ep})}
-\frac{\va_+(0,c)^2}{\sigma_+(c)}\Big)Im(I_3(c_{\ep})),
\end{align*}
and by Remark \ref{rmk: epsilon},
\begin{align*}
&\Big|Re\Big(\frac{c_{\ep}^2P(c_{\ep})}{\sigma_+(c_{\ep})}
-\frac{c^2P(c)}{\sigma_+(c)}\Big)\Big|\leq C|\ep|,\\
&\Big|Re\Big(\frac{\va_+(0,c_{\ep})^2}{\sigma_+(c_{\ep})}
-\frac{\va_+(0,c)^2}{\sigma_+(c)}\Big)\Big|\leq C|\ep|,\\
&\Big|Im\Big(\frac{c_{\ep}^2P(c_{\ep})}{\sigma_+(c_{\ep})}\Big)\Big|
=\Big|Im\Big(\frac{c_{\ep}^2P(c_{\ep})}{\sigma_+(c_{\ep})}
-\frac{c^2P(c)}{\sigma_+(c)}\Big)\Big|\leq C|\ep|,\\
&\Big|Im\Big(\frac{\va_+(0,c_{\ep})^2}{\sigma_+(c_{\ep})}\Big)\Big|
=\Big|Im\Big(\frac{\va_+(0,c_{\ep})^2}{\sigma_+(c_{\ep})}
-\frac{\va_+(0,c)^2}{\sigma_+(c)}\Big)\Big|\leq C|\ep|.
\end{align*}
Thus by taking $\ep_0$ small enough, we have for any $|\ep|\leq \ep_0$, $|Im(I_-(c_{\ep}))|\geq |Im(J_3(c_{\ep}))|\geq\frac{3\pi}{4}$, $\frac{c^2P(c)}{\sigma_+(c)}\geq C^{-1}$, $\frac{\va_+(0,c)^2}{\sigma_+(c)}\geq C^{-1}$.
Thus by taking $\delta_0$ small enough, we have for $|c_{\ep}-W_{\pm}(1)|\leq \delta_0$,
\begin{align*}
|\mathcal{D}(c_{\ep})|
&\geq\Big|\frac{c^2P(c)}{\sigma_+(c)}Im(I_-(c_{\ep}))Re(I_3(c_{\ep}))\Big|
-\Big|\frac{\va_+(0,c)^2}{\sigma_+(c)}Im(I_3(c_{\ep}))\Big|\\
&\ \
-\Big|Re\Big(\frac{c_{\ep}^2P(c_{\ep})}{\sigma_+(c_{\ep})}
-\frac{c^2P(c)}{\sigma_+(c)}\Big)Im(I_-(c_{\ep}))Re(I_3(c_{\ep}))\Big|\\
&\ \
-\Big|Im\Big(\frac{c_{\ep}^2P(c_{\ep})}{\sigma_+(c_{\ep})}\Big)
Re(I_-(c_{\ep}))Re(I_3(c_{\ep}))\Big|
-\Big|Im\Big(\frac{\va_+(0,c_{\ep})^2}{\sigma_+(c_{\ep})}\Big)Re(I_3(c_{\ep}))\Big|\\
&\ \ -\Big|Re\Big(\frac{c_{\ep}^2P(c_{\ep})I_-(c_{\ep})}
{\sigma_+(c_{\ep})}\Big)Im(I_3(c_{\ep}))\Big|
+\Big|
-Re\Big(\frac{\va_+(0,c_{\ep})^2}{\sigma_+(c_{\ep})}
-\frac{\va_+(0,c)^2}{\sigma_+(c)}\Big)Im(I_3(c_{\ep}))\Big|\\
&\geq C^{-1}|Re(I_3(c_{\ep}))|-C\ep|Re(I_3(c_{\ep}))|-C\\
&\geq C^{-1} l(c_{\ep}-W_{\pm}(1)).
\end{align*}
For $|c_{\ep}-W_{\pm}(-1)|\leq \delta_0$, by Lemma \ref{lem: I limit-}, we get
\begin{align*}
\Big|\frac{c_{\ep}^2P(c_{\ep})I_+(c_{\ep})\big(J_1(c_{\ep})+J_2(c_{\ep})\big)}
{\sigma_-(c_{\ep})}\Big|+\Big|\frac{\va_-(0,c_{\ep})^2\big(J_1(c_{\ep})+J_2(c_{\ep})\big)}
{\sigma_-(c_{\ep})}\Big|+\Big|\va_+(0,c_{\ep})^2I_+(c_{\ep})\Big|\leq C.
\end{align*}
By the same method, we have
\begin{align*}
|\mathcal{D}(c_{\ep})|&\geq\Big|Im\Big(\frac{c_{\ep}^2P(c_{\ep})I_+(c_{\ep})J_3(c_{\ep})}
{\sigma_-(c_{\ep})}\Big)
-Im\Big(\frac{\va_-(0,c_{\ep})^2J_3(c_{\ep})}{\sigma_-(c_{\ep})}\Big)\Big|\\
&\quad
-\Big|\frac{c_{\ep}^2P(c_{\ep})I_+(c_{\ep})\big(J_1(c_{\ep})+J_2(c_{\ep})\big)}
{\sigma_-(c_{\ep})}-\frac{\va_-(0,c_{\ep})^2\big(J_1(c_{\ep})+J_2(c_{\ep})\big)}
{\sigma_-(c_{\ep})}-\va_+(0,c_{\ep})^2I_+(c_{\ep})\Big|\\
&\geq \Big|\frac{c^2P(c)}{\sigma_-(c)}Im(I_+(c_{\ep}))Re(J_3(c_{\ep}))\Big|
-\Big|\frac{\va_-(0,c)^2}{\sigma_-(c)}Im(J_3(c_{\ep}))\Big|\\
&\quad
-\Big|Re\Big(\frac{c_{\ep}^2P(c_{\ep})}{\sigma_-(c_{\ep})}
-\frac{c^2P(c)}{\sigma_-(c)}\Big)Im(I_+(c_{\ep}))Re(J_3(c_{\ep}))\Big|\\
&\quad
-\Big|Im\Big(\frac{c_{\ep}^2P(c_{\ep})}{\sigma_-(c_{\ep})}\Big)
Re(I_+(c_{\ep}))Re(J_3(c_{\ep}))\Big|
-\Big|Im\Big(\frac{\va_-(0,c_{\ep})^2}{\sigma_-(c_{\ep})}\Big)Re(J_3(c_{\ep}))\Big|\\
&\quad -\Big|Re\Big(\frac{c_{\ep}^2P(c_{\ep})I_+(c_{\ep})}
{\sigma_-(c_{\ep})}\Big)Im(J_3(c_{\ep}))\Big|
-\Big|
Re\Big(\frac{\va_-(0,c_{\ep})^2}{\sigma_-(c_{\ep})}
-\frac{\va_-(0,c)^2}{\sigma_-(c)}\Big)Im(J_3(c_{\ep}))\Big|-C\\
&\geq C^{-1}|Re(J_3(c_{\ep}))|-C\ep|Re(J_3(c_{\ep}))|-C\ep\\
&\geq C^{-1} l(c_{\ep}-W_{\pm}(-1)).
\end{align*}
\noindent{\bf{Case 2. $W_+(1)=W_-(-1)$ and $W_+(-1)\neq W_-(1)$.}}\\
For $|c_{\ep}-W_+(1)|=|c_{\ep}-W_-(-1)|< \delta_0$, by Lemma \ref{lem: I bdd} and \eqref{eq: P est}, we get
\begin{align*}
\big|\mathcal{D}(c_{\ep})\big|
&=\Big|c_{\ep}^2P(c_{\ep})I_+(c_{\ep})I_-(c_{\ep})-\va_+(0,c_{\ep})^2I_+(c_{\ep})
-\va_-(0,c_{\ep})^2I_-(c_{\ep})\Big|\nonumber\\
&\geq |c_{\ep}|^2|P(c_{\ep})||I_+(c_{\ep})||I_-(c_{\ep})|-C
|I_+(c_{\ep})|-C|I_-(c_{\ep})|\nonumber\\
& \geq C^{-1}l(c_{\ep}-W_+(1))l(c_{\ep}-W_-(-1))-Cl(c_{\ep}-W_+(1))
-Cl(c_{\ep}-W_-(-1))\\
&\geq C^{-1}l(c_{\ep}-W_+(1))l(c_{\ep}-W_-(-1)).
\end{align*}
And for $|c_{\ep}-W_-(1)|<\delta_0$, we can obtain by the same argument as in Case 1 that
\begin{align*}
\big|\mathcal{D}(c_{\ep})\big|\geq C^{-1}l(c_{\ep}-W_-(1))
\end{align*}
and for $|c_{\ep}-W_+(-1)|<\delta_0$,
\begin{align*}
\big|\mathcal{D}(c_{\ep})\big|\geq C^{-1}l(c_{\ep}-W_+(-1)).
\end{align*}
\noindent{\bf{Case 3. $W_+(1)\neq W_-(-1)$ and $W_+(-1)=W_-(1)$.}}\\
The proof is similar to Case 2 and we have for $|c_{\ep}-W_+(-1)|=|c_{\ep}-W_-(1)|< \delta_0$,
\begin{align*}
\big|\mathcal{D}(c_{\ep})\big|
&=\Big|c_{\ep}^2P(c_{\ep})I_+(c_{\ep})I_-(c_{\ep})-\va_+(0,c_{\ep})^2I_+(c_{\ep})
-\va_-(0,c_{\ep})^2I_-(c_{\ep})\Big|\nonumber\\
&\geq |c_{\ep}|^2|P(c_{\ep})||I_+(c_{\ep})||I_-(c_{\ep})|-C
|I_+(c_{\ep})|-C|I_-(c_{\ep})|\nonumber\\
& \geq C^{-1}l(c_{\ep}-W_+(-1))l(c_{\ep}-W_-(1))-Cl(c_{\ep}-W_+(-1))
-Cl(c_{\ep}-W_-(1))\\
&\geq C^{-1}l(c_{\ep}-W_+(-1))l(c_{\ep}-W_-(1)).
\end{align*}
And we have, for $|c_{\ep}-W_+(1)|<\delta_0$,
\begin{align*}
\big|\mathcal{D}(c_{\ep})\big|\geq C^{-1}l(c_{\ep}-W_+(1))
\end{align*}
and for $|c_{\ep}-W_-(-1)|<\delta_0$,
\begin{align*}
\big|\mathcal{D}(c_{\ep})\big|\geq C^{-1}l(c_{\ep}-W_-(-1)).
\end{align*}
\noindent{\bf Case 4. $W_+(1)=W_-(-1)$ and $W_-(1)=W_+(-1)$.} \\
For the case $|c_{\ep}-W_+(1)|<\delta_0$, by Lemma \ref{lem: I bdd} and \eqref{eq: P est}, we get
\begin{align}
\big|\mathcal{D}(c_{\ep})\big|
&=\Big|c_{\ep}^2P(c_{\ep})I_+(c_{\ep})I_-(c_{\ep})-\va_+(0,c_{\ep})^2I_+(c_{\ep})
-\va_-(0,c_{\ep})^2I_-(c_{\ep})\Big|\nonumber\\
&\geq |c_{\ep}|^2|P(c_{\ep})||I_+(c_{\ep})||I_-(c_{\ep})|-C
|I_+(c_{\ep})|-C|I_-(c_{\ep})|\nonumber\\
& \geq C^{-1}\Big(1+\Big|\ln\big|W_+(1)-c_{\ep}\big|\Big|\Big)^2
-C\Big(1+\Big|\ln\big|W_+(1)-c_{\ep}\big|\Big|\Big)\nonumber\\
&\geq C^{-1}\Big(1+\Big|\ln\big|W_+(1)-c_{\ep}\big|\Big|\Big)^2.
\end{align}
Similarly, we can deduce that for the case of $|c_{\ep}-W_-(1)|<\delta_0$,
\begin{equation*}
|\mathcal{D}(c_{\ep})|\geq C^{-1}\Big(1+\Big|\ln\big|W_-(1)-c_{\ep}\big|\Big|\Big)^2.
\end{equation*}
For the case $c_{\ep}\in\Omega_{\ep_0}\setminus D_0$ with $|c|\geq \delta_0$, $|c_{\ep}-W_+(1)|\geq \delta_0$ and $|c_{\ep}-W_-(1)|\geq \delta_0$, by the fact that $\mathcal{D}(c_{\ep})$ is continuous to the boundary and Lemma \ref{lem: D>0}, we have $|\mathcal{D}(c_{\ep})|\geq C^{-1}$ in this case.
Thus, from the above argument, we can deduce that for $c_{\ep}\in\Omega_{\ep_0}\setminus D_0$,
\begin{eqnarray*}
|\mathcal{D}(c_{\ep})|\geq \frac{l\big(W_+(1)-c_{\ep}\big)l
\big(W_-(1)-c_{\ep}\big)l\big(W_+(-1)-c_{\ep}\big)l
\big(W_-(-1)-c_{\ep}\big)}{C|c_{\ep}|}.
\end{eqnarray*}
This completes the proof of the proposition.
\end{proof}
\begin{remark}\label{rmk: D}
The above proposition implies for $\ep_0$ small enough
and $c_{\ep}\in \Omega_{\ep_0}\setminus D_0$, $|\mathcal{D}(c_{\ep})|$
is lower bounded and $\frac{1}{\mathcal{D}}$ is well-defined
in $\Omega_{\ep_0}\setminus D_0$ and
\begin{eqnarray*}
\Big|\frac{1}{\mathcal{D}(c_{\ep})}\Big|\leq \frac{C|c_{\ep}|}{l\big(W_+(1)-c_{\ep}\big)l
\big(W_-(1)-c_{\ep}\big)l\big(W_+(-1)-c_{\ep}\big)l
\big(W_-(-1)-c_{\ep}\big)}.
\end{eqnarray*}
And we have for $c_{\ep}=c+i\ep$ with
$c\in D_0\setminus \big\{0,W_+(1),W_-(1),W_+(-1),W_-(-1)\big\}$,
\begin{eqnarray*}
\lim_{\ep\rightarrow0{\pm}}\frac{1}{\mathcal{D}(c_{\ep})}
=\frac{1}{\mathcal{D}^{re}(c)\pm i\mathcal{D}^{im}(c)}.
\end{eqnarray*}
Moreover, $\frac{1}{\mathcal{D}}$ can be continuous extend
to the boundary with $\frac{1}{\mathcal{D}(0)}=\frac{1}{\mathcal{D}(W_-(1))}=\frac{1}{\mathcal{D}(W_-(-1))}=\frac{1}{\mathcal{D}(W_+(-1))}
=\frac{1}{\mathcal{D}(W_+(1))}=0$.
And then we have for $c\in D_0$,
\begin{eqnarray*}
\Big|\frac{1}{\mathcal{D}^{re}(c)\pm i\mathcal{D}^{im}(c)}\Big|\leq \frac{C |c|}
{l\big(W_+(1)-c\big)l
\big(W_-(1)-c\big)l\big(W_+(-1)-c\big)l
\big(W_-(-1)-c\big)}.
\end{eqnarray*}
\end{remark}
The upper bound of $\frac{1}{\mathcal{D}^{re}(c)\pm i\mathcal{D}^{im}(c)}$ follows from Lemma \ref{lem: I bdd} and Lemma \ref{lem: D>0}. We omit the proof of this remark.
\begin{lemma}\label{lem: limit up bdd}
For $c\in D_0\setminus\{0, W_+(1), W_+(-1), W_-(1),W_-(-1)\}$, there exists a positive constant $C$ such that
\begin{eqnarray*}
|I^{re}_{\pm}(c)|\leq \frac{Cl\big(W_+(\pm1)-c\big)
l\big(W_-(\pm1)-c\big)}{|c|},
\end{eqnarray*}
and
\begin{eqnarray*}
|\mathcal{D}^{re}(c)|
\leq\frac{Cl\big(W_+(1)-c\big)
l\big(W_-(1)-c\big)l\big(W_+(-1)-c\big)
l\big(W_-(-1)-c\big)}{|c|},
\end{eqnarray*}
and
\begin{eqnarray*}
|\mathcal{D}^{im}(c)|
\leq\frac{Cl\big(W_+(1)-c\big)
l\big(W_-(1)-c\big)}{|c|}+\frac{Cl\big(W_+(-1)-c\big)
l\big(W_-(-1)-c\big)}{|c|}.
\end{eqnarray*}
\end{lemma}
\begin{proof}
From Remark \ref{rmk: sigma bdd} and the fact that $|\Pi_{\pm}(c)|+|R^1_{\pm}(c)|+|R^2_{\pm}(c)|\leq C$ for $c\in D_0\setminus\{0, W_+(1), W_+(-1), W_-(1),W_-(-1)\}$, we can easily get the estimate of $I^{re}_{\pm}$.
The estimate of $\mathcal{D}^{re}(c)$ and $\mathcal{D}^{im}(c)$ for $c\in D_0\setminus\{0, W_+(1), W_+(-1), W_-(1),W_-(-1)\}$ follows from the estimate of $I^{re}_{\pm}$ and Proposition \ref{prop: sol. hom. [0,1]}, Proposition \ref{prop: sol. hom. [-1,0]} and \eqref{eq: P est}.
\end{proof}
\subsection{Determine the coefficients}
Now we solve the inhomogeneous Sturmian equation,
for $c\in \Omega_{\ep_0}\setminus D_0$,
\begin{equation}\label{eq: inhomo Sturmian}
\left\{\begin{array}{l}
\pa_y\Big(\mathcal{H}(y,c)\pa_y\Th(y,c)\Big)-\al^2\mathcal{H}(y,c)\Th(y,c)=F(y,c)\\
\Th(-1,c)=\Th(1,c)=0,\\
\end{array}\right.
\end{equation}
here $F(y,c)=cG(\al,y,c)$ and recall $G(\al, y,c)=G_1(\al, y,c)-\f{\widehat{\phi}_0(\al,0)f(\al,y,c)}{b(y)^3}$ defined as in \eqref{eq: G_1} and \eqref{eq: f}.
In particular, we can get that for $c=0$,
\begin{eqnarray*}
F(y,0)=-\widehat{\phi}_0(0)\left(\Big(\big(u(y)^2-b(y)^2\big)\Big(\frac{\chi}{b}(y)\Big)'\Big)'
-\al^2\big(u(y)^2-b(y)^2\big)\frac{\chi}{b}(y)\right).
\end{eqnarray*}
For $c\in \Omega_{\ep_0}\setminus\{0\}$, let
\begin{eqnarray*}
T_{\pm}(F)(c)=\int_0^{\pm1}\frac{\int_{y_{c_{\pm}}}^{y}F(z,c_{\ep})\va_{\pm}(z,c_{\ep})dz}
{\mathcal{H}(y,c_{\ep})\va_{\pm}(y,c_{\ep})^2}dy,
\end{eqnarray*}
and
\begin{eqnarray*}
L(F)(c)=\va_-(0,c)\int^{y_{c_+}}_0F(y,c)\va_+(y,c)dy
-\va_+(0,c)\int^{y_{c_-}}_0F(y,c)\va_-(y,c)dy.
\end{eqnarray*}
For $y\in[0,1]$ and $c\in \Om_{\ep_0}\setminus D_0$, let
\begin{equation}\label{eq: Theta_0+}
\begin{split}
\Th_+^0(y,c)&=\va_+(y,c)\int_0^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c)dz}
{\mathcal{H}(y',c)\va_+(y',c)^2}dy'\\
&\quad+\wt{\mu}_+(F)(c) \va_+(y,c)\int_0^y\frac{1}{\mathcal{H}(y',c)\va_+(y',c)^2}dy'
+\nu_+(F)(c)\va_+(y,c)
\end{split}
\end{equation}
and
\begin{equation}\label{eq: Theta_1+}
\begin{split}
\Th_+^1(y,c)&=\va_+(y,c)\int_1^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c)dz}
{\mathcal{H}(y',c)\va_+(y',c)^2}dy'\\
&\quad+\mu_+(F)(c) \va_+(y,c)\int_1^y\frac{1}{\mathcal{H}(y',c)\va_+(y',c)^2}dy',
\end{split}
\end{equation}
with
\begin{equation}\label{mu+}
\begin{split}
\mu_+(F)(c)=\wt{\mu}_+(F)(c)&=\frac{1}{\mathcal{D}(c)}\big[-c^2P(c)T_+(F)(c)I_-(c)
-\va_-(0,c)L(F)(c)I_-(c)\\
&\quad
+\va_+(0,c)^2T_+(F)(c)-(\va_+\va_-)(0,c)T_-(F)(c)\big],
\end{split}
\end{equation}
and
\begin{equation}\label{nu+}
\begin{split}
\nu_+(F)(c)&=\frac{1}{\mathcal{D}(c)}\big[\va_-(0,c)L(F)(c)I_+(c)I_-(c)
-(\va_+\va_-)(0,c)T_-(F)(c)I_+(c)\\
&\quad +\va_-(0,c)^2T_+(F)(c)I_-(c)\big].
\end{split}
\end{equation}
For $y\in[-1,0]$ and $c\in \Om_{\ep_0}\setminus D_0$, let
\begin{equation}\label{eq: Theta_-1-}
\begin{split}
\Th_-^{-1}(y,c)&=\va_-(y,c)\int_{-1}^y\frac{\int_{y_{c_-}}^{y'}(F\va_-)(y'',c)dy''}{\mathcal{H}(y',c)\va_-(y',c)^2}dy'\\
&\quad+\mu_-(F)(c) \va_-(y,c)\int_{-1}^y\frac{1}{\mathcal{H}(y',c)\va_-(y',c)^2}dy',
\end{split}
\end{equation}
and
\begin{equation}\label{eq: Theta_0-}
\begin{split}
\Th_-^0(y,c)&=\va_-(y,c)\int_0^y\frac{\int_{y_{c_-}}^{y'}(F\va_-)(y'',c)dy''}
{\mathcal{H}(y',c)\va_-(y',c)^2}dy'\\
&\quad +\wt{\mu}_-(F)(c) \va_-(y,c)\int_0^y\frac{1}{\mathcal{H}(y',c)\va_-(y',c)^2}dy'
+\nu_-(F)(c)\va_-(y,c),
\end{split}
\end{equation}
with
\begin{equation}\label{mu-}
\begin{split}
\mu_-(F)(c)=\wt{\mu}_-(F)(c)&=\frac{1}{\mathcal{D}(c)}\big[c^2P(c)T_-(F)(c)I_+(c)
+\va_+(0,c)L(F)(c)I_+(c)\\
&\ \
+(\va_+\va_-)(0,c)T_+(F)(c)-\va_-(0,c)^2T_-(F)(c)\big],
\end{split}
\end{equation}
\begin{equation}\label{nu-}
\begin{split}
\nu_-(F)(c)&=\frac{1}{\mathcal{D}(c)}\big[\va_+(0,c)L(F)(c)I_+(c)I_-(c)
+\va_+(0,c)^2T_-(F)(c)I_+(c)\\
&\quad+(\va_+\va_-)T_+(F)(c)I_-(c)\big].
\end{split}
\end{equation}
\begin{proposition}\label{prop: inhom solu}
Let $c\in \Omega_{\ep_0}\setminus D_0$. Then for $y\in[0,1]$, $\Th_{+}^0(y,c)\equiv \Th_{+}^1(y,c)\buildrel\hbox{\footnotesize def}\over = \Th_+(y,c)$ and for $y\in[-1,0]$, $\Th_{-}^0(y,c)\equiv \Th_{-}^1(y,c)\buildrel\hbox{\footnotesize def}\over = \Th_-(y,c)$. Moreover,
\begin{eqnarray*}
\Th(y,c)=\left\{\begin{array}{l}
\Th_+(y,c), \ \ y\in[0,1],\\
\Th_-(y,c), \ \ y\in[-1,0],\\
\end{array}\right.
\end{eqnarray*}
is the unique $C^1([-1,1])$ solution to \eqref{eq: inhomo Sturmian}.
\end{proposition}
\begin{proof}
Recall the solution $\va_{\pm}(y,c)$ of \eqref{eq: homo eq} obtained in Proposition \ref{prop: sol. hom. [0,1]} and Proposition \ref{prop: sol. hom. [-1,0]}. Then it is easy to check that the solution of \eqref{eq: inhomo Sturmian} satisfies
\begin{eqnarray*}
\pa_y\Big(\mathcal{H}(y,c)\va_{\pm}(y,c)^2\pa_y\big(\frac{\Th}{\va_{\pm}}\big)(y,c)\Big)
=\va_{\pm}(y,c)F(y,c).
\end{eqnarray*}
Then the solution of \eqref{eq: inhomo Sturmian} must have the following forms: \\
For $y\in[0,1]$ and $c\in \Om_{\ep_0}\setminus D_0$,
\begin{eqnarray*}
\begin{split}
\Th(y,c)&=\va_+(y,c)\int_0^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(y'',c)dy''}{\mathcal{H}(y',c)\va_+(y',c)^2}dy'\\
&\quad+\wt{\mu}_+(F)(c) \va_+(y,c)\int_0^y\frac{1}{\mathcal{H}(y',c)\va_+(y',c)^2}dy'
+\nu_+(F)(c)\va_+(y,c)
\end{split}
\end{eqnarray*}
and
\begin{eqnarray*}
\begin{split}
\Th(y,c)&=\va_+(y,c)\int_1^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(y'',c)dy''}{\mathcal{H}(y',c)\va_+(y',c)^2}dy'\\
&\quad+\mu_+(F)(c) \va_+(y,c)\int_1^y\frac{1}{\mathcal{H}(y',c)\va_+(y',c)^2}dy'.
\end{split}
\end{eqnarray*}
For $y\in[-1,0]$ and $c\in \Om_{\ep_0}\setminus D_0$,
\begin{eqnarray*}
\begin{split}
\Th(y,c)&=\va_-(y,c)\int_{-1}^y\frac{\int_{y_{c_-}}^{y'}(F\va_-)(y'',c)dy''}{\mathcal{H}(y',c)\va_-(y',c)^2}dy'\\
&\quad+\mu_-(F)(c) \va_-(y,c)\int_{-1}^y\frac{1}{\mathcal{H}(y',c)\va_-(y',c)^2}dy',
\end{split}
\end{eqnarray*}
and
\begin{eqnarray*}
\begin{split}
\Th(y,c)&=\va_-(y,c)\int_0^y\frac{\int_{y_{c_-}}^{y'}(F\va_-)(y'',c)dy''}{\mathcal{H}(y',c)\va_-(y',c)^2}dy'\\
&\quad +\wt{\mu}_-(F)(c) \va_-(y,c)\int_0^y\frac{1}{\mathcal{H}(y',c)\va_-(y',c)^2}dy'
+\nu_-(F)(c)\va_-(y,c).
\end{split}
\end{eqnarray*}
Using the boundary condition and the fact that $\Th(y,c)$ is a $C^1([-1,1])$ function, we obtain that the coefficients are determined by the following equation:
\begin{equation*}
\left\{\begin{array}{l}
\mu_+(F)(c)=\wt{\mu}_+(F)(c), \ \
\mu_-(F)(c)=\wt{\mu}_-(F)(c),\\
I_+(c)\mu_+(F)(c)+\nu_+(F)(c)=-T_+(F)(c),\\
I_-(c)\mu_-(F)(c)-\nu_-(F)(c)=T_-(F)(c),\\
\va_+(0,c)\nu_+(F)(c)-\va_-(0,c)\nu_-(F)(c)=0,\\
\va_-(0,c)\mu_+(F)(c)-\va_+(0,c)\mu_-(F)(c)
+c^2(\va_-\va_+\pa_y\va_+)(0,c)\nu_+(F)(c)\\
\ \ -c^2(\va_+\va_-\pa_y\va_-)(0,c)\nu_-(F)(c)=L(F)(c),
\end{array}\right.
\end{equation*}
which is
\begin{gather*}
W
\begin{bmatrix}
\mu_+(F)(c)\\
\mu_-(F)(c)\\
\nu_+(F)(c)\\
\nu_-(F)(c)
\end{bmatrix}
=
\begin{bmatrix}
-T_+(F)(c)\\
T_-(F)(c)\\
0\\
L(F)(c)
\end{bmatrix}.
\end{gather*}
Therefore we get from Lemma \ref{stern lemma},
\begin{eqnarray*}
\det(W)=c^2P(c)I_+(c)I_-(c)-\va_+(0,c)^2I_+(c)-\va_-(0,c)^2I_-(c)
= \mathcal{D}(c)\neq0.
\end{eqnarray*}
Thus, by solving the matrix equations (\ref{matrix}),
we can deduce that $\mu_{\pm}(F)(c), \wt{\mu}_{\pm}(F)(c),\nu_{\pm}(F)(c)$
satisfy (\ref{mu+}), (\ref{nu+}), (\ref{mu-}) and (\ref{nu-}). The fact $\Th_{+}^0(y,c)\equiv \Th_{+}^1(y,c)$ and $\Th_{-}^0(y,c)\equiv \Th_{-}^1(y,c)$ can be obtained by the construction. The uniqueness of the solution can be obtained by Lemma \ref{stern lemma}.
Thus we proved the proposition.
\end{proof}
\subsection{The behavior of inhomogeneous solution}
\begin{lemma}\label{lem: T(F)}
Suppose $c_{\ep}=c+i\ep\in D_{\ep_0}$ with $c\in D_0\setminus\{0\}$ and $F\in C([-a,a]\times\Om_{\ep})$. Then we have
\begin{equation*}
\lim_{\ep\rightarrow0} T_{\pm}(F)(c_{\ep})=T_{\pm}(F)(c).
\end{equation*}
Suppose $c_{\ep}\in\Omega_{\ep_0}\setminus D_0$. Then there exists a constant $C>0$ such that,
\begin{eqnarray*}
\big|T_{\pm}(F)(c_{\ep})\big|\leq C\|F\|_{L^{\infty}}l\big(c_{\ep}\big).
\end{eqnarray*}
\end{lemma}
\begin{proof}
It is easy to check that $|\mathcal{H}(y,c_{\ep})|\geq C^{-1}\big(\big|y^2-y_{c_{\pm}}^2\big|+\ep^2\big)$.
By Proposition \ref{prop: sol. hom. [0,1]} and Proposition \ref{prop: sol. hom. [-1,0]}, we have $C\geq |\va_{\pm}(y,c_{\ep})|\geq \frac{1}{2}$ and
\begin{eqnarray*}
\Big|\frac{\int_{y_{c_{\pm}}}^{y}F(z,c_{\ep})\va_{\pm}(z,c_{\ep})dz}{\mathcal{H}(y,c_{\ep})\va_{\pm}(y,c_{\ep})^2}
\Big|
\leq C\|F(y,c_{\ep})\|_{L^{\infty}}\|\va_{\pm}\|_{L^{\infty}}
\frac{|y-y_{c_{\pm}}|}{|\mathcal{H}(y,c_{\ep})|}\leq \f{C}{|y|+|c_{\ep}|},
\end{eqnarray*}
which directly implies
\begin{eqnarray*}
\big|T_{\pm}(F)(c_{\ep})\big|\leq C\|F\|_{L^{\infty}}\Big(\big|\ln|c_{\ep}|\big|+1\Big).
\end{eqnarray*}
Since $F(y,c_{\ep})$, $\mathcal{H}(y,c_{\ep})$ and $\va_{\pm}(y,c_{\ep})$ are continuous functions, then by Lebesgue's dominated convergence theorem, as $\ep\rightarrow0$, it holds that
\begin{equation*}
\lim_{\ep\rightarrow0} T_{\pm}(F)(c_{\ep})=T_{\pm}(F)(c).
\end{equation*}
Thus we complete the proof of the lemma.
\end{proof}
\begin{remark}\label{rmk: T(F)}
For $c\in D_0\setminus \{0\}$, there is a constant $C>0$ such that,
\begin{eqnarray*}
\big|T_{\pm}(F)(c)\big|\leq C\|F\|_{L^{\infty}}l\big(c_{\ep}\big).
\end{eqnarray*}
\end{remark}
\begin{proof}
For $c\in D_0\setminus \{0\}$, we have $|\mathcal{H}(y,c)|\geq C^{-1}\big|y^2-y_{c_{\pm}}^2\big|$, then
\begin{eqnarray*}
\big|T_{\pm}(F)(c)\big|\leq C\|F\|_{L^{\infty}}\left|\int_{0}^{\pm 1}\f{1}{|y|+|c|}dy\right|\leq C\|F\|_{L^{\infty}}l\big(c_{\ep}\big).
\end{eqnarray*}
Thus we complete the proof of the remark.
\end{proof}
In the following paper, for $c\in D_0\setminus\{0, W_+(1), W_+(-1), W_-(1),W_-(-1)\}$, let
\begin{align*}
\mathcal{U}_+^{re}(F)(c)&=-c^2P(c)T_+(F)(c)I^{re}_-(c)-\va_-(0,c)L(F)(c)I^{re}_-(c)\\
&\quad
+\va_+(0,c)^2T_+(F)(c)-(\va_+\va_-)(0,c)T_-(F)(c),
\end{align*}
\begin{equation*}
\mathcal{U}_+^{im}(F)(c)=-\frac{\pi c^2P(c)T_+(F)(c)\chi_-(c)}{\sigma_-(c)}
-\frac{\pi\va_-(0,c)L(F)(c)\chi_-(c)}{\sigma_-(c)},
\end{equation*}
\begin{align*}
\mathcal{U}^{re}_-(F)(c)&=c^2P(c)T_-(F)(c)I^{re}_+(c)+\va_+(0,c)L(F)(c)I^{re}_+(c)\\
&\ \
+(\va_+\va_-)(0,c)T_+(F)(c)-\va_-(0,c)^2T_-(F)(c),
\end{align*}
\begin{align*}
\mathcal{U}^{im}_-(F)(c)=\frac{\pi c^2P(c)T_-(F)(c)\chi_+(c)}{\sigma_+(c)}
+\frac{\pi\va_+(0,c)L(F)(c)\chi_+(c)}{\sigma_+(c)},
\end{align*}
\begin{align*}
\mathcal{V}_+^{re}(F)(c)&=\va_-(0,c)L(F)(c)I^{re}_+(c)I^{re}_-(c)
+(\va_+\va_-)(0,c)T_-(F)(c)I^{re}_+(c)\\
&\quad
+\va_-(0,c)^2T_+(F)(c)I^{re}_-(c)
-\frac{\pi^2\va_-(0,c)L(F)(c)\chi_+(c)\chi_-(c)}{\sigma_+(c)\sigma_-(c)},
\end{align*}
\begin{align*}
\mathcal{V}_+^{im}(F)(c)&=\pi\va_-(0,c)L(F)(c)
\Big(\frac{I^{re}_+(c)\chi_-(c)}{\sigma_-(c)}+\frac{I^{re}_-(c)\chi_+(c)}{\sigma_+(c)}\Big)\\
&\quad
+\frac{\pi(\va_+\va_-)(0,c)T_-(F)(c)\chi_+(c)}{\sigma_+(c)}
+\frac{\pi\va_-(0,c)^2T_+(F)(c)\chi_-(c)}{\sigma_-(c)},
\end{align*}
\begin{align*}
\mathcal{V}_-^{re}(F)(c)&=
\va_+(0,c)L(F)(c)I^{re}_+(c)I^{re}_-(c)
+\va_+(0,c)^2T_-(F)(c)I^{re}_+(c)\\
&\quad
+(\va_+\va_-)(0,c)T_+(F)(c)I^{re}_-(c)-\frac{\pi^2\va_+(0,c)L(F)(c)\chi_+(c)\chi_-(c)}{\sigma_+(c)\sigma_-(c)},
\end{align*}
\begin{align*}
\mathcal{V}_-^{im}(F)(c)&=\pi\va_+(0,c)L(F)(c)
\Big(\frac{I^{re}_+(c)\chi_-(c)}{\sigma_-(c)}+\frac{I^{re}_-(c)\chi_+(c)}{\sigma_+(c)}\Big)\\
&\quad
+\frac{\pi\va_+(0,c)^2T_-(F)(c)\chi_+(c)}{\sigma_+(c)}
+\frac{\pi(\va_+\va_-)(0,c)T_+(F)(c)\chi_-(c)}{\sigma_-(c)}.
\end{align*}
And we also introduce for $c\in D_0\setminus\{0, W_+(1), W_+(-1), W_-(1),W_-(-1)\}$,
\begin{align*}
\mu_+^+(F)(c)&=\frac{\mathcal{U}^{re}_+(F)(c)
+i\mathcal{U}^{im}_+(F)(c)}{\mathcal{D}^{re}(c)+ i\mathcal{D}^{im}(c)},
\ \ \mu_+^-(F)(c)=\frac{\mathcal{U}^{re}_+(F)(c)- i\mathcal{U}^{im}_+(F)(c)}{\mathcal{D}^{re}(c)- i\mathcal{D}^{im}(c)},\\
\mu_-^+(F)(c)&=\frac{\mathcal{U}^{re}_-(F)(c)+i\mathcal{U}^{im}_-(F)(c)}{\mathcal{D}^{re}(c)+ i\mathcal{D}^{im}(c)},\ \
\mu_-^-(F)(c)=\frac{\mathcal{U}^{re}_-(F)(c)-i\mathcal{U}^{im}_-(F)(c)}{\mathcal{D}^{re}(c)- i\mathcal{D}^{im}(c)},\\
\nu_+^+(F)(c)&=\frac{\mathcal{V}^{re}_+(F)(c)+i\mathcal{V}^{im}_+(F)(c)}{\mathcal{D}^{re}(c)+ i\mathcal{D}^{im}(c)}, \ \
\nu_+^-(F)(c)=\frac{\mathcal{V}^{re}_+(F)(c)-i\mathcal{V}^{im}_+(F)(c)}{\mathcal{D}^{re}(c)- i\mathcal{D}^{im}(c)},\\
\nu_-^+(F)(c)&=\frac{\mathcal{V}^{re}_-(F)(c)+i\mathcal{V}^{im}_-(F)(c)}
{\mathcal{D}^{re}(c)+ i\mathcal{D}^{im}(c)},\ \
\nu_-^-(F)(c)=\frac{\mathcal{V}^{re}_-(F)(c)-i\mathcal{V}^{im}_-(F)(c)}
{\mathcal{D}^{re}(c)- i\mathcal{D}^{im}(c)}.
\end{align*}
\begin{proposition}\label{prop: mu nu lim}
(1) Let $c_{\ep}=c\pm i\ep\in \Omega_{\ep_0}\setminus D_0$, $0<\ep<\ep_0$. There holds that
\begin{align*}
&|\mu_+(F)(c_{\ep})|\leq \frac{C\|F\|_{L^{\infty}}|c_{\ep}|l\big(c_{\ep}\big)
\Big(l\big(W_+(-1)-c_{\ep}\big)+l\big(W_-(-1)-c_{\ep}\big)\Big)}
{l\big(W_+(1)-c_{\ep}\big)l\big(W_-(1)-c_{\ep}\big)
l\big(W_+(-1)-c_{\ep}\big)l\big(W_-(-1)-c_{\ep}\big)},\\
&|\mu_-(F)(c_{\ep})|\leq \frac{C\|F\|_{L^{\infty}}|c_{\ep}|l\big(c_{\ep}\big)
\Big(l\big(W_+(1)-c_{\ep}\big)+l\big(W_-(1)-c_{\ep}\big)\Big)}
{l\big(W_+(1)-c_{\ep}\big)l\big(W_-(1)-c_{\ep}\big)
l\big(W_+(-1)-c_{\ep}\big)l\big(W_-(-1)-c_{\ep}\big)},\\
&|\nu_{\pm}(F)(c_{\ep})|\leq C\|F\|_{L^{\infty}}l\big(c_{\ep}\big).
\end{align*}
(2) For $c_{\ep}=c+i\ep$, $0<\ep<1$, $c\in D_0\setminus\{0, W_+(1), W_+(-1), W_-(1),W_-(-1)\}$, there holds that
\begin{eqnarray*}
\lim_{\ep\rightarrow0^{\pm}}\mu_+(F)(c_{\ep})=\mu_+^{\pm}(F)(c),\ \
\lim_{\ep\rightarrow0^{\pm}}\nu_+(F)(c_{\ep})=\nu_+^{\pm}(F)(c),
\end{eqnarray*}
\begin{eqnarray*}
\lim_{\ep\rightarrow0^{\pm}}\mu_-(F)(c_{\ep})=\mu_-^{\pm}(F)(c),\ \
\lim_{\ep\rightarrow0^{\pm}}\nu_-(F)(c_{\ep})=\nu_-^{\pm}(F)(c).
\end{eqnarray*}
\end{proposition}
\begin{proof}
By Proposition \ref{prop: sol. hom. [0,1]} and Proposition \ref{prop: sol. hom. [-1,0]}, we have
$
|\pa_y\va_{\pm}(0,c_{\ep})|\leq C|c_{\ep}|,
$
which gives for $c_{\ep}\in\Omega_{\ep_0}$,
\begin{equation*}
\big|P(c_{\ep})\big|=\big|(\va_-^2\va_+\pa_y\va_+)(0,c_{\ep})
-(\va_+^2\va_-\pa_y\va_-)(0,c_{\ep})\big|\leq C|c_{\ep}\big|,
\end{equation*}
and we also have
\begin{equation}\label{eq: L(F) bdd}
\begin{split}
\big|L(F)(c_{\ep})\big|&=\Big|
\va_-(0,c_{\ep})\int_0^{y_{c_+}}(F\va_+)(y,c_{\ep})dy
-\va_+(0,c_{\ep})\int^{y_{c_-}}_0(F\va_-)(y,c_{\ep})dy\Big|\\
&\leq C|c_{\ep}|\|F\|_{L^{\infty}}.
\end{split}
\end{equation}
From which and by using Lemma \ref{prop: D est}, Lemma \ref{lem: I bdd} and Lemma \ref{lem: T(F)} gives rise to
\begin{align*}
&|\mu_+(F)(c_{\ep})|\leq
\frac{C\|F\|_{L^{\infty}}|c_{\ep}|\big(\big|\ln|c_{\ep}|\big|+1\big)
\Big(l\big(W_+(-1)-c_{\ep}\big)+l\big(W_-(-1)-c_{\ep}\big)\Big)}
{l\big(W_+(1)-c_{\ep}\big)l\big(W_-(1)-c_{\ep}\big)
l\big(W_+(-1)-c_{\ep}\big)l\big(W_-(-1)-c_{\ep}\big)},\\
&|\mu_-(F)(c_{\ep})|\leq \frac{C\|F\|_{L^{\infty}}|c_{\ep}|\big(\big|\ln|c_{\ep}|\big|+1\big)
\Big(l\big(W_+(1)-c_{\ep}\big)+l\big(W_-(1)-c_{\ep}\big)\Big)}
{l\big(W_+(1)-c_{\ep}\big)l\big(W_-(1)-c_{\ep}\big)
l\big(W_+(-1)-c_{\ep}\big)l\big(W_-(-1)-c_{\ep}\big)}.
\end{align*}
Similarly, we get
\begin{eqnarray*}
|\nu_{\pm}(F)(c_{\ep})|\leq C\|F\|_{L^{\infty}}\big(\big|\ln|c_{\ep}|\big|+1\big).
\end{eqnarray*}
On the other hand, for $c\in D_0\setminus\{0, W_+(1), W_+(-1), W_-(1),W_-(-1)\}$,
by Lemma \ref{lem: D>0}, Lemma \ref{lem: T(F)}, \eqref{I+ lim} and \eqref{I- lim},
we can easily get that
\begin{eqnarray*}
&\ & \lim_{\ep\rightarrow0^{\pm}}\mu_+(F)(c_{\ep})
=\mu_+^{\pm}(F)(c),\ \ \lim_{\ep\rightarrow0^{\pm}}\nu_+(F)(c_{\ep})=\nu_+^{\pm}(F)(c),\\
&\ & \lim_{\ep\rightarrow0^{\pm}}\mu_-(F)(c_{\ep})
=\mu_-^{\pm}(F)(c),\ \ \lim_{\ep\rightarrow0^{\pm}}\nu_-(F)(c_{\ep})=\nu_-^{\pm}(F)(c).
\end{eqnarray*}
Thus we complete the proof of Proposition \ref{prop: mu nu lim}.
\end{proof}
\begin{remark}\label{rmk: munu lim up rem}
From Proposition \ref{prop: mu nu lim}, we have for $c\in D_0\setminus\{0, W_+(1), W_+(-1), W_-(1),W_-(-1)\}$,
\begin{align*}
|\mu^{\pm}_+(F)(c)|&\leq\frac{C\|F\|_{L^{\infty}}|c|l(c)
\Big(l\big(W_+(-1)-c\big)+l\big(W_-(-1)-c\big)\Big)}
{l\big(W_+(1)-c\big)l\big(W_-(1)-c\big)
l\big(W_+(-1)-c\big)l\big(W_-(-1)-c\big)},\\
|\mu^{\pm}_-(F)(c)|&\leq\frac{C\|F\|_{L^{\infty}}|c|l(c)
\Big(l\big(W_+(1)-c\big)+l\big(W_-(1)-c\big)\Big)}
{l\big(W_+(1)-c\big)l\big(W_-(1)-c\big)
l\big(W_+(-1)-c\big)l\big(W_-(-1)-c\big)},\\
|\nu^{\pm}_{\pm}(F)(c)|&\leq C\|F\|_{L^{\infty}}l\big(c\big).
\end{align*}
\end{remark}
\begin{lemma}\label{lem: XYST bdd}
For $c\in D_0\setminus\{0, W_+(1), W_+(-1), W_-(1),W_-(-1)\}$, we have the following estimates:
\begin{align*}
|\mathcal{U}^{re}_+(F)(c)|&\leq C\|F\|_{L^{\infty}}l\big(W_+(-1)-c\big)l\big(W_-(-1)-c\big)l(c),\\
|\mathcal{U}^{re}_-(F)(c)|&\leq C\|F\|_{L^{\infty}}l\big(W_+(1)-c\big)l\big(W_-(1)-c\big)l(c),\\
|\mathcal{U}^{im}_{\pm}(F)(c)|&\leq C\|F\|_{L^{\infty}},\\
|\mathcal{V}^{re}_{\pm}(F)(c)|&\leq \frac{C\|F\|_{L^{\infty}}}{|c|}l\big(W_+(1)-c\big)l\big(W_-(1)-c\big)
l\big(W_+(-1)-c\big)l\big(W_-(-1)-c\big)l(c),\\
|\mathcal{V}^{im}_{\pm}(F)(c)|&\leq \frac{C\|F\|_{L^{\infty}}}{|c|}\Big(l\big(W_+(1)-c\big)l\big(W_-(1)-c\big)
+l\big(W_+(-1)-c\big)l\big(W_-(-1)-c\big)+l(c)\Big).
\end{align*}
\end{lemma}
\begin{proof}
By Lemma \ref{lem: limit up bdd}, Remark \ref{rmk: T(F)}, \eqref{eq: P est} and \eqref{eq: L(F) bdd}, we can get
\begin{align*}
|\mathcal{U}^{re}_+(F)(c)|&\leq C\|F\|_{L^{\infty}}|c|^2\big(\big|\ln|c|\big|+1\big)\Big(1+\big|\ln|W_+(-1)-c|\big|
+\big|\ln|c-W_-(-1)|\big|\Big)\nonumber\\
&\quad +C\|F\|_{L^{\infty}}\Big(1+\big|\ln|W_+(-1)-c|\big|+
\big|\ln|c-W_-(-1)|\big|\Big)+C\big(\big|\ln|c|\big|+1\big)\nonumber\\
&\leq C\|F\|_{L^{\infty}}\Big(1+\big|\ln|W_+(-1)-c|\big|+\big|\ln|c-W_-(-1)|\big|
+\big|\ln|c|\big|\Big)
\end{align*}
and
\begin{align*}
|\mathcal{U}^{re}_-(F)(c)|&\leq C\|F\|_{L^{\infty}}|c|^2\big(\big|\ln|c|\big|+1\big)\Big(1+\big|\ln|W_+(1)-c|\big|
+\big|\ln|c-W_-(1)|\big|\Big)\nonumber\\
&\quad +C\|F\|_{L^{\infty}}\Big(1+\big|\ln|W_+(1)-c|\big|+
\big|\ln|c-W_-(1)|\big|\Big)+C\big(\big|\ln|c|\big|+1\big)\nonumber\\
&\leq C\|F\|_{L^{\infty}}\Big(1+\big|\ln|W_+(1)-c|\big|+\big|\ln|c-W_-(1)|\big|
+\big|\ln|c|\big|\Big)
\end{align*}
From Remark \ref{rmk: sigma bdd}, Remark \ref{rmk: T(F)}, \eqref{eq: P est} and \eqref{eq: L(F) bdd}, we have
\begin{align*}
|\mathcal{U}^{im}_{\pm}(F)(c)|\leq C\|F\|_{L^{\infty}}+C\|F\|_{L^{\infty}}|c|^2\big(1+\big|\ln|c|\big|\big)\leq C\|F\|_{L^{\infty}}.
\end{align*}
Similarly, combining Lemma \ref{lem: limit up bdd}, Remark \ref{rmk: sigma bdd}, Remark \ref{rmk: T(F)} and \eqref{eq: L(F) bdd}, we deduce
\begin{align*}
|\mathcal{V}^{re}_{\pm}(F)(c)|
&\leq \frac{C\|F\|_{L^{\infty}}}{|c|}\Big\{\Big(1+\big|\ln|W_+(1)-c|\big|+\big|
\ln|c-W_-(1)|\big|\Big)\\
&\ \ \times\Big(1+\big|\ln|W_+(-1)-c|\big|+\big|
\ln|c-W_-(-1)|\big|\Big)\Big\}\\
&\quad +\frac{C\|F\|_{L^{\infty}}}{|c|}\big(1+\big|\ln|c|\big|\big)
\Big\{1+\big|\ln|W_+(1)-c|\big|+\big|
\ln|c-W_-(1)|\big|\\
&\ \ +\big|\ln|W_+(-1)-c|\big|+\big|
\ln|c-W_-(-1)|\big|\Big\}+\frac{C\|F\|_{L^{\infty}}}{|c|}\\
&\leq \frac{C\|F\|_{L^{\infty}}}{|c|}l\big(W_+(1)-c\big)l\big(W_-(1)-c\big)
l\big(W_+(-1)-c\big)l\big(W_-(-1)-c\big)l(c),
\end{align*}
and
\begin{align*}
|\mathcal{V}^{im}_{\pm}(F)(c)|
&\leq C\|F\|_{L^{\infty}}\Big(1+\big|\ln|W_+(1)-c|\big|+\big|
\ln|c-W_-(1)|\big|+\big|\ln|W_+(-1)-c|\big|\\
&\ \ +\big|
\ln|c-W_-(-1)|\big|\Big)+\frac{C\|F\|_{L^{\infty}}\big(1+\big|\ln|c|\big|\big)}{|c|}\\
&\leq \frac{C\|F\|_{L^{\infty}}}{|c|}\Big(1+\big|\ln|W_+(1)-c|\big|
+\big|\ln|c-W_-(1)|\big|\\
&\ \ +\big|\ln|W_+(-1)-c|\big|+\big|
\ln|c-W_-(-1)|\big|+\big|\ln|c|\big|\Big)\\
&\leq \frac{C\|F\|_{L^{\infty}}}{|c|}\Big(l\big(W_+(1)-c\big)+l\big(W_-(1)-c\big)
+l\big(W_+(-1)-c\big)\\
&\ \ +l\big(W_-(-1)-c\big)+l(c)\Big).
\end{align*}
Thus we prove the lemma.
\end{proof}
\begin{proposition}\label{prop: F y bdd lim}
1. Let $c_{\ep}=c+i\ep\in D_{\ep_0}\cup D_0$. Then it holds that,\\
for $0\leq y<y_{c_+}\leq 1$ or $0\leq y\leq 1<y_{c_+}\leq a_+$
\begin{align*}
&\left|\va_+(y,c_{\ep})
\int_0^y\frac{1}{\mathcal{H}(y',c_{\ep})\va_+(y',c_{\ep})^2}dy'\right|
\leq \frac{C\left(\big|\ln|y-y_{c_+}|\big|+1\right)}{|c_{\ep}|},\\
&\left|\va_+(y,c_{\ep})
\int_0^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c_{\ep})dz}
{\mathcal{H}(y',c_{\ep})\va_+(y',c_{\ep})^2}dy'\right|
\leq C\|F\|_{L^{\infty}}\big(\big|\ln(|y|+|c_{\ep}|)\big|+1\big),
\end{align*}
and for $0\leq y_{c_+}<y\leq 1$,
\begin{align*}
&\left|\va_+(y,c_{\ep})
\int_1^y\frac{1}{\mathcal{H}(y',c_{\ep})\va_+(y',c_{\ep})^2}dy'\right|
\leq \frac{C\left(\big|\ln|y-y_{c_+}|\big|+1\right)}{|y|},\\
&\left|\va_+(y,c_{\ep})
\int_1^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c_{\ep})dz}
{\mathcal{H}(y',c_{\ep})\va_+(y',c_{\ep})^2}dy'\right|
\leq C\|F\|_{L^{\infty}}\big(\big|\ln(|y|+|c_{\ep}|)\big|+1\big);
\end{align*}
and for $-1\leq y<y_{c_-}\leq 0$,
\begin{align*}
&\left|\va_-(y,c_{\ep})
\int_{-1}^y\frac{1}{\mathcal{H}(y',c_{\ep})\va_-(y',c_{\ep})^2}dy'\right|\leq \frac{C\left(\big|\ln|y-y_{c_-}|\big|+1\right)}{|y|},\\
&\left|\va_-(y,c_{\ep})
\int_{-1}^y\frac{\int_{y_{c_-}}^{y'}(F\va_-)(z,c_{\ep})dz}
{\mathcal{H}(y',c_{\ep})\va_-(y',c_{\ep})^2}dy'\right|
\leq C\|F\|_{L^{\infty}}\big(\big|\ln(|y|+|c_{\ep}|)\big|+1\big),
\end{align*}
and for $-1\leq y_{c_-}<y \leq0$ or $a_-\leq y_{c_-}<-1 \leq y\leq 0$,
\begin{align*}
&\left|\va_-(y,c_{\ep})
\int_0^y\frac{1}{\mathcal{H}(y',c_{\ep})\va_-(y',c_{\ep})^2}dy'\right|
\leq \frac{C\left(\big|\ln|y-y_{c_-}|\big|+1\right)}{|c_{\ep}|},\\
&\left|\va_-(y,c_{\ep})
\int_0^y\frac{\int_{y_{c_-}}^{y'}(F\va_-)(z,c_{\ep})dz}
{\mathcal{H}(y',c_{\ep})\va_-(y',c_{\ep})^2}dy'\right|
\leq C\|F\|_{L^{\infty}}\big(\big|\ln(|y|+|c_{\ep}|)\big|+1\big).
\end{align*}
2. Let $c_{\ep}=c+i\ep\in D_{\ep_0}$. Then it holds that for $0\leq y<y_{c_+}\leq 1$ or $0\leq y\leq1<y_{c_+}\leq a_+$,
\begin{eqnarray*}
\lim_{\ep\rightarrow0}\va_+(y,c_{\ep})
\int_0^y\frac{1}{\mathcal{H}(y',c_{\ep})\va_+(y',c_{\ep})^2}dy'=\va_+(y,c)
\int_0^y\frac{1}{\mathcal{H}(y',c)\va_+(y',c)^2}dy',
\end{eqnarray*}
\begin{eqnarray*}
\lim_{\ep\rightarrow0}\va_+(y,c_{\ep})
\int_0^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c_{\ep})dz}
{\mathcal{H}(y',c_{\ep})\va_+(y',c_{\ep})^2}dy'=\va_+(y,c)
\int_0^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c)dz}
{\mathcal{H}(y',c)\va_+(y',c)^2}dy',
\end{eqnarray*}
and for $0\leq y_{c_+}<y\leq 1$,
\begin{eqnarray*}
\lim_{\ep\rightarrow0}\va_+(y,c_{\ep})
\int_1^y\frac{1}{\mathcal{H}(y',c_{\ep})\va_+(y',c_{\ep})^2}dy'=\va_+(y,c)
\int_1^y\frac{1}{\mathcal{H}(y',c)\va_+(y',c)^2}dy',
\end{eqnarray*}
\begin{eqnarray*}
\lim_{\ep\rightarrow0}\va_+(y,c_{\ep})
\int_1^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c_{\ep})dz}
{\mathcal{H}(y',c_{\ep})\va_+(y',c_{\ep})^2}dy'=\va_+(y,c)
\int_1^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c)dz}
{\mathcal{H}(y',c)\va_+(y',c)^2}dy'.
\end{eqnarray*}
For $-1\leq y<y_{c_-}\leq 0$,
\begin{eqnarray*}
\lim_{\ep\rightarrow0}\va_-(y,c_{\ep})
\int_{-1}^y\frac{1}{\mathcal{H}(y',c_{\ep})\va_-(y',c_{\ep})^2}dy'=\va_-(y,c)
\int_{-1}^y\frac{1}{\mathcal{H}(y',c)\va_-(y',c)^2}dy',
\end{eqnarray*}
\begin{eqnarray*}
\lim_{\ep\rightarrow0}\va_-(y,c_{\ep})
\int_{-1}^y\frac{1}{\mathcal{H}(y',c_{\ep})\va_-(y',c_{\ep})^2}dy'=\va_-(y,c)
\int_{-1}^y\frac{1}{\mathcal{H}(y',c)\va_-(y',c)^2}dy',
\end{eqnarray*}
and for $-1\leq y_{c_-}<y \leq0$ or $a_-\leq y_{c_-}<-1 \leq y\leq 0$,
\begin{eqnarray*}
\lim_{\ep\rightarrow0}\va_-(y,c_{\ep})
\int_0^y\frac{1}{\mathcal{H}(y',c_{\ep})\va_-(y',c_{\ep})^2}dy'=\va_-(y,c)
\int_0^y\frac{1}{\mathcal{H}(y',c)\va_-(y',c)^2}dy',
\end{eqnarray*}
\begin{eqnarray*}
\lim_{\ep\rightarrow0}\va_-(y,c_{\ep})
\int_0^y\frac{1}{\mathcal{H}(y',c_{\ep})\va_-(y',c_{\ep})^2}dy'=\va_-(y,c)
\int_0^y\frac{1}{\mathcal{H}(y',c)\va_-(y',c)^2}dy'.
\end{eqnarray*}
\end{proposition}
\begin{proof}
We consider the case of $y\in[0,1]$ and the case of $y\in[-1,0]$ can be proved by the same argument.
By Proposition \ref{prop: sol. hom. [0,1]}, we get for $c\in D_{\ep_0}\cup D_0$
\begin{eqnarray*}
\left|\frac{1}{\mathcal{H}(y',c)\va_+(y',c)^2}\right|\leq \f{C}{|y'-y_{c_+}|(|y'|+|c|)},
\end{eqnarray*}
and
\begin{eqnarray*}
\left|\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c)dz}
{\mathcal{H}(y',c)\va_+(y',c)^2}\right|\leq \f{C\|F\|_{L^{\infty}}}{|y'|+|c|},
\end{eqnarray*}
which implies for $0\leq y<y_{c_+}\leq 1$ or $0\leq y\leq1<y_{c_+}\leq a_+$,
\begin{align*}
&\Big|\va_+(y,c)
\int_0^y\frac{1}{\mathcal{H}(y',c)\va_+(y',c)^2}dy'\Big|\leq
\frac{C\left(\big|\ln|y_{c_+}-y|\big|+1\right)}{|c|}\\
&\Big|\va_+(y,c)\int_0^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c)dz}
{\mathcal{H}(y',c)\va_+(y',c)^2}dy'\Big|\leq C\|F\|_{L^{\infty}}\big(\big|\ln(|y|+|c|)\big|+1\big).
\end{align*}
and for $0\leq y_{c_+}<y\leq 1$,
\begin{align*}
&\Big|\va_+(y,c)
\int_1^y\frac{1}{\mathcal{H}(y',c)\va_+(y',c)^2}dy'\Big|\leq
\frac{C\left(\big|\ln|y_{c_+}-y|\big|+1\right)}{|y|}\\
&\Big|\int_1^y\va_+(y,c)\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c)dz}
{\mathcal{H}(y',c)\va_+(y',c)^2}dy'\Big|\leq C\|F\|_{L^{\infty}}\big(\big|\ln(|y|+|c|)\big|+1\big)
\end{align*}
Since $F(y,c_{\ep})$, $\mathcal{H}(y,c_{\ep})$ and $\va_{\pm}(y,c_{\ep})$ are continuous functions, then by Lebesgue's dominated convergence theorem, as $\ep\rightarrow0$, we can obtain the second part.
\end{proof}
\begin{proposition}\label{prop: Br Bl Theta bdd}
Let $c_{\ep}\in B_{\ep_0}^l$ or $c_{\ep}\in B_{\ep_0}^r$. Then it holds that
\begin{eqnarray*}
|\Th(y,c_{\ep})|\leq C\|F\|_{L^{\infty}}.
\end{eqnarray*}
\end{proposition}
\begin{proof}
We only show the proof of the case $0\leq y\leq 1$ and $c_{\ep}\in B_{\ep_0}^r$, the proofs of the other three cases are similar.
In this case, $c_{\ep}=\max\{W_+(1),W_-(-1)\}+\ep e^{i\th}$ and $c_r=\max\{W_+(1),W_-(-1)\}$ with $y_{c_+}=W_+^{-1}\big(\max\{W_+(1),W_-(-1)\}\big)=a_+$ and by Proposition \ref{prop: inhom solu}, we can write $\Th(y,c_{\ep})$ in the following way,
\begin{align*}
\Th(y,c_{\ep})&=\va_+(y,c_{\ep})\int_0^y\frac{\int_{a_+}^{y'}(F\va_+)(z,c_{\ep})dz}
{\mathcal{H}(y',c_{\ep})\va_+(y',c_{\ep})^2}dy'\\
&\quad+\mu_+(F)(c_{\ep}) \va_+(y,c_{\ep})\int_0^y\frac{y_{c_+}}{\mathcal{H}(y',c_{\ep})\va_+(y',c_{\ep})^2}dy'
+\nu_+(F)(c_{\ep})\va_+(y,c_{\ep}).
\end{align*}
Then we have
\begin{align*}
&\left|\va_+(y,c_{\ep})\int_0^y\frac{\int_{a_+}^{y'}(F\va_+)(z,c_{\ep})dz}
{\mathcal{H}(y',c_{\ep})\va_+(y',c_{\ep})^2}dy'\right|\\
&\leq C\|F\|_{L^{\infty}}\left|\int_0^y\f{|y'-a_+|}{|W_+(y')-W_+(1)-\ep e^{i\th}||W_-(y')-W_+(1)-\ep e^{i\th}|}dy'\right|\leq C\|F\|_{L^{\infty}}.
\end{align*}
By Proposition \ref{prop: mu nu lim}, we have
\begin{align*}
&\left|\mu_+(F)(c_{\ep}) \va_+(y,c_{\ep})\int_0^y\frac{1}{\mathcal{H}(y',c_{\ep})\va_+(y',c_{\ep})^2}dy'\right|\\
&\leq \f{C\|F\|_{L^{\infty}}}{1+|\ln |\ep||}\left|\int_0^y\f{dy'}{|W_+(y')-c_r-\ep e^{i\th}||W_-(y')-c_r-\ep e^{i\th}|}\right|
\leq C\|F\|_{L^{\infty}},
\end{align*}
and
\begin{eqnarray*}
\left|\nu_+(F)(c_{\ep})\va_+(y,c_{\ep})\right|\leq C\|F\|_{L^{\infty}}.
\end{eqnarray*}
Thus we prove the proposition.
\end{proof}
\section{The proof of main theorem}
Now we present the proof of Theorem \ref{main thm}.
\begin{proof}
We recall that $\big(cI-M_{\al}\big)^{-1}\Big(\begin{array}{l} \widehat{\psi}_0\\ \widehat{\phi}_0\end{array}\Big)(\al,y)=\Big(\begin{array}{l} \Psi_1\\ \Phi_1\end{array}\Big)(\al,y,c)$ and let $\Phi_1(\al,y,c)=b(y)\Phi(\al,y,c)+\widehat{\phi}_0(\al,0)\chi(y)/c$, then $\Psi_1(\al,y,c)=(u(y)-c)\Phi(\al,y,c)
+(u(y)-c)\widehat{\phi}_0(\al,0)\f{\chi(y)}{cb(y)}+\f{\widehat{\phi}_0(\al,y)}{b(y)}$ and $\Phi(\al, y,c)$ satisfies
\begin{eqnarray*}
\pa_y\Big[\Big(\big(u(y)-c\big)^2-b(y)^2\Big)\pa_y{\Phi(\al, y,c)}\Big]
-\al^2\Big(\big(u(y)-c\big)^2-b(y)^2\Big){\Phi(\al,y,c)}=G(\al,y,c).
\end{eqnarray*}
In Proposition \ref{prop: inhom solu} we proved that $\Th(y,c)$ satisfies
\begin{eqnarray*}
\pa_y\Big[\Big(\big(u(y)-c\big)^2-b(y)^2\Big)\pa_y\Th(y,c)\Big]
-\al^2\Big(\big(u(y)-c\big)^2-b(y)^2\Big)\Th(y,c)=F(y,c)=cG(\al,y,c).
\end{eqnarray*}
Thus $\Th(y,c)=c{\Phi(\al,y,c)}$ and by \eqref{eq: matrix-form}, we obtain that for $y\in[-1,1]$,
\begin{align*}
&\widehat{\psi}(t,\al,y)\\
&=\frac{1}{2\pi i}
\int_{\pa\Omega_{\ep_0}}e^{-i\al tc_{\ep}}\Big(\frac{u(y)-c_{\ep}}{c_{\ep}}\Theta(y,c_{\ep})
+\frac{u(y)-c_{\ep}}{c_{\ep}b(y)}\widehat{\phi}_0(\al,0)\chi(y)
+\frac{\widehat{\phi}_0(\al,y)}{b(y)}\Big)dc_{\ep}\\
&=\frac{1}{2\pi i}
\int_{\pa\Omega_{\ep_0}}e^{-i\al tc_{\ep}}\Big(\frac{u(y)-c_{\ep}}{c_{\ep}}\Theta(y,c_{\ep})
+\frac{u(y)\widehat{\phi}_0(\al,0)\chi(y)}{c_{\ep}b(y)}
+\frac{\widehat{\phi}_0(\al,y)-\chi(y)\widehat{\phi}(\al,0)}{b(y)}\Big)dc_{\ep},
\end{align*}
and
\begin{align*}
\widehat{\phi}(t,\al,y)&=\frac{1}{2\pi i}
\int_{\pa\Omega_{\ep_0}}e^{-i\al tc_{\ep}}\Big(\frac{b(y)}{c_{\ep}}\Theta(y,c_{\ep})
+\frac{\widehat{\phi}_0(\al,0)\chi(y)}{c_{\ep}}\Big)dc_{\ep}.
\end{align*}
By the fact that $\Theta(y,c_{\ep})$ is an analytic function in $\Om_{\ep_0}\setminus D_0$, we obtain
\begin{equation}\label{eq: psi and phi}
\begin{split}
\widehat{\psi}(t,\al,y)&=\frac{u(y)}{b(y)}\widehat{\phi}_0(\al,0)\chi(y)+\lim_{\ep_0\rightarrow0+}\frac{1}{2\pi i}
\int_{\pa\Omega_{\ep_0}}e^{-i\al tc_{\ep}}\Big(\frac{u(y)-c_{\ep}}{c_{\ep}}\Theta(y,c_{\ep})\Big)dc_{\ep},\\
\widehat{\phi}(t,\al,y)&=\widehat{\phi}_0(\al,0)\chi(y)+\lim_{\ep_0\rightarrow0+}\frac{1}{2\pi i}
\int_{\pa\Omega_{\ep_0}}e^{-i\al tc_{\ep}}\Big(\frac{b(y)}{c_{\ep}}\Theta(y,c_{\ep})\Big)dc_{\ep}.
\end{split}
\end{equation}
In the following, we denote
$M_+=\max\{W_+(1),W_-(-1)\}$ and $m_-=\min\{W_+(-1),W_-(1)\}$
for brevity.
\noindent{\bf{Proof \ of \ 1.}}
For $y=0$ and $\chi(0)=1$, we get that,
\begin{eqnarray*}
\widehat{\phi}(t,\al,0)=\lim_{\ep_0\rightarrow0+}\frac{1}{2\pi i}
\int_{\pa\Omega_{\ep_0}}e^{-i\al tc_{\ep}}\frac{\widehat{\phi}_0(\al,0)}{c_{\ep}}dc_{\ep}
=\widehat{\phi}_0(\al,0),
\end{eqnarray*}
and
\begin{align*}
\widehat{\psi}(t,\al,0)&=
\frac{u'(0)}{b'(0)}\widehat{\phi}_0(\al,0)
-\lim_{\ep_0\rightarrow0+}\frac{1}{2\pi i}
\int_{\pa\Omega_{\ep_0}}e^{-i\al tc_{\ep}}\Theta(0,c_{\ep})dc_{\ep}\\
&=\frac{u'(0)}{b'(0)}\widehat{\phi}_0(\al,0)
-\lim_{\ep_0\rightarrow0+}\frac{1}{2\pi i}
\int_{\pa\Omega_{\ep_0}}e^{-i\al tc_{\ep}}\nu_+(F)(c_{\ep})\va_+(0,c_{\ep})dc_{\ep}.
\end{align*}
We have for the second term
\begin{align*}
&\frac{1}{2\pi i}\int_{\pa\Omega_{\ep_0}}e^{-i\al tc_{\ep}}
\nu_+(F)(c_{\ep})\va_+(0,c_{\ep})dc_{\ep}\\
&=
\frac{1}{2\pi i}\int_{\{|c_{\ep}|\leq \sqrt{2}\ep_0\}
\cap\pa\Omega_{\ep_0}}e^{-i\al tc_{\ep}}\nu_+(F)(c_{\ep})\va_+(0,c_{\ep})dc_{\ep}\\
&\quad+\frac{1}{2\pi i}\int_{\{|c_{\ep}|>\sqrt{2}\ep_0\}
\cap\pa\Omega_{\ep_0}}e^{-i\al tc_{\ep}}\nu_+(F)(c_{\ep})\va_+(0,c_{\ep})dc_{\ep}\\
&=I(c_{\ep_0})+J(c_{\ep_0}).
\end{align*}
By Proposition \ref{prop: mu nu lim}, for $c_{\ep}\in \pa\Omega_{\ep_0}$ and $|c_{\ep}|\leq \sqrt{2}\ep_0$, we have
\begin{eqnarray*}
|\nu_+(F)(c_{\ep})|\leq C\|F\|_{L^{\infty}}\big(1+\big|\ln|c_{\ep}|\big|\big)\leq C\|F\|_{L^{\infty}}\left(\big|\ln|\ep_0|\big|+1\right).
\end{eqnarray*}
Thus, we deduce that
$I(c_{\ep})\leq C\|F\|_{L^{\infty}}\ep_0\left(\big|\ln|\ep_0|\big|+1\right),$
and then $\lim\limits_{\ep_0\rightarrow0+}I(c_{\ep_0})=0.$
As for $J(c_{\ep_0})$, we have
\begin{align*}
J(c_{\ep_0})
&=-\frac{1}{2\pi i}\int_{\ep_0}^{M_+}e^{-i\al t(c+i\ep_0)}\nu_+(F)(c+i\ep_0)\va_+(0,c+i\ep_0)dc\\
&\quad-\frac{1}{2\pi i}\int_{m_-}^{\ep_0}e^{-i\al t(c+i\ep_0)}\nu_+(F)(c+i\ep_0)\va_+(0,c+i\ep_0)dc\\
&\quad+\frac{1}{2\pi i}\int_{m_-}^{\ep_0}e^{-i\al t(c-i\ep_0)}\nu_+(F)(c-i\ep_0)\va_+(0,c-i\ep_0)dc\\
&\quad+\frac{1}{2\pi i}\int_{\ep_0}^{M_+}e^{-i\al t(c-i\ep_0)}\nu_+(F)(c-i\ep_0)\va_+(0,c-i\ep_0)dc\\
&\quad+\frac{1}{2\pi i}\int_{\pa B^{l}_{\ep_0}}e^{-i\al tc_{\ep}}\nu_+(F)(c_{\ep})\va_+(0,c_{\ep})dc_{\ep}
+\frac{1}{2\pi i}\int_{\pa B^r_{\ep_0}}e^{-i\al tc_{\ep}}\nu_+(F)(c_{\ep})\va_+(0,c_{\ep})dc_{\ep},
\end{align*}
by Proposition \ref{prop: mu nu lim} and the Lebesgue's dominated convergence theorem, we have
\begin{align*}
\lim_{\ep_0\to 0+}J(c_{\ep_0})
&=\frac{1}{2\pi i}\int_{m_-}^{M_+}e^{-i\al tc}\big(\nu_+^-(F)(c)-\nu_+^+(F)(c)\big)\va_+(0,c)dc.
\end{align*}
By Remark \ref{rmk: munu lim up rem}, we have
$\big|\nu_+^{\pm}(F)(c)\big|\leq C\|F\|_{L^{\infty}}\big(\big|\ln|c|\big|+1\big)\in L_c^1$ and
\begin{eqnarray*}
\big(\nu_+^-(F)(c)-\nu_+^+(F)(c)\big)\va_+(0,c)\in L_c^1,
\end{eqnarray*}
and the Riemann-Lebesgue Lemma implies that
$
\lim\limits_{t\to +\infty}\lim\limits_{\ep_0\rightarrow0+}J(c_{\ep_0})\rightarrow0.
$
From which, it implies that
\begin{eqnarray*}
\widehat{\psi}(t,\al,0)\rightarrow
\frac{u'(0)}{b'(0)}\widehat{\phi}_0(\al,0)
, \quad \textrm{as} \quad t\rightarrow+\infty.
\end{eqnarray*}
\noindent{\bf{Proof \ of \ 2.}} For the case of $0<y\leq 1$, for any $0<\ep\leq\ep_0$, let
$$
K(t,\al,y)=\lim_{\ep_0\to 0+}\frac{1}{2\pi i}\int_{\pa\Omega_{\ep_0}}e^{-i\al tc_{\ep}}\frac{u(y)-c_{\ep}}{c_{\ep}}\Theta_+(y,c_{\ep})dc_{\ep},
$$
and then
$
\widehat{\psi}(t,\al,y)
=\frac{u(y)}{b(y)}\widehat{\phi}_0(\al,0)\chi(y)+K(t,\al,y).
$
We divide $K(t,\al,y)$ into 6 parts and let
\begin{align*}
{\color{green}K_1(t,\al,y)}
&=\lim_{\ep\to0+}\frac{1}{2\pi i}
\int_{W_+(y)}^{M_+}
e^{-i\al t(c-i\ep)}\frac{u(y)-c+i\ep}{c-i\ep}\Theta_+^0(y,(c-i\ep))dc\\
&\quad+\lim_{\ep\to0+}\frac{1}{2\pi i}\int_{M_+}^{W_+(y)}
e^{-i\al t(c+i\ep)}\frac{u(y)-c-i{\ep}}{c+i{\ep}}\Theta_+^0(y,c+i{\ep})dc\\
&\quad+\lim_{\ep\to0+}\frac{1}{2\pi i}\int_{\pa B_{\ep}^r}
e^{-i\al tc_{\ep}}\frac{u(y)-c_{\ep}}{c_{\ep}}\Theta_+(y,c_{\ep})dc_{\ep},\\
{\color{blue}K_2(t,\al,y)}
&=\lim_{\ep\to0+}\frac{1}{2\pi i}
\int_{W_-(y)}^{m_-}
e^{-i\al t(c+i{\ep})}\frac{u(y)-c-i{\ep}}{c+i{\ep}}\Theta_+^0(y,c+i{\ep})dc\\
&\quad+\lim_{\ep\to0+}\frac{1}{2\pi i}\int_{m_-}^{ W_-(y)}
e^{-i\al t(c-i{\ep})}\frac{u(y)-c+i{\ep}}{c-i{\ep}}\Theta_+^0(y,c-i{\ep})dc\\
&\quad+\lim_{\ep\to0+}\frac{1}{2\pi i}\int_{\pa B_{\ep}^l}
e^{-i\al tc_{\ep}}\frac{u(y)-c_{\ep}}{c_{\ep}}\Theta_+(y,c_{\ep})dc_{\ep},
\end{align*}
and
\begin{align*}
&\lim_{\ep\to0+}\frac{1}{2\pi i}
\int_{W_-(\frac{y}{2})}^{W_+(\frac{y}{2})}e^{-i\al t(c-i{\ep})}
\frac{u(y)-c+i{\ep}}{c-i{\ep}}\Theta_+^1(y,c-i{\ep})dc\\
&\quad+\lim_{\ep\to0+}\frac{1}{2\pi i}\int_{W_+(\frac{y}{2})}^{ W_-(\frac{y}{2})}e^{-i\al t(c+i{\ep})}
\frac{u(y)-c-i{\ep}}{c+i{\ep}}\Theta_+^1(y,c+i{\ep})dc\\
&=\lim_{\ep\to0+}\frac{1}{2\pi i}
\int_{W_-(\frac{y}{2})}^{W_+(\frac{y}{2})}e^{-i\al t(c-i{\ep})}
\frac{u(y)-c+i{\ep}}{c-i{\ep}}\int_1^y\frac{\va_+(y,c-i{\ep})\int_{y_{c_+}}^{y'}(F\va_+)(z,c-i{\ep})dz}
{\mathcal{H}(y',c-i{\ep})\va_+(y',c-i{\ep})^2}dy'dc\\
&\quad+\lim_{\ep\to0+}\frac{1}{2\pi i}\int_{W_+(\frac{y}{2})}^{ W_-(\frac{y}{2})}e^{-i\al t(c+i{\ep})}
\frac{u(y)-c-i{\ep}}{c+i{\ep}}\int_1^y\frac{\va_+(y,c+i{\ep})\int_{y_{c_+}}^{y'}(F\va_+)(z,c+i{\ep})dz}
{\mathcal{H}(y',c+i{\ep})\va_+(y',c+i{\ep})^2}dy'dc\\
&\quad+\lim_{\ep\to0+}\frac{1}{2\pi i}
\int_{W_-(\frac{y}{2})}^{W_+(\frac{y}{2})}e^{-i\al t(c-i{\ep})}
\frac{u(y)-c+i{\ep}}{c-i{\ep}}\int_{1}^y\frac{\mu_+(F)(c-i{\ep})\va_+(y,c-i{\ep})}
{(\mathcal{H}\va_+^2)(y',c-i{\ep})}dy'dc\\
&\quad+\lim_{\ep\to0+}\frac{1}{2\pi i}\int_{W_+(\frac{y}{2})}^{ W_-(\frac{y}{2})}e^{-i\al t(c+i{\ep})}
\frac{u(y)-c-i{\ep}}{c+i{\ep}}\int_{1}^y\frac{\mu_+(F)(c+i{\ep})\va_+(y,c+i{\ep})}
{(\mathcal{H}\va_+^2)(y',c+i{\ep})}dy'dc\\
&\buildrel\hbox{\footnotesize def}\over ={\color{red}K_3(t,\al,y)+K_4(t,\al,y)},
\end{align*}
and let
\begin{align*}
K_5(t,\al,y)&=\lim_{\ep\to0+}\frac{1}{2\pi i}
\int_{W_+(\frac{y}{2})}^{W_+(y)}
e^{-i\al t(c-i{\ep})}\frac{u(y)-c+i{\ep}}{c-i{\ep}}\Theta_+^1(y,c-i{\ep})dc\\
&\quad+\lim_{\ep\to0+}\frac{1}{2\pi i}
\int_{W_+(y)}^{ W_+(\frac{y}{2})}
e^{-i\al t(c+i{\ep})}\frac{u(y)-c-i{\ep}}{c+i{\ep}}\Theta_+^1(y,c+i{\ep})dc\\
K_6(t,\al,y)&=\lim_{\ep\to0+}\frac{1}{2\pi i}
\int_{W_-(\frac{y}{2})}^{W_-(y)}
e^{-i\al t(c+i{\ep})}\frac{u(y)-c-i{\ep}}{c+i{\ep}}\Theta_+^1(y,c+i{\ep})dc\\
&\quad+\lim_{\ep\to0+}\frac{1}{2\pi i}
\int_{W_-(y)}^{ W_-(\frac{y}{2})}
e^{-i\al t(c-i{\ep})}\frac{u(y)-c+i{\ep}}{c-i{\ep}}\Theta_+^1(y,c-i{\ep})dc,
\end{align*}
so that $K(t,\al,y)=\sum\limits_{i=1}^{6}K_i(t,\al,y)$.
And for convenience, we give a picture to show that how we depart the contour domain:\\
\begin{tikzpicture}[domain=0:4]
\draw[thick] (-4,0) -- (4,0);
\draw[thick] (-4,2) -- (4,2);
\draw[thick] (-4,-2) -- (4,-2);
\draw[blue,thick] (-4,2) -- (-2,2);
\draw[blue,thick] (-4,-2) -- (-2,-2);
\draw (-0.05,-0.2) node {$0$};
\draw (-3.8,0.3) node {\small $m_-$};
\draw (4.2,0.3) node {\small $M_+$};
\draw[dashed, color=red] (-1,2) -- (-1,-2);
\draw[thick,red] (-1,2) -- (1,2);
\draw[->,dashed, color=red](-1,2) -- (-1,1);
\draw[->,dashed,red] (-1,-2) -- (0,-2);
\draw[red,thick] (-1,-2) -- (1,-2);
\draw (-1,-0.3) node {\small $W_-(\frac{y}{2})$};
\draw[dashed, color=red] (1,2) -- (1,-2);
\draw[->,dashed, color=red] (1,0) -- (1,1);
\draw[->,thick,red] (1,2) -- (0,2);
\draw[->] (-1,2) -- (-1.5,2);
\draw[->] (-2,-2) -- (-1.5,-2);
\draw[->] (2,2) -- (1.5,2);
\draw[->] (1,-2) -- (1.5,-2);
\draw[->,blue] (-2,2) -- (-3,2);
\draw[->,blue] (-4,-2) -- (-3,-2);
\draw[->,green] (4,2) -- (3,2);
\draw[->,green] (2,-2) -- (3,-2);
\draw[thick,green] (2,-2) -- (4,-2);
\draw[thick,green] (2,2) -- (4,2);
\draw (1,-0.3) node {\small $W_+(\frac{y}{2})$};
\draw[dashed] (-2,2) -- (-2,-2);
\draw (-2.6,-0.3) node {\small $W_-(y)$};
\draw[dashed] (2,2) -- (2,-2);
\draw (2.6,-0.3) node {\small $W_+(y)$};
\draw[green, thick] (4, 2) arc [start angle = 90, end angle = -90,
x radius = 12.6mm, y radius = 19.95mm];
\draw[<-,green] (4, 2) arc [start angle = 90, end angle = -90,
x radius = 12.6mm, y radius = 19.95mm];
\draw (5.8,0) node {\small $\pa B_{\epsilon}^r$};
\draw[blue, thick] (-4, 2) arc [start angle = 90, end angle = -90,
x radius = -12.6mm, y radius = 19.95mm];
\draw[->,blue] (-4, 2) arc [start angle = 90, end angle = -30,
x radius = -12.6mm, y radius = 19.95mm];
\draw (-5.7,0) node {\small $\pa B_{\epsilon}^l$};
\draw (0,2.5) node {$\pa\Om_{\ep}$};
\draw (-0.7,0.7) node {\small \color{red}$\Gamma_{\ep}$};
\draw (-4.3,-2.2) node {\small $m_--i\epsilon$};
\draw (-4.3,2.2) node {\small$m_-+i\epsilon$};
\draw (4.3,-2.2) node {\small$M_+-i\epsilon$};
\draw (4.3,2.2) node {\small $M_++i\epsilon$};
\end{tikzpicture}
\noindent As for $K_1$, we have
\begin{align*}
&K_1(t,\al,y)\\
&=\lim_{\ep\to0+}\frac{1}{2\pi i}
\int_{W_+(y)}^{M_+}e^{-i\al t(c-i\ep)}\frac{u(y)-c+i\ep}{c-i\ep}
\Big\{\va_+(y,c-i\ep)\int_0^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c-i\ep)dz}
{(\mathcal{H}\va_+^2)(y',c-i\ep)}dy'\\
&\quad +\mu_+(F)(c-i\ep)\va_+(y,c-i\ep)
\int_0^y\frac{1}{(\mathcal{H}\va_+^2)(y',c-i\ep)}dy'
+\nu_+(F)(c-i\ep)\va_+(y,c-i\ep)\Big\}dc\\
&\quad -\lim_{\ep\to0+}\frac{1}{2\pi i}
\int_{W_+(y)}^{M_+}e^{-i\al t(c+i\ep)}
\frac{u(y)-c-i\ep}{c+i\ep}\Big\{\va_+(y,c+i\ep)
\int_0^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c+i\ep)dz}
{(\mathcal{H}\va_+^2)(y',c+i\ep)}dy'\\
&\quad +\mu_+(F)(c+i\ep)\va_+(y,c+i\ep)
\int_0^y\frac{1}{(\mathcal{H}\va_+^2)(y',c+i\ep)}dy'
+\nu_+(F)(c+i\ep)\va_+(y,c+i\ep)\Big\}dc\\
&\quad -\lim_{\ep\to0+}\frac{\ep}{2\pi}
\int_{\frac{\pi}{2}}^{\frac{3\pi}{2}}e^{-i\al t\big(M_++\ep e^{i\theta}\big)}
\frac{u(y)-M_++\ep e^{i\theta}}{M_+-\ep e^{i\theta}}
\Theta_+\big(y,M_+-\ep e^{i\theta}\big) e^{i\theta}d\theta\nonumber\\
&=K_{11}(t,\al,y)+K_{12}(t,\al,y)+K_{13}(t,\al,y).
\end{align*}
Proposition \ref{prop: Br Bl Theta bdd} implies
$
K_{13}(t,\al,y)=0.
$\\
For $c\in \big[W_+(y),M_+\big]$ with $y\in (0,1]$ fixed, by Proposition \ref{prop: mu nu lim}, Proposition \ref{prop: F y bdd lim}, Remark \ref{rmk: D} and the Lebesgue's dominated convergence theorem, we obtain that
\begin{align*}
&K_1(t,\al,y)=K_{11}(t,\al,y)+K_{12}(t,\al,y)\\
&=\frac{1}{2\pi i}\int_{W_+(y)}^{M_+}e^{-i \al tc}\frac{u(y)-c}{c}
\Big\{\big(\mu_+^-(F)(c)-\mu_+^+(F)(c)\big)
\int_0^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'\\
\nonumber&\quad +\big(\nu_+^-(F)(c)-\nu_+^+(F)(c)\big)\va_+(y,c)\Big\}dc\\
&= \frac{1}{\pi}\int_{W_+(y)}^{M_+}e^{-i \al tc}\frac{u(y)-c}{c}
\bigg(\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_+(F)(c)-\mathcal{D}^{re}(c)
\mathcal{U}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_0^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'\\
&\quad
+\frac{\mathcal{D}^{im}(c)\mathcal{V}^{re}_+(F)(c)-\mathcal{D}^{re}(c)
\mathcal{V}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}\va_+(y,c)\bigg)dc.
\end{align*}
Here for $c\in[W_+(y), M_+]$, by Remark \ref{rmk: D}, Lemma \ref{lem: limit up bdd}, Lemma \ref{lem: XYST bdd} and Proposition \ref{prop: F y bdd lim}, we have
\begin{align*}
&\Big|\frac{u(y)-c}{c}
\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_+(F)(c)-\mathcal{D}^{re}(c)
\mathcal{U}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_0^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'\Big|\\
&\quad\leq C\|F\|_{L^{\infty}}\big(\big|\ln|y-y_{c_+}|\big|+1\big)\in L_c^1\big(W_+(y), M_+\big),\\
&\Big|\frac{u(y)-c}{c}\frac{\mathcal{D}^{im}(c)\mathcal{V}^{re}_+(F)(c)
-\mathcal{D}^{re}(c)\mathcal{V}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}\va_+(y,c)\Big|\\
&\quad\leq C\|F\|_{L^{\infty}}\in L_c^1\big(W_+(y), M_+\big).
\end{align*}
And then the Riemann-Lebesgue lemma gives
\begin{eqnarray*}
\lim_{t\rightarrow+\infty}K_{11}(t,\al,y)+K_{12}(t,\al,y)=0.
\end{eqnarray*}
Thus we get
$\lim\limits_{t\to+\infty}K_1(t,\al,y)=0.$
By the same argument, we obtain
\begin{align*}
&K_2(t,\al,y)\\
&=\frac{1}{\pi}\int_{m_-}^{W_-(y)}e^{-i \al tc}
\frac{u(y)-c}{c}\Big\{\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_+(F)(c)
-\mathcal{D}^{re}(c)\mathcal{U}^{im}_+(F)(c)}{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_0^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'\\
&\quad
+\frac{\mathcal{D}^{im}(c)\mathcal{V}^{re}_+(F)(c)
-\mathcal{D}^{re}(c)\mathcal{V}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}\va_+(y,c)\Big\}dc,
\end{align*}
and
\begin{align*}
&\Big|\frac{u(y)-c}{c}
\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_+(F)(c)-\mathcal{D}^{re}(c)
\mathcal{U}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_0^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'\Big|\\
&\quad\leq C\|F\|_{L^{\infty}}\big(\big|\ln|y-y_{c_+}|\big|+1\big)\in L_c^1\big(m_-,W_-(y)\big),\\
&\Big|\frac{u(y)-c}{c}\frac{\mathcal{D}^{im}(c)\mathcal{V}^{re}_+(F)(c)
-\mathcal{D}^{re}(c)\mathcal{V}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}\va_+(y,c)\Big|\\
&\quad\leq C\|F\|_{L^{\infty}}\in L_c^1\big(m_-,W_-(y)\big).
\end{align*}
Thus
$\lim\limits_{t\rightarrow+\infty}K_2(t,\al,y)=0.$
We rewrite $K_5$ as follows
\begin{align*}
&K_5(t,\al,y)\\
&=\lim_{\ep\rightarrow0^+}\frac{1}{2\pi i}
\int_{W_+(\frac{y}{2})}^{W_+(y)}e^{-i\al t(c-i\ep)}
\frac{u(y)-c+i\ep}{c-i\ep}
\Big\{\va_+(y,c-i\ep)\int_1^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c-i\ep)dz}
{(\mathcal{H}\va_+^2)(y',c-i\ep)}dy'\\
&\quad +\mu_+(F)(c-i\ep)
\int_1^y\frac{\va_+(y,c-i\ep)}{(\mathcal{H}\va_+^2)(y',c-i\ep)}dy'\Big\}dc\\
&\quad -\lim_{\ep\rightarrow0^+}\frac{1}{2\pi i}
\int_{W_+(\frac{y}{2})}^{W_+(y)}e^{-i\al t(c+i\ep)}
\frac{u(y)-c-i\ep}{c+i\ep}
\Big\{\va_+(y,c+i\ep)\int_1^y\frac{\int_{y_{c_+}}^{y'}(F\va_+)(z,c+i\ep)dz}
{(\mathcal{H}\va_+^2)(y',c+i\ep)}dy'\\
&\quad +\mu_+(F)(c+i\ep)
\int_1^y\frac{\va_+(y,c+i\ep)}{(\mathcal{H}\va_+^2)(y',c+i\ep)}dy'\Big\}dc.
\end{align*}
For $c\in[W_+(\frac{y}{2}), W_+(y)]$, then $0<\frac{y}{2}\leq y_{c_+}\leq y\leq1$. Thus by Proposition \ref{prop: F y bdd lim}, Remark \ref{rmk: D} and the Lebesgue's dominated convergence theorem, we get
\begin{align*}
&K_5(t,\al,y)\\
&=\frac{1}{2\pi i}
\int_{W_+(\frac{y}{2})}^{W_+(y)}e^{-i\al tc}\frac{u(y)-c}{c}\big(\mu_+^-(F)(c)-\mu_+^+(F)(c)\big)
\int_1^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'dc\\
&=\frac{1}{\pi}
\int_{W_+(\frac{y}{2})}^{W_+(y)}e^{-i\al tc}\frac{u(y)-c}{c}
\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_+(F)(c)-\mathcal{D}^{re}(c)\mathcal{U}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_1^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'dc.
\end{align*}
And for $c\in[W_+(\frac{y}{2}), W_+(y)]$, by Remark \ref{rmk: D} and Lemma \ref{lem: limit up bdd} and Lemma \ref{lem: XYST bdd}, we obtain
\begin{align*}
&\Big|\frac{u(y)-c}{c}
\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_+(F)(c)
-\mathcal{D}^{re}(c)\mathcal{U}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_1^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'\Big|\\
&\leq C(y)\|F\|_{L^{\infty}}\big(\big|\ln|y-y_{c_+}|\big|+1\big)\in L_c^1\big(W_+(y/2), W_+(y)\big),
\end{align*}
and then the Riemann-Lebesgue lemma implies $\lim\limits_{t\rightarrow+\infty}K_5(t,\al,y)=0.$\\
By the same argument, we can also get
\begin{align*}
&K_6(t,\al,y)\\
&=\frac{1}{\pi}
\int^{W_-(\frac{y}{2})}_{W_-(y)}e^{-i\al tc}\frac{u(y)-c}{c}
\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_+(F)(c)-\mathcal{D}^{re}(c)\mathcal{U}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_1^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'dc,
\end{align*}
and $\lim\limits_{t\rightarrow+\infty}K_6(t,\al,y)=0.$
In the following, we mainly calculate the term $K_3(t,\al,y)$ and $K_4(t,\al,y)$.
Recall that
\begin{align*}
K_4(t,\al,y)&=\lim_{\ep\to0+}\frac{1}{2\pi i}
\int_{W_-(\frac{y}{2})}^{W_+(\frac{y}{2})}e^{-i\al t(c-i{\ep})}
\frac{u(y)-c+i{\ep}}{c-i{\ep}}\int_{1}^y\frac{\mu_+(F)(c-i{\ep})\va_+(y,c-i{\ep})}
{(\mathcal{H}\va_+^2)(y',c-i{\ep})}dy'dc\\
&\quad+\lim_{\ep\to0+}\frac{1}{2\pi i}\int_{W_+(\frac{y}{2})}^{ W_-(\frac{y}{2})}e^{-i\al t(c+i{\ep})}
\frac{u(y)-c-i{\ep}}{c+i{\ep}}\int_{1}^y\frac{\mu_+(F)(c+i{\ep})\va_+(y,c+i{\ep})}
{(\mathcal{H}\va_+^2)(y',c+i{\ep})}dy'dc,
\end{align*}
By Proposition \ref{prop: mu nu lim} and by Proposition \ref{prop: F y bdd lim} and the Lebesgue's dominated convergence theorem, we obtain
\begin{align*}
K_4(t,\al,y)=\frac{1}{2\pi i}
\int_{W_-(\frac{y}{2})}^{W_+(\frac{y}{2})}e^{-i\al tc}
\frac{u(y)-c}{c}(\mu_+^-(F)(c)-\mu_+^+(F)(c))
\int_1^y\frac{\va_+(y,c)}
{(\mathcal{H}\va_+^2)(y',c)}dy'dc.
\end{align*}
And from Remark \ref{rmk: munu lim up rem} and Proposition \ref{prop: F y bdd lim}, we have for $c\in[W_-(\frac{y}{2}), W_+(\frac{y}{2})]$,
\begin{align*}
&\Big|\frac{u(y)-c}{c}\big(\mu_+^-(F)(c)-\mu_+^+(F)(c)\big)
\int_1^y\frac{\va_+(y,c)}
{(\mathcal{H}\va_+^2)(y',c)}dy'dc\Big|\\
&\ \leq \frac{C\|F\|_{L^{\infty}}\big(\big|\ln|y-y_{c_+}|\big|+1\big)}{|y|}\big(\big|\ln|c|\big|+1\big)
\in L_c^1\big(W_-(y/2), W_+(y/2)\big).
\end{align*}
And then the Riemann-Lebesgue lemma gives $
\lim\limits_{t\rightarrow+\infty}K_4(t,\al,y)=0.$
Let
\begin{eqnarray*}
H(y,c_{\ep})=\int_1^y\frac{\va_+(y,c_{\ep})}
{(\mathcal{H}\va_+^2)(y',c_{\ep})}\int_{y_{c_+}}^{y'}(F\va_+)(z,c_{\ep})dz
dy',
\end{eqnarray*}
and $\Gamma_{\ep}$ be the boundary of $\big\{c:~W_-(\f{y}{2})\leq Re c\leq W_+(\f{y}{2}),\ |Im c|\leq \ep\big\}$. Then we have
\begin{align*}
&K_3(t,\al,y)\\
&=\lim_{\ep\to0+}\frac{1}{2\pi i}
\int_{W_-(\frac{y}{2})}^{W_+(\frac{y}{2})}e^{-i\al t(c-i{\ep})}
\frac{u(y)-c+i{\ep}}{c-i{\ep}}\big(H(y,c-i\ep)-H(y,0)\big)dc\\
&\quad+\lim_{\ep\to0+}\frac{1}{2\pi i}\int_{W_+(\frac{y}{2})}^{ W_-(\frac{y}{2})}e^{-i\al t(c+i{\ep})}
\frac{u(y)-c-i{\ep}}{c+i{\ep}}\big(H(y,c+i\ep)-H(y,0)\big)dc\\
&\quad+H(y,0)\lim_{\ep\to0+}\frac{1}{2\pi i}\int_{\Gamma_{\ep}}
e^{-i\al tc}\frac{u(y)-c}{c}dc\\
&\quad
-H(y,0)\lim_{\ep\to0+}\frac{1}{2\pi}
\int_{-\ep}^{\ep}e^{-i\al t\big(W_+(\frac{y}{2})+i\tau\big)}
\frac{u(y)-W_+(\frac{y}{2})-i\tau}{W_+(\frac{y}{2})+i\tau}d\tau\\
&\quad
-H(y,0)\lim_{\ep\rightarrow0^+}\frac{1}{2\pi }
\int_{\ep}^{-\ep}e^{-i\al t\big(W_-(\frac{y}{2})+i\tau\big)}
\frac{u(y)-W_-(\frac{y}{2})-i\tau}{W_-(\frac{y}{2})+i\tau}d\tau\\
&=K_{31}(t,\al,y)+K_{32}(t,\al,y)+K_{33}(t,\al,y)+K_{34}(t,\al,y)+K_{35}(t,\al,y),
\end{align*}
where we can easily get that
\begin{eqnarray*}
K_{33}(t,\al,y)=H(y,0)\lim_{\ep\rightarrow0+}\frac{1}{2\pi i}
\int_{\Gamma_{\ep}}e^{-i\al tc_{\ep}}\frac{u(y)-c_{\ep}}{c_{\ep}}dc_{\ep}=u(y)H(y,0),
\end{eqnarray*}
and $K_{34}(t,\al,y)=K_{35}(t,\al,y)=0$.
As for $K_{31}$ and $K_{32}$, we have for $c_{\ep}=c\pm i\ep$,
\begin{align*}
&H(y,c_{\ep})-H(y,0)\\
&=\big(\va_+(y,c_{\ep})-\va_+(y,0)\big)
\int_1^y\frac{1}{(\mathcal{H}\va_+^2)(y',c_{\ep})}
\int_{y_{c_+}}^{y'}(F\va_+)(z,c_{\ep})dzdy'\\
&\quad +\va_+(y,0)\int_1^y\Big(\frac{1}{(\mathcal{H}\va_+^2)(y',c_{\ep})}
-\frac{1}{(\mathcal{H}\va_+^2)(y',0)}\Big)
\int_{y_{c_+}}^{y'}(F\va_+)(z,c_{\ep})dzdy'\\
&\quad-\va_+(y,0)\int_1^y\frac{1}{(\mathcal{H}\va_+^2)(y',0)}
\int_0^{y_{c_+}}(F\va_+)(z,c_{\ep})dzdy'\\
&\quad +\va_+(y,0)\int_1^y\frac{1}{(\mathcal{H}\va_+^2)(y',0)}
\int_0^{y'}\big[(F\va_+)(z,c_{\ep})-(F\va_+)(z,0)\big]dzdy'.
\end{align*}
Then by Proposition \ref{prop: sol. hom. [0,1]} and Proposition \ref{prop: F y bdd lim}, we have
\begin{eqnarray*}
|H(y,c_{\ep})-H(y,0)|\leq C(y)|c_{\ep}|,
\end{eqnarray*}
and $\lim\limits_{\ep\to 0+}(H(y,c\pm i{\ep})-H(y,0))=H(y,c)-H(y,0)$
and then the Lebesgue's dominated convergence theorem gives
\begin{eqnarray*}
K_{31}(t,\al,y)=\frac{1}{2\pi i}
\int_{W_-(\frac{y}{2})}^{W_+(\frac{y}{2})}e^{-i\al tc}
\frac{u(y)-c}{c}\big(H(y,c)-H(y,0)\big)dc=-K_{32}(t,\al,y).
\end{eqnarray*}
Thus we obtain $K_3(t,\al,y)=u(y)H(y,0)$ with
\begin{eqnarray*}
H(y,0)=\va_+(y,0)\int_1^y\frac{\int_0^{y'}F(z,0)\va_+(z,0)dz}
{\big(u(y')^2-b(y')^2\big)\va_+(y',0)^2}dy'.
\end{eqnarray*}
Therefore we get for $y\in(0,1]$,
\begin{align*}
\widehat{\psi}(t,\al,y)
&=\frac{u(y)}{b(y)}\widehat{\phi}_0(\al,0)\chi(y)+K(t,\al,y)\\
&=\frac{u(y)}{b(y)}\widehat{\phi}_0(\al,0)\chi(y)+u(y)H(y,0)+R_1^+(t,\al,y)+R_2^+(t,\al,y)+R_3^+(t,\al,y),
\end{align*}
where
\begin{align*}
R_{1}^+(t,\al,y)&=\frac{1}{\pi}\int_{W_+(y)}^{M_+}e^{-i \al tc}\frac{u(y)-c}{c}
\bigg(\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_+(F)(c)-\mathcal{D}^{re}(c)
\mathcal{U}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_0^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'\\
&\quad
+\frac{\mathcal{D}^{im}(c)\mathcal{V}^{re}_+(F)(c)-\mathcal{D}^{re}(c)
\mathcal{V}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}\va_+(y,c)\bigg)dc,\\
R_2^+(t,\al,y)
&=\frac{1}{\pi}\int_{m_-}^{W_-(y)}e^{-i \al tc}
\frac{u(y)-c}{c}\bigg(\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_+(F)(c)
-\mathcal{D}^{re}(c)\mathcal{U}^{im}_+(F)(c)}{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_0^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'\\
&\quad
+\frac{\mathcal{D}^{im}(c)\mathcal{V}^{re}_+(F)(c)
-\mathcal{D}^{re}(c)\mathcal{V}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}\va_+(y,c)\bigg)dc,\\
R_3^+(t,\al,y)
&=\frac{1}{\pi}
\int_{W_-(y)}^{W_+(y)}e^{-i\al tc}\frac{u(y)-c}{c}
\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_+(F)(c)-\mathcal{D}^{re}(c)\mathcal{U}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_1^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'dc,
\end{align*}
and $R_i^+(t,\al,y)\to 0$ as $t\to +\infty$ for $i=1,2,3,$ and $y>0$.
Similarly, we also get that for $0<y\leq 1$,
\begin{align*}
\widehat{\phi}(t,\al,y)
&=\widehat{\phi}_0(\al,0)\chi(y)+b(y)H(y,0)+R_4^+(t,\al,y)+R_5^+(t,\al,y)+R_6^+(t,\al,y),
\end{align*}
where
\begin{align*}
R_{4}^+(t,\al,y)&=\frac{1}{\pi}\int_{W_+(y)}^{M_+}e^{-i \al tc}\frac{b(y)}{c}
\bigg(\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_+(F)(c)-\mathcal{D}^{re}(c)
\mathcal{U}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_0^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'\\
&\quad
+\frac{\mathcal{D}^{im}(c)\mathcal{V}^{re}_+(F)(c)-\mathcal{D}^{re}(c)
\mathcal{V}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}\va_+(y,c)\bigg)dc,\\
R_5^+(t,\al,y)
&=\frac{1}{\pi}\int_{m_-}^{W_-(y)}e^{-i \al tc}
\frac{b(y)}{c}\bigg(\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_+(F)(c)
-\mathcal{D}^{re}(c)\mathcal{U}^{im}_+(F)(c)}{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_0^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'\\
&\quad
+\frac{\mathcal{D}^{im}(c)\mathcal{V}^{re}_+(F)(c)
-\mathcal{D}^{re}(c)\mathcal{V}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}\va_+(y,c)\bigg)dc,\\
R_6^+(t,\al,y)
&=\frac{1}{\pi}
\int_{W_-(y)}^{W_+(y)}e^{-i\al tc}\frac{b(y)}{c}
\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_+(F)(c)-\mathcal{D}^{re}(c)\mathcal{U}^{im}_+(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_1^y\frac{\va_+(y,c)}{(\mathcal{H}\va_+^2)(y',c)}dy'dc,
\end{align*}
and $R_i^+(t,\al,y)\to 0$ as $t\to +\infty$ for $i=4,5,6,$ and $y>0$.\\
\noindent{\bf{Proof \ of \ 3.}} For $y\in [-1,0)$, we can get the conclusion by the same method and we obtain that
\begin{align*}
\widehat{\psi}(t,\al,y)
&=\frac{u(y)}{b(y)}\widehat{\phi}_0(\al,0)\chi(y)+u(y)\widetilde{H}(y,0)-R_1^-(t,\al,y)-R_2^-(t,\al,y)-R_3^-(t,\al,y),
\end{align*}
and
\begin{align*}
\widehat{\phi}(t,\al,y)
&=\widehat{\phi}_0(\al,0)\chi(y)+b(y)\widetilde{H}(y,0)-R_4^-(t,\al,y)-R_5^-(t,\al,y)-R_6^-(t,\al,y),
\end{align*}
where
\begin{eqnarray*}
\widetilde{H}(y,0)=\va_-(y,0)\int_{-1}^y\frac{1}{\big(u(y')^2-b(y')^2\big)\va_-(y',0)^2}
\int_0^{y'}F(z,0)\va_-(z,0)dzdy'.
\end{eqnarray*}
and
\begin{align*}
R_{1}^-(t,\al,y)&=\frac{1}{\pi}\int_{W_+(y)}^{m_-}e^{-i \al tc}\frac{u(y)-c}{c}
\bigg(\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_-(F)(c)-\mathcal{D}^{re}(c)
\mathcal{U}^{im}_-(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_0^y\frac{\va_-(y,c)}{(\mathcal{H}\va_-^2)(y',c)}dy'\\
&\quad
+\frac{\mathcal{D}^{im}(c)\mathcal{V}^{re}_-(F)(c)-\mathcal{D}^{re}(c)
\mathcal{V}^{im}_-(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}\va_-(y,c)\bigg)dc,\\
R_2^-(t,\al,y)
&=\frac{1}{\pi}\int_{M_+}^{W_-(y)}e^{-i \al tc}
\frac{u(y)-c}{c}\bigg(\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_-(F)(c)
-\mathcal{D}^{re}(c)\mathcal{U}^{im}_-(F)(c)}{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_0^y\frac{\va_-(y,c)}{(\mathcal{H}\va_-^2)(y',c)}dy'\\
&\quad
+\frac{\mathcal{D}^{im}(c)\mathcal{V}^{re}_-(F)(c)
-\mathcal{D}^{re}(c)\mathcal{V}^{im}_-(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}\va_-(y,c)\bigg)dc,\\
R_3^-(t,\al,y)
&=\frac{1}{\pi}
\int_{W_-(y)}^{W_+(y)}e^{-i\al tc}\frac{u(y)-c}{c}
\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_-(F)(c)-\mathcal{D}^{re}(c)\mathcal{U}^{im}_-(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_{-1}^y\frac{\va_-(y,c)}{(\mathcal{H}\va_-^2)(y',c)}dy'dc,\\
R_{4}^-(t,\al,y)&=\frac{1}{\pi}\int_{W_+(y)}^{m_-}e^{-i \al tc}\frac{b(y)}{c}
\bigg(\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_-(F)(c)-\mathcal{D}^{re}(c)
\mathcal{U}^{im}_-(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_0^y\frac{\va_-(y,c)}{(\mathcal{H}\va_-^2)(y',c)}dy'\\
&\quad
+\frac{\mathcal{D}^{im}(c)\mathcal{V}^{re}_-(F)(c)-\mathcal{D}^{re}(c)
\mathcal{V}^{im}_-(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}\va_-(y,c)\bigg)dc,\\
R_5^-(t,\al,y)
&=\frac{1}{\pi}\int_{M_+}^{W_-(y)}e^{-i \al tc}
\frac{b(y)}{c}\bigg(\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_-(F)(c)
-\mathcal{D}^{re}(c)\mathcal{U}^{im}_-(F)(c)}{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_0^y\frac{\va_-(y,c)}{(\mathcal{H}\va_-^2)(y',c)}dy'\\
&\quad
+\frac{\mathcal{D}^{im}(c)\mathcal{V}^{re}_-(F)(c)
-\mathcal{D}^{re}(c)\mathcal{V}^{im}_-(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}\va_-(y,c)\bigg)dc,\\
R_6^-(t,\al,y)
&=\frac{1}{\pi}
\int_{W_-(y)}^{W_+(y)}e^{-i\al tc}\frac{b(y)}{c}
\frac{\mathcal{D}^{im}(c)\mathcal{U}^{re}_-(F)(c)-\mathcal{D}^{re}(c)\mathcal{U}^{im}_-(F)(c)}
{\mathcal{D}^{re}(c)^2+\mathcal{D}^{im}(c)^2}
\int_{-1}^y\frac{\va_-(y,c)}{(\mathcal{H}\va_-^2)(y',c)}dy'dc,
\end{align*}
and $R_i^-(t,\al,y)\to 0$ as $t\to +\infty$ for $i=1,2,...,6,$ and $y<0$.
At last, we give the calculation of the term $H(y,0)$.
Firstly, due to the fact that
$$
F(y,0)=-\widehat{\phi}_0(\al,0)
\left(\pa_y\Big[\big(u(y)^2-b(y)^2\big)\pa_y\Big(\frac{\chi(y)}{b(y)}\Big)\Big]-\al^2 \big(u(y)^2-b(y)^2\big)\frac{\chi(y)}{b(y)}\right),
$$
we have by using the integration by parts,
\begin{align*}
&\int_0^{y'}(F\va_+)(z,0)dz\\
&=\widehat{\phi}_0(\al,0)\Big\{\int_0^{y'} \big(u(z)^2-b(z)^2\big)\pa_z\Big(\frac{\chi(z)}{b(z)}\Big)\pa_z\va_+(z,0)dz
- \big(u(z)^2-b(z)^2\big)\pa_z\Big(\frac{\chi(z)}{b(z)}\Big)\va_+(z,0)\Big|_0^{y'}\\
&\ \ +\int_0^{y'}\al^2 \big(u(z)^2-b(z)^2\big)\frac{\chi(z)}{b(z)}\va_+(z,0)dz\Big\}\\
&=-\widehat{\phi}_0(\al,0)\Big\{\int_0^{y'} \pa_z\Big(\big(u(z)^2-b(z)^2\big)\pa_z\va_+(z,0)\Big)\frac{\chi(z)}{b(z)}dz
-\big(u(z)^2-b(z)^2\big)\pa_z\va_+(z,0)\frac{\chi(z)}{b(z)}\Big|_0^{y'}\\
&\ \
+\big(u(z)^2-b(z)^2\big)\pa_z\Big(\frac{\chi(z)}{b(z)}\Big)\va_+(z,0)\Big|_0^{y'}
-\int_0^{y'}\al^2 \big(u(z)^2-b(z)^2\big)\frac{\chi(z)}{b(z)}\va_+(z,0)dz\Big\}\\
&=\widehat{\phi}_0(\al,0)\Big\{\big(u(z)^2-b(z)^2\big)
\pa_z\va_+(z,0)\frac{\chi(z)}{b(z)}\Big|_0^{y'}
- \big(u(z)^2-b(z)^2\big)\pa_z\Big(\frac{\chi(z)}{b(z)}\Big)\va_+(z,0)\Big|_0^{y'}\Big\}\\
&=-\widehat{\phi}_0(\al,0)b'(0)\frac{u'(0)^2-b'(0)^2}{b'(0)^2}-\widehat{\phi}_0(\al,0)(u(y')^2-b(y')^2)\pa_{y'}\big(\f{\chi}{b}\big)(y')\va_+(y',0)\\
&\quad+\widehat{\phi}_0(\al,0)(u(y')^2-b(y')^2)\f{\chi(y')}{b(y')}\pa_{y'}\va_+(y',0).
\end{align*}
Thus we obtain for $y>0$,
\begin{eqnarray*}
H(y,0)=-\widehat{\phi}_0(\al,0)\frac{u'(0)^2-b'(0)^2}{b'(0)}
\va_+(y,0)\int_1^y\frac{1}{\big(u(y')^2-b(y')^2\big)\va_+(y',0)^2}dy'-\f{\chi(y)\widehat{\phi}_0(\al,0)}{b(y)},
\end{eqnarray*}
and then for $0<y\leq 1$, we get
\begin{align*}
&\f{u(y)}{b(y)}\widehat{\phi}_0(\al,0)\chi(y)+u(y)H(y,0)=
-\frac{u'(0)^2-b'(0)^2}{b'(0)}\widehat{\phi}_0(\al,0)
\int_1^y\frac{u(y)\va_+(y,0)}{\big(u(y')^2-b(y')^2\big)\va_+(y',0)^2}dy',\\
&\widehat{\phi}_0(\al,0)\chi(y)+b(y)H(y,0)=
-\frac{u'(0)^2-b'(0)^2}{b'(0)}\widehat{\phi}_0(\al,0)
\int_1^y\frac{b(y)\va_+(y,0)}{\big(u(y')^2-b(y')^2\big)\va_+(y',0)^2}dy'.
\end{align*}
Similar as the case of $y\in(0,1]$, we have
\begin{align*}
&\f{u(y)}{b(y)}\widehat{\phi}_0(\al,0)\chi(y)+u(y)\widetilde{H}(y,0)=
-\frac{u'(0)^2-b'(0)^2}{b'(0)}\widehat{\phi}_0(\al,0)
\int_{-1}^y\frac{u(y)\va_+(y,0)}{\big(u(y')^2-b(y')^2\big)\va_-(y',0)^2}dy',\\
&\widehat{\phi}_0(\al,0)\chi(y)+b(y)\widetilde{H}(y,0)=
-\frac{u'(0)^2-b'(0)^2}{b'(0)}\widehat{\phi}_0(\al,0)
\int_{-1}^y\frac{b(y)\va_+(y,0)}{\big(u(y')^2-b(y')^2\big)\va_-(y',0)^2}dy'.
\end{align*}
Thus we complete the proof of Theorem \ref{main thm}.
\end{proof}
\section*{Acknowledgement}
C. Zhai is partially supported by China Postdoctoral Science Foundation under Grant 2018M630014. Z. Zhang is partially supported by NSF of China under Grant 11425103.
|
\section{Introduction}
In this paper we determine the three-dimensional (3D) carrier-envelope-phase (CEP) map, or spatial variation of the CEP, of broadband, few-cycle pulses focused by mirrors or lenses, and describe several mechanisms to tailor it for specific applications. The CEP map is relevant to all nonlinear light-matter interaction experiments as high-harmonic or attosecond pulse generation, photoelectron emission from metal surfaces, nano-objects or structures; and in the control of chemical reactions in femtochemistry, where the CEP is a crucial parameter determining the outcome of the experiment \cite{Krausz2014NPhot,Baltuska2003Nature, Haworth2007NPhys, Ishii2014NC, Kruger2011Nature}. In gas high-order harmonic generation, for example, the XUV emission may even originate from a large portion of the focal volume, influencing its spectral content through phase matching
\cite{Ishii2014NC,Chini2014NPhot, Popmintchev2010NPhot, Rudawski2015EPJD}. Thus knowing and controlling the focal CEP variation, along with a consequent choice of the material sample position and extent, can be considered as relevant as controlling the CEP of the laser source.
It has happened slowly and with difficulty to stop identifying the variation of the CEP through a focus with Gouy's phase shift, probably because the former has its origin in the latter \cite{Tritschler2005OL, Yang2008COL}. Gouy's phase shift is an additional $-\pi$ phase shift of a monochromatic constituent through the focus that changes its phase velocity, while the CEP shift is the shift of the phase of the carrier oscillations at the time of pulse peak due to the different and space-dependent phase and group velocities \cite{Major2015JO}. The precise on-axis and off-axis CEP map of ideally focused few-cycle, transform-limited pulsed Gaussian beams was described in \cite{Porras2009OL}. The deviation from Gouy's phase shift is governed by the so-called factor $g$ of the input beam, also called Porras factor in \cite{Hoff2017NPhys}, which is determined by the variation of the beam width with frequency \cite{Porras2009OL}. On the experimental side, earlier measurements did not report conclusive deviations of CEP shift from Gouy's phase shift \cite{Lindner2004PRL, Tritschler2005OL, Wang2007OL}. Recent measurements based on spectral interferometry reported however strong deviations from Gouy's phase shift and their relation with the wavelength-dependent properties of the input beam \cite{Major2015AO}. Very recently, direct measurement of the on-axis and off-axis CEP map based on the CEP-sensitivity of the back-scattered photoelectrons from a nanotip scanning the focal volume \cite{Hoff2017NPhys} confirmed the strong deviations from the Gouy's phase shift and corroborated the predictions in \cite{Porras2009OL}, in particular, the predicted dependence with the $g$-factor. These observations underscore the need to properly characterize the laser source in use ---especially measuring its factor $g$--- to get advanced control on the CEP and hence on the allied CEP-sensitive light-matter interactions \cite{Hoff2017NPhys}. Unfortunately, there are no studies on how the factor $g$ determining the CEP map relates to the femtosecond laser source characteristics, so that no effective control can be exercised on the CEP under close-to-ideal focusing conditions as with spherical or parabolic mirrors.
Soon after the demonstration of CEP-sensitivity of different physical processes, methods have been proposed to attain some control on axial variations of the CEP based on dispersive material for specific applications \cite{Zapata-Rodriguez2008OE,Porras2009OE}.
We have previously studied the focusing of few-cycle pulses with lenses \cite{Porras2012APB,Porras2012JOSAB,MajorPhD} and determined the precise conditions for lenses to focus to transform-limited, few-cycle pulses traveling undistorted in the focal volume. We have also described how the on-axis CEP variation differs from that using mirrors as an effect of the small chromatic aberration introduced by the lens \cite{Porras2012APB}. Although mirrors are thought to be generally preferable, chromatic focusing has recently been proposed as an additional control knob for high-order harmonic generation \cite{Holgado2017PRA}. Also, we have previously found that a small pulse chirp, lengthening the pulse by, e. g., less that one femtosecond, strongly affects the on-axis CEP variation \cite{Porras2012JOSAB}. Chirped femtosecond pulses has also been used to control high-order harmonic generation \cite{Salieres1998PRL, Chang1998PRA, Lee2001PRL, Mauritssom2004PRA, Holgado2016PRA}, chemical reactions \cite{Goswami2009CP} and THz-generation \cite{Vidal2014JOSAB}. Chromatic aberration and chirp thus offer two control parameters, often used in femtosecond light-matter interaction experiments, to control the CEP map.
However, the capability of chromatic focusing and chirp to modify and control the CEP map has only been analyzed along the optical axis, and not off-axis, while nonlinear interactions generally take place in a volume, not just on-axis
\cite{Karimi2013OE, Krausz2014NPhot, Ishii2014NC,Hoff2017JOP}. It is thus of interest to these experiments to investigate how the complete 3D CEP map depends on $g$, chromatic aberration and chirp, so that specific axial and transversal CEP variations (e. g., no axial and/or radial CEP variation) required for particular CEP-sensitive interactions at specific target positions and extents can be tailored acting on these parameters.
Thus, in Sec. \ref{sec:3DCEPmap} we evaluate the complete 3D CEP map, recalling the conditions under which the few-cycle pulses can be focused without appreciable distortion. In Sec. \ref{sec:3DCEPmaptailor} we analyze the conditions under which the CEP map can be made flat, axially, transversally or both, in specific locations in the focal volume, with a focusing mirror or with a lens using small chirps and chromatic aberrations. In Sec. \ref{sec:doublet} we exemplify the results by showing these small pulse chirp and the tunable small chromatic aberration are easily provided by a separable achromatic doublet that would allow to generate these tailored CEP maps in practice.
\section{The three-dimensional CEP map}\label{sec:3DCEPmap}
If the beam to be focused is of sufficiently high quality, its spectrum can be assumed to be of the form of the collimated Gaussian beam $\hat E_{\rm in}(\omega,r)=\hat p_{\rm in}(\omega)\exp(-r^2/s_L^2)$, where $r$ is the radial coordinate from the optical axis, $\hat p_{\rm in}(\omega)$ is the on-axis spectrum, of bandwidth $\Delta\omega$ about the laser carrier frequency $\omega_0$. The beam spot size $s_L$ and the associated Rayleigh distance $L=\omega s_L^2/2c$ depend in general on frequency $\omega$. Since applications require careful focusing to a nearly diffraction-limited beam, we can assume that the spectrum after the focusing system is given, from well-known Gaussian beam formulas in the Debye approximation ($L\gg f$, or no appreciable focal shift), by $\hat E(\omega,r,z)=\hat p(\omega)(-f/q)\exp(i\omega r^2/2cq)\exp(i\omega z/c)$, where $z$ is the axial distance from the exit plane of focusing system, $q=Z-iL_R$ and $L_R=f^2/L$ are the complex beam parameter and Rayleigh distance of the focused beam, $f$ is the focal length, which may depend on frequency, $Z=z-f$ is an axial coordinate with origin at the focal point at each frequency, and $\hat p(\omega)$ may differ from $\hat p_{\rm in}(\omega)$ in the possible spectral changes introduced by the focusing system. The space-dependent parts of the spectral amplitude and phase at each point after the focusing system are then given by
\begin{eqnarray}
a(\omega,r,z)&=& \frac{f}{L_R}\frac{1}{[1+(Z/L_R)^2]^{1/2}}\exp[-r^2/s^2(Z)]\, , \\
\varphi(\omega,r,z)&=& -\frac{\pi}{2}- \tan^{-1}\left(\frac{Z}{L_R}\right) + \frac{\omega r^2}{2cR(Z)} + \frac{\omega}{c}z
\end{eqnarray}
where $s^2(Z)=s_f^2[1+(Z/L_R)^2]^{1/2}$, $s_f^2= 2cL_R/\omega$, $R(Z)=Z[1+ (L_R/Z)^2]$.
It has been shown \cite{Porras2012APB} that the above Gaussian-beam formulas describe accurately the spectrum of the focused pulsed beam if $s_L$ is small enough compared to the mirror or lens radius, for aperture effects and spherical aberration to be negligible. Under this condition, and if the focusing system is a lens, pulse broadening introduced by the lens material dispersion is substantially the same as that introduced by a slab of the lens material and thickness equal to the lens central thickness $D$, i. e., $\hat p(\omega) = \hat p_{\rm in}(\omega)\exp[i D n(\omega)]$, where $n(\omega)$ is the lens material refractive index \cite{Porras2012APB}. It is then possible to have a transform-limited, few-cycle pulse after the lens by simply pre-compensating for the dispersion introduced by that plate, as usually done with standard pulse-shaping techniques (e. g., GVD and/or TOD compensation, depending on the pulse duration and lens thickness $D$). We may also have a pulse with positive/negative chirp after the lens by under/over-compensating that material dispersion. It has also been shown \cite{Porras2012APB} that the distortions in the pulse shape about the focus originating from the lens chromatic aberration, measured by the parameter $\gamma= \omega (df/d\omega)/L_R$ (evaluated at the carrier frequency), are negligible if $|\gamma|\ll \omega_0/\Delta \omega$, a condition that is easily given with real lenses. Under these conditions, the only source of distortion in the pulse shape is the intrinsic distortion due to the transverse limitation of a focused beam, i. e., to diffraction, distortions that are generally very small, and have been shown to be accurately described by the first-order theory of diffraction effects in few-cycle pulses \cite{Porras2002PRE}. In this theory, the time-domain electric field, or inverse Fourier transform of $\hat E(\omega,r,z)$, is expressed as the enveloped carrier oscillations $E(t,r,z)=A(t,r,z)\exp\{-i[\omega_0 t-\varphi(\omega_0,r,z)]\}$, with an envelope given by
\begin{equation}\label{A}
A(t,r,z)\simeq a(\omega_0,r,z)p(\tau) + i a'(\omega_0,r,z)\frac{dp(\tau)}{d\tau}\, .
\end{equation}
Prime signs stand for derivative with respect to frequency, $p(\tau)$ is the inverse Fourier transform of $\hat p(\omega)$ evaluated at the local time $\tau= t-\varphi'(\omega_0,r,z)$. The second term in the rhs of Eq. (\ref{A}) accounts for the small diffraction-induced changes in the pulse shape about the focus.
The CEP at a point $(r,z)$ is the phase of the carrier oscillations at the time of pulse peak, i. e., $\Phi(r,z)=-\omega_0[\tau(r,z)+\varphi'(\omega_0,r,z)]+ \varphi(\omega_0,r,z)+ \phi(r,z)$, where $\tau(r,z)$ is the time at which $|A(\tau,r,z)|$ peaks and $\phi(r,z)$ the argument of $A(\tau,r,z)$ at that time. Thus, if we take the focal point at the carrier frequency ($r=0, z=f(\omega_0)\equiv f_0$) (focal point for short) as a reference point, the CEP shift from the focal point is given by $\Delta\Phi(r,z)=\Phi(r,z)-\Phi(0,f_0)$, which can be separated in the two contributions
\begin{eqnarray}\label{PHI1}
\Delta\Phi_1(r,z)&=& [-\omega_0\varphi'(\omega_0,r,z)+\varphi(\omega_0,r,z)]\nonumber \\ &-&[-\omega_0\varphi'(\omega_0,0,f_0)+\varphi(\omega_0,0,f_0)]\,,
\end{eqnarray}
and
\begin{eqnarray}\label{PHI2}
\Delta\Phi_2(r,z)&=& -\omega_0[\tau(r,z)-\tau(0,f_0)] + \phi(r,z) - \phi (0,f_0) \nonumber \\
&\equiv& -\omega_0 \Delta\tau(r,z) + \Delta \phi(r,z)\, .
\end{eqnarray}
The first contribution is independent of the shape $p(\tau)$ of the pulse, the only one considered in \cite{Porras2009OL} for focusing without chromatic aberration, and has been verified experimentally to describe the actual CEP map with transform-limited (unchirped Gaussian) pulses \cite{Hoff2017NPhys} (which implies that the second contribution is negligible with transform-limited pulses). The effect of chromatic aberration on the CEP was considered in \cite{Porras2012APB}, but limited to on-axis points ($r=0$). The second contribution includes the effect of the small pulse reshaping, and depends on the pulse shape itself. For on-axis points only, it has been shown that any small amount of chirp in the focused pulse drastically alters the CEP map \cite{Porras2012JOSAB}. Here we evaluate the two contributions for off-axis points with chirped or unchirped Gaussian pulses, providing thus a complete description of the CEP map about the focus.
Evaluation of Eq. (\ref{PHI1}) is a straightforward but lengthy (and tedious) exercise of algebra and calculation of derivatives, of which we summarize the most relevant points. With the normalized variables $\zeta=Z/L_R$ and $\rho=r/s(Z)$ the spectral phase reads as $\varphi=-\pi/2-\tan^{-1}(\zeta) + \rho^2\zeta + (\omega/c)z$. Both $\zeta$ and $\rho$ depend on $\omega$ because $f$, and $L$, and hence $L_R$, $Z=z-f$ and all other quantities in the Gaussian-beam formulas where they appear, depend on $\omega$. Careful evaluation of their derivatives yield $\zeta'=-(f'/L_R)- \zeta (L'_R/L_R)$ and $(\rho^{2})^{\prime}=\rho^2[1/\omega-(L'_R/L_R)-2\zeta\zeta'/(1+\zeta^2)]$. After long algebra, the first contribution to the CEP shift from the focal point is obtained to be
\begin{eqnarray}\label{CEPSHIFT1}
\Delta\Phi_1(r,z) &=& -\tan^{-1}\left(\frac{Z}{L_R}\right) \nonumber \\
&+& \frac{1-2\frac{r^2}{s^2(Z)}}{1+\left(\frac{Z}{L_R}\right)^2} \left[g\left(\frac{Z}{L_R}\right)+ \gamma \left(\frac{Z}{L_R}\right)^2 \right] \nonumber \\
&+& \gamma \frac{r^2}{s^2(Z)} ,
\end{eqnarray}
where
\begin{equation} \label{eq:g_gamma}
g= -\frac{L'_R}{L_R}\omega\,, \quad \gamma= \frac{f'}{L_R}\omega \,,
\end{equation}
and where {\it all frequency-dependent quantities are evaluated at the carrier frequency $\omega_0$,} e. g., $Z=z-f_0$ is the axial distance from the focal point.
Similarly, the spectral amplitude, expressed in the dimensionless variables, is given by $a= (f/L_R) \exp(-\rho^2)/(1+\zeta^2)^{1/2}$. After some algebra, its derivative with respect to frequency can be written as $a'=h(\rho,\zeta)a$, where
\begin{equation}
h(\rho,\zeta)=h(\zeta)-\rho^2\left(\frac{1}{\omega}- \frac{L'_R}{L_R}- \frac{2\zeta\zeta'}{1+\zeta^2}\right)\, ,
\end{equation}
\begin{equation}
h(\zeta)= \frac{f'}{f}\left(1+ \frac{f}{L_R}\frac{\zeta}{1+\zeta^2}\right) - \frac{L'_R}{L_R}\left(1- \frac{\zeta^2}{1+\zeta^2}\right) \, ,
\end{equation}
and where all frequency-dependent quantities are evaluated at the carrier frequency.
The complex envelope in Eq. (\ref{A}) then yields $A\simeq\left[p(\tau)+ i h(\rho,\zeta)dp/d\tau\right]a$, which, being a first-order approximation (second order derivatives are neglected), is conveniently replaced with $A \simeq p\left[\tau+ih(\rho,\zeta)\right]a$, having the same first-order approximation. Assuming that the on-axis pulse immediately after the lens is the Gaussian pulse $p(\tau)=(\Delta T/b)\exp(-\tau^2/b^2)$, where $b^2=\Delta T^2 -2iC$, $C$ is a residual chirp, and $\Delta T$ is the transform-limited duration (actual duration $\Delta T_C =\Delta T[1 + (2C/\Delta T^2)^2]^{1/2}$), the reshaped real amplitude's temporal shape in the focal region is found to be
\begin{equation}
|A|\simeq \frac{\Delta T}{b}\exp\left[\frac{h^2(\rho,\zeta)}{\Delta T^2}\right]\exp\left\{-\frac{\left[\tau- \left(\frac{2C}{\Delta T^2}\right)h(\rho,\zeta)\right]^2}{\Delta T_C^2}\right\}.
\end{equation}
Thus, reshaping due to the strong localization in the focal region consists of a drift of the time of pulse peak from point to point of space given by $\tau(\rho,\zeta)= (2C/\Delta T^2)h(\rho,\zeta)$. This expression allows to evaluate $-\omega_0[\tau(r,z)-\tau(0,f_0)]= -\omega_0(2C/\Delta T^2)[h(\rho,\zeta)-h(0,0)]$ in Eq. (\ref{PHI2}) for the second contribution to the CEP shift from the focal point. Similar calculations allow to conclude that the difference of phases of the envelopes at the drifted peaks, $\phi(r,z)-\phi(0,f_0)$, does not give a significant contribution to the CEP shift. After some algebra, Eq. (\ref{PHI2}) for the second contribution to the CEP shift then yields
\begin{eqnarray}\label{CEPSHIFT2}
\Delta\Phi_2(r,z) &=& \frac{2C}{\Delta T^2}\frac{1-2\frac{r^2}{s^2(Z)}}{1+\left(\frac{Z}{L_R}\right)^2} \left[-\gamma\left(\frac{Z}{L_R}\right)+ g \left(\frac{Z}{L_R}\right)^2 \right] \nonumber \\
&+& \frac{2C}{\Delta T^2} (1+g) \frac{r^2}{s^2(Z)}\,,
\end{eqnarray}
where again all quantities are evaluated at the carrier frequency. The total CEP shift form the focal point is finally obtained to be
\begin{eqnarray}\label{CEPSHIFT}
\Delta\Phi(r,z) &=& -\tan^{-1}\left(\frac{Z}{L_R}\right) \nonumber \\
&+& \frac{1-2\frac{r^2}{s^2(Z)}}{1+\left(\frac{Z}{L_R}\right)^2} \left[G\left(\frac{Z}{L_R}\right)+ \Gamma \left(\frac{Z}{L_R}\right)^2 \right] \nonumber \\
&+& \left[\Gamma + \frac{2C}{\Delta T^2}\right] \frac{r^2}{s^2(Z)} ,
\end{eqnarray}
where
\begin{equation}\label{GGAMMA}
G\equiv g-\gamma\frac{2C}{\Delta T^2}\,, \quad \Gamma\equiv \gamma+ g\frac{2C}{\Delta T^2}\,.
\end{equation}
These equations provide the complete 3D CEP map for the focusing conditions of interest in many applications to a nearly diffraction-limited and transform-limited pulsed beam. The most relevant novelty compared to the on-axis formula is the term in the third row that makes the CEP at the focal plane to present, in general, a quadratic variation with radial distance. Out of the focal plane, the CEP has also a positive or negative quadratic variation, depending on the specific value of $Z$. Along caustic surfaces $r/s(z)=\mbox{const}$, the CEP variation is more pronounced or less pronounced than Gouy's phase, depending on the value of $g$ of the input beam and the specific caustic surface, and as described elsewhere \cite{Porras2009OL,Porras2012JOSAB,Hoff2017NPhys}. The CEP shift along the caustic surface $r/s(Z)=1/\sqrt{2}$ always equals to Gouy's phase shift. For further development, we rewrite (\ref{CEPSHIFT}) more compactly as
\begin{equation}\label{CEPSHIFTC}
\Delta \Phi = -\tan^{-1}\zeta + \frac{\zeta G + \zeta^2 \Gamma}{1+\zeta^2}(1- 2\rho^2) + P \rho^2,
\end{equation}
where
\begin{equation}\label{P}
P=\Gamma + 2C/\Delta T^2=\gamma + (g+1)\frac{2C}{\Delta T^2}.
\end{equation}
\section{Tailoring the three-dimensional CEP map}\label{sec:3DCEPmaptailor}
The CEP map can be used and manipulated in a number of ways, depending on the particular application. For example, it is generally desirable to have a constant CEP in the volume of the target to avoid CEP-integration effects that may wash out the sensitivity to the CEP of the light-matter interaction. One may then evaluate the CEP standard deviation in the target volume and minimize it. In the following we assume a laser source of certain factor $g$ (that cannot easily be modified), a target at position $Z$ of narrow axial thickness compared to $L_R$, and large transversal size compared to $s(Z)$. We can also consider an light-matter interaction that depends on a certain power $n$ of the intensity $I^n\propto \exp[-2n r^2/s^2(Z)]= \exp(-2n\rho^2)$. This is an alternative way to consider a transversally limited target of effective radius $s(Z)/\sqrt{n}$ smaller than the beam radius, e. g., negligible transversal size for $n\rightarrow \infty$. The effective CEP in the thin sample, or average CEP with $I^n$ is found to be
\begin{equation}\label{EFFCEP}
\langle\Delta\Phi\rangle = -\tan^{-1}\zeta + \frac{\zeta G + \zeta^2 \Gamma}{1+\zeta^2}\left(1- \frac{1}{n}\right) + \frac{P}{2n}\,,
\end{equation}
with a transversal standard deviation given by
\begin{equation}\label{SV}
\langle (\Delta\Phi - \langle\Delta\Phi\rangle)^2\rangle^{1/2} = \frac{1}{n}\left|\frac{G\zeta+\Gamma \zeta^2}{1+\zeta^2}- \frac{P}{2} \right|.
\end{equation}
For a given sample position $\zeta$, we may wish to have no transversal variation of the CEP, no local axial variation of the CEP, or both, acting on the two control parameters at hand, namely, pulse chirp and lens chromatic aberration. No transversal variation requires
\begin{equation}\label{NOR}
2\zeta G + 2\zeta^2\Gamma - (1+\zeta^2)P=0\,,
\end{equation}
locally vanishing axial variation of the effective CEP at position $\zeta$, and in particular vanishing on-axis CEP variation for $n\rightarrow\infty$, requires
\begin{equation}\label{NOA}
-(1+\zeta^2)+ (1-\zeta^2)G\left(1-\frac{1}{n}\right) + 2\zeta\Gamma\left(1- \frac{1}{n}\right)=0\,,
\end{equation}
as obtained by equating to zero the derivative of Eq. (\ref{EFFCEP}) with respect to $\zeta$.
\subsection{Design of the CEP map with focusing mirrors}
Since in most of experiments mirrors are used to focus few-cycle pulses, we first analyze if the above conditions can be given with a mirror, in which case $\gamma=0$. Then $G=g$, $\Gamma = g(2C/\Delta T^2)$ and $P=(1+g)(2C/\Delta T^2)$. From Eq. (\ref{NOA}), no axial CEP variation at the focus $\zeta=0$ occurs, irrespective of the chirp, only with specific value $g=1/(1-1/n)$ of the input pulse, or $g=1$ for the on-axis CEP, and therefore it is not generally possible. From Eq. (\ref{NOR}), the CEP across the focal plane is constant only if $P=(g+1)(2C/\Delta T^2) = 0$, i. e., requires a transform-limited pulse. Any chirp induces a CEP variation across the focal plane. Axially and radially constant CEP occurs only given with input laser pulses with $g=1$ and $C=0$.
In many experiments the target (e. g., a gas nozzle) is placed slightly out of focus, preferably in its second half. Equation (\ref{NOR}) for transversally flat CEP at $0<|\zeta|<1$ in the case of a mirror yields the needed relative pulse chirp
\begin{equation}\label{NORM}
\frac{2C}{\Delta T^2} = \frac{2\zeta g}{g(1-\zeta^2)+(1+\zeta^2)}\,.
\end{equation}
Figure \ref{Fig1}(a) and (b) are examples of CEP maps in the focal volume of input focused pulses with $g=-0.5$, having radially flat CEP at $\zeta=0$ with transform-limited pulses ($C=0$), and radially flat CEP at the middle of the second half of the focal region with slightly chirped pulses ($2C/\Delta T^2 =-0.571$). Figure \ref{Fig1}(c) shows the corresponding on-axis axial variations, and FIG. \ref{Fig1}(d) the relative chirp $2C/\Delta T^2$ to have transversally flat CEP at different positions $\zeta_{\rm frozen}$ in the second half of the focal region for a few (supposedly) typical values of the factor $g$ of the input beam. The relative chirp providing radially flat CEP in the first half of the focal region has opposite sign. As seen, chirping the pulse could be useful in practice (does not entail important pulse broadening) to have transversally flat CEP at a desired axial position in the focal region for input pulses with $g>-1$.
\begin{figure}[t]
\includegraphics[width=8cm]{Fig1ab.eps}\\
\includegraphics*[width=4.3cm]{Fig1c.eps}\includegraphics*[width=4.3cm]{Fig1d.eps}
\caption{\label{Fig1} For spherical mirror focusing, focal CEP maps (cut along $y=0$ plane) for input beam with $g=-0.5$, (a) transform-limited pulse, $2C/\Delta T^2=0$, and (b) chirped pulse, $2C/\Delta T^2=-0.571$. The CEP is transversally flat at the focal plane in (a) and at $Z=L_R/2$ in (b). (c) Axial variation of the CEP in the two above cases (solid curves) compared to Gouy's phase shift (dashed curve). (d) Needed relative chirp as a function of the position $Z_{\rm frozen}$ at which the CEP is constant transversally, for a few values of $g$ of the input pulse.}
\end{figure}
We instead may wish to have a CEP with no axial variation at a certain position $0<|\zeta|<1$ in the focal volume. Condition (\ref{NOA}) yields the chirp
\begin{equation}\label{NOAM}
\frac{2C}{\Delta T^2}=\frac{1+\zeta^2- (1-\zeta^2)g\left(1-\frac{1}{n}\right)}{2\zeta g \left(1-\frac{1}{n}\right)}\,.
\end{equation}
Unfortunately, the chirp values are considerably large for most of values of $g$. Only for $g>1$, reasonably small chirp produces locally constant CEP within the focal region, as shown in FIG. \ref{Fig2}(a) for the effective CEP with $n=4$ and in the second half of the focal region with positive chirps (opposite chirps produce the same effect in the first half).
Of particular interest is the situation in which the CEP is constant transversally and axially. Equating (\ref{NORM}) and (\ref{NOAM}), this effect is seen to be possible in the focal region only with an input beam with $g\ge 1.15$ at a certain position $\zeta$ in $[-1,1]$ (where the curves in FIG. \ref{Fig1}(d) and FIG. \ref{Fig2}(a) intersect) that depends on the particular value of $g$. This relevant position in the second half of the focus and the required chirp are depicted in FIG. \ref{Fig2}(b) as functions of $g$. The curves ends at $g=1/(1-1/n)$ at focus with no chirp. For example, with an input pulse with $g=1.26$, the chirp $2C/\Delta T^2=0.574$ leads to the axially frozen effective CEP with $n=4$ at $\zeta=0.5$, as seen in FIG. \ref{Fig2}(c). The transversally and axially flat CEP map at $\zeta=0.5$ is seen in FIG. \ref{Fig2}(d).
\begin{figure}[t!]
\includegraphics[width=4.3cm]{Fig2a.eps}\includegraphics*[width=4.4cm]{Fig2b.eps}
\includegraphics[width=4.1cm]{Fig2c.eps}\includegraphics*[width=4.4cm]{Fig2d.eps}
\caption{\label{Fig2} (a) Needed relative chirp as a function of the position $Z_{\rm frozen}$ at which the effective CEP with $n=4$ is locally constant axially, for a few values of $g$ of the input beam for which this chirp is relatively low. (b) Needed relative chirp to freeze the effective CEP with $n=4$ both axially and transversally, and the position within the focus at which this happens, as functions of the values of $g$ for which this effect is possible. (c) On-axis and effective CEP with $n=4$ (solid curves) for $g=1.26$ and $2C/\Delta T^2=0.574$ and Gouy's phase (dashed curve), showing locally constant effective CEP at $Z=L_R/2$. (d) CEP map in the focal region for $g=1.26$ and $2C/\Delta T^2=0.574$, showing also transversally constant CEP at $Z=L_R/2$.}
\end{figure}
These results underscore the need to characterize the femtosecond laser source in use by measuring its factor $g$. By doing this we will not only know the possible CEP maps about the focus of the mirror, but will also know if the CEP map can be adapted to particular applications using small amounts of chirp as a control knob, or suitably positioning the sample according to the CEP maps.
\subsection{Design of the CEP map with lenses}
\begin{figure}[b!]
\includegraphics*[width=4.3cm]{Fig3a.eps}\includegraphics*[width=4.3cm]{Fig3b.eps}
\includegraphics*[width=4.3cm]{Fig3c.eps}\includegraphics*[width=4.3cm]{Fig3d.eps}
\caption{\label{Fig3} (a) Chromatic aberration parameter $\gamma$ as a function of the relative chirp $2C/\Delta T^2$ (a) to freeze axially the on-axis CEP at the focal point for different values of $g$ of the input pulse. (b) The same but to have transversally flat CEP at the focal plane. (c,d) Chromatic aberration and chirp to produce on-axis and transversally frozen CEP at focus. Positive chromatic aberration requires negative chirp (solid curves), and vice versa (dashed curves).}
\end{figure}
\begin{figure}[t!]
\includegraphics*[width=8cm]{Fig4abcd.eps}
\caption{\label{Fig4} CEP maps in the focal volume ($y=0$ sections) for $g=0$ and the indicated values of $\gamma$ and $2C/\Delta T^2$. The CEP at $\zeta=0$ is axially constant in (a), transversally constant in (b), axially and transversally constant in (c) and (d), with a maximum in (c) and a minimum in (d).}
\end{figure}
The use of a lens offers more possibilities. According to Eq. (\ref{NOA}), we can have a CEP map with no axial variation at focus with input pulses of any value of $g$ if the chromatic aberration $\gamma$ and the relative chirp $2C/\Delta T^2$ verify
\begin{equation}\label{NOAL}
\gamma \frac{2C}{\Delta T^2} = g - m\,,
\end{equation}
where $m= (1-1/n)^{-1}$. Figure \ref{Fig3}(a) shows the $\gamma-2C/\Delta T^2$ curves determined by Eq. (\ref{NOAL}), limited to relative chirps $|2C/\Delta T^2|<1$ so as not to enlarge excessively the pulse, for different values of its $g$-parameter in the particular case of $n\rightarrow\infty$ ($m=1$), i .e., to freeze the on-axis CEP variation at focus. For other values of $n$ the curves are qualitatively similar. As an example, the particular CEP map for an input pulse with $g=0$ having an on-axis maximum at $\zeta=0$ using the relative chirp $2C/\Delta T^2=0.5$ and the chromatic aberration $\gamma=-2$ is plotted in FIG. \ref{Fig4}(a). With chirp and chromatic aberration of opposite signs, the CEP map would feature a minimum.
Instead, we may wish to have the same CEP at all points of the focal plane. According to Eq. (\ref{NOR}), transversally constant CEP imposes the linear relation
\begin{equation}\label{NORL}
\gamma = -(g+1) \frac{2C}{\Delta T^2}\,.
\end{equation}
The $\gamma-2C/\Delta T^2$ lines determined by Eq. (\ref{NORL}) are depicted in FIG. \ref{Fig3}(b) for the same values of the $g$ parameter as in FIG. \ref{Fig3}(a). The particular CEP map with flat CEP at the focal plane for an input pulse with $g=0$ using relative chirp $2C/\Delta T^2=0.5$ and $\gamma=-0.5$ is plotted in FIG. \ref{Fig4}(b) for illustration purposes. Chirp and chromatic aberration of opposite signs would also produce flat CEP at the focal plane.
While axially and transversally flat CEP at focus with a focusing mirror requires a transform-limited input pulse with a particular value of $g$, with a lens this particularly relevant CEP map is possible with a variety of values of $g$. From Eqs. (\ref{NOAL}) and (\ref{NORL}) we obtain the needed chirp and chromatic aberration as
\begin{equation}\label{FF}
\frac{2C}{\Delta T^2} = \pm \sqrt{\frac{m-g}{1+g}}\,, \quad \gamma =\mp \sqrt{(1+g)(m-g)}\,.
\end{equation}
These values are the points of intersection of the curves in Figures \ref{Fig3}(a) and \ref{Fig3}(b) for each specific value of $g$, intersection points that exist for input pulses with $-1\le g\le m$, and are represented in FIG. \ref{Fig3}(c) and \ref{Fig3}(d) as functions of $g$ in the particular case of $n\rightarrow\infty$ ($m=1$). Positive and negative chromatic aberrations correspond to negative and positive chirp in Eq. (\ref{FF}), or solid and dashed curves in Figures \ref{Fig3}(c) and \ref{Fig3}(d). Thus, given the factor $g$ of the input pulse, there are two specific values of chromatic aberration and chirp to tailor an axially and transversally flat CEP map at focus. Only for $0\le g\le 1$ [black curves in Figures \ref{Fig3}(a) and \ref{Fig3}(b)], however, the chirp can be to be considered small enough according to our criterion of $|2C/\Delta T^2|<1$.
As a couple of examples, the CEP maps corresponding to input pulses characterized by $g=0$ with $\gamma=+1, 2C/\Delta T^2=-1$ (with a local axial maximum at $\zeta=0$) and with $\gamma=-1, 2C/\Delta T^2+1$ (with a local minimum) are plotted in Figures \ref{Fig4}(c) and \ref{Fig4}(d).
Also, it is generally possible to freeze the CEP at particular locations out-of-focus. The required chromatic aberrations can be straightforwardly obtained from Eqs. (\ref{NOR}) and (\ref{NOA}), but the obtained expressions for $\gamma$ and $2C/\Delta T^2$ are long and cumbersome. In Ref. \cite{Porras2012JOSAB}, it has been shown that irrespective of the value of $g$, a small chromatic aberration $\gamma$ in $[-1,0]$ or in $[0,1]$ flattens the axial CEP variation in the first or second half of the focus, respectively, and a small chirp enhances further the flatness of the axial CEP variation. Following this criterion, it is easy in practice, and faster than deriving particular expressions, to adopt a inspection procedure by directly plotting the 3D CEP map given by Eq. (\ref{CEPSHIFTC}) with the given value of $g$, and different values values of $\gamma$ and $2C/\Delta T^2$. We have found that for optimum values of $\gamma$ and $2C/\Delta^2$ to flatten the axial CEP variation in one half of the focus, the transversal CEP variation at about the middle of that half vanishes. This result is illustrated in FIG. \ref{Fig5} for focused input pulses with different values of $g$, including negative ones. The vertical dashed lines in these figures indicate the positions about the middle of the first half of the focus where the CEP is axially and transversally flat. For each $g$, chromatic aberrations and chirp of opposite signs would equally flatten axially and transversally the CEP in the second half of the focus.
\begin{figure*}[t!]
\includegraphics*[width=3.6cm]{Fig5a.eps}\includegraphics*[width=3.6cm]{Fig5c.eps}\includegraphics*[width=3.6cm]{Fig5e.eps}\includegraphics*[width=3.6cm]{Fig5g.eps}\includegraphics*[width=3.6cm]{Fig5i.eps}
\includegraphics*[width=3.6cm]{Fig5b.eps}\includegraphics*[width=3.6cm]{Fig5d.eps}\includegraphics*[width=3.6cm]{Fig5f.eps}\includegraphics*[width=3.6cm]{Fig5h.eps}\includegraphics*[width=3.6cm]{Fig5j.eps}
\caption{\label{Fig5} Upper plots: On-axis ($n\rightarrow\infty$) CEP shift for input pulses with the indicated values of $g$, focused by a lens with the indicated values of chromatic aberration $\gamma$ and chirp $2C/\Delta T^2$ such that the CEP is approximately flat in the first half of the focus (red curves). The dashed curves represent Gouy's phase shift for reference, and the black curves the CEP shift in case of focusing without chromatic aberration and without chirp. Lower plots: Corresponding CEP maps ($y=0$ sections) showing that the CEP is also transversally flat at positions about $-L_R/2$, as indicated by the vertical dashed lines in the upper plots.}
\end{figure*}
\section{Implementation of tunable chromatic aberration for CEP map control}\label{sec:doublet}
In addition to small chirps $|2C/\Delta T^2|\le 1$, we propose a simple "separable" close-to-achromatic doublet to easily tune the chromatic aberration parameter $\gamma$, and hence to focus few-cycle pulses with tailored CEP map. Substantial chromatic aberration can severely distort the pulse \cite{Bor1992OC}, but condition $|\gamma|< \omega_0/\Delta\omega$ ensures negligible pulse distortion, and according to the above analysis $|\gamma|\le 1$ generally suffices to the purpose of CEP map design.
For two thin lenses of focal lengths $f_1$ and $f_2$ separated a distance $d$, the focal length is
\begin{equation}\label{eq:ftot}
\frac{1}{f} = \frac{1}{f_{1}} + \frac{1}{f_{2}} - \frac{d}{f_{1} f_{2}}\,.
\end{equation}
Lensmaker's formula for thin lenses and Eq. (\ref{eq:g_gamma}) yields
\begin{equation} \label{eq:gamma_lenssystem}
\gamma = -L \left[ \frac{a}{f_{1}} + \frac{b}{f_{2}} - \frac{a+b}{f_{1} f_{2}} d \right] = \gamma(0) + L \frac{a+b}{f_1f_2}d\,,
\end{equation}
where $a = n'_{1} \omega/(n_{1} -1)$ and $b = n'_{2} \omega/(n_{2} -1)$, $n_{1,2}$ are the lens refractive indexes, $L=\omega s^2/2c=f^2/L_R$ is the Rayleigh distance of the input pulse, and, as above, all quantities and their derivatives are evaluated at $\omega_0$. In the second equality $\gamma(0)=-L(a/f_1 + b/f_2)$ is the chromatic aberration parameter when the two lenses are joint. Thus, chromatic aberration varies linearly with increasing separation $d$, giving an easy way to control it and hence the CEP map. Since usually achromatic doublets consist of a focusing and a defocusing element, meaning $f_{1} f_{2} < 0$, $\gamma$ decreases with lens separation.
Suppose we wish to focus the input pulse of the Rayleigh distance $L$ to a half-focal depth $L_R$, as required by a particular experimental setup, and with chromatic aberration $\gamma(0)$ when the two lenses are joint. Setting the required focal length $\sqrt{LL_R}$ in Eq. (\ref{eq:ftot}) with $d=0$ and from $\gamma(0)=-L(a/f_1 + b/f_2)$, we obtain the required focal lengths as
\begin{equation}\label{eq:ad_gorig}
\frac{1}{f_{1}} = \frac{\gamma(0)/L + b/\sqrt{L L_{R}}}{b-a}\,, \quad \frac{1}{f_{2}} = \frac{-\gamma(0)/L - a/\sqrt{L L_{R}}}{b-a}\,.
\end{equation}
By slightly separating the two lenses, the chromatic aberration $\gamma$ can be tuned following the second equality in Eq. (\ref{eq:gamma_lenssystem}), and if the system is designed such that the required separations $d$ for efficient $\gamma$ tuning are minimal compared to $f_1$ and $f_2$, the variation of the total focal length $f$ is also minimal, and the variation of the focused Rayleigh distance $L_R= f^2/L$ is also very small.
As an example of design, we consider pulses at 800 nm carrier wavelength, input spot size $s_L=17$ mm (20 mm FWHM in intensity, and $L=1.135\times 10^6$ mm) to be focused to $L_R=3$ mm. Chromatic aberration is wished to be tuned from $\gamma=+1$ to $\gamma=-1$. With a N-BAF10 lens and a SF10 lens, Eqs. (\ref{eq:ad_gorig}) with $\gamma(0)=+1$ give $f_{1} = 589.2\,\mathrm{mm}$ and $f_{2} = -867.9\,\mathrm{mm}$, and the total focal length with zero separation is $f=184.5$ cm. From Eq. (\ref{eq:gamma_lenssystem}), $\gamma$ varies linearly ranges from $+1$ to $-1$ when $d$ ranges from $d=0$ to $d=10$ mm. At the same time, the total focal length obtained from Eq. (\ref{eq:ftot}) experiences a small variation from $184.5\,\mathrm{cm}$ to $178.1\,\mathrm{cm}$, corresponding to a negligible decrease of the focused Rayleigh range $L_{R}$ from $3\,\mathrm{mm}$ to $2.8\,\mathrm{mm}$. Due to the small change of the focal length the doublet can be placed and moved on a linear stage, so the target remains at the same position in the focal volume. This means that the proposed doublet system, along with small chirps of the focused pulse, allows one to study light-matter interactions with practically the same parameters except the CEP map, that can be tuned to meet the needs of the specific experiment.
\begin{figure}[tb!]
\includegraphics*[width=4.3cm]{Fig6a.eps}\includegraphics*[width=4.3cm]{Fig6b.eps}
\caption{\label{Fig6} (a) On-axis CEP shift about the focus for two different lens separation distances evaluated from Eq. (\ref{CEPSHIFTC}) (curves) and from the numerical simulation described in the text (symbols). (b) The same but at the focal plane.}
\end{figure}
We have confirmed numerically the usefulness of the above system for focusing without distortion few-cycle pulses. Similarly to previous works \cite{Porras2012APB, Porras2012JOSAB, MajorPhD}, our numerical calculations use a ray-tracing algorithm to analyze propagation from the input to the output plane of the focusing optics, while the electric field strengths in specific points in the focal volume are evaluated from scalar diffraction theory. The model takes into account the truncation of the beam by the aperture of the optics, and considers the lens varying thickness along with the resulting spherical and chromatic aberrations \cite{MajorPhD}. The above system is realized with central lens thicknesses $D_1=3.5$ mm, $D_2=2.0$ mm, $3$ inch diameter, radii of curvature $120\,\mathrm{cm}$ and $58\,\mathrm{cm}$ for the first lens, and $-58\,\mathrm{cm}$ and $881\,\mathrm{cm}$ for the second lens, making it possible zero separation (negative value means a concave surface). In the ray-tracing calculations, $d$ is assimilated to the distance between the back and front surfaces of the two lenses. The input pulsed beam at $800$ nm carrier wavelength is Gaussian temporally and transversally, of duration $\Delta T= 8.49$ fs ($10$ fs FWHM), width $s_L=17$ mm ($20$ mm FWHM), and $g$ factor equal to $0.5$. With precompensation of the second and third order dispersion introduced by the central thicknesses $D_1$ and $D_2$, the pulse after the lenses is Gaussian and transform-limited to all practical purposes and of the original duration at all points of the focal volume, similarly to the examples of previous works \cite{Porras2012APB, Porras2012JOSAB, Major2015AO, MajorPhD} where the usefulness of lenses to focus few-cycle pulses is stressed. The symbols in Figures \ref{Fig6}(a) and \ref{Fig6}(b) represent the CEP shift along the beam axis and at along the focal plane, respectively, extracted from the numerically evaluated electric fields when the lens separation is $d=0$ mm and $d=10$, and the solid curves represent the predictions of Eq. (\ref{CEPSHIFT}) with $g=0.5$, $C=0$ and $\gamma=\pm 1$, demonstrating also the usefulness of the separable doublet system for CEP map tailoring. Additional calculations confirm applicability of the system with pulse durations down to $5\,\mathrm{fs}$ (two-cycle pulse). With shorter pulses, precompensation of higher order dispersion terms than the second and third might be necessary to retain the pulse shape.
\section{Conclusions}
We have completed partial calculations provided by different authors in previous works in order to provide the complete spatial distribution of the CEP of few-cycle, Gaussian-Gaussian pulsed beams focused by mirrors or lenses. This amounts to specify the complete electric field under the envelope at each position of the focal volume. Our results underscore the importance of characterizing the few-cycle laser source in use by measuring its factor $g$, since it is the fundamental parameter that specifies the CEP map in the focal volume. Once $g$ is determined, small residual chirps and small chromatic aberration can be used to modify and adapt the CEP map for specific applications or target positions without appreciably deteriorating the pulse. We have shown, for example, how to suppress the CEP variation both longitudinally and transversally at different locations of the focal volume. We have also proposed and tested numerically a quasi-achromatic doublet that can perform this job without distorting the few-cycle pulse, and only changing the CEP map. This analysis, or others more adapted to each particular situation, can help to reach conditions that are beneficial in research areas where light-matter interactions strongly depend on the CEP, such as high-harmonic generation, attosecond pulse production, ultrafast nanooptics or plasmonics. As possible extensions of this work, it would be of interest to remove the restriction to Gaussian beams in order to evaluate the CEP map of other focused beams in use in these phase-sensitive light-matter interactions, e. g., of vortex beams \cite{Gauthier2017Nature} or radially/azimuthally polarized beams \cite{Biss2003OL}.
\section*{Acknowledgments}
Projects of the Spanish Ministerio de Econom\'{\i}a y Competitividad No. MTM2015-63914-P and No. FIS2017-87360-P. The ELI-ALPS project (GINOP-2.3.6-15-2015-00001) is supported by the European Union and co-financed by the European Regional Development Fund.
|
\section{Introduction}
The upsurge of wireless applications, including Internet-of-Things (IoT), Tactile Internet, tele-medicine and mobile edge artificial intelligence,
is driving the paradigm shift of wireless networks from
content delivery to skillset-delivery
networks \cite{Fettweis_JSAC2016}. Network densification \cite{shi2015largecooperative} has emerged as a promising approach to support innovative mobile applications with stringent requirements
such as ultra-low latency, ultra-high data rate and massive devices connectivity. Unfortunately, interference in
dense wireless network deployment becomes a key capacity limiting factor
given large numbers of transmitters and receivers.
Network cooperation through sharing channel state information (CSI) and messages among transmitting nodes is a viable technology to improve the spectral efficiency and energy efficiency in ultra-dense wireless networks.
Under shared CSI among transmitters, interference alignment \cite{cadambe2008interference} is shown to mitigate interferences base
on linear coding schemes, capable of achieving half the cake for each user in $K$-user interference channel. Cooperative transmission \cite{gesbert2010multi} with message sharing has shown to be able to further improve system throughput. In particular, through centralized signal processing and interference management with full message sharing via the cloud data center, cloud radio access network (Cloud-RAN) \cite{Yuanming_TWC2014} can harness the advantages of network densification. By pushing the storage resources to the network edge \cite{yang2016low_globalsip}, cache-aided wireless network \cite{maddah2014fundamental} provides a cost effective way to enable transmission cooperation.
Unfortunately, most existing works on network cooperation
lead to significant channel signaling overhead. This is practically
challenging in ultra-dense networks.
A growing body of recent works has
hence been focusing on CSI acquisition overhead reduction for interference coordination in wireless networks. Among them, delay effect
in CSI acquisition has been considered in \cite{maddah2012completely}.
Both \cite{shi2015optimal} and \cite{razaviyayn2016stochastic} have studied
transceiver design using partial CSI, requiring instantaneous
CSI for strong links and only
distribution CSI of the remaining weak links.
In addition, finite precision CSI
feedback \cite{davoodi2017generalized} and the compressed channel estimation \cite{bajwa2010compressed} can further reduce
CSI acquisition overhead.
However, the applicability of the aforementioned results in practical
systems remains unclear, which motivates a recent proposal on topological interference management (TIM) \cite{Jafar_TIT2013TIM}. The main idea of
TIM is to manage the interference based only on the network connectivity information, which can significantly reduce the CSI acquisition overhead.
By requiring only network topologies, TIM becomes one of the most promising and powerful schemes for interference management in ultra-dense wireless networks. By further enabling message sharing, the work of \cite{yi2015topological} shows that transmitter cooperation based only on network topology information can strictly improve the degrees-of-freedom (DoFs). However, their results are only applicable to some specific network connectivity patterns.
In this paper, we propose a generalized low-rank optimization approach for investigating the benefits of topological cooperation for any network topology. We begin by first establishing the generalized interference alignment conditions based only on the network connectivity information
with message sharing among transmitters. A low-rank model is further developed to maximize the achievable DoFs by exploiting the relationship between the model matrix rank and the achievable DoFs. The developed low-rank matrix optimization model thus generalizes the low-rank matrix completion model \cite{shi2016low} without message sharing among transmitters. Unfortunately,
the resulting generalized low-rank optimization problem in complex field is non-convex and highly intractable due to poor structure,
for which novel and efficient algorithms need to be developed.
Low-rank matrix optimization models have wide range of applications in machine learning, high-dimensional statistics, signal processing and wireless networks \cite{shi2016low,davenport2016overview,sridharan2015linear,papailiopoulos2012interference}. A wealth of recent works focus on both convex
approximation and non-convex algorithms to solve the non-convex and highly intractable low-rank optimization problems.
Nuclear norm is a well-known convex proxy for non-convex rank function with optimality guarantees under statistical models \cite{candes2009exact}.
To further reduce the storage and computation overhead for low-rank optimization, non-convex approach based on matrix factorization shows
good promises \cite{jain2013low}. With suitable statistical models, the non-convex methods can also find globally optimal solution for some structured optimization problems such as matrix completion \cite{ge2016matrix}. In particular, the work of \cite{yi2016topological} adopted an alternating minimization algorithm to exploit topological transmitter cooperation gains. This algorithm stores the iterative
results in the factored form and optimizes over one factor
while fixing the other.
Nevertheless, the nuclear norm based convex relaxation approach
in fact fails to solve the formulation of generalized low-rank matrix optimization problem because of the poor structures. Actually, the nuclear norm minimization approach always yields a full-rank matrix solution. Alternating minimization \cite{jain2013low} algorithm by factorizing the fixed-rank matrix is particularly
useful when the resulting problem is biconvex with respect
to the two factors in matrix factorization. However, the convergence
of the alternating minimization algorithm heavily depends
on the initial points with slow convergence rates.
It may also yield poor performance in achievable DoFs, as it
only guarantees convergence to the first-order stationary points \cite{shi2016low,yi2016topological}. In contrast, Riemannian optimization \cite{Absil_2009optimizationonManifolds} approach has shown to be effective
in improving the achievable DoFs by solving the low-rank matrix
optimization problems, as the Riemannian trust-region algorithm guarantees convergence to the second-order stationary points with high precision solutions \cite{shi2016low}. Furthermore, the Riemannian optimization algorithms are robust to initial points in ensuring convergence \cite{boumal2016global} with fast convergence rates. However, no available Riemannian optimization algorithms have been developed for the general non-square low-rank problems in the
\textit{complex field}. In this work,
we develop Riemannian optimization algorithms for solving the presented generalized low-rank optimization problem in the complex field.
\subsection{Contributions}
In this paper, we develop a generalized interference
alignment condition to enable transmitter cooperation based
only on the network topology information.
We present a generalized low-rank model to maximize the achievable DoFs.
To address the special challenges in the resulting generalized low-rank optimization problem, we develop Riemannian optimization algorithms by exploiting the non-compact Stiefel manifold of
fixed-rank matrices in complex field. Specifically, we propose to solve the generalized low-rank optimization problem by solving a sequence of fixed rank subproblems with rank increase. By applying
semidefinite lifting technique \cite{ge2017no}, the fixed rank
subproblem is reformulated as a positive semidefinite matrix problem
in complex field with rank constraint. By applying the Burer-Monteiro \cite{burer2003nonlinear} parameterization approach to factorize the
positive semidefinite matrix, the resulting problem turns out to be a
Riemannian optimization problem on complex non-compact Stiefel manifold. Therefore, the generalized low-rank optimization problem can be successfully solved by developing Riemannian optimization algorithms on the complex-valued non-compact Stiefel manifold.
We summarize the main contributions of this work as follows:
\begin{enumerate}
\item We establish a generalized interference alignment condition to enable transmitter cooperation with message sharing based
only
on network connectivity information. We develop a generalized low-rank model to maximize the achievable DoFs.
\item We develop first-order and second-order Riemannian optimization algorithms for solving the generalized low-rank optimization problem in \emph{complex field}. We exploit
the complex compact Stiefel manifold of complex fixed-rank matrices using the semidefinite lifting and Burer-Monteiro factorization techniques.
\item Numerical results demonstrate that the proposed second-order Riemannian trust-region algorithm is able to achieve
the highest DoFs with high precision second-order stationary point solutions. Furthermore, its computing time is comparable to the first-order Riemannian conjugate gradient algorithm in medium network sizes. Overall, the Riemannian algorithms show much better performance than the alternating minimization algorithm.
\end{enumerate}
\subsection{Organization and Notation}
The remainder of this paper is organized as follows. In Section \ref{sec:system}, we first introduce the system model, before
establishing the generalized topological interference alignment conditions.
We develop the generalized low-rank model
in Section \ref{sec:GLRM},
and derive a positive semidefinite reformulation with the Burer-Monteiro approach to address the low-rank optimization problem in complex field.
We derive Riemannian algorithms on complex non-compact Stiefel manifold
in Section \ref{sec:manifold}. Section \ref{sec:simulation}
provides simulation results. Finally, we conclude this work
in Section \ref{sec:conclusion},
We use $[K]$ denote the set $\{1,2,\cdots,K\}$. $\mathbb{S}_+^{N}$ denotes the set of all $N\times N$ Hermitian positive semidefinite matrices. And $\langle \cdot,\cdot \rangle$ denotes inner product, i.e., $\langle \bs{A}, \bs{X}\rangle={\textrm{Tr}}(\bs{A}^{\sf{H}}\bs{X})$.
\section{System Model and Problem Formulation}\label{sec:system}
In this section, we establish the generalized interference alignment condition for partially connected $K$-user interference channel with transmitter cooperation.
\setlength\arraycolsep{2pt}
\subsection{System Model}
Consider a partially-connected interference channel with $K$
single-antenna transmitters and $K$ single-antenna receivers.
Transmitters aim to deliver
a set of independent messages $W_1,W_2$, $\cdots,W_K$
to receivers $1,\,2,\,\cdots,\,K$, respectively.
Transmitter $k$ has message $W_k$ and is always connected with receiver $k$.
The channel coefficient $h_{kl}\in\mathbb{C}$
between the $l$-th transmitter and the $k$-th user is nonzero only for $(k,l)\in\mathcal{E}$.
Block fading channel model is
considered in this paper, i.e., $h_{kl}$ remains stationary
in $r$ consecutive
channel uses, during which the input-output relationship is given by
\begin{eqnarray}
\bs{y}_k=\sum_{(k,i)\in\mathcal{E}}{h}_{ki}{\bs{x}}_i+\bs{z}_k, ~\forall k\in[K],
\end{eqnarray}
where ${\bs{x}}_{i}\in\mathbb{C}^{r}$ is the transmitted signal at transmitter $i$, $\bs{y}_k\in\mathbb{C}^{r}$ is the received signal
at receiver $k$, and $\bs{z}_k\in\mathbb{C}^{r}$ is the additive
isotropic white Gaussian noise, i.e., $\bs{z}_k\sim\mathcal{CN}(\bs{0},\bs{\Sigma}_k)$ with
$\bs{\Sigma}_k\in\mathbb{C}^{r\times r}$. Partial connectivity
of the interference channel provides opportunities to enable
cooperative transmission based
only on the network connectivity information. Specifically, transmitter
cooperation is enabled with message sharing, for which we denote the index set of
messages available at transmitter $k$ as $\mathcal{S}_k\subseteq[K]$. A
$5$-user example of such system is shown in Fig. \ref{fig:system}.
Let $R(W_k)$ be the achievable data rate of message $W_k$, i.e.,
there exists a coding scheme such that the rate of message $W_k$
is $R(W_k)$ and the decoding error probability can be arbitrarily small.
Let $\textrm{SNR}$ denote the signal-to-noise-ratio.
For each message delivery, the degree-of-freedom (DoF) \cite{Jafar_TIT2013TIM}, the first order characterization of channel capacity, is defined as
\begin{equation}
\textrm{DoF}(W_k) = \lim_{\textrm{SNR}\rightarrow\infty}\frac{R(W_k)}{\log(\textrm{SNR})},~\forall k\in[K].
\end{equation}
The set of achievable DoF allocation is denoted as $\{\textrm{DoF}(W_1),\cdots,\textrm{DoF}(W_K)\}$,
whose closure is called the DoF region.
This paper adopts DoF as the performance metric
and designs a linear coding scheme.
\subsection{Linear Coding Strategy}
Linear coding scheme is attractive for interference management owing
to its low complexity. Specifically, its optimality in terms
of DoF has been shown via interference alignment \cite{cadambe2008interference}. Its effectiveness has been demonstrated in the problems of topological interference management (TIM) and index coding \cite{Jafar_TIT2013TIM}. We thus focus on linear coding scheme
to design low complexity and efficient approaches for maximizing
achievable DoFs.
Suppose each message $W_k$ is represented by a complex vector $\bs{s}_k\in\mathbb{C}^{d_k}$ with $d_k$ data streams. Let ${\bs{V}}_{kj}\in\mathbb{C}^{r\times d_j}$ be the precoding matrix at transmitter $k$ for message $W_j$. Then the
transmitted signal is given by
\begin{equation}
\bs{x}_k = \sum_{j\in\mathcal{S}_k}\bs{V}_{kj}\bs{s}_j.
\end{equation}
Consequently, the received signal at user $k$ is
\begin{equation}
\bs{y}_k = \!\!\!\!\!\!\!\!\!\!\!\!\sum_{j:(k,j)\in\mathcal{E},k\in\mathcal{S}_j}\!\!\!\!\!\!\!\!\!\! h_{kj}\bs{V}_{jk}\bs{s}_k+ \sum_{i\ne k}\sum_{j:(k,j)\in\mathcal{E},i\in\mathcal{S}_j}\!\!\!\! \!\!\! h_{kj}\bs{V}_{ji}\bs{s}_i+\bs{z}_k.
\end{equation}
We let $\bs{U}_k\in\mathbb{C}^{r\times d_k}$ be the decoding matrix at receiver $k$.
\begin{figure}
\centering
\includegraphics[width=0.65\columnwidth]{cache_network}
\caption{The architecture of the partially-connected $K$-user interference channel with transmitter cooperation. $\mathcal{S}_i$ denotes the index set of messages available at transmitter $i$.}\label{fig:system}
\end{figure}
In densified wireless networks, interference is a key bottleneck to support high data rate and low latency. To alleviate interferences by aligning the intersection of interference spaces, the following interference alignment conditions were presented in \cite{Jafar_TIT2013TIM,bresler2014feasibility}
\begin{eqnarray}
\label{equa:IA1}
\sum\limits_{j:(k,j)\in\mathcal{E},k\in\mathcal{S}_j}h_{kj}\bs{U}_k^{\sf{H}}\bs{V}_{jk} &\ne &0,~\forall k\in [K], \\
\label{equa:IA2}
\sum\limits_{j:(k,j)\in\mathcal{E},i\in\mathcal{S}_j}h_{kj}\bs{U}_k^{\sf{H}}\bs{V}_{ji} &=&0,~i\ne k.
\end{eqnarray}
Correspondingly, the message at receiver $k$ is decoded via
\begin{equation}
\hat{\bs{s}}_k=(\!\!\!\!\!\!\sum\limits_{j:(k,j)\in\mathcal{E},k\in\mathcal{S}_j}\!\!\!\!\!\!h_{kj}\bs{U}_k^{\sf{H}}\bs{V}_{jk})^{-1}\bs{U}_k^{\sf{H}}\bs{y}_k.
\end{equation}
If there exists $\bs{U}_k$'s, $\bs{V}_{ji}$'s satisfying interference alignment conditions (\ref{equa:IA1}) and (\ref{equa:IA2}), DoF tuple $({d_1}{r}^{-1},\cdots,{d_K}{r}^{-1})$ is then achievable. We thus can achieve the highest DoF by finding the minimal channel use number $r$.
\subsection{Topology-Based Alignment Condition}
Note that equations (\ref{equa:IA1}) and (\ref{equa:IA2}) are always feasible by increasing $r$.
However, the interference alignment conditions (\ref{equa:IA1}) and (\ref{equa:IA2}) require the knowledge of channel coefficients $h_{ij}$s at the transmitters. In practice, obtaining dense network channel state information (CSI)
at transmitters often requires large signaling overhead,
which presents a severe obstacle to their application in
densified wireless networks.
One desirable way to address the CSI acquisition overhead issue is to
establish new interference alignment conditions based only on the network connectivity information, for which we present the following generalized interference alignment conditions for \textit{topological cooperation}
\begin{eqnarray}
\label{siso_cgm1}
&&\textrm{det}\left(\sum_{j:(k,j)\in\mathcal{E},k\in\mathcal{S}_j}{\bs{U}}_k^{\sf{H}}{\bs{V}}_{jk}\right)\ne
0,~~\forall k\in [K],\\
\label{siso_cgm2}
&&\bs{U}_k^{\sf{H}}\bs{V}_{ji} =\bs{0},~~i\in\mathcal{S}_j,i\ne k, (k,j)\in\mathcal{E}.
\end{eqnarray}
Here, ``topological cooperation" refers to the fact that for cooperation enabled transmitters, we design transceivers to manage interferences based on network topology information instead of instantaneous channel state information.
\begin{proposition}
For generic channel coefficients $h_{ij}$'s randomly distributed according
to some continuous probability distribution \cite{razaviyayn2012degrees}, if (\ref{siso_cgm1}) and (\ref{siso_cgm2}) hold for some $\bs{U}_{k},\bs{V}_{ji}$s based only on the network topology information, then they shall satisfy the channel dependent interference alignment conditions (\ref{equa:IA1}) and (\ref{equa:IA2}) with probability 1.
\end{proposition}
\begin{proof}
Let $\bs{Z}_{1},\cdots,\bs{Z}_T\in\mathbb{C}^{d\times d}$ denote the set of matrices $\{\bs{U}_k^{\sf{H}}\bs{V}_{jk}:(k,j)\in\mathcal{E},k\in\mathcal{S}_j\}$ given $k$. Our goal is to prove that if $\textrm{det}\Big(\sum_{t=1}^{T}\bs{Z}_{t}\Big)\ne 0$, the probability of $\textrm{det}\Big(\sum_{t=1}^{T}h_t\bs{Z}_{t}\Big)\ne 0$ is $1$ for generic $h_1,\cdots,h_T$. Note that $\sum_{t=1}^{T}\bs{Z}_{t} = \begin{bmatrix}
\bs{Z}_1 & \cdots & \bs{Z}_T
\end{bmatrix}\begin{bmatrix}
\bs{I} & \cdots & \bs{I}
\end{bmatrix}^H$. Thus, the condition $\textrm{det}\Big(\sum_{t=1}^{T}\bs{Z}_{t}\Big)\ne 0$ implies that $\begin{bmatrix}
\bs{Z}_1 & \cdots & \bs{Z}_T
\end{bmatrix}$ has full rank, i.e., the dimension of $\textrm{span}\{\bs{Z}_j\}$ is $r$. Since the solution to the determinant equation $\textrm{det}\Big(\sum_{t=1}^{T}h_t\bs{Z}_{t}\Big)= 0$ is an algebraic hypersurface \cite{bresler2014feasibility}, the probability of the linear combination $\sum_{t=1}^{T}h_t\bs{Z}_{t}$ with generic coefficients $h_1,\cdots,h_T$ lying on the algebraic hypersurface is hence zero. Therefore, $\textrm{det}\Big(\sum_{t=1}^{T}h_t\bs{Z}_{t}\Big)\ne 0$ holds with probability $1$.
\end{proof}
By leveraging conditions (\ref{siso_cgm1}) and (\ref{siso_cgm2}), interferences can be aligned based only on network topology CSI instead of full
CSI.
This significantly reduces the overhead of CSI acquisition. In particular, the topological interference alignment condition without message sharing is given by \cite{shi2016low}
\begin{eqnarray}
\label{TIM_con1}
&&\textrm{det}\left({\bs{U}}_k^{\sf{H}}{\bs{V}}_{k}\right)\ne
0,~~\forall k\in [K],\\
\label{TIM_con2}
&&\bs{U}_k^{\sf{H}}\bs{V}_{j} =\bs{0},~~j\ne k, (k,j)\in\mathcal{E},
\end{eqnarray}
which is a special case of (\ref{siso_cgm1}) and (\ref{siso_cgm2}) with $\mathcal{S}_j=\{j\}$ and ${\bs{V}}_{j}$ denoting as ${\bs{V}}_{jj}$. Conditions (\ref{siso_cgm1}) and (\ref{siso_cgm2}) thus manifest the benefits of transmitter cooperation, as solutions to (\ref{TIM_con1}) and (\ref{TIM_con2}) are always solutions to (\ref{siso_cgm1}) and (\ref{siso_cgm2}), but not conversely.
\begin{remark}
This work assumes that there are equal number of transmitters and receivers. Nevertheless, the principle applies for arbitrary number of transmitters and receivers. This is because both Proposition 1 and the low-rank matrix representation for precoding and decoding matrices in Section 3 hold for any number of transmitters and receivers. For simplicity of notation we consider a system with $K$ transmitters and receivers in this paper.
\end{remark}
\section{Generalized Low-Rank Optimization for Topological Cooperation}\label{sec:GLRM}
This section develops a generalized low-rank optimization framework to maximize achievable DoFs under topological cooperation. To address the challenges of the present generalized low-rank optimization problem in complex field and
to exploit the algorithmic benefits of Riemannian optimization, we propose to reformulate an optimization problem over the complex non-compact Stiefel manifold by using the semidefinite lifting and Burer-Monteiro approaches.
\subsection{Generalized Low-Rank Model for Topological Cooperation}
Without loss of generality, we restrict $\sum_{j:(k,j)\in\mathcal{E},k\in\mathcal{S}_j}{\bs{U}}_k^{\sf{H}}{\bs{V}}_{jk}=\bs{I}$ in condition (\ref{siso_cgm1}). By letting $m=\sum_{k}d_k,~ n=K\sum_{k}d_k$ and defining
\begin{align*}
\bs{U} &= \begin{bmatrix}
\bs{U}_1 & \cdots & \bs{U}_{K}
\end{bmatrix}\in\mathbb{C}^{r\times m}, \\
\bs{V}_{j}&=[{\bs{V}}_{j1},\cdots,{\bs{V}}_{jK}]\in\mathbb{C}^{r\times m}, \\
\bs{V}&= \begin{bmatrix}
\bs{V}_1 & \cdots & \bs{V}_{K}
\end{bmatrix}\in\mathbb{C}^{r\times n},\\
\bs{X}&=[\bs{X}_{kj}^{i}]=[\bs{U}_k^{\sf{H}}\bs{V}_{ji}]={\bs{U}}^{\sf{H}}{\bs{V}}\in\mathbb{C}^{m\times n},
\end{align*}
the rank of matrix $\bs{X}$ is given as
\begin{equation}
\textrm{rank}(\bm{X})=r = d_k/\text{DoF}(W_k).
\end{equation}
We thus can maximize the achievable DoF for interference-free message delivery by solving the following \emph{generalized low-rank optimization} problem
\begin{eqnarray}
\mathscr{P}:\mathop{\textrm{minimize}}_{{\bs{X}}\in\mathbb{C}^{m\times n}}&& \textrm{rank}({\bs{X}}) \nonumber\\
\textrm{subject to}&& \mathcal{A}({\bs{X}})={\bs{b}},\label{GLRM_rank}
\end{eqnarray}
where the affine constraint $\mathcal{A}({\bs{X}})={\bs{b}}$ captures
\begin{eqnarray}
&& \sum_{j:(k,j)\in\mathcal{E},k\in\mathcal{S}_j}\bs{X}_{kj}^{k}=\bs{I},~\forall k\in [K] \label{prob:GLRM1eq1}\\
&&\bs{X}_{kj}^{i} =\bs{0},~i\ne k,i\in\mathcal{S}_j, (k,j)\in\mathcal{E} \label{prob:GLRM1eq2
\end{eqnarray}
and $\mathcal{A}:\mathbb{C}^{m\times n}\mapsto\mathbb{C}^{l}$.
For the simpler case without message sharing, the topological interference alignment problem can be formulated as the following low-rank matrix completion problem \cite{shi2016low,hassibi2014topological}
\begin{eqnarray}\label{prob:TIM}
\mathop{\textrm{minimize}}_{\bm{X}\in\mathbb{C}^{m\times m}}&& \textrm{rank}(\bs{X}) \nonumber\\
\textrm{subject to}&& \bs{X}_{kk}=\bs{I},~\forall k\in [K] \nonumber\\
&&\bs{X}_{kj} = \bs{0},~j\ne k, (k,j)\in\mathcal{E},
\end{eqnarray}
which is a special case of problem $\mathscr{P}$. The resulting low-rank matrix completion model is demonstrated in Fig. \ref{fig:tim_lrmc}.
\begin{figure}[t]
\centering
\subfloat[TIM]{\includegraphics[width=0.47\columnwidth]{TIM}\label{fig:TIM}} \hfil
\subfloat[Matrix Completion Model]{\includegraphics[width=0.45\columnwidth]{LRMC}\label{fig:LRMC}}
\caption{Matrix completion model for the topological interference alignment without message sharing for single data stream $d_k=1$. In this case, all the diagonal entries are set to 1 as $\mathcal{S}_k=\{k\}$. As an example, the $(2,3)$-th entry is zero because the second receiver is connected with the third transmitter as interference. And the $(3,2)$-th entry of the matrix can be arbitrary value as $(3,2)\not\in\mathcal{E}$.}
\label{fig:tim_lrmc}
\end{figure}
\subsection{Problem Analysis}
Basically, methods for solving low-rank problems can be divided into two categories. One uses convex relaxation approach and the other one uses nonconvex approach based on matrix factorization. In addition, penalty decomposition method is proposed in \cite{zhang2011penalty} for low-rank optimization problems. The inner iterations adopt a block coordinated descent method, whereas the outer iterations update the weight of rank function. However, each inner iteration requires the computation of singular value decomposition, which leads to large computation overhead ($\mathcal{O}(mnl+m^2n+m^3)$). Therefore, it is not suitable for our tranceiver design problem in ultra-dense networks.
\subsubsection{Convex Relaxation Methods}
Nuclear norm is a well-known convex proxy \cite{davenport2016overview} for rank function.
The nuclear norm relaxation approach for problem $\mathscr{P}$ is given by
\begin{eqnarray}
\mathop{\textrm{minimize}}_{{\bs{X}}\in\mathbb{C}^{m\times n}}&& \|{\bs{X}}\|_* \nonumber\\
\textrm{subject to}&& \mathcal{A}({\bs{X}})={\bs{b}}.\label{prob:nuc}
\end{eqnarray}
It can be solved by an equivalent semidefinite programming (SDP) problem
\begin{eqnarray}
\mathop{\textrm{minimize}}_{\bs{X},\bs{W}_1,\bs{W}_2}&& \textrm{Tr}(\bs{W}_1)+\textrm{Tr}(\bs{W}_2) \nonumber\\
\textrm{subject to}&& \mathcal{A}({\bs{X}})={\bs{b}}, \label{prob:nuc} \\
&& \begin{bmatrix}
\bs{W}_1 & \bs{X} \\ \bs{X}^{\sf{H}} & \bs{W}_2
\end{bmatrix}\succeq \bs{0} \nonumber.
\end{eqnarray}
Unfortunately, the SDP solution requires computing singular value decomposition at each iteration, which is not scalable to large problem sizes in ultra-dense networks. Specifically, with high precision second-order interior point method, the convergence rate is fast while the computational cost for each iteration is $\mathcal{O}((mn+l)^3)$ due to computing the Newton step \cite{boyd2004convex}. Using the first-order algorithm alternating direction method of multipliers (ADMM) \cite{o2016conic}, the computational cost is $\mathcal{O}(mnl+m^2n+m^3)$ at each iteration. Furthermore, the nuclear norm relaxation approach always yields a full rank solution due to the poor structure of the affine operator $\mathcal{A}$.
\begin{proposition}\label{proposition:nuc}
The nuclear norm relaxation approach (\ref{prob:nuc}) for the generalized low-rank optimization problem $\mathscr{P}$ always yields a full rank solution.
\end{proposition}
\begin{proof}
See Appendix \ref{append:nuc}.
\end{proof}
Therefore, the nuclear norm relaxation based approach is inapplicable for the poorly structured low-rank optimization problem $\mathscr{P}$. We thus call problem $\mathscr{P}$ as the generalized low-rank optimization problem.
\subsubsection{Nonconvex Approaches}
A rank $r$ matrix ${\bs{X}}$ can be factorized as ${\bs{X}}={\bs{L}}{\bs{R}}^{\sf{H}}$, where ${\bs{L}}\in\mathbb{C}^{m\times r}$ and ${\bs{R}}\in\mathbb{C}^{n\times r}$. Nonconvex approaches to low-rank optimization leverage matrix factorizations and design various updating strategies for two factors $\bm{U}$ and $\bm{V}$. By solving a sequence of the fixed rank least square subproblems based on matrix factorization
\begin{eqnarray}\label{GLRM_fixedrank}
\mathop{\textrm{minimize}}_{{\bs{X}}\in\mathbb{C}^{m\times n}} && f_0({\bs{X}})=\frac{1}{2}\|\mathcal{A}({\bs{X}})-{\bs{b}}\|_2^2 \nonumber\\
\textrm{subject to}&& \textrm{rank}({\bs{X}})=r,
\end{eqnarray}
and increasing $r$, we can find the minimal rank $r$ for the original problem $\mathscr{P}$.
Specifically, for the rank constrained problem (\ref{GLRM_fixedrank}) with convex objective function, the alternating minimization \cite{jain2013low} algorithm can function as follows:
\begin{eqnarray}
\bm{L}_{k+1} &=& \arg\min_{{\bs{L}}}~~ f_0(\bs{L}\bs{R}_{k}^{\sf{H}}) \label{GLRM_fixedrank_altmin1},\\
\bm{R}_{k+1} &=& \arg\min_{{\bs{R}}}~~ f_0(\bs{L}_{k+1}\bs{R}^{\sf{H}}). \label{GLRM_fixedrank_altmin2}
\end{eqnarray}
It essentially optimizes the bi-convex objective function $f_0(\bm{L}\bm{R}^{\sf{H}})$ by freezing one of ${\bs{L}}$ and ${\bs{R}}$ alternatively. However, the convergence of the alternating minimization algorithm are sensitive to initial points and its convergence rate
can be slow. Furthermore, it may yield poor performance for achievable DoFs maximization by only converging to first-order stationary point.
In contrast, Riemannian optimization algorithms are capable
of updating the two factors $\bm{L}$ and $\bm{R}$
\emph{simultaneously} by exploiting the
quotient manifold geometry of fixed-rank matrices based
on matrix factorization. First-order Riemannian conjugate
gradient and second-order Riemannian trust-region algorithm
can help find first-order stationary points and second-order
stationary points, respectively. It has been shown in \cite{boumal2016global} that Riemannian optimization algorithms converge to first-order and second-order stationary points from arbitrary initial points.
Furthermore, Riemannian trust-region algorithm can
achieve high achievable DoFs with second-order stationary points
while also enjoys locally super-linear \cite{Absil_2009optimizationonManifolds} convergence rates.
\begin{remark}
The invariance of matrix factorization $(\bs{L}\bs{M}^{\sf{H}},\bs{M}^{-1}\bs{R})$ for any full rank matrix $\bs{M}$ makes the critical points of $f_0$ parameterized with $\bs{L}$ and $\bs{R}$ are not isolated in Euclidean space. This indeterminancy profoundly affects the convergence of second-order optimization algorithms \cite{shi2017useradmission,journee2010lowrank}. To address this issue, we shall develop efficient algorithms on the quotient manifold instead of Euclidean space.
\end{remark}
Unfortunately, available Riemannian algorithms for non-square fixed-rank matrix optimization problems only operate in $\mathbb{R}$ field \cite{Absil_2009optimizationonManifolds} and do not directly apply to
solve problem (\ref{GLRM_fixedrank}) in the complex filed.
Inspired by the fact that the complex non-compact Stiefel manifold is well defined \cite{yatawatta2013radio}, we propose to reformulate complex matrix optimization problem (\ref{GLRM_fixedrank}) on the complex non-compact Stiefel manifold by using the Burer-Monteiro approach.
Specifically, applying semidefinite lifting, the original problem (\ref{GLRM_fixedrank}) is equivalently reformulated into rank constrained positive semidefinite matrix optimization, before factorizing the semidefinite matrices using the Burer-Monteiro approach. The original complex matrix optimization problem (\ref{GLRM_fixedrank}) is thus reformulated as the Riemannian optimization problem over the well-defined non-compact Stiefel manifold in complex field.
\subsection{Semidefinite Lifting and the Burer-Monteiro Approach}\label{sec:liftBM}
Burer-Monteiro approach is a well-known nonconvex parameterization method for solving positive semidefinite (PSD) matrices problems \cite{burer2003nonlinear}. A rank $r$ PSD matrix ${\bs{Z}}\in\mathbb{S}^{N}$ can be factorized as ${\bs{Z}}={\bs{Y}}{\bs{Y}}^{\sf{H}}$ with ${\bs{Y}}\in\mathbb{C}^{N\times r}$. The linear operator $\mathcal{A}$ can be represented as a set of matrices $\bs{A}_i\in\mathbb{C}^{n\times m}$, i.e.,
\begin{equation}\label{eq:defA}
\mathcal{A}({\bs{X}}) = [\langle \bs{A}_i, \bs{X}\rangle], i=1,\cdots,l.
\end{equation}
Then the objective function in (\ref{GLRM_fixedrank}) can be rewritten as
\begin{equation}
f_0(\bs{X})=\frac{1}{2}\sum_{i=1}^{l}|\langle \bs{A}_i, \bs{X}\rangle-b_i|^2.
\end{equation}
By semidefinite lifting \cite{ge2017no} ${\bs{X}}$ to
\begin{equation}
{\bs{Z}}=\begin{bmatrix}
{\bs{Z}}_{11} & {\bs{Z}}_{12} \\
{\bs{Z}}_{21} & {\bs{Z}}_{22}
\end{bmatrix}:=\begin{bmatrix}
{\bs{L}}{\bs{L}}^{\sf{H}} & {\bs{L}}{\bs{R}}^{\sf{H}} \\
{\bs{R}}{\bs{L}}^{\sf{H}} & {\bs{R}}{\bs{R}}^{\sf{H}}
\end{bmatrix},
\end{equation}
we can reformulate problem (\ref{GLRM_fixedrank}) as a complex PSD matrix problem with rank constraint:
\begin{eqnarray}\label{GLRM_fixedrank_PSD}
\mathop{\textrm{minimize}}_{{\bs{Z}}\in\mathbb{S}_+^{N}}&& \frac{1}{2}\|\mathcal{B}({\bs{Z}})-{\bs{b}}\|_2^2 \nonumber\\
\textrm{subject to}&& \textrm{rank}({\bs{Z}})= r,
\end{eqnarray}
where $N=m+n$ and
\begin{equation}
\mathcal{B}(\bs{Z})=\mathcal{A}(\bs{Z}_{12})=\mathcal{A}({\bs{L}}{\bs{R}}^{\sf{H}}).
\end{equation}
Here we use $\mathcal{B}$ to denote
a set of matrices $\bs{B}_i\in\mathbb{C}^{N\times N}$
\begin{equation}\label{eq:defB}
\bs{B}_i = \begin{bmatrix}
\bs{0} & \bs{A}_i \\ \bs{0} & \bs{0}
\end{bmatrix}, \langle \bs{B}_i, \bs{Z}\rangle = \langle \bs{A}_i, \bs{X}\rangle.
\end{equation}
We define ${\bs{Y}} = \begin{bmatrix}
{\bs{L}} \\ {\bs{R}}
\end{bmatrix}\in\mathbb{C}^{N\times r}$.
The search space $\{\bm{Z}:\bm{Z}\in\mathbb{S}_+^{N},\textrm{rank}({\bs{Z}})= r\}$ admits a well-defined manifold structure, by factorizing $\bm{Z}=\bm{Y}\bm{Y}^{\sf{H}}$ based on the principles of Burer-Monteiro approach. Problem (\ref{GLRM_fixedrank_PSD}) thus can be transformed as
\begin{eqnarray}\label{GLRM_BM}
\mathop{\textrm{minimize}}_{{\bs{Y}}\in\mathbb{C}_*^{N\times r}}&& f({\bs{Y}})=\frac{1}{2}\|\mathcal{B}(\bs{Y}\bs{Y}^{\sf{H}})-\bs{b}\|_2^2.
\end{eqnarray}
This is a Riemannian optimization problem with a smooth ($C^{\infty}$) objective function over the complex \textit{non-compact Stiefel manifold} $\mathbb{C}_*^{N\times r}$, i.e., the set of all $N\times r$ full column rank matrices in complex field.
In summary, we propose to solve the generalized low-rank optimization
problem by solving a sequence of \textit{complex} fixed-rank optimization problem using the Riemannian optimization technique. This is
achieved by lifting the complex fixed-rank optimization problem
into the complex positive semidefinite matrix optimization problem,
followed by parameterizing it using the Burer-Monteior approach.
This yields the Riemannian optimization problem over complex
non-compact Stiefel manifold. After obtaining a solution $\bs{Y}$ from (\ref{GLRM_BM}), we can recover the solution
$\bs{X}={\bs{L}}{\bs{R}}^{\sf{H}}$ to the original problem (\ref{GLRM_fixedrank}). The whole algorithm of addressing the transmitter cooperation problem based only on the network topology information is demonstrated in Algorithm \ref{algorithm:riemalgo}.
\SetNlSty{textbf}{}{:}
\IncMargin{1em}
\begin{algorithm}[tb]
\textbf{Input:} $\{S_j\},\mathcal{E},K,\{d_k\}$, accuracy $\varepsilon$. \\
Construct $\mathcal{B}$ and $\bm{b}$ following (\ref{prob:GLRM1eq1}) (\ref{prob:GLRM1eq2}) (\ref{eq:defA}) (\ref{eq:defB}). Let $N=m+n=(K+1)\sum_{k}d_k$.
\\
\For{$r=1,\cdots,N$}{
Solve (\ref{GLRM_BM}) with Riemannian optimization algorithm.\\
\If{$f({\bs{Y}}^{[r]})< \varepsilon$}{\Return $\bs{Y}^{[r]}$}
}
\textbf{Output:} $\bm{Y}^{[r]}$ and rank $r$.
\caption{Optimization Framework for Transmitter Cooperation Based on Network Topology Information
\label{algorithm:riemalgo}
\end{algorithm}
\section{Matrix Optimization on Complex Non-compact Stiefel Manifold}\label{sec:manifold}
In this section, we shall develop Riemannian conjugate gradient and Riemannian trust-region algorithms for solving problem (\ref{GLRM_BM}). Riemannian optimization generalizes the concepts of gradient and Hessian in Euclidean space to Riemannian gradient and Hessian on manifolds. They are represented in the tangent space, which is the linearization of the search space.
\subsection{Quotient Geometry of Fixed-Rank Problem}
For problem (\ref{GLRM_BM}), the optima are not isolated because of ${\bs{Y}}{\bs{Y}}^{\sf{H}}$ remains invariant under the \textit{canonical projection} \cite[Sec 3.4.1]{Absil_2009optimizationonManifolds}
\begin{equation}\label{eq:canonicalprojection}
\pi:~{\bs{Y}}\mapsto {\bs{Y}}{\bs{Q}}
\end{equation}
for any unitary matrix ${\bs{Q}}\in\mathcal{U}(r)$
where $\mathcal{U}(r)$ denotes the set of
$r\times r$ unitary matrices.
To address this non-uniqueness we consider problem (\ref{GLRM_BM}) over the equivalent class
\begin{equation}
[{\bs{Y}}] = \{{\bs{Y}}{\bs{Q}}:{\bs{Y}}\in\overline{\mathcal{M}}=\mathbb{C}_*^{N\times r},{\bs{Q}}\in\mathcal{U}(r)\},
\end{equation}
that is
\begin{eqnarray}\label{GLRM_equivalentclass}
\mathop{\textrm{minimize}}_{[{\bs{Y}}]\in\mathcal{M}}&& f([{\bs{Y}}]).
\end{eqnarray}
Then the whole set of feasible solutions can be represented by isolated points in the \textit{quotient manifold}, i.e. $\mathcal{M}=\overline{\mathcal{M}}/\sim
:=\overline{\mathcal{M}}/\mathcal{U}(r)$ with \textit{canonical projection} \cite[Sec 3.4]{Absil_2009optimizationonManifolds} $\pi$. Here $\sim$ is the equivalence relation and $\overline{\mathcal{M}}/\sim:=\{[\bs{Y}]:\bs{Y}\in\overline{\mathcal{M}}\}$. $\overline{\mathcal{M}}$ is considered as an abstract manifold.
\subsection{Riemannian Ingredients for Iterative Algorithms on Riemannian Manifolds}
By studying the unconstrained problem on the quotient manifold instead of the constrained problem in Euclidean space, Riemannian optimization can exploit the non-uniqueness of matrix factorization with Burer-Monteiro approach. We now develop conjugate gradient and trust-region algorithms on the Riemannian manifold.
To achieve this goal, we first linearize the search space,
by defining the concept of \textit{tangent space} \cite[Sec 3.5]{Absil_2009optimizationonManifolds} and associated ``inner product'' on the tangent space. Next, we will derive the expressions for Riemannian gradient and Riemannian Hessian in this subsection.
Specifically, tangent space $\mathcal{T}_{\bs{Y}}\overline{\mathcal{M}}$ is a vector space consisting of all tangent vectors to $\overline{\mathcal{M}}$ at $\bs{Y}$.
\begin{proposition}
The tangent space of $\overline{\mathcal{M}}=\mathbb{C}_*^{N\times r}$ at $\bs{Y}$ is given by $\mathcal{T}_{\bs{Y}}\overline{\mathcal{M}}=\mathbb{C}^{N\times r}$.
\end{proposition}
\begin{proof}
$\overline{\mathcal{M}}$ is an open submanifold \cite[Sec 3.5.2]{Absil_2009optimizationonManifolds} of $\mathbb{C}^{N\times r}$ and hence, $\mathcal{T}_{\bs{Y}}\overline{\mathcal{M}}=\mathbb{C}^{N\times r}$ for all $\bs{Y}\in\overline{\mathcal{M}}$.
\end{proof}
In order to eliminate the non-uniqueness along the equivalent class $[\bs{Y}]$, we will decompose the
tangent space into two orthogonal parts, i.e., \textit{vertical space} and \textit{horizontal space}. Vertical space $\mathcal{V}_{\bs{Y}}$ is the tangent space of equivalent class $[\bs{Y}]$, while horizontal space $\mathcal{H}_{\bs{Y}}$ is the orthogonal complement of vertical space in
the tangent space. That is,
\begin{equation}
\mathcal{T}_{\bs{Y}}\overline{\mathcal{M}} = \mathcal{V}_{\bs{Y}} \oplus \mathcal{H}_{\bs{Y}},
\end{equation}
where $\oplus$ denotes the direct sum of two subspace. In this way, we can always find the unique ``lifted'' representation of the tangent vectors of $\mathcal{T}_{[\bs{Y}]}\mathcal{M}$ in $\mathcal{T}_{\bs{Y}}\overline{\mathcal{M}}$ at any element of $[\bs{Y}]$, i.e., for any $\bs{\xi}\in\mathcal{T}_{\bs{Y}}\mathcal{M}$ we define a unique horizontal lift $\overline{\bs{\xi}}\in\mathcal{H}_{\bs{Y}}$ at $\bs{Y}$ such that
\begin{equation}
\overline{\bs{\xi}}:= \Pi_{\bs{Y}}^{h}\bs{\xi},
\end{equation}
where \textit{horizontal projection} $\Pi_{\bs{Y}}^{h}(\cdot)$ is the orthogonal projection from $\mathcal{T}_{\bs{Y}}\mathcal{M}$ onto $\mathcal{H}_{\bs{Y}}$.
\begin{proposition}\label{proposition:vertical}
The vertical space at $\bs{Y}$ is given by
\begin{equation}\label{eq:vspace}
\mathcal{V}_{\bs{Y}}\triangleq \{\bs{Y}\bs{\Omega}:\bs{\Omega}^{\sf{H}}=-\bs{\Omega},\bs{\Omega}\in\mathbb{C}^{r\times r} \}.
\end{equation}
\end{proposition}
\begin{proof}
See Appendix \ref{append:vertical}.
\end{proof}
According to the definition, horizontal space should be derived from
\begin{equation}
\mathcal{H}_{\bs{Y}} = \{\overline{\bs{\xi}}\in\mathcal{T}_{\bs{Y}}\overline{\mathcal{M}}:\overline{g}_{\bs{Y}}(\overline{\bs{\xi}},\overline{\bs{\zeta}})=0,~\forall \overline{\bs{\zeta}}\in\mathcal{V}_{\bs{Y}}\},
\end{equation}
where $\overline{g}$ is \textit{Riemannian metric} for the abstract manifold $\overline{\mathcal{M}}$. Riemannian metric is the generalization of ``inner product'' in Euclidean space to a manifold. It is a bilinear, symmetric positive-definite operator
\begin{equation}
\overline{g}:~\mathcal{T}_{\bs{Y}}\overline{\mathcal{M}} \times \mathcal{T}_{\bs{Y}}\overline{\mathcal{M}} \mapsto \mathbb{R}.
\end{equation}
In this paper, we can choose
\begin{equation}\label{eq:metric}
\overline{g}_{\bs{Y}}(\overline{\bs{\xi}},\overline{\bs{\zeta}}):={\textrm{Tr}}(\Re(\overline{\bs{\xi}}^{\sf{H}}\overline{\bs{\zeta}}))=\frac{1}{2}\textrm{Tr}(\overline{\bs{\xi}}^{\sf{H}}\overline{\bs{\zeta}}+\overline{\bs{\zeta}}^{\sf{H}}\overline{\bs{\xi}})
\end{equation}
as a Riemannian metric for the abstract manifold $\overline{\mathcal{M}}$, where $\bs{Y}\in\overline{\mathcal{M}}$ and $\overline{\bs{\xi}},\overline{\bs{\zeta}}\in\mathcal{T}_{\bs{Y}}\overline{\mathcal{M}}$. The manifold $\overline{\mathcal{M}}$ is called a \textit{Riemannian manifold} when its tangent spaces are endowed with a Riemannian metric. From another perspective, $\overline{\mathcal{M}}$ can also be viewed as a K\"{a}hler manifold whose K\"{a}hler form is a real closed (1,1)-form \cite{barth2015compact}.
Therefore, we can obtain the explicit expressions for the horizontal space and horizontal projection.
\begin{proposition}\label{proposition:horizontal}
The horizontal space is
\begin{equation}\label{eq:hspace}
\mathcal{H}_{\bs{Y}}=\{\bs{\xi}\in\mathbb{C}^{N\times r}:\bs{\xi}^{\sf{H}}\bs{Y}=\bs{Y}^{\sf{H}}\bs{\xi}\},
\end{equation}
and the orthogonal projection onto the horizontal space is
\begin{equation}\label{eq:hproj}
\Pi_{\bs{Y}}^h \bs{\xi}_{\bs{Y}} = \bs{\xi}_{\bs{Y}}- \bs{Y}\bs{\Omega},
\end{equation}
where $\bs{\Omega}^{\sf{H}}=-\bs{\Omega}\in\mathbb{C}^{r\times r}$ is the solution of Lyapunov equation
\begin{equation}\label{eq:lyap}
\bs{Y}^{\sf{H}}\bs{Y}\bs{\Omega}+\bs{\Omega}\bs{Y}^{\sf{H}}\bs{Y}=\bs{Y}^{\sf{H}}\bs{\xi}_{\bs{Y}}-\bs{\xi}_{\bs{Y}}^{\sf{H}}\bs{Y}.
\end{equation}
\end{proposition}
\begin{proof}
See Appendix \ref{append:horizontal}.
\end{proof}
Given the Riemannian metric for the abstract manifold, quotient manifold is naturally endowed with a Riemannian metric
\begin{equation}
g_{[\bs{Y}]}(\bs{\xi}_{[\bs{Y}]},\bs{\zeta}_{[\bs{Y}]}) : = \overline{g}_{\bs{Y}}(\overline{\bs{\xi}}_{\bs{Y}},\overline{\bs{\zeta}}_{\bs{Y}})
\end{equation}
such that the expression $\overline{g}_{\bs{Y}}(\overline{\bs{\xi}}_{\bs{Y}},\overline{\bs{\zeta}}_{\bs{Y}})$ remains for any elements in the equivalent class $[\bs{Y}]$.
Hence $\mathcal{M}$ is a \textit{Riemannian quotient manifold} of the abstract manifold $\overline{\mathcal{M}}$ with the Riemannian metric $g$, and the canonical projection $\pi:(\overline{\mathcal{M}},\overline{g})\mapsto (\mathcal{M},g)$ is a \textit{Riemannian submersion} \cite[Sec 3.6.2]{Absil_2009optimizationonManifolds}.
Riemannian optimization generalizes the gradient and Hessian into Riemannian gradient and Riemannian Hessian.
\begin{table*}[ht]
\centering
\begin{tabular}{c|c r}
{$\mathop{\textrm{minimize}}_{[{\bs{Y}}]\in\overline{\mathcal{M}}}~ f({\bs{Y}})$} & & \\\hline
Computation space & $\overline{\mathcal{M}}=C_*^{N\times r},\mathcal{M}=\overline{\mathcal{M}}/\sim $ & \\
Canonical projection & $\pi:\bs{Y}\mapsto \bs{Y}\bs{Q},{\bs{Q}}\in\mathcal{U}(r)\qquad $ & (\ref{eq:canonicalprojection}) \\
Remannian metric & $g_{\bs{Y}}(\bs{\xi}_{\bs{Y}},\bs{\eta}_{\bs{Y}}) = \textrm{Tr}(\bs{\xi}_{\bs{Y}}^{\sf{H}}\bs{\eta}_{\bs{Y}}+\bs{\eta}_{\bs{Y}}^{\sf{H}}\bs{\xi}_{\bs{Y}})\qquad$ & (\ref{eq:metric}) \\
Vertical space & $\mathcal{V}_{\bs{Y}}=\{\bs{Y}\bs{\Omega}:\bs{\Omega}^{\sf{H}}=-\bs{\Omega},\bs{\Omega}\in\mathbb{C}^{r\times r} \}\qquad$ & (\ref{eq:vspace}) \\
Horizontal space & $\mathcal{H}_{\bs{Y}}=\{\bs{\xi}\in\mathbb{C}^{N\times r}:\bs{\xi}^{\sf{H}}\bs{Y}=\bs{Y}^{\sf{H}}\bs{\xi}\}\qquad$ & (\ref{eq:hspace}) \\
Projection onto horizontal space & $ \Pi_{\mathcal{M}}^h\bs{\xi}_{\bs{Y}}=\bs{\xi}_{\bs{Y}}- \bs{Y}\bs{\Omega}\qquad$ &(\ref{eq:hproj}) \\
Remannian gradient & $\textrm{grad}f(\bs{Y})=\sum_{i=1}^{l}(C_i\bs{B}_i+C_i^*\bs{B}_i^{\sf{H}})\bs{Y}\qquad$ & (\ref{eq:finalrgrad}) \\
Remannian Hessian &
$\textrm{Hess}f(\bs{Y})[\bs{\eta}_{\bs{Y}}]\qquad$ & (\ref{eq:finalrhess}) \\
Retraction & $\mathcal{R}_{\bs{Y}}(\bs{\xi})=\pi(\bs{Y}+\overline{\bs{\xi}})
\qquad
$ &(\ref{eq:retraction})
\end{tabular}
\caption{Riemannian ingredients}
\label{tab:riemannian_ingredients}
\end{table*}
\subsubsection{Riemannian Gradient} Riemannian gradient is a necessary ingredient to develop the Riemannian conjugate gradient and Riemannian trust-region algorithm. For quotient manifold, the horizontal representation of Riemannian gradient, denoted by $\textrm{grad}f(\bs{Y})$, arises from
\begin{equation}\label{eq:rgrad}
\textrm{grad}f(\bs{Y})=\Pi_{\bs{Y}}^h \overline{\textrm{grad}}f(\bs{Y}),
\end{equation}
in which $\overline{\textrm{grad}}f(\bs{Y})$ is the Riemannian gradient in the abstract manifold $\overline{\mathcal{M}}$ at $\bs{Y}$. Note that $\overline{\textrm{grad}}f (\bs{Y})$ is given by
\begin{equation}\label{eq:gradbar}
\overline{g}_{\bs{Y}}(\overline{\textrm{grad}}f(\bs{Y}),\overline{\bs{\xi}})=\textrm{D}f(\bs{Y})[\overline{\bs{\xi}}],~\forall \overline{\bs{\xi}}\in\mathcal{T}_{\bs{Y}}\overline{\mathcal{M}},
\end{equation}
where $\textrm{D}f(\bs{Y})[\bs{\xi}]:=\lim_{t\rightarrow 0} t^{-1} \left[ f(\bs{Y}+t\bs{\xi})-f(\bs{Y})\right]$ is the directional derivative of $f$, whereas $\overline{\bs{\xi}}$ is the horizontal lift of $\bs{\xi}$. Then we conclude that
\begin{equation}\label{eq:finalrgrad}
\textrm{grad}f(\bs{Y})=\overline{\textrm{grad}}f(\bs{Y})=\sum_{i=1}^{l}(C_i\bs{B}_i+C_i^*\bs{B}_i^{\sf{H}})\bs{Y},
\end{equation}
in which $C_i=\langle \bs{B}_i,\bs{Y}\bs{Y}^{\sf{H}} \rangle-b_i$. The derivation process is described in detail in Appendix \ref{append:gradhess}.
\subsubsection{Riemannian Hessian}
For the purpose of developing a second-order algorithm, we need to think of the Riemannian Hessian as an linear operator closely connected to the directional derivative of the gradient. \textit{Riemannian connection} defines a ``directional derivative'' on the Riemannian manifold. To be specific, Euclidean directional derivative is a Riemannian connection on $\mathbb{C}^{N\times r}$. Since the quotient manifold $\mathcal{M}$
has a Riemannian metric that is invariant along the horizontal space, the Riemannian connection \cite[Proposition 5.3.4]{Absil_2009optimizationonManifolds} can be derived from
\begin{equation}
\nabla_{\bm{\eta}}\bm{\xi} = \Pi_{\bm{Y}}^{h}(\textrm{D}\overline{\bm{\xi}}[\overline{\bm{\eta}}]),
\end{equation}
for any $\bm{\eta}\in \mathcal{V}_{\bs{Y}},\; \bm{\xi}\in\mathfrak{X}(\mathcal{M})$ and $\mathfrak{X}(\mathcal{M})$ is the set of smooth vector fields on $\mathcal{M}$. The horizontal representation of Riemannian Hessian operator \cite[Definition 5.5.1]{Absil_2009optimizationonManifolds} is given as
\begin{equation}\label{eq:rhess}
\textrm{Hess}f(\bs{Y})[\bs{\xi}_{\bs{Y}}]:=\nabla_{\bs{\xi}_{\bs{Y}}}\overline{\textrm{grad}}f.
\end{equation}
Then the Riemannian Hessian is given by
\begin{align}\label{eq:finalrhess}
\textrm{Hess}f(\bs{Y})[\bs{\eta}_{\bs{Y}}]=\Pi_{\bs{Y}}^{h}\biggl(&\sum_{i=1}^{l}({C_{\bs{\eta}}}_{i}\bs{B}_i\bs{Y} +C_i\bs{B}_i\bs{\eta}_Y \nonumber\\&+ {C_{\bs{\eta}}}_{i}^*\bs{B}_i^{\sf{H}}\bs{Y} +C_i^*\bs{B}_i^{\sf{H}}\bs{\eta}_\bs{Y})\biggr),
\end{align}
where ${C_{\bs{\eta}}}_{i}=\langle \bs{B}_i,\bs{Y}\bs{\eta}_\bs{Y}^{\sf{H}} +\bs{\eta}_\bs{Y}\bs{Y}^{\sf{H}}\rangle$.
We relegate the derivation details of this expression to Appendix \ref{append:gradhess}.
\subsection{Riemannian Optimization for Fixed-Rank Problem}
Riemannian optimization generalizes the optimization algorithms in Euclidean space to a manifold. Similarly, we need to compute search directions in the tangent space and appropriate stepsizes. To ensure each iteration is always on the given manifold, \textit{retraction} \cite[Sec 4.1]{Absil_2009optimizationonManifolds} is defined as a pull-back from the tangent space onto the manifold. To be specific, the updating formula in the $i$-th iteration is given by
\begin{equation}
\bs{Y}_{k+1} = \mathcal{R}_{\bs{Y}_k}(\alpha_k\bs{\eta}_{k}),
\end{equation}
where $\alpha_k>0$ is the step size, $\bs{\eta}_k \in \mathcal{T}_{\bs{Y}_k}\mathcal{M}$ is the search direction, and $\mathcal{R}$ denotes retraction operation which maps an element from the set of all tangent spaces $\mathcal{T}\mathcal{M}=\cup_{\bs{Y}\in\mathcal{M}}\mathcal{T}_{\bs{Y}}\mathcal{M}$ to $\mathcal{M}$. The retraction operation is shown in Fig. \ref{fig:retraction}.
\begin{proposition}
Choices of $\overline{\mathcal{R}}$ and $\mathcal{R}$
\begin{equation}\label{eq:retraction}
\overline{\mathcal{R}}_{\bs{Y}}(\overline{\bs{\xi}}): = \bs{Y}+\overline{\bs{\xi}},~~
\mathcal{R}_{\bs{Y}}(\bs{\xi}):=\pi(\overline{\mathcal{R}}_{\bs{Y}}(\overline{\bs{\xi}}))
\end{equation}
define retractions on $\overline{\mathcal{M}}$ and $\mathcal{M}$, respectively.
\end{proposition}
\begin{proof}
Since $\overline{\mathcal{M}}=\mathbb{C}_*^{N\times r}$ is an embedded manifold and also an open submanifold of $\mathcal{E}=\mathbb{C}^{N\times r}$, following \cite[Sec 4.1.1]{Absil_2009optimizationonManifolds} we can choose the identity mapping
\begin{equation}
\phi(\bs{F}) = \bs{F}
\end{equation}
as a diffeomorphism so that $\phi:\overline{\mathcal{M}} \times \overline{\mathcal{N}}\rightarrow \mathcal{E}_*,\overline{\mathcal{N}}=\emptyset$ and $\textrm{dim}(\overline{\mathcal{M}})+\textrm{dim}(\overline{\mathcal{N}})=\textrm{dim}(\mathcal{E})$. Therefore, we conclude that
\begin{equation}
\overline{\mathcal{R}}_{\bs{Y}}(\overline{\bs{\xi}}): =\bs{Y}+\overline{\bs{\xi}}
\end{equation}
defines a retraction on $\overline{\mathcal{M}}$. Adding with that equivalent classes are orbits of the Lie group $\mathcal{U}_r$ which acts linearly \cite[Sec 4.1.2]{Absil_2009optimizationonManifolds} on the abstract manifold $\overline{\mathcal{M}}$,
\begin{equation}
\mathcal{R}_{\bs{Y}}(\bs{\xi}):=\pi(\overline{\mathcal{R}}_{\bs{Y}}(\overline{\bs{\xi}}))
\end{equation}
defines a retraction on $\mathcal{M}$.
\end{proof}
In this subsection, we will introduce Riemannian conjugate gradient (RCG) method and Riemannian trust-region (RTR) method.
\subsubsection{Riemannian Conjugate Gradient Method}\label{sec:RCG}
When the search direction is chosen as the negative Riemannian gradient and the step size is determined by backtracking line search following the Armijo rule \cite[4.6.3]{Absil_2009optimizationonManifolds}, we have the Riemannian gradient descent algorithm. Riemannian conjugate gradient method can be expressed as
\begin{equation}
\bs{\eta}_{k+1} = -\textrm{grad}f_k+\beta_k \mathfrak{T}_{\alpha_k\bs{\eta}_{k}}(\bs{\eta}_{k}),
\end{equation}
where $\mathfrak{T}_{\bs{\eta}_{\bs{Y}}}(\bs{\xi}_{\bs{Y}})$ is the \textit{vector transport} operator so that $\bs{\xi}_{\bs{Y}}$ is transported from $\mathcal{T}_{\bs{Y}}\mathcal{M}$ to $\mathcal{T}_{\mathcal{R}_{\bs{Y}}(\bs{\eta}_{\bs{Y}})}\mathcal{M}$ for $\bs{\xi}_{\bs{Y}}\in\mathcal{T}_{\bs{Y}}\mathcal{M}$. This is shown in Fig. \ref{fig:transport}.
\begin{figure}[t]
\centering
\subfloat[Retraction]{\includegraphics[width=2.2in]{retraction}\label{fig:retraction}}\hfil
\subfloat[Vector Transport]{\includegraphics[width=2.2in]{transport}\label{fig:transport}}
\caption{Riemannian retraction and vector transport operation.}
\end{figure}
A vector transport is defined by
\begin{equation}\label{eq:transport}
\mathfrak{T}_{\bs{\eta}_{\bs{Y}}}(\bs{\xi}_{\bs{Y}}):=D\mathcal{R}_{\bs{Y}}(\bs{\eta}_{\bs{Y}})[\bs{\xi}_{\bs{Y}}]=\bs{\xi}_{\bs{Y}}.
\end{equation}
Among many good choices for $\beta_k$, we choose
\begin{equation}
\beta_{k} =\frac{g_{\bs{Y}_k}(\textrm{grad}f_k,\textrm{grad}f_k-\mathfrak{T}_{\alpha_{k-1}\bs{\eta}_{k-1}}(\textrm{grad}f_{k-1}))}{g_{\bs{Y}_k}(\bs{\eta}_{k-1},\textrm{grad}f_k-\mathfrak{T}_{\alpha_{k-1}\bs{\eta}_{k-1}}(\textrm{grad}f_{k-1}))}, \nonumber
\end{equation}
which is a generalized version of Hestenes-Stiefel \cite{hestenes1952methods}.
\subsubsection{Riemannian Trust-Region Algorithm}\label{sec:RTR}
When the search direction is chosen by solving the local second-order approximation of $f(\bs{Y})$, it results in the Riemannian trust-region algorithm. We will find the updating vector $\bs{\eta}$ by solving the trust-region subproblem
\begin{eqnarray}
\mathop{\textrm{minimize}}_{\bs{\eta}\in\mathcal{T}_{\bs{Y}_k}\mathcal{M}}&&m_k(\bs{\eta})=f_k+g_{\bs{Y}_k}(\textrm{grad}f_k,\bs{\eta}) +\frac{1}{2}g_{\bs{Y}_k}(\textrm{Hess}f_k[\bs{\eta}],\bs{\eta}) \nonumber\\
\textrm{subject to}&&\|\bs{\eta}\|_g\leq \Delta_k, \label{prob:tr_sub}
\end{eqnarray}
where $\Delta_k>0$ is the radius of the trust region and $\|\bs{\eta}\|_g=\sqrt{g_{\bs{Y}_k}(\bs{\eta},\bs{\eta})}$. Note that the solution $\bs{\eta}$ becomes a candidate for updating. This is because we will select a proper $\Delta_k$, find the corresponding solution $\bs{\eta}_k$ of the trust-region subproblem and then update $\bs{Y}$ through
\begin{equation}
\bs{Y}_{k+1} = \mathcal{R}_{\bs{Y}_k}(\bs{\eta}_{k}).
\end{equation}
The criterion for choosing $\Delta_k$ is based on evaluating
\begin{equation}
\rho_k = \frac{f_k-f(\mathcal{R}_{\bs{Y}_k}(\bs{\eta}_k))}{m_k(\bs{0})-m_k(\bs{\eta}_k)}.
\end{equation}
When $\rho_k$ is very small, the radius of trust region $\Delta_k$ should be reduced because in this case the second-order approximation is too inaccurate. If $\rho_k$ is not very small, we shall accept $\Delta_k$ and $\bs{\eta}_k$ and reduce the trust region. If $\rho_k$ is close to 1, it m
means that the second-order approximation models original objective function well. Hence, we can accept this step and expand
the trust region. Likewise, if $\rho_k\ll 1$, we should also reject this step and increase $\Delta_k$. The trust-region subproblem can be solved by the truncated conjugate gradient \cite[Sec 7.3.2]{Absil_2009optimizationonManifolds} algorithm (see Algorithm \ref{algorithm:tcg}).
\SetNlSty{textbf}{}{:}
\IncMargin{1em}
\begin{algorithm}[tb]
\textbf{Input:} $\mathcal{B},\bm{b},\bs{Y}_k,\Delta_k$. Parameters $\kappa,\theta>0$\\
Initialize: $\bm{\eta}^{0},\bs{r}_0=\textrm{grad}f_k,\bs{\delta}_0=-\bs{r}_0$\\
\While{$\|\bs{r}_{j+1}\|_{g}>\|\bs{r}_{0}\|_{g}\min(\|\bs{r}_{0}\|_{g}^{\theta},\kappa)$}{
\If{$g_{\bs{Y}_k}(\bs{\delta}_j,\textrm{Hess}f_k[\bs{\delta}_j])\leq 0 $}
{Compute $\tau=\arg\min m_k(\bs{\eta}_k)$ where $\bs{\eta}_k=\bs{\eta}^j+\tau\bs{\delta}_j$ and $\|\bs{\eta}_k\|_g= \Delta_k$,\\
\Return $\bs{\eta}$}
Set $\bs{\eta}^{j+1}=\bs{\eta}^j+\alpha_j\bs{\delta}_j$ where $\alpha_j=\|\bs{r}_j\|_g^2/g_{\bs{Y}_k}(\bs{\delta}_j,\textrm{Hess}f_k[\bs{\delta}_j])$\\
\If{$\|\bs{\eta}^{j+1}\|_g\geq {\Delta}_k $}
{Set $\tau$ as the solution to $\|\bs{\eta}_k\|_g= \Delta_k$ where $\bs{\eta}_k=\bs{\eta}^j+\tau\bs{\delta}_j$,\\
\Return $\bs{\eta}_k$}
Set $\bs{r}_{j+1}=\bs{r}_j+\alpha_j\textrm{Hess}f_k[\bs{\delta}_j]$,\\
Set $\beta_{j+1}=\|\bs{r}_{j+1}\|_{g}^2/\|\bs{r}_{j}\|_{g}^2$,\\
Set $\bs{\delta}_{j+1}=-\bs{r}_{j+1}+\beta_{j+1}\bs{\delta}_j$,\\
$j\leftarrow j+1$
}
\textbf{Output:} $\bs{\eta}=\bs{\eta}_k$.
\caption{Truncated Conjugate Gradient Algorithm for (\ref{prob:tr_sub})}
\label{algorithm:tcg}
\end{algorithm}
Riemannian trust-region algorithm harnesses the second-order information of the problem. It admits a superlinear \cite[Theorem 7.4.11]{Absil_2009optimizationonManifolds} convergence rate locally and is robust to initial points. Since the objective function $f$ is exactly a quadratic function which satisfies the Lipschitz gradient condition and other assumptions in \cite{boumal2016global}, we can always find an approximate second-order critical points by the Riemannian trust-region algorithm.
\subsection{Computational Complexity Analysis}
Riemannian conjugate gradient and Riemannian trust-region algorithm involve computing optimization ingredients at each iteration, for which we show their computational complexity.
\begin{itemize}
\item Evaluate the objective value $f(\bs{Y})$. Since $\mathcal{B}(\cdot)$ involves a series of sparse matrix multiplication, we can compute it efficiently and the complexity of computing $f(\bs{Y})$ is $\mathcal{O}(mnl)$.
\item Compute the Riemannian gradient $\textrm{grad}f$ (\ref{eq:finalrgrad}). This includes computing matrix multiplication in $\mathcal{O}(mnl)$ and horizontal projection $\Pi_{\bs{Y}}^h$ (\ref{eq:hproj}). Since complexity of solving the Lyapunov equation (\ref{eq:lyap}) is $\mathcal{O}(r^3+(m+n)r^2)$, the overall complexity is $\mathcal{O}(mnl+(m+n)r^2+r^3)$.
\item Compute the Riemannian Hessian $\textrm{Hess}f$ (\ref{eq:finalrhess}). Its cost is also $\mathcal{O}(mnl+(m+n)r^2+r^3)$.
\item Computing the Riemannian metric $\overline{g}$ (\ref{eq:metric}). This complexity is dominant by matrix multiplications, which is $\mathcal{O}((m+n)r^2)$.
\item Computational complexity of retraction $\overline{\mathcal{R}}$ (\ref{eq:retraction}) is $\mathcal{O}((m+n)r)$ and vector transport $\mathfrak{T}$ (\ref{eq:transport}) is insignificant.
\end{itemize}
From the above results, we conclude that the computational complexity of Riemannian conjugate gradient algorithm for each iteration is $\mathcal{O}(mnl+(m+n)r^2+r^3)$. And each iteration of truncated conjugate gradient algorithm also involves computation with complexity $\mathcal{O}(mnl+(m+n)r^2+r^3)$.
\section{Simulations} \label{sec:simulation}
This section presents numerical experiments to demonstrate the efficacy of the generalized low-rank optimization approach for topological cooperation via the newly presented Riemannian optimization algorithms. We will investigate the performance of different algorithms from the perspective of convergence rate and achievable DoF. We evaluate our model
in different settings and demonstrate that the generalized low-rank approach can effectively enable transmitter cooperation based
only on network topology information.
Our simulations compare the following matrix-factorization-based algorithms
for solving the generalized low-rank optimization problem $\mathscr{P}$:
\begin{itemize}
\item ``AltMin'' \cite{yi2016topological}: Alternating minimization algorithm (\ref{GLRM_fixedrank_altmin1}) (\ref{GLRM_fixedrank_altmin2}) is adopted in \cite{yi2016topological} for topological transmitter cooperation problem by alternatively updating factors. For fixed $r$, (\ref{GLRM_fixedrank_altmin1}) (\ref{GLRM_fixedrank_altmin2}) are solved with gradient descent followed by backtracking line search.
\item ``RCG'': The Riemannian conjugate gradient method is developed in Sec \ref{sec:RCG}. We implement this algorithm with \textit{Manopt} \cite{boumal2014manopt} software package.
\item ``RTR'': The Riemannian trust-region method is developed in Sec \ref{sec:RTR} and also implemented with \textit{Manopt}.
\end{itemize}
All algorithm are adopted with random initialization strategy for each rank $r$, and we find the minimal $r$ by increasing $r$ from 1 to $N$ until ${m}^{-0.5}\cdot\|\mathcal{A}(\bs{X})-\bs{b}\|<10^{-3}$. In our numerical experiments, the network topology and shared messages at each transmitter are generated uniformly at random with probabilities
\begin{equation}
\textrm{Prob}((k,j)\in\mathcal{E})=\left\{\begin{aligned}
& p, && j\ne k \\
& 1, && j=k
\end{aligned}\right.,
\end{equation}
and
\begin{equation}
\textrm{Prob}(j\in\mathcal{S}_k)=\left\{\begin{aligned}
& q, && j\ne k \\
& 1, && j=k
\end{aligned}\right.,
\end{equation}
respectively.
\begin{figure}[h]
\centering
\subfloat[]{\includegraphics[width=\columnwidth]{convergence_iter}}\hfil
\subfloat[]{\includegraphics[width=\columnwidth]{convergence_time}}
\caption{Convergence rate and computing time of all algorithms with full transmitter cooperation.}
\label{fig:convergence}
\end{figure}
\subsection{Convergence Rate}
Consider a partially connected $20$-user interference channel with full transmitter cooperation. The network topology is generated randomly and each link is connected with probability $p=0.3$. Each message is split into $3$ data streams. In this simulation, $K=20, d_1=\cdots=d_K=3$, $p=0.3$ and $r=12$. Fig. \ref{fig:convergence} shows
the convergence behaviors of all 3 algorithms in terms of
iterations and time.
The results indicate that the proposed RTR algorithm
exhibits a superlinear convergence rate, and the computing
rate is comparable with first order RCG algorithm.
In addition, the proposed RTR can yield a more accurate
solution with the second-order stationary point when
compared against first order algorithms that guarantee convergence
only to first-order stationary points.
The overall test results show that the proposed RTR and RCG
algorithms are much more efficient than other contemporary algorithms
in terms of convergence rates and solution performance.
To further show that the interferences are nulled, we choose the interference leakage as the metric and plot it in Fig. \ref{fig:leakage} for the same setting of Fig. \ref{fig:convergence}. The interference leakage cost is given by
\begin{equation}
IL = \sum_{i\ne k}\sum_{j:(k,j)\in\mathcal{E},i\in\mathcal{S}_j}\!\!\!\! \!\!\! \|h_{kj}\bs{U}_{k}^{\sf{H}}\bs{V}_{ji}\Big(h_{kj}\bs{U}_{k}^{\sf{H}}\bs{V}_{ji}\Big)^{\sf{H}}\|_F^2,
\end{equation}
where the channel coefficients follow standard complex Gaussian distribution. Fig. \ref{fig:leakage} demonstrates that there is a rapid decline of interference leakage as the objective value decreases with the proposed Riemannian optimization algorithms.
\begin{figure}[h]
\centering
\subfloat[]{\includegraphics[width=\columnwidth]{leakage_iter}}\hfil
\subfloat[]{\includegraphics[width=\columnwidth]{leakage_time}}
\caption{Convergence of interference leakage for all algorithms with full transmitter cooperation.}
\label{fig:leakage}
\end{figure}
\subsection{DoF over Network Topologies}
Consider a partially connected $20$-user interference channel without message splitting ($d_k=1$). The network topologies are generated randomly with different $p$. Fig. \ref{fig:connectedlinks} demonstrates the DoF over $p$ with full transmitter cooperation. Each DoF result
is averaged over 100 times. This result shows that, among the 3
solutions, the
proposed RTR algorithm achieves the best performance with
second-order stationary points.
The Riemannian algorithms RTR and RCG significantly outperform the alternating minimization algorithm owing to their good convergence guarantee.
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{topology}
\caption{DoF over the number of connected links.}
\label{fig:connectedlinks}
\end{figure}
To further justify the effectiveness of the Riemannian optimization framework, we check the recovered DoF returned by the proposed RTR algorithm for all specific network topologies $\mathcal{E}$ and specific message sharing pattern $\{\mathcal{S}_k:k\in[K]\}$ provided in \cite{yi2015topological}. Specifically, transmitter cooperation improves the symmetric DoF from $1/3$ to $2/5$ for Example 1 in Fig. 1(a), from $2/5$ to $1/2$ for Example 4 in Fig. 3(a), and from $1/3$ to $2/5$ for Example 7 in Fig. 6(a), compared with the cases without cooperation. And the optimal symmetric DoF is $1/2$ for Example 6 in Fig. 5(a) with transmitter cooperation. All these optimal symmetric DoF results can be achieved by the proposed RTR algorithm numerically. However, theoretically identifying the network topologies and the message sharing patterns for which the Riemannian trust region algorithm can provide optimal symmetric DoFs is still a challenging open problem.
\subsection{Transmitter Cooperation Gains}
We investigate the achievable DoFs in partially connected $20$-user interference channels. We randomly generate the network topologies with $p=0.2$ and simulate different algorithms under different transmitter cooperation level $q$ with single data stream. For each cooperation level, we take average over 500 channel realizations.
Fig.~\ref{fig:cooperationlevel} shows that the
second-order algorithm RTR can achieve
the highest DoF among all algorithms.
Comparing the first-order algorithms, the proposed RCG
outperforms AltMin. With the high convergence rate and second-order stationary points solutions of RTR, the gap between RTR algorithm and other algorithms grows with $q$, which indicates
that the proposed RTR algorithm is capable
of fully leveraging the benefits of transmitter cooperation.
\begin{figure}[h]
\centering
\subfloat[]{\includegraphics[width=\columnwidth]{cooperation}\label{fig:cooperationlevel}}\hfil
\subfloat[]{\includegraphics[width=\columnwidth]{sumrate}\label{fig:sumrate}}
\caption{Benefits of transmitter cooperation: (a) DoF over different transmitters cooperation levels $q$. (b) Sum-rate over the transmit power.}
\end{figure}
To further illustrate the transmitter cooperation benefit, we evaluate the achievable sum-rate using the proposed RTR algorithm in
Fig.~\ref{fig:cooperationlevel}. In this single data stream test
setting, $\bs{U}_k,\bs{V}_{ji}$ degenerates to vectors
$\bs{u}_k,\bs{v}_{ji}$. Assume that each single data stream
symbol $s_i$ has unit power, i.e., $\mathbb{E}(|s_i|^2)=1$.
Suppose the noise is i.i.d. Gaussian, i.e., $\bs{\Sigma}_k = \sigma^2\bs{I}_r$, with $\sigma^2$ at $-120$dB. The distance $d_{ij}$ between each connected transmitter-receiver pair $(j,i)$ is uniformly distributed in $[0.1, 0.2]$ km. The fading channel model is given as
\begin{equation}
h_{ij} = 10^{-L(d_{ij})/20}c_{ij}, (i,j)\in\mathcal{E},
\end{equation}
where the pass loss is given by $L(d_{ij}) = 128.1+37.6\log_{10}d_{ij}$ and the small scale fading coefficient is given by $c_{ij}\sim\mathcal{CN}(0,1)$. Then the sum rate per channel use is given by
\begin{equation}
C_{\text{sum}}=\frac{1}{r}\sum_{k=1}^{K}C_k = \frac{1}{r}\sum_{k=1}^{K}\log(1+\textrm{SINR}_k),
\end{equation}
where
\begin{equation*}
\textrm{SINR}_k = \frac{|\sum_{(k,j)\in\mathcal{E},k\in\mathcal{S}_j}h_{kj}\bs{u}_k^{\sf{H}}\bs{v}_{jk}|^2}{\sum_{i\ne k}|\sum_{(k,j)\in\mathcal{E},i\in\mathcal{S}_j}h_{kj}\bs{u}_k^{\sf{H}}\bs{v}_{ji}|^2+\|\bs{u}_k\|_2^2\sigma^2}.
\end{equation*}
Based on the singular value decomposition (SVD) $\bs{X}^*=\bs{U}\bs{\Sigma}\bs{V}^{\sf{H}}$, the transmit beamformer is simply chosen as $\bs{U}\bs{\Sigma}^{\frac{1}{2}}$, and receive beamformer is given by $\bs{V}\bs{\Sigma}^{\frac{1}{2}}$ with power normalization. Fig. \ref{fig:sumrate} shows the achievable sum-rate over different transmit power $P$. Each point is averaged
across 100 channel realizations with random
$\mathcal{E}$ ($p=0.2$). The result also demonstrates that the
proposed RTR algorithm is capable of achieving high data
rates by leveraging transmitter cooperation.
In summary, our numerical experiments demonstrate that the
proposed Riemannian trust-region algorithm is capable of obtaining high-precision solutions with second-order stationary points,
leading to high achievable DoFs and data rates.
Furthermore, the computation time of RTR algorithm in medium scale problems is comparable with first order algorithms, which demonstrates that the proposed RTR algorithm is a powerful algorithm capable of
harnessing the benefit of topological cooperation in
problems involving medium network sizes.
\section{Conclusions} \label{sec:conclusion}
This work investigates the opportunities of transmitter
cooperation based only on topological information
with message sharing. Our contributions include the
derivation of a generalized topological interference alignment condition, followed by the development of
a low-rank matrix optimization approach to maximize the achievable DoFs.
To solve the resulting generalized low-rank optimization problem
which is nonconvex in complex field, we developed Riemannian
optimization algorithms by exploiting the complex non-compact Stiefel manifold for fixed-rank matrices in complex field. In particular, we
adopted
the semidefinite lifting technique and Burer-Monteiro factorization approach. Our experiments demonstrated that the proposed Riemannian algorithms considerably outperformed the
alternating minimization algorithm. Additionally,
the proposed Riemannian trust-region algorithm achieves high DoFs with high-precision
second-order stationary point solutions, with computation
complexity comparable with the first-order Riemannian
conjugate gradient algorithm.
\appendices
\section{Proof of Proposition \ref{proposition:nuc}}\label{append:nuc}
For simplicity, we only give the proof of the single data stream case, while it can be readily extended to general multiple data stream cases. In this case, $\bs{X}$ is given by
\begin{equation}
\bs{X} = \begin{bmatrix}
\bs{u}_1^{\sf{H}}\bs{v}_{11} & \cdots & \bs{u}_1^{\sf{H}}\bs{v}_{KK} \\
\vdots & \ddots & \vdots \\
\bs{u}_K^{\sf{H}}\bs{v}_{11} & \cdots & \bs{u}_K^{\sf{H}}\bs{v}_{KK}
\end{bmatrix}\in\mathbb{C}^{K\times K^2}.
\end{equation}
Let $\bs{x}_i$ denote the transpose of the $i$-th row of matrix $\bs{X}$. Problem (\ref{prob:nuc}) can be rewritten as
\begin{eqnarray}\label{prob:re_nuc}
\mathop{\textrm{minimize}}_{{\bs{X}}\in\mathbb{C}^{m\times n}}&& \|{\bs{X}}\|_* \nonumber\\
\textrm{subject to}&& \bm{1}^{\sf{H}}\bs{x}_k^{\mathcal{D}_k}=1,~\forall k\in [K] \nonumber\\
&&\bs{x}_{k}^{\mathcal{G}_k} = \bs{0},~\forall k\in [K],
\end{eqnarray}
in which $\bs{x}_k^{\mathcal{D}_k},\bs{x}_k^{\mathcal{G}_k}$ are vectors whose elements are sampled from $\bs{x}_k$, and $\mathcal{D}_k,\mathcal{G}_k$ are the index sets of sampling. $\mathcal{D}_k$ and $\mathcal{G}_k$ are given by
\begin{equation}
\mathcal{D}_k = \{(j-1)*K+k:(k,j)\in\mathcal{E},k\in\mathcal{S}_j\}
\end{equation}
and
\begin{equation}
\mathcal{G}_k = \{(j-1)*K+i:i\ne k, (k,j)\in\mathcal{E},i\in\mathcal{S}_j\},
\end{equation}
respectively.
\begin{lemma}\label{prop:optimal}
The optimal solution of (\ref{prob:re_nuc}), denoted by $\bs{X}^\star$, is given by
\begin{equation}
{\bs{x}^\star}_k^{\mathcal{D}_k} = \frac{1}{|\mathcal{D}_k|}{\bs{1}},
\end{equation}
where $\bs{1}$ denotes the vector of all ones. The remaining entries of $\bs{X}^\star$ are all zeros.
\end{lemma}
\begin{proof}
Let $g(\bs{X})=\|\bs{X}\|_*=\textrm{Tr}(\sqrt{\bs{X}\bs{X}^{\sf{H}}})$. To proof Lemma \ref{prop:optimal}, it is equivalent to prove that $t=0$ is a minimum of the convex function
\begin{equation}
h(t) = \|\bs{X}^\star+t\bs{X}\|_*,\mathcal{A}(\bs{X}^\star+t\bs{X})=\bs{b},t\in\mathbb{R}.
\end{equation}
This can be deduced from the fact that any feasible point can be expressed as $\bs{X}^\star+t\bs{X}$, and if $t=0$ is a minimum of $h(t)$, then $g(\bs{X}^\star)\leq g(\bs{X})$ always holds.
Based on the structure of $\mathcal{D}_k$ and $\mathcal{G}_k$, we have
\begin{equation}\label{eq:proof2}
\bs{X}^\star{\bs{X}^\star}^{\sf{H}}=\begin{bmatrix}
\frac{1}{|\mathcal{D}_1|} & 0 & \cdots & 0 \\
0 & \frac{1}{|\mathcal{D}_2|} & \cdots & 0 \\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 & \cdots & \frac{1}{|\mathcal{D}_K|}
\end{bmatrix}.
\end{equation}
Then $\bs{X}^\star$ is a full rank matrix, and we can find $|t|\leq \epsilon$ such that $\bs{X}^\star+t\bs{X}$ is invertible. Therefore, the derivative of $h$ is given by
\begin{align}
h^\prime(t)= \frac{1}{2}\big\langle &\left((\bs{X}^\star+t\bs{X})(\bs{X}^\star+t\bs{X})^{\sf{H}}\right)^{-\frac{1}{2}}, \nonumber\\&\bs{X}{\bs{X}^\star}^{\sf{H}}+\bs{X}^\star\bs{X}^{\sf{H}}+2t\bs{X}\bs{X}^{\sf{H}} \big\rangle.
\end{align}
Then we have
\begin{equation}
h^\prime(0)=\frac{1}{2}\big\langle \left(\bs{X}^\star{\bs{X}^\star}^{\sf{H}}\right)^{-\frac{1}{2}},\bs{X}{\bs{X}^\star}^{\sf{H}}+\bs{X}^\star\bs{X}^{\sf{H}} \big\rangle.
\end{equation}
Since $ \mathcal{A}(\bs{X}^\star+t\bs{X})=\bs{b}$, i.e.,
\begin{equation}
\bm{1}^{\sf{H}}({\bs{x}^\star}_k^{\mathcal{D}_k}+t\bs{x}_k^{\mathcal{D}_k}) = 1
\end{equation}
then we have $\langle{\bs{x}^\star}_k^{\mathcal{D}_k},\bs{x}_k^{\mathcal{D}_k}\rangle = 0$. Therefore,
\begin{equation}\label{eq:proof1}
\textrm{diag}(\bs{X}{\bs{X}^\star}^{\sf{H}}) = \textrm{diag}(\bs{X}^\star\bs{X}^{\sf{H}}) = \bs{0}.
\end{equation}
From (\ref{eq:proof1}) (\ref{eq:proof2}), we can deduce that $h^\prime(0)=0$, and thus $t=0$ is a minimum of $h(t)$.
\end{proof}
From Lemma \ref{prop:optimal} we know that the optimal solution $\bs{X}^{\star}$ of (\ref{prob:nuc}) is full rank. So nuclear norm relaxation approach always fails.
\section{Proof of Proposition \ref{proposition:vertical}}\label{append:vertical}
The elements in the vertical space $\mathcal{V}_{\bs{Y}}$ must be tangential to the equivalent class $[\bs{Y}]=\{\bs{Y}\bs{Q}:\bs{Q}^{\sf{H}}\bs{Q}=\bs{I}\}$. Let $\bs{Y}(t)=\bs{Y}_0\bs{Q}(t)$ be a curve in $[\bs{Y}_0]$ through $\bs{Y}_0$ at $t=0$, i.e., $\bs{Q}(0)=\bs{I}$. Then we have
\begin{equation}\label{eq:tangentcurve}
\bs{Y}(t)\bs{Y}(t)^{\sf{H}}=\bs{Y}_0\bs{Q}(t)\bs{Q}(t)^{\sf{H}}\bs{Y}_0^{\sf{H}}=\bs{Y}_0\bs{Y}_0^{\sf{H}}.
\end{equation}
By differentiating (\ref{eq:tangentcurve}) we get
\begin{equation}\label{eq:tangentcurve2}
\dot{\bs{Y}}(t)\bs{Y}(t)^{\sf{H}}+\bs{Y}(t)\dot{\bs{Y}}(t)^{\sf{H}}=\bs{0}.
\end{equation}
So we deduce that $\dot{\bs{Y}}(0)$ belongs to
\begin{equation}\label{eq:vertical1}
\{\bs{Z}\in\mathbb{C}^{N\times r}: \bs{Z}\bs{Y}_0^{\sf{H}}+\bs{Y}_0\bs{Z}^{\sf{H}}=\bs{0}\},
\end{equation}
of which $\mathcal{T}_{\bs{Y}}\mathcal{M}$ is a subset. On the other side, let $F:\bs{Y}\mapsto \bs{Y}\bs{Y}^{\sf{H}}$, then (\ref{eq:vertical1}) is $\textrm{ker}(\textrm{D}F(\bs{Y_0}))$ and $F^{-1}(\bs{Y}_0\bs{Y}_0^{\sf{H}})=[\bs{Y}_0]$. Therefore, from \cite[Sec 3.5.7]{Absil_2009optimizationonManifolds} we know
\begin{equation}
\mathcal{T}_{\bs{Y}}\mathcal{M}=\{\bs{Z}\in\mathbb{C}^{N\times r}: \bs{Z}\bs{Y}_0^{\sf{H}}+\bs{Y}_0\bs{Z}^{\sf{H}}=\bs{0}\}.
\end{equation}
Without loss of generality, we can set
\begin{equation}\label{eq:tangentcurve3}
\dot{\bs{Y}}(t) = \bs{Y}(t)\bs{\Omega}(t),~~\bs{\Omega}(t)\in\mathbb{C}^{r\times r},
\end{equation}
since $\bs{Y}(t)\in\mathbb{C}_*^{N\times r}$ is full rank. Then we can replace equation (\ref{eq:tangentcurve3}) in (\ref{eq:tangentcurve2}) and obtain
\begin{equation}
\bs{Y}(t)(\bs{\Omega}(t)+\bs{\Omega}(t)^{\sf{H}})\bs{Y}(t)^{\sf{H}}=\bs{0}.
\end{equation}
Therefore, the vertical space is given by
\begin{equation}
\mathcal{V}_{\bs{Y}}=\{\bs{Y}_0\bs{\Omega}:\bs{\Omega}^{\sf{H}}=-\bs{\Omega}\}.
\end{equation}
\section{Proof of Proposition \ref{proposition:horizontal}}\label{append:horizontal}
The horizontal space is given by
\begin{equation}
\mathcal{H}_{\bs{Y}} = \{\overline{\bs{\xi}}\in\mathcal{T}_{\bs{Y}}\overline{\mathcal{M}}:\overline{g}_{\bs{Y}}(\overline{\bs{\xi}},\overline{\bs{\zeta}})=0,~\forall \overline{\bs{\zeta}}\in\mathcal{V}_{\bs{Y}}\},
\end{equation}
that is
\begin{equation}
\overline{g}_{\bs{Y}}(\overline{\bs{\xi}},\bs{Y}\bs{\Omega})=0, \forall \bs{\Omega}^{\sf{H}}=-\bs{\Omega}.
\end{equation}
Since
\begin{align}
\overline{g}_{\bs{Y}}(\overline{\bs{\xi}},\bs{Y}\bs{\Omega})&={\textrm{Tr}}(\overline{\bs{\xi}}^{\sf{H}}\bs{Y}\bs{\Omega}+\bs{\Omega}^{\sf{H}}\bs{Y}^{\sf{H}}\overline{\bs{\xi}}) \nonumber\\
&={\textrm{Tr}}(\overline{\bs{\xi}}^{\sf{H}}\bs{Y}\bs{\Omega}-\bs{\Omega}\bs{Y}^{\sf{H}}\overline{\bs{\xi}}) \nonumber\\
&={\textrm{Tr}}((\overline{\bs{\xi}}^{\sf{H}}\bs{Y}-\bs{Y}^{\sf{H}}\overline{\bs{\xi}})\bs{\Omega}),
\end{align}
we know the horizontal space consists of all the elements $\overline{\bs{\xi}}$ that satisfies ${\textrm{Tr}}((\overline{\bs{\xi}}^{\sf{H}}\bs{Y}-\bs{Y}^{\sf{H}}\overline{\bs{\xi}})\bs{\Omega})=0$ for all $\bs{\Omega}^{\sf{H}}=-\bs{\Omega}$. Therefore, the horizontal space is
\begin{equation}
\mathcal{H}_{\bs{Y}}=\{\bs{\xi}\in\mathbb{C}^{N\times r}:\bs{\xi}^{\sf{H}}\bs{Y}=\bs{Y}^{\sf{H}}\bs{\xi}\}.
\end{equation}
Suppose for a vecor $\bs{\xi}\in\mathcal{T}_{\bs{Y}}\mathcal{M}$ its projection onto the vertical space is given by $\bs{\xi}^{v}=\bs{Y}\bs{\Omega}_{\bs{\xi}}$, then the horizontal projection is given by $\bs{\xi}^{h}=\bs{\xi}-\bs{Y}\bs{\Omega}_{\bs{\xi}}$, and
\begin{equation}
{\bs{\xi}^{h}}^{\sf{H}}\bs{Y}=\bs{Y}^{\sf{H}}{\bs{\xi}^{h}}.
\end{equation}
So we can find the $\bs{\Omega}_{\bs{\xi}}$ from
\begin{align}
&(\bs{\xi}-\bs{Y}\bs{\Omega}_{\bs{\xi}})^{\sf{H}}\bs{Y}=\bs{Y}^{\sf{H}}(\bs{\xi}-\bs{Y}\bs{\Omega}_{\bs{\xi}}) \nonumber\\\Rightarrow ~& \bs{Y}^{\sf{H}}\bs{Y}\bs{\Omega}_{\bs{\xi}}+\bs{\Omega}_{\bs{\xi}}\bs{Y}^{\sf{H}}\bs{Y} = \bs{Y}^{\sf{H}}\bs{\xi}-\bs{\xi}^{\sf{H}}\bs{Y}.
\end{align}
Then we conclude that the horizontal projection of $\bs{\xi}$ is given by
\begin{equation}
\Pi_{\bs{Y}}^{h}\bs{\xi}=\bs{\xi}-\bs{Y}\bs{\Omega},
\end{equation}
where $\bs{\Omega}$ is the solution to the Lyapunov equation
\begin{equation}
\bs{Y}^{\sf{H}}\bs{Y}\bs{\Omega}+\bs{\Omega}\bs{Y}^{\sf{H}}\bs{Y} = \bs{Y}^{\sf{H}}\bs{\xi}-\bs{\xi}^{\sf{H}}\bs{Y}.
\end{equation}
\section{Computing the Riemannian Gradient and Hessian}\label{append:gradhess}
We first rewrite the objective function of (\ref{GLRM_BM}) as
\begin{equation}
f(\bs{Y}) = \frac{1}{2}\sum_{i=1}^{l}|\langle \bs{B}_i,\bs{Y}\bs{Y}^{\sf{H}} \rangle-b_i|^2.
\end{equation}
The complex gradient of $f(\bs{Y})$ is given by
\begin{align}
f^\prime(\bs{Y}) &= \sum_{i=1}^{l} (\langle \bs{B}_i,\bs{Y}\bs{Y}^{\sf{H}} \rangle-b_i)\bs{B}_i\bs{Y} +(\langle \bs{B}_i^{\sf{H}},\bs{Y}\bs{Y}^{\sf{H}} \rangle-b_i^*)\bs{B}_i^{\sf{H}}\bs{Y} \nonumber \\
&=\sum_{i=1}^{l}(C_i\bs{B}_i+C_i^*\bs{B}_i^{\sf{H}})\bs{Y},
\end{align}
in which $C_i=\langle \bs{B}_i,\bs{Y}\bs{Y}^{\sf{H}} \rangle-b_i$. The Riemannian gradient $\overline{\textrm{grad}}f(\bs{Y})$ is derived from (\ref{eq:gradbar}), and we find that
\begin{align}
\textrm{D}f(\bs{Y})[\bs{\xi}] &= \frac{1}{2}\sum_{i=1}^{l} \langle \bs{B}_i,\bs{\xi}\bs{Y}^{\sf{H}}+\bs{Y}\bs{\xi}^{\sf{H}} \rangle^*(\langle \bs{B}_i,\bs{Y}\bs{Y}^{\sf{H}}\rangle-b_i) \nonumber\\&\qquad+(\langle \bs{B}_i,\bs{Y}\bs{Y}^{\sf{H}}\rangle-b_i)^*\langle \bs{B}_i,\bs{\xi}\bs{Y}^{\sf{H}}+\bs{Y}\bs{\xi}^{\sf{H}} \rangle \nonumber\\
&=\overline{g}_{\bs{Y}}((C_i\bs{B}_i+C_i^*\bs{B}_i^{\sf{H}})\bs{Y},\bs{\xi}).
\end{align}
Therefore, $\overline{\textrm{grad}}f(\bs{Y})=f^\prime(\bs{Y})$. Then we observe that $\overline{\textrm{grad}}f(\bs{Y})^{\sf{H}}\bs{Y}=\bs{Y}^{\sf{H}}\overline{\textrm{grad}}f(\bs{Y})$, i.e., $\overline{\textrm{grad}}f(\bs{Y})$ is already in the horizontal space $\mathcal{V}_{\bs{Y}}$. So the horizontal representation of Riemannian gradient is given by
\begin{equation}
\textrm{grad}f(\bs{Y})=\sum_{i=1}^{l}(C_i\bs{B}_i+C_i^*\bs{B}_i^{\sf{H}})\bs{Y}.
\end{equation}
To derive the Riemannian Hessian (\ref{eq:rhess}), we compute
\begin{align}
\textrm{D}\overline{\textrm{grad}}f(\bs{Y})[\bs{\eta}_{\bs{Y}}]=\sum_{i=1}^{l}(&{C_{\bs{\eta}}}_{i}\bs{B}_i\bs{Y} +C_i\bs{B}_i\bs{\eta}_Y \nonumber\\&
+ {C_{\bs{\eta}}}_{i}^*\bs{B}_i^{\sf{H}}\bs{Y} +C_i^*\bs{B}_i^{\sf{H}}\bs{\eta}_\bs{Y}),
\end{align}
where ${C_{\bs{\eta}}}_{i}=\langle \bs{B}_i,\bs{Y}\bs{\eta}_\bs{Y}^{\sf{H}} +\bs{\eta}_\bs{Y}\bs{Y}^{\sf{H}}\rangle$.
We conclude that
\begin{align}
\textrm{Hess}f(\bs{Y})[\bs{\eta}_{\bs{Y}}]=\Pi_{\bs{Y}}^{h}\biggl(&\sum_{i=1}^{l}({C_{\bs{\eta}}}_{i}\bs{B}_i\bs{Y} +C_i\bs{B}_i\bs{\eta}_Y \nonumber\\&+ {C_{\bs{\eta}}}_{i}^*\bs{B}_i^{\sf{H}}\bs{Y} +C_i^*\bs{B}_i^{\sf{H}}\bs{\eta}_\bs{Y})\biggr).
\end{align}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The development of factorization method \cite{fac}
closely related to the method of Darboux transformations \cite{Darboux,MatSal} and
supersymmetric mechanics \cite{Witten,Cooper}
allows essentially extend the list of
exactly solvable quantum mechanical systems.
A particular class of the systems which can be solved following this approach
are given by multi-soliton reflectionless potentials \cite{reflectionless},
the Calogero model \cite{Calogero1,Calogero2,Perel}
and its $\mathcal{P}\mathcal{T}$-regularizations \cite{PT1}.
All these quantum mechanical systems are intimately
related with completely integrable classical field theories. Another
class of such systems corresponds to rational deformations (extensions)
of the quantum harmonic oscillator (QHO) and conformal mechanics model
of de Alfaro, Fubini and Furlan (AFF)
\cite{AFF,WeiJor,Gomez,Grandati2,CarPly,CarPly2,CIP}.
This extremely large variety of new exactly solvable systems
is very interesting in the light of hidden symmetries study
\cite{Cariglia}.
Each of them possesses some
peculiar
characteristics,
which can appropriately be described in terms of
higher order, sometimes explicitly depending on time,
integrals of motion.
These new families of solvable systems can be generated from
the free particle, the harmonic oscillator and the AFF model
which, together with their supersymmetric extensions,
are characterized by (super)conformal symmetry \cite{SCM1,SCM2,IvaKriLev2,SCM3,RJAA,Olaf}.
This symmetry plays important role in the description of various physical
aspects and phenomena
such as non-relativistic holography and AdS/CFT correspondence
\cite{DTson,KBJM,MT,CDPH},
geodesic motion in background of charged black holes \cite{CDKKTV,AIPT,GibTow},
quark confinement \cite{BTGE}, and Bose-Einstein condensates
\cite{BEC1,BEC2},
to name a few.
For quantum systems with multi-soliton potentials
and those corresponding to
rational extensions of the harmonic oscillator and AFF models,
the whole picture is complicated by the presence of hidden symmetries.
For example, the spectrum of multi-soliton potential systems
is divided into finite number of bound states and a
continuous part.
These spectral peculiarities together with reflectionless nature of the systems are reflected by the
higher order Lax-Novikov integral of motion.
The presence of this integral
leads to
extension of the $\mathcal{N}=2$ super-Poincar\'e symmetry
of such systems to the exotic nonlinear $\mathcal{N}=4$
supersymmetry, in which the Lax-Novikov integrals of
the super-partners compose the central charge
of superalgebra \cite{CJNP,AraMatPly,AraPly}.
Another example where the Lax-Novikov integral plays important
role is provided by the Calogero model \cite{CorLecPly}.
In the case of its $\mathcal{PT}$-regularized two-particle version
being perfectly invisible zero-gap quantum system,
this higher order integral generates a nonlinear extension
of the (super)conformal algebra \cite{JM1,JM2}.
On the other hand, any exactly solvable system belonging to the
family of the QHO and AFF systems with rationally extended potentials has an
infinite discrete spectrum with a finite (possibly zero) number of missing energy levels
in its lower part. The infinite equidistant tower of discrete states in such systems
can be considered as analog of the conduction band in the spectrum
of finite-gap quantum systems, while grouped and separated by gaps discrete levels
in the lower part of the spectrum can be treated as analogs of the valence bands.
Instead of Lax-Novikov integral, any such a system is characterized by the presence
of at least two pairs of higher order ladder operators which form
the complete spectrum generating sets: they allow to obtain an arbitrary physical state from
another \cite{CarPly2,CIP}. These higher order ladder operators also
encode the information on the number of valence bands (or, number of gaps in the spectrum),
as well as the total number of states in them. It also was observed
that each separate pair of these ladder operators generate
a non-linear deformation of the conformal algebra \cite{CIP}.
Having this observation in mind, the problem on which we work in this paper is to find the complete
spectrum generating algebra for this family of the systems.
Nonlinear conformal algebras
appeared earlier in the context of $W$ symmetries \cite{JYN,W}.
Symmetries described by finite $W$ algebras
appear, particularly, in anisotropic harmonic oscillator systems with commensurable
frequencies \cite{BDKL}, and in the Coulomb problem in spaces of
constant curvature, that corresponds to the so called Higgs oscillator and its
generalizations \cite{Higgs,Zhe,Evnin}.
In this work, we are dealing with rationally extended Hamiltonians
which can be obtained via factorization method
based on the iterative Darboux-Crum-Krein-Adler (DCKA)
mapping \cite{Darboux,MatSal,Crum,Krein,Adler}
applied to the QHO eigenstates.
In particular,
we are interested in those cases in which two different
selections of seed states produce the same, but
modulo a nonzero discrete shift, system.
One choice of the seed states for DCKA transformation
consists in selecting only physical eigenstates,
while the second choice corresponds to selection
of a certain ``complementary" set of non-physical states
of harmonic oscillator with negative energy
produced by the
spatial Wick rotation
$x\rightarrow ix$ of physical states.
This is what we name the Darboux duality property \cite{CIP}.
As is well known, any Darboux (or DCKA) transformation allows to construct
the $\mathcal{N}=2$ superextended system described
by linear (or nonlinear) Poincar\'e superalgebra.
The presence of another such a transformation with the described properties
allows us to extend superalgebra generally
to a nonlinearly extended superconformal algebra,
which, via the repeated (anti)-commutation relations,
creates the sets of the ladder operators of the partner systems
which compose new bosonic generators, and new
fermionic generators of the superalgebra. In this way
we finally obtain the closed spectrum generating
nonlinear superalgebra of the $\mathcal{N}=2$ superextended system.
The resulting closed nonlinear superalgebra contains generally
several different copies of either deformed superconformal $\mathfrak{osp}(2|2)$ algebra
in the case of superextended rationally deformed AFF models, or
deformed super-Schr\"odinger algebra in the case of super-extension
of rationally deformed harmonic oscillator systems.
The rest of the paper is organized as follows.
In Section \ref{Dar}, we recall how in general case
nonlinear $\mathcal{N}=2$ super-Poincar\'e symmetry
is generated by DCKA transformations.
Using this construction, in Section \ref{isoharconf}
we show how the requirement of inclusion of
the spectrum generating sets of operators expands
linear Poincar\'e supersymmetry
of the $\mathcal{N}=2$ superextended AFF model and harmonic oscillator
up to superconformal $\mathfrak{osp}(2|2)$ and
super-Schr\"odinger symmetries, respectively.
Interesting peculiarity is that for the AFF model
the $\mathcal{N}=2$ supersymmetry can be realized in both
unbroken and broken phases, for which the complete sets of
the true and time-depending integrals are related by an automorphism
of the superconformal $\mathfrak{osp}(2|2)$
symmetry. In the case of the harmonic oscillator,
linear $\mathcal{N}=2$ super-Poincar\'e symmetry
is realized only in the unbroken phase.
The mechanism of generation of superconformal and super-Schr\"odinger
symmetries in the $\mathcal{N}=2$ superextended AFF model and harmonic oscillator
is generalized in Section \ref{gen} for rational deformations
of these systems, where, as we shall see,
nonlinearly extended
versions of the indicated symmetries appear.
Concrete simple examples illustrating the general
construction are presented
in Section \ref{examples}. Last Section \ref{conclusion}
is devoted to the discussion and outlook.
Several Appendices provide necessary technical
details for the main text.
\section{Darboux transformations and supersymmetry}
\label{Dar}
The procedure we will apply is based on the Darboux-Crum-Adler-Krein transformations \cite{Darboux,MatSal,Crum,Krein,Adler}.
Here, we summarize its basic ingredients.
Consider a Hamiltonian operator
$L=
-\frac{d^2}{dx^2}+V(x)$, and choose
the set of its eigenstates $\psi_1,\ldots,\psi_n$ of eigenvalues
$\lambda_1,\ldots,\lambda_n$
so that
their Wronskian $W$ takes nonzero values on the domain of
the quantum system $L$.
In such a way we generate a new, super-partner quantum system
\begin{equation}
\label{DarbouxHamil}
\breve{L}=-
\frac{d^2}{dx^2}+\breve{V}(x)\,,\qquad
\breve{V}(x)=V(x)-
2
\big(\ln(W(\psi_1,\psi_2,\ldots,\psi_n)\big)''\,,
\end{equation}
defined on the same domain as the initial system $L$.
An arbitrary solution $\psi_\lambda$ to the second order differential equation
$L\psi_\lambda=\lambda\psi_\lambda$,
$\lambda\neq \lambda_j$, $j=1,\ldots,n$,
is mapped into the wave function
\begin{equation}
\label{Darbouxstates}
\psi_{[n],\lambda}=\frac{W(\psi_1,\psi_2,\ldots,\psi_n,\psi_\lambda)}{W(\psi_1,\psi_2,\ldots,\psi_n)}\,,
\end{equation}
which is an
eigenstate of $\breve{L}$ of the same eigenvalue,
$\breve{L}\psi_{[n],\lambda}=\lambda\psi_{[n],\lambda}$,
and has the same nature of physical or non-physical
state as the pre-image $\psi_\lambda$.
{}From the pair $L$ and $\breve{L}$,
we construct the $\mathcal{N}=2$ superextended system
described by the $2\times 2$ matrix Hamiltonian and the supercharges
given by
\begin{equation}\label{Hlambda*}
\mathcal{H}=
\left(
\begin{array}{cc}
H_1\equiv \breve{L}-\lambda_*& 0 \\
0 & H_2\equiv L-\lambda_*
\end{array}
\right),\qquad
\mathcal{Q}_1=
\left(
\begin{array}{cc}
0& \mathbb A_n \\
\mathbb A_n^\dagger & 0
\end{array}
\right),
\qquad
\mathcal{Q}_2=i\sigma_3\mathcal{Q}_1\,.
\end{equation}
The system also is characterized by the integral $\Sigma=\frac{1}{2}\sigma_3$.
Here $\lambda_*$ is a constant,
while $\mathbb A_n$ and $\mathbb A_n^\dagger$
are the operators defined recursively by
relations
\begin{equation}\label{Andef}
\mathbb A_n=A_nA_{n-1}\ldots A_1,\qquad
A_{k}=
\mathbb A_{k-1}\psi_{k}\frac{d}{dx}\frac{1}{\mathbb A_{k-1}\psi_{k}},\qquad k=1,\ldots,n,
\end{equation}
where $A_0=1$ is assumed.
These operators intertwine the partner Hamiltonians,
$\mathbb A_n L=\breve{L}\mathbb A_n$,
$\mathbb A_n^\dagger \breve{L}=L\mathbb A_n^\dagger$. They
provide an alternative representation
for relation (\ref{Darbouxstates}),
$\psi_{[n,\lambda]}=\mathbb A_n\psi_\lambda$,
and generate the polynomials in Hamiltonian operators,
\begin{eqnarray}\label{An+An-}
&\mathbb A_n^\dagger \mathbb A_n=\prod_{j=1}^n(L-\lambda_j)\equiv P_n(L)\,,\qquad
\mathbb A_n \mathbb A_n^\dagger =P_n(\breve{L})\,.&
\end{eqnarray}
Up to (inessential for us here) multiplicative factor, we also have a relation
$\mathbb A_n^\dagger \psi_{[n,\lambda]}=\psi_\lambda.$
By the construction, the kernel of $\mathbb A_{n}$
is spanned by $n$ seed states $\psi_j$, $j=1,\ldots,n$,
while
$\ker \mathbb A_{n}^\dagger=\text{span}\,\{\mathbb A_{n}\widetilde{\psi_1},\ldots,\mathbb A_{n}\widetilde{\psi_n}\}$, where
\begin{equation}
\label{secondlinearindependent}
\widetilde{\psi_\lambda(x)}=\psi_\lambda(x) \int^x \frac{d\xi}{(\psi_\lambda(\xi))^2}\
\end{equation}
is a linearly independent solution to the
equation
$L{\psi}=\lambda{\psi}$
corresponding to the same value of $\lambda$,
$L\widetilde{\psi_\lambda}=\lambda\widetilde{\psi_\lambda}$.
Operator $\Gamma=\sigma_3$,
$\Gamma^2=1$, plays a role
of the $\mathbb Z_2$-grading operator that
identifies $\mathcal{H}$ and $\Sigma$
as even (bosonic) generators of the
superalgebra,
$[\Gamma,\mathcal{H}]=0$, $[\Gamma,\Sigma]=0$,
while $\mathcal{Q}_a$ are identified
as odd (fermionic)
generators,
$\{\Gamma,\mathcal{Q}_a\}=0$.
The integrals satisfy the (anti)-commutation relations
\begin{equation}\label{N2susy}
[\mathcal{H},\mathcal{Q}_a]=0\,, \quad
[\mathcal{H},\Sigma]=0\,,\quad
\{\mathcal{Q}_a,\mathcal{Q}_b\}=2\delta_{ab}P_n(\mathcal{H}+\lambda_*)\,,
\quad
[\Sigma,\mathcal{Q}_a]=-i\epsilon_{ab}\mathcal{Q}_b
\end{equation}
of the $\mathcal{N}=2$ superalgebra
which has a linear Lie-superalgebraic nature in the case of $n=1$ and is nonlinear
for $n>1$. Integral $\Sigma$ is a generator of a $U(1)$
$\mathcal{R}$-symmetry.
We will apply this construction to the
AFF model as well as to the QHO.
The peculiarity of these systems which we will exploit
is that
under Wick rotation $x\rightarrow ix$
their Hamiltonian operators transform in a simple way
just by multiplying by minus one.
As a consequence, under such a transformation
solution to the second order
differential equation $L\psi(x)=\lambda \psi(x)$
transforms into the function $\psi(ix)$ which satisfies
the equation $L\psi(ix)=-\lambda \psi(ix)$.
This property allows us to
produce from the AFF as well as from the QHO models
the pairs of exactly the same super-partner systems
but with the mutually shifted spectra
by choosing alternative sets of the seed states.
As a result,
the $\mathcal{N}=2$ supersymmetric structure
of the corresponding extended systems related to
the AFF model will
expand upto superconformal $\mathfrak{osp}(2| 2)$
symmetry in the case of $n=1$, while
for $n>1$ the corresponding
systems will be described by some nonlinear
extension of the $\mathfrak{osp}(2| 2)$.
When $L$ corresponds to
the QHO,
the $\mathfrak{osp}(2|2)$ will expand up to the
super-Schr\"odinger symmetry and
its nonlinear extensions in the cases of $n=1$ and $n>1$, respectively.
Essential ingredients in superconformal extensions
will be the sets of the spectrum generating ladder operators
\cite{CarPly2,CIP}.
\section{Superconformal and super-Schr\"odinger symmetries}
\label{isoharconf}
In this section we review and compare the superconformal symmetries
of the $\mathcal{N}=2$ superextended AFF and quantum harmonic
oscillator models. The
approach exposed here
will form a basis for the analysis of nonlinearly extended
superconformal symmetries
of rational deformations of these systems.
\subsection{Intertwining and ladder operators}
The dynamics of a particle
in a harmonic oscillator potential
is described by the Hamiltonian operator
$L=
-\frac{d^2}{dx^2}+x^2
\equiv L_0$
with
spectrum
$E_n=
2n+1
$,
$n=0,1,\ldots,$
and (nonnormalized) eigenstates
$\psi_{n}(x)=H_n(x)e^{-x^2/2}$, where $H_n(x)$
are the Hermite polynomials.
The choice of the ground state $\psi_0(x)=e^{-x^2/2}$ as the seed state for
the Darboux transformation generates according to (\ref{Andef}) the first order
differential operators
$A_1=\frac{d}{dx}+x\equiv a^-$, and $A_1^\dagger\equiv a^+$.
In this case the partner system (\ref{DarbouxHamil})
is just the shifted QHO
$\breve{L}=L_0+2$ with the shift constant equal to the
distance between neighbour energy levels
in its spectrum.
Because of this,
the intertwining operators $a^-$ and $a^+$ are nothing else
that the harmonic oscillator's ladder operators
satisfying relations $a^+a^-=L_0-1$, $a^-a^+=L+1$,
and so, $[a^-,a^+]=2$, and
$[L_0,a^\pm]=\pm 2a^\pm$.
If instead of taking the ground state
as the seed state for generation of the Darboux transformation
we choose a non-physical eigenstate
$\psi_0(ix)=e^{x^2/2}$ of eigenvalue $-1=-E_0$,
we obtain the same ladder operators
as the intertwining-factorizing operators $A_1=-a^+$ and $A^\dagger_1=-a^-$.
In this case, however, their relation to operators
$A_1$ and $A_1^\dagger$ is interchanged, and consequently
$\breve{L}=L_0-2$. This difference will
be essential to expand the $\mathcal{N}=2$ supersymmetry
of the superextended QHO up to
its dynamical super-Schr\"odinger symmetry.
\vskip0.1cm
The AFF model with a confining harmonic potential term, sometimes also
called an isotonic oscillator \cite{WeiJor},
is described by the Hamiltonian operator
\begin{equation}
\label{isoham}
L^{iso}_\nu=
-\frac{d^2}{dx^2}+x^2+\frac{\nu(\nu+1)}{x^2}
\,,\qquad \nu\in\mathbb R\,,
\end{equation}
defined on the domain
$\{ \psi\in L^2((0,\infty),dx)\vert \psi(0^+)=0\}$.
The case $\nu=0$ is understood as a limit $\nu\rightarrow 0$
which corresponds to the half-harmonic oscillator with the infinite potential well put at $x=0$.
For $\nu\geq-1/2$
the spectrum of the system and corresponding (non-normalized)
eigenstates are
\begin{equation}\label{Enun}
E_{\nu,n}=
2\nu+4n+3, \qquad
\psi_{\nu,n}=x^{\nu+1}\mathcal{L}_n^{(\nu+1/2)}(x^2)e^{-x^2/2}\,,\qquad
n=0,1,\ldots,
\end{equation}
where $\mathcal{L}_{n}^\alpha$
are the generalized Laguerre polynomials \cite{Perel}.
The coupling constant $g(\nu)=\nu(\nu+1)$
has a symmetry $g(\nu)=g(-\nu-1)$.
Applying the transformation $\nu\rightarrow -\nu-1$
to eigenstates (\ref{Enun}), one obtains
$\psi_{-\nu-1,n}=x^{-\nu}\mathcal{L}_n^{(-\nu-1/2)}(x^2)e^{-x^2/2}$,
which satisfy relation $L^{iso}_\nu\psi_{-\nu-1,n}=(-2\nu+4n+1)\psi_{-\nu-1,n}$.
For $\nu<0$ these are eigenfunctions (\ref{Enun})
of the system $L^{iso}_{\vert \nu\vert -1}$.
On the other hand, for $\nu>0$ these are non-physical eigenstates
of $L^{iso}_{ \nu}$,
which are non-normalizable singular at $x=0^+$ functions.
In spite of the non-physical nature, such states will play important role
in our constructions.
In the limit $\nu \rightarrow 0$,
system (\ref{isoham}) transforms into the
half-harmonic oscillator with the infinite potential barrier at $x=0$,
which we will denote below by $L_{0^+}$.
In accordance with the relations $H_{2n+1}(x)=2(-4)^nn!x\mathcal{L}_n^{(1/2)}(x^2)$ and
$H_{2n}(x)=(-4)^nn!\mathcal{L}_{n}^{(-1/2)}(x^2)$,
physical eigenstates $\psi_{\nu,n}$ of the system (\ref{isoham})
in the limit $\nu\rightarrow 0$
take the form of the odd states of the quantum harmonic
oscillator $L_0$ being eigenstates of $L_{0^+}$ on its domain,
while the indicated non-physical eigenstates
transform in this limit into the even eigenstates of $L_0$, which do not
satisfy the Dirichlet boundary condition for the
half-harmonic oscillator system $L_{0^+}$.
Let us take as the initial system $L=L_{\nu-1}^{iso}$
with $\nu\geq1/2$
and choose its ground state $\psi_{\nu-1,0}=x^\nu e^{-x^2/2}$
to generate
the Darboux transformation. In this way
we obtain the first order differential operators
\begin{equation} \label{A1nu-def}
A_1=\frac{d}{dx}+x-\frac{\nu}{x}\equiv A^-_\nu\,,\qquad
A_1^\dagger =-\frac{d}{dx}+x-\frac{\nu}{x}\equiv A^+_\nu\,,
\end{equation}
for which $\text{ker}\,A^-_\nu=\psi_{\nu-1,0}$ and
$\text{ker}\,A^+_\nu=(\psi_{\nu-1,0})^{-1}$.
They satisfy relations
\begin{eqnarray}\label{AA+nu}
&A^+_\nu A^-_\nu=L_{\nu-1}^{iso}-2\nu-1\,,\qquad
A^-_\nu A^+_\nu=L_{\nu}^{iso}-2\nu+1\,,&\\
& A^-_\nu L_{\nu-1}^{iso}=(L_{\nu}^{iso}+2)A^-_\nu\,,\qquad
A^+_\nu L_{\nu}^{iso}=(L_{\nu-1}^{iso}-2)A^+_\nu\,.&\label{AL+A}
\end{eqnarray}
Unlike the case of the harmonic oscillator,
the partner Hamiltonian operators are described
by potentials characterized by different coupling constants,
and though we have a shape invariance \cite{shape1,shape2},
the first order differential operators $A^+_\nu$
and $A^-_\nu$ are not ladder operators
for $L_{\nu}^{iso}$ or $L_{\nu-1}^{iso}$.
One can use the
symmetry of the coupling constant
and consider the first order operators
by changing $\nu\rightarrow -\nu-1$
in (\ref{A1nu-def}).
The obtained in such a way operators
will factorize and intertwine
the pair of the Hamiltonian operators
$L_{\nu}^{iso}$ and
$L_{\nu+1}^{iso}$.
We then make additionally a shift $\nu\rightarrow \nu-1$, and
obtain the first order differential operators
\begin{equation}\label{A-nudef}
A^-_{-\nu}=-A^+_{\nu}{}_{\vert}{}_{\nu\rightarrow -\nu}=
\frac{d}{dx}-x-\frac{\nu}{x}\,,\qquad
A^+_{-\nu}=-A^-_{\nu }{}_{\vert}{}_{\nu\rightarrow -\nu}=
- \frac{d}{dx}-x-\frac{\nu}{x}\,,
\end{equation}
for which
$\text{ker}\,A^+_{-\nu}=\psi_{-\nu-1,0}=x^{-\nu}e^{-x^2/2}$
and $\text{ker}\,A^-_{-\nu}=(\psi_{-\nu-1,0})^{-1}$.
It may be noted that $(\psi_{-\nu-1,0}(x))^{-1}=\psi_{\nu-1}(ix)$
is a non-physical eigenstate of $L_{\nu-1}^{iso}$ of eigenvalue
$-2\nu-3=-E_{\nu-1,0}$. Operators (\ref{A-nudef}) satisfy the relations
\begin{eqnarray}\label{AA-nu}
&A^+_{-\nu} A^-_{-\nu}=L_{\nu-1}^{iso}+2\nu+1\,,\qquad
A^-_{-\nu} A^+_{-\nu}=L_{\nu}^{iso}+2\nu-1\,,&\\
& A^-_{-\nu} L_{\nu-1}^{iso}=(L_{\nu}^{iso}-2)A^-_{-\nu}\,,\qquad
A^+_{-\nu} L_{\nu}^{iso}=(L_{\nu-1}^{iso}+2)A^+_{-\nu}\,.&\label{AL-A}
\end{eqnarray}
In
(\ref{AA-nu}) the additive factorization constants
are opposite in sign to those in (\ref{AA+nu}),
and the same is true for the displacement constants
in intertwining relations (\ref{AL-A}) and (\ref{AL+A}).
Combining intertwining relations (\ref{AL-A}) and (\ref{AL+A}),
one can construct ladder operators for
the system (\ref{isoham}),
\begin{eqnarray}\label{Cnu+-1}
&\mathfrak{C}_{\nu}^-=A^-_\nu A^+_{-\nu}=A^+_{-\nu-1}A^-_{\nu+1}=-(a^-)^2-\frac{\nu(\nu+1)}{x^2}\,,&\\
&\mathfrak{C}_{\nu}^+=A^-_{-\nu} A^+_\nu=A^+_{\nu+1}A^-_{-\nu-1}=-(a^+)^2-\frac{\nu(\nu+1)}{x^2}\,.&
\label{Cnu+-2}
\end{eqnarray}
They satisfy relations
$[L^{iso}_\nu,\mathfrak{C}_{\nu}^\pm]=\pm 4\mathfrak{C}_{\nu}^\pm$
and $\mathfrak{C}_{\nu}^\pm\psi_{\nu,n}\propto \psi_{\nu,n\pm 1}$,
$\mathfrak{C}_{\nu}^-\psi_{\nu,0}=0$
in correspondence with Eq. (\ref{Enun}).
\vskip0.1cm
The above comments on the choice of the seed states
in the form of non-physical eigenstates
obtained by Wick rotation $x\rightarrow ix$ from the
ground states of the harmonic and isotonic oscillators
correspond to a particular case of the already noted
specific symmetry
of both quantum systems.
In correspondence with this,
functions $ \psi_n(ix)\equiv \psi_{-n}(x)$
and $\psi_{\nu,n}(ix)\equiv \psi_{\nu,-n}(x) $ are formal (non-physical)
eigenstates of $L_0$ and $L^{iso}_\nu$ of eigenvalues $\lambda=-(2n+1)$
and $\lambda=-(2\nu+4n+3)$, respectively,
which are non-normalizable, exponentially increasing at infinity
functions.
Regardless of their non-physical nature,
these states also can be used as seed states for DCKA transformations
and will play a key role in our further constructions.
Furthermore, we have
solutions
(\ref{secondlinearindependent}), which
by means of relation (\ref{Darbouxstates}), or equivalently,
$\psi_{[n,\lambda]}=\mathbb A_n\psi_\lambda$,
may be transformed into normalizable wave functions.
When this happens,
the transformed system has a
corresponding eigenstate
in the lower part of the spectrum.
\vskip0.1cm
For the sake of brevity
we will use the following notations for physical and non-physical
eigenstates of the QHO\,:
\begin{equation}
\label{notation}
n\equiv \psi_n(x),\qquad -n\equiv \psi_{-n}=\psi_n(ix)\,,\qquad \widetilde{n}\equiv \widetilde{\psi_n}\,,
\qquad \widetilde{-n}\equiv \widetilde{\psi_{-n}}.
\end{equation}
Each rational deformation of the QHO and AFF systems
we will consider below can be generated by different DCKA transformations
from the harmonic and the half-harmonic oscillators.
When all the seed states
are eigenstates with positive
index $n>0$, following \cite{CIP} we call the corresponding DCKA scheme positive.
When the seed states carry only negative integer index of
non-physical eigenstates
and the DCKA transformation produces the same
system as the positive scheme modulo the global shift of the spectrum,
we call such a scheme negative, and refer to the
``complementary" schemes as dual.
For the details on the origin of such a duality see \cite{CIP}.
\subsection{$\mathfrak{osp}(2|2)$ and super-Schr\"odinger symmetries}\label{Section3.2}
We start with the $\mathcal{N}=2$ superextended AFF
model described by the matrix Hamiltonian operator
\begin{equation}
\label{hamiliso}
\mathcal{H}_\nu=
\left(
\begin{array}{cc}
L_{\nu}^{iso}-2\nu+1& 0 \\
0 & L_{\nu-1}^{iso}-2\nu-1
\end{array}
\right).
\end{equation}
With this definition, the equidistant spectrum of the system
is given by $\mathcal{E}_n=4n,$ $n=0,1,\ldots$,
where $n=0$ corresponds to the nondegenerate ground state
$(0,\psi_{\nu-1,0})^t$ of zero energy,
while all energy levels with $n\geq 1$ are doubly degenerate.
The described properties of the AFF model
allow us to identify the Lie-superalgebraic symmetry of
the extended system (\ref{hamiliso}) generated
by the even,
$\mathcal{H}_\nu$, $\mathcal{R}_\nu=\Sigma-\nu \mathbb{I}$, $\mathcal{C}_\nu^\pm$,
and odd,
$\mathcal{Q}_\nu^a$,
$\mathcal{S}_\nu^a$, operators,
where $\mathbb{I}$ is the unit $2\times 2$ matrix, and
\begin{eqnarray}\label{lader}
&\mathcal{C}_\nu^{\pm}=
\left(
\begin{array}{cc}
\mathfrak{C}_{\nu}^\pm& 0 \\
0 & \mathfrak{C}_{\nu-1} ^\pm
\end{array}
\right),&\\
\label{supercharge1}
&\mathcal{Q}_{\nu}^{1}=
\left(
\begin{array}{cc}
0& A^-_\nu \\
A^+_\nu & 0
\end{array}
\right),
\qquad
\mathcal{S}_\nu^1=
\left(
\begin{array}{cc}
0& A^-_{-\nu} \\
A^+_{-\nu} & 0
\end{array}
\right),&\\
&\mathcal{Q}_\nu^{2}=
i\sigma_3\mathcal{Q}_\nu^1\,,
\qquad \mathcal{S}_\nu^2=
i\sigma_3\mathcal{S}_\nu^1\,.&
\end{eqnarray}
Here the bosonic generators $\mathcal{C}_\nu^{\pm}$ are composed
from the ladder operators
given by Eqs. (\ref{Cnu+-1}),
(\ref{Cnu+-2})
and their analogs with index $\nu$ changed to $\nu-1$.
The supercharges $\mathcal{Q}_\nu^a$
correspond to the odd integrals defined in (\ref{Hlambda*}), and
the odd generators $\mathcal{S}_\nu^a$
are constructed from the intertwining
operators (\ref{A-nudef}).
The Lie superalgebraic relations have the form
\begin{eqnarray}\label{HRQ0}
&[\mathcal{H}_\nu,\mathcal{R}_\nu]=[\mathcal{H}_\nu,\mathcal{Q}_\nu^a]=0\,,&\\
\label{evencommutation}
&[\mathcal{H}_\nu,\mathcal{C}_\nu^{\pm}]=\pm4\mathcal{C}_\nu^{\pm}\,, \qquad
[\mathcal{C}_\nu^{-},\mathcal{C}_\nu^{+}]=8\mathcal{H}_\nu-16\mathcal{R}_\nu\,,&\\
\label{evenodd}
&[\mathcal{H}_\nu,\mathcal{S}_\nu^a]=-4i\epsilon^{ab}\mathcal{S}_\nu^b\,,\qquad
[\mathcal{R}_\nu,\mathcal{Q}_\nu^a]=-i\epsilon^{ab}\mathcal{Q}_\nu^b\,,
\qquad
[\mathcal{R}_\nu,\mathcal{S}_\nu^a]=-i\epsilon^{ab}\mathcal{S}^b_\nu\,,&\\
\label{fq1}
&[\mathcal{C}_\nu^-,\mathcal{Q}_\nu^a]=2(\mathcal{S}_\nu^a+i\epsilon^{ab}\mathcal{S}_\nu^b), \qquad
[\mathcal{C}_\nu^+,\mathcal{Q}_\nu^a]=-2(\mathcal{S}_\nu^a-i\epsilon^{ab}\mathcal{S}_\nu^b)\,,&\\
\label{fq3}
&[\mathcal{C}_\nu^-,\mathcal{S}_\nu^a]=2(\mathcal{Q}_\nu^a-i\epsilon^{ab}\mathcal{Q}_\nu^b)\,, \qquad
[\mathcal{C}_\nu^+,\mathcal{S}_\nu^a]=-2(\mathcal{Q}_\nu^a+i\epsilon^{ab}\mathcal{Q}_\nu^b)\,,&\\
\label{anti1}
&\{ \mathcal{Q}_\nu^a,\mathcal{Q}_\nu^b\}=2\delta^{ab}\mathcal{H}_\nu\,, \qquad
\{ \mathcal{S}_\nu^a,\mathcal{S}_\nu^b\}=2\delta^{ab}(\mathcal{H}_\nu -4\mathcal{R}_\nu)\,,&\\
\label{anti2}
&\{\mathcal{Q}^a_\nu,\mathcal{S}^b_\nu\}=\delta^{ab}(\mathcal{C}_\nu^{+}+\mathcal{C}_\nu^-)+
i\epsilon^{ab}(\mathcal{C}_\nu^+-\mathcal{C}_\nu^-)\,,&
\end{eqnarray}
and correspond to superconformal $\mathfrak{osp}(2|2)$
symmetry. Since both supercharges $\mathcal{Q}_\nu^a$
annihilate the unique ground state,
the $\mathcal{N}=2$ super-Poincar\'e symmetry
is unbroken.
The noncommutativity of $\mathcal{S}_\nu^a$
with $\mathcal{H}_\nu$ appears because
the operators $A^\pm_{-\nu}$
intertwine $L^{iso}_\nu$ and $L^{iso}_{\nu-1}$
with different shift constants in comparison
with $A^\pm_{\nu}$. Nonzero commutators
of $\mathcal{C}_\nu^\pm$ with $\mathcal{H}_\nu$
correspond to the commutation relations
of the respective ladder operators with Hamiltonians of the super-partner
subsystems.
So, the generators $\mathcal{S}_\nu^a$ and $\mathcal{C}_\nu^\pm$
are not integrals of motion
of the system $\mathcal{H}_\nu$.
They can be promoted to the \emph{dynamical integrals}
of motion via time-dressing by the evolution
operator,
\begin{eqnarray}
\label{recipeintegrals}
&\mathcal{O}_\nu\rightarrow \widetilde{\mathcal{O}_\nu}= e^{-i\mathcal{H}_\nu t}\mathcal{O}_\nu
e^{i\mathcal{H}_\nu t}\,.&
\end{eqnarray}
The transformed in this way operators
are explicitly depending on time
integrals of motion
$\widetilde{\mathcal{S}_\nu^1}=\mathcal{S}_\nu^1\cos 4t -\mathcal{S}_\nu^2\sin4t$,
$\widetilde{\mathcal{S}_\nu^2}=\mathcal{S}_\nu^2\cos 4t +\mathcal{S}_\nu^1\sin4t$,
and $\widetilde{\mathcal{C}_\nu^\pm}=e^{\mp 4it}\mathcal{C}_\nu^\pm$,
which satisfy the Heisenberg equation of
motion of the form $\frac{d}{dt}\widetilde{\mathcal{O}_\nu}=\partial \widetilde{\mathcal{O}_\nu}/\partial t-i
[\widetilde{\mathcal{O}_\nu},\mathcal{H}_\nu]=0$.
The substitution of $\mathcal{S}_\nu^a$ and $\mathcal{C}_\nu^\pm$
for the time-dressed integrals $\widetilde{\mathcal{S}_\nu^a}$ and
$\widetilde{\mathcal{C}_\nu^\pm}$ does not change the form of
the $\mathfrak{osp}(2|2)$ superalgebraic relations
since transformation (\ref{recipeintegrals}) is unitary.
\vskip0.1cm
We have fixed the Hamiltonian operator (\ref{hamiliso})
coherently with the supercharges $\mathcal{Q}_\nu^a$
constructed in terms of the intertwining operators $A^\pm_\nu$.
The matrix operators $\mathcal{S}_\nu^a$ constructed
in terms of the other pair of intertwining operators $A^\pm_{-\nu}$
turn out in this case to be the dynamical integrals of motion.
If
we choose, instead, the operator
$\mathcal{H}'=\mathcal{H}_\nu -4\mathcal{R}_\nu=\text{diag}\,(L^{iso}_\nu+2\nu-1,\,
L^{iso}_{\nu-1}+2\nu+1) $
as a Hamiltonian of the superextended system,
then fermionic operators $\mathcal{S}_\nu^a$
will be true, not depending explicitly on time, integrals of motion (supercharges),
while $\mathcal{Q}_\nu^a$ will transform (after unitary time-dressing procedure)
into dynamical integrals of motion.
In effect, these changes are a part of the transformation
$\mathcal{H}_\nu\rightarrow \mathcal{H}'_\nu=\mathcal{H}_\nu -4\mathcal{R}_\nu$,
$\mathcal{Q}_\nu^1 \rightarrow {\mathcal{Q}'}_\nu^1=\mathcal{S}_\nu^2$,
$\mathcal{Q}_\nu^2 \rightarrow {\mathcal{Q}'}_\nu^2=\mathcal{S}_\nu^1$,
$\mathcal{S}_\nu^1 \rightarrow {\mathcal{S}'}_\nu^1=\mathcal{Q}_\nu^2$,
$\mathcal{S}_\nu^2 \rightarrow {\mathcal{S}'}_\nu^2=\mathcal{Q}_\nu^1$
$\mathcal{R}_\nu \rightarrow \mathcal{R}'_\nu=-\mathcal{R}_\nu$,
$\mathcal{C}_\nu^\pm \rightarrow
{\mathcal{C}'}_\nu^\pm= \mathcal{C}_\nu^\pm$,
which is the automorphism of the superalgebra
(\ref{HRQ0})--(\ref{anti2}).
The superextended system described by the
Hamiltonian $\mathcal{H}'_\nu=\mathcal{H}_\nu -4\mathcal{R}_\nu$
has, however, essentially different properties in comparison with the
system given by the matrix Hamiltonian $\mathcal{H}_\nu$.
Its spectrum is $\mathcal{E}_n=4n+2\nu+2$, $n=0,1,\ldots$,
and all its energy levels including the lowest one,
$\mathcal{E}_0=+2\nu+2>0$, are doubly degenerate,
and neigther of its two lowest eigenstates is
annihilated by both supercharges ${\mathcal{Q}'}_\nu^a$.
This means that the system $\mathcal{H}'_\nu$, unlike
$\mathcal{H}_\nu$, is characterized by the
spontaneously broken $\mathcal{N}=2$
super-Poincar\'e symmetry.
\vskip0.1cm
One can change the basis by considering the linear combinations
\begin{eqnarray}
&\pi_{\nu}^a=\frac{1}{2\sqrt{2}}(\mathcal{S}_\nu^a -\epsilon^{ab}\mathcal{Q}_\nu^b)\, ,
\qquad
\zeta_\nu^{a}=\frac{1}{2\sqrt{2}}(\mathcal{Q}_\nu^a-\epsilon^{ab}\mathcal{S}_\nu^b)\,, &\\
&\mathcal{D}_\nu=\frac{1}{8}(\mathcal{C}_\nu^++\mathcal{C}_\nu^-)\,,&\\
&\mathcal{L}_\nu=\frac{1}{4}\mathcal{H}_\nu-\frac{1}{2}\mathcal{R}_\nu+
\frac{i}{2}(\mathcal{C}_\nu^+-\mathcal{C}_\nu^-)\,,\qquad
\mathcal{K}_\nu=\frac{1}{4}\mathcal{H}_\nu-\frac{1}{2}\mathcal{R}_\nu-
\frac{i}{2}(\mathcal{C}_\nu^+-\mathcal{C}_\nu^-)\,. &
\end{eqnarray}
Then nonzero (anti)commutation relations
of the superalgebra take the form
\begin{eqnarray}
&[\mathcal{D}_\nu,\mathcal{L}_\nu]=i\mathcal{L}_\nu\,,\qquad
[\mathcal{D}_\nu,\mathcal{K}_\nu]=-i\mathcal{K}_\nu\,,\qquad
[\mathcal{K}_\nu,\mathcal{L}_\nu]=2i\mathcal{D}_\nu\,,&\\
&\{\pi_\nu^{a},\pi_\nu^{b}\}=2\delta^{ab}\mathcal{L}_\nu\,,
\qquad
\{\zeta_\nu^{a},\zeta_\nu^{b}\}=2\delta^{ab}\mathcal{K}_\nu\,,
\qquad \{\pi_\nu^{a},\zeta_\nu^{b}\}=2\delta^{ab}\mathcal{D}_{\nu}+\epsilon^{ab}\mathcal{R}_\nu\,,&\\
&[\mathcal{L}_\nu,\zeta_\nu^{a}]=-i\pi_\nu^{a}\,,\qquad [\mathcal{K}_\nu,\pi_\nu^{a}]=i\zeta_\nu^{a}\,,\qquad
[\mathcal{D}_\nu,\pi_\nu^{a}]=\frac{i}{2}\pi_\nu^{a}\,,\qquad
[\mathcal{D}_\nu,\zeta_\nu^{a}]=-\frac{i}{2}\zeta_\nu^{a}\,,&\\
&[\mathcal{R}_\nu,\pi_\nu^{a}]=i\epsilon^{ab}\pi_\nu^{b}\,,\qquad
[\mathcal{R}_\nu,\zeta_\nu^{a}]=i\epsilon^{ab}\zeta_\nu^{b}\,.&
\end{eqnarray}
This is a form of superconformal algebra which appears in a free
nonrelativisitc spin-$1/2$ particle system and
superextended AFF model without the confining harmonic potential term \cite{AnaPLy}.
One can note that
if we restore the harmonic oscillator frequency $\omega$, which was fixed here equal to $2$,
and take its zero limit,
we recover the supersymmetric two-particle Calogero model
(with the omitted center of mass degree of freedom)
and its superconformal symmetry.
\vskip0.1cm
The superconformal symmetry of the superextended QHO
described by matrix Hamiltonian $\mathcal{H}=\text{diag}\,(L_0+1,L_0-1)$
can be generated in analogous way by taking $\psi_0$ and $\psi_{-0}$
as the seed states to produce ladder operators as the intertwining-factorizing
operators in two different ways as it was described above.
The superconformal structure of the supersymmetric harmonic oscillator
can be obtained, however, in a more direct way just by \emph{formally} putting
$\nu=0$ in the Hamiltonian operator (\ref{isoham}), in the first order intertwining operators (\ref{A1nu-def})
and (\ref{A-nudef}), in the ladder operators
(\ref{Cnu+-1}), (\ref{Cnu+-2}) of the AFF model, and in the generator $\mathcal{R}_\nu$
of the $U(1)$ $\mathcal{R}$-symmetry.
In such a way we reproduce the superconformal $\mathfrak{osp}(2|2)$
algebra for the supersymmetric QHO
having exactly the same form as for the superextended AFF system.
The important point, however, is that the QHO
is defined on the whole real line instead of the half-line in the case of the
AFF model. Because of this the distance between its energy levels is twice
less than in the AFF model and the superconformal $\mathfrak{osp}(2|2)$
symmetry expands upto the superextended Schr\"odinger symmetry
whose superalgebra includes the superextended Heisenberg
algebra in the form of invariant subsuperalgebra.
The operators $\mathcal{C}_0^\pm=-(a^\pm)^2\equiv \mathcal{J}_\pm$, unlike the case of the
superextended
AFF model (\ref{hamiliso}), are not enough now
to produce all the Hilbert space of the super-oscillator if we start from its ground state
$(0,\psi_0)^t$
and also use other generators of the $\mathfrak{osp}(2|2)$\,:
in this way we generate only the states of the form
$(0,\psi_{2n})^t$ and $(\psi_{2n+1},0)^t$, $n=0,1,\ldots$,
and their linear combinations.
To produce the missing states, it is sufficient to extend the set of
generators by matrices $\Sigma_a=\sigma_a$, $a=1,2$.
They anti-commute with the grading operator $\Gamma=\sigma_3$,
and are identified as odd generators. {}From the commutation
relations $[\mathcal{H},\Sigma_\pm]=\pm 2\Sigma_\pm$,
where $\Sigma_\pm=\frac{1}{2}(\Sigma_1\pm i\Sigma_2)$,
we find that by unitary transformation of the form (\ref{recipeintegrals})
they can be promoted to the dynamical integrals
$\widetilde{\Sigma_\pm}=e^{\mp 2it}\Sigma_\pm$.
The anti-commutators of $\Sigma_a$ between themselves
generate central charge $\mathbb{I}$, while their anti-commutators with
fermionic generators $\mathcal{Q}_a$ and $\mathcal{S}_a$
produce a new pair of the bosonic generators
\begin{equation}
\mathcal{G}_{\pm}=
\left(
\begin{array}{cc}
a^\pm & 0 \\
0 & a^\pm
\end{array}
\right).
\end{equation}
Operators $\mathcal{G}_\pm$ together with
the identity matrix operator $\mathbb{I}$
generate the $\mathfrak{h}_1$ Heisenberg algebra
which is an invariant subalgebra of the
complete superextended Schr\"odinger symmetry
of the super-oscillator. Operators
$\mathcal{G}_\pm$, $\mathbb{I}$ and $\Sigma_a$
generate the superextended
Heisenberg algebra.
The complete set of the (anti)-commutation relations
of the superextended Schr\"odinger symmetry
is given by those of the $\mathfrak{osp}(2|2)$
superalgebra and by relations
\begin{eqnarray}
\label{qho3}
&[\mathcal{H},\mathcal{G}_\pm]=\pm2\mathcal{G}_{\pm}\,,\qquad
[\mathcal{G}_\mp,\mathcal{J}_{\pm}]=\mp 2\mathcal{G}_{\pm}\,,\qquad
[\mathcal{G}_-,\mathcal{G}_{+}]=2\mathbb{I}\,,&\\
&\{\Sigma_{a},\Sigma_b\}=2\delta_{ab}\mathbb{I}\,,\qquad [\mathcal{H},\Sigma_a]=
2i\epsilon_{ab}\Sigma_{b} \,,\qquad [\sigma_3,\Sigma_a]=-2i\epsilon_{ab}\Sigma_{b}\,,&\\
&\{\Sigma_{a},\mathcal{Q}_b\}=\delta_{ab}(\mathcal{G}_{+}+\mathcal{G}_{-})+
i\epsilon_{ab}(\mathcal{G}_{+}-\mathcal{G}_{-})\,,&\\
&\{\Sigma_{a},\mathcal{S}_b\}=\delta_{ab}(\mathcal{G}_{+}+\mathcal{G}_{-})-
i\epsilon_{ab}(\mathcal{G}_{+}-\mathcal{G}_{-})\,,&\\
&[\mathcal{G}_-,\mathcal{Q}_a]=\Sigma_a+i\epsilon_{ab}\Sigma_b\, ,\qquad
[\mathcal{G}_+,\mathcal{Q}_a]=-\Sigma_a+i\epsilon_{ab}\Sigma_b\,,&\\
\label{qho4}
&[\mathcal{G}_-,\mathcal{S}_a]=\Sigma_a-i\epsilon_{ab}\Sigma_b \,,\qquad
[\mathcal{G}_+,\mathcal{S}_a]=-(\Sigma_a+i\epsilon_{ab}\Sigma_b)\,,&\\
\label{SigGJ0}
&[\Sigma_{a},\mathcal{G}_\pm]=[\Sigma_{a},\mathcal{J}_\pm]=0\,.&
\end{eqnarray}
Let us also note that generators $\Sigma_a$
intertwine the supercharges
$\mathcal{Q}_a$ with the dynamical fermionic
integrals $\mathcal{S}_a$\,:
$\Sigma_1\mathcal{Q}_a=
\epsilon_{ab}\mathcal{S}_b\Sigma_1$,
$\Sigma_2\mathcal{Q}_a=-\epsilon_{ab}\mathcal{S}_b\Sigma_2$.
It is also worth to note that if we try to expand the set of symmetry generators
of the superextended AFF system by $\Sigma_a$, their
(anti)-commutation with generators of the $\mathfrak{osp}(2|2)$ symmetry
would lead to infinite-dimensional superalgebraic structure.
In other way this can be understood from the point of view of the
time-dressing procedure. In this case
$[\mathcal{H}_\nu,\Sigma_{a}]\sim 1/x^2\cdot \epsilon_{ab}\Sigma_b$,
and the unitary transformation (\ref{recipeintegrals}) results
in a nonlocal operator being an infinite series in derivative
$d/dx$ and time parameter $t$.
\vskip0.1cm
If similarly to the case of the superextended AFF model,
we change the Hamiltonian $\mathcal{H}_0$ to
$\mathcal{H}_0'=\mathcal{H}_0-4\mathcal{R}_0=\text{diag}\,(L_0-1,L_0+1)$,
we obtain exactly the same matrix operator but with just
permuted Hamiltonian operators of the subsystems.
This transformation is a part of the corresponding automorphism of
the super-Schr\"odinger algebra which, unlike the AFF model case,
does not change the unbroken nature of the $\mathcal{N}=2$ super-Poincar\'e symmetry.
\vskip0.1cm
In the limit of zero frequency, the super-Schr\"odinger
symmetry of the free nonrelativistic spin-$1/2$ quantum particle is recovered
\cite{BeckHussin,DuvHorSh,InzPly}.
This also is coherent with the known relation between the
free particle system, harmonic oscillator, and AFF model \cite{AFF,BTGE,Nied}.
If we change the coordinate and time variable of the free particle
of mass $M$ as
$(x,t)\rightarrow (\xi,\tau)$,
$x(\xi,\tau)=\xi(\beta\omega)^{1/2}/\cos (\omega\tau)$, $t(\tau)= \beta\tan (\omega\tau)$,
where $\beta$ is a real constant of dimension of time,
its action $\mathcal{A}=\frac{1}{2}M\int\dot{x}^2dt$ will be transformed into
the action $\mathcal{A}_\omega=\int L(\xi,\xi')d\tau$ of the harmonic oscillator
of frequency $\omega$
with
\begin{eqnarray}
&L(\xi,\xi')=\frac{1}{2}M\left((\xi')^2-\omega^2\xi^2\right)+\frac{d}{d\tau}\left(\frac{1}{2}M\omega
\xi^2\tan(\omega\tau)\right)\,,&
\end{eqnarray}
where $\xi'=d\xi/d\tau$.The total derivative term is essential for getting
the QHO's propagator from the free particle system's propagator
\cite{OleS,BTGE}.
In the context of (super)conformal symmetry we discuss,
it is important that
the
action potential term $\mathcal{A}_g=g\int \frac{dt}{x^2}$
is invariant under the indicated transformation,
$\mathcal{A}_g\rightarrow g\int \frac{d\tau}{\xi^2}$.
Then the same transformation applied to the
Calogero system given by Lagrangian $L=\frac{1}{2}M\dot{x}^2+\frac{g}{x^2}$
will produce the Lagrangian for the AFF model with
the confining harmonic oscillator potential term.
This is a particular case
of the so called classical Arnold
transformation \cite{AT}.
\vskip0.1cm
For the superconformal symmetries of the
$\mathcal{N}=2$ superextended AFF and QHO systems
there was important the equidistant nature of their spectra,
and that the sets of the spectrum generating operators of the initial
quantum models are composed by the conjugate pairs of
differential operators of the second and first order, respectively.
The rational deformations of these systems are characterized
in a generic case by the sets of three pairs of the spectrum generating
ladder operators which are differential operators of the orders not lower than
three. This difference, as we shall see, leads to the
essentially different \emph{nonlinear} structures of the superconformal
and super-Schr\"odinger symmetries of the corresponding
$\mathcal{N}=2$ superextended systems.
\vskip0.1cm
To conclude this section we note that as it follows from relations
(\ref{HRQ0})--(\ref{anti2}), the complete $\mathfrak{osp}(2|2)$ superconformal structure
of the superextended AFF system
can be obtained by extending the set of its generators
of $\mathcal{N}=2$ supersymmetry (\ref{N2susy})
by any one of the dynamical integrals $\mathcal{S}^a_\nu$, $\mathcal{C}^+_\nu$
or $\mathcal{C}^-_\nu$ (or by any of the operators $\mathcal{S}_a$, $\mathcal{J}_+$ or
$\mathcal{J}_-$ in the case of super-oscillator).
To recover additional (anti)-commutation relations
(\ref{qho3})--(\ref{anti2})
of the super-Schr\"odinger symmetry of super-oscillator system,
it is sufficient then to extend the set of
generators of its $\mathfrak{osp}(2|2)$ superalgebra
by any one of the dynamical integrals
$\Sigma_a$, $a=1,2$, $\mathcal{G}_+$ or $\mathcal{G}_-$.
This observation will play a key role
in what follows for generation of nonlinearly deformed and
extended superconformal and super-Schr\"odinger
symmetries of the $\mathcal{N}=2$ super-extensions
of the rationally deformed AFF and harmonic oscillator systems.
\section{Symmetry generators}
\label{gen}
We pass now to the study of extensions and deformations
of the superconformal and super-Schr\"odinger symmetries
which appear in the $\mathcal{N}=2$ superextended systems
described by the pairs of the Hamiltonian operators
($L_0$, $L_{0,def}$) and ($L_{0^+}$, $L^{iso}_{m,def}$).
Here $L_{0,def}$ and $L^{iso}_{m,def}$ correspond
to rational deformations of the QHO
and the AFF model with integer
values of the parameter $\nu=m$, $m\in \mathbb N$.
In general case, rationally
deformed harmonic and isotonic oscillator systems
are characterized by finite number of gaps (missing energy levels)
in their spectra, and their description requires
more than one pair of spectrum generating ladder operators.
It is because of this expansion of the sets of ladder operators
and of their higher differential order that
there appear
nonlinearly deformed superconformal
and super-Schr\"odinger structures which generalize
the corresponding Lie superalgebraic symmetries.
This section is devoted to the description of the complete sets of
generators of the indicated symmetries.
They will be constructed on the basis of
the QHO.
\subsection{Generation of rationally extended oscillator systems}
We first consider rational deformations (extensions)
of the harmonic oscillator.
They are constructed following the well known recipe of the
Krein-Adler theorem \cite{Krein,Adler}.
Consider a `positive' scheme $(n_1,n_1+1,\ldots,n_\ell,n_\ell+1)$,
where $n_j\in \mathbb N $, $j=1,\ldots,\ell$, correspond to the
set of the chosen seed states, see notations (\ref{notation}).
It generates the deformed system
\begin{eqnarray}
\label{DQHO}
&L_{(n_1,n_1+1,\ldots,n_\ell,n_\ell+1)}=L_{0} +4\ell +
\frac{F(x)}{Q(x)},&
\end{eqnarray}
where $F(x)$ and $Q(x)$ are real-valued even polynomials,
with $Q(x)$ being positive definite and having the degree of $F(x)$ plus two.
Physical eigenstates of this system are obtained by the map
(\ref{Darbouxstates}) applied to the physical eigenstates
of harmonic oscillator with indexes different from those of the seed states.
The resulting system has certain number of gaps
in the spectrum at
location of energy levels of the seed states,
and each gap corresponds to even number of missing neighbour
energy levels.
\vskip0.1cm
Another class of systems which can be constructed by
Darboux-Crum transformations on the basis
of the QHO
corresponds to deformations of the AFF model
(\ref{isoham}) with $\nu=m$, $m\in \mathbb N$.
For this one can first construct rationally deformed harmonic oscillator system
as in (\ref{DQHO}), then introduce an infinite potential barrier at $x=0$,
and finally take first $m$ physical states (satisfying Dirichlet boundary condition at $x=0$)
as the seed states for additional Darboux-Crum transformation.
This composite generating procedure admits other interpretations due to
the iterative properties of DCKA transformations.
One way to see this is to take the half-harmonic oscillator
$L_{0^+}$ as a starting system and use
the set of states $(n_1,n_1+1,\ldots,n_\ell,n_\ell+1,2k_1+1,\ldots, 2k_m+1)$,
where even indexes inside the set
$n_1,n_1+1,\ldots,n_\ell,n_\ell+1$ represent non-physical eigenstates
of $L_{0^+}$ and
$k_i$, $i=1,\ldots,m$, are identified as
$m$ odd states which were not considered in the first set of $2n_\ell$ states.
The Hamiltonian operator
\begin{eqnarray}
\label{REIO}
&L_{(n_1,n_1+1,\ldots,n_
\ell,n_\ell+1,2k_1+1,\ldots, 2k_m+1)}=L_{m}^{iso}+
2m +4\ell + \frac{\widetilde{F(x)}}{\widetilde{Q(x)}},&
\end{eqnarray}
appears as a final result of the described procedure,
where polynomials $\widetilde{F(x)}$ and $\widetilde{Q(x)}$
have the properties similar to those of $F(x)$ and $Q(x)$ in
(\ref{DQHO}).
With the described method one can obtain
rational deformations of the AFF model with
integer values of index $\nu=m$ only, on which we will focus here.
Physical states of the system (\ref{REIO}) are obtained
by applying the mapping
(\ref{Darbouxstates}) to \emph{odd} states of harmonic oscillator
with index different from that of the seed states. In general such a
system has gaps in its spectrum.
If, however, the set
$n_1,n_1+1,\ldots,n_
\ell,n_\ell+1,2k_1+1,\ldots, 2k_m+1$ contains all
the $\ell+m$ odd indexes from $1$ to $2k_m+1$,
the generated deformed AFF system
will have no gaps in its spectrum
and we obtain a system completely isospectral
to $L_{0^+}+4\ell+2m$. Such completely isospectral
(gapless) rational deformations are impossible
in the case of the harmonic oscillator system.
\vskip0.1cm
The method of mirror diagrams developed and employed in \cite{CIP} is a
technique
such that a dual scheme with
non-physical `negative' eigenstates (\ref{notation}), which
are
obtained by the
transformation $x\rightarrow ix$ from physical states,
is derived from a positive scheme with physical states of $L_0$ only,
and vice versa.
Both schemes are related in such a way that
the second logarithmic derivatives
of their Wronskians are equal up to an additive constant.
Formally one scheme can be obtained from another in the following way.
Let positive scheme is
$\Delta^+\equiv (l_1^+,\ldots,l_{n_+}^+)$, where
$l_i^+$ with $i=1,\ldots, n_+$ are certain positive numbers
which have to be chosen according to the
already described rules.
Then negative scheme is given by
$\Delta^-=(-\check{0},\ldots,-\check{n}_i^-,
\ldots,-l_{n_+}^+)$, where $-\check{n}_i^-\equiv l_i^+-l_{n_+}^+$,
means that the corresponding number $-n_i^-\equiv l_i^+-l_{n_+}^+$ is omitted
from the set $\Delta^-$.
On the contrary, if we are given the negative scheme
$\Delta^-\equiv (-l_{1}^-,\ldots,-l_{n_-}^-)$, where $-l_{j}^-$ with
$j=1,\ldots, n_-$
are certain negative numbers,
then positive scheme is given by
$\Delta^+=(\check{0}, \ldots,\check{n}^+_{j},\ldots,l_{n_-}^-)$, where symbols
$\check{n}_j^+=l_{n_-}^- -l_{j}^-$ represent the
states missing from the list of the chosen seed states.
For dual schemes the relation
\begin{equation}\label{genDual}
e^{-n_+x^2/2}W(-l_{1}^-,\ldots,-l_{n_-}^-)=e^{n_-x^2/2}
W(l_1^+,\ldots,l_{n_+}^+)\,,
\end{equation}
is valid modulo a multiplicative constant, from which one finds that
the Hamiltonians of dual schemes satisfy the relation
\begin{equation}
\label{L+L-}
L_{(+)}-L_{(-)}=2N\,,\qquad N\equiv n_+ +n_-=l_{n^+}^++1=l_{n^-}^-+1\,,
\end{equation}
where $L_{(+)}\equiv L_{(l_{1}^+,\ldots,l_{n_+}^+)}$ and
$ L_{(-)}\equiv L_{(-l_{1}^-,\ldots,-l_{n_-}^-)}$.
By means of negative scheme we do not remove any energy level
from the spectrum, but, instead, energy levels
can, but not obligatorily, be introduced in its lower part.
Particular in special case
of completely isospectral deformations of the (shifted) $L_m^{iso}$ systems,
all $m$ seed states composing negative scheme
are non-physical odd eigenstates of $L_0$.
Analogous picture with dual negative schemes also is valid for rational
deformations of the harmonic oscillator system
(\ref{DQHO}), where, however,
completely isospectral deformations of the (shifted) operator $L_0$
are impossible.
\subsection{Basic intertwining operators}
According to \cite{CarPly2,CIP},
with each of the dual schemes it is necessary
first to associate two basic pairs of the intertwining operators.
Here, we discuss general properties of such operators
without taking care of the concrete nature of the system built by the DCKA transformation.
On the way, however, some important distinctions between rational deformations of
the AFF model and harmonic oscillator have to be taken into account, and for this reason,
it is convenient to speak of two classes of the systems.
We distinguish them by introducing the class index $c$,
where $c = 1$ and $c=2$ will correspond to deformed harmonic oscillator
and deformed AFF conformal mechanics model, respectively.
Although in essence both Hamiltonians $L_{(+)}$ and $L_{(-)}$ are the same operator,
the corresponding DCKA transformations are very different from each other.
The positive scheme generates the pair of Hermitian conjugate intertwining operators
to be differential operators of order $n_+$, which for the sake of simplicity
we denote as
$\mathbb A_{n_+}\equiv A_{(+)}^-$ and $(A_{(+)}^-)^\dagger \equiv A_{(+)}^+$.
In the same way, the negative scheme produces a different pair of intertwining operators,
this time of order $n_-$,
$\mathbb A_{n_-}\equiv A_{(-)}^-$ and $(A_{(-)}^-)^\dagger \equiv A_{(-)}^+$.
These operators satisfy intertwining relations
\begin{equation}
\label{inter0}
A_{(+)}^-L=L_{(+)}A_{(+)}^-\,, \qquad A_{(-)}^-L=L_{(-)}A_{(-)}^-\,,
\end{equation}
and
Hermitian conjugate relations
with
$A_{(+)}^+$ and $A_{(-)}^+$,
where $L$ is $L_0$ or $L_{0^+}$ in dependence on the class of
the rationally deformed system $L_{(\pm)}$.
In this way, operators
$A_{(-)}^-$ and $A_{(+)}^-$
transform eigenstates
of the harmonic oscillator
into eigenstates of $L_{(\pm)}$, but they
do this in different ways.
Applying operator identities
(\ref{inter0}) to an arbitrary physical or non-physical (formal) eigenstate
$\varphi_n$ of $L$ different from any seed state of the positive scheme
and using Eq. (\ref{L+L-}), one can derive the equality
\begin{equation}
\label{relation-operators}
A_{(-)}^-\varphi_n=A_{(+)}^-\varphi_{n+N}
\end{equation}
to be valid modulo a multiplicative constant.
As a result,
both operators acting on the same state of the harmonic oscillator produce
different states
of the new system.
Analogous difference between intertwining operators was
already observed in Section \ref{isoharconf} in the case of the
non-deformed harmonic oscillator and
AFF model.
The Hermitian conjugate operators $A_{(-)}^+$ and $A_{(+)}^+$ do a similar job but
in the opposite direction.
Eq. (\ref{relation-operators}) suggests that some
peculiarities
should be taken into account for class 2 systems\,:
the infinite potential barrier at $x=0$
assumes that
physical states of $L_{0^+}$ and
$L_{(\pm)}$ systems
are described by odd wave functions.
Then,
in order $ A_{(+)}^-$ would transform
physical states of $L_{0^+}$ into physical states
of $L_{(\pm)}$, index $N$ has to be taken even in (4.6) for odd values of index $n$.
This means that
$A_{(-)}^-$ transforms physical states into physical
only if $N$ is even.
In the case of odd $N$, it is necessary to take
$A_{(-)}^-a^-$ or $A_{(-)}^-a^+$ as a physical intertwining operator.
It is convenient to take into account this peculiarity
by denoting
the remainder of the division $N/c$ by
$r(N,c)$\,: it takes value $1$
in the class $c=2$ of the systems with odd $N$ and equals zero
in all other cases.
The alternate products of the described
intertwining operators are of the form (\ref{An+An-}),
and
for further analysis it
is useful to write down them explicitly:
\begin{eqnarray}
\label{A-A-A+A+Poly}
&A_{(\pm)}^{+}A_{(\pm)}^{-}=P_{n_\pm}(L)\,,\qquad A_{(\pm)}^{-}A_{(\pm)}^{+}=P_{n_\pm}(L_{(\pm)})\,,&\\
\label{polyA}
&P_{n_+}(\eta)\equiv \prod_{k=1}^{n_+}(\eta-2l_k^+-1)\,, \qquad
P_{n_-}(\eta)\equiv \prod_{k=1}^{n_-}(\eta+2l_k^-+1)\,.&
\end{eqnarray}
Here $l_k^+$ are indexes of physical states of eigenvalues $2l_k^++1$
in
the set $\Delta_+$,
and $-l^-_k$ correspond to non-physical states of eigenvalues $-2l^-_k-1$
in
the negative scheme $\Delta_-$.
In the same
vein, it is useful to write
$(a^+)^k(a^-)^k=T_k(L_0)$ and $(a^-)^k(a^+)^k=T_k(L_0+2k)$, where
\begin{eqnarray}
\label{Tk}
&T_{k}(\eta)\equiv \prod_{s=1}^{k}(\eta-2s+1)\,, \qquad T_{k}(\eta+2k)\equiv \prod_{s=1}^{k}(\eta+2s-1)\,.&
\end{eqnarray}
We also have the operator identities
\begin{eqnarray}
\label{ide}
(a^-)^{N}=(-1)^{n_-}A_{(-)}^+A_{(+)}^-\,,
\qquad f(L_{(-)})A_{(+)}^-(a^+)^{n_-}=(-1)^{n_-}h(L_{(-)})A_{(-)}^-(a^-)^{n_+}
\end{eqnarray}
and their Hermitian conjugate versions,
where $f(\eta)$ and $h(\eta)$ are polynomials whose explicit structure
is given in Appendix \ref{show}.
In one-gap deformations of the harmonic oscillator and gapless deformations
of $L_1^{iso}$ these polynomials reduce to $1$.
\subsection{Extended sets of ladder and intertwining operators}
\label{interladder}
The intertwining operators considered in the previous subsection
allow us to construct ladder operators for the deformed systems, see
\cite{CarPly2,CIP}\,:
\begin{itemize}
\item [1.] Operators of the type $\mathfrak{A}$\,: $\mathfrak{A}_{c}^\pm=A_{(-)}^-(a^\pm)^c A_{(-)}^+$,
which have a structure of the Darboux-dressed ladder operators of the harmonic oscillator $L_0$ ($c=1$)
and of the half-harmonic oscillator $L_{0^+}$ ($c=2$).
These operators detect all the separated states by annihilating them but act
as ladder operators in the equidistant part of the spectrum.
For gapless (completely isospectral) deformations of $L_{0^+}$ these are
the spectrum generating operators
by which one can generate any physical eigenstate of the system from its any other physical eigenstate.
\item[2.] Operators of the $\mathfrak{B}$ type\,: $\mathfrak{B}_{c}^\pm=A_{(+)}^-(a^\pm)^c A_{(+)}^+$.
Their function is to identify each valence band\,: the operators $\mathfrak{B}_{c}^+$ and $ \mathfrak{B}_{c}^-$
annihilate, respectively, the highest and the lowest states in each valence band.
In the equidistant part of the spectrum they also act as ordinary ladder operators.
For gapless deformations, this pair
is also a spectrum generating set, but
$|\mathfrak{B}_{2}^\pm| > |\mathfrak{A}_{2}^\pm|$,
where $|.|$ denotes differential order of operator.
\item [3.] Operators of the $\mathfrak{C}$ type connect the equidistant part
of the spectrum with separated states.
They are given by $\mathfrak{C}_{N+r(N,c)}^-=A_{(+)}^-(a^-)^{r(N,c)}A_{(-)}^+$
and
$\mathfrak{C}_{N+r(N,c)}^+=A_{(-)}^-(a^+)^{r(N,c)}A_{(+)}^+$.
Index $N+r(N,c)$ in the notation for these ladder operators
indicate that their action is similar
to the action of the operators $(a^\pm)^{N+r(N,c)}$
in the case of the systems $L_0$ and $L_{0^+}$, respectively.
\end{itemize}
As it was shown in
\cite{CarPly2,CIP},
the minimal complete set of the spectrum generating ladder operators
for deformed systems of both classes $c=1,2$
is formed by any of the two sets
$(\mathfrak{A}, \mathfrak{C})$ or $(\mathfrak{B}, \mathfrak{C})$.
However, in nonlinear algebras produced by operators
from any of these two sets, new structures are generated
of the nature similar to powers of $a^\pm$, see Appendix~\ref{apen-red}.
Because of this,
instead of talking about three types of the ladder operators,
it is convenient to talk about three families of the operators given by
\begin{eqnarray}
\label{genlad}
&\mathfrak{A}_{k}^\pm\equiv A_{(-)}^-(a^\pm)^k A_{(-)}^+\,,\qquad \mathfrak{B}_{ k}^\pm\equiv A_{(+)}^-(a^\pm)^k A_{(+)}^+\,,
&\\
&\mathfrak{C}_{N\pm k'}^-\equiv A_{(+)}^-(a^\mp)^{k'}A_{(-)}^+ \,,\qquad
\mathfrak{C}_{N\pm k'}^+\equiv (\mathfrak{C}_{N\pm k'}^-)^\dagger\,,&\label{genlad+}
\end{eqnarray}
where, formally, $k$ can take any nonnegative integer value and $k'$ is such that
$N-k'\geq 0$, otherwise operators (\ref{genlad+}) reduce to $\mathfrak{A}_k^\pm$,
see (\ref{C-}). At $k=0$ and $N-k'=0$ all these operators reduce to certain polynomials in
$L_{(\pm)}$.
Independently on the class of the system, or on whether
the operators are physical or not,
the three families $\mathfrak{D}_{\rho,j}^\pm=( \mathfrak{A}_{j}^\pm, \mathfrak{B}_{j}^\pm,\mathfrak{C}_{j}^\pm )$,
$\rho=1,2,3$, $j=1,2,\ldots$, satisfy the commutation relations of the form
\begin{equation}
\label{sl2rh}
[L_{(\pm)},\mathfrak{D}_{\rho,j}^\pm]=\pm2j \mathfrak{D}_{\rho,j}^\pm \,,
\qquad [\mathfrak{D}_{\rho,j}^-,\mathfrak{D}_{\rho,j}^+]=\mathcal{P}_{\rho,j}(L_{(-)})\,,
\end{equation}
where $\mathcal{P}_{\rho,j}(L_{(-)})$
is a certain polynomial of the corresponding Hamiltonian operator of the system,
whose order
as a polynomial is equal to differential order
of $\mathfrak{D}_{\rho,j}^\pm$ minus one,
see
Appendix \ref{apen-red}.
Algebra
(\ref{sl2rh}) can be considered as a deformation of
$\mathfrak{sl}(2,\mathbb R)$
\cite{JM2}.
\vskip0.1cm
Equation (\ref{sl2rh}) suggests that for physical operators
the factor $2j$ has to be a multiple of the distance
between two consecutive energy levels in the equidistant part of each system,
that is $\Delta E=2c$. Then, for $\mathfrak{A}$ and $\mathfrak{B}$ families,
the physical operators are those
whose index is $j=lc$ with $l \in \mathbb N$, while
for $\mathfrak{C}$ family index should be $j=N+r(N,c)+cs$, where $s$ is
integer such that $j>0$.
For class $c=2$ systems,
the squares of non-physical operators of the $\mathfrak{C}$ type are physical and
reduce
to products of other symmetry generators; see
Appendix \ref{apen-red}.
The unique special case is
$\mathfrak{C}_{2N}^\pm=(-1)^{n_-}(\mathfrak{C}_{N}^\pm)^2$
with odd $N$, in which operators
$\mathfrak{C}_{2N}^\pm$ are not secondary and have to be taken into account.
On the other hand, each ladder operator belonging to the
families $\mathfrak{A}$, $\mathfrak{B}$ and $\mathfrak{C}$ can be constructed
by ``gluing" two complementary generators of
alternative Darboux transformations.
Such generators
have
the form
$A_{(\pm)}^- (a^{\pm})^n$ with $n=cs$ and
$A_{(\pm)}^- (a^{\mp})^n$ with $n=cs+r(N,c)$, where $s=0,1,2\ldots$,
and their physical nature in the sense
of the comments related to (\ref{relation-operators}) is
guaranteed by the indicated selection of values for parameter $n$;
for $s=0$ we recover the basic
intertwining operators.
{}From the point of view of DCKA transformations,
factors $(a^-)^n$ can be produced by selecting the set of eigenstates $(0,1,\ldots,n-1)$
as the seed states,
whose effect is to shift harmonic oscillator Hamiltonian operator for the constant
$2n$. Analogously, factor $(a^+)^n$ corresponds to a negative scheme
$(-0,-1,\ldots,-(n-1))$ which produces a shift for $-2n$.
Despite these families of ladder and intertwining operators seem to be infinite,
one can use Eq. (\ref{ide}) to reduce them
to finite sets
of operators.
The operators reducible to the products
of other physical operators of lower order will not be considered
by us here as basic elements of the set of generators of symmetry.
In other words, we admit that some generators can appear
as coefficients in corresponding (super)algebraic relations.
The related details are presented in Appendix
\ref{apen-red}, and we describe below
the resulting picture.
First, we turn to the sequences of operators
(\ref{genlad}) and (\ref{genlad+}).
For gapless deformations of the AFF model, the
spectrum generating
set is given by any pair of the conjugate operators $\mathfrak{A}^{\pm}_{2}$,
$\mathfrak{B}^{\pm}_{2}$,
or $\mathfrak{C}^{\pm}_{2}$. In
general case, only the operators
\begin{eqnarray}
\label{ladgen}
\left\{
\begin{array}{cc}
\mathfrak{A}_{k}^\pm\,, & 0<k<N\,,\\
\mathfrak{B}_{k}^\pm\,, & 0<k<N\,,\\
\mathfrak{C}_{k}^\pm\,, & 0<k<2N+r(N,c)\,,\\
\end{array}
\right.
\end{eqnarray}
are the basic symmetry generators while the other
can be written in terms of them and polynomials in $L_{(\pm)}$.
For one-gap deformations of the harmonic
oscillator, the set of basic ladder operators can be reduced further
to the set
\begin{eqnarray}
\label{basicsubsetonegap}
\left\{
\begin{array}{cc}
\mathfrak{A}_{k}^\pm\,, & 0<k<n_+\,,\\
\mathfrak{B}_{k}^\pm\,, & 0<k<n_-\,,\\
\mathfrak{C}_{k}^\pm\,, & M<k<n_+\,,\\
\end{array}
\right. \qquad
M=\left\{\begin{array}{ccc}
\max\,(n_-,n_+) & \text{if} & n_-\neq n_+\,,\\
N/2 & \text{if} & n_-=n_+\,,
\end{array}
\right.
\end{eqnarray}
and also we have relations
$\mathfrak{A}_{n_+}^{\pm}=(-1)^{n_-}\mathfrak{C}_{n_+}^{\pm}$ and $\mathfrak{B}_{n_-}^{\pm}=(-1)^{n_-}\mathfrak{C}_{n_-}^{\pm}$.
Consider now the issue of reduction of sequences of the intertwining operators.
For general deformations,
only the
operators
\begin{eqnarray}
\label{genA}
\left\{
\begin{array}{cc}
A_{(\pm)}^-(a^\pm)^{n}\,, & 0\leq n<N\,,\\
A_{(\pm)}^-(a^\mp)^{n}\,,& 0<n<N+r(N,c)\,,
\end{array}
\right.
\end{eqnarray}
and their Hermitian conjugate counterparts
can be considered as basic, see Appendix \ref{apen-red}.
One can note
that the total number of the basic intertwining operators $\#_f=2[(4N-2+r(N,c))/c]$
is greater than the number of the basic ladder operators
$\#_{lad} = 2[(4N-3+r(N,c))/c]$
which can be constructed with their help.
In particular case of gapless deformations of the AFF model,
the indicated set of Darboux generators can be reduced to
those which produce, by `gluing' procedure, one conjugate pair of the
spectrum generating
ladder operators of the form
$\mathfrak{D}_{2,\rho}^\pm$.
For $c=1$ one-gap systems, identity
(\ref{ide}) allows us to reduce further the set of the basic intertwining
operators, which, together with corresponding Hermitian conjugate operators,
is given by any of the two options,
\begin{equation}
\label{frakS}
\mathfrak{S}_{z}\equiv \left\{ \begin{array}{lcc}
A_{(-)}^-{(a^+)}^{|z|}\,, & -N<z\leq 0\,,\\\vspace{-0.4cm}
\\ A_{(-)}^-{(a^-)}^{z}\,, & 0< z \leq n_+\,,\\\vspace{-0.4cm}
\\ A_{(+)}^-{(a^+)}^{N-z} \,, & n_+ < z \leq N\,,\\\vspace{-0.4cm}
\\ A_{(+)}^-{(a^-)}^{N-z}\,, & N<z<2N\,, \\
\end{array}
\right.\\\qquad \text{or}\qquad \mathfrak{S}_{z}^{'}\equiv \mathfrak{S}_{N-z}\,,
\end{equation}
see Appendix \ref{apen-red}.
Here we have reserved $z=0$ and $z=N$ values for index $z$ to the dual schemes'
intertwining operators: in the first choice, $\mathfrak{S}_{0}=A_{(-)}^-$ and
$\mathfrak{S}_{N}=A_{(+)}^-$, and
for the second choice we have $\mathfrak{S}_{0}'=A_{(+)}^-$
and $\mathfrak{S}_{N}'=A_{(-)}^-$. Written in this way, these operators satisfy the
intertwining relations $
\mathfrak{S}_{z}L=(L_{(-)}+2z)\mathfrak{S}_{z}$ or $\mathfrak{S}_{z}'L=(L_{(+)}-2z)\mathfrak{S}_{z}'$,
and their Hermitian conjugate versions.
Then, to study supersymmetry, we have to choose either positive or negative scheme
to define the $\mathcal{N}=2$ superextended Hamiltonian.
We take $\mathfrak{S}_{z}$ if we work with a negative scheme,
and $\mathfrak{S}_{z}'$ if positive scheme is chosen for the construction of super-extension.
\subsection{Supersymmetric extension}
\label{susyextension}
For each of the two dual schemes, one can construct an $\mathcal{N}=2$ superextended
Hamiltonian operator of the form (\ref{Hlambda*}) by choosing appropriately $H_1=\breve{L}-\lambda_*$
and $H_0=L-\lambda_*$.
We put $\breve{L}=L_{(+)}$ and $\lambda_*=\lambda_+= 2l_1^++1$
for positive scheme, and choose $\breve{L}=L_{(-)}$ and $\lambda_*=\lambda_-= -2l_1^--1$
for negative scheme.
For both options, we set $L=L_0$ if we are dealing with a rational extension of
harmonic oscillator, and $L=L_{0^+}$ if we work with a deformation of the AFF model.
We name the matrix Hamiltonian associated with negative scheme as $\mathcal{H}$, and
denote by $\mathcal{H}'$ the Hamiltonian of positive scheme.
The spectrum of these systems can be found using
the properties of the corresponding intertwining operators described in
Section \ref{Dar}, see also refs. \cite{CarPly2,CIP}.
The two Hamiltonians are connected by relation
$\mathcal{H}-\mathcal{H}'=-N(1+\sigma_3)-\lambda_-+\lambda_+$,
and $\sigma_3$ plays a role of
the $\mathcal{R}$ symmetry
generator for both superextended systems.
In this subsection we finally construct the corresponding
spectrum generating superalgebra for $\mathcal{H}$ and $\mathcal{H'}$.
The resulting structures are based on the physical operators $\mathfrak{D}_{\rho,j}^\pm$.
As we shall see,
the supersymmetric versions of the $c=1$ systems
are described by a nonlinearly extended super-Schr\"odinger symmetry with
bosonic generators to be differential operators of even and odd orders, while in the case
of the $c=2$ systems we obtain
nonlinearly extended superconformal symmetry in which bosonic generators
are of even order only.
\vskip0.1cm
We construct a pair of fermionic operators on the basis
of each intertwining operator from the set (\ref{genA}) and
its Hermitian conjugate counterpart.
Let us consider first
the extended nonlinear super-Schr\"odinger symmetry of a one-gap deformed
harmonic oscillator,
and then we generalize the picture.
If we choose the negative scheme, then we use $\mathfrak{S}_z$ defined in (\ref{frakS}) to construct
the set of operators
\begin{eqnarray}
\label{gencharge}
\mathcal{Q}_1^{z}=
\left(
\begin{array}{cc}
0& \mathfrak{S}_z \\
\mathfrak{S}_z^\dagger & 0
\end{array}
\right)\,,
\qquad
\mathcal{Q}_2^{z}= i\sigma_3\mathcal{Q}_1^{z}\,, \qquad -N<z<2N\,.
\end{eqnarray}
They satisfy the (anti)-commutation relations
\begin{equation}
\label{SUSY}
[\mathcal{H},\mathcal{Q}_a^z]=2iz\epsilon_{ab}\mathcal{Q}_{b}^{z}\,,
\qquad \{\mathcal{Q}_a^z,\mathcal{Q}_b^z\}=2\delta_{ab}\mathbb{P}_{z}(\mathcal{H},\sigma_3)\,,\qquad
[\Sigma,\mathcal{Q}_{a}^z]=-i\epsilon_{ab}\mathcal{Q}_{b}^{z}\,,
\end{equation}
where $\mathbb{P}_z$ are some polynomials whose structure is described in
Appendix \ref{apen-comm}.
For the choice of the positive scheme to fix extended Hamiltonian, according to
(\ref{frakS}), the corresponding fermionic operators are given by
$\mathcal{Q}_1^{'z}\equiv \mathcal{Q}_1^{N-z}$. They satisfy relations
of the same form (\ref{SUSY}) but with replacement
$\mathcal{H}\rightarrow \mathcal{H}'$, $\Sigma=\frac{1}{2}\sigma_3 \rightarrow \Sigma'=-\frac{1}{2}\sigma_3$,
$\mathbb{P}_{z}(\mathcal{H},\sigma_3)\rightarrow
\mathbb{P}_{z}'(\mathcal{H}',\sigma_3)=\mathbb{P}_{N-z}(\mathcal{H}'-N(1+\sigma_3)-\lambda_-+\lambda_+,\sigma_3)$,
$\mathcal{Q}_{1}^{z}\rightarrow \mathcal{Q}_{2}^{'z}$ and
$\mathcal{Q}_{2}^{z}\rightarrow \mathcal{Q}_{1}^{'z}$.
The fermionic operators $\mathcal{Q}_a^{0}$ (or $\mathcal{Q}_a^{'0}$)
are the supercharges
of the (nonlinear in general case) $\mathcal{N}=2$ Poincar\'e supersymmetry,
which are
integrals of motion of the system $\mathcal{H}$
(or $\mathcal{H}'$), and
$\mathbb{P}_{0}=
P_{n_-}(\mathcal{H} + \lambda_{-})\,$
(or $\mathbb{P}_{0}=P_{n_+}(\mathcal{H}' +
\lambda_{+})$) with polynomials $P_{n_\pm}$ defined in (\ref{polyA}).
The operators $\mathcal{Q}_a^{'0}$ are analogous here to supercharges
in (\ref{supercharge1}).
On the other hand, we have here the
fermionic operators $\mathcal{Q}_a^{'N}$ as analogs
of dynamical integrals $\mathcal{S}^a_\nu$ there.
We recall that in the simple linear case considered in Section \ref{isoharconf},
the interchange between positive and negative schemes
corresponds to the automorphism of superconformal algebra,
and this observation will be helpful for us
for the analysis of the nonlinearly extended
super-Schr\"odinger
structures.
Here, actually,
each of the $(\#_f-2)/2$ pairs of fermionic operators distinct from supercharges
provides a possible dynamical extension of the super-Poincar\'e symmetry.
As we will see, all of them are necessary to obtain a closed
nonlinear
spectrum generating
superalgebra of the superextended system.
To construct any extension of the deformed Poincar\'e supersymmetry,
we calculate $\{\mathcal{Q}_{a}^{0},\mathcal{Q}_{a}^{z} \}$, in the negative scheme,
or $\{\mathcal{Q}_{a}^{'0},\mathcal{Q}_{a}^{'z} \}$ in the positive one.
In the first case we have
\begin{equation}
\label{Cn+kQN}
\{\mathcal{Q}_a^{0},\mathcal{Q}_{b}^{z}\}=\delta_{ab}(\mathcal{G}_{-z}^{(2\theta(z)-1)}+
\mathcal{G}_{+z}^{(2\theta(z)-1)})+i\epsilon_{ab}(\mathcal{G}_{-z}^{(2\theta(z)-1)}- \mathcal{G}_{+z}^{(2\theta(z)-1)})\,,
\end{equation}
where $z\in (-N,0)\cup(0,2N)$, $\theta(z)=1\, (0)$ for $z>0\, (z<0)$,
and $\mathcal{G}^{(2\theta(z)-1)}_{\pm z}$
are given by
\begin{equation}
\label{superC}
\mathcal{G}_{+z}^{(2\theta(z)-1)}=
\left(
\begin{array}{cc}
\mathfrak{S}_{0}(\mathfrak{S}_{z})^{\dagger} & 0 \\
0& (\mathfrak{S}_{z})^{\dagger}\mathfrak{S}_{0}
\end{array}
\right)\,,\qquad
\mathcal{G}_{-z}^{(2\theta(z)-1)}=(\mathcal{G}_{+z}^{(2\theta(z)-1)})^\dagger\,.
\qquad
\end{equation}
Following definition (\ref{frakS}), one finds directly that
$\mathfrak{S}_{0}(\mathfrak{S}_{z})^{\dagger}$ is equal to $\mathfrak{A}_{|z|}^-$ when $-N<z<0$,
while for $0<z\leq n_+$, this operator is equal to $\mathfrak{A}_{z}^+$,
and takes the form of $\mathfrak{C}_{z}^+$ for $n_+<z<2N$.
The operators $(\mathfrak{S}_{z})^{\dagger}\mathfrak{S}_{0}$
reduce to
\begin{equation}
\label{S*S+}
(\mathfrak{S}_{z})^{\dagger}\mathfrak{S}_{0}= \left\{ \begin{array}{lcc}
P_{n_-}(L-2k)(a^-)^{|z|}\,, & -N<z<0\,,\\\vspace{-0.4cm}
\\(a^+)^{z}P_{n_-}(L)\,, & 0<z\leq n_+\,,\\\vspace{-0.4cm}
\\ (-1)^{n_-}(a^+)^zT_{N-z}(L+2N)\,, & n_+<z<N\,,\\\vspace{-0.4cm}
\\ (-1)^{n_-}(a^+)^z \,, & N\leq z<2N\,.
\end{array}
\right.\\
\end{equation}
Note that $\mathcal{G}_{\pm k}^{(-1)}$ and $\mathcal{G}_{\pm k}^{(+1)}$ with $k=|z|\leq n_-$
are two different matrix extensions of the same operator $\mathfrak{A}_k^\pm$.
For a superextended system based on the positive scheme, we obtain
\begin{equation}
\label{Q0'Qk'}
\{\mathcal{Q}_a^{'0},\mathcal{Q}_{b}^{'z}\}=
\delta_{ab}(\mathcal{G}^{'(2\theta(z)-1)}_{-z}+\mathcal{G}^{'(2\theta(z)-1)}_{+z})
-i\epsilon_{ab}(\mathcal{G}^{'(2\theta(z)-1)}_{-z}- \mathcal{G}^{'(2\theta(z)-1)}_{+z})\,,
\end{equation}
where, again, $z\in (-N,0)\cup(0,2N)$, and $\mathcal{G}^{'(2\theta(z)-1)}_{\pm z}$
are given by
\begin{equation}
\label{superC'}
\mathcal{G}^{'(2\theta(z)-1)}_{-z}=
\left(
\begin{array}{cc}
\mathfrak{S'}_{0}(\mathfrak{S'}_{z})^{\dagger} & 0 \\
0& (\mathfrak{S'}_{z})^{\dagger}\mathfrak{S'}_{0}
\end{array}
\right)\,,\qquad
\mathcal{G}^{'(2\theta(z)-1)}_{+z}=(\mathcal{G}^{'(2\theta(z)-1)}_{-z})^\dagger\,.
\qquad
\end{equation}
Now,
$\mathfrak{S'}_{0}(\mathfrak{S'}_{z})^{\dagger}=\mathfrak{B}^{+}_{|z|}$ when $-N<z<0$,
while for positive index $z$ this operator reduces to
$\mathfrak{B}^{-}_{z}$
when $0<z\leq n_-,$ and to $\mathfrak{C}^{-}_{z}$
when $n_-<z<2N$. For the other matrix element we have
\begin{eqnarray}
(\mathfrak{S}_{z}')^{\dagger}\mathfrak{S}_{0}'= \left\{ \begin{array}{lcc}
(a^+)^{|z|}P_{n_+}(L)\,, & -N<z<0\,,\\\vspace{-0.4cm}
\\(a^-)^{z}P_{n_+}(L)\,, & 0<z\leq n_-\,,\\\vspace{-0.4cm}
\\ (-1)^{n_-}T_{N-k}(L)(a^-)^z\,, & n_-<z<N\,,\\\vspace{-0.4cm}
\\ (-1)^{n_-}(a^-)^z\,, & N<z<2N\,.
\end{array}
\right.
\end{eqnarray}
Here, again, there are two different matrix extensions of the
operators of the $\mathfrak{B}$-family given by
$\mathcal{G}_{\pm k}^{'(+1)}$ and $\mathcal{G}_{\pm k}^{'(-1)}$ when $k\leq n_-$.
By comparing both schemes
one can note two other special features.
It turns out that $\mathcal{G}_{\pm k}^{(1)}=\mathcal{G}_{\pm k}^{'(1)}$ when
$k\geq N$, and this corresponds to the automorphism discussed
in Section~\ref{isoharconf}. In the same way,
for $\max(n_-,n_+)<k<N$, operators $\mathcal{G}_{\pm k}^{(1)}$ and $\mathcal{G}^{'(1)}_{\pm k}$
are different matrix extensions of $\mathfrak{C}^{\pm}_k$.
{}From here and in what follows
we do not specify whether we have the superextended system
corresponding to the negative or the positive scheme, and will just use, respectively,
the unprimed or primed
notations for operators of the alternative dual schemes.
In particular, we have
\begin{equation}
[\mathcal{H},\mathcal{G}_{\pm k}^{(2\theta(z)-1)}]=\pm 2k \mathcal{G}_{\pm k}^{(2\theta(z)-1)}
\,,\qquad k\equiv|z|\,, \qquad z \in (-N,0)\cup(0,2N)\,,
\end{equation}
that shows explicitly that our new bosonic operators have the nature of ladder operators
of the superextended system $\mathcal{H}$.
Commutators $[\mathcal{G}_{-k}^{(1)},\mathcal{G}_{+k}^{(1)}]$ and
$[\mathcal{G}_{-k}^{(-1)},\mathcal{G}_{+k}^{(-1)}]$
produce polynomials in $\mathcal{H}$ and $\sigma_3$,
which can be calculated by using the polynomials
$\mathcal{P}_{\rho,j}$ defined in (\ref{sl2rh}).
The algebra generated by $\mathcal{H}$, $\mathcal{G}_{\pm k}^{(2\theta(z)-1)}$
and $\sigma_3$ is identified
as a deformation of $\mathfrak{sl}(2,\mathbb R)\oplus \mathfrak{u}(1)$,
where a concrete form of deformation depends on
the system, $\mathcal{H}$, and on $z$.
Each of these nonlinear
bosonic algebras expands further up to
a certain closed nonlinear deformation of superconformal $\mathfrak{osp}(2|2)$ algebra
generated by the subset
of operators
\begin{equation}
\label{U1}
\mathcal{U}_{0,z}^{(2\theta(z)-1)}\equiv \{\mathcal{H},\sigma_3,
\mathbb{I},\mathcal{G}_{\pm |z|}^{(2\theta(z)-1)}, \mathcal{Q}_{a}^{0},
\mathcal{Q}_{a}^{z}\}\,,\qquad z \in (-N,0)\cup(0,2N)\,,
\end{equation}
see Appendix \ref{apen-comm}.
The deficiency of any of these nonlinear superalgebras
is that none of them is a spectrum generating algebra for the
superextended system\,: application
of operators from the set (\ref{U1}) and of their products
does not allow to connect two arbitrary eigenstates in the spectrum
of $\mathcal{H}$.
To find the spectrum generating superalgebra
for this kind of the superextended systems, one can try
to include into the superalgebra simultaneously
the operators $\mathcal{G}^{(1)}_{\pm N}$
and, say, $\mathcal{G}^{(1)}_{\pm 1}$ or $\mathcal{G}^{(-1)}_{\pm 1}$.
The operators $\mathcal{G}^{(1)}_{\pm N}$
provide us with matrix extension of the operators $\mathfrak{C}_{N}^\pm$
being ladder operators for deformed subsystems $L_{(-)}$ or $L_{(+)}$.
Analogously, operators $\mathcal{G}^{(1)}_{\pm 1}$ or $\mathcal{G}^{(-1)}_{\pm 1}$
supply us with matrix extensions of the ladder operators
$\mathfrak{A}_{ 1}^\pm$ or $\mathfrak{B}_{ 1}^\pm$
($\mathfrak{A}_{ 2}^\pm$ or $\mathfrak{B}_{ 2}^\pm$)
when systems $L_{(\pm)}$ are of the class $c=1$ or $c=2$ with even (odd) $N$.
Therefore, it is enough to unify the sets of generators
$\mathcal{U}_{0,1}^{(1)}$ and $\mathcal{U}_{0,N}^{(1)}$.
Having in mind the commutation relations
between operators of the three families
$\mathfrak{A}$, $\mathfrak{B}$ and $\mathfrak{C}$, one can find, however,
that the commutators of the operators
$\mathcal{G}^{(1)}_{\pm N}$ with $\mathcal{G}^{(1)}_{\pm 1}$
generate other bosonic matrix operators $\mathcal{G}^{(1)}_{\pm k}$.
The commutation of these operators with supercharges $\mathcal{Q}_{a}^{0}$
generates the rest of the fermionic operators we considered, see
Appendix \ref{apen-comm} for details.
The set of higher order generators
is completed by considering all non-reducible bosonic and fermionic generators,
which do not decompose into the products of other generators.
In correspondence with that was noted above, we
arrive finally at two different extensions of the sets of
operators with index less than $N$. By this reason it is convenient also
to introduce the operators
\begin{eqnarray}
\label{G_k}
&\mathcal{G}_{\pm k}^{(0)}\equiv \Pi_- (a^\pm)^k, \qquad
k=1,\ldots,N-1, \qquad
\Pi_-=\frac{1}{2}(1-\sigma_3)\,,&
\end{eqnarray}
which help us to fix in a unique way the bosonic set of generators. For our purposes we choose
to write all the operators $\mathcal{G}_{\pm k}^{(-1)}$ in terms of $\mathcal{G}_{\pm k}^{(1)}$ and
$\mathcal{G}_{\pm k}^{(0)}$ when $k\leq n_+$ in the negative scheme,
and when $k\leq n_-$ in the extended system associated with the positive scheme.
For indexes outside the indicated scheme-dependent range, we neglect operators $\mathcal{G}_{\pm k}^{(-1)}$
because they are not basic in correspondence with discussion on reduction of ladder operators in the
previous subsection \ref{interladder}.
As a result, we have to drop out in (\ref{U1}) all the operators $\mathcal{G}_{\pm |z|}^{(2\theta(z)-1)}$ with
$z\in (-N,0)$.
By taking anti-commutators of fermionic operators $\mathcal{Q}_{a}^{N}$
with $\mathcal{Q}_{a}^{z}$, $z\neq 0$, we produce bosonic
dynamical integrals $\mathcal{J}_{\pm |z-N|}^{(1-2\theta(z-N))}$,
which have exactly the same
structure of the even generators
$\mathcal{G}_{\pm |z|}'^{(2\theta(z)-1)}$
in the extension associated with the dual scheme.
In this way we obtain the subsets of
operators
\begin{eqnarray}
\label{I1}
\mathcal{I}_{N,z}^{(1-2\theta(z-N))}\equiv \{\mathcal{H},\sigma_3,\mathbb{I},
\mathcal{J}_{\pm |z-N|}^{(1-2\theta(z-N))}, \mathcal{Q}_{a}^{N},\mathcal{Q}_{a}^{z}\}\,
\qquad z \in (-N,0)\cup(0,2N)\,,
\end{eqnarray}
which also generate closed nonlinear super-algerabraic structures.
With the help of (\ref{G_k}), we find similarly to the subsets (\ref{U1}),
that a part of the sets (\ref{I1}) also can be reduced.
Having in mind the ordering relation between $n_-$ and $n_+$,
the superextended systems associated with
the negative schemes
can be characterized finally by the following irreducible,
in the sense of subection \ref{interladder},
subsets of symmetry generators\,:
\begin{table}[htbp]
\begin{center}
\begin{tabular}{ll}
\hline
$n_-\leq n_+$& $n_+<n_- $\\\hline
$\mathcal{U}_{0,k}^{(1)}\,, 0<k<2N$ & $\mathcal{U}_{0,k}^{(1)}\,, k\in (0,n_+)\cup (n_-,2N)$ \\
$\mathcal{I}_{N,z}^{(1-2\theta(N-z))}\,, z\in(-N,0)\qquad\qquad$ & $\mathcal{I}_{N,z}^{(1-2\theta(N-z))}\,, z\in(-N,0)\cup$\\
$\cup(n_+,N)$ & $[n_+,N)$
\\\hline
\end{tabular}
\end{center}
\end{table}
\noindent
For more details,
see Appendix \ref{apen-red}.
A similar result can be obtained for superextended systems
associated with positive schemes, where
the roles played by families $\mathfrak{A}$ and $\mathfrak{B}$,
and of numbers $n_-$ and $n_+$
are interchanged.
Finally, we arrive at the following picture. Any operator that can be obtained via
(anti)commutation relations of the basic generators from the above Table but does not belong to them
can be written as their product.
As a result, the listed basic operators produce the spectrum generating superalgebra
for one-gap rationally deformed super-extended systems.
It is worth it to stress that in this set of generators the unique
true integrals of motion, in addition to $\mathcal{H}$ and $\sigma_3$,
are the supercharges $\mathcal{Q}_a^{0}$, while the rest has to
be promoted to the dynamical integrals by using transformation
(\ref{recipeintegrals}).
For gapless rational extensions of the systems of class $c=2$,
only the subset $\mathcal{U}_{0,2}^{(1)}$ has to be considered
instead of the family of sets $\mathcal{U}_{0,k}^{(1)}$.
For super-extensions of rationally deformed systems of arbitrary form in the sense of the class $c$
and arbitrary number of gaps and their dimensions,
the identification of their generalized super-Schr\"odinger or superconformal
structures is realized in a similar way. The procedure is based
on the sets of operators (\ref{ladgen}) and
(\ref{genA}), which include the operators
(\ref{basicsubsetonegap}) and (\ref{frakS})
of the discussed one-gap case as subsets.
As a result,
for every irreducible pair of ladder operators (\ref{ladgen}) with index less than $N$
we have two super-extensions which are related by operators of the form (\ref{G_k}). When we
put together the
subsets containing the spectrum generating set of operators,
we obtain all the other structures.
\section{Examples
}
\label{examples}
In this section we employ the developed machinery
for two simple examples of rationally extended systems.
We consider here the cases in which the negative scheme is characterized by only
one seed state, $n_-= 1$,
and then $N=n_++1$.
The systems generated by Darboux transformations of this kind could be a one-gap rational extension
of the harmonic oscillator, or a gapless deformation of $L_{1}^{iso}$, and according
to the general picture described in the previous section,
one can
make some
general assertions.
\vskip0.1cm
\begin{itemize}
\item[(i)]\emph{Peculiarities of one-gap deformations of the QHO}\,:
The superextended Hamiltonian constructed on the
base of the negative scheme with $n_-=1$
is characterized by unbroken
$\mathcal{N}=2$
Poincar\'e supersymmetry whose
supercharges, being the first
order differential operators, generate
a Lie superalgebra.
The $\mathfrak{B}$ family of ladder operators in the sense of (\ref{basicsubsetonegap})
does not play any
role in this scheme.
On the other hand, the super-Hamiltonian provided by the positive scheme
possesses $n_+$
singlet states while the ground state is a doublet.
The $\mathcal{N}=2$ super-Poincar\'e
algebra of such a system is nonlinear
as its supercharges are of differential order
$n_+=2\ell \geq 2$.
\vskip0.1cm
\item[(ii)]\emph{Peculiarities of gapless deformations of $L_1^{iso}$}\,:
The negative scheme produces a super-Hamiltonian with spontaneously broken
supersymmetry, whose all energy levels are doubly degenerate;
its $\mathcal{N}=2$ super-Poincar\'e algebra has linear nature.
To construct the spectrum generating algebra we only need a matrix extension
of the operators $\mathfrak{A}_2^\pm$. In a superextended system produced by
the positive scheme, $n_+>1$ physical and non-physical states of $L_{0^+}$ of positive energy
(the latter being even eigenstates of harmonic oscillator)
are used as seed states for Darboux transformation.
Its supersymmetry is spontaneously broken,
and the $\mathcal{N}=2$ super-Poincar\'e algebra is nonlinear.
The nonlinearly deformed super-Poincar\'e symmetry cannot be expanded
up to spectrum generating superalgebra by combining it with matrix
extension of the $\mathfrak{A}^\pm_2$,
but this can be done by using matrix extensions of the
$\mathfrak{B}_2^\pm$ or $\mathfrak{C}_2^\pm$ ladder operators,
see (\ref{superC'}).
The resulting spectrum generating superalgebra is a certain nonlinear
deformation
of the $\mathfrak{osp}(2|2)$ superconformal symmetry.
\vskip0.1cm
\end{itemize}
In what follows,
we first consider the simplest example of the deformed gapless AFF model, and
then we discuss the super-extension with a rational one-gap deformation
of the harmonic oscillator. As we shall see, in the second case
the algebraic structure of the deformed superextended Schr\"odinger symmetry
is more complicated than that of the deformed superconformal
symmetry associated with gapless deformation of the AFF system.
\subsection{Gapless deformation of AFF model}
The super-extension we consider here is based on the
dual schemes $(1,2,3)\sim (-3)$ composed from
eigenstates of the half-harmonic oscillator $L_{0^+}$.
We have here $N=4$,
and seed states $\psi_2$ and $\psi_{-3}$ are non-physical.
The Darboux transformation of the negative scheme generates
the Hamiltonian
\begin{eqnarray}
\label{L(-3)}
&L_{(-)}\equiv L_{(-3)}=-\frac{d^2}{dx^2}+x^2-2(\ln \psi_{-3})''=
L_{1}^{iso}+8 \frac{2 x^2-3}{(3 + 2 x^2)^2}-2\,,&
\end{eqnarray}
and the pair of the corresponding intertwining operators is
\begin{eqnarray}
\label{A-3}
&A_{(-)}^-\equiv A_{(-3)}^-=A_{-1}^--\frac{4x}{2x^2+3}\,,\qquad
A_{(-)}^+\equiv A_{(-3)}^+=A_{-1}^+-\frac{4x}{2x^2+3}\,.&
\end{eqnarray}
For them the second intertwining relation from (\ref{inter0}) and its Hermitian conjugate version are valid.
Following the way described in the previous section, we obtain
$P_{n_-}(L_{(-)})=L_{(-)}+7\equiv H_1$ and
$P_{n_-}(L_{0^+})=L_{0^+}+7\equiv H_0$.
Physical eigenstates $\phi_l$ of the deformed system $H_1$ are obtained
from odd eigenfunctions
of $H_0$, $\phi_l=A_{(-)}^-\psi_{2l+1}$,
$l=0,1,2,\ldots$. The state $\phi_l=A_{(-)}^-\widetilde{\psi_{(-3)}}=1/\psi_{(-3)}$
diverges in $x=0$, and therefore is not a physical state of $L_{(-)}$.
Since the state
$\psi_{-3}$ is not physical, $H_1$ is completely isospectral to $H_0$.
In fact, this is the simplest example of this category.
In positive scheme we have the intertwining operators
$A_{(+)}^\pm=A_{(1,2,3)}^\pm$, which are constructed iteratively
using three seed states according to the prescription (\ref{Andef}).
These operators satisfy the first intertwining relation from
(\ref{inter0}) as well as its conjugate version,
and their products are
$P_{n_+}(L_{0^+})=(L_{0^+}-3)(L_{0^+}-5)(L_{0^+}-7)$
and
$P_{n_+}(L_{(+)})=(L_{(+)}-3)(L_{(+)}-5)(L_{(+)}-7)$,
where we use Eq.
(\ref{A-A-A+A+Poly}) and $L_{(+)}=L_{(-)}+8$.
Consider now the superextended system constructed on the base
of the negative scheme.
At the end of this subsection we discuss the super-extension
based on the positive scheme.
The superextended Hamiltonian and its spectrum are given by
\begin{equation}
\label{hamilisodef}
\mathcal{H}=
\left(
\begin{array}{cc}
H_1& 0 \\
0 & H_0
\end{array}\right),\qquad \mathcal{E}_{n}=4n+10\,,
\qquad n=0,1,\ldots\,.
\end{equation}
Due to complete isospectrality of $ H_1$ and $H_0$, all the energy levels
of the system (\ref{hamilisodef})
including the lowest one $\mathcal{E}_{0}=10>0$
are doubly degenerate. The Witten's index
equals zero, and we have here the case
of spontaneously broken $\mathcal{N}=2$ super-Poincar\'e symmetry
generated by Hamiltonian $\mathcal{H}$, the supercharges $\mathcal{Q}^0_a$
constructed in terms of $A_{(-)}^\pm$,
and by $\Sigma=\frac{1}{2}\sigma_3$.
The complete set of fermionic operators $\mathcal{Q}_a^{z}$ with $-3\leq z\leq 7$ is given by
Eq. (\ref{gencharge}), and the ladder operators $\mathcal{G}_{\pm k}^{(1)}$ with $k=1,\ldots,7$
are constructed according to (\ref{superC}).
However,
due to the isospectrality of the subsystems, we need here only the restricted set
of operators
$\mathcal{U}_{0,2}^{(1)}=\{\mathcal{H},\mathbb{I},\mathcal{G}_{\pm2}^{(1)},\sigma_3,\mathcal{Q}_a^{0},\mathcal{Q}_a^{2} \}$
to obtain the spectrum generating superalgebra of the system,
where
\begin{eqnarray}
\label{Q2Cpm2}
\mathcal{Q}^z_1=
\left(
\begin{array}{cc}
0& A^-_{(-)}(a^-)^z \\
(a^+)^zA^+_{(-)} & 0
\end{array}\right),\,\, z=0,2\,;\quad
\mathcal{G}_{-2}^{(1)}=
\left(
\begin{array}{cc}
A_{(-)}^-(a^-)^2A^+_{(-)}& 0 \\
0 & H_0(a^-)^2
\end{array}\right),
\end{eqnarray}
and $\mathcal{Q}^z_2=i\sigma_3 \mathcal{Q}^z_1$, $\mathcal{G}_{+2}^{(1)}=(\mathcal{G}_{-2}^{(1)})^\dagger$.
They satisfy the superalgebraic relations
\begin{eqnarray}
\label{nonlinear3}
&[\mathcal{H},\mathcal{Q}_a^{0}]=0\,,\qquad
[\mathcal{H},\mathcal{Q}_a^{2}]=4i\epsilon_{ab}\mathcal{Q}_b^{2}\,,\qquad
[\sigma_3,\mathcal{Q}_a^{z}]=-2i\epsilon_{ab}\mathcal{Q}_b^{z}\,,\quad z=0,2\,,&\\
\label{nonlinear2}
&\{\mathcal{Q}_a^{0},\mathcal{Q}_a^{0}\}=2\delta_{ab}\mathcal{H}\,,\qquad
\{\mathcal{Q}_a^{0},\mathcal{Q}_b^{2}\}=\delta_{ab}(\mathcal{G}_{-2}^{(1)}+\mathcal{G}_{+2}^{(1)})+i
\epsilon_{ab}(\mathcal{G}_{-2}^{(1)}-\mathcal{G}_{+2}^{(1)})\,,&\\
\label{nonlinear1}
&[\mathcal{H},\mathcal{G}_{\pm 2}^{(1)}]=\pm4\mathcal{G}_{\pm2}^{(1)}\,,\qquad
[\mathcal{G}_{\mp 2}^{(1)},\mathcal{Q}_a^{0}]=\pm 2(\mathcal{Q}_a^{2}\mp i\epsilon_{ab}\mathcal{Q}_b^{2})\,,&\\
\label{nonlinear4}
&[\mathcal{G}_{-2}^{(1)},\mathcal{G}_{+2}^{(1)}]=8(\mathcal{H}-4)(\mathcal{H}(2\mathcal{H}-9)+\Pi_-
(\mathcal{H}^2-4\mathcal{H}+24))\,,&\\
\label{nonlinear5}
&[\mathcal{G}_{\mp 2}^{(1)},\mathcal{Q}_a^{2}]= \pm 2(-80 + 4 \mathcal{H} + \mathcal{H}^2)(\mathcal{Q}_a^{0}\pm
i\epsilon_{ab}\mathcal{Q}_b^{0})\,,&\\
\label{nonlinear7}
&\{\mathcal{Q}_a^{2},\mathcal{Q}_b^{2}\}=2\delta_{ab}
(\eta+1)(\eta+3)(\eta+7)|_{\eta=\mathcal{H}+2\sigma_3-9}\,,&
\end{eqnarray}
where $\Pi_-=\frac{1}{2}(1-\sigma_3)$.
The common eigenstates of $\mathcal{H}$ and $\mathcal{Q}^0_1$ are
\begin{equation}
\label{states-3}
\Psi_{n}^{+}=
\left(
\begin{array}{c}
(\mathcal{E}_n)^{-1/2} A_{(-)}^-\psi_{2n+1} \\
\psi_{2n+1}
\end{array}
\right),
\qquad
\Psi_{n}^{-}=\sigma_3\Psi_{n}^+,
\end{equation}
where $\mathcal{Q}^0_1\Psi^\pm_n=\pm\sqrt{\mathcal{E}_n}\Psi^\pm_n$,
and we have here the relations
$\Psi_{n}^\pm=(\mathcal{G}_{+2}^{(1)})^n\Psi_0^{\pm}$ and $\mathcal{G}_{-2}^{(1)}\Psi_0^{\pm}=0$.
As a result one can generate all the complete set of eigenstates of the system
by applying the generators of superalgebra to any of the two ground states
$\Psi_0^+$ or $\Psi_0^-$, and therefore the restricted set of
generators we have chosen is the complete spectrum generating
set for the superextended system (\ref{hamilisodef}).
We also note here that the action of generators not included in the set
$\mathcal{U}_{0,2}^{(1)}$ can be reconstructed by considering the action of $\mathcal{U}_{0,2}^{(1)}$'s elements
in the Hilbert space of the system.
\vskip0.1cm
The complete set of (anti)-commutation relations
(\ref{nonlinear1})-(\ref{nonlinear7}) corresponds to a nonlinear deformation of superconformal
algebra $\mathfrak{osp}(2|2)$.
The first relation from (\ref{nonlinear1})
and equation (\ref{nonlinear4}) represent
a nonlinear deformation of $\mathfrak{sl}(2,\mathbb{R})$
with commutator $[\mathcal{G}_{-2}^{(1)},\mathcal{G}_{+2}^{(1)}]$
to be a cubic polynomial in $\mathcal{H}$.
{}From the superalgebraic relations it follows that like in the linear case
of superconformal $\mathfrak{osp}(2|2)$ symmetry
discussed in Section \ref{Section3.2},
here the extension of the
set of generators $\mathcal{H}$, $\mathcal{Q}^0_a$ and $\Sigma$ of
the $\mathcal{N}=2$ Poincar\'e super-symmetry
by any one of the dynamical integrals $\mathcal{Q}^2_a$, $a=1,2$,
$\mathcal{G}_{+2}^{(1)}$ or $\mathcal{G}_{-2}^{(1)}$ recovers
all the complete set of generators of the nonlinearly deformed
superconformal $\mathfrak{osp}(2|2)$ symmetry.
\vskip0.2cm
Due to a gapless deformation of the AFF model,
here similarly to the case of the non-deformed superconformal $\mathfrak{osp}(2|2)$ symmetry,
the super-extension
based on the positive scheme
is characterized by essentially different physical properties.
For the positive scheme
we take $\lambda_*=3$ in
(\ref{Hlambda*}), and identify
$\mathcal{H}'=\text{diag}\,(L_{(+)}-3,L_{0^+}-3)$
as the extended Hamiltonian.
This $\mathcal{H}'$ is related to
$\mathcal{H}$ defined by Eq. (\ref{hamilisodef})
by the equality $\mathcal{H}'=\mathcal{H}-6+4\sigma_3$.
For extended system $\mathcal{H}'$,
supercharges ${\mathcal{Q}'}_a^{0}$
have the form similar to $\mathcal{Q}_a^{0}$ in (\ref{Q2Cpm2})
but with $A^\pm_{(-)}$ changed for the intertwining operators
$A^\pm_{(+)}$.
Being differential operators of the third order,
they satisfy relations
$[\mathcal{H}', {\mathcal{Q}'}_a^{0}]=0$ and
$\{{\mathcal{Q}'}_a^{0},{\mathcal{Q}'}_b^{0}\}=2\delta_{ab}P_{n_+}(\mathcal{H}'+3)$
with $P_{n_+}(\mathcal{H}'+3)=\mathcal{H}'(\mathcal{H}'-2)(\mathcal{H}'-4)$.
The linear $\mathcal{N}=2$ super-Poincar\'e algebra
of the system (\ref{hamilisodef}) is changed here for the nonlinearly
deformed superalgebra with anti-commutator to be polynomial
of the third order in Hamiltonian. This system has two nondegenerate states
$(0,\psi_{1})^t$ and $(0,\psi_{3})^t$ of energies, respectively, $0$ and $4$,
and both them are annihilated by both supercharges ${\mathcal{Q}'}_a^{0}$.
All higher energy levels $\mathcal{E}'_n=4n$ with $n=2,3,\ldots$
are doubly degenerate.
Thus, the nonlinearly deformed $\mathcal{N}=2$ super-Poincar\'e
symmetry of this system can be identified as partially unbroken \cite{KliPly}
since the supercharges have differential order three but
annihilate only two nondegenerate physical states.
Here instead of the spectrum generating set $\mathcal{U}_{0,2}^{(1)}$ formed by true
and dynamical integrals the same role is played by the set
of integrals
$\mathcal{U}_{0,2}'^{(1)}=\{\mathcal{H}',\mathcal{G}'^{(1)}_{\pm 2}, \mathbb{I},\sigma_3, {\mathcal{Q}'}_a^{0},
{\mathcal{Q}'}_a^{2}\}$, where fermionic generators are
${\mathcal{Q}'}_a^{z}=\mathcal{Q}_a^{4-z}$ with $z=0,2$ according with (\ref{frakS}) and (\ref{gencharge}).
Bosonic dynamical integrals $\mathcal{G}'^{(1)}_{\pm 2}$ are given here by
\begin{eqnarray}
\label{Cpm2Hprime}
\mathcal{G}_{-2}'^{(1)}=
\left(
\begin{array}{cc}
A_{(+)}^-(a^+)A^+_{(-)}& 0 \\
0 & (L_{0^+}-1)(a^-)^2
\end{array}\right),\qquad
\mathcal{G}_{+2}'^{(1)}=(\mathcal{G}_{-2}'^{(1)})^\dagger\,,
\end{eqnarray}
where Eq. (\ref{superC'}) have been used for the case
of the present positive scheme. They are generated via anticommutation of
${\mathcal{Q}'}^0_a$ with ${\mathcal{Q}'}^2_b$.
The set of operators $\mathcal{U}_{0,2}'^{(1)}$ generates the nonlinearly deformed
superconformal $\mathfrak{osp}(2|2)$ symmetry
given by superalgebra of the form
(\ref{nonlinear3})--(\ref{nonlinear7})
but with coefficients to be polynomials of higher order in Hamiltonian
$\mathcal{H}'$
in comparison with the case of the system (\ref{hamilisodef}).
\subsection{Rationally extended harmonic oscillator}\label{SecDefQHO}
The example we discuss in this subsection corresponds to the rational extension
of the harmonic oscillator based on the dual schemes $(1,2)\sim (-2)$,
for which $N=3$.
Different aspects of this system were extensively studied in literature
\cite{CarPly,CarPly2}, but it was not considered yet in the light of the
nonlinearly extended super-Schr\"odinger
symmetry we investigate here.
The Hamiltonian produced via Darboux transformation based on the
negative scheme is
\begin{eqnarray}
\label{H_-2}
&L_{(-)}=L_{(-2)}=-\frac{d^2}{dx^2}+x^2+8\frac{2 x^2-1 }{(1 + 2 x^2)^2}-2\,,&
\end{eqnarray}
whose spectrum is
$E_0=-5$, $E_{n+1}=2n+1$, $n=0,1,\ldots$.
In this system a gap of size 6
separates the ground state energy from the equidistant part of the spectrum,
where levels are separated from each other by a distance $\Delta E=2$.
The pair of ladder operators of the $\mathfrak{C}$-family
connects here the isolated ground state with the equidistant part of the spectrum,
and together with the ladder operators $\mathfrak{A}^\pm_1$ they form
the complete spectrum generating set of operators for the system.
The intertwining operators of the negative scheme are
\begin{eqnarray}
&A_{(-)}^-\equiv A_{(-2)}^{-}=\frac{d}{dx}-x-\frac{4x}{2x^2+1},\qquad A_{(-)}^+\equiv
A_{(-2)}^{+}=(A_{(-2)}^{-})^{\dagger}\,.&
\end{eqnarray}
We also have the intertwining operators $A_{(+)}^\pm \equiv A_{(1,2)}^\pm$
constructed on the base of the seed states of the positive scheme $(1,2)$.
These four operators satisfy their respective intertwining relations
of the form (\ref{inter0}), and
their alternate products (\ref{polyA}) reduce here to polynomials
$P_{n_-}(L_{(-)})=L_{(-2)}+5\equiv H_1$, $P_{n_-}(L)=L+5\equiv H_0$ and
$P_{n_+}(L_{(+)})=(L_{(+)}-3)(L_{(+)}-5)$, $P_{n_+}(L)=(L+3)(L+5)$, where $L=L_0$
is the Hamiltonian operator of the harmonic oscillator, and
$L_{(+)}$ is the Hamiltonian
produced by positive scheme, which is related with $L_{(-)}$, according to (\ref{L+L-}),
by $L_{(+)}-L_{(-)}=6$.
Here, the eigenstate $A_{(-2)}^{-}\widetilde{\psi_{-2}}=1/\psi_{-2}$ is the isolated
ground state of zero energy of the shifted Hamiltonian operator $H_1$.
The superextended Hamiltonian and its spectrum are
\begin{equation}
\label{superdefHO}
\mathcal{H}=
\left(
\begin{array}{cc}
H_1& 0 \\
0 & H_0
\end{array}\right),\qquad \mathcal{E}_{0}=0\,, \qquad \mathcal{E}_{n+1}=2n+6\,,\qquad n=0,1,\ldots\,.
\end{equation}
The ground state of zero energy is non-degenerate and corresponds to the ground state
$(A_{(-2)}^- \widetilde{\psi_{-2}},0)^t$. Other energy levels
are doubly degenerate
and correspond to eigenstates of the extended Hamiltonian (\ref{superdefHO})
and supercharge $\mathcal{Q}^0_1$, see below\,:
\begin{equation}
\Psi_{n+1}^{+}=
\left(
\begin{array}{c}
(\mathcal{E}_{n+1})^{-1/2} A_{(-2)}^-\psi_{n} \\
\psi_{n}
\end{array}
\right),\qquad
\Psi_{n+1}^{-}=\sigma_3\Psi_{n+1}^{+}\,.
\end{equation}
Witten's index equals one, and the system (\ref{superdefHO})
is characterized by unbroken $\mathcal{N}=2$ Poincar\'e supersymmetry.
Now we use the construction of Section \ref{gen} to produce generators
of the extended nonlinearly deformed super-Schr\"odinger symmetry
of the system.
Following (\ref{gencharge}) and (\ref{superC}),
we construct the odd operators
$\mathcal{Q}_{a}^{z}$ with $z=-2,-1,0,\ldots,5$, and
matrix bosonic ladder operators $\mathcal{G}_{\pm k}^{(1)}$ with $k=1,\ldots,5$.
Also we must consider the operators $\mathcal{G}_{\pm k}^{(0)}$ with $k=1,2$
defined in (\ref{G_k}).
To obtain all the
ingredients, we have to use the version of relation (\ref{reqgen1}) for this system
translated to the supersymmetric extension of $\mathfrak{C}_{N+k}^\pm$ which is
\begin{equation}
\label{req}
\mathcal{G}_{\pm(3l+n)}^{(1)}=(-\mathcal{G}_{\pm 3}^{(1)})^l\mathcal{G}_{\pm n}^{(1)} \,,\qquad n=3,4,5\,,\qquad l=0,1,\ldots\,.
\end{equation}
Then we generate the even part of the superalgebra\,:
\begin{equation}
\label{slr1}
[\mathcal{H},\mathcal{G}_{\pm n}^{(1)}]=\pm 2n \mathcal{G}_{\pm n}^{(1)}\,,
\qquad
[\mathcal{H},\mathcal{G}_{\pm l}^{(0)}]=\pm 2l\mathcal{G}_{\pm l}^{(0)}\,,
\end{equation}
\begin{equation}
\label{slr3}
[\mathcal{G}_\alpha^{(1)},\mathcal{G}_\beta^{(1)}]=P_{\alpha,\beta} \mathcal{G}_{\alpha+\beta}^{(1)}+
M_{\alpha,\beta}\mathcal{G}_{\alpha+\beta}^{(0)}\,,
\quad
\alpha,\beta=\pm1,\ldots,\pm5\,,
\end{equation}
\begin{equation}
\label{slr4}
[\mathcal{G}_\alpha^{(0)},\mathcal{G}_\beta^{(1)}]=\Pi_-(F_{\alpha,\beta}\mathcal{G}_{\alpha+\beta}^{(1)}+
N_{\alpha,\beta}\mathcal{G}_{\alpha+\beta}^{(0)})\,,
\quad
\alpha=1,2\,,\quad\beta=\pm1,\ldots,\pm5\,,
\end{equation}
\begin{equation}
[\mathcal{G}_{-1}^{(0)},\mathcal{G}_{+1}^{(0)}]=2\Pi_-,\qquad [\mathcal{G}_{\pm 1}^{(0)},\mathcal{G}_{\mp 2}^{(0)}]=
\pm6\mathcal{G}_{\pm 1}^{(0)}\,,
\qquad
[\mathcal{G}_{-2}^{(0)},\mathcal{G}_{+2}^{(0)}]=8\Pi_-(\mathcal{H}-5)\,,
\end{equation}
where we put $\mathcal{G}_0^{(1)}=\mathcal{G}_0^{(0)}=1$ and $P_{\alpha,\beta}$, $F_{\alpha,\beta}$,
$M_{\alpha,\beta}$ and $N_{\alpha,\beta}$ are some polynomials in $\mathcal{H}$ and $\Pi_-=\frac{1}{2}(1-\sigma_3)$,
some of which are numerical coefficients, whose explicit form is listed
in Appendix \ref{list}.
We note that
in equations (\ref{slr3}) and (\ref{slr4}), the operators
$\mathcal{G}_{\pm n}^{(1)}$ with $1<n\leq 7$ can appear,
where for $n>5$ we use relation (\ref{req}) (admitting $\mathcal{G}_{\pm3}^{(0)}$
as coefficients in the algebra).
Additionally we note that the operators
$\mathcal{G}_{\pm m}^{(0)}$ with $m>2$ in both equations where they appear
are
absorbed in generators $\mathcal{G}_{\pm m}^{(1)}$.
For eigenstates we have the relations
\begin{eqnarray}
\label{C3psi}
&\Psi_{3j+k}^\pm= (\mathcal{G}_{+3}^{(1)})^j\Psi_k^\pm\,,
\qquad
\Psi_{0}=\mathcal{G}_{-3}^{(1)}\Psi_1^\pm,\qquad j=1,2,\ldots\,,\qquad k=1,2,3\,,&\\
\label{C2psi}
&\Psi_{j}^\pm= (\mathcal{G}_{+ 1}^{(1)})^j\Psi_1^\pm\,,
\qquad
\mathcal{G}^{(1)}_{\pm 1}\Psi_{0}=\mathcal{G}_{-1}^{(1)}\Psi_{1}^\pm=0\,.&
\end{eqnarray}
Eq. (\ref{C3psi}) shows that we can connect the
isolated ground state with the equidistant part of the
spectrum using $\mathcal{G}_{\pm 3}^{(1)}$, which are not spectrum generating operators.
Eq. (\ref{C2psi}) indicates that the states in the equidistant part of
the spectrum can be connected by $\mathcal{G}_{\pm 1}^{(1)}$, but this part
of the spectrum cannot be connected by them with
the ground state. Thus
we have to use a combination of both pairs of these operators.
On the other hand, the odd operators $\mathcal{Q}_a^z$ satisfy relations (\ref{SUSY}),
where $\mathbb{P}_0=\mathcal{H}$, and, therefore,
we have again the linear $\mathcal{N}=2$ Poincar\'e supersymmetry
as a subsuperalgebra generated by $\mathcal{H}$, $\mathcal{Q}^0_a$ and $\Sigma$.
The general anti-commutation structure is given by
\begin{equation}
\label{susy3}
\{\mathcal{Q}_a^{n},\mathcal{Q}_b^{m}\}=\delta_{ab}(\mathbb{C}_{nm}+(\mathbb{C}_{nm})^{\dagger})+
i\epsilon_{ab}(\mathbb{C}_{nm}-(\mathbb{C}_{nm})^{\dagger})\,,
\end{equation}
where $\mathbb{C}_{n,m}=\mathbb{C}_{n,m}(\mathcal{G}_{|n-m|}^{(1)},\mathcal{G}_{|n-m|}^{(0)})$
are some linear combinations of the indicated
ladder operators with coefficients to be polynomials in
$\mathcal{H}$, $\mathcal{G}_{\pm3}^{(0)}$ and $\sigma_3$.
Some of these relations
define ladder operators, see Eq.
(\ref{Cn+kQN}).
For $n=N=3$ and $m=-1,-2$ we can use
(\ref{superC'}) knowing that $\mathcal{Q}_{a}'^{z}=\mathcal{Q}_{a}^{3-z}$,
see subsection \ref{susyextension}.
For structure of anti-commutation relations with
other combinations of indexes, see Appendix \ref{list}.
To complete the description
of the generated nonlinear supersymmetric structure,
we write down the commutators between the independent lowering operators and
supercharges\,:
\begin{eqnarray}
\label{susy4}
&[\mathcal{G}_{-m}^{(1)},\mathcal{Q}_a^{n}]=\mathbb{Q}_{m,n}^{1}(\mathcal{Q}_{a}^{n-m}+i
\epsilon_{ab}\mathcal{Q}_b^{n-m})+\mathbb{Q}_{m,n}^{2}(\mathcal{Q}_{a}^{m+n}-i
\epsilon_{ab}\mathcal{Q}_b^{m+n})\,,&\\
\label{susy5}
&[\mathcal{G}_{-m}^{(0)},\mathcal{Q}_a^{n}]=\mathbb{G}_{m,n}^{1}(\mathcal{Q}_{a}^{n-m}+i
\epsilon_{ab}\mathcal{Q}_b^{n-m})+\mathbb{G}_{m,n}^{2}(\mathcal{Q}_{a}^{m+n}-
i\epsilon_{ab}\mathcal{Q}_b^{m+n})\,.&
\end{eqnarray}
Here $\mathbb{Q}_{m,n}^{j}$ and $\mathbb{G}_{m,n}^{j}$ with $j=1,2$ are
polynomials in $\mathcal{H}$ or numerical coefficients, some of which are listed
in the sets of general commutation relations in Appendix \ref{apen-comm},
while other are given explicitly in Appendix \ref{list}.
As the odd fermionic operators are Hermitian, then
$[\mathcal{G}_{+m}^{(1)},\mathcal{Q}_a^{z} ]=-([\mathcal{G}_{- m}^{(1)},
\mathcal{Q}_a^{z} ])^{\dagger}$,
and we do not write them explicitly.
In matrix language, Eq. (\ref{susy4}) can be written as
\begin{eqnarray}
&[\mathcal{G}_{-m}^{(1)},\mathcal{Q}_a^{n}]=
\left(
\begin{array}{cc}
0& \mathfrak{S}_{n+m}^- \\
\mathfrak{S}_{n-m}^+ & 0
\end{array}
\right)\,,&
\end{eqnarray}
and an important point here is that the number $n-m$ could take values less than -2 and $n+m$ could
be greater than 5, but fermionic operators are defined with the index $z$ taking integer values
in the interval $I=[-2,+5]$.
It is necessary to remember that we cut the series of $\mathfrak{S}_z^\pm$
because operators outside the defined interval are reduced to combinations (products) of other basic operators.
In this way, we formally apply the definition of $\mathfrak{S}_z^\pm$ outside of the indicated
interval and
use the relation in Appendix \ref{apen-red} to show that these ``new" generated operators reduce to combinations
of operators with index values in the interval $I$
and of the generators $\mathfrak{C}_{\pm 3}$.
Finally, the subsets which produce closed subsuperalgebras here are those defined by $\mathcal{U}_{0,z}^{(1)}$
in (\ref{U1}), with $z=1,\ldots,5$ in addition to $\mathcal{I}_{N,-k}^{(1)}$ given in (\ref{I1}) with $k=1,2$.
With respect to the positive scheme, the super-Hamiltonian is given by $\mathcal{H}'=\text{diag}\,
(L_{(+)}-3,L_0-3)$.
It has two positive energy singlet states of the form $(0,\psi_n)$ with $n=1,2$; besides,
there are two ground states $\Psi_0^+=(\phi_0,\psi_0)$ and $\Psi_0^-=\sigma_3\Psi_0^+$ of energy $-2$.
According
to the construction from the previous section,
the fermionic operators here are $\mathcal{Q}^{'z}_a=\mathcal{Q}^{3-z}_a$, and
the basic subsets which generate
closed subsuperalgebras are $\mathcal{U}_{0,k}'^{(1)}$ and $\mathcal{I}_{N,l}'^{(1-2\theta(l))}$
with $k=3,4,5$ and $l=-1,-2,4,5$.
One can note that considering $\mathcal{G}_{\pm 3}^{(1)}$ as coefficients,
the subset $\{\mathcal{H}, \mathcal{G}_{\pm 3}^{(1)},\sigma_3,
\mathcal{Q}_a^{-2},\mathcal{Q}_a^{1},\mathcal{Q}_a^{4}, \mathbb{I} \} $
also generates a closed nonlinear superalgebraic structure.
\subsection{Further possibilities for super-extensions}\label{SecFigs}
As we have seen in Section \ref{gen}, there is a big variety of Darboux transformations,
performed by the operators (\ref{genA}), which match the (half-) harmonic oscillator
with a specific rationally deformed system of class $c$, but with different shift constants.
In fact, these operators satisfy the intertwining relations
$A_{(\pm)}^-(a^\pm)^n L=(L_{(\pm)}\mp 2n)A_{(\pm)}^-(a^\pm)^n$ and
$A_{(\pm)}^-(a^\mp)^n L=(L_{(\pm)}\pm 2n)A_{(\pm)}^-(a^\mp)^n$.
Such kind of transformations also allow us to construct an
$\mathcal{N}=2$ superextended Hamiltonian of the form (\ref{Hlambda*}) by choosing
$(L_{(\pm)}\pm 2n)$ as $\breve{L}$, and its spectrum will be
different from the spectrum of $\mathcal{H}$ and $\mathcal{H}'$.
We shall not discuss here
the corresponding superconformal structures of such superextended systems, but
restrict ourselves just by some general comments related to their spectra.
As any DCKA transformation can be decomposed step by step, it is convenient to start
by considering the mapping performed by intertwining operators
$(a^\pm)^n$.
Due to the shape
invariance of the QHO $L_0$,
the two Darboux schemes $(0,1,\ldots,n-1)$ and
$(-0,-1,\ldots,-(n-1))$ produce the same but shifted
Hamiltonian operators $L_0+2n$ and $L_0-2n$, respectively.
Then, following (\ref{Hlambda*}), we find that
the superextended systems
$(L_0+2n,L_0)$ and $(L_0-2n,L_0)$
have similar spectra.
This is summarized by the diagram
on Figure \ref{figure1},
\begin{figure}[H]
\begin{minipage}[r][2in][t]{0.5\textwidth}
\begin{center}
\includegraphics[scale=0.22]{figure1.eps}
\end{center}
\end{minipage}
\begin{minipage}[r][2in][t]{3in}
\caption{\small{The diagram shows the shifted harmonic oscillators $L_0\pm 2n$
as super-partners
of $L_0$.
Symbol X indicates non-physical seed states $\psi_{-k}$,
blue discs with boldface border indicate non-physical states
$\widetilde{\psi_{-k}}$, see Eq. (\ref{secondlinearindependent}),
and circles with inscribed yellow triangles indicate physical
seed states $\psi_n$.
Action of the corresponding intertwining operators is shown by arrows. } }
\label{figure1}
\end{minipage}
\end{figure}
\vskip0.7cm
\noindent
{}From the diagram
it is easy to reed the spectrum of the corresponding superextended system. If
we chose $\breve{L}=L_0+2n$, the intertwining operator $(a^-)^n$ annihilates the first $n$ physical states
of $L_0$,
and the superextended system $(L_0,L_0+2n)$
has $n$ singlet eigenstates given by
$(\psi_j,0)^t$ with $j=0,\ldots,n-1$.
On the other hand, taking
$(a^+)^n$ as intertwining operator, we generate the superextended system
$(L_0,L_0-2n)$ characterized by $n$ singlet states
$(0,\psi_j)^t$.
In correspondence with these arguments,
the intertwining operators of the form $A_{(\pm)}^-(a^\pm)^n$ and $A_{(\pm)}^-(a^\mp)^n$ have the following interpretation:
we first shift the initial system
as much as we want with the corresponding power of $a^\pm$,
and then we produce the nontrivial deformation.
As a result, the final superextended system will present
singlet states in correspondence with this shift. To illustrate the process, consider the two
examples presented diagrammatically in Figure \ref{figure2}.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=0.6]{figure2.eps}
\caption{\small{The diagrams show some modifications of the systems discussed in this section.
On the left panel $a)$, the operator $L_{(+)}=L_{(-)}-6$, and $L_{(-)}$ is given by (\ref{H_-2}).
On the right panel $b)$, the operator $L_{(-)}$ is defined in (\ref{L(-3)}), and $L_{(+)}=L_{(-)}-8$.
Circles with smaller yellow circles inside indicate
even states which are non-physical
eigenstates;
small yellow crossed circles correspond to
non-physical Wick-rotated even
eigenstates.}}
\label{figure2}
\end{center}
\end{figure}
\vskip-0.7cm
\noindent
This diagram follows the same logic of Fig. \ref{figure1},
but without showing the action of
intertwining operators. According to the left panel a) of Fig. \ref{figure2},
one can construct the following
super-extensions. By matching $L_0$ with $L_{(+)}$,
we have the positive scheme discussed at the end of
the previous subsection.
Matching $L_0$ with $L_{(+)}-2$ results in the system with two singlet states of the form
$(0,\psi_{n})^t$ with $n=0,1$, and one more singlet of the form
$(\phi_0,0)^t$, where $\phi_0$ is the ground state of $L_{(+)}$.
By matching $L_0$ with $L_{(+)}-4$,
we obtain the system with singlet states $(0,\psi_0)$ and
$(\phi_0,0)^t$.
Finally, by matching $L_{0}$ with $L_{(+)}-6=L_{(-)}$,
we reproduce the negative scheme.
In all these cases we use the chain of relations
$$
\phi_0=A_{(+)}^-(a^+)^3\widetilde{\psi_{-2}}=
A_{(+)}^-(a^+)^2\widetilde{\psi_{-1}}=
A_{(+)}^-(a^+)\widetilde{\psi_{-0}}=A_{(+)}^-\psi_{0}=A_{(-)}^-\widetilde{\psi_{-2}}
$$
to be equalities modulo numerical multiplicative factors.
A similar analysis can be done for deformations
of the $L_1^{iso}$ with the help of the right panel b).
Any of these shifted deformed systems can be paired
each other as super-partners. For example, we can pair
$L_{(+)}$ and $L_{(-)}$ from panel a),
and the intertwining operator for them
will be the operator $\mathfrak{C}_{3}^\pm$.
In this sense, each deformed system of a class $c$
can be paired with its proper copy
but with the Hamiltonian displaced for additive constant $2cj$,
$j=1,2,\ldots$,
by using $\mathfrak{A}_{cj}^\pm$,
$\mathfrak{B}_{cj}^\pm$, or $\mathfrak{C}_{cj}^\pm$
as intertwining operators.
\section{Discussion and outlook}
\label{conclusion}
In conclusion, we indicate some open problems to be interesting
for further investigation.
\vskip0.2cm
We considered super-extensions
of rationally deformed harmonic oscillator and AFF models
paired, respectively, with the undeformed harmonic
or half-harmonic oscillator systems.
But one could construct superextended systems composed
from the pairs of rationally deformed systems having in mind that the corresponding
dual schemes can be interpreted as periodic Darboux chains \cite{VesSha}, in which
the harmonic oscillator will just be an intermediate
system between $L_{(-)}$ and $L_{(+)}$.
In the corresponding superextended Hamiltonian
(\ref{Hlambda*}) we then will have $\breve{L}=L_{(+)}$ and $L=L_{(-)}$,
for which the intertwining operators are given by the sets of the ladder operators
of the three families considered here.
Such a generalization was discussed very schematically
in Section \ref{SecFigs}.
The use of some singular systems in the middle of the
Darboux chain could allow to construct intertwining operators of less order.
Situation like that appears in the case of exotic supersymmetry
in finite-gap systems obtained from the free particle \cite{AraMatPly}.
\vskip0.2cm
Recently, the $\mathcal{PT}$-regularized
systems (\ref{isoham}) with special
values of the coupling constant $g=n(n+1)$ but without the confining potential term
were studied in \cite{JM1,JM2} together with their rational deformations, which
are intimately related to the Korteweg-de Vries hierarchy of completely
integrable systems. It was found there that such systems
reveal unusual properties, in particular,
in the context of superconformal symmetries.
It would be interesting to investigate the
$\mathcal{PT}$-regularized versions
of the superextended rationally deformed AFF systems studied by us here.
We note that the
$\mathcal{PT}$-regularized version of the undeformed model (\ref{isoham})
was considered earlier in \cite{PT1}.
The indicated problem also is interesting in another aspect.
If in the $\mathcal{PT}$-regularized AFF model we reconstruct the frequency
parameter $\omega$ and consider the limit $\omega\rightarrow0$,
we reproduce the completely invisible zero-gap systems with $g=n(n+1)$ investigated in Refs. \cite{JM1,JM2}.
Based on certain ladder operators we studied here,
one can also reproduce the higher derivative Lax-Novikov
integrals which underly peculiar properties of the systems
from \cite{JM1,JM2}. However, in the same na\"ively
applied limit $\omega\rightarrow 0$ additional terms
appearing in the rationally deformed versions of the
AFF model turn into zero. Therefore, the question is
whether the rational terms responsible, in particular,
for the origin of the extreme wave solutions in \cite{JM1}
can be reproduced from the rational terms
we have here by considering limit
$\omega\rightarrow 0$ in some indirect way.
\vskip0.2cm
We were dealing here only with rationally deformed AFF models with special coupling constant
values $g=n(n+1)$. It would be interesting to extend
our analysis for the case of rational deformations
of the AFF model of general form (\ref{isoham}).
For this, it will be necessary to generalize
the three families of the ladder operators for the case of arbitrary values
of the parameter $\nu$.
Some rational deformations of the system (\ref{isoham}) with arbitrary coupling constant values
were considered in \cite{Grandati2}.
If the technique of dual schemes can be generalized for the case of $\nu\in\mathbb R$,
our approach can also be extended for rational deformations
of the AFF model (\ref{isoham}).
Some preliminary considerations indicate that this indeed can be done.
In this context the case of gapless deformations
is of a special interest due to their complete isospectrality
to the original system (\ref{isoham}) with corresponding value of the parameter $\nu$,
which attracted considerable
interest in the physics of anyons some time ago \cite{LeiMyr,Poly}.
This also could shed a light on the properties of the
$\mathcal{PT}$-regularized two-particle Calogero systems
with arbitrary values of $g=\nu(\nu+1)$ but without the confining potential term.
One can expect that they have to be essentially different from the peculiar
properties of the class of such systems
with $g=n(n+1)$ studied in \cite{JM1,JM2}.
\vskip0.2cm
Deformed superconformal symmetry we investigated is
based essentially on the higher-derivative integrals of motion.
Such integrals are associated usually with the hidden
symmetries \cite{Cariglia}. In their study,
the Eisenhart-Duval lift \cite{Eisenhart-lift,Duval,DGH,CGGH}
plays a very important role by allowing to
look at the system from the perspective of Riemannian geometry
and to establish relation between different systems
possessing hidden symmetries.
It would be very interesting
to consider the systems we studied by applying to
them the method of the Eisenhart-Duval lift.
\vskip0.3cm
\noindent {\large{\bf Acknowledgements} }
\vskip0.2cm
L.I. acknowledges the CONICYT scholarship 21170053
and the project FCI-PM-02 (USACH).
|
\section{Introduction}
For understanding the strong interaction sector of the standard model
(SM) it is not sufficient to compute masses of stable
particles. Gaining insight into interactions of two or more
hadrons and resonances is a must. Due to the non-perturbative nature
of low energy quantum chromodynamics (QCD), computations of interaction
properties from lattice
QCD are highly desirable. While ultimately the phase shift in a given
partial wave is to be computed, also the scattering length is in many
cases a useful quantity, in particular when the two-particle interaction
is weak.
Due to the importance of chiral symmetry in QCD the investigation of
systems with two pseudoscalar mesons is of particular interest. Here,
chiral perturbation theory (ChPT) is able to provide a description of
the pion mass dependence. And any non-perturbative computation in turn
allows to check this dependence. Naturally, ChPT works best for two
pion systems, while convergence is unclear for pion-kaon or two kaon
systems.
The two pion system is studied well experimentally, also in the
different isospin channels. However, as soon as one or both pions are
replaced by kaons, experimental results become sparse. On the other
hand, this gap starts to be filled by lattice QCD calculations.
For the pion-kaon system with isospin $I=3/2$ there are
by now a few lattice results available focusing on the scattering
length~\cite{Beane:2006gj,Sasaki:2013vxa,Fu:2011wc,Janowski:2014uda,Lang:2012sv}.
The most recent computation in Ref.~\cite{Janowski:2014uda} uses one lattice at
physical pion and kaon masses and lattice spacing
$a\approx\SI{0.114}{\femto\metre}$. For the sea and valence sector they use
$N_f=2+1$ M{\"o}bius domain wall fermions and an Iwasaki gauge action.
In Ref.~\cite{Sasaki:2013vxa} a systematic study of the elastic scattering lengths for
the light pseudoscalar mesons was carried out with $N_f=2+1$ order $\order{a}$-improved
Wilson quarks at pion masses ranging from
\SIrange{170}{710}{\mega\electronvolt} and a lattice spacing
$a\approx\SI{0.09}{\femto\metre}$. Furthermore
Refs.~\cite{Fu:2011wc,Beane:2006gj} use $N_f=2+1$ flavors on the MILC
configurations with a rooted staggered sea quark action. Whereas
Ref.~\cite{Fu:2011wc} calculates the scattering length at a lattice spacing
$a\approx\SI{0.15}{\femto\metre}$, a slightly smaller lattice spacing
$a\approx\SI{0.125}{\femto\metre}$ has been used in Ref.~\cite{Beane:2006gj}.
The pion masses in Ref.~\cite{Beane:2006gj} range from
\SIrange{290}{600}{\mega\electronvolt} using domain wall valence quarks with a
chiral extrapolation done in mixed-action chiral perturbation theory (MAChPT)~\cite{PhysRevD.73.074510,PhysRevD.75.054501}.
The range of pion masses, \SIrange{330}{466}{\mega\electronvolt}, for the Asqtad
improved staggered fermions of
Ref.~\cite{Fu:2011wc} is a bit smaller compared to Ref.~\cite{Beane:2006gj}.
In Ref.~\cite{Lang:2012sv} the phaseshifts and scattering lengths for
$\pi$-$K$-scattering in $I=3/2$ and $I=1/2$ in the $s$-wave and the $p$-wave has
been determined. The gauge action is a $N_f=2$ tree level improved Wilson-Clover
action. The authors include the strange quark as a valence quark only which then
corresponds to pion and kaon masses of
$M_{\pi}=\SI{266}{\mega\electronvolt}$ and $M_K=\SI{522}{\mega\electronvolt}$,
respectively.
In this paper we are going to present results for the s-wave
scattering length of the pion-kaon system in the elastic region with
isospin $I=3/2$. The investigation is based on gauge configurations
produced by the European Twisted Mass Collaboration (ETMC) with
$N_f=2+1+1$ dynamical quark flavors~\cite{Baron:2010th}. In
contrast to previous computations, we are able to perform a continuum
extrapolation owing to 11 ensembles with $M_\pi$ ranging from
\SIrange{230}{450}{\mega\electronvolt} distributed over 3 different
lattice spacing values. We employ in total 4 different extrapolation
methods to also estimate systematic uncertainties associated with our
computation.
Finally, since this paper is the fourth in a series of
publications~\cite{Helmes:2015gla,Liu:2016cba,Helmes:2017smr}
concerning elastic scattering of two pions in different channels and
kaon-kaon with $I=1$, we are able to compare results of two
pseudoscalar mesons at maximal isospin involving different amounts of
strangeness. The leading order ChPT predictions for the dependence on
the reduced mass divided by the relevant decay constant are
identical for the three systems and differences appear only at NLO.
This paper is organized as follows: We first introduce the lattice
details of our calculation. After the discussion of the analysis
methods we present the main result, followed by a detailed discussion
of the analysis details. We close with a discussion and
summary. Technical details can be found in the appendix.
\section{Lattice Action and Operators}
\subsection{Action}
The lattice details for the investigation presented here are very
similar to the ones we used to study the kaon-kaon scattering
length~\cite{Helmes:2017smr}. We use $N_f=2+1+1$ flavor lattice QCD
ensembles generated by the ETM Collaboration, for which details can be
found in Refs.~\cite{Chiarappa:2006ae,Baron:2010th,Baron:2010bv}.
The parameters relevant for this paper are compiled in
\Cref{tab:setup}: we give for each ensemble the inverse gauge coupling
$\beta=6/g_0^2$, the bare quark mass parameters $\mu_\ell, \mu_\sigma$
and $\mu_\delta$, the lattice volume and the number of configurations on
which we estimated the relevant quantities.
\begin{table}
\centering
\begin{tabular*}{.9\textwidth}{@{\extracolsep{\fill}}lcccccc}
\hline\hline
ensemble & $\beta$ & $a\mu_\ell$ & $a\mu_\sigma$ & $a\mu_\delta$ &
$(L/a)^3\times T/a$ & $N_\mathrm{conf}$ \\
\hline\hline
A30.32 & $1.90$ & $0.0030$ & $0.150$ & $0.190$ & $32^3\times64$ & $259$ \\
A40.24 & $1.90$ & $0.0040$ & $0.150$ & $0.190$ & $24^3\times48$ & $376$ \\
A40.32 & $1.90$ & $0.0040$ & $0.150$ & $0.190$ & $32^3\times64$ & $246$ \\
A60.24 & $1.90$ & $0.0060$ & $0.150$ & $0.190$ & $24^3\times48$ & $303$ \\
A80.24 & $1.90$ & $0.0080$ & $0.150$ & $0.190$ & $24^3\times48$ & $300$ \\
A100.24 & $1.90$ & $0.0100$ & $0.150$ & $0.190$ & $24^3\times48$ & $304$ \\
\hline
B35.32 & $1.95$ & $0.0035$ & $0.135$ & $0.170$ & $32^3\times64$ & $241$ \\
B55.32 & $1.95$ & $0.0055$ & $0.135$ & $0.170$ & $32^3\times64$ & $251$ \\
B85.24 & $1.95$ & $0.0085$ & $0.135$ & $0.170$ & $32^3\times64$ & $288$ \\
\hline
D30.48 & $2.10$ & $0.0030$ & $0.120$ & $0.1385$ & $48^3\times96$ & $364$ \\
D45.32sc & $2.10$ & $0.0045$ & $0.0937$ & $0.1077$ & $32^3\times64$ & $289$ \\
\hline\hline
\end{tabular*}
\caption{The gauge ensembles used in this study. For the labeling of
the ensembles we adopted the notation in
Ref.~\cite{Baron:2010bv}. In addition to the relevant input
parameters we give the lattice volume and the number of evaluated
configurations, $N_\mathrm{conf}$.}
\label{tab:setup}
\end{table}
The ensembles were generated using the Iwasaki gauge action and employ the
$N_f=2+1+1$ twisted mass fermion
action~\cite{Frezzotti:2003ni,Frezzotti:2003xj,Frezzotti:2004wz}.
For orientation, the $\beta$-values \numlist{1.90;1.95;2.10} correspond to
lattice spacing values of $a\sim\,$\SIlist{0.089;0.082;0.062}{\femto\metre},
respectively, see also \Cref{tab:r0values}. The
ensembles were generated at so-called maximal twist, which guarantees
automatic order $\order{a}$ improvement for almost all physical
quantities~\cite{Frezzotti:2003ni}. The
renormalized light quark mass $m_\ell$ is directly proportional to the
light twisted quark mass via
\begin{equation}
m_\ell\ =\ \frac{1}{Z_P} \mu_\ell\,,
\end{equation}
with $Z_P$ the pseudoscalar renormalization constant. The relation of
the bare parameters $\mu_\sigma$ and $\mu_\delta$ to the renormalized
charm and strange quark masses reads
\begin{equation}
m_{c,s}\ =\ \frac{1}{Z_P}\mu_\sigma\ \pm\ \frac{1}{Z_S} \mu_\delta\,,
\end{equation}
with $Z_S$ the non-singlet scalar renormalization constant.
As noted in Refs.~\cite{Baron:2010bv, Carrasco:2014cwa}, the renormalized sea strange quark masses across the ``A'', ``B'' and ``D'' ensembles vary by up to about 20\% and in a few cases differ from the physical strange quark mass to the same extent.
For D30.48 and D45.32sc at the finest lattice spacing, the sea strange quark
mass on the former ensemble overshoots the physical strange quark mass while it
is consistent on the latter ensemble.
In order to correct for these mis-tunings and to avoid the complicated
flavor-parity mixing in the unitary non-degenerate strange-charm
sector~\cite{Baron:2010th}, we adopt a mixed action ansatz with so-called
Osterwalder-Seiler (OS)~\cite{Frezzotti:2004wz} valence quarks, while keeping
order $\order{a}$ improvement intact.
We denote the OS bare strange quark parameter with $\mu_s$.
It is related to the renormalized strange quark mass by
\begin{equation}
m_s\ =\ \frac{1}{Z_P} \mu_s\,.
\end{equation}
For each ensemble we investigate three values of $\mu_s$ which are compiled in \Cref{tab:mus}.
More details on the mixed action approach can be found in Ref.~\cite{Helmes:2017smr}.
As a smearing and contraction scheme we employ the stochastic Laplacian-Heaviside approach, described in Ref.~\cite{Morningstar:2011ka}.
Details of our parameter choices can be found in Refs.~\cite{Helmes:2015gla,Helmes:2017smr}.
\begin{table}[t!]
\centering
\begin{tabular*}{.7\textwidth}{@{\extracolsep{\fill}}lrrr}
\hline\hline
$\beta$ & $1.90$ & $1.95$ & $2.10$ \\
\hline\hline
$a\mu_s$ & 0.0185 & 0.0160 & 0.013/0.0115\\
& 0.0225 & 0.0186 & 0.015\\
& 0.0246 & 0.0210 & 0.018\\
\hline\hline
\end{tabular*}
\caption{Values of the bare strange quark mass $a\mu_s$ used for the
three $\beta$-values. The lightest strange quark mass on the ensemble D30.48 is $a\mu_s = 0.0115$
instead of $a\mu_s = 0.013$.}
\label{tab:mus}
\end{table}
\subsection{Lattice Operators and Correlation Functions}
For reasons which will become clear later we need to estimate the
masses of the pion, the kaon and the $\eta$ meson on our
ensembles. The masses for the pion and kaon are obtained from the large
Euclidean time dependence of two point functions of the form
\begin{equation}
C_X(t-t') = \langle\interp{X}(t) \, \interp{X}^\dagger(t')\rangle\,,
\label{eq:two_corr}
\end{equation}
where $X\in\{\pi, K\}$. The operators for the charged pion and kaon
projected to zero momentum read
\begin{equation}
\mathcal{O}(X)(t)\ =\ \sum_{\mathbf{x}} O_X(\mathbf{x}, t)
\end{equation}
with
\begin{align}
O_\pi(\mathbf{x}, t)\ &=\
i\bar{d}(\mathbf{x},t)\,\gamma_5\,u(\mathbf{x},t)\,,\\
O_K(\mathbf{x}, t)\ &=\
i\bar{s}(\mathbf{x},t)\,\gamma_5\,u(\mathbf{x},t)\,.
\end{align}
For the $\eta$ (and $\eta^\prime$) meson we use the two operators
\begin{align}
O_\ell(\mathbf{x}, t)\ &=\ \frac{i}{\sqrt{2}}(\bar{u}(\mathbf{x},t)\,\gamma_5\, u(\mathbf{x},t) +
\bar{d}(\mathbf{x},t)\,\gamma_5\, d(\mathbf{x},t))\,,\\
O_s(\mathbf{x}, t)\ &=\ i\bar{s}(\mathbf{x},t)\,\gamma_5\, s(\mathbf{x},t)\,.
\end{align}
From these we build a two-by-two correlator matrix by taking the
disconnected diagrams into account. The $\eta$
(principal) correlator is determined by solving a generalized
eigenvalue problem as described in detail in
Ref.~\cite{Ottnad:2015hva}.
A complete discussion of the analysis of the $\eta$ (and $\eta'$) meson is beyond the scope of this paper and the full analysis will be presented in a future publication~\cite{Jost:2018nn}.
In addition to the aforementioned meson masses, we also need to
estimate the energy $E_{\pi K}$ of the interacting pion-kaon two particle system.
For the case of maximal isospin, i.e. $I=3/2$, the corresponding
two particle operator reads
\begin{align}
\interp{\pi K}(t) =
-\sum_{\mathbf{x},\mathbf{x'}}\bar{d}(\mathbf{x},t)\,\gamma_5\,u(\mathbf{x},t)
\bar{s}(\mathbf{x'},t)\,\gamma_5\,u(\mathbf{x'},t)\,.
\label{eq:inter_pik}
\end{align}
It is used to construct the two-particle correlation function
\begin{align}
C_{\pi K}(t-t') = \langle\interp{\pi K}(t)\interp{\pi K}^{\dagger}(t')\rangle\,.
\label{eq:four_corr}
\end{align}
$E_{\pi K}$ can then be determined from the large Euclidean time
dependence of $C_{\pi K}$.
\section{Analysis Methods}
\label{sec:analysis_methods}
We focus in this work on pion-kaon scattering in the elastic region.
For small enough squared scattering momentum $p^2$ one can perform
the effective range expansion for partial wave $\ell$
\begin{equation}
p^{2\ell+1} \cot(\delta_\ell)\ =\ -\frac{1}{a_\ell} +\mathcal{O}(p^2)
\end{equation}
with phase shift $\delta_\ell$ and scattering length $a_\ell$.
For the pion-kaon system it is, to a very good approximation,
sufficient to study the s-wave, i.e. $\ell=0$.
In lattice QCD the phase shift or the scattering length can only be
computed from finite volume induced energy shifts. The relevant energy
shift here is given by
\begin{equation}
\delta E = E_{\pi K} - M_\pi - M_K\,.
\end{equation}
Using again the effective range expansion, one arrives at the
L{\"u}scher formula~\cite{Luscher:1986pf}
\begin{align}
&\delta E=- \frac{2 \pi a_{0}} {\mu_{\pi K} L^3} \left(1 +c_1 \frac{a_{0}}{L} + c_2
\frac{a_{0}^2}{L^2}\right) +
\mathcal{O}(L^{-6})
\label{eq:scat_lusch}
\end{align}
relating $\delta E$ directly to the scattering length $a_0$,
the reduced mass of the pion-kaon system
\begin{align}
\mu_{\pi K} = \frac{M_\pi M_K}{M_\pi+M_K}\,,
\label{eq:red_mass}
\end{align}
and the spatial extent of the finite volume $L$. The
coefficients read~\cite{Luscher:1986pf}
\[
c_1 = -2.837297\,,\quad c_2=6.375183\,.
\]
Given $\delta E$, $\mu_{\pi K}$ and $L$, L{\"u}scher's formula allows
one to determine the scattering length $a_0$ by solving \Cref{eq:scat_lusch} for $a_0$.
In what follows, we will describe how we extract $\delta E$ and the
other relevant bare quantities from correlation functions. Then we
will give details on our
approach to inter- or extrapolate the results to physical conditions
and to the continuum limit.
In order to gain some understanding of systematic uncertainties, we
perform the analysis in two different ways once the bare data has
been extracted.
Combined chiral and continuum extrapolations are performed at fixed strange quark mass using next to leading order ChPT (NLO ChPT) and a variant thereof referred to as the $\Gamma$-method, as described in Ref.~\cite{Beane:2006gj}.
\subsection{Physical Inputs}
\begin{table}
\begin{tabular*}{.5\linewidth}{@{\extracolsep{\fill}}lrr}
\hline\hline
$\beta$ &
$a\ [\mathrm{fm}]$ & $r_0/a$ \\
\hline\hline
$1.90$ & $0.0885(36)$ & $5.31(8)$ \\
$1.95$ & $0.0815(30)$ & $5.77(6)$ \\
$2.10$ & $0.0619(18)$ & $7.60(8)$ \\
\hline\hline
\end{tabular*}
\caption{Values of the Sommer parameter $r_0/a$ and the
lattice spacing $a$
at the three values of $\beta$. See
Ref.~\cite{Carrasco:2014cwa} for details.}
\label{tab:r0values}
\end{table}
For the analysis presented below, we require physical inputs for the
pion, the kaon and $\eta$-meson masses as well as the pion decay constant.
To this end, we employ the values in the isospin symmetric limit, $\overline{M}_{\pi}$ and $\overline{M}_{K}$, as determined in chiral perturbation theory~\cite{Gasser:1983yg} and given in Ref.~\cite{Aoki:2016frl} as
\begin{equation}
\begin{split}
\overline{M}_{\pi} & = 134.8(3)\,\text{MeV}\,, \\
\overline{M}_{K} & = 494.2(3)\,\text{MeV} \,.
\end{split}
\label{eq:meson_masses_phys}
\end{equation}
For the $\eta$ meson mass we use the average obtained by the Particle Data
Group~\cite{PhysRevD.98.030001}:
\begin{equation}
\label{eq:etamass}
\overline{M}_\eta = 547.86(2) \,\text{MeV}\,.
\end{equation}
For the decay constant, we use the phenomenological average determined by the Particle Data Group given in Ref.~\cite{Olive:2016xmw} as
\begin{equation}
f_{\pi^{-}}^{\text{(PDG)}} = 130.50(13)\,\text{MeV}\,.
\label{eq:decay_const_phys}
\end{equation}
As an intermediate lattice scale, we employ the Sommer parameter
$r_0$~\cite{Sommer:1993ce}. It was determined in
Ref.~\cite{Carrasco:2014cwa} from the ensembles we use here to be
\begin{align}
r_0 = \SI{0.474+-0.011}{\femto\metre}\,.
\label{eq:r0_cont}
\end{align}
In the parts of the analysis which require $r_0$, we use
parametric bootstrap samples with central value
and width given in \Cref{eq:r0_cont}.
Where $r_0/a$ values enter as fit parameters, we constrain the corresponding
fit parameters using Gaussian priors in the augmented $\chi^2$ function given as
\begin{equation}
\chi^2_\mathrm{aug}\ =\ \chi^2 + \sum_\beta
\left(\frac{(r_0/a)(\beta) - P_{r}(\beta)}{\Delta
r_0/a(\beta)}\right)^2\,.
\label{eq:chi_aug}
\end{equation}
\subsection{Energy Values from Correlation Functions}
\label{sec:pollution}
The energies of the two point correlation functions as given in \Cref{eq:two_corr} are
extracted from fits of the form
\begin{equation}
C_X(t) = A_0^2(e^{-E_{X}t}+e^{-E_X(T-t)})\,,
\label{eq:c2_fit}
\end{equation}
to the data.
While for $M_K$ and $M_\pi$ the signal extends up to $T/2$, for the $\eta$ we have to face more noise .
We deal with this by applying the excited state subtraction method used and described in Refs.~\cite{Michael:2013gka,Ottnad:2015hva}.
In the determination of the energy shift $\delta E$, the total energy $E_{\pi K}$
of the interacting $\pi$-$K$ system must be computed.
However, in the spectral decomposition of the two-particle correlation function,
unwanted time dependent contributions, so-called thermal pollutions, appear.
Taking into account that our $\pi$-$K$ correlation function is symmetric around the $T/2$ point, the leading contributions in the spectral decomposition can be cast into the form
\begin{equation} \label{eq:pollution}
\begin{split}
C_{\pi K}(t) &\approx \braket{\Omega|\pi^+K^+|\pi K}\braket{\pi K|(\pi^+ K^+)^{\dagger}|\Omega}\left(e^{-E_{\pi K}t}+e^{-E_{\pi K}(T-t)}\right) \\
&+ A_1 \left(e^{-E_{\pi} T}e^{(E_{\pi}-E_K)t} + e^{-E_K T}e^{(E_{K}-E_{\pi})t} \right) \,,
\end{split}
\end{equation}
where only the first line corresponds to the energy level we are interested in.
However, at finite $T$-values, the second contribution might be
sizable, in particular at times close to $T/2$.
Moreover, the thermal pollution cannot be separated easily from the signal we are interested in.
We have studied two different methods, labeled \E{1} and \E{2}, to extract $E_{\pi K}$ from $C_{\pi K}(t)$, where \E{1} has already been discussed in Ref.~\cite{Dudek:2012gj}.
\begin{itemize}
\item {\E{1}: weighting and shifting:}\\
To render one of the polluting terms
in \Cref{eq:pollution} time independent, the correlation function first gets
weighted by a factor $\exp((E_K - E_\pi)t)$. We chose this factor,
because $\exp(-E_\pi T)$ is significantly larger than $\exp(-E_K T)$.
The resulting constant term can then be removed by the shifting procedure,
which thus replaces $C_{\pi K}(t)$ by
\begin{equation}
\begin{split}
C^w_{\pi K}(t) &= e^{(E_K-E_\pi)t} C_{\pi K}(t)\,, \\
\widetilde{C}^w_{\pi K}(t) &= C^w_{\pi K}(t) - C^w_{\pi K}(t+\delta t) \,,
\end{split}
\end{equation}
where $\delta t$ is a fixed number of time slices.
Subsequently, we multiply $\widetilde{C}^w_{\pi K}(t)$ by
$\exp(-(E_K-E_\pi)t)$, which (mostly) recovers the original time
dependence in the contribution of interest
\begin{align}
C^{\text{E1}}_{\pi K}(t) =e^{-(E_K - E_\pi)t}\widetilde{C}^w_{\pi K}(t)\,.
\label{eq:reweight}
\end{align}
We then extract the total energy of the $\pi$-$K$ system, $E_{\pi K}$, from a fit of the form,
\begin{equation}
\begin{split}
C^{\text{E1}}_{\pi K}(t) &= A_0^2\left(e^{-E_{\pi K}t}+e^{-E_{\pi
K}(T-t)}-e^{(E_K - E_\pi)
\delta t}\left(e^{-E_{\pi K}(t+\delta t)}+e^{-E_{\pi K}(T-(t+\delta
t))}\right)\right) \\
&+\widetilde{A}_1 e^{(E_{K}-E_{\pi})t}\,.
\label{eq:e_fit}
\end{split}
\end{equation}
Note that in contrast to Ref.~\cite{Dudek:2012gj}, where correlator matrices with various thermal pollutions are considered, we are able to take $\tilde{A}_1$ as an additional fit parameter in order to account for this sub-leading term.
\item {\E{2}: dividing out the pollution:}\\
To improve on method \E{1}, we assume that the decomposition given in \Cref{eq:pollution} allows one to neglect any further thermal pollutions.
This leads to dividing out the time dependent part
\begin{align}
p(t) = e^{(E_K-E_\pi)t}e^{-E_KT}+e^{-(E_K-E_\pi)t}e^{-E_\pi T}\,,
\label{eq:c4_poll}
\end{align}
explicitly. With
\begin{equation}
C'_{\pi K}(t)\ =\ \frac{C_{\pi K}(t)}{p(t)}
\end{equation}
we then proceed to calculate
\begin{align}
&\widetilde{C}_{\pi K}(t) = C'_{\pi K}(t)-C'_{\pi K}(t+\delta t)\,,\\
&C^{\text{E2}}_{\pi K}(t) = p(t)\widetilde{C}_{\pi K}(t)\,,
\label{eq:e2_meth}
\end{align}
from which $E_{\pi K}$ can be extracted through a fit of the form
\begin{align}
C^{\text{E2}}_{\pi K}\left(t\right)=A_{0}^2\left(e^{-E_{\pi K}t}+e^{-E_{\pi K}\left(T-t\right)}-\frac{p\left(t\right)}{p\left(t+1\right)}\cdot
\left(e^{-E_{\pi K}\left(t+1\right)}+e^{-E_{\pi K}\left(T-\left(t+1\right)\right)}\right)\right) \,.
\label{eq:e2_fit}
\end{align}
\end{itemize}
We remark that for both methods \E{1} and \E{2} the energies $E_\pi$
and $E_K$, i.e. $M_\pi$ and $M_K$ for zero momentum, are required as an input.
They are determined from the corresponding two-point correlation functions.
For the error analysis bootstrap samples are used to fully preserve all correlations.
\label{sec:method_extra}
After solving \Cref{eq:scat_lusch} for $a_0$ up to order $\order{L^{-5}}$ on every ensemble for each strange quark mass of \Cref{tab:mus}, we have three parameters in which we want to extra- or interpolate: the lattice spacing $a$, the strange quark mass $m_s$ and the light quark mass $m_\ell$.
To evaluate $a_0$ at the physical point we follow a two step procedure.
We first fix the strange quark mass to its physical value and do a combined chiral and continuum extrapolation afterwards.
For the extrapolation in the light quark mass, we use the continuum ChPT expression in terms of meson masses and decay constants, as detailed in Refs.~\cite{Kubis:2001ij,PhysRevD.75.054501,Fu:2011wc}.
\subsection{Fixing the strange quark mass}
\label{sec:strange_quark}
In order to fix the strange quark mass we adopt the following procedure:
we match the quantity
\begin{align}
M_s^2=M_K^2-0.5 M_\pi^2\,,
\label{eq:fix_A}
\end{align}
which is proportional to the strange quark mass at the leading order of ChPT, to its physical value
\begin{equation}
(M_s^{\text{phys}})^2\ =\ \overline{M}_K^2 - 0.5 \overline{M}_{\pi}^2 \,,
\label{eq:fix_A_phys}
\end{equation}
using our determinations of $M_K^2$ at three valence strange quark masses on a per-ensemble basis.
For each ensemble, we then interpolate all valence strange quark mass dependent
observables, i.e. $\mu_{\pi K}\,a_0^{3/2}$, $M_K$, $M_{\eta}$ and $\mu_{\pi K}$, in $M_s^2$ to this reference value.
\subsection{Chiral extrapolation}
With the strange quark mass fixed, the extrapolation to the physical point can
be carried out using ChPT. The first NLO calculation of the scattering
amplitude and scattering lengths was done in Ref.~\cite{Bernard:1990kw}.
From the formulae for the isospin even (odd) scattering lengths $a^+$
($a^-$) in Ref.~\cite{Kubis:2001ij} the NLO ChPT
formulae for $\mu_{\pi K}\,a_0^{I}\,,I\in\Set{1/2,3/2}$, can be derived as sketched in \Cref{app:pik_chpt}, giving
\begin{equation}
\begin{split}
\mu_{\pi K}\,a_0^{3/2} &= \frac{\mu^2_{\pi K}}{4\pi f_\pi^2}
\left[\frac{32M_\pi M_K}{f_\pi^2} L_{\pi K}(\Lambda_\chi)-1-
\frac{16 M_\pi^2}{f_\pi^2}L_{5}(\Lambda_\chi)\right. \\
&\left.+\frac{1}{16\pi^2 f_\pi^2} \chi_{\text{NLO}}^{3/2}(\Lambda_\chi,M_\pi,M_K,M_\eta)\right] +c\cdot f(a^2)\,.
\end{split}
\label{eq:a0_alg_fit}
\end{equation}
\Cref{eq:a0_alg_fit} depends on the masses of the pion and the kaon, their reduced mass as defined in \Cref{eq:red_mass}, the $\eta$ mass and the pion decay constant.
In addition, the equation depends on the low energy constants (LECs) $L_5$ and $L_{\pi K}$ while $\chi_{\text{NLO}}^{3/2}$ is a known function, see \Cref{app:nlo}.
We express \Cref{eq:a0_alg_fit} in terms of the meson masses and decay constants as they are determined on the lattice, which has the benefit that their ratios can be determined with high statistical precision without the need for explicit factors of the lattice scale.
Formally we fix the scale-dependent LECs at the renormalization scale $\Lambda_\chi = f_{\pi^-}^\text{(PDG)}$.
However, in practice we employ $a \Lambda_\chi = a f_\pi(\beta,\mu_\ell) / K^\text{FSE}_{f_\pi}$ in all chiral logarithms, where the values for the finite-size correction factor $K^\text{FSE}_{f_\pi}$ are given in \Cref{sec:meson_masses_eshift_scat_length}.
Doing so should only induce higher order corrections in the chiral expansion.
Automatic order $\order{a}$ improvement of Wilson twisted mass fermions at
maximal twist guarantees that the leading lattice artifacts are of
order $\order{a^2}$ or better.
For instance, for the $I=2$ $\pi \pi$ s-wave scattering length, discretization
effects start only at order $\order{a^2 M_\pi^2}$~\cite{Buchoff:2008hh}.
A corresponding theoretical result for $\pi K$ is missing so far.
However, our numerical data suggest that also for $\pi K$ lattice artefacts are very small.
Still, we include a term $c\cdot f(a^2)$ accounting for possible
discretization effects, with fit parameter $c$ and $f(a^2)$ either equal to
$a^2/r_0^2$ or to $a^2 M_X^2$, with $M_X^2$ one of the masses or mass
combinations $M_{\pi}^2\,,\,M_K^2, M_K^2+0.5M_{\pi}^2, \mu_{\pi K}^2$. In the
following analysis we will include the term $c\cdot f(a^2)$ into our fit for
every choice of $f(a^2)$ and thus investigate a possible dependence of our data
on the lattice spacing.
To summarize, our fit parameters are the LECs $L_5$ and $L_{\pi K}$, and $c$, where $L_{\pi K}$ is the combination of renormalized LECs
\begin{align}
L_{\pi K} = 2L_1+2L_2+L_3-2L_4-\frac{L_5}{2}+2L_6+L_8\,.
\label{eq:lec_lin_comb}
\end{align}
Let us mention already here that the fits to the data described in the next section turn out to be not sensitive to $L_5$.
Therefore, we include it as a prior in the fit with the value taken from Ref.~\cite{Aoki:2016frl}.
In slight abuse of language we will denote this extrapolation method
as NLO ChPT.
\subsection{Extrapolations Using the $\Gamma$ Method at Fixed $m_s$}
\label{sec:lin_chpt}
Next, we describe an alternative way to extrapolate our data.
For this we fix the strange quark mass as described in the previous subsection and interpolate the other quantities.
Using the interpolated data we can compute the following quantity~\cite{Beane:2006gj}
\begin{align}
\Gamma\left(\frac{M_\pi}{f_\pi},\,\frac{M_K}{f_\pi}\right) =
-\frac{f^2_\pi}{16M^2_\pi}\left(\frac{4\pi f^2_\pi}{\mu^2_{\pi
K}}[\mu_{\pi K}\,a_0^{3/2}] + 1
+\chi_{\text{NLO}}^{-}-2\frac{M_KM_\pi}{f^2_\pi}\chi_{\text{NLO}}^{+}\right)\,,
\label{eq:gamma_calc}
\end{align}
with next to leading order ChPT functions $\chi^{\pm}_{\text{NLO}}$
given in \Cref{app:nlo}.
Using the ChPT expressions for the isospin even/odd scattering lengths $a^+$ and $a^-$, $\Gamma$ can also be expressed as~\cite{Beane:2006gj}
\begin{align}
\Gamma\left(\frac{M_\pi}{f_\pi},\,\frac{M_K}{f_\pi}\right) =
L_5-2\frac{M_K}{M_\pi}L_{\pi K}\,.
\label{eq:gamma_fit}
\end{align}
For the derivation see \Cref{app:pik_chpt}.
Hence, a linear fit of \Cref{eq:gamma_fit} to the data obtained via \Cref{eq:gamma_calc} allows one to determine the LECs $L_5$ and $L_{\pi K}$.
Given $L_5$ and $L_{\pi K}$ from the fit one can compute $\mu_{\pi K}\,a_0^{3/2}$ at the physical point using \Cref{eq:a0_alg_fit}.
Again, it turns out we are not sensitive to $L_5$ in our fits.
Therefore, we use a prior as discussed before.
This extrapolation method we denote as $\Gamma$ method.
\section{Results}
\label{sec:results}
\begin{figure}[t]
\includegraphics{branch_comparison.pdf}
\caption{Comparison of values for $\mu_{\pi K}\,a_0^{3/2}$ in the continuum
limit at the physical point obtained with the different methods
used in this paper. The fit ranges decrease with increasing index as
described in \Cref{tab:chpt_ranges}. The inner error band
represents the statistical error only, while the outer error band
represents the statistical and systematic errors added in quadrature.}
\label{fig:methods}
\end{figure}
In this section we present our main result for $\mu_{\pi K}\,a_0^{3/2}$ in the continuum limit and extrapolated to the physical point.
We use two thermal state pollution removal methods, \E{1} and \E{2}, for
$E_{\pi K}$. Next we employ the two (related) ChPT extrapolations,
$\Gamma$-method and NLO ChPT, as discussed before. For each of the two
ChPT extrapolation methods we use three fit ranges as compiled in
\Cref{tab:chpt_ranges}. Hence, we have twelve estimates for each quantity
at the physical point in the continuum limit available, which we use
to estimate systematic uncertainties. We remark that the fit for the
$\Gamma$ method is in terms of $M_K/M_\pi$ and for NLO ChPT in terms
of $\mu_{\pi K}/f_\pi$. Thus, we vary the fit range at the lower end
for the $\Gamma$ method and at the upper end for NLO ChPT.
For $\mu_{\pi K}\,a_0^{3/2}$ the twelve estimates are shown in
\Cref{fig:methods}. The final result is obtained as the weighted average over all of these,
as shown in the figure as the horizontal bold line. The
weight is computed according to
\begin{equation}
w = \frac{(1- 2 \cdot |p - 1/2|)^2}{\Delta^2}
\end{equation}
with $p$ the $p$-value of the corresponding ChPT fit and $\Delta$ the
statistical uncertainty obtained from the fit.
The statistical uncertainty of the final results is determined from
the bootstrap procedure. For $\mu_{\pi K}\,a_0^{3/2}$ this is shown in
\Cref{fig:methods} as the inner error band.
In addition, we determine three systematic uncertainties: The first is
obtained from the difference between using only \E{1} or only \E{2}
results. The second from the difference between using only the $\Gamma$
method or only NLO ChPT. Finally, we use the maximal difference of the
weighted average to the twelve estimates as a systematic uncertainty
coming from the choice of fit ranges.
The results of all twelve fits can be found in \Cref{tab:lin_chpt_res_ex}
for the $\Gamma$ method and \Cref{tab:nlo_chpt_res_ex} for NLO ChPT fits.
The fit range indices used in \Cref{fig:methods} are resolved in
\Cref{tab:chpt_ranges}.
\begin{table}
\centering
\begin{tabular*}{.5\textwidth}{@{\extracolsep{\fill}}lrrr}
\hline\hline
Method & index & Begin & End \\
\hline
\multirow{3}{*}{$\Gamma$} &1 &1.2 &2.0 \\
&2 &1.4 &2.0 \\
&3 &1.5 &2.0 \\
\hline
\multirow{3}{*}{NLO} &1 &1.2 &1.6 \\
&2 &1.2 &1.41 \\
&3 &1.2 &1.35 \\
\hline\hline
\end{tabular*}
\caption{Fit ranges used for extrapolations $\Gamma$ and NLO ChPT. The index
column refers to \Cref{fig:methods}.}
\label{tab:chpt_ranges}
\end{table}
With this procedure and all
errors added in quadrature we quote
\begin{align}
\mu_{\pi K}\,a_0^{3/2,\text{phys}} = -0.0463(17)\,,\qquad
L_{\pi K} = 0.0038(3)\,.
\end{align}
This translates to
\begin{align}
M_{\pi}\,a_0^{3/2,\text{phys}} = -0.059(2)\,,\qquad
M_{\pi}\,a_0^{1/2,\text{phys}}=0.163(3)
\end{align}
as our final results. The error budget is compiled in
\Cref{tab:errorbudget}.
While the dominating contribution to the error for both $\mu_{\pi K}\,a_0^{3/2}$
and $L_{\pi K}$
summed in quadrature is coming from the fit range and the statistical
uncertainty, also the choice of the thermal state removal method
contributes significantly.
The contribution from the different chiral extrapolation methods is negligible.
If the errors were added (not in quadrature),
the total error would become a factor $\sim1.7$ larger.
We remark that these results have been obtained with $L_5$ as an
input, because the fits are not sufficiently sensitive to determine
$L_5$ directly. We use the most recent determination from a
$N_f=2+1+1$ lattice calculation by HPQCD~\cite{Aoki:2016frl}, which is
extrapolated to the continuum limit. At our renormalization scale it reads
\begin{equation}
L_5\ =\ \num{5.4(3)e-3}\,.
\label{eq:l5_prior}
\end{equation}
\begin{table}[ht]
\centering
\begin{tabular*}{.7\textwidth}{@{\extracolsep{\fill}}lrr}
\hline\hline
& $\mu_{\pi K}\,a_0^{3/2}\cdot 10^5$ & $L_{\pi K}\cdot 10^5$ \\
\hline
statistical & $82\ (28\%)$ & $15\ (32\%)$\\
fit range & $139\ (47\%)$ & $19\ (41\%)$\\
\E{1} v. \E{2} & $64\ (22\%)$ & $12\ (24\%)$\\
NLO ChPT v. $\Gamma$ & $9\ (3\%)$ & $1\ (3\%)$\\
\hline
$\sum$ & $294\ (100\%)$ & $47\ (100\%)$ \\
sqrt $\sum$ in quadrature & 173 &27 \\
\hline\hline
\end{tabular*}
\caption{Error budget for the final results of $\mu_{\pi K}\,a_0^{3/2}$ and
$L_{\pi K}$. }
\label{tab:errorbudget}
\end{table}
\section{Analysis Details and Discussion}
\subsection{Error Analysis, Thermal Pollution and Choice of Fit Ranges}
The error analysis is performed using the stationary blocked bootstrap
procedure~\cite{Dimitris:1994}.
In order to determine an appropriate average block length, we compute the integrated autocorrelation time $\tau_{\mathrm{int}}$ for the correlation functions $C_{X}(t)$ at all source-sink separations, with $X$ being $\pi$, $K$, $\eta$ or $\pi K$.
In the case of $\pi K$, $C_{X}(t)$ is of course first suitably transformed for the extraction of the interaction energy as discussed in \cref{sec:pollution}.
The computation of $\tau_{\mathrm{int}}$ is detailed in Ref.~\cite{Wolff:2003sm}.
The average block length is then chosen to be the ceiling of the maximum integrated autocorrelation time
observed over all correlation functions at all source-sink separations
\begin{displaymath}
b=\lceil \max_{X,t} \left( \tau^{(X,t)}_{\mathrm{int}} \right) \rceil
\end{displaymath}
on a per-ensemble basis.
We have confirmed explicitly that this method produces a block length at which the
estimate of the statistical error
plateaus and are thus confident that we properly take into account
the effect of autocorrelations on our quoted statistical errors.
Using the so-determined block length on a per-ensemble basis,
we generate $N=1500$ samples from which
we estimate statistical errors throughout our analysis.
As discussed in \Cref{sec:pollution}, we employ methods \E{1} and \E{2} to remove unwanted thermal pollutions from the $\pi K$ two particle correlation function.
Both methods allow us to describe the data rather well, but the choice of best fit range depends on the method used to remove the thermal pollution.
This in turn affects the value of the extracted $E_{\pi K}$ and, subsequently, the value of $\mu_{\pi K}\,a_0^{3/2}$ obtained from the energy difference.
To demonstrate the quality of our fits, we look at the ratio
\begin{equation}
\frac{C^{\text{[E1,E2]}}_{\pi K}(t)}{f^{\text{[E1,E2]}}(t) } \,,
\label{eq:C_over_fit_compare}
\end{equation}
where $C^{\text{[E1,E2]}}$ are defined in
\Cref{eq:reweight,eq:e2_meth}, respectively, and the fit functions
$f^{\text{[E1,E2]}}(t)$ are given in \Cref{eq:e_fit,eq:e2_fit}, respectively.
The ratio is shown in \Cref{fig:C_over_fit_compare} for the two ensembles A40.24 and A40.32 for two fit ranges for which both methods describe the data well.
\begin{figure}
\begin{subfigure}{0.49\textwidth}
\includegraphics{{ratio_cmp_paper_A40.24}.pdf}
\caption{Fit range: $[14,20]$
\\ $aE_{\pi K} = \lbrace 0.389(1)^{\text{E1}}, 0.389(1)^{\text{E2}} \rbrace$}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics{{ratio_cmp_paper_A40.32}.pdf}
\caption{Fit range: $[20,28]$
\\$aE_{\pi K} = \lbrace 0.3793(6)^{\text{E1}}, 0.3789(5)^{\text{E2}} \rbrace$}
\end{subfigure}
\caption{Plot of \Cref{eq:C_over_fit_compare} for ensembles A40.24 and A40.32
for the lightest strange quark mass for the fit ranges used for the analysis, comparing the quality of the data description by methods \E{1} and \E{2}.}
\label{fig:C_over_fit_compare}
\end{figure}
The choice of the fit ranges to determine energy levels is always
difficult. In the past, we have used many fit ranges and weighted them
according to their fit
qualities~\cite{Helmes:2015gla,Helmes:2017smr}. However, this
procedure relies on properly estimated variance-covariance matrices,
which is notoriously difficult. For the pion-kaon correlation
functions needed in this paper we have observed several cases where
the fit including the variance-covariance matrix did not properly
describe the data after visual inspection. Therefore, we use fits here
assuming independent data points with the correlation still taken into
account by the bootstrap procedure.
As a consequence, we cannot apply the weighting procedure used in
Refs.~\cite{Helmes:2015gla,Helmes:2017smr} any longer
and have to choose fit ranges. The procedure is as follows:
Due to exponential error growth of $C_{\pi K}$ we fix $t_f=T/2-4a$ and vary
$t_i$, beginning from a region where excited states do not contribute
significantly anymore. From these fits we choose one fit range where the ratio of
\Cref{eq:C_over_fit_compare} is best compatible with 1. The statistical error is
calculated from the bootstrap samples as discussed before. We then estimate the systematic
uncertainty from the remaining fit ranges. To this end we determine the
difference of the mean value to the upper and lower bound of values for $E_{\pi
K}$. This procedure results in an asymmetric estimate of the systematic
uncertainty of $E_{\pi K}$. The results are compiled in \Cref{tab:raw_data_energies_comp}.
Since $C_K$ and $C_\pi$ do not suffer from exponential error growth at late times we set $t_f=T/2$.
\Cref{tab:fit_ranges} gives an overview of our chosen values of $t_i$ and $t_f$ for all ensembles.
\begin{table}
\centering
\begin{tabular*}{.6\textwidth}{@{\extracolsep{\fill}}lllllll}
\hline\hline
& \multicolumn{2}{c}{$aM_\pi$} & \multicolumn{2}{c}{$aM_K$} & \multicolumn{2}{c}{$aE_{\pi K}$} \\
\hline
Ensemble & $t_i$ & $t_f$ & $t_i$ & $t_f$ & $t_i$ & $t_f$\\
\hline
A30.32 & 13 & 32 & 13 & 32 & 21 & 28\\
A40.24 & 11 & 24 & 11 & 24 & 14 & 20\\
A40.32 & 13 & 32 & 13 & 32 & 20 & 28\\
A60.24 & 11 & 24 & 11 & 24 & 16 & 20\\
A80.24 & 11 & 24 & 11 & 24 & 16 & 20\\
A100.24 & 11 & 24 & 11 & 24 & 15 & 20\\
B35.32 & 13 & 32 & 13 & 32 & 22 & 28\\
B55.32 & 13 & 32 & 13 & 32 & 19 & 28\\
B85.24 & 11 & 24 & 11 & 24 & 15 & 20\\
D45.32sc & 14 & 32 & 14 & 32 & 22 & 28\\
D30.48 & 22 & 48 & 22 & 48 & 35 & 44\\
\hline\hline
\end{tabular*}
\caption{Typical minimal and maximal values of the starting and end points of fit
ranges for the Correlation functions under investigation.}
\label{tab:fit_ranges}
\end{table}
It is always difficult to include systematic uncertainties in the
analysis chain. Since we see systematic uncertainties on extracted
energies on the same level as the statistical one, we adopt the
following procedure to include this uncertainty:
Because $\mu_{\pi K}\,a_0^{3/2}$ is derived from $E_{\pi K}$ we chose to scale
the statistical error for $\mu_{\pi K}\,a_0^{3/2}$ on each ensemble after the data have been
interpolated in the strange quark mass. To this end we define a scaling
factor $s$ via the standard error $\Delta X$ and the average of the
systematic uncertainties $\overline{Q}_X$ over the 3 strange quark
masses for each ensemble:
\begin{align}
s =
\sqrt{ \frac{ (\Delta X)^2+\overline{Q}_X^2}{(\Delta X)^2} }\,,
\label{eq:scalefac_err}
\end{align}
where the average $\overline{Q}_X$ is the simple mean over the six
systematic errors.
\subsection{\boldmath Meson Masses, Energy Shift $\delta E$ and
Scattering Length $\mu_{\pi K}\,a_0^{3/2}$}
\label{sec:meson_masses_eshift_scat_length}
In order to extract $\delta E$ we first determine $M_K$ and $M_\pi$ from fitting \Cref{eq:c2_fit} to our data for $C_{\pi}(t)$ and $C_{K}(t)$.
We then calculate the reduced mass $\mu_{\pi K}$ via \Cref{eq:red_mass} for all combinations of fit ranges.
$M_\pi$, $M_K$ and $\mu_{\pi K}$ are listed in \Cref{tab:raw_data_masses_comp}.
The two methods \E{1} and \E{2} give us two estimates of $E_{\pi K}$ as outlined in \Cref{sec:pollution}, from which we determine $\delta E$ and hence the scattering length using \Cref{eq:scat_lusch}.
The values for $E_{\pi K}$ and $\delta E$ are collected in \Cref{tab:raw_data_energies_comp,tab:raw_data_deltae_comp}.
We introduce factors $K^{\text{FSE}}_X$ for
$X~\in~\{M_K,\,M_\pi,\,f_\pi\}$ to correct our lattice data for finite size
effects. They have already been calculated in
Ref.~\cite{Carrasco:2014cwa} and are listed in \Cref{tab:ext_dat}. We apply
these factors for e.g. $M_\pi$ like
\[
M^*_\pi=\frac{M_\pi}{K^{\text{FSE}}_{M_\pi}}\,.
\]
We correct every quantity of the set named above and drop the asterisk in what follows to improve legibility.
For $M_\eta$ statistical uncertainties are too big to resolve finite
volume effects, see also Ref.~\cite{Ottnad:2017bjt}.
For the two methods \E{1} and \E{2} we solve \Cref{eq:scat_lusch} for $a_0$ up to order
$\mathcal{O}(L^{-5})$ numerically.
The values for $a_0$ and its product with the reduced mass, $\mu_{\pi K}\,a_0^{3/2}$ are collected
in \Cref{tab:raw_data_scatlen_comp}.
Since the finite size behavior of the scattering length is unknown, we do not
apply finite size corrections to the reduced mass appearing in $\mu_{\pi K}\,a_0^{3/2}$, either.
\subsection{Strange quark mass fixing}
Before extrapolating to the continuum limit and the physical point, we
interpolate all data to reference strange quark masses as discussed
before. The data for the three strange quark masses are strongly
correlated because the same stochastic light perambulators were used
for all light-strange observables. As a consequence, the
variance-covariance matrix was sometimes not sufficiently well
estimated such that its inverse was unreliable.
As a result, we resort to performing these fits using uncorrelated $\chi^2$ which results in best fit parameters which describe the data much better.
It should be noted that all statistical covariance is still fully taken into account by the bootstrap procedure and our final statistical errors on all fit parameters are correctly estimated.
As an example, we show in \Cref{fig:ma_interp}
the interpolation of $\mu_{\pi K}\,a_0^{3/2}$ in $M_K^2 - 0.5M_\pi^2$ for
ensemble B55.32 comparing methods \E{1} and \E{2}. The
large uncertainty in the interpolation variable stems mainly from the uncertainty in the scaling
quantity $r_0$. Furthermore the errors of the three data points are highly
correlated.
The interpolation to the reference point is shown as a red
diamond. In general, the strange quark mass dependence of $\mu_{\pi K}\,a_0^{3/2}$ is mild and stems mainly
from the reduced mass $\mu_{\pi K}$.
The values thus determined are compiled in
\Cref{sec:chpt_A_input}. They serve as input data for the subsequent chiral
extrapolations.
\begin{figure}[h]
\centering
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\textwidth]{{eval_lfit_mua32_M1A_E1_B55.32}.pdf}
\subcaption{B55.32 for \E{1}}
\label{fig:ma_interp_M1AE2}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\textwidth]{{eval_lfit_mua32_M1A_E2_B55.32}.pdf}
\caption{B55.32 for \E{2} }
\label{fig:ma_interp_mB}
\end{subfigure}
\caption{Interpolations of $\mu_{\pi K}\,a_0^{3/2}$ for the two different methods \E{1} and
\E{2}.}
\label{fig:ma_interp}
\end{figure}
\subsection{Chiral and Continuum Extrapolations}
\label{sec:interp}
Having interpolated all our lattice data to a fixed reference strange quark mass corresponding to the physical strange quark mass at leading order, we will describe below possible systematic errors in our chiral and continuum extrapolations.
\subsubsection{\boldmath $\Gamma$-Method}
\label{sec:res_gamma}
In this section we present results employing the determination of $L_{\pi K}$ using the linear fit introduced in \Cref{sec:lin_chpt}.
\Cref{fig:lin_chpt_a} shows the chiral extrapolations in terms of $M_K/M_{\pi}$
for pollution removal \E{1} and \E{2}.
Since we work at fixed strange quark mass, the light quark mass decreases from left to right in the figure.
In order to check how our extraction of $L_{\pi K}$ is affected by the range of
included pion masses we employ three different fit ranges
\begin{align}
\frac{M_K}{M_{\pi}}\in\Set{(0;2.0),(1.5;2.0),(1.4;2.0)}\,.
\label{eq:fit_intervals_gamma}
\end{align}
In \Cref{tab:lin_chpt_res_ex} we compile the results for the fits corresponding
to the data points of \Cref{fig:lin_chpt_a} for all 3 fit ranges. As the fit
range is restricted to our lightest ensembles, the value extracted
for $L_{\pi K}$ tends up, while the absolute value of the extracted
$\mu_{\pi K}\,a_0^{3/2}$ decreases. It is worth noting that this behavior is only
observed for the pollution removal \E{2}, whereas for \E{1} the values for
$L_{\pi K}$ and $\mu_{\pi K}\,a_0^{3/2}$ stay constant within their statistical
errors.
\begin{figure}
\begin{subfigure}{0.49\textwidth}
\includegraphics{pi_K_I32_fse_true_gamma_M1A_E1_fr0.pdf}
\caption{Pollution Removal \E{1}}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics{pi_K_I32_fse_true_gamma_M1A_E2_fr0.pdf}
\caption{Pollution Removal \E{2}}
\end{subfigure}
\caption{Linearized chiral extrapolations of $\Gamma$ with data
interpolated to the reference strange quark mass. The data for different
lattice spacings are color encoded. In addition we show the linear fit
(solid curve, gray error band)}
\label{fig:lin_chpt_a}
\end{figure}
\begin{table}
\centering
\begin{tabular*}{.9\textwidth}{@{\extracolsep{\fill}}l
c
c
S[fixed-exponent=-3,table-omit-exponent]
S[fixed-exponent=-2,table-omit-exponent]}
\hline\hline
Removal & {Fit range} & $p$-value & {$L_{\pi K}\times 10^{3}$} & {$\mu_{\pi K}\,a_0^{3/2} \times 10^{2}$}\\
\hline
\multirow{3}{*}{\E{1}} & \numrange{0.0}{2.0}&0.8&0.0037 +- 0.0002 &-0.0468 +- 0.0009 \\
& \numrange{1.4}{2.0}&0.7&0.0037 +- 0.0002 & -0.047 +- 0.001 \\
& \numrange{1.5}{2.0}&0.6&0.0036 +- 0.0003 & -0.047 +- 0.002 \\
\hline
\multirow{3}{*}{\E{2}} & \numrange{0.0}{2.0}&0.3&0.0037 +- 0.0001 &-0.0465 +- 0.0007\\
& \numrange{1.4}{2.0}&0.5&0.0038 +- 0.0002 &-0.0457 +- 0.0009\\
& \numrange{1.5}{2.0}&0.4&0.0040 +- 0.0002 & -0.045 +- 0.001\\
\hline\hline
\end{tabular*}
\caption{Fit results for the Gamma method. The fits shown in
\Cref{fig:lin_chpt_a} correspond to the largest fit range in the table.}
\label{tab:lin_chpt_res_ex}
\end{table}
\subsubsection{Chiral Perturbation Theory at NLO}
\label{sec:i32_chpt}
We first fit the continuum ChPT formula \Cref{eq:a0_alg_fit} with
$c=0$, i.e. assuming no lattice artefacts.
In the right column of plots in figure \Cref{fig:pik_chpt_a1}, we show
the lattice data for $\mu_{\pi K}\,a_0^{3/2}$ interpolated to the reference strange quark
mass as a function of $\mu_{\pi K}/f_{\pi}$ for the two thermal
pollution removal methods
\E{1} and \E{2}. The solid line corresponds to the leading order,
parameter free ChPT prediction. Plotting our best fit curve with NLO
ChPT together with the data is difficult, because $\mu_{\pi K}\,a_0^{3/2}$ depends on
meson masses and $f_\pi$ besides $\mu_{\pi K}/f_{\pi}$. Therefore, in
order to demonstrate that the fit is able to describe our data, we
indicate the relative deviation $\delta_r(\mu_{\pi K}\,a_0^{3/2})$ between the fitted
points and the original data
\begin{align}
\delta_r(\mu_{\pi K}\,a_0^{3/2})=
\frac{(\mu_{\pi K}\,a_0^{3/2})^{\text{meas}}-(\mu_{\pi K}\,a_0^{3/2})^{\text{fit}}}
{(\mu_{\pi K}\,a_0^{3/2})^{\text{meas}}}\,,
\label{eq:rel_dev_ma}
\end{align}
in \Cref{fig:pik_chpt_a1e1_rel_dev,fig:pik_chpt_a1e2_rel_dev}.
The indicated error bars are statistical only and it is clear that
within these uncertainties, our data is reasonably well described by
the fit.
\begin{figure}[h!]
\begin{subfigure}{0.49\textwidth}
\includegraphics{pi_K_I32_fse_true_nlo_chpt_M1A_E1_fr0.pdf}
\caption{\E{1}:$\mu_{\pi K}\,a_0^{3/2}$ in terms of $\mu_{\pi K}/f_{\pi}$ together with LO ChPT formula.}
\label{fig:pik_chpt_a1e1}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics{pi_K_I32_fse_true_nlo_chpt_M1A_E1_fr0_rel_dev.pdf}
\caption{\E{1}:Relative deviation between the measured and the fitted values
of $\mu_{\pi K}\,a_0^{3/2}$.}
\label{fig:pik_chpt_a1e1_rel_dev}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics{pi_K_I32_fse_true_nlo_chpt_M1A_E2_fr0.pdf}
\caption{\E{2}:$\mu_{\pi K}\,a_0^{3/2}$ in terms of $\mu_{\pi K}/f_{\pi}$ together with LO ChPT formula.}
\label{fig:pik_chpt_a1e2}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics{pi_K_I32_fse_true_nlo_chpt_M1A_E2_fr0_rel_dev.pdf}
\caption{\E{2}:Relative deviation between the measured and the fitted values
of $\mu_{\pi K}\,a_0^{3/2}$.}
\label{fig:pik_chpt_a1e2_rel_dev}
\end{subfigure}
\caption{Chiral extrapolation of $\mu_{\pi K}\,a_0^{3/2}$ for \E{1} and \E{2}. Different colors and symbols denote different values of $\beta$. In black we plot the LO ChPT formula. The golden diamond gives our final result using the given method at the physical point.}
\label{fig:pik_chpt_a1}
\end{figure}
As in \Cref{sec:res_gamma}, to investigate the validity of \Cref{eq:a0_alg_fit} across our entire range of pion masses, we studied three different fit intervals for $\mu_{\pi K}/f_{\pi}$, namely
\begin{align}
\frac{\mu_{\pi K}}{f_{\pi}}\in\Set{(0;1.35),(0;1.41),(0;1.60)}\,,
\end{align}
where now the first range corresponds to only using our lightest pion masses and the third range includes all of our ensembles.
The resulting trend in the extracted values of $L_{\pi K}$ and $\mu_{\pi K}\,a_0^{3/2}$ is shown in \Cref{tab:nlo_chpt_res_ex}.
Just as in the study of the $\Gamma$-method, including heavier pion masses leads
to smaller values of $L_{\pi K}$ and correspondingly smaller values of $\mu_{\pi K}\,a_0^{3/2}$.
\begin{table}
\centering
\begin{tabular*}{.9\textwidth}{@{\extracolsep{\fill}}l
c
c
S[fixed-exponent=-3,table-omit-exponent]
S[fixed-exponent=-2,table-omit-exponent]}
\hline\hline
Removal & {Fit range} & $p$-value & {$L_{\pi K}\times 10^{3}$} & {$\mu_{\pi K}\,a_0^{3/2} \times 10^{2}$}\\
\hline
\multirow{3}{*}{\E{1}} &\numrange{0.0}{1.35} &0.6& 0.0036 +- 0.0003 & -0.047 +- 0.002\\
&\numrange{0.0}{1.41} &0.7& 0.0036 +- 0.0002 & -0.047 +- 0.001\\
&\numrange{0.0}{1.60} &0.7& 0.0037 +- 0.0002 & -0.047 +- 0.001\\
\hline
\multirow{3}{*}{\E{2}} &\numrange{0.0}{1.35} &0.5&0.0039 +- 0.0002 & -0.045 +- 0.001 \\
&\numrange{0.0}{1.41} &0.5&0.0038 +- 0.0002 & -0.046 +- 0.001 \\
&\numrange{0.0}{1.60} &0.4&0.0037 +- 0.0001 & -0.0464 +- 0.0007 \\
\hline\hline
\end{tabular*}
\caption{Fit results for the NLO ChPT fit. The fits shown in
\Cref{fig:pik_chpt_a1} correspond to the largest fit range in the table.}
\label{tab:nlo_chpt_res_ex}
\end{table}
To investigate possible discretization effects we now allow $c\neq 0$ in \Cref{eq:a0_alg_fit}.
Fitting \Cref{eq:a0_alg_fit} for the different choices of $f(a^2)$, we are neither able to obtain a statistically significant result for the fit parameter $c$, nor do we see significant differences in the extracted values of $L_{\pi K}$ and $\mu_{\pi K}\,a_0^{3/2}$.
We conclude that we are not able to resolve lattice artifacts in this
quantity given our statistical uncertainties.
\section{Discussion}
\label{sec:discussion}
Let us first discuss the main systematics of our computation:
In contrast to the pion-pion or kaon-kaon systems, there are time
dependent thermal pollutions in the correlation functions relevant for the
extraction of the pion-kaon s-wave scattering length. This very fact
turns out to represent one of the major systematic uncertainties in
the present computation. We have investigated two methods to
remove the leading thermal pollutions, denoted as \E{1} and \E{2}.
With both we are able to describe the data for the correlation
functions. However, there is uncertainty left, because we remove only
the leading pollution and the removal procedure requires input
estimated from other two point functions. Thus, we eventually decided
to use both methods \E{1} and \E{2} and include the differences in the
systematic uncertainty.
Secondly, we perform a mixed action simulation for the strange quark. We
use this to correct for small mis-tuning in the sea strange quark mass
value used for the gauge configuration generation. This leads --- at
least in principle --- to a small mismatch in the renormalization
condition used for the continuum extrapolation. We cannot resolve the
corresponding effect on our results quantitatively given our
statistical uncertainties. But, since we
study quantities which mainly depend on the valence quark
properties we expect them to be small.
Thirdly, in the ChPT determination of $L_{\pi K}$ the remaining LEC, $L_5$,
entered as a prior to numerically stabilize our fits. The HPQCD value of
Ref.~\cite{Dowdall:2013rya} stems from an independent lattice
simulation, but is extrapolated to the continuum limit.
In Ref.~\cite{Dowdall:2013rya} $L_5$ is given at scale $M_\eta$,
which we translated to our renormalization scale given by the pion
decay constant.
In addition, the extrapolation from our data to the physical point is
quite long. Here, a computation directly with physical pion mass would
improve our confidence in the result. The final error on our
determination is only as small as it is due to the highly constraining
ChPT description of $\mu_{\pi K}\,a_0^{3/2}$.
Finally, although we are not able to resolve lattice artefacts in
our determination of $\mu_{\pi K}\,a_0^{3/2}$ our statistical errors and limited set
of gauge ensembles especially at the finest lattice spacing might
make us unable to resolve possible lattice artefacts.
\section{Summary}
\label{sec:summary}
\begin{figure}[t!]
\centering
\begin{subfigure}{.49\textwidth}
\includegraphics[height=210pt]{final_comparison.pdf}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\includegraphics[height=210pt]{scattering_maxi3.pdf}
\end{subfigure}
\caption{Left: Comparison of physical results for $M_\pi a_0^{I=3/2}$
from various lattice
computations~\cite{Beane:2006gj,Sasaki:2013vxa,Fu:2011wc,Janowski:2014uda,Lang:2012sv}.
The unfilled point denotes the LO extrapolation to the physical point using
the data of Ref.~\cite{Lang:2012sv}.
Right: s-wave scattering lengths for the pion-pion, pion-kaon and
kaon-kaon maximum isospin channels as a function of the squared
reduced mass $\mu^2$ of the system divided by $f_\pi^2$ for pion-pion and
pion-kaon and by $f_K^2$ for kaon-kaon.}
\label{fig:GBsummary}
\end{figure}
In this paper we have presented a first lattice computation of the
pion-kaon s-wave scattering length for isospin $I=3/2$ extrapolated to
the continuum limit. By varying our methodology we estimate the
systematic uncertainties in our results. Our errors cover statistical
uncertainties, continuum and chiral extrapolations as well as the
removal of thermal pollutions.
In the left panel of \Cref{fig:GBsummary} we compare the results
presented in this paper with previous lattice determinations. The
inner (darker)
error bars show the purely statistical errors whereas the outer (lighter) ones correspond
to the statistical and systematic errors added in quadrature.
Even though the four other determinations lack the extrapolation to the continuum
limit, overall agreement within errors is observed. However,
concerning the final uncertainty, our determination improves
significantly on the previous determinations by controlling more
sources of uncertainty.
As mentioned in the introduction, the three two particle systems
pion-pion, pion-kaon and kaon-kaon are very similar. Therefore, it is
interesting to compare the data for pion-pion~\cite{Helmes:2015gla},
kaon-kaon~\cite{Helmes:2017smr} and pion-kaon in a single plot.
This is done in the right panel of \Cref{fig:GBsummary} where we show $\mu\cdot a_0$ as a
function of $(\mu/f)^2$. Here $\mu$ is the reduced mass of the
corresponding two particle system and $f$ is the pion decay constant
$f_\pi$ for the pion-pion and pion-kaon and $f_K$ for
the kaon-kaon system. The dashed line in the right panel of \Cref{fig:GBsummary}
is the leading order, parameter-free ChPT prediction all three systems share. The
three symbols (and colors) represent our data for the three different
systems, respectively.
It can be seen that for all three systems, the deviations from LO ChPT are small.
For the pion-kaon system, a parametrization in terms of $f_K \cdot f_\pi$ would bring the points even closer to the LO line, while increasing the deviation of the final result from the LO estimate.
For the kaon-kaon system, instead, a parametrization in terms of $f_\pi$ rather than $f_K$ (which is perfectly valid at this order of ChPT) would render the deviation from the LO line more severe.
It is somewhat surprising that ChPT appears to work so well for all
three systems, especially for the heavier points in our simulations
and even more for the kaon-kaon system,
where the expansion parameter becomes large. A possible reason for
this finding might be the fact that all three systems are only weakly
interacting.
\section*{Acknowledgements}
We thank the
members of ETMC for the most enjoyable collaboration. The computer
time for this project was made available to us by the John von
Neumann-Institute for Computing (NIC) on the Juqueen and Jureca
systems in J{\"u}lich.
We thank K.~Ottnad for providing us with the data for $f_\pi$ and S.~Simula for the estimates of the finite size
corrections to $M_\pi$, $M_K$ and $f_\pi$.
This project was funded by the DFG as a project in
the Sino-German CRC110. The open source software
packages tmLQCD~\cite{Jansen:2009xp}, Lemon~\cite{Deuzeman:2011wz}, QUDA~\cite{Clark:2009wm,Babich:2011np,Clark:2016rdz},
R~\cite{R:2005}, python~\cite{python:2001} and SciPy~\cite{scipy:2018} have been used.
\bibliographystyle{h-physrev5}
|
\section{Introduction and preliminaries}
\ \ Throughout this paper, let $G$ be a finite simple graph with vertex set $V=V(G)$ and edge set $E=E(G)$. We use \cite{w} as a reference for terminology and notation which are not defined here. The {\em open neighborhood} of a vertex $v$ is denoted by $N(v)$, and the {\em closed neighborhood} of $v$ is $N[v]=N(v)\cup\{v\}$. The {\em minimum} and {\em maximum degree} of $G$ are respectively denoted by $\delta=\delta(G)$ and $\Delta=\Delta(G)$. The {\em corona} of two graphs $G_{1}$ and $G_{2}$ is the graph $G_{1}\circ G_{2}$ formed from one copy of $G_{1}$ and $|V(G_{1})|$ copies of $G_{2}$ where the $ith$ vertex of $G_{1}$ is adjacent to every vertex in the $ith$ copy of $G_{2}$.
Let $S\subseteq V(G)$. For a real-valued function $f:V(G)\rightarrow \mathbb{R}$ we define $f(S)=\sum_{v\in S}f(v)$. Also, $\omega(f)=f(V(G))$ is the weight of $f$. A {\em signed dominating function} ({\em signed 2-independence function}), abbreviated SDF (S2IF), of $G$ is a function $f:V(G)\rightarrow\{-1,1\}$ such that $f(N[v])\geq1$ ($f(N[v])\leq1$), for every $v\in V(G)$. The {\em signed domination number} ({\em signed 2-independence number}), abbreviated SDN (S2IN), of $G$ is $\gamma_{s}(G)=\min\{f(V(G))|f \mbox{ is a SDF of}\ G\}$ ($\alpha_{s}^{2}(G)=\max\{f(V(G))|f \mbox{ is a S2IF of}\ G\}$). These parameters were introduced in \cite{dthhs} and \cite{z}, respectively.
A set $S\subseteq V(G)$ is a {\em dominating set} if each vertex in $V(G)\backslash S$ has at least one neighbor in $S$.
Gallant et al. \cite{gghr} introduced the concept of {\em limited packing} in graphs. They exhibited some real-world applications of it to network security, NIMBY, market saturation and codes. In fact as it is defined in \cite{gghr}, a set of vertices $B\subseteq V(G)$ is called a {\em $k$-limited packing} in $G$ provided that for all $v\in V(G)$, we have $|N[v]\cap B|\leq k$. The {\em limited packing number}, denoted $L_{k}(G)$, is the largest number of vertices in a $k$-limited packing set. In \cite{hh}, Harary and Haynes introduced the concept of {\em tuple domination} in graphs. A set $D\subseteq V(G)$ is a {\em $k$-tuple dominating set} in $G$ if $|N[v]\cap D|\geq k$, for all $v\in V(G)$. The {\em $k$-tuple domination number}, denoted $\gamma_{\times k}(G)$, is the smallest number of vertices in a $k$-tuple dominating set. In fact the authors showed that every graph $G$ with $\delta\geq k-1$ has a $k$-tuple dominating set and hence a $k$-tuple domination number.
A function $f:V(G)\rightarrow\{-1,1\}$ is said to be a {\em nonnegative signed dominating function} (NNSDF) of $G$ if $f(N[v])\geq0$ for each $v\in V(G)$. The {\em nonnegative signed domination number} (NNSDN) of $G$, $\gamma^{NN}_{s}(G)$, is the minimum weight of an NNSDF of $G$. This concept was introduced in \cite{hfx}. For more information the reader can consult \cite{as}.
In this paper, we continue the investigating of the concept of nonnegative signed domination in graphs. In section $2$, we present sharp lower and upper bounds on NNSDN of regular graphs, by using the properties of the above graph parameters. Specifically, we prove that $\gamma^{NN}_{s}(G)\leq n/3$ for a cubic graph $G$ of order $n$. In section $3$, we show that the lower bound $2(-1+\sqrt{1+2n})-n$ for NNSDN of a bipartite graph $G$ of order $n$, given in \cite{as}, is not true as it stands. We correct it by giving a more general result on ($r+1$)-clique-free graphs ($r\geq2$) as an application of the well-known theorem of Tur\'{a}n from the extremal graph theory. Also, we characterize all such graphs attaining the new bound. Finally, in section 4 we give lower and upper bounds on NNSDN with emphasis on trees as:
\begin{center}$-n+2\lceil\frac{\Delta+1}{2}\rceil\leq \gamma^{NN}_s(T)$,\ and $\gamma^{NN}_s(T)\leq n-\ell-s'$ for $n\geq3$\end{center}
where $\ell$ and $s'$ are the number of leaves and the support vertices with odd number of leaves, respectively. Moreover, we give the characterizations of all trees attaining these bounds.
For convenience, throughout the paper we make use of the following notation. Let $f:V(G)\rightarrow\{-1,1\}$ be an NNSDF of graph $G$. Define $V_{+}$ and $V_{-}$ as the set of all vertices of $G$ that are assigned $1$ and $-1$ under $f$, respectively. We consider $[V_{-},V_{+}]$ as the set of edges having one end point in $V_{-}$ and the other in $V_{+}$.
\section{Regular graphs}
\ \ Favaron \cite{f} and Wang \cite{wa} proved that for any $r$-regular graph $G$,
\begin{equation}\label{EQ1}
\gamma_{s}(G)\leq\left \{
\begin{array}{lll}
(\frac{r+1}{r+3})n & \mbox{if} & r\equiv0\ {mod\ 2} \vspace{1.5mm}\\
(\frac{(r+1)^{2}}{r^{2}+4r-1})n & \mbox{if} & r\equiv1\ (mod\ 2)
\end{array}
\right.(\cite {f})
\end{equation}
and
\begin{equation}\label{EQ2}
\alpha_{s}^{2}(G)\geq\left \{
\begin{array}{lll}
(\frac{-r^{2}+r+2}{r^{2}+r+2})n & \mbox{if} & r\equiv0\ (mod\ 2) \vspace{1.5mm}\\
(\frac{1-r}{1+r})n & \mbox{if} & r\equiv1\ (mod\ 2).
\end{array}
\right.(\cite{wa})
\end{equation}
Moreover, they showed that these bounds are sharp. Also, the following sharp lower and upper bounds on $\gamma_{s}(G)$ and $\alpha_{s}^{2}(G)$ of an $r$-regular graph $G$ were given in \cite{dthhs,hw}, and \cite{z}, respectively.
\begin{equation}\label{EQ3}
\gamma_{s}(G)\geq\left \{
\begin{array}{lll}
\frac{n}{r+1} & \mbox{if} & r\equiv0\ (mod\ 2) \vspace{1.5mm}\\
\frac{2n}{r+1} & \mbox{if} & r\equiv1\ (mod\ 2)
\end{array}
\right.(\cite {dthhs,hw})
\end{equation}
and
\begin{equation}\label{EQ4}
\alpha_{s}^{2}(G)\leq\left \{
\begin{array}{lll}
\frac{n}{r+1} & \mbox{if} & r\equiv0\ (mod\ 2) \vspace{1.5mm}\\
0 & \mbox{if} & r\equiv1\ (mod\ 2).
\end{array}
\right.(\cite {z})
\end{equation}\label{EQ5}
We are now in a position to exhibit the main theorem of this section.
\begin{theorem}\label{T2.1}
Let $G$ be an $r$-regular graph of order $n$. Then\vspace{1mm}\\
\emph{(i)}\ $\gamma^{NN}_{s}(G)=n-2L_{\lfloor\frac{r+1}{2}\rfloor}(G)$.\vspace{1mm}\\
\emph{(ii)}\ $\gamma^{NN}_{s}(G)=2\gamma_{\times\lceil\frac{r+1}{2}\rceil}(G)-n$.
\end{theorem}
\begin{proof}
(i) Let $B$ be a maximum $\lfloor\frac{r+1}{2}\rfloor$-limited packing in $G$. We define $f:V(G)\rightarrow\{-1,1\}$ by
$$f(v)=\left \{
\begin{array}{lll}
-1 & \mbox{if} & v\in B \\
\ 1 & \mbox{if} & v\in V(G)\setminus B.
\end{array}
\right.$$
Then $f(N[v])=|N[v]\cap(V(G)\setminus B)|-|N[v]\cap B|=|N[v]|-2|N[v]\cap B|\geq r+1-2\lfloor\frac{r+1}{2}\rfloor\geq0$, for all $v\in V(G)$. Therefore, $f$ is an NNSDF of $G$. So,
\begin{equation*}
\gamma^{NN}_{s}(G)\leq f(V(G))=n-2L_{\lfloor\frac{r+1}{2}\rfloor}(G).
\end{equation*}
If $f$ is a minimum NNSDF, then $|N[v]\cap V_{-}|\leq \lfloor\frac{r+1}{2}\rfloor$, for all $v\in V(G)$. Thus, $V_{-}$ is a $\lfloor\frac{r+1}{2}\rfloor$-limited packing in $G$. Therefore,
\begin{equation*}
(n-\gamma^{NN}_{s}(G))/2=|V_{-}|\leq L_{\lfloor\frac{r+1}{2}\rfloor}(G).
\end{equation*}
So, the first equality holds.\\
(ii) Suppose that $D$ is a minimum $\lceil\frac{r+1}{2}\rceil$-tuple dominating set in $G$. We define $f:V(G)\rightarrow\{-1,1\}$ by
$$f(v)=\left \{
\begin{array}{lll}
\ 1 & \mbox{if} & v\in D \\
-1 & \mbox{if} & v\in V(G)\setminus D.
\end{array}
\right.$$
Then $f(N[v])=|N[v]\cap D|-|N[v]\cap(V(G)\setminus D)|=2|N[v]\cap D|-|N[v]|\geq 2\lceil\frac{r+1}{2}\rceil-r-1\geq0$, for each vertex $v$. So, $f$ is an NNSDF of $G$. This shows that
\begin{equation*}
\gamma^{NN}_{s}(G)\leq f(V(G))=2\gamma_{\times\lceil\frac{r+1}{2}\rceil}(G)-n.
\end{equation*}
Let $f$ be a minimum NNSDF of $G$. Then $|N[v]\cap V_{+}|\geq \lceil\frac{r+1}{2}\rceil$, for each vertex $v$. Hence $V_{+}$ is a $\lceil\frac{r+1}{2}\rceil$-tuple dominating set in $G$. It follows that
\begin{equation*}
(n+\gamma^{NN}_{s}(G))/2=|V_{+}|\geq \gamma_{\times\lceil\frac{r+1}{2}\rceil}(G).
\end{equation*}
This completes the proof of (ii).
\end{proof}
If $f$ is an NNSDF of $G$, then $f(N[v])\geq1$ for each vertex $v$ of even degree. Hence,
\begin{equation}
\gamma_s(G)=\gamma^{NN}_s(G)
\end{equation}
when $G$ is an $r$-regular graph and $r$ is even. Similar to Part (ii) of Theorem \ref{T2.1}, an analogous equality for $\alpha_{s}^{2}(G)$ can be proved as follows:
$$\alpha_{s}^{2}(G)=n-2\gamma_{\times\lceil\frac{r}{2}\rceil}(G).$$
Therefore,
\begin{equation}\label{EQ6}
\gamma^{NN}_s(G)=-\alpha_{s}^{2}(G)
\end{equation}
when $G$ is an $r$-regular graph and $r$ is odd.\\
By Theorem \ref{T2.1} and the inequalities (1)--(\ref{EQ6}), we conclude the following theorem.
\begin{theorem}\label{T2.2}
For any $r$-regular graph $G$ of order $n$,
$$\left(\frac{1}{r+1}\right)n\leq \gamma^{NN}_{s}(G)
\leq\left(\frac{r+1}{r+3}\right)n,\ \ \ r\equiv0\ (mod\ 2)$$
and
$$0\leq \gamma^{NN}_{s}(G)\leq\left(\frac{r-1}{r+1}\right)n,\ \ \ r\equiv1\ (mod\ 2).$$
Furthermore, these bounds are sharp.
\end{theorem}
Balister et al. \cite{bbg} proved that if $G$ is a cubic graph of order $n$, then $L_{2}(G)\geq n/3$. Taking into account this fact and using the first part of Theorem \ref{T2.1}, the upper bound $n/2$ given in Theorem \ref{T2.2} for a cubic graph $G$ of order $n$ can be improved as follows.
\begin{theorem}\label{T2.3}
If $G$ is a cubic graph of order $n$, then $\gamma^{NN}_{s}(G)\leq n/3$.
\end{theorem}
The upper bound in Theorem \ref{T2.3} is the best possible. To see this fact, let $G_{6}$ be the graph depicted in the following figure. It is easy to see that $\gamma^{NN}_{s}(G_{6})=2$. By taking multiple copies of $G_{6}$, we have infinite collection of cubic graphs $G$ with $\gamma^{NN}_{s}(G)=|V(G)|/3$.
\begin{picture}(269.518,188.518)(0,0)
\put(131,175){\circle*{6}}
\put(160,175){\circle*{6}}
\put(116,157){\circle*{6}}
\put(175,157){\circle*{6}}
\put(131,139){\circle*{6}}
\put(160,139){\circle*{6}}
\multiput(133,176)(.037,0){670}{\line(2,0){.9}}
\multiput(131,139)(.037,0){670}{\line(2,0){.9}}
\multiput(158,177)(.027,-.029){670}{\line(2,0){.9}}
\multiput(115,157)(.024,.031){670}{\line(2,0){.9}}
\multiput(175,157)(-.024,-.029){670}{\line(2,0){.9}}
\multiput(131,139)(-.024,.028){670}{\line(2,0){.9}}
\multiput(115,157)(.084,0){670}{\line(2,0){.9}}
\multiput(131,139)(.043,.056){670}{\line(2,0){.9}}
\multiput(160,139)(-.044,.053){670}{\line(2,0){.9}}
\end{picture}\vspace{-52mm}\\
\begin{center}
The graph $G_{6}$.
\end{center}
\section{($r+1$)-clique-free graphs}
\ \ We need the following well-known theorem of Tur\'{a}n from the extremal graph theory.
\begin{lemma}\label{L1}
\emph{(\cite{T}, Tur\'{a}n's Theorem)}. If $G$ is an \emph{(}$r+1$\emph{)}-clique-free graph of order $n$, then
$$|E(G)|\le \frac{r-1}{2r}\cdot n^2,$$
with equality if and only if $G$ is the Tur\'{a}n graph $T_{n,r}$ and $r$ divides $n$.
\end{lemma}
The following lower bound was exhibited in \cite{as} for the NNSDN of a bipartite graph $G$ of order $n$.
\begin{equation}\label{EQ9}
\gamma^{NN}_{s}(G)\geq2(-1+\sqrt{1+2n})-n.
\end{equation}
The above inequality is not true as it stands. It is easy to see that the family
$$\Sigma=\{K_{p,p}\circ \overline{K_{p+1}}\ |\ p\geq1\}$$
serves as an infinite family of counterexamples to (\ref{EQ9}) (see Figure $1$).
In what follows we bound $\gamma^{NN}_s(G)$ from below for all ($r+1$)-clique-free graphs ($r\geq2$). Moreover, as the special case $r=2$ of such graphs, the given lower bound on $\gamma^{NN}_s(G)$ in the next theorem can be considered for a bipartite graph $G$ instead of (\ref{EQ9}).
\begin{theorem}
Let $r\geq2$ be an integer. If $G$ is an \emph{(}$r+1$\emph{)}-clique-free graph of order $n$, then
$$\gamma^{NN}_{s}(G)\geq-\frac{2r}{r-1}+\frac{2}{r-1}\sqrt{r^{2}+r(r-1)n}-n.$$
Furthermore, the equality holds if and only if $G=H\circ \overline{K_{(r-1)p+1}}$ in which $H$ is a complete $r$-partite graph with $p$ vertices in each partite set.
\end{theorem}
\begin{proof}
Let $f$ be a minimum NNSDF of $G$. Then, each vertex $v$ in $V_{-}$ has at least one neighbor in $V_{+}$. Also, $|N(v)\cap V_{-}|\leq|N(v)\cap V_{+}|+1$ for all $v\in V_{+}$. Moreover, by lemma \ref{L1} we have \begin{equation}\label{EQ1}
\begin{array}{lcl}
|V_{-}|&\leq& |[V_{-},V_{+}]|=\sum_{v\in V_{+}}|N(v)\cap V_{-}|\leq \sum_{v\in V_{+}}(|N(v)\cap V_{+}|+1)\\
&=&2|E(G[V_{+}])|+|V_{+}|\leq(r-1)|V_{+}|^2/r+|V_{+}|.
\end{array}
\end{equation}
This follows that
$$(r-1)|V_{+}|^2/r+2|V_{+}|-n\geq0.$$
Solving the above inequality for $|V_{+}|$ we obtain
$$(n+\gamma^{NN}_{s}(G))/2=|V_{+}|\geq\frac{r}{r-1}(-1+\sqrt{1+(r-1)n/r}).$$
This implies the desired lower bound.
Let $G=H\circ \overline{K_{(r-1)p+1}}$. Then, the function $g$ assigning $1$ to the vertices in $V(H)$ and $-1$ to the other vertices defines an NNSDF with weight $-r^{2}p^{2}+rp^{2}$. Therefore, $\gamma^{NN}_s(G)\leq-r^{2}p^{2}+rp^{2}$. On the other hand, since $n=rp+rp((r-1)p+1))$, we have
$$\gamma^{NN}_{s}(G)\geq-\frac{2r}{r-1}+\frac{2}{r-1}\sqrt{r^{2}+r(r-1)n}-n=-r^{2}p^{2}+rp^{2}.$$
Therefore, $\gamma^{NN}_{s}(G)=-r^{2}p^{2}+rp^{2}$.
Conversely, suppose that the equality holds. Then $|V_{-}|=|[V_{-},V_{+}]|$,
$|N(v)\cap V_{-}|=|N(v)\cap V_{+}|+1$ for all
$v\in V_{+}$, and $|E(G[V_{+}])|=(r-1)|V_{+}|^2/2r$, by (\ref{EQ1}).
This implies that every vertex in $V_{-}$ has exactly one neighbor in $V_{+}$ (and therefore the vertices in
$V_{-}$ are independent) and every vertex $v$ in $V_{+}$ has exactly
$|N(v)\cap V_{+}|+1$ neighbors in $V_{-}$. Moreover, $G[V_{+}]$ is the Tur\'{a}n
graph $T_{|V_{+}|,r}$ in which $r\mid|V_{+}|$. This implies that each partite set of it has the cardinality $|V_{+}|/r$.
Thus, $G=G[V_{+}]\circ\overline{K_{(r-1)|V_{+}|/r+1}}$. This completes the proof.
\end{proof}\vspace{23mm}
\begin{picture}(269.518,188.518)(0,0)
\put(102,179){\circle*{6}}
\put(147,179){\circle*{6}}
\put(102,211){\circle*{6}}
\put(147,211){\circle*{6}}
\put(192,179){\circle*{6}}
\put(192,211){\circle*{6}}
\put(216,225){\circle*{6}}
\put(208,232){\circle*{6}}
\put(200,238){\circle*{6}}
\put(192,244){\circle*{6}}
\put(165,244){\circle*{6}}
\put(154,244){\circle*{6}}
\put(142,244){\circle*{6}}
\put(131,244){\circle*{6}}
\put(165,147){\circle*{6}}
\put(154,147){\circle*{6}}
\put(142,147){\circle*{6}}
\put(131,147){\circle*{6}}
\put(216,159){\circle*{6}}
\put(208,154){\circle*{6}}
\put(200,150){\circle*{6}}
\put(192,146){\circle*{6}}
\put(78,224){\circle*{6}}
\put(84,231){\circle*{6}}
\put(91,237){\circle*{6}}
\put(100,243){\circle*{6}}
\put(99,148){\circle*{6}}
\put(91,152){\circle*{6}}
\put(84,156){\circle*{6}}
\put(75,160){\circle*{6}}
\multiput(147,211)(.024,.048){670}{\line(2,0){.9}}
\multiput(147,211)(-.024,.047){670}{\line(2,0){.9}}
\multiput(147,211)(.009,.048){670}{\line(2,0){.9}}
\multiput(147,211)(-.008,.048){670}{\line(2,0){.9}}
\multiput(102,179)(-.004,-.049){670}{\line(2,0){.9}}
\multiput(102,179)(-.04,-.026){670}{\line(2,0){.9}}
\multiput(102,179)(-.02,-.043){670}{\line(2,0){.9}}
\multiput(102,179)(-.027,-.03){670}{\line(2,0){.9}}
\multiput(147,179)(-.025,-.049){670}{\line(2,0){.9}}
\multiput(147,179)(.025,-.049){670}{\line(2,0){.9}}
\multiput(147,179)(-.009,-.049){670}{\line(2,0){.9}}
\multiput(147,179)(.01,-.049){670}{\line(2,0){.9}}
\multiput(192,211)(.037,.022){670}{\line(2,0){.9}}
\multiput(192,211)(-.002,.052){670}{\line(2,0){.9}}
\multiput(192,211)(.025,.033){670}{\line(2,0){.9}}
\multiput(192,211)(.011,.04){670}{\line(2,0){.9}}
\multiput(102,211)(-.005,.052){670}{\line(2,0){.9}}
\multiput(102,211)(-.04,.021){670}{\line(2,0){.9}}
\multiput(102,211)(-.019,.04){670}{\line(2,0){.9}}
\multiput(102,211)(-.029,.03){670}{\line(2,0){.9}}
\multiput(191,179)(-.002,-.052){670}{\line(2,0){.9}}
\multiput(191,179)(.04,-.032){670}{\line(2,0){.9}}
\multiput(191,179)(.012,-.042){670}{\line(2,0){.9}}
\multiput(191,179)(.026,-.04){670}{\line(2,0){.9}}
\multiput(191,179)(0,.052){670}{\line(2,0){.9}}
\multiput(191,179)(-.069,.048){670}{\line(2,0){.9}}
\multiput(191,179)(-.14,.05){670}{\line(2,0){.9}}
\multiput(102,179)(0,.052){670}{\line(2,0){.9}}
\multiput(102,179)(.07,.052){670}{\line(2,0){.9}}
\multiput(102,179)(.142,.052){670}{\line(2,0){.9}}
\multiput(147,179)(-.076,.052){670}{\line(2,0){.9}}
\multiput(147,179)(-.001,.052){670}{\line(2,0){.9}}
\multiput(147,179)(.074,.052){670}{\line(2,0){.9}}
\put(105,171){$1$}
\put(152,174){$1$}
\put(105,211){$1$}
\put(151,209){$1$}
\put(184,171){$1$}
\put(185,213){$1$}
\put(217,227){$-1$}
\put(209,234){$-1$}
\put(201,240){$-1$}
\put(193,246){$-1$}
\put(154,248){$-1$}
\put(143,248){$-1$}
\put(131,248){$-1$}
\put(120,248){$-1$}
\put(156,137){$-1$}
\put(144,137){$-1$}
\put(132,137){$-1$}
\put(121,137){$-1$}
\put(218,151){$-1$}
\put(206,145){$-1$}
\put(197,140){$-1$}
\put(185,137){$-1$}
\put(62,225){$-1$}
\put(68,232){$-1$}
\put(77,240){$-1$}
\put(86,246){$-1$}
\put(89,137){$-1$}
\put(81,142){$-1$}
\put(72,146){$-1$}
\put(62,150){$-1$}
\end{picture}\vspace{-52mm}\\
\begin{center}
Figure $1$. A member of $\Sigma$ for $p=3$ with $\gamma^{NN}_{s}(G)=-18$.
\end{center}
\section{Trees}
\ \ It has been proved by Henning in \cite{h} that $\alpha_{s}^{2}(T)\geq0$, for any tree $T$. Hence, $\alpha_{s}^{2}(T)$ is bounded from below not depending on the order or any other parameters, for any tree $T$. A similar result cannot be presented for $\gamma^{NN}_{s}(T)$. In fact, the following observation shows that $\gamma^{NN}_{s}(T)$ is not bounded from both above and below.
\begin{Observation}
For any integer $k$, there is a tree $T$ with $\gamma^{NN}_{s}(T)=k$.
\end{Observation}
\begin{proof}
Let $k<0$. Consider $T=P_{|k|}\circ \overline{K_{2}}$. Assigning $-1$ to the leaves and $1$ to the support vertices gives a minimum NNSDF of $T$ with weight $\gamma^{NN}_{s}(T)=k$. On the other hand, it is easy to see that $\gamma^{NN}_{s}(K_{1,2t-1})=0$ for each positive integer $t$. Moreover, $\gamma^{NN}_{s}(P_{n})=n-2\lceil n/3\rceil$ (see \cite{hfx}) shows that $\gamma^{NN}_{s}(P_{3k})=k$, for each positive integer $k$.
\end{proof}
Let $\Theta$ be the collection of all trees $T$ with maximum degree $\Delta(T)$ formed from the star $K_{1,\Delta(T)}$,
with a central vertex $u$ of maximum degree, for which all vertices in $V(T)\setminus N[u]$ are leaves with their support
vertices in a set $S\subseteq N(u)$ with $|S|=\lfloor\frac{\Delta}{2}\rfloor$. Moreover, deg$(s)\leq3$ for all $s\in S$.
\begin{theorem}\label{T1}
For any graph $G$ of order $n$ with maximum degree $\Delta$, $\gamma^{NN}_{s}(G)\geq-n+2\lceil\frac{\Delta+1}{2}\rceil$. In particular, the equality holds for a tree $T$ if and only if $T\in \Theta$.
\end{theorem}
\begin{proof}
Let $f$ be a minimum NNSDF of $G$ and $u$ be a vertex of the maximum degree in $V(G)$. Since $f(N[u])\geq0$, at least $\lceil\frac{\Delta+1}{2}\rceil$ vertices in $N[u]$ belong to $V_{+}$. This implies that\\
$$\gamma^{NN}_{s}(G)=f(V(G))=-n+2|V_{+}|\geq-n+2\left\lceil\frac{\Delta+1}{2}\right\rceil.$$
Suppose that the equality holds for a tree $T$. This shows that $|V_{+}|=\lceil\frac{\Delta+1}{2}\rceil$ and therefore
$V_{+}\subseteq N[u]$. If $\Delta=1$, then $T=K_{2}$. Moreover, it is easy to see that $G$ is formed from $K_{1,2}$ by
adding at most one pendant edge to just one leaf of $K_{1,2}$ if $\Delta=2$. In each of the two cases $T\in \Theta$. So, we may assume that $\Delta\geq3$. Since $\lceil\frac{\Delta+1}{2}\rceil<\Delta$, there exists a vertex $w$ in $N(u)$ with $f(w)=-1$. Therefore, $f(u)=1$ and the other $\lceil\frac{\Delta+1}{2}\rceil-1=\lfloor\frac{\Delta}{2}\rfloor$ vertices of $V_{+}$ appears in $N(u)$. Since $T$ is a tree, the subset $N(u)$ is independent and every vertex in $V(G)\setminus N[u]$ has at most one neighbor in $N(u)$. On the other hand, the condition $f(N[v])\geq0$ for each vertex $v$, implies that every vertex not in $N[u]$ has exactly one neighbor in $S=N(u)\cap V_{+}$, with $|S|=\lfloor\frac{\Delta}{2}\rfloor$. Furthermore, every vertex in $S$ has at most two neighbors except $u$ and all vertices in $V(G)\setminus N[u]$ are leaves with their support vertices in $S$. Hence, $T\in \Theta$.
Now let $T\in \Theta$ and $u$ be a vertex of maximum degree. It is easy to see that the function $f$ assigning 1 to all vertices of $S\cup\{u\}$ and $-1$ to the other vertices defines an NNSDF of $T$ with weight $-n+2\lceil\frac{\Delta+1}{2}\rceil$. So, $\gamma^{NN}_{s}(T)\leq-n+2\lceil\frac{\Delta+1}{2}\rceil$. This completes the proof.
\end{proof}
Note that the lower bound given in Theorem \ref{T1} is sharp not only for all trees in $\Theta$ but for some other collections of graphs, for example the complete graphs.
For a tree $T$, let $L(T)$ and $S(T)$ be the set of leaves and support vertives, respectively. For any support vertex $v$ of $T$ consider $L_{v}$ as the set of all leaves adjacent to $v$ and $\ell_{v}=|L_{v}|$.\\
For characterizing all trees attaining the next bound we introduce $\Omega$ to be the collection of all trees $T$ satisfying:\vspace{.35mm}\\
(i) $T$ is a star with even number of leaves,\vspace{.4mm}\\
or\vspace{.4mm}\\
(ii) $T$ has no support vertex with even number of leaves and $S(T)$ is a dominating set in $T$ in which everey vertex has
at most one neighbor in $V(T)\setminus L(T)$.
\begin{theorem}
Let $T$ be a tree of order $n\geq3$ with $\ell$ leaves and $s'$ be the number of support vertices with odd number of leaves. Then, $\gamma^{NN}_{s}(T)\leq n-\ell+s'$. Moreover, the equality holds if and only if $T\in \Omega$.
\end{theorem}
\begin{proof}
Let $f$ be a minimum NNSDF of $T$ and $u$ be a support vertex. Then there exists a vertex $u'$ in $N[u]$ with $f(u')=-1$, for otherwise the function $f'$ assigning $-1$ to a leaf of the support vertex $u$ and $f'(x)=f(x)$ to the other vertices $x$ is an NNSDF with the weight $f'(V(T))<f(V(T))$, a contradiction. Also, without loss of generality, we may assume that $L_{u}\cap V_{-}$ is not empty. Otherwise the function $p$ assigning $-1$ to a leaf of $u$, $1$ to $u'$ and $p(x)=f(x)$ to the other vertices $x$ would be an NNSDF with $p(V(T))=\gamma^{NN}_{s}(T)$ (if $f(u)=-1$, then we consider $u'$ as $u$). Thus, we may always assume that $L_{u}\cap V_{-}\neq \emptyset$ and $f(u)=1$, for all support vertices $u$ of $T$.
Suppose that there exists a support vertex $v$ which is adjacent to at most $\lceil\frac{\ell_{v}}{2}\rceil-1$ vertices in $V_{-}$. Then, $f(N[v])\geq|N[v]|-2(\lceil \ell_{v}/2\rceil-1)\geq2$. Let $v'$ be a leaf adjacent to $v$ with $f(v')=1$. Then, it is easy to see that the function defined by $g(v')=-1$ and $g(x)=f(x)$ for each $x\in V(T)\setminus\{v'\}$ is an NNSDF of $T$ with weight $g(V(T))<f(V(T))$, a contradiction. Therefore, every support vertex $v$ is adjacent to at least $\lceil\frac{\ell_{v}}{2}\rceil$ neighbors in $V_{-}$. Moreover, without loss of generality, we may assume that these $\lceil\frac{\ell_{v}}{2}\rceil$ neighbors belong to $L_{v}$.
Let $S(T)=\{v_{1},...,v_{s}\}$ and $S'=\{v_{1},...,v_{s'}\}$ be the set of support vertices with odd number of leaves.
Then\\
$\gamma^{NN}_{s}(T)=n-2|V_{-}|\leq n-2\left(\lceil\frac{\ell_{v_{1}}}{2}\rceil+...
+\lceil\frac{\ell_{v_{s'}}}{2}\rceil+\lceil\frac{\ell_{v_{s'+1}}}{2}\rceil+...+\lceil\frac{\ell_{v_{s}}}{2}\rceil\right)$
\begin{center}$=n-2\left(\frac{\ell_{v_{1}}+1}{2}+...+\frac{\ell_{v_{s'}}+1}{2}+\frac{\ell_{v_{s'+1}}}{2}+...
+\frac{\ell_{v_{s}}}{2}\right)=n-(\ell+s').$\end{center}
Let $\gamma^{NN}_s(T)=n-\ell-s'$ and $f$ be a minimum NNSDF of $T$ which assigns $-1$ to exactly $\lceil\frac{\ell_{v}}{2}\rceil$ leaves for each support vertex $v$ and $1$ to other vertices. Assume that $T$ is not a star with even number of leaves. We prove that $T$ satisfies (ii). If $T$ has a support vertex $v$ with even number of leaves, then the function $g$ assigning $-1$ to exactly $\lceil\frac{\ell_{v}}{2}\rceil+1$ leaves of $v$, $1$ to its other leaves and $g(x)=f(x)$ to the other vertices $x$ would be an NNSDF of $T$ with weight $g(V(T))<f(V(T))$ contradicting the fact that $f(V(T))=\gamma^{NN}_s(T)$. Therefore, $T$ has no support vertex with even number of leaves. We now show that $S(T)$ is a dominating set in $T$. If a vertex $w$ in $V(T)\setminus(S(T)\cup L(T))$ has no neighbor in $S(T)$, then $k(w)=-1$ and $k(x)=f(x)$ for $x\in V(T)\setminus\{w\}$ would be an NNSDF with $k(V(T))< f(V(T))$. This contradiction implies that $S(T)$ is a dominating set. Suppose to the contrary that there exists a vertex $u\in S(T)$ such that $|N(u)\cap(V(T)\setminus L(T))|\geq2$. Then the function $h$ assigning $-1$ to exactly $\lceil\frac{\ell_{u}}{2}\rceil+1$ leaves of $u$, $1$ to its other leaves and $h(x)=f(x)$ for $x\in V(T)\setminus L_{u}$ defines an NNSDF of $T$ with weight $h(V(T))<f(V(T))$, a contradiction. The above argument shows that $T$ satisfies (ii).
Suppose that $T\in \Omega$ and $f$ is a minimum NNSDF of $T$ which assigns $1$ to $v$ and $-1$ to at least $\lceil\frac{\ell_{v}}{2}\rceil$ leaves of each support vertex $v$. Both (i) and (ii) shows that $f$ assigns $-1$ to exactly $\lceil\frac{\ell_{v}}{2}\rceil$ leaves of each support vertex $v$. Now let $f(u)=-1$ for some vertex $u$ in $V(T)\setminus(S(T)\cup L(T))$. Since $S(T)$ dominates $V(T)$, there exists a vertex $v\in S(T)$ adjacent to $u$. Then (ii) implies that $N(v)\cap(V(T)\setminus(S(T)\cup L(T)))=\{u\}$ and hence $f(N[v])\leq-1$, a contradiction. Therefore, $f$ assigns $1$ to all vertices in $V(T)\setminus(S(T)\cup L(T))$. Thus, $\gamma^{NN}_s(T)=n-\ell-s'$. This completes the proof.
\end{proof}
\section{Concluding remarks}
\ \ As applications of the concepts of limited packing and tuple domination we exhibited sharp lower and upper bounds
on NNSDN of regular graphs. We made use of Tur\'{a}n's Theorem for bounding the NNSDN of $(r+1)$-clique-free
graphs from below, and bounded this parameter for trees and characterized all graphs (trees) attainig the bounds.
It is worth giving exact formulas or bounds for this parameter of some other certain families of graphs. For example,
grids, nearly regular graphs, claw-free graphs, etc. We now conclude the paper with the following problem:\vspace{0.5mm}\\
\textbf{Problem.} How can we classify the other families of graphs by NNSDN?
|
\section*{Videos, Datasets, and Code}
Videos, data, and code to reproduce our results are available at: \url{https://github.com/idsia-robotics/proximity-quadrotor-learning}.
\section{Introduction}
\label{sec:introduction}
Robot control systems are traditionally structured in two distinct modules: perception and control. Perception processes the robot’s input in order to derive a high-level state, which represents meaningful and relevant information for the task the robot needs to solve. State information is then used by a controller, which determines the low-level control signals to be provided to the hardware.
For mobile robots operating in real-world unstructured environments, perception is often challenging. This is especially true when this step involves the interpretation of complex, high-dimensional data such as images. Many recent successful systems deal with this problem by adopting supervised machine learning (ML) techniques which operate on sensing data as input.
When designing such a system, we face the choice between at least two approaches.
\begin{itemize}
\item A mediated approach, in which one trains a supervised ML model to predict the high-level state given the inputs; then, control signals are derived from the state using a designed controller (or another learned model).
\item An end-to-end approach, in which a supervised ML model is trained to directly predict the control signals from the sensing data, without passing through an intermediate high-level representation of the state.
\end{itemize}
Which architecture is preferable? In some situations, one might not have a choice: for example, if one can not (or does not want to) collect ground truth information about the high-level state, training a perception model for a mediated approach is not possible (and this fact motivates many end-to-end systems). However, sometimes datasets annotated with such a ground truth can be acquired, potentially at a cost. Is it worth it?
There are several advantages in a mediated approach: 1) in many mobile robotics applications, hand-designing a controller, given a high-level state representation, is feasible and gives the designer explicit control on the resulting robot behavior; 2) a mediated approach is more transparent and may be easier to inspect and debug: given an unexpected robot behavior, the designer can inspect the high-level state to determine whether the problem is in the perception or controller.
In contrast, end-to-end approaches are appealing because they are conceptually very simple and can potentially be more computationally efficient, especially if the high-level state representation is complex and high-dimensional. Moreover, as mentioned above, end-to-end approaches don’t depend on high-level state's ground truth for training.
The considerations above disregard one key issue, i.e., the difficulty of learning models for the two approaches. Is learning a perception model (which outputs high-level state) easier than learning an end-to-end model (which directly outputs control signals)? Does one of the two models require a larger amount of training data to reach the same performance?
This is a key issue to consider when designing ML-based robot controllers: we investigate this question for one specific task. After reviewing related literature (Section~\ref{sec:rw}), we model and formalize the end-to-end architecture and two variants of the mediated architecture (Section~\ref{sec:model:general}); we instantiate these three architectures for one specific task, that we consider in the remainder of the paper, i.e., controlling a quadrotor in order to hover in front of a person who is freely moving (Section~\ref{sec:model}); such behavior could be implemented by a quadrotor tasked to monitor a person or expecting commands from them. The \textbf{main contribution} of the paper is a set of experiments showing that, for this specific task, training models for the three architectures has the same difficulty (the setup and results for these experiments are described in Sections ~\ref{sec:setup} and \ref{sec:results} respectively). The \textbf{main limitation} is that we limit our study to a single reactive control task; nonetheless, it is a challenging, real-world task which has several characteristics in common with other important subproblems in mobile robotics.
A \textbf{secondary contribution} of this paper is the design, implementation and validation of the drone control system described above, which constitutes a useful component in applications involving proximal interaction of humans and quadrotors; collected datasets, source code and training models are available for download.
\subsection{Related Work}
\label{sec:rw}
End-to-end learning approaches map raw sensor data (lidar, camera images, ego-motion sensors) to control actions: either as direct low-level control outputs~\cite{LeCun2005}, or as target points (e.g., desired velocity) for optimal low-level controllers~\cite{Kaufmann2018}.
Supervised end-to-end approaches have been successfully applied to a variety of challenging control scenarios: off-road obstacle avoidance~\cite{LeCun2005, Bajracharya2008}; autonomous driving~\cite{Chen2017,Xu2017,Chi2017,Heylen2018,bojarski2016end}; vision-based manipulation~\cite{Levine2016}; quadrotor control in forested environments~\cite{Ross2013,Giusti2016,Smolyanskiy2017}, cities~\cite{Loquercio2018,Kumaar2018}, and cluttered environments~\cite{Kaufmann2018}.
These can be considered instances of imitation learning (also known as learning-from-demonstration): in fact, one trains a policy to solve a sequential task using only demonstrations by an expert (represented as state/action pairs), as training data~\cite{attia2018global}. In end-to-end approaches, the state coincides with the robot sensing; this yields reactive controllers unless recurrent models are used to capture temporal dynamics~\cite{Chi2017}.
The ground truth used for learning may have various origins: skilled drivers or operators~\cite{Ross2013,Kumaar2018,LeCun2005,Bajracharya2008}; people walking~\cite{Giusti2016,Smolyanskiy2017} or driving a vehicle that is not the target robot~\cite{Loquercio2018}; random controllers that sometimes lead to collisions, which the model learns to avoid~\cite{Gandhi2017}; hand-designed controllers~\cite{Kaufmann2018};
controllers learned through reinforcement learning~\cite{Levine2016}; or the future position of vehicles in a large driving dataset~\cite{Xu2017}.
Most of these approaches are stateless and reactive, but applications of Deep Recurrent Neural Networks allows to capture temporal dynamics in end-to-end approaches~\cite{Chi2017}.
Compared to mediated approaches, end-to-end learning has been found in some cases to be slower to converge~\cite{Glasmachers2017,Shalev2017} and to require more training samples~\cite{Shalev2016}. Direct perception methods~\cite{Chen2015} are a further alternative for learning higher-level representations of the environment than mediated methods.
The task we consider in this paper is aimed to proximity human-robot interaction~\cite{Ng2011,Naseer2013}: a drone should be able to fly at an appropriate distance in front of people, following them~\cite{Jahidul2018} while waiting for possible command gestures~\cite{Peshkova2017}. The use of visual markers~\cite{Vasconcelos2016} simplifies this task but, in general, more sophisticated techniques, like Tracking-Learning-Detection~\cite{Kalal2012}, are needed~\cite{Bartal2015}. For instance, several deep learning approaches have been proposed to estimate the 6D pose of a person's head~\cite{Patacchiola2017} or body~\cite{Bogo2016} from monocular cameras (as well of general objects~\cite{Xiang2017}). Adopting such perception modules would be a reasonable alternative to solve our task; however, in this paper the task acts as a model of a larger class of tasks.
\section{Task and Model}
\label{sec:model}
\newcommand{\mathbf{x}}{\mathbf{x}}
\newcommand{\mathbf{x}_\text{im}}{\mathbf{x}_\text{im}}
\newcommand{\mathbf{x}_\text{odom}}{\mathbf{x}_\text{odom}}
\newcommand{\mathbf{u}}{\mathbf{u}}
\newcommand{u_{\text{a}x}}{u_{\text{a}x}}
\newcommand{u_{\text{a}y}}{u_{\text{a}y}}
\newcommand{u_{\text{v}z}}{u_{\text{v}z}}
\newcommand{u_{\omega z}}{u_{\omega z}}
\newcommand{\mathbf{s}}{\mathbf{s}}
\newcommand{\mathbf{s}_\text{pose}}{\mathbf{s}_\text{pose}}
\newcommand{\mathbf{s}_\text{odom}}{\mathbf{s}_\text{odom}}
\newcommand{s_x}{s_x}
\newcommand{s_y}{s_y}
\newcommand{s_z}{s_z}
\newcommand{s_\theta}{s_\theta}
\newcommand{C}{C}
\newcommand{C_\text{acq}}{C_\text{acq}}
We consider one specific task: controlling a quadrotor to stay at a fixed distance ($\Delta =\SI{1.5}{m}$) and at eye-level height in front of a user who is free to move in an environment.
The available inputs are the video feed from a forward-pointing camera $\mathbf{x}_\text{im}$ and the current linear velocity of the quadrotor $\mathbf{x}_\text{odom} = (v_x, v_y)$, obtained through ego-motion sensors;
the controller outputs $\mathbf{u}$ are composed by:
\begin{itemize}
\item the desired pitch and roll of the drone, which map to the acceleration along the drone's $x$ and $y$ axes ($u_{\text{a}x}$, $u_{\text{a}y}$ respectively);
\item the desired velocity along the $z$ axis ($u_{\text{v}z}$);
\item the desired angular velocity around the $z$ axis ($u_{\omega z}$).
\end{itemize}
The state $\mathbf{s}$ for the given task can be compactly represented using two pieces of information $\mathbf{s} = (\mathbf{s}_\text{pose}, \mathbf{s}_\text{odom})$.
\begin{itemize}
\item A 3D transformation $\mathbf{s}_\text{pose}$, representing the pose (location and orientation) of the user's head with respect to the quadrotor; more specifically, only the heading component of the head's orientation is relevant to the task (in fact, tilting the head or looking up or down should not affect the drone behavior). Therefore, we represent with angle $s_\theta$, $-\pi \leq s_\theta \leq \pi$, the relative angle between the user's and drone's orientation, such that $s_\theta = 0$ if the user's face points along the negative $x$ axis of the drone (i.e., the user is in front of the drone and faces towards it). Therefore, $\mathbf{s}_\text{pose} = (s_x, s_y, s_z, s_\theta)$.
\item In addition, the state also contains the quadrotor's current measured velocity, which is available directly as an input $\mathbf{s}_\text{odom} = \mathbf{x}_\text{odom}$.
\end{itemize}
In this specific task, we can also design a controller $C$ that produces control signals given the state: $\mathbf{u} = f_{C}(\mathbf{s})$, as detailed in Section~\ref{sec:controller}. Note that for other tasks (e.g. grasping), designing a controller might be very complex.
\subsection{General Approaches}
\label{sec:model:general}
We now describe the three approaches that we will compare in the following sections (see Figure~\ref{fig:approaches}). Even though for clarity we use the task-specific notation introduced above, these approaches can be easily generalized to other control tasks; then, $\mathbf{x}_\text{im}$ corresponds to the subset of the inputs that we want to learn how to interpret in order to obtain high-level state information $\mathbf{s}_\text{pose}$. $\mathbf{x}_\text{odom}$ represents inputs that we can (directly or through some processing step) turn into meaningful high-level state information $\mathbf{s}_\text{odom}$. For example, in an autonomous driving scenario: $\mathbf{x}_\text{im}$ could correspond to lidar data; $\mathbf{s}_\text{pose}$ to the relative position of other cars on the road and their velocities; $\mathbf{s}_\text{odom}$ to the car odometry and to the car's position on the lane (obtained through some existing modules using inputs from $\mathbf{x}_\text{odom}$); $\mathbf{u}$ to steering and acceleration/brake controls.
In \textbf{approach A1 (mediated)}, we learn a model M1 mapping $\mathbf{x}_\text{im}$ to $\mathbf{s}_\text{pose}$. Let $f_\text{M1}$ denote the function implemented by M1; $\mathbf{s}_\text{pose}^\text{A1} = f_\text{M1}(\mathbf{x}_\text{im})$ denotes the estimate of $\mathbf{s}_\text{pose}$. $\mathbf{s}_\text{pose}^\text{A1}$ is then joined to the odometry (available directly in the input) to form $\mathbf{s}^\text{A1} = (\mathbf{s}_\text{pose}^\text{A1}, \mathbf{s}_\text{odom})$. The control signals are then computed as $\mathbf{u}^\text{A1} = f_{C}(\mathbf{s}^\text{A1})$. In order to train M1, we require ground truth information on $\mathbf{s}_\text{pose}$. Moreover, this approach requires the availability of a controller $C$.
In \textbf{approach A2 (end-to-end)}, we directly learn a function $f_\text{M2}$ to predict $\mathbf{u}$ from $\mathbf{x} = (\mathbf{x}_\text{im}, \mathbf{x}_\text{odom})$. We denote the result as
$\mathbf{u}^\text{A2} = f_\text{M2}(\mathbf{x})$. Training M2 requires that a ground truth for $\mathbf{u}$ is available. Such a ground truth can be obtained in two ways. \emph{Learning strategy 1}: acquiring the ground truth for $\mathbf{u}$ directly (e.g., recording a skilled human pilot); \emph{Learning strategy 2}: obtaining the ground truth for $\mathbf{u}$ through a designed controller $C$, which is given ground truth $\mathbf{s}_\text{pose}$ information.
In \textbf{approach A3 (mediated with learned controller)} we learn a model M1 just like in approach A1. This model obtains a state estimate $\mathbf{s}^\text{A3} = \mathbf{s}^\text{A1}$. However, instead of using an hand-designed controller $C$ to produce the control signals from the estimated state, we learn a mapping $f_\text{M3}$ from $\mathbf{s}$ to $\mathbf{u}$. The control signals are then computed as $\mathbf{u}^\text{A3} = f_\text{M3}(\mathbf{s}^\text{A3})$. In order to train M1, we require ground truth information on $\mathbf{s}_\text{pose}$. Moreover, in order to train M3 we need a ground truth on $\mathbf{u}$, e.g. from a human controller. Differently from A1, this approach does not require one to design the controller $C$.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{img/approaches-complete.png}
\caption{Yellow background: learned models. Ground truth data used for training them is depicted with a dotted arrow. Requirements (in terms of ground truth or of a designed controller) are drawn with a green background.}
\label{fig:approaches}
\end{figure}
\begin{comment}
\begin{table}[]
\centering
\caption{Approach requirements}
\label{tab:req}
\begin{tabular}{@{}llll@{}}
\toprule
Approach & State labels & Controller $C$ & Output labels \\ \midrule
A1 & Required & Required & Not required \\
A2 strategy 1 & Not required & Not required & Required \\
A2 strategy 2 & Required & Required & Not required \\
A3 & Required & Not required & Required \\ \bottomrule
\end{tabular}
\end{table}
\end{comment}
\begin{figure*}[t]
\centering \vspace{2mm}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{img/0.png}
\label{fig:0_test}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{img/3.png}
\label{fig:3_test}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{img/9.png}
\label{fig:9_test}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{img/6.png}
\label{fig:6_test}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{img/11.png}
\label{fig:11_test}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{img/12.png}
\label{fig:12_test}
\end{subfigure}
\caption{Instances from the dataset. Each shows ground truth (green triangle) and outputs (blue rectangles) of A1, A2, A3 for each control variable: $u_{\text{a}x}$ (left), $u_{\text{a}y}$ (bottom), $u_{\text{v}z}$ (right), $u_{\omega z}$ (top). More details and videos in supplementary material. Note that some images in the dataset are affected by acquisition, compression or transmission artifacts
}
\label{fig:dataset}
\end{figure*}
\section{Experimental Setup}
\label{sec:setup}
\subsection{Hardware and Infrastructure}
All experiments take place in a
\SI[product-units = single]{10 x 10}{\metre} room fitted with an optical motion capture system. Within the room, a flying area of \SI[product-units = single]{7 x 7}{\metre} is defined and fenced by a black lightweight net.
We use a Parrot Bebop 2 quadrotor; the quadrotor is networked via wifi to a server, and controlled through a ROS interface~\cite{bebop_autonomy}, which exposes at \SI{5}{\hertz} the velocity of the drone $\mathbf{x}_\text{odom}$ computed from the visual odometry of a bottom-looking camera and, at \SI{30}{\hertz}, the front-facing camera feed. The drone uses a sonar and an IMU to estimate altitude and attitude; an onboard low-level controller accepts inputs $\mathbf{u}$ and updates the motors speed.
The quadrotor is outfitted with a mocap target, so that its exact 6D pose is acquired in real time. Moreover, we outfit the user's head with another mocap target; this is implemented by having the user wear one of two objects on their head, on top of which the mocap target is fixed: either a black baseball hat, or a thin elastic band. While the former is clearly visible when the user is seen from the front, the latter is almost invisible.
\subsection{Controller}
\label{sec:setup:controller}
We implement a simple baseline controller $C$ as a stateless function $f_{C}$. From the estimation of the person's pose, we compute a target point $\bar{p} = (s_x, s_y, s_z) + \Delta e(\pi+s_\theta)$ in front them, at distance $\Delta$ along unit vector $e(\pi+s_\theta)$. Then, we compute a desired velocity $\bar{v} =\left[ \frac{\bar p}{2 \tau} \right]^{v_{\max}}_{-v_{\max}}$ to reach $\bar p$ in $2 \tau$ time, limiting the components magnitude to $v_{\max}$. Finally, we compute a control output to reach velocity $\bar{v}$ and rotate towards the user in $\tau$ time:
\label{sec:controller}
\begin{align*}
(u_{\text{a}x}, u_{\text{a}x}) &=\left[ \left(\left(\bar{v}_x, \bar{v}_y \right) - \left(v_x, v_y \right)\right)/\tau \right]^{a_{\max}}_{-a_{\max}}\\
u_{\text{v}z} &= \bar{v}_z \\
u_{\omega z} &=\left[ \theta(\bar p) / \tau \right]^{\omega_{\max}}_{-\omega_{\max}},
\end{align*}
where $\theta(\bar p)$ is the azimuth of vector $\bar p$.
Parameters are fixed to $ \tau = \SI{0.5}{\second},\, v_{\max} = \SI{1.5}{\metre \per \second},\, a_{\max} = \SI{1.0}{\metre \per \square \second},\, \omega_{\max} = \SI{2.0}{\radian \per \second}$.
\subsection{Dataset Acquisition}
\label{sec:setup:dataset}
In order to ease the acquisition of the dataset, the drone follows a controller $C_\text{acq}$ that, given the pose of the user and the pose of the quadrotor provided by the mocap system, generates control signals that keep the quadrotor at a parametrized distance (that we adjusted between 1.0 and 2.0 meters during the acquisitions) in front of the user's head. Such a controller is similar of the one described in Section~\ref{sec:controller}
The outputs of $C_\text{acq}$ are not part of the dataset: they are only used to facilitate the acquisition of the sessions and make for an engaging experience for our test subjects.
We recorded 15 different sessions, each with a different user (age 23 to 38, different ethnicity, height ranging from 160 to 197 cm, variable levels of physical fitness, different clothing styles, hairstyles and colors). During the acquisition, users are instructed to move around the room freely: the quadrotor will then fly to stay in front of their head. After an initial period with cautious motion, users start challenging the controller and moving rather aggressively (videos of recording sessions are available as supplementary material), so that the quadrotor struggles to keep up; this ensures that many different relative positions for the head and the quadrotor are represented in the data. In some recordings, the users wear the hat, while in others the headband (which is almost not visible in the camera frames). The room is equipped with 4 distinct light sources (2 overhead, 2 movable spotlights placed in different corners of the room), which we toggle and move at different times of the recording to add variability to the scene. During some of the recordings, more people are in the flying area; moreover, often the quadrotor camera also sees people and different background objects standing outside of the flying area, such as computers, screens, desks, and windows. While acquiring the data, we take care that, if more than one person is visible in the frame, the person wearing the mocap target is the one closest to the camera. On occasion, the video link is temporarily corrupted and some frames are acquired with visual artifacts; we purposefully do not remove such frames from the recording as most of them can be still understood by an human observer. On average each session is $3$ minutes long, totaling more than $45$ minutes of flight.
All sessions are recorded in ROS bagfiles, from which we extract 79k dataset instances at a rate of 30Hz. Each instance contains: the video frame $\mathbf{x}_\text{im}$, the pose of the user head relative to the drone $\mathbf{s}_\text{pose}$, and the drone velocity $\mathbf{x}_\text{odom} = \mathbf{s}_\text{odom} = (v_x, v_y)$. The controller $C$ is then applied to $(\mathbf{s}_\text{pose},\mathbf{s}_\text{odom})$ to obtain the ground truth value $\mathbf{u}$ (containing the four control variables). Note that the ground truth control signal does not necessarily correspond to the control signal that the drone was receiving during the acquisition, which came from a different controller.
We use all data from three sessions (16k instances) as a test set to quantitatively evaluate the performance of the three approaches. We randomly split the 63k instances from the remaining 12 sessions into a training set (50k instances) and a validation set (13k instances).
\subsection{Machine Learning Models}
\label{sec:setup:ML}
\subsubsection{Model M1}
Model M1 is implemented as a deep neural network using a ResNet~\cite{resnet} architecture, which has been recently adopted for similar quadrotor control tasks~\cite{Loquercio2018}.
The network accepts as input a $108 \times 60 \times 3$ RGB image, and produces the 4 components of $\mathbf{s}$. The internal architecture is a ResNet-8~\cite{resnet} followed by two dense layers with 256 and 128 neurons, respectively.
\subsubsection{Model M2}
Model M2 has a similar architecture as M1, but additionally accepts $\mathbf{x}_\text{odom}$ through two additional input neurons that skip the ResNet-8 layers. These neurons, concatenated to the ResNet-8 output, are input to the two dense layers with 256 and 128 neurons. The outputs correspond to the 4 components of $\mathbf{u}$.
\subsubsection{Model M3}
Model M3 is implemented through a simple multilayer perceptron that maps 6 input values ($\mathbf{s}^\text{A3} = \mathbf{s}^\text{A1} = (\mathbf{s}_\text{pose}^\text{A1}, \mathbf{s}_\text{odom})$) to the 4 components of $\mathbf{u}$; it contains 2 hidden layers, with 256 and 128 neurons respectively.
\subsubsection{Training}
We train the three models with the same setup: 1) we use Mean Absolute Error as loss function and ADAM~\cite{adam} as optimizer with a learning rate of 0.001; 2) we speed up learning by reducing the learning rate when the validation loss plateaus for more than 5 epochs; 3) we use early stopping (10 epochs of patience on the validation loss, with a maximum of 200 epochs).
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{img/quantitative2.png}
\caption{Quantitative results for the four control variables $\mathbf{u}$: estimation quality versus number of training samples.}
\label{fig:quantitative}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{img/y_data.png}%
\caption{We report each of the four control variables on a separate row. Left: the time series of the ground truth control (green thick line) and of the outputs of the ML approaches (thin lines), for a 16 second segment extracted from the test set, in which the user was moving very aggressively. Right: the scatter plot comparing, on the y axis, the output of A1 (left), A2 (center), and A3 (right) to the ground truth (x axis), for the whole testing set; a perfect model would yield all points on the diagonal.}
\label{fig:outputs}
\end{figure*}
\section{Experimental Results}
\label{sec:results}
We report three sets of experiments. First, we quantitatively evaluate the prediction quality of different approaches for all instances in the testing set, versus the ground truth for $\mathbf{u}$, also evaluating the impact of the training set size; second, we compare the trajectories of a quadrotor when it is controlled by an ideal controller, with the trajectories obtained by A1, A2 and A3. Finally, we perform qualitative robustness tests on the system controlled by each of the three approaches.
\subsection{Quantitative results on testing instances}
For each component of $\mathbf{u}^\text{A1}$, $\mathbf{u}^\text{A2}$ and $\mathbf{u}^\text{A3}$ we compute the coefficient of determination $R^2$~\cite{rsquared} of the estimate.
This measure corresponds to the proportion of variance in $\mathbf{u}$ that is explained by the model: a perfect estimator yields $R^2 = 1$; a dummy estimator that always returns the mean of the variable to be estimated yields $R^2 = 0$; even worse estimators yield $R^2 < 0$. $R^2$ allows us to compare the quality of our estimate for different components of $\mathbf{u}$ even though each has a different variance.
We compute these metrics for the three approaches trained on different amounts of training data; in particular, we are interested to compare how hard it is for the different approaches to achieve a given performance. For a given training set size $T$, we randomly sample $T$ out of the 50k training instances, which we use to train the models for each of A1, A2 and A3. Validation and testing datasets in all cases remain the same. To account for the variability due the sampling of the training set (which is very large for low values of $T$), we repeat each experiment for up to 50 replicas.
\begin{comment}
\begin{table}\centering
\caption{Training size $T$ and number of replicas.}
\label{table:replicas}
\begin{tabular}{@{}lrrrrrrrr@{}}
\toprule
$T$ & 128 & 512 & 1k & 2k & 5k & 10k & 20k & 50k \\
Replicas & 50 & 20 & 10 & 5 & 2 & 2 & 2 & 2 \\ \bottomrule
\end{tabular}
\end{table}
\end{comment}
Figure~\ref{fig:quantitative} reports the $R^2$ values resulting from the experiments described above. We observe that prediction quality increases with $T$, but plateaus after $T=10000$; there is no clear difference in prediction performance among A1, A2 and A3.
$u_{\omega z}$ and $u_{\text{v}z}$ yield $R^2$ coefficients of $0.88$ and $0.82$, respectively. Prediction quality is significantly worse for $u_{\text{a}x}$ and especially for $u_{\text{a}y}$, with $R^2$ values of $0.59$ and $0.57$, respectively. This is easily explained: $u_{\omega z}$ is mostly dependent on the position of the body in the frame, which is easy to perceive in $\mathbf{x}_\text{im}$: $u_{\omega z}$ is learned significantly better than a dummy regressor, with as few as 128 training instances.
$u_{\text{v}z}$ mostly depends on the vertical position of the head in the frame, which is also easy to perceive in $\mathbf{x}_\text{im}$, but has less variability in the datasets. Predicting $u_{\text{a}x}$ and $u_{\text{a}y}$ is harder: the former relies on an accurate perception of the distance of the user (which is confounded by their height and body size), and the latter on an estimate of the relative orientation of the head ($s_\theta$), which is arguably hard to get on low resolution inputs.
Note that prediction quality is not necessarily related to the quality of the resulting robot behavior. For example, a model which always yields a very tiny but systematic overestimate on $u_{\text{a}x}$ yields an $R^2$ value close to 1, but would cause the quadrotor to crash on the user in a very short time. On the contrary, a model that produces predictions affected by large amounts of uncorrelated noise with zero mean could yield acceptable behaviors (especially with short control timesteps) but very disappointing quantitative metrics. To provide a better idea of the usability of the predictions for control, Figure~\ref{fig:outputs} illustrates the outputs of the three approaches (trained with $T=50000$) on a short segment extracted from a recording belonging to the testing set.
We observe that the predicted signals closely track the ground truth; outputs for $u_{\text{a}x}$ and $u_{\text{a}y}$ exhibit a larger amount of high-frequency noise, which appears non-systematic.
\begin{figure}[t]
\centering
\vspace{1mm}
\includegraphics[height=6cm]{img/girl1.pdf}
\caption{Trajectories followed by the quadrotor starting in the red pose while approaching a person standing still in different poses (black circles), from left to right: initially facing the quadrotor at \SI{90}{\degree}, \SI{45}{\degree}, and \SI{0}{\degree}. Trajectories represent one run and are colored by the controller in use.}
\label{fig:traj_controllers}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[height=4.7cm]{img/girl2A1.pdf}
\includegraphics[height=4.7cm]{img/girl2A2.pdf}
\caption{Five runs for each controller (left: A1, right: A2) towards a person initially facing the quadrotor at \SI{90}{\degree} and a single run with the ground truth controller (green).}
\label{fig:traj}
\end{figure}
\subsection{Analysis of flying performance}
Figure~\ref{fig:traj_controllers} compares the trajectories resulting from the ground truth controller, with those resulting from A1, A2, and A3. The quadrotor approaches a user (not in the dataset), who is standing still, from different relative poses: the final pose of the robot always faces the user, and is reached in about 5 seconds, after which the robot stabilizes in a short time.
In Figure~\ref{fig:traj}, the quadrotor's approach towards a different user (also not in the dataset) was run 5 times for A1 and A2. This is the only case in our tests where we could notice a (small) difference in the behavior of different approaches, with A2 being somewhat smoother and closer to the ground truth trajectory.
Supplementary videos report more details on this experiment and extensive tests where we challenge the system robustness by having multiple people in the frame, quick movements, distracting objects, sudden and extreme lighting variations. During the tests, we cycled the control between the three approaches, and found that they behaved indistinguishably from each other; in all cases the drone behavior was predictable and perceived as safe by the users.
\section{Conclusions}
\label{sec:conclusions}
We considered the task of controlling a quadrotor to hover in front of a freely moving user using input data from an onboard camera. On this task, we compared mediated approaches to end-to-end approaches; we found equivalent quantitative performance, learning difficulty, perceived quality of the robot behaviors and robustness to challenging inputs; in only one occasion we measured repeatably different trajectories, which were slightly smoother in the end-to-end approach, but correct in all cases and perceived as very similar.
\balance
\bibliographystyle{IEEEtran}
|
\section{Constructing monogamous measures of entanglement from other monogamous measures}
For any entanglement monotone $E$, and any monotonically increasing
function $g:\mbb{R}_{+}\to\mbb{R}_+$ with the property that $g(x)=0$ iff $x=0$, denote
\begin{equation}
E_g( \rho^{AB}) \equiv\min\sum_{j}p_jg[E(|\psi_j\rangle\langle\psi_j|^{AB})],
\end{equation}
where the minimum is taken over all pure state decompositions of
$\rho^{AB}=\sum_{j=1}^{n}p_j|\psi_j\rangle\langle\psi_j|^{AB}$.
Observe that if in addition $g$ is convex then we must have
\begin{equation}
E_g( \rho^{AB}) \geq g[E( \rho^{AB})].
\end{equation}
Therefore, if $E_g$ is an entanglement measure and is
monogamous on pure tripartite states then $E$ is also
monogamous on pure tripartite states. To see why,
note that if $E(\psi^{A|BC})=E(\rho^{AB})$ we also have
\begin{eqnarray}
&&E_g( |\psi\rangle^{A|BC}) =g[ E( |\psi\rangle^{A|BC})]\nonumber\\
&=&g[ E( \rho^{AB})]\leq E_g( \rho^{AB}).
\end{eqnarray}
But since $E_g$ is a measure of entanglement we
must have $E_g(|\psi\rangle^{A|BC})\geq E_g(\rho^{AB})$ so that
we get $E_g(|\psi\rangle^{A|BC})=E_g(\rho^{AB})$. Since we assume
here that $E_g$ is monogamous, thus $E_g(\rho^{AC})=0$,
which implies that $E(\rho^{AC})=0$.
As a simple example of this, consider the function
$g(x)=x^2$ and take $E=C$ be the concurrence as defined
in Ref.~\cite{Rungta2003pra}. Then, $E_g=C^2$ is the
tangle which is monogamous (it is given in terms of the
linear entropy, which is strictly concave).
Hence, the above analysis implies that the concurrence $C$ is also monogamous.
|
\section{INTRODUCTION}
\label{sec:Introduction}
Accurate, real-time, and robust perception of the environment is an indispensable component in autonomous driving systems. For perception in high-end autonomous vehicles, LiDAR (Light Detection And Ranging) sensors play an important role. LiDAR sensors can directly provide distance measurements, and their resolution and field of view exceed those of radar and ultrasonic sensors~\cite{moosmann2009segmentation}. LiDAR sensors are robust under almost all lighting conditions: day or night, with or without glare and shadows~\cite{wu2017squeezeseg}. As such, LiDAR-based perception has attracted significant research attention.
Recently, deep learning has been shown to be very effective for LiDAR perception tasks. Specifically, Wu et al. proposed SqueezeSeg \cite{wu2017squeezeseg}, which focuses on the problem of point-cloud segmentation. SqueezeSeg projects a 3D LiDAR point cloud onto a spherical surface, and uses a 2D CNN to predict point-wise labels for the point cloud. SqueezeSeg is extremely efficient -- the fastest version achieves an inference speed of over 100 frames per second. However, SqueezeSeg still has several limitations: first, its accuracy still needs to be improved to be practically useful. One important reason for accuracy degradation is \textit{dropout noise} -- missing points from the sensed point cloud caused by limited sensing range, mirror diffusion of the sensing laser, or jitter in incident angles. Such dropout noise can corrupt the output of SqueezeSeg's early layers, which reduces accuracy. Second, training deep learning models such as SqueezeSeg requires tens of thousands of labeled point clouds; however, collecting and annotating this data is even more time consuming and expensive than collecting comparable data from cameras. GTA-V is used to synthesize LiDAR point cloud as an extra source of training data~\cite{wu2017squeezeseg}; however, this approach suffers from the domain shift problem \cite{torralba2011unbiased} - models trained on synthetic data usually fail catastrophically on the real data, as shown in Fig.~\ref{fig:DomainShift}. Domain shift comes from different sources, but the absence of dropout noise and intensity signals in GTA-V are two important factors. Simulating realistic dropout noise and intensity is very difficult, as it requires sophisticated modeling of both the LiDAR device and the environment, both of which contain a lot of non-deterministic factors. As such, the LiDAR point clouds generated by GTA-V do not contain dropout noise and intensity signals. The comparison of simulated data and real data is shown in Fig.~\ref{fig:DomainShift} (a), (b).
\begin{figure}[!t]
\begin{center}
\centering \includegraphics[width=0.9\linewidth]{DomainShift.pdf}
\caption{An example of \emph{domain shift}. The point clouds are projected onto a spherical surface for visualization (car in red, pedestrian in blue). Our domain adaptation pipeline improves the segmentation from (c) to (d) while trained on synthetic data.}
\vspace{-0.9cm}
\label{fig:DomainShift}
\end{center}
\end{figure}
In this paper, we focus on addressing the challenges above. First, to improve the accuracy, we mitigate the impact of dropout noise by proposing the Context Aggregation Module (CAM), a novel CNN module that aggregates contextual information from a larger receptive field and improves the robustness of the network to dropout noise. Adding CAM to the early layers of SqueezeSegV2 not only significantly improves its performance when trained on real data, but also effectively reduces the domain gap, boosting the network's real-world test accuracy when trained on synthetic data. In addition to CAM, we adopt several improvements to SqueezeSeg, including using focal loss \cite{lin2018focal}, batch normalization \cite{ioffe2015batch}, and LiDAR mask as an input channel. These improvements together boosted the accuracy of SqueezeSegV2 by 6.0\% - 8.6\% in all categories on the converted KITTI dataset \cite{wu2017squeezeseg}.
Second, to better utilize synthetic data for training the model, we propose a domain adaptation training pipeline that contains the following steps: first, before training, we render intensity channels in synthetic data through \textit{learned intensity rendering}. We train a neural network that takes the point coordinates as input, and predicts intensity values. This rendering network can be trained in a "self-supervised" fashion on unlabeled real data. After training the network, we feed the synthetic data into the network and render the intensity channel, which is absent from the original simulation. Second, we use the synthetic data augmented with rendered intensity to train the network. Meanwhile, we follow \cite{morerio2018minimal} and use \textit{geodesic correlation alignment} to align the batch statistics between real data and synthetic data. 3) After training, we propose \textit{progressive domain calibration} to further reduce the gap between the target domain and the trained network. Experiments show that the above domain-adaptation training pipeline significantly improves the accuracy of the model trained with synthetic data from 29.0\% to 57.4\% on the real world test data.
The contributions of this paper are threefold: 1) We improve the model structure of SqueezeSeg with CAM to increase its robustness to dropout noise, which leads to significant accuracy improvements of 6.0\% to 8.6\% for different categories. We name the new model SqueezeSegV2. 2) We propose a domain-adaptation training pipeline that significantly reduces the distribution gap between synthetic data and real data. Model trained on synthetic data achieves 28.4\% accuracy improvement on the real test data. 3) We create a large-scale 3D LiDAR point cloud dataset, GTA-LiDAR, which consists of 100,000 samples of synthetic point cloud augmented with rendered intensity. The source code and dataset will be open-sourced.
\section{RELATED WORK}
\label{sec:RelatedWork}
\textbf{3D LiDAR Point Cloud Segmentation} aims to recognize objects from point clouds by predicting point-wise labels. Non-deep-learning methods~\cite{moosmann2009segmentation, douillard2011segmentation, zermas2017fast} usually involve several stages such as ground removal, clustering, and classification. SqueezeSeg~\cite{wu2017squeezeseg} is one early work that applies deep learning to this problem. Piewak et al.~\cite{piewak2018boosting} adopted a similar problem formulation and pipeline to SqueezeSeg and proposed a new network architecture called LiLaNet. They created a dataset by utilizing image-based semantic segmentation to generate labels for the LiDAR point cloud. However, the dataset was not released, so we were not able to conduct a direct comparison to their work. Another category of methods is based on PointNet~\cite{qi2017pointnet, qi2017pointnet++}, which treats a point cloud as an unordered set of 3D points. This is effective with 3D perception problems such as classification and segmentation. Limited by its computational complexity; however, PointNet is mainly used to process indoor scenes where the number of points is limited. Frustum-PointNet~\cite{qi2017frustum} is proposed for out-door object detection, but it relies on image object detection to first locate object clusters and feeds the cluster, instead of the whole point cloud, to the PointNet.
\textbf{Unsupervised Domain Adaptation (UDA)} aims to adapt the models from one labeled source domain to another unlabeled target domain. Recent UDA methods have focused on transferring deep neural network representations~\cite{patel2015visual,csurka2017domain}. Typically, deep UDA methods employ a conjoined architecture with two streams to represent the models for the source and target domains, respectively. In addition to the task related loss computed from the labeled source data, deep UDA models are usually trained jointly with another loss, such as a discrepancy loss~\cite{long2015learning,sun2017correlation,zhuo2017deep,zhang2017curriculum,morerio2018minimal}, adversarial loss~\cite{liu2016coupled,ganin2016domain,tzeng2017adversarial,shrivastava2017learning,bousmalis2017unsupervised,hoffman2018cycada}, label distribution loss~\cite{zhang2017curriculum} or reconstruction loss~\cite{ghifary2015domain,ghifary2016deep}.
The most relevant work is the exploration of synthetic data~\cite{shrivastava2017learning,zhang2017curriculum,hoffman2018cycada}. By enforcing a self-regularization loss, Shrivastava et al.~\cite{shrivastava2017learning} proposed SimGAN to improve the realism of synthetic data using unlabeled real data. Another category of relevant work employs a discrepancy loss~\cite{long2015learning,sun2017correlation,zhuo2017deep,morerio2018minimal}, which explicitly measures the discrepancy between the source and target domains on corresponding activation layers of the two network streams. Instead of working on 2D images, we try to adapt synthetic 3D LiDAR point clouds by a novel adaptation pipeline.
\textbf{Simulation} has recently been used for creating large-scale ground truth data for training purposes. Richter et al.~\cite{playingfordata} provided a method to extract semantic segmentation for the synthesized in-game images. In \cite{drivinginthematrix}, the same game engine is used to extract ground truth 2D bounding boxes for objects in the image. Yue et al.~\cite{yue2018lidar} proposed a framework to generate synthetic LiDAR point clouds. Richter et al.~\cite{playingforbenchmarks} and Kr\"ahenb\"uhl~\cite{philip2018free} extracted more types of information from video games.
\begin{figure*}[!t]
\begin{center}
\centering \includegraphics[width=0.65\linewidth]{SqueezeSegV2Framework.pdf}
\caption{Network structure of the proposed SqueezeSegV2 model for road-object segmentation from 3D LiDAR point clouds.}
\label{fig:SqueezeSegV2Framework}
\end{center}
\end{figure*}
\section{Improving the model structure}
\label{sec:Better}
We propose SqueezeSegV2, by improving upon the base SqueezeSeg model, adding Context Aggregation Module (CAM), adding LiDAR mask as an input channel, using batch normalization \cite{ioffe2015batch}, and employing the focal loss \cite{lin2018focal}. The network structure of SqueezeSegV2 is shown in Fig.~\ref{fig:SqueezeSegV2Framework}.
\subsection{Context Aggregation Module}
\label{ssec:CAM}
LiDAR point cloud data contains many missing points, which we refer to as dropout noise, as shown in Fig.~\ref{fig:DomainShift}(b). Dropout noise is mainly caused by 1) limited sensor range, 2) mirror reflection (instead of diffusion reflection) of sensing lasers on smooth surfaces, and 3) jitter of the incident angle. Dropout noise has a significant impact on SqueezeSeg, especially in early layers of a network. At early layers where the receptive field of the convolution filter is very small, missing points in a small neighborhood can corrupt the output of the filter significantly. To illustrate this, we conduct a simple numerical experiment, where we randomly sample an input tensor and feed it into a $3\times3$ convolution filter. We randomly drop out some pixels from the input tensor, and as shown in Fig.~\ref{fig:CAM_Comparison}, as we increase the dropout probability, the difference between the errors of the corrupted output and the original output also increases.
\begin{figure}[!t]
\begin{center}
\centering \includegraphics[width=1.0\linewidth]{CAM.pdf}
\caption{Structure of Context Aggregation Module.}
\vspace{-0.6cm}
\label{fig:CAM}
\end{center}
\end{figure}
This problem not only impacts SqueezeSeg when trained on real data, but also leads to a serious domain gap between synthetic data and real data, since simulating realistic dropout noise from the same distribution is very difficult.
To solve this problem, we propose a novel Context Aggregation Module (CAM) to reduce the sensitivity to dropout noise. As shown in Fig.~\ref{fig:CAM}, CAM starts with a max pooling with a relatively large kernel size. The max pooling aggregates contextual information around a pixel with a much larger receptive field, and it is less sensitive to missing data within its receptive field. Also, max pooling can be computed efficiently even with a large kernel size. The max pooling layer is then followed by two cascaded convolution layers with a ReLU activation in between. Following~\cite{hu2018squeeze}, we use the \textit{sigmoid} function to normalize the output of the module and use an element-wise multiplication to combine the output with the input. As shown in Fig.~\ref{fig:CAM_Comparison}, the proposed module is much less sensitive to dropout noise -- with the same corrupted input data, the error is significantly reduced.
In SqueezeSegV2, we insert CAM after the output of the first three modules (1 convolution layer and 2 FireModules), where the receptive fields of the filters are small. As can be seen in later experiments, CAM 1) significantly improves the accuracy when trained on real data, and 2) significantly reduces the domain gap while trained on synthetic data and testing on real data.
\begin{figure}[!t]
\begin{center}
\centering \includegraphics[width=0.8\linewidth]{CAM_Comparison.pdf}
\caption{We feed a random tensor to a convolutional filter, one with CAM before a $3\times3$ convolution filter and the other one without CAM. We randomly add dropout noise to the input, and measure the output errors. As we increase the dropout probability, the error also increases. For all dropout probabilities, adding CAM improve the robustness towards the dropout noise and therefore, the error is always smaller.}
\vspace{-0.5cm}
\label{fig:CAM_Comparison}
\end{center}
\end{figure}
\subsection{Focal Loss}
\label{ssec:Focal}
LiDAR point clouds have a very imbalanced distribution of point categories -- there are many more background points than there are foreground objects such as cars, pedestrians, etc. This imbalanced distribution makes the model focus more on easy-to-classify background points which contribute no useful learning signals, with the foreground objects not being adequately addressed during training.
To address this problem, we replace the original cross entropy loss from SqueezeSeg~\cite{wu2017squeezeseg} with a focal loss~\cite{lin2018focal}. The focal loss modulates the loss contribution from different pixels and focuses on hard examples. For a given pixel label $t$, and the predicted probability of $p_t$, focal loss~\cite{lin2018focal} adds a modulating factor $(1-p_t)^{\gamma}$ to the cross entropy loss. The focal loss for that pixel is thus
\begin{equation}
FL(p_t) =-(1-p_t)^{\gamma}\log{(p_t)}
\end{equation}
When a pixel is mis-classified and $p_t$ is small, the
modulating factor is near 1 and the loss is unaffected. As
$p_t \rightarrow1$, the factor goes to 0, and the loss for well-classified
pixels is down-weighted. The focusing parameter $\gamma$ smoothly adjusts the rate at which well-classified examples are down-weighted.
When $\gamma = 0$, the Focal Loss is equivalent to the Cross Entropy Loss. As $\gamma$ increases, the effect of the modulating factor is likewise increased. We choose $\gamma$ to be $2$ in our experiments.
\subsection{Other Improvements}
\textbf{LiDAR Mask}: Besides the original (x, y, z, intensity, depth) channels, we add one more channel -- a binary mask indicating if each pixel is missing or existing. As we can see from Table~\ref{tab:SqueezeSegV2}, the addition of the mask channel significantly improves segmentation accuracy for cyclists.
\textbf{Batch Normalization}: Unlike SqueezeSeg~\cite{wu2017squeezeseg}, we also add batch normalization (BN)~\cite{ioffe2015batch} after every convolution layer. The BN layer is designed to alleviate the issue of internal covariate shift -- a common problem for training a deep neural network. We observe an improvement in car segmentation after using BN layers in Table~\ref{tab:SqueezeSegV2}.
\section{Domain Adaptation Training}
\label{sec:Stronger}
In this section, we introduce our unsupervised domain adaptation pipeline that trains SqueezeSegV2 on synthetic data and improves its performance on real data. We construct a large-scale 3D LiDAR point cloud dataset, GTA-LiDAR, with 100,000 LiDAR scans simulated on GTA-V. To deal with the \emph{domain shift} problem, we employ three strategies: learned intensity rendering, geodesic correlation alignment, and progressive domain calibration, as shown in Fig.~\ref{fig:DAFramework}.
\vspace{-0.1cm}
\begin{figure}[!t]
\begin{center}
\subfigure[Pre-training: Learned Intensity Rendering]{
\label{fig:LIR}
\includegraphics[width=1.\linewidth]{LIR.pdf}
}
\subfigure[Training: Geodesic Correlation Alignment]{
\label{fig:GCA}
\includegraphics[width=1.\linewidth]{GCA.pdf}
}
\subfigure[Post-training: Progressive Domain Calibration]{
\label{fig:PDC}
\includegraphics[width=1.\linewidth]{PDC.pdf}
}
\caption{Framework of the proposed unsupervised domain adaptation method for road-object segmentation from the synthetic GTA-LiDAR dataset to the real-world KITTI dataset.}
\vspace{-0.6cm}
\label{fig:DAFramework}
\end{center}
\end{figure}
\subsection{The GTA-LiDAR Dataset}
\label{ssec:GTA-LiDAR}
We synthesize 100,000 LiDAR point clouds in GTA-V to train SqueezeSegV2. We use the framework in \cite{philip2018free} to generate depth semantic segmentation maps, and use the method in \cite{yue2018lidar} to do Image-LiDAR registration in GTA-V. Following \cite{yue2018lidar}, we collect 100,000 point cloud scans by deploying a virtual car to drive autonomously in the virtual world. GTA-V provides a wide variety of scenes, car types, traffic conditions, etc., which ensures the diversity of our synthetic data. Each point in the synthetic point cloud contains one label, one distance and $x, y, z$ coordinates. However, it does not contain intensity, which represents the magnitude of the reflected laser signal. Also, the synthetic data does not contain dropout noise as in the real data. Because of such distribution discrepancies, the model trained on synthetic data fails to transfer to real data.
\subsection{Learned Intensity Rendering}
\label{ssec:Intensity}
The synthetic data only contains $x, y, z, depth$ channels and does not have intensity. As shown in SqueezeSeg~\cite{wu2017squeezeseg}, intensity is an important signal. The absence of intensity can lead to serious accuracy loss. Rendering realistic intensity is a non-trivial task, since a multitude of factors that affect intensity, such as surface materials and LiDAR sensitivity, are generally unknown to us.
\begin{figure*}[!t]
\begin{center}
\centering \includegraphics[width=1.0\linewidth]{figures/PredictedIntensity.pdf}
\caption{Rendered v.s. ground truth intensity in the KITTI dataset.}
\vspace{-10pt}
\label{fig:PredictedIntensity}
\end{center}
\end{figure*}
To solve this problem, we propose a method called \textit{learned intensity rendering}. The idea is to use a network to take the $x, y, z, depth$ channels of the point cloud as input, and predict the intensity. Such rendering network can be trained with unlabeled LiDAR data, which can be easily collected as long as a LiDAR sensor is available. As shown in Fig.~\ref{fig:LIR}, we train the rendering network in a self-supervision fashion, splitting the $x, y, z$ channels as input to the network and the intensity channel as the label. The structure of the rendering network is almost the same as SqueezeSeg, except that the CRF layer is removed.
The intensity rendering can be seen as a regression problem, where the $\ell_2$ loss is a natural choice. However, $\ell_2$ fails to capture the multi-modal distribution of the intensity -- given the same input of $x, y, z$, the intensity can differ. To model this property, we designed a hybrid loss function that involves both classification and regression. We divide the intensity into $n=10$ regions, with each region having a reference intensity value. The network first predicts which region the intensity belongs to. Once the region is selected, the network further predicts a deviation from the reference intensity. This way, the categorical prediction can capture the multi-modal distribution of the intensity, and the deviation prediction leads to more accurate estimations. We train the rendering network on the KITTI \cite{geiger2012we} dataset with the hybrid loss function and measure its accuracy with mean squared error (MSE). Compared to $\ell_2$ loss, the converged MSE drops significantly by 3X from 0.033 to 0.011. A few rendered results using two different losses are shown in Fig.~\ref{fig:PredictedIntensity}. After training the rendering network, we feed synthetic GTA-LiDAR data into the network to render point-wise intensities.
\subsection{Geodesic Correlation Alignment}
\label{ssec:LE}
After rendering intensity, we train SqueezeSegV2 on the synthetic data with focal loss. However, due to distribution discrepancies between synthetic data and real data, the trained model usually fails to generalize to real data.
To reduce this domain discrepancy, we adopt geodesic correlation alignment during training. As shown in Fig.~\ref{fig:GCA}, at every step of training, we feed in one batch of synthetic data and one batch of real data to the network. We compute the focal loss on the synthetic batch, where labels are available. Meanwhile, we compute the geodesic distance \cite{morerio2018minimal} between the output distributions of two batches. The total loss now contains both the focal loss and the geodesic loss. Where the focal loss focuses on training the network to learn semantics from the point cloud, the geodesic loss penalizes discrepancies between batch statistics from two domains. Note that other distances such as the Euclidean distance can also be used to align the domain statistics. However, we choose the geodesic distance over the Euclidean distance since it takes into account the manifold’s curvature. More details can be found in \cite{morerio2018minimal}.
We denote the input synthetic data as $X_{sim}$, synthetic labels as $Y_{sim}$ the input real data as $X_{real}$. Our loss function can be computed as
\begin{equation}
FL(X_{sim}, Y_{sim}) + \lambda \cdot GL(X_{sim}, X_{real}),
\end{equation}
where $FL$ denotes focal loss between the synthetic label and network prediction, $GL$ denotes the geodesic loss between batch statistics of synthetic and real data. $\lambda$ is a weight coefficient and we set it to 10 in our experiment. Note that in this step, we only require unlabeled real data, which is much easier to obtain than annotated data as long as a LiDAR sensor is available.
\begin{algorithm}[!t]
{
\small
\KwIn{Unlabeled real data $\mathcal{X}$, model $\mathcal{M}$}
$\mathcal{X}^{(0)} \leftarrow \mathcal{X}$ \\
\For{layer $l$ in the model $\mathcal{M}$}{
$\mathcal{X}^{(l)} \leftarrow \mathcal{M}^{(l)}(\mathcal{X}^{(l-1)})$\\
$\mu^{(l)} \leftarrow \mathbb{E} (\mathcal{X}^{(l)}$), $\sigma^{(l)} \leftarrow \sqrt{Var(\mathcal{X}^{(l)})}$ \\
Update the BatchNorm parameters of $\mathcal{M}^{(l)}$ \\
$\mathcal{X}^{(l)} \leftarrow (\mathcal{X}^{(l)} - \mu^{(l)}) / \sigma^{(l)}$ \\
}
\KwOut{Calibrated model $\mathcal{M}$}
}
\small\caption{Progressive Domain Calibration}
\label{Alg:MTSSRLearning}
\end{algorithm}
\subsection{Progressive Domain Calibration}
\label{ssec:PBA}
After training SqueezeSegV2 on synthetic data with geodesic correlation alignment, each layer of the network learns to recognize patterns from its input and extract higher level features. However, due to the non-linear nature of the network, each layer can only work well if its input is constrained within a certain range. Taking the ReLU function as an example, if somehow its input distribution shifts below 0, the output of the ReLU becomes all zero. Otherwise, if the input shifts towards larger than 0, the ReLU becomes a linear function. For deep learning models with multiple layers, distribution discrepancies from the input data can lead to distribution shift at the output of each layer, which is accumulated or even amplified across the network and eventually leads to a serious degradation of performance, as illustrated in Fig. \ref{fig:PDC}.
To address this problem, we employ a post training procedure called progressive domain calibration (PDC). The idea is to break the propagation of the distribution shift through each layer with progressive layer-wise calibration. For a network trained on synthetic data, we feed the real data into the network. Staring from the first layer, we compute its output statistics (mean and variance) under the given input, and then re-normalize the output's mean to be 0 and its standard deviation to be 1, as shown in Fig.~\ref{fig:PDC}. Meanwhile, we update the batch normalization parameters (mean and variance) of the layer with the new statistics. We progressively repeat this process for all layers of the network until the last layer. Similar to geodesic correlation alignment, this process only requires unlabeled real data, which is presumably abundant. This algorithm is summarized in Algorithm~\ref{Alg:MTSSRLearning}.
A similar idea was proposed in \cite{li2018adaptive}, but PDC is different since it performs calibration progressively, making sure that the calibrations of earlier layers do not impact those of later layers.
\section{Experiments}
\label{sec:Experiments}
In this section, we introduce the details of our experiments.
We train and test SqueezeSegV2 on a converted KITTI \cite{geiger2012we} dataset as \cite{wu2017squeezeseg}. To verify the generalization ability, we further train SqueezeSegV2 on the synthetic GTA-LiDAR dataset and test it on the real world KITTI dataset.
\begin{figure*}[!t]
\begin{center}
\centering \includegraphics[width=0.9\linewidth]{SqueezeSegV1V2.pdf}
\vspace{-4pt}
\caption{Segmentation result comparison between SqueezeSeg~\cite{wu2017squeezeseg} and our SqueezeSegV2 (red: car, green: cyclist). Note that in first row, SqueezeSegV2 produces much more accurate segmentation for the cyclist. In the second row, SqueezeSegV2 avoids a falsely detected car that is far away.}
\label{fig:SqueezeSegV1V2}
\vspace{-5pt}
\end{center}
\end{figure*}
\begin{figure*}[!t]
\begin{center}
\centering \includegraphics[width=0.9\linewidth]{DomainAdaptationResult.pdf}
\vspace{-6pt}
\caption{Segmentation result comparison before and after domain adaptation (red: car, blue: pedestrian).}
\label{fig:DomainAdaptationResult}
\vspace{-12pt}
\end{center}
\end{figure*}
\subsection{Experimental Settings}
\label{ssec:Settings}
We compare the proposed method with SqueezeSeg~\cite{wu2017squeezeseg}, one state-of-the-art model for semantic segmentation from 3D LiDAR point clouds. We use KITTI~\cite{geiger2012we} as the real world dataset. KITTI provides images, LiDAR scans, and 3D bounding boxes organized in sequences. Following~\cite{wu2017squeezeseg}, we obtain the point-wise labels from 3D bounding boxes, all points of which are considered part of the target object. In total, 10,848 samples with point-wise labels are collected. For SqueezeSegV2, the dataset is split into a training set with 8,057 samples and a testing set with 2,791 samples.
For domain adaptation, we train the model on GTA-LiDAR, and test it on KITTI for comparison.
Similar to~\cite{wu2017squeezeseg}, we evaluate our model's performance on class-level segmentation tasks by a point-wise comparison of the predicted results with ground-truth labels. We employ intersection-over-union (IoU) as our evaluation metric, which is defined as $IoU_c = \frac{|\mathcal{P}_c \cap \mathcal{G}_c|}{|\mathcal{P}_c \cup \mathcal{G}_c|}$, where $\mathcal{P}_c$ and $\mathcal{G}_c$ respectively denote the predicted and ground-truth point sets that belong to class-$c$. $|\cdot|$ denotes the cardinality of a set.
\begin{table}[!t]
\begin{center}
\caption{Segmentation performance (IoU, \%) comparison between the proposed SqueezeSegV2 (+BN+M+FL+CAM) model and state-of-the-art baselines on the KITTI dataset.
\begin{tabular}
{c | c c c c}
\hline
& Car & Pedestrian & Cyclist & Average\\
\hline
SqueezeSeg~\cite{wu2017squeezeseg} & 64.6 & 21.8 & 25.1 & 37.2 \\
+BN & 71.6 & 15.2 & 25.4 & 37.4 \\
+BN+M & 70.0 & 17.1 & 32.3 & 39.8\\
+BN+M+FL & 71.2 & 22.8 & 27.5 & 40.5 \\
+BN+M+FL+CAM & \textbf{73.2} & \textbf{27.8} & \textbf{33.6} & \textbf{44.9} \\
\hline
PointSeg \cite{wang2018pointseg} & 67.4 & 19.2 & 32.7 & 39.8 \\
\hline
\end{tabular}
\label{tab:SqueezeSegV2}
\end{center}
\textbf{+BN} denotes using batch normalization. \textbf{+M} denotes adding LiDAR mask as input. \textbf{+FL} denotes using focal loss. \textbf{+CAM} denotes using the CAM module.
\end{table}
\subsection{Improved Model Structure}
\label{ssec:Results_SqueezeSegV2}
The performance comparisons, measured in IoU, between the proposed SqueezeSegV2 model and baselines are shown in Table~\ref{tab:SqueezeSegV2}. Some segmentation results are shown in Fig.~\ref{fig:SqueezeSegV1V2}.
From the results, we have the following observations. (1) both batch normalization and the mask channel can produce better segmentation results - batch normalization boosts segmentation of cars, whereas the mask channel boosts segmentation of cyclists. (2) Focal loss improves segmentation of pedestrians and cyclists. The number of points corresponding to pedestrians and cyclists is low relative to the large number of background points. This class imbalance causes the network to focus less on the pedestrian and cyclist classes. Focal loss mitigates this problem by focusing the network on optimization of these two categories. (3) CAM significantly improves the performance of all the classes by reducing the network's sensitivity to dropout noise.
\subsection{Domain Adaptation Pipeline}
\label{ssec:Results_SqueezeSegUDA}
\begin{table}[!t]
\begin{center}
\small
\caption{Segmentation performance (IoU, \%) of the proposed domain adaptation pipeline from GTA-LiDAR to the KITTI.
\begin{tabular}
{c | c c }
\hline
& \multicolumn{1}{c}{Car} & \multicolumn{1}{c}{Pedestrian}\\
\hline
SQSG trained on GTA \cite{wu2017squeezeseg} & 29.0 & - \\
SQSG trained on GTA-LiDAR & 30.0 & 2.1 \\
+LIR & 42.0 & 16.7 \\
+LIR+GCA & 48.2 & 18.2\\
+LIR+GCA+PDC & 50.3 & 18.6 \\
+LIR+GCA+PDC+CAM & \textbf{57.4} & \textbf{23.5} \\
\hline
SQSG trained on KITTI w/o intensity \cite{wu2017squeezeseg} & 57.1 & - \\
\hline
\end{tabular}
\label{tab:DA_Result}
\end{center}
\textbf{SQSG} denotes SqueezeSeg. \textbf{+LIR} denotes using learned intensity rendering. \textbf{+GCA} denotes using geodesic correlation alignment. \textbf{+PDC} denotes using progressive domain calibration. \textbf{+CAM} denotes using the CAM module.
\vspace{-0.4cm}
\end{table}
The performance comparisons, measured in IoU, between the proposed domain adaptation pipeline and baselines are shown in Table~\ref{tab:DA_Result}. Some segmentation results are shown in Fig.~\ref{fig:DomainAdaptationResult}. From the results, we have the following observations. (1) Models trained on the source domain without any adaptation does not perform well. Due to the influence of \emph{domain discrepancy}, the joint probability distributions of observed LiDAR and road-objects greatly differ in the two domains. This results in the model's low transferability from the source domain to the target domain. (2) All adaptation methods are effective, with the combined pipeline performing the best, demonstrating its effectiveness. (3) Adding the CAM to the network also significantly boosts the performance on the real data, supporting our hypothesis that dropout noise is a significant source of domain discrepancy. Therefore, improving the network to make it more robust to dropout noise can help reduce the domain gap. (4) Compared with \cite{wu2017squeezeseg} where a SqueezeSeg model is trained on the real KITTI dataset but without intensity, our SqueezeSegV2 model trained purely on synthetic data and unlabeled real data achieves a better accuracy, showing the effectiveness of our domain adaptation training pipeline. (5) Compared with our latest SqueezeSegV2 model trained on the real KITTI dataset, there is still an obvious performance gap. Adapting the segmentation model from synthetic LiDAR point clouds is still a challenging problem.
\section{CONCLUSION}
\label{sec:Conclusion}
In this paper, we proposed SqueezeSegV2 with better segmentation performance than the original SqueezeSeg and a domain adaptation pipeline with stronger transferability. We designed a context aggregation module to mitigate the impact of dropout noise. Together with other improvements such as focal loss, batch normalization and a LiDAR mask channel, SqueezeSegV2 sees accuracy improvements of 6.0\% to 8.6\% in various pixel categories over the original SqueezeSeg. We also proposed a domain adaptation pipeline with three components: learned intensity rendering, geodesic correlation alignment, and progressive domain calibration. The proposed pipeline significantly improved the real world accuracy of the model trained on synthetic data by 28.4\%, even out-performing a baseline model \cite{wu2017squeezeseg} trained on the real dataset.
\section*{Acknowledgement}
This work is partially supported by Berkeley Deep Drive (BDD), and partially sponsored by individual gifts from Intel and Samsung. We would like to thank Alvin Wan and Ravi Krishna for their constructive feedback.
\addtolength{\textheight}{-8cm}
\newpage
\small\bibliographystyle{IEEEtran}
|
\section{The stability equation}\label{appA}
The non-zero elements in the stability equation \eqref{Stability} are given below. For simplicity, the derivative of a thermodynamic quantity with respect to $\rho_0$ at constant $T_0$ (and vice-versa) has been denoted as $\frac{\partial}{\partial \rho_0}$, instead of $\left.\frac{\partial}{\partial\rho_{0}}\right|_{T_{0}}$. The elements are,
\begin{equation}
\left.\begin{array}{l}
\mathscr{L}_{t}\left(1,1\right)=1\\
\mathscr{L}_{t}\left(2,1\right)=u_{0}\\
\mathscr{L}_{t}\left(2,2\right)=\mathscr{L}_{t}\left(3,3\right)=\mathscr{L}_{t}\left(4,4\right)=\rho_{0}\\
\mathscr{L}_{t}\left(5,1\right)=e_{0}+\rho_{0}\frac{\partial e_{0}}{\partial\rho_{0}}\\
\mathscr{L}_{t}\left(5,5\right)=\rho_{0}\frac{\partial e_{0}}{\partial T_{0}}
\end{array}\right\}
\end{equation}
\begin{equation}
\left.\begin{array}{l}
\mathscr{L}_{x}\left(1,1\right)=u_{0}\\
\mathscr{L}_{x}\left(1,2\right)=\rho_{0}\\
\mathscr{L}_{x}\left(2,1\right)=u_{0}u_{0}+\frac{\partial p_{0}}{\partial\rho_{0}}\\
\mathscr{L}_{x}\left(2,2\right)=2\rho_{0}u_{0}\\
\mathscr{L}_{x}\left(2,3\right)=-\frac{1}{Re}\frac{\partial\mu_{0}}{\partial y}\\
\mathscr{L}_{x}\left(2,5\right)=\frac{\partial p_{0}}{\partial T_{0}}\\
\mathscr{L}_{x}\left(3,1\right)=-\frac{1}{Re}\frac{\partial\mu_{0}}{\partial\rho_{0}}\frac{\partial u_{0}}{\partial y}\\
\mathscr{L}_{x}\left(3,2\right)=-\frac{1}{Re}\frac{\partial\lambda_{0}}{\partial y}\\
\mathscr{L}_{x}\left(3,3\right)=\mathscr{L}_{x}\left(4,4\right)=\rho_{0}u_{0}\\
\mathscr{L}_{x}\left(3,5\right)=-\frac{1}{Re}\frac{\partial\mu_{0}}{\partial T_{0}}\frac{\partial u_{0}}{\partial y}\\
\mathscr{L}_{x}\left(5,1\right)=e_{0}u_{0}+\rho_{0}u_{0}\frac{\partial e_{0}}{\partial\rho_{0}}\\
\mathscr{L}_{x}\left(5,2\right)=\rho_{0}e_{0}+p_{0}\\
\mathscr{L}_{x}\left(5,3\right)=-\frac{2}{Re}\mu_{0}\frac{\partial u_{0}}{\partial y}\\
\mathscr{L}_{x}\left(5,5\right)=\rho_{0}u_{0}\frac{\partial e_{0}}{\partial T_{0}}
\end{array}\right\}
\end{equation}
\begin{equation}
\left.\begin{array}{l}
\mathscr{L}_{y}\left(1,3\right)=\rho_{0}\\
\mathscr{L}_{y}\left(2,1\right)=-\frac{1}{Re}\frac{\partial\mu_{0}}{\partial\rho_{0}}\frac{\partial u_{0}}{\partial y}\\
\mathscr{L}_{y}\left(2,2\right)=-\frac{1}{Re}\frac{\partial\mu_{0}}{\partial y}\\
\mathscr{L}_{y}\left(2,3\right)=\rho_{0}u_{0}\\
\mathscr{L}_{y}\left(2,5\right)=-\frac{1}{Re}\frac{\partial\mu_{0}}{\partial T_{0}}\frac{\partial u_{0}}{\partial y}\\
\mathscr{L}_{y}\left(3,1\right)=\frac{\partial p_{0}}{\partial\rho_{0}}\\
\mathscr{L}_{y}\left(3,3\right)=-\frac{2}{Re}\frac{\partial\mu_{0}}{\partial y}-\frac{1}{Re}\frac{\partial\lambda_{0}}{\partial y}\\
\mathscr{L}_{y}\left(3,5\right)=\frac{\partial p_{0}}{\partial T_{0}}\\
\mathscr{L}_{y}\left(4,4\right)=-\frac{1}{Re}\frac{\partial\mu_{0}}{\partial y}\\
\mathscr{L}_{y}\left(5,1\right)=-\frac{1}{RePrEc}\left(\frac{\partial\kappa_{0}}{\partial\rho_{0}}\frac{\partial T}{\partial y}\right)\\
\mathscr{L}_{y}\left(5,2\right)=-\frac{2}{Re}\mu_{0}\frac{\partial u_{0}}{\partial y}\\
\mathscr{L}_{y}\left(5,3\right)=\rho_{0}e_{0}+p_{0}\\
\mathscr{L}_{y}\left(5,5\right)=-\frac{1}{RePrEc}\left(\frac{\partial\kappa_{0}}{\partial y}+\frac{\partial\kappa_{0}}{\partial T_{0}}\frac{\partial T_{0}}{\partial y}\right)
\end{array}\right\}
\end{equation}
\begin{equation}
\left.\begin{array}{l}
\mathscr{L}_{z}\left(1,4\right)=\rho_{0}\\
\mathscr{L}_{z}\left(2,4\right)=\rho_{0}u_{0}\\
\mathscr{L}_{z}\left(3,4\right)=-\frac{1}{Re}\frac{\partial\lambda_{0}}{\partial y}\\
\mathscr{L}_{z}\left(4,1\right)=\frac{\partial p_{0}}{\partial\rho_{0}}\\
\mathscr{L}_{z}\left(4,3\right)=-\frac{1}{Re}\frac{\partial\mu_{0}}{\partial y}\\
\mathscr{L}_{z}\left(4,5\right)=\frac{\partial p_{0}}{\partial T_{0}}\\
\mathscr{L}_{z}\left(5,4\right)=\rho_{0}e_{0}+p_{0}
\end{array}\right\}
\end{equation}
\begin{equation}
\left.\begin{array}{l}
\mathscr{L}_{q}\left(1,3\right)=\frac{\partial\rho_{0}}{\partial y}\\
\mathscr{L}_{q}\left(2,1\right)=-\frac{1}{Re}\frac{\partial\mu_{0}}{\partial\rho_{0}}\frac{\partial^{2}u_{0}}{\partial y^{2}}-\frac{1}{Re}\frac{\partial u_{0}}{\partial y}\left(\frac{\partial^{2}\mu_{0}}{\partial\rho_{0}^{2}}\frac{\partial\rho_{0}}{\partial y}+\frac{\partial^{2}\mu_{0}}{\partial\rho_{0}\partial T_{0}}\frac{\partial T_{0}}{\partial y}\right)\\
\mathscr{L}_{q}\left(2,3\right)=\rho_{0}\frac{\partial u_{0}}{\partial y}+u_{0}\frac{\partial\rho_{0}}{\partial y}\\
\mathscr{L}_{q}\left(2,5\right)=-\frac{1}{Re}\frac{\partial\mu_{0}}{\partial T_{0}}\frac{\partial^{2}u_{0}}{\partial y^{2}}-\frac{1}{Re}\left(\frac{\partial^{2}\mu_{0}}{\partial T_{0}^{2}}\frac{\partial T_{0}}{\partial y}+\frac{\partial^{2}\mu_{0}}{\partial T_{0}\partial\rho_{0}}\frac{\partial\rho_{0}}{\partial y}\right)\frac{\partial u_{0}}{\partial y}\\
\mathscr{L}_{q}\left(3,1\right)=\frac{\partial^{2}p_{0}}{\partial\rho_{0}^{2}}\frac{\partial\rho_{0}}{\partial y}+\frac{\partial^{2}p_{0}}{\partial\rho_{0}\partial T_{0}}\frac{\partial T_{0}}{\partial y}\\
\mathscr{L}_{q}\left(3,5\right)=\frac{\partial^{2}p_{0}}{\partial T_{0}^{2}}\frac{\partial T_{0}}{\partial y}+\frac{\partial^{2}p_{0}}{\partial\rho_{0}\partial T_{0}}\frac{\partial\rho_{0}}{\partial y}\\
\mathscr{L}_{q}\left(5,1\right)=-\frac{1}{RePrEc}\left(\frac{\partial^{2}T_{0}}{\partial y^{2}}\frac{\partial\kappa_{0}}{\partial\rho_{0}}+\left(\frac{\partial^{2}\kappa_{0}}{\partial\rho_{0}^{2}}\frac{\partial\rho_{0}}{\partial y}+\frac{\partial^{2}\kappa_{0}}{\partial\rho_{0}\partial T_{0}}\frac{\partial T_{0}}{\partial y}\right)\frac{\partial T_{0}}{\partial y}\right)-\frac{1}{Re}\frac{\partial\mu_{0}}{\partial\rho_{0}}\left(\frac{\partial u_{0}}{\partial y}\right)^{2}\\
\mathscr{L}_{q}\left(5,3\right)=e_{0}\frac{\partial\rho_{0}}{\partial y}+\rho_{0}\frac{\partial e_{0}}{\partial y}\\
\mathscr{L}_{q}\left(5,5\right)=-\frac{1}{RePrEc}\left(\frac{\partial^{2}T_{0}}{\partial y^{2}}\frac{\partial\kappa_{0}}{\partial T_{0}}+\left(\frac{\partial^{2}\kappa_{0}}{\partial T_{0}^{2}}\frac{\partial T_{0}}{\partial y}+\frac{\partial^{2}\kappa_{0}}{\partial\rho_{0}\partial T_{0}}\frac{\partial\rho_{0}}{\partial y}\right)\frac{\partial T_{0}}{\partial y}\right)-\frac{1}{Re}\frac{\partial\mu_{0}}{\partial T_{0}}\left(\frac{\partial u_{0}}{\partial y}\right)^{2}
\end{array}\right\}
\end{equation}
\begin{equation}
\left.\begin{array}{l}
\mathscr{V}_{xx}\left(2,2\right)=\mathscr{V}_{yy}\left(3,3\right)=\mathscr{V}_{zz}\left(4,4\right)=-\frac{2\mu_{0}+\lambda_{0}}{Re}\\
\mathscr{V}_{xx}\left(3,3\right)=\mathscr{V}_{xx}\left(4,4\right)=-\frac{\mu_{0}}{Re}\\
\mathscr{V}_{yy}\left(2,2\right)=\mathscr{V}_{yy}\left(4,4\right)=-\frac{\mu_{0}}{Re}\\
\mathscr{V}_{zz}\left(2,2\right)=\mathscr{V}_{zz}\left(3,3\right)=-\frac{\mu_{0}}{Re}\\
\mathscr{V}_{xx}\left(5,5\right)=\mathscr{V}_{yy}\left(5,5\right)=\mathscr{V}_{zz}\left(5,5\right)=-\frac{\kappa_{0}}{RePrEc}\\
\mathscr{V}_{xy}\left(2,3\right)=\mathscr{V}_{xy}\left(3,2\right)=-\frac{\mu_{0}+\lambda_{0}}{Re}\\
\mathscr{V}_{xz}\left(2,4\right)=\mathscr{V}_{xz}\left(4,2\right)=-\frac{\mu_{0}+\lambda_{0}}{Re}\\
\mathscr{V}_{yz}\left(3,4\right)=\mathscr{V}_{yz}\left(4,3\right)=-\frac{\mu_{0}+\lambda_{0}}{Re}
\end{array}\right\}
\end{equation}
\red{The second-order finite differences were used to determine the gradients of the properties. For instance, the gradients of viscosity
\begin{equation}
\frac{\partial\mu\left(T_{0},\rho_{0}\right)}{\partial T}=\frac{\mu\left(T_{0}+\Delta T,\rho_{0}\right)-\mu\left(T_{0}-\Delta T,\rho_{0}\right)}{2\Delta T}
\end{equation}
\begin{equation}
\frac{\partial\mu\left(T_{0},\rho_{0}\right)}{\partial\rho}=\frac{\mu\left(T_{0},\rho_{0}+\Delta\rho\right)-\mu\left(T_{0},\rho_{0}-\Delta\rho\right)}{2\Delta\rho}
\end{equation}
\begin{equation}
\frac{\partial^{2}\mu\left(T_{0},\rho_{0}\right)}{\partial T\partial\rho}=\frac{\frac{\partial\mu}{\partial\rho}\left(T_{0}+\Delta T,\rho_{0}\right)-\frac{\partial\mu}{\partial\rho}\left(T_{0}-\Delta T,\rho_{0}\right)}{2\Delta T}
\end{equation}
}
\red{An example of the sensitivity of $\dfrac{\pp^2 \mu^*}{\pp T^*\pp\rho^*}$ to $\Delta T^*$ and $\Delta \rho^*$ is shown in figure~\ref{Fig18}. In fact, the gradients of the thermodynamic \& transport properties became rather robust when $\Delta T^*\leq1$ K and $\Delta \rho^*\leq1$ Kg/m$^3$. In this study, the results are obtained with $\Delta T^*=0.1$ K and $\Delta \rho^*=0.1$ Kg/m$^3$.}
\begin{figure}
\begin{center}
\includegraphics[width=0.6\linewidth]{Figure_18}
\end{center}
\caption{Sensitivity of $\dfrac{\pp^2 \mu^*}{\pp T^*\pp\rho^*}$ to $\Delta T^*$ and $\Delta \rho^*$.}
\label{Fig18}
\end{figure}
\section{Cubic equation of state}\label{appB}
\red{
The material dependent parameters for CO2 are provided in table~\ref{Table5}. These parameters are necessary inputs for the cubic equation of states detailed below and can be easily replaced for other fluids.}
\begin{table}
\begin{center}
\begin{tabular}{ccccc}
gas constant & heat capacity ratio & acentric factor & critical pressure & critical temperature\vspace{5pt}\\
$R_g^*=188.9$ J/(Kg K) & $\gamma=1.289$ & $\omega=0.228$ & $p_c^*=73.9$ bar & $T_c^*=304.1$ K\\
\end{tabular}
\caption{The material dependent parameters for CO2.}
\label{Table5}
\end{center}
\end{table}
\subsection{The van der Waals equation of state}
The \citet{VdW_1873} equation of state (EoS) is the simplest cubic equation of state that is capable of accounting phase separation and the critical point \citep[see the introduction in][]{Zappoli_2015,Engineering_Thermodynamics_7Ed}. The EoS can be written as
\begin{equation}
p=\frac{R_{g}T}{\vartheta-b}-\frac{a}{\vartheta^{2}},
\end{equation}
\noindent
where $R_{g}$ is the specific gas constant, $a$ is a measure of the attraction forces between molecules, and $b$ accounts for the finite volume occupied by the molecules. The constants $a$ and $b$ can be determined at the critical point where
\begin{equation}
\left.\frac{\partial p}{\partial\vartheta}\right|_{T_c}=\left.\frac{\partial^{2}p}{\partial\vartheta^{2}}\right|_{T_c}=0~~\Rightarrow~~
a=\frac{27}{64}\frac{R_{g}^{2}T_{c}^{2}}{p_{c}},~~
b=\frac{R_{g}T_{c}}{8p_{c}}
\end{equation}
Using the Maxwell relations and the departure function, it is possible to obtain the internal energy as
\begin{equation}
e=C_vT-a\rho.
\end{equation}
The required derivatives for stability equations are
\begin{equation}
\left.\frac{\partial p}{\partial T}\right|_{\rho}=\frac{\rho R_{g}}{1-\rho b},~
\left.\frac{\partial p}{\partial\rho}\right|_{T}=\frac{R_{g}T}{\left(1-\rho b\right)^{2}}-2a\rho,
\end{equation}
\begin{equation}
\left.\frac{\partial}{\partial\rho}\right|_{T}\left(\left.\frac{\partial p}{\partial T}\right|_{\rho}\right)=\frac{R_{g}}{\left(1-\rho b\right)^{2}},~
\left.\frac{\partial^{2}p}{\partial T^{2}}\right|_{\rho}=0,~
\left.\frac{\partial^{2}p}{\partial\rho^{2}}\right|_{T}=\frac{2R_{g}Tb}{\left(1-\rho b\right)^{3}}-2a,
\end{equation}
\begin{equation}
\left.\frac{\partial e}{\partial T}\right|_{\rho}=C_v+\left.\frac{\partial C_v}{\partial T}\right|_{\rho}T,~
\left.\frac{\partial e}{\partial\rho}\right|_{T}=-a.
\end{equation}
\subsection{The Redlich-Kwong equation of state}
The Redlich-Kwong \citep{RK_1949} equation of state is given as
\begin{equation}
p=\frac{R_{g}T}{\vartheta-b}-\frac{a\alpha}{\vartheta\left(\vartheta+b\right)},
\end{equation}
where $\alpha=\sqrt{T_c/T}$. Similarly, by satisfying the critical condition, the constants $a$ and $b$ are
\begin{equation}
a=\frac{0.42748R_{g}T_{c}^{2}}{p_{c}},~
b=\frac{0.08664R_{g}T_{c}}{p_{c}}.
\end{equation}
The internal energy is
\begin{equation}
e=C_vT+\frac{3}{2}\frac{a}{b}\alpha\ln\frac{1}{1+b\rho}.
\end{equation}
The derivatives in the stability equations are
\begin{equation}
\left.\frac{\partial p}{\partial T}\right|_{\rho}=\frac{\rho R_{g}}{1-b\rho}+\frac{1}{2}T^{-\frac{3}{2}}T_{c}^{\frac{1}{2}}\frac{a\rho^{2}}{1+b\rho},~
\left.\frac{\partial p}{\partial\rho}\right|_{T}=\frac{R_{g}T}{\left(1-\rho b\right)^{2}}-T^{-\frac{1}{2}}T_{c}^{\frac{1}{2}}\frac{2a\rho+ab\rho^{2}}{\left(1+\rho b\right)^{2}},
\end{equation}
\begin{equation}
\left.\frac{\partial}{\partial\rho}\right|_{T}\left(\left.\frac{\partial p}{\partial T}\right|_{\rho}\right)=\frac{R_{g}}{\left(1-b\rho\right)^{2}}+\frac{1}{2}T^{-\frac{3}{2}}T_{c}^{\frac{1}{2}}\frac{2a\rho+ab\rho^{2}}{\left(1+b\rho\right)^{2}},
\end{equation}
\begin{equation}
\left.\frac{\partial^{2}p}{\partial T^{2}}\right|_{\rho}=-\frac{3}{4}T^{-\frac{5}{2}}T_{c}^{\frac{1}{2}}\frac{a\rho^{2}}{1+b\rho},~
\left.\frac{\partial^{2}p}{\partial\rho^{2}}\right|_{T}=\frac{2bR_{g}T}{\left(1-\rho b\right)^{3}}-T^{-\frac{1}{2}}T_{c}^{\frac{1}{2}}\frac{2a}{\left(1+\rho b\right)^{3}},
\end{equation}
\begin{equation}
\left.\frac{\partial e}{\partial T}\right|_{\rho}=C_v+\left.\frac{\partial C_v}{\partial T}\right|_{\rho}T-\frac{3}{4}\frac{a}{b}T^{-\frac{3}{2}}T_{c}^{\frac{1}{2}}\ln\frac{1}{1+b\rho},~
\left.\frac{\partial e}{\partial\rho}\right|_{T}=-\frac{3}{2}T^{-\frac{1}{2}}T_{c}^{\frac{1}{2}}\frac{a}{1+b\rho}.
\end{equation}
\subsection{The Peng-Robinson equation of state}
The Peng-Robinson \citep{PR_1976} equations of state modifies the original RK and SRK (RK modified by \citet{SRK_1972}) EoS, giving better results regarding the liquid density, vapor pressure and equilibrium ratios. It is one of the most used EoS. It is given as
\begin{equation}
p=\frac{R_{g}T}{\vartheta-b}-\frac{a\alpha}{\vartheta^{2}+2b\vartheta-b^{2}}.
\end{equation}
The constants $a$, $b$ and parameter $\alpha$ are given by
\begin{equation}
a=\frac{0.457235R_{g}^{2}T_{c}^{2}}{p_{c}},~
b=\frac{0.077796R_{g}T_{c}}{p_{c}},~
\alpha=\left(1+K\left(1-\sqrt{T/T_{c}}\right)\right)^{2}.
\end{equation}
Here $K=0.37464+1.54226\omega-0.26992\omega^{2}$, $\omega$ is the acentric factor of the species. The internal energy
\begin{equation}
e=C_vT+\frac{a}{2\sqrt{2}b}\left[\left(1+K\right)^{2}-K(1+K)\sqrt{T/T_{c}}\right]\ln\frac{1+b\left(1-\sqrt{2}\right)\rho}{1+b\left(1+\sqrt{2}\right)\rho}.
\end{equation}
The derivatives used in the linear stability equations are give by
\begin{equation}
\left.\frac{\partial p}{\partial T}\right|_{\rho}=\frac{\rho R_{g}}{1-\rho b}+K\sqrt{\frac{\alpha}{TT_{c}}}\frac{a\rho^{2}}{1+2b\rho-b^{2}\rho^{2}},
\end{equation}
\begin{equation}
\left.\frac{\partial p}{\partial\rho}\right|_{T}=\frac{R_{g}T}{\left(1-\rho b\right)^{2}}-\alpha\frac{2a\rho+2ab\rho^{2}}{\left(1+2b\rho-b^{2}\rho^{2}\right)^{2}},
\end{equation}
\begin{equation}
\left.\frac{\partial}{\partial\rho}\right|_{T}\left(\left.\frac{\partial p}{\partial T}\right|_{\rho}\right)=\frac{R_{g}}{\left(1-\rho b\right)^{2}}+K\sqrt{\frac{\alpha}{TT_{c}}}\frac{2a\rho+2ab\rho^{2}}{\left(1+2b\rho-b^{2}\rho^{2}\right)^{2}},
\end{equation}
\begin{equation}
\left.\frac{\partial^{2}p}{\partial T^{2}}\right|_{\rho}=-\frac{K\left(1+K\right)}{2\sqrt{T^{3}T_{c}}}\frac{a\rho^{2}}{1+2b\rho-b^{2}\rho^{2}},
\end{equation}
\begin{equation}
\left.\frac{\partial^{2}p}{\partial\rho^{2}}\right|_{T}=\frac{2R_{g}bT}{\left(1-\rho b\right)^{3}}-\alpha\frac{2a\left(2b^{3}\rho^{3}+3b^{2}\rho^{2}+1\right)}{\left(1+2b\rho-b^{2}\rho^{2}\right)^{3}},
\end{equation}
\begin{equation}
\left.\frac{\partial e}{\partial T}\right|_{\rho}=C_v+\left.\frac{\partial C_v}{\partial T}\right|_{\rho}T+\frac{a}{4\sqrt{2}b}\left[-K(1+K)\sqrt{1/TT_{c}}\right]\ln\frac{1+b\left(1-\sqrt{2}\right)\rho}{1+b\left(1+\sqrt{2}\right)\rho},
\end{equation}
\begin{equation}
\left.\frac{\partial e}{\partial\rho}\right|_{T}=-\frac{a}{1+2b\rho-b^{2}\rho^{2}}\left[\left(1+K\right)^{2}-K(1+K)\sqrt{T/T_{c}}\right].
\end{equation}
\section{Influence of the bulk viscosity}\label{appC}
The dynamics of a fluid are described by the Navier-Stokes equation, which in its simplest form contain a linear relation between deformation of a fluid element and the resulting stress, with the shear viscosity $\mu$ the coefficient of proportionality. Phenomenologically, another coefficient is possible, the second viscosity $\lambda$, which was introduced by \citet{Stokes1845}. Stokes anticipated that the second viscosity might play a role in compressible fluids. However, for the cases he considered, the fluids can be assumed incompressible with negligible dilatational effects, such that the bulk viscosity within the second viscosity can be ignored. This is known as the Stokes approximation. Consequently, setting the bulk viscosity $\mu_b$ to zero, has been broadly adopted in numerical simulations of compressible flows \citep[see a succinct review by][]{Graves1999}.
\citet{Cramer2012}'s numerical estimates indicate that $\mu_b/\mu$ of some common gases can reach $O(10^3)$. To investigate the influence of $\mu_b$ on the results of the linear stability, we performed simulations with $\mu_b=1000\mu$. The results are shown in figure~\ref{Fig19} and \ref{Fig20}, which show the comparison of the linear stability results for $\mu_b=1000\mu$ and $\mu_b=0$, using the RP model (the other parameters are kept the same). Figure~\ref{Fig19} shows that the neutral curves are barely affected. A discernible difference only exists in the transcritical case ($T_w^*=300$ K, \PrEc=0.06), where the neutral curve with $\mu_b=1000\mu$ becomes slightly more expanded. On the other hand, the algebraic instability does not vary with bulk viscosity. Only the modal growth region ($G_{max}=\infty$) in figure~\ref{Fig20}(b) becomes larger with $\mu_b=1000\mu$ and is consistent with the results shown in figure~\ref{Fig19}(b).
\begin{figure}
\begin{center}
\includegraphics[width=0.45\linewidth,clip]{Figure_19a}
\includegraphics[width=0.465\linewidth,clip]{Figure_19b}\\\vspace{10pt}
\includegraphics[width=0.465\linewidth,clip]{Figure_19c}
\end{center}
\caption{Influence of bulk viscosity on neutral curves in (a) subcritical (b) transcritical and (c) supercritical case.}
\label{Fig19}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\linewidth,clip]{Figure_20a}
\includegraphics[width=0.45\linewidth,clip]{Figure_20b}\\\vspace{10pt}
\includegraphics[width=0.45\linewidth,clip]{Figure_20c}
\end{center}
\caption{Influence of bulk viscosity on the algebraic growth. $\PrEc=0.07$, $\Rey=1000$. (a) subcritical (b) transcritical and (c) supercritical case.}
\label{Fig20}
\end{figure}
The above comparisons support the Stokes' hypothesis used in this study. In fact, $\mu_b$ is frequency dependent, this means that depending on the perturbation one prescribes in the stability analysis, the bulk viscosity will have different values. Therefore for a more rigorous investigation we would need reliable frequency resolved data for the bulk viscosity, either from theories, experiments \citep{Karim1952}, or molecular dynamics simulations \citep{Hoover1980} .
\section{\red{Influence of the reference scaling}}\label{appD}
\red{
Previous studies have shown that the scaling of the governing equations may have a large influence on the results. For example, if the viscosity at the cold wall is used as a reference value, \citet{sahu2010a} concluded that increasing the temperature difference between both walls destabilizes the flow, while \citet{SAMEEN2007} concluded the opposite behaviour if the viscosity at the hot wall is used. On the other hand, \citet{Rinaldi2018} investigated the edge state solutions of viscosity-stratified flows where they showed that a different reference value for viscosity does not qualitatively change the results. In this appendix, we show how the definition of the non-dimensional quantities will influence the results presented in the paper. }
\red{We introduce the averaged values of the thermodynamic \& transport properties:
\begin{gather}
T^*_{av}=\dfrac{1}{h^*}\int T^*\dd y,~
\rho^*_{av}=\dfrac{1}{h^*}\int \rho^*\dd y,\\
\mu^*_{av}=\dfrac{1}{h^*}\int \mu^*\dd y,~
\kappa^*_{av}=\dfrac{1}{h^*}\int \kappa^*\dd y.
\end{gather}
When the governing equations are scaled by the above averaged values, one obtains the averaged Reynolds number, $\overline{\Rey}$, and the product of the averaged Prandtl and Eckert number, $\overline{\PrEc}$:
\begin{equation}
\overline{\Rey}=\frac{\rho_{av}^{*}u_{r}^{*}h^{*}}{\mu_{av}^{*}},~
\overline{\PrEc}=\frac{\mu_{av}^{*}u_{r}^{*2}}{\kappa_{av}^{*}T_{av}^{*}}.
\end{equation}
We name it the \emph{average scaling}, to distinguish from the \emph{wall scaling} presented in \S\ref{Sec2-1} of the paper.
Note that the reference velocity is not independent, and is given by
\begin{equation}
u_r^*=\sqrt{\PrEc\frac{\kappa^*_wT_w^*}{\mu_w^*}}=\sqrt{\overline{\PrEc}\frac{\kappa^*_{av}T_{av}^*}{\mu_{av}^*}}.
\end{equation}
Using both scalings resulted in qualitatively similar conclusion as shown in figure~\ref{Fig21} for the modal instability. That is, the flow becomes more unstable in the subcritical regime, inviscid unstable in the transcritical regime, and more stable in the supercritical regime.
}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\linewidth]{Figure_21a}
\includegraphics[width=0.45\linewidth]{Figure_21b}\\\vspace{10pt}
\includegraphics[width=0.45\linewidth]{Figure_21c}
\end{center}
\caption{Influence of the reference scaling on the neutral curve for the (a) subcritical (b) transcritical and (c) supercritical cases. Solid lines show the results with wall scaling (in the $\alpha-\Rey$ diagram), while dashed lines indicate the average scaling (in the $\alpha-\overline{\Rey}$ diagram). These changes will}
\label{Fig21}
\end{figure}
\red{Regarding the algebraic instability using the average scaling, as can be seen from figure~\ref{Fig22}, the maximum growth shows a minor reduction in the subcritical regime. Increases in $G_{\max}$ are noticed for the trans- and supercritical regimes. Comparisons with the ideal gas have been summarized in table~\ref{Table6}. The ideal gas are not sensitive to the wall temperature under both scalings. With average scaling, the conclusion for the algebraic instability will not change.}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\linewidth]{Figure_22a}
\includegraphics[width=0.45\linewidth]{Figure_22b}\\\vspace{10pt}
\includegraphics[width=0.45\linewidth]{Figure_22c}
\end{center}
\caption{Influence of the reference scaling on the maximum algebraic growth $G_{\max}$ for the (a) subcritical (b) transcritical and (c) supercritical cases. The left and right half show the results with wall scaling and average scaling respectively. The non-ideal gas model is RP, $\PrEc=0.07$, $\Rey=1000$.}
\label{Fig22}
\end{figure}
\begin{table}
\begin{center}
\begin{tabular}{lcc}
& \multicolumn{2}{c}{Scaling} \\
& wall-based & average-based\vspace{5pt}\\
$T^*_w=290$ K (IG) & 178.1 & 189.9\\
$T^*_w=290$ K (RP) & 316.7 & 271.3\vspace{5pt}\\
$T^*_w=300$ K (IG) & 178.1 & 189.9\\
$T^*_w=300$ K (RP) & 1040.8 & 1582.7\vspace{5pt}\\
$T^*_w=310$ K (IG) & 178.1 & 189.9\\
$T^*_w=310$ K (RP) & 190.3 & 268.2
\end{tabular}
\caption{\red{Maximum algebraic growth $G_{\max}$ over $\alpha$ and $\beta$. $\PrEc=0.07$, $\Rey=1000$ (wall scaling) or $\overline{\Rey}=1000$ (average scaling). The values for the $T^*_w=300$ case are for $\alpha=0$ where modal instability is absent.}}
\label{Table6}
\end{center}
\end{table}
\section{The laminar base flow}\label{Sec3}
The base flow \red{is driven by a constant body force in the streamwise direction} and is obtained by solving \eqref{NS1}, \eqref{NS2} and \eqref{NS3} with the assumption that the flow is fully developed, spanwise and streamwise independent, steady and parallel, \textit{i.e.~} $
\partial\left(\right)/\partial x=0,~
\partial\left(\right)/\partial z=0,~
\partial\left(\right)/\partial t=0,~
v=w=0$.
The N-S equations are thus simplified as
\begin{equation}\label{EqBaseflow1}
\frac{\pp}{\pp y}\left(\mu\frac{\pp u}{\pp y}\right)=-\Rey F=-\hat{F},
\end{equation}
\begin{equation}\label{EqBaseflow2}
\frac{\pp p}{\pp y}=0,
\end{equation}
\begin{equation}\label{EqBaseflow3}
\frac{\pp}{\pp y}\left(\frac{\kappa}{\PrEc}\frac{\pp T}{\pp y}+\mu u\frac{\pp u}{\pp y}\right)=-\Rey F\cdot u=-\hat{F}\cdot u.
\end{equation}
It is worth noting that the above equations are independent of density, therefore, $\rho_0$ can be separately determined by the EoS. We assume the body force, $\hat{F}$, which drives the flow, to be uniform. \red{To obtain a solution of the base flow, an initial temperature field is assumed, \eg $T=T_w=$ constant, $\mu$ and $\kappa$ are determined from REFPROP according to the temperature and pressure. First, the velocity is solved using equation \eqref{EqBaseflow1}, followed by an update of temperature by solving \eqref{EqBaseflow3}. $\mu$ and $\kappa$ are then updated using the obtained temperature. This procedure is repeated until the solution is converged.}
\subsection{The isothermal limit}\label{Sec3-1}
When $\PrEc\rightarrow0$, the viscous heating is negligible if compared to the thermal conduction. Therefore, the temperature, as well as the other thermodynamic properties, remain constant, namely $T_0=1,~\rho_0=1,~\mu_0=1,~\kappa_0=1$.
The flow is thus simply governed by ${\pp^{2}u}/{\pp y^{2}}=-\hat{F}$. Choosing $u^*_r$ as the centerline velocity, leads to setting $\hat{F}=2$. As a result, the dimensionless base flow is independent of any parameters (\eg $T_w$, $\hat{F}$ and $PrEc$) and is given by $u_0=y(2-y)$. A sketch of this base flow, which is free from any non-ideal gas effects, is shown in figure~\ref{Fig3} (dashed lines).
\begin{figure}
\begin{center}
\vspace{20pt}
\includegraphics[width=0.65\linewidth]{Figure_3}
\end{center}
\caption{Sketch of the laminar base flow. Dashed lines show the isothermal limit with $\PrEc\rightarrow0$, such that $u_0=y(2-y)$, $T_0=1$, $\rho_0=1$. Solid lines represent a transcritical case, in which $T_0$ crosses $T_{pc}$ and $u_0$ is inflectional. }
\label{Fig3}
\end{figure}
\subsection{The compressible base flow}\label{Sec3-2}
Equations \eqref{EqBaseflow1}, \eqref{EqBaseflow2} and \eqref{EqBaseflow3} show that the compressible base flow is determined by $\PrEc$, $\hat{F}$, and $T^*_w$. Either by increasing $\PrEc$ or $\hat{F}$, the compressibility effects become more significant. Without loss of the generality, a constant body force $\hat{F}=2$ is specified in this work, while $\PrEc$ is varied from the isothermal limit (we assume $\PrEc=10^{-5}$) to a typical compressible state with $\PrEc=0.1$. For example, setting $\PrEc=0.1$ and $T_w^*=290,~300,$ or $310$ K, the Mach number is $\Ma=0.40,~0.58,$ or $1.35$, respectively. In this work, the wall temperature $T^*_w$ is considered in a range from 265 to 320 K. Note, given our non-dimensionalization, the base flow is free from the choice of the Reynolds number.
Figure~\ref{Fig4} shows the contours of the centerline temperature $T_{\rm center}^*$ and velocity $u_{\rm center}=u_{\rm center}^*/u_r^*$ as a function of wall temperature $T_w^*$ and $\PrEc$.
Regardless of the wall temperature, an increase of $\PrEc$ is accompanied with an increase of $T_{\rm center}^*$ and $u_{\rm center}$ as compressible effects become more prominent. Interestingly, a distinct right-angled triangular area emerges in each subplot of figure~\ref{Fig4}. At the hypotenuse of this triangle, the centerline temperature, $T_{\rm center}^*$, and velocity, $u_{\rm center}$, suddenly increase, forming a discontinuity in the $PrEc$--$T_w^*$ plane.
It is also interesting to note that the hypotenuse of the triangle almost coincides with the line where $T^*_{\rm center}$ reaches the pseudo critical temperature $T_{pc}^*=307.7$ K (shown with the dot-dashed line). Likewise, the upper boundary of the triangle coincides with the dotted line where $T^*_w=T_{pc}^*$.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\linewidth,clip]{Figure_4a}
\includegraphics[width=0.48\linewidth,clip]{Figure_4b}
\end{center}
\caption{Centerline (a) temperature $T^*_{\rm center}$ and (b) velocity $u_{\rm center}$ as functions of wall temperature $T_w^*$ and $\PrEc$. Model RP, $p^*=80$ bar, $\hat{F}=2$.}
\label{Fig4}
\end{figure}
For a more detailed discussion, we will now define three cases with different wall temperatures that are summarized in table~\ref{Table2} and highlighted by dashed lines in figure~\ref{Fig4}. These cases will also be used in the subsequent sections regarding the linear modal and algebraic instability analysis. The wall temperature for these cases has been set to 290, 300 and 310 K, such that their temperature profile in the considered range of $PrEc$ is either subcritical, transcritical or supercritical, respectively.
Their base flow profiles are plotted in figure~\ref{Fig5}, together with the incompressible limit, indicated by the dashed line in each subplot. The profiles on the left half (black lines) and right half (blue lines) represent the base flow of the non-ideal (RP) and ideal (IG) gases, respectively. As $\PrEc$ uniformly increases from 0.01 to 0.1 it can be seen that the temperature and velocity increase, while the density decreases. For the transcritical case, however, a sudden jump of the base flow profiles can be observed. This jump occurs between $\PrEc=$ 0.05115 and 0.05116, as highlighted by the orange and red lines in figure~\ref{Fig5}(b,e,h). Note, the jump is caused by an inflectional velocity profile as highlighted by the red line in figure~\ref{Fig5}(h).
\begin{table}
\begin{center}\def~{\hphantom{0}}
\begin{tabular}{rcccc}
Case & $T_w^*$ & $\PrEc$ & $\Ma$ & Temperature range \vspace{5pt} \\
Subcritical & 290 K &$\PrEc\leq0.1$ & $\Ma\leq0.40$ & 290 K (wall) - 304.9 K (center)\\
Transcritical & 300 K &$\PrEc\leq0.1$ & $\Ma\leq0.58$ & 300 K (wall) - 366.2 K (center)\\
Supercritical & 310 K &$\PrEc\leq0.1$ & $\Ma\leq1.35$ & 310 K (wall) - 328.6 K (center)\\
\end{tabular}
\caption{Cases investigated in this study.}
\label{Table2}
\end{center}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth,clip]{Figure_5a}\\
\includegraphics[width=0.8\linewidth,clip]{Figure_5b}\\
\includegraphics[width=0.8\linewidth,clip]{Figure_5c}
\end{center}
\caption{Temperature (a-c), density (d-f) and velocity (g-i) profiles of the base flow. The wall temperature is (a,d,g) $T^*_w=290$ K,
(b,e,h) $T^*_w=300$ K,
(c,f,i) $T^*_w=310$ K respectively. $\PrEc$ increases uniformly from 0.01 to 0.1. The black and blue lines on the left and right half denote the non-ideal (RP) and ideal (IG) gases respectively. The dashed lines in each subplot shows the isothermal limit. The REFPROP library is used for the transport and thermodynamic properties of the non-ideal gas. The orange and red lines in subplot (b,e,h) show the profiles at $\PrEc=$ 0.05115 and 0.05116 respectively. The dash-dotted lines (the triangle) in subplot (g,h,i) show the lines of constant gradient $|{\partial u}/{\partial y}|=\hat{F}$.}
\label{Fig5}
\end{figure}
The discontinuous behaviour with respect to $PrEc$ can be explained as follows. Integrating \eqref{EqBaseflow1} gives $\mu{\partial u}/{\partial y}=-\hat{F}y+C.$ Applying the symmetry condition at the channel center ($y=1$), it follows that $C=\hat{F}$. Therefore, \eqref{EqBaseflow1} can be written as
\begin{equation}\label{EqBaseflow5}
\mu\frac{\partial^{2}u}{\partial y^{2}}=-\hat{F}-\frac{\partial\mu}{\partial y}\frac{\partial u}{\partial y}=
-\hat{F}\left(1+\frac{1}{\mu}\frac{\partial\mu}{\partial y}(1-y)\right).
\end{equation}
Based on \eqref{EqBaseflow5}, it can be seen that an inflectional velocity profile occurs if the viscosity gradient is large enough to change the sign within the parenthesis in (\ref{EqBaseflow5}), namely if
\begin{equation}\label{EqBaseflow6}
\frac{1}{\mu}\left|\frac{\partial\mu}{\partial y}\right|>1.
\end{equation}
In the cases considered herein, it appears that the viscosity gradient at the wall is large enough to cause an inflectional profile to occur when the temperature in the channel center reaches $T_{pc}^*$. Recall figure \ref{Fig2}(c), a sharp gradient of the viscosity (${\partial\mu}/{\partial T}\ll0$) is seen close to the pseudo-critical point. As $\PrEc$ increases, ${\partial T}/{\partial y}$ increases at the wall, such that ${\partial\mu}/{\partial y}\cong({\partial\mu}/{\partial T})({\partial T}/{\partial y})$ can drop below -1 at the wall, leading to inflectional velocity profiles. The jump of the base flow solution can thus be explained by referring to figure \ref{Fig5}(h). Since, $\mu |{\partial u}/{\partial y}|$ at the wall is equal to the constant forcing $\hat{F}$, regardless of $\PrEc$ and wall temperature, the velocity profiles with/without inflectional points are isolated by the line of constant gradient $\hat{F}$ (the dash-dotted lines that form a triangle in figure \ref{Fig5}(g-i)). Therefore, a velocity profile without an inflection point cannot reach the apex of the triangle ($|{\partial u}/{\partial y}|$ decreases towards channel center) and the sudden increase of the centerline velocity appears once an inflection point is formed.
In general, the base flow solutions can be summarized as follows:
\begin{itemize}
\item In the {\bf subcritical} case, the wall temperature is much lower than $T_{pc}^*$, and in the range of $\PrEc$ considered, $T_{\rm center}^*$ is always less than $T_{pc}^*$. Hence, the velocity profile is not inflectional.
\item In the {\bf transcritical} case, the wall temperature is close to $T_{pc}^*$, such that for large enough $\PrEc$, $T_{\rm center}^*$ reaches $T_{pc}^*$. Consequently, a jump of the solution with respect to $\PrEc$ occurs and the velocity profile becomes inflectional. From figure~\ref{Fig4}(b), it can be inferred that the lower the wall temperature $T_w^*$, the larger the discontinuity will be.
\item In the {\bf supercritical} case, the wall temperature is higher than $T_{pc}^*$. The properties of the fluid are gas-like \red{(compressed vapour)} and the velocity is not inflectional.
\end{itemize}
\section{Introduction}\label{Sec1}
Many processes in industrial applications constitute of flows with fluids that do not follow the ideal gas law. For example, flows in vapour power systems, re-entry of spacecrafts, supercritical dyeing, refrigeration and heat pump systems \citep[examples in supercritical fluids can be found in][]{Brunner2010}. The non-ideality of fluids is especially significant in the thermodynamic region close to the vapour critical point. As such, it is of great importance to understand the fundamental physics that are related to flows with these fluids.
Recently, researchers have studied how non-ideal gas effects influence turbulence and heat transfer. For example, \citet{Kawai2015,Kawai2016} studied turbulent boundary layers with supercritical pressures and transcritical temperatures. They found that the mean velocity profiles (with density weighted Van Driest transformation) coincide with the same log-law as seen in an ideal gas. \citet{Sciacovelli2017,Sciacovelli2017b,Sciacovelli2016} comprehensively studied turbulence dynamics and near wall turbulence of flows with molecularly complex fluids in the dense gas regime using direct numerical simulations. They found that dense-gas flows with a heavy fluorocarbon exhibit almost negligible friction heating (in channel flows), weakening of compressive (and enhancement of expanding) structures (in homogeneous isotropic turbulence). \citet{Patel2016} studied the influence of variable properties on fully developed turbulent channel flows and derived a velocity transformation that allows to collapse velocity profiles for heated or cooled non-ideal fluids. Moreover, \citet{Rinaldi2017} provided an explanation of near wall turbulence modulation, especially the intercomponent energy transfer that has been observed by, \eg \citet{Morinishi2004}, \citet{Pirozzoli2008}, \citet{Duan2010}. \citet{Nemati2016,Peeters2016} studied turbulent heat transfer to supercritical CO$_2$, indicating that both the mean and instantaneous property variations have significant effects on turbulent structures and their self-regeneration processes in near-wall turbulence. \citet{Alferez2017} have studied the refraction properties of compression shock waves in non-ideal gases. One of the new regimes found is that, due to the non-ideality of the fluid it is possible that acoustic modes can be completely damped by a compression shock, leading to so-called `quiet shocks'.
For ideal gases, the thermodynamic properties are associated with a simple equation of state (EOS). Additionally, the transport properties (namely, the viscosity and thermal conductivity) can be estimated as unary functions of the temperature (\eg the widely used power law or Sutherland's law). To assess to which degree the ideal gas law holds, it is possible to evaluate the compressibility factor, defined as $Z=p^*/\rho^* R^*T^*$. Figure~\ref{Fig1} shows the $T-\vartheta$ diagram (temperature - specific volume diagram, $\vartheta=1/\rho$) of carbon dioxide CO$_2$. The white circle in each subplot indicates the critical point, which for CO$_2$ is at a pressure and temperature of $p^*_c=73.9$ bar and $T^*_c=304.25$ K. In this paper, we denote dimensional and critical quantities with superscript `$*$' and subscript `$c$', respectively. Figure \ref{Fig1}(a) shows the critical isobar (black thin dashed line), four isobars of 40 to 100 bar (black thin lines), and the compressibility factor $Z$ (colored contour lines). Close to the critical point, the non-ideality is clearly indicated by low values of $Z$, while the boundary between ideal and non-ideal gas behavior is roughly indicated by the thick dashed line of $Z=0.99$. The distribution of the thermodynamic and transport properties (specific heat capacity at constant volume $C_v^*$, dynamic viscosity $\mu^*$ and thermal conductivity $\kappa^*$) are shown in figure \ref{Fig1}(b,c,d). In the ideal gas region, these contour lines become quasi-parallel to the $x$-axis, indicating that they can be regarded as functions of temperature only. On the other hand, near the critical point, the gradients of these properties with respect to temperature and specific volume become significant.
\begin{figure}
\begin{center}
\includegraphics[width=0.49\linewidth]{Figure_1a}
\includegraphics[width=0.49\linewidth]{Figure_1b}\\
\includegraphics[width=0.49\linewidth]{Figure_1c}
\includegraphics[width=0.49\linewidth]{Figure_1d}
\end{center}
\caption{$T-\vartheta$ diagram of CO$_2$ along with the critical point (white circle), the saturation curves (blue and red solid lines), the liquid-vapour region (grey area), the critical isobar (black thin dashed line), isobars of 40, 60, 80 and 100 bar (black thin lines), compressibility factor $Z=0.99$ (black thick dashed line) as well as colored contour lines of (a) $Z=p^*/\rho^* R^*T^*$, (b) the specific heat capacity at constant volume $C_v^*$, (c) the dynamic viscosity $\mu^*$, and (d) the thermal conductivity $\kappa^*$.}
\label{Fig1}
\end{figure}
In view of its great simplicity, most of the present knowledge on stability and laminar-turbulent transition is limited to the ideal gas \citep{Fedorov2011} or incompressible flows, where thermodynamic properties are constant. On the other hand, numerical simulations of real gas effects (high-enthalpy effects) in hypersonic flows has just gone through an initial stage \citep{Zhong2012,Marxen2013,Marxen2014}. In fact, the well-known Orr-Sommerfeld equation \citep[][often termed the O-S equation]{Orr1907, Sommerfeld1908} was derived by applying the linear stability theory (LST) to the incompressible parallel plane shear flow. Solved as an eigenvalue problem (in the time/space-asymptotic limit), the growth rate and profiles of the perturbation are obtained from the most unstable mode as its eigenvalue and eigenvector. This is known as the modal stability problem. The critical Reynolds number $\Rey_c$, below which the flow is stable, regardless of the wavenumber and frequency of the perturbation, is often determined and emphasized in such modal stability analysis. For example, in plane Poiseuille flow, $\Rey_c$ is numerically determined to be 5772.22 \citep{Thomas1953, Orszag1971}. Here the Reynolds number is based on the half-channel height and the centerline flow velocity. Due to the non-normality of most practical linear systems, the modal stability analysis cannot cover the full behavior of the linear instability \citep{Schmid2001, Schmid2014}. Instead of solving the eigenvalue problem, the stability equation can be formulated as an initial value problem under the framework of constrained optimization. Maximizing the energy growth in a finite domain of time or space, leads to the optimal perturbation, which grows transiently even below the critical Reynolds number $\Rey_c$. This is termed the transient growth or algebraic growth. Accordingly, a ``critical'' Reynolds number can be defined for the algebraic growth as well. Also for plane Poiseuille flow, this number is 49.6 \citep{Joseph1969, Busse1969}.
Studies of viscosity stratified flows, where the viscosity depends on temperature, has recently received attention, readers may refer to \citet{Govindarajan2014} for a review. Based on the modified O-S equations, early studies show that including a linear temperature profile destabilizes the Poiseuille flow \citep{Potter1972} and stabilizes/destabilizes the water boundary layer flow (depend on wall heating/cooling) \citep{Wazzan1972}. However, viscosity and temperature perturbations were ignored in both studies and were later examined by \citet{Pinarbasi1995}. \citet{Wall1996,Wall1997} investigated the effects of different viscosity models, indicating that the flow can either be more stable or unstable. The study on wall heating and viscosity-stratification has also been extended to transient growth, secondary instability \red{as well as instabilites in other types of flows} \citep{Chikkadi2005,SAMEEN2007,SAMEEN2011,sahu2011,sahu2014}. For compressible plane Couette flow, \citet{Malik2008} showed that the flow is more stable with viscosity stratification, while recently, a further study on this flow is given by \citet{Saikia2017}, in which the effects of individual/combined viscosity-thermal conductivity stratification are elucidated. The influence of viscosity gradients on the edge state is recently studied by \citet{Rinaldi2018}, showing that in minimal channel flows, the kinetic energy level and the driving force of self-sustained cycle of the edge state depends on viscosity distribution. The above studies are based on the incompressible flow assumption or the ideal gas equation-of-state (EoS), at the same time, transport properties are estimated as functions of temperature only.
Since there is very limited knowledge on flow stability with highly non-ideal fluids, we investigate Poiseuille flows with fluids close to the thermodynamic vapour-liquid critical point. In \S \ref{Sec2}, the gas model, the formulation of the stability analysis and the related numerical methods are outlined in detail. The results and discussions on the base flow are provided in \S \ref{Sec3}, followed by the modal growth and algebraic instability in \S \ref{Sec4} and \ref{Sec5}, respectively. The paper is concluded in \S \ref{Sec6}.
\section{Conclusion}\label{Sec6}
Linear stability of highly non-ideal plane Poiseuille flows is studied. We have chosen carbon dioxide (CO$_2$) at supercritical pressures ($p^*=$80 bar) as an example of a fluid in a highly non-ideal thermodynamic region.
The investigation is based on the fully compressible Navier-Stokes equations in which the product of two dimensionless parameters, namely the Prandtl $Pr$ and Eckart $Ec$ numbers, determines the viscous heating and consequently the temperature increase between the two isothermal walls. The investigated range of $\PrEc$ is from the isothermal limit ($\PrEc\rightarrow0$) to typical compressible flows with $\PrEc=0.1$.
Three cases with wall temperatures in the vicinity of the pseudo-critical point ($T^*_{pc}=307.7$ K) have been investigated in more detail. In particular, the wall temperatures are set such that the temperature profile is subcritical ($T_w^*=290$ K), transcritical ($T_w^*=300$ K) and supercritical ($T_w^*=310$ K). In all cases, the thermodynamic and transport properties are strongly dependent on the thermodynamic state of the fluid (\eg temperature, density) and they influence the stability in a coupled way through the base flow and the linear stability operator.
In the isothermal limit, the three cases with different wall temperatures have the same base flow as the ideal gas. When $\PrEc$ increases, the base flows of the three cases deviate from the ideal gas solution. In the subcritical regime, as $\PrEc$ increases, the flow becomes more unstable with regard to both the modal and algebraic growth, while for ideal gas the trend is opposite. When $\PrEc$ is large enough, or $T_w$ is closer to (but lower than) $T_{pc}$, the flow falls in the transcritical regime. Due to the large gradient of the viscosity near $T_{pc}$, the base flow becomes inflectional and inviscid unstable. As a consequence, the stability of the non-ideal gas flow is very different from the ideal gas. The neutral curve is expanded, which results in a very low critical Reynolds number. Moreover, the algebraic growth is also enhanced. It should be expected that the laminar-turbulent transition is more likely to be dominated by modal growth in this regime. When $T_w>T_{pc}$, the fluid is in the thermodynamic supercritical regime. In this case, the results of the modal growth show that the non-ideal gas is substantially more stable than the ideal gas. However, the transient growth shows only a weak dependence on the non-ideal gas effects. Additionally, we show that the linear stability analysis with simple cubic equations of state give qualitatively similar results than using the more accurate multi-parameter equation of state implemented in the REFPROP library. \red{Discussions on the reference scaling indicate that the conclusion is not influenced by the choice of the reference variables.} This investigation constitutes the first systematic study of linear stability with highly non-ideal fluids close to the thermodynamic critical point. Future studies will focus on the validation of the results using direct numerical simulations.
\section{Governing equations}\label{Sec2}
\subsection{Flow conservation equations}\label{Sec2-1}
The laws of conservation of mass, momentum and energy (the Navier-Stokes (N-S) equations), in dimensionless form, are given by
\begin{equation}\label{NS1}
\frac{\partial\rho}{\partial t}+\frac{\partial\left(\rho u_{j}\right)}{\partial x_{j}}=0,
\end{equation}
\begin{equation}\label{NS2}
\frac{\partial\left(\rho u_{i}\right)}{\partial t}+\frac{\partial\left(\rho u_{i}u_{j}+p\delta_{ij}-\tau_{ij}\right)}{\partial x_{j}}=F_i,
\end{equation}
\begin{equation}\label{NS3}
\frac{\partial\left(\rho E\right)}{\partial t}+\frac{\partial\left(\rho Eu_{j}+pu_{j}+q_{j}-u_{i}\tau_{ij}\right)}{\partial x_{j}}=u_jF_j,
\end{equation}
where $x_i=(x,y,z)$ are the coordinates in the streamwise, wall-normal and spanwise directions, $u_i=(u,v,w)$ are the velocity components, $t$ the time, $\rho$ the fluid density, $E=e+u_iu_i/2$ the total energy, $e$ the internal energy, $F_i$ the body force and $p$ is the pressure. The viscous stress tensor, $\tau_{ij}$, and the heat flux, $q_{j}$, are given by
\begin{equation}\label{stresstensor}
\tau_{ij}=\frac{\mu}{\Rey}\left(\frac{\partial u_{i}}{\partial x_{j}}+\frac{\partial u_{j}}{\partial x_{i}}\right)+\frac{\lambda}{\Rey}\delta_{ij}\frac{\partial u_{k}}{\partial x_{k}}, \quad
q_{j}=-\frac{\kappa}{\Rey\PrEc}\frac{\partial T}{\partial x_{j}}.
\end{equation}
Here $\mu$ is the dynamic viscosity, $\lambda=\mu_b -2/3 \mu$ the second viscosity, $\mu_b$ the bulk viscosity, and $\kappa$ is the thermal conductivity. Results presented in the following sections are subject to $\mu_b=0$. However, we will discuss the influence of the bulk viscosity on the linear stability in Appendix \ref{appC}.
The above equations have been non-dimensionalized by reference values, as follows
\begin{gather}
u=\frac{u^{*}}{u_{r}^{*}},~
x_{i}=\frac{x_{i}^{*}}{h^{*}},~
t=\frac{t^{*}u_{r}^{*}}{h^{*}},~
p=\frac{p^{*}}{\rho_{w}^{*}u_{r}^{*2}},~
\rho=\frac{\rho^{*}}{\rho_{w}^{*}},~\nonumber \\
T=\frac{T^{*}}{T_{w}^{*}},~
E=\frac{E^{*}}{u_{r}^{*2}},
\mu=\frac{\mu^{*}}{\mu_{w}^{*}},~
\kappa=\frac{\kappa^{*}}{\kappa_{w}^{*}},
\end{gather}
which leads to the definition of the Reynolds number, $\Rey$, Prandtl number, $\Pran$, Eckert number, $\Ec$ and the Mach number, $\Ma$, which are given as
\begin{equation}
\Rey=\frac{\rho_{w}^{*}u_{r}^{*}h^{*}}{\mu_{w}^{*}},~
\Pran=\frac{\mu_{w}^{*}C_{pw}^{*}}{\kappa_{w}^{*}},~
\Ec=\frac{u_{r}^{*2}}{C_{pw}^{*}T_{w}^{*}},~
\Ma=\frac{u_{r}^{*}}{c^*_w}.
\end{equation}
The subscript $w$ denotes wall values, $h^*$ is the half channel height, $c^*_w$ is the speed of sound at the wall, $u^*_r$ is the reference velocity. Note that for an ideal gas $\Ec=(\gamma-1)\Ma^2$, where $\gamma$ is the heat capacity ratio. In this study, both walls are at the same temperature. \red{Discussions on the choice of different reference scalings are provided in Appendix \ref{appD}. }
\subsection{Fluid equations of state}
In order to find a closed form of the conservation equations, an equation of state for the fluid has to be specified. As a representative example of non-ideal fluids, the study is performed with CO$_2$ at a pressure of $p^*= 80$ bar, which is above the critical pressure, within the highly non-ideal thermodynamic region (see the isobar in figure \ref{Fig1}). To account for the non-ideal gas effects, the NIST REFPROP library \citep{Lemmon_2002} has been used to obtain the most accurate thermodynamic and transport properties along with their gradients.
\red{The multi-parameter EoS (in functional forms) used in REFPROP are developed with an optimization algorithm. The EoS are suitable for a broad variety of fluids while high accuracy can be maintained. Readers shall refer to \citet{Span2003a} for the derivation of the EoS. To build the linear stability equations (see Appendix \ref{appA}), the temperature $T_0$ and density $\rho_0$ are provided as input, while the required properties and their derivatives are obtained as output from REFPROP.}
Moreover, as a direct method to determine the thermodynamic properties, several cubic EoS (see Appendix \ref{appB}), i.e. van der Waals \citep{VdW_1873}, Redlich-Kwong \citep{RK_1949} and Peng-Robinson \citep{PR_1976}, are used for the stability analysis as comparison. All results with the non-ideal EOS are also compared with an ideal gas model (IG). \red{A constant specific heat ratio $\gamma=1.289$ is used for the IG model.} All the fluid models are summarized in table \ref{Table1}.
\begin{table}
\begin{center}
\begin{tabular}{lll}
Fluid model & EoS & Transport properties\vspace{5pt}\\
RP & REFPROP & REFPROP\\
PR & Peng-Robinson & REFPROP\\
RK & Redlich-Kwong & REFPROP\\
VW & van der Waals & REFPROP\\
IG & ideal gas & Power/Sutherland law
\end{tabular}
\caption{Fluid models studied in this paper. Gradients of the properties (with respect to temperature and density) are calculated analytically (see Appendix \ref{appB}) or numerically with a finite-difference algorithm (REFPROP). As shown in figure~\ref{Fig2}(c,d), there is no discernible difference using the power or Sutherland law for the IG model, therefore results presented in this study for IG are based on the power law.}
\label{Table1}
\end{center}
\end{table}
Figure \ref{Fig2} shows the thermodynamic and transport properties of CO$_2$ at a pressure of 80 bar. The pentagram in subplot (a) shows the pseudo-critical temperature ($T_{pc}^*=307.7$~K, RP model), which is defined as the point on a supercritical isobar where $C_p^*$ reaches a maximum. Near $T_{pc}^*$, all properties show large gradients, which do not exist in an ideal gas. As shown in figure \ref{Fig2}(a,b), the Peng-Robinson (PR) EoS is closest to the highly accurate multiparameter EoS of CO$_2$ as implemented in REFPROP (RP).
In general, the cubic EoSs do capture key features of the thermodynamic property variations.
In figure~\ref{Fig2}(c,d), the power law~\eqref{power} and Sutherland law~\eqref{suther}, which fall on top of each other, are compared to the distributions from RP. The power and Sutherland laws for dynamic viscosity and thermal conductivity are given as
\begin{equation}\label{power}
\frac{\mu^*}{\mu^*_{\textrm{ref}}}=\left(\frac{T^*}{T^*_\textrm{ref}}\right)^{n_{1}},~
\frac{\kappa^*}{\kappa^*_\textrm{ref}}=\left(\frac{T^*}{T^*_\textrm{ref}}\right)^{n_{2}},~n_1=0.79,~
n_2=1.30,~
{T_{\textrm{ref}}^{*}}=273~\textrm{K},
\end{equation}
\begin{equation}\label{suther}
\frac{\mu^{*}}{\mu_{\textrm{ref}}^{*}}=\left(\frac{T^{*}}{T_{\textrm{ref}}^{*}}\right)^{\frac{3}{2}}\frac{T_{\textrm{ref}}^{*}+S_{1}^{*}}{T^{*}+S_{1}^{*}},~
\frac{\kappa}{\kappa_\textrm{ref}}=\left(\frac{T^{*}}{T_{\textrm{ref}}^{*}}\right)^{\frac{3}{2}}\frac{T_{\textrm{ref}}^{*}+S_{2}^{*}}{T^{*}+S_{2}^{*}},
\end{equation}
where
\begin{equation}
\left.\begin{array}{c}
{T_{\textrm{ref}}^{*}}=273~\textrm{K},~
\mu^*_{\textrm{ref}}=1.37\times10^{-5}~\textrm{Pa}\cdot\textrm{s},~
\kappa^*_{\textrm{ref}}=0.0146~\textrm{W}/(\textrm{m}\cdot\textrm{K}),~\\
n_1=0.79,~
n_2=1.30,~
S_1^*=222~\textrm{K},~
S_2^*=1800~\textrm{K}.
\end{array}\right\}
\end{equation}
In general, as temperature increases from subcritical to supercritical values, the fluid continuously transitions from \red{compressed liquid} to \red{compressed vapour} and finally reaches values that can be described by an ideal gas.
\begin{figure}
\begin{center}
\vspace{10pt}
\includegraphics[width=0.95\linewidth]{Figure_2}
\end{center}
\caption{Thermodynamic and transport properties of CO$_2$ at $p^*=80$ bar for the fluid models in table~\ref{Table1}. Sub-figures show the distribution of (a) density $\rho^*$, (b) heat capacity at constant pressure $C_p^*$, (c) viscosity $\mu^*$ and, (d) thermal conductivity $\kappa^*$ versus temperature $T^*$. The pentagram shows the pseudo-critical temperature $T^*_{pc}$ (RP model). The shaded area indicates the pseudo-critical transition.}
\label{Fig2}
\end{figure}
\subsection{The linearized stability equations}\label{Sec2-2}
Following the common procedure, the flow field is decomposed into the base flow and a perturbation, as
\begin{equation}\label{decompose}
\left.\begin{array}{c}
\rho=\rho_{0}+\rho^{\prime}\\
u_{i}=u_{i0}+u_{i}^{\prime}\\
T=T_{0}+T^{\prime}\\
p=p_{0}+p^{\prime}\\
E=E_{0}+E^{\prime}\\
\mu=\mu_{0}+\mu^{\prime}\\
\kappa=\kappa_{0}+\kappa^{\prime}
\end{array}\right\}
\end{equation}
It is known that for simple compressible systems (\eg pure substances, uniform mixture of nonreacting gases), the thermodynamic state is defined by two independent thermodynamic properties. In this study, we keep $\rho$ and $T$ as the two basic thermodynamic variables, while the other thermodynamic and transport properties (\eg $E$, $p$, $\mu$, $\kappa$) are determined as functions of $\rho$ and $T$. For example, the pressure perturbation $p^{\prime}$ is expanded by a Taylor-series with respect to $\rho_0$ and $T_0$ in the following way
\begin{multline}
p^{\prime}=\left.\frac{\partial p_{0}}{\partial\rho_{0}}\right|_{T_{0}}\rho^{\prime}+\left.\frac{\partial p_{0}}{\partial T_{0}}\right|_{\rho_{0}}T^{\prime}\\
+\frac{1}{2}\left(\left.\frac{\partial^{2}p_{0}}{\partial\rho_{0}^{2}}\right|_{T_{0}}\rho^{\prime}\rho^{\prime}+2\left(\left.\frac{\partial}{\partial\rho_{0}}\right|_{T_{0}}\left.\frac{\partial}{\partial T_{0}}\right|_{\rho_{0}}p_{0}\right)\rho^{\prime}T^{\prime}+\left.\frac{\partial^{2}p_{0}}{\partial T_{0}^{2}}\right|_{\rho_{0}}T^{\prime}T^{\prime}\right)+\cdots
\end{multline}
For the sake of brevity, the partial derivative of a quantity with respect to $T$ at constant $\rho$, will be written as $\left.\partial/\partial T\right|_{\rho_{0}} \equiv \partial/\partial T$, and accordingly $\left.\partial/\partial \rho\right|_{T_{0}} \equiv \partial/\partial \rho$. The stability equation is derived by substituting \eqref{decompose} into the N-S equations \eqref{NS1}, \eqref{NS2} and \eqref{NS3}, and then subtracting the governing equations of the base flow. With the nonlinear terms neglected, the linear stability equations are formulated as
\begin{multline}\label{Stability}
\mathscr{L}_{t}\frac{\partial\boldsymbol{q}}{\partial t}+\mathscr{L}_{x}\frac{\partial\boldsymbol{q}}{\partial x}+\mathscr{L}_{y}\frac{\partial\boldsymbol{q}}{\partial y}+\mathscr{L}_{z}\frac{\partial\boldsymbol{q}}{\partial z}+\mathscr{L}_{q}\boldsymbol{q} \\
+\mathscr{V}_{xx}\frac{\partial^{2}\boldsymbol{q}}{\partial x^{2}}+\mathscr{V}_{xy}\frac{\partial^{2}\boldsymbol{q}}{\partial x\partial y}+\mathscr{V}_{xz}\frac{\partial^{2}\boldsymbol{q}}{\partial x\partial z}+\mathscr{V}_{yy}\frac{\partial^{2}\boldsymbol{q}}{\partial y^{2}}+\mathscr{V}_{yz}\frac{\partial^{2}\boldsymbol{q}}{\partial y\partial z}+\mathscr{V}_{zz}\frac{\partial^{2}\boldsymbol{q}}{\partial z^{2}}=0.
\end{multline}
Here $\boldsymbol{q}=\left(\rho^{\prime},u^{\prime},v^{\prime},w^{\prime},T^{\prime}\right)^{T}$ is the perturbation vector and $\mathscr{L}_t$, $\mathscr{L}_x$, $\mathscr{L}_y$, $\mathscr{L}_z$, $\mathscr{L}_q$, $\mathscr{V}_{xx}$, $\mathscr{V}_{yy}$, $\mathscr{V}_{zz}$, $\mathscr{V}_{xy}$, $\mathscr{V}_{yz}$, and $\mathscr{V}_{xz}$ are matrices of size $5\times5$. The detailed expressions for these matrices are provided in Appendix~\ref{appA}. As can be seen, they are functions of the base flow, the thermodynamic and transport properties, $\Rey$ and $\PrEc$. The gradients of the properties are either calculated analytically using cubic EoS (see Appendix~\ref{appB}) or numerically employing finite-difference method within the REFPROP library.
\subsection{Modal and algebraic stability}\label{Sec2-3}
In modal stability, the perturbation is assumed to have the form
\begin{equation}\label{Wave}
\boldsymbol{q}\left(x,y,z,t\right)=\hat{\boldsymbol{q}}\left(y\right)\exp\left(i\alpha x+i\beta z-i\omega t\right)+c.c.
\end{equation}
where $c.c.$ stands for the complex conjugate. Substituting \eqref{Wave} into \eqref{Stability} results in
\begin{multline}\label{Stability2}
(-i\omega\mathscr{L}_{t}+i\alpha\mathscr{L}_{x}+\mathscr{L}_{y}D+i\beta\mathscr{L}_{z}+\mathscr{L}_{q}\\
-\alpha^{2}\mathscr{V}_{xx}+i\alpha\mathscr{V}_{xy}D-\alpha\beta\mathscr{V}_{xz}+\mathscr{V}_{yy}D^{2}+i\beta \mathscr{V}_{yz}D-\beta^{2}\mathscr{V}_{zz})\hat{\boldsymbol{q}}=0,
\end{multline}
where $D={\rm d}/{\rm d}y$. The equation \eqref{Stability2} is solved as an eigenvalue problem, which describes the development of the perturbations in temporal or spatial domain, \textit{i.e.~}
\begin{equation}\label{Eigen}
\begin{cases}
L_{T}\hat{\boldsymbol{q}}=\omega\mathscr{L}_{t}\hat{\boldsymbol{q}} & \mathrm{(temporal)},\\
L_{S}\hat{\boldsymbol{q}}=\alpha\left(\beta\mathscr{V}_{xz}-i\mathscr{V}_{xy}D-i\mathscr{L}_{x}\right)\hat{\boldsymbol{q}}+\alpha^{2}\mathscr{V}_{xx}\hat{\boldsymbol{q}} & (\mathrm{spatial}),
\end{cases}
\end{equation}
where
\begin{eqnarray}
L_{T}&=&\alpha\mathscr{L}_{x}-i\mathscr{L}_{y}D+\beta\mathscr{L}_{z}-i\mathscr{L}_{q} \nonumber\\
&+&i\alpha^{2}\mathscr{V}_{xx}+\alpha\mathscr{V}_{xy}D+i\alpha\beta\mathscr{V}_{xz}-iD^{2}\mathscr{V}_{yy}+\beta D\mathscr{V}_{yz}+i\beta^{2}\mathscr{V}_{zz}, \\
L_{S}&=&-i\omega\mathscr{L}_{t}+\mathscr{L}_{y}D+i\beta\mathscr{L}_{z}+\mathscr{L}_{q}+D^{2}\mathscr{V}_{yy}+i\beta D\mathscr{V}_{yz}-\beta^{2}\mathscr{V}_{zz}.
\end{eqnarray}
Here we consider the temporal problem only, therefore $\alpha$ and $\beta$ are the prescribed streamwise and spanwise wave numbers. $\omega=\omega_r+i\omega_i$ is solved as the eigenvalue, where $\omega_r$ and $\omega_i$ give the angular frequency and growth rate of the perturbation. \red{The domain $0\leq y \leq 2$ is discretized with Chebyshev collocation points, defined by
\begin{equation}
y_j=1-\cos\dfrac{\pi j}{N}~~~j=0,1,...,N-1,N.
\end{equation}
The differentiation of \eqref{Eigen} is accomplished with the matrix form of Chebyshev collocation derivatives. Numerical tests indicate that, typically $N=200$ (used here) are sufficient to give a grid-independent solution of the physical modes.}
With regard to the algebraic stability, following \citet{Schmid2001}, the optimal energy amplification is defined as:
\begin{equation}
G\left(t\right)=\underset{\boldsymbol{q}_{0}}{\max}\frac{E\left(\boldsymbol{q}\left(t\right)\right)}{E\left(\boldsymbol{q}_{0}\right)},~
G\left(x\right)=\underset{\boldsymbol{q}_{0}}{\max}\frac{E\left(\boldsymbol{q}\left(x\right)\right)}{E\left(\boldsymbol{q}_{0}\right)},~
\end{equation}
Here $E\left(\boldsymbol{q}\right)$ is the disturbance energy with the definition as given in \eqref{Norm}. \red{The perturbation $\boldsymbol{q}$ is expanded by the eigenvector obtained from the modal stability. The calculation of the optimal energy amplification $G$ and the corresponding optimal perturbation (the input), as well as the resulting perturbation (the output), lead to a singular value problem, which is solved with the same Chebyshev differentiation method as in the modal growth \citep{Schmid2001,Schmid2014}. }
The (modal and algebraic) perturbations are solved subjected to the boundary condition: $u^{\prime}=v^{\prime}=w^{\prime}=T^{\prime}=0$ at the lower ($y=0$) and upper wall ($y=2$).
\section{Linear modal instability}\label{Sec4}
Depending on the cases discussed below, we will use the definition of dynamic and thermodynamic modes, as
\begin{equation}
\begin{cases}
\rho^{\prime}=0,\:\mathrm{and}\:T^{\prime}=0 & \textrm{(dynamic modes)} \\
\rho^{\prime}\neq0,\:\mathrm{or}\:T^{\prime}\neq0 & \textrm{(thermodynamic modes). }
\end{cases}
\end{equation}
\subsection{The isothermal limit}\label{Sec4-1}
\begin{figure}
\begin{center}
\includegraphics[width=0.465\linewidth,clip]{Figure_6a}
\includegraphics[width=0.48\linewidth,clip]{Figure_6b} \\
\includegraphics[width=0.45\linewidth,clip]{Figure_6c} \hspace{4pt}
\includegraphics[width=0.45\linewidth,clip]{Figure_6d}
\end{center}
\caption{Eigen spectrum (a) and neutral curve (b) for the isothermal limit. The eigen spectrum is subject to $\alpha=1$, $\beta=0$, and $\Rey=10000$. The neutral curve is solved for 2-D perturbations ($\beta=0$). Symbols show results using different fluid models (RP, PR, RK, VW, IG) and incompressible equations (IC). IC in subplot (b) shows the results given by \citet[][pp. 71]{Schmid2001}. In subplot (c) and (d), profiles of the unstable mode ($\omega=0.2375+0.0037i$) and one of the stable modes ($\omega=0.4164-0.1382i$, highlighted in orange in the spectrum) are shown. The perturbations are normalized by $|u^\prime|$. An offset of -0.1 and -0.2 is applied to $|\rho^\prime|$ and $|T^\prime|$. The solid lines are results with fluid model IG.}
\label{Fig6}
\end{figure}
With the base flow obtained in section~\ref{Sec3-1}, we solve the stability equations \eqref{Stability} for the isothermal limit with different fluid models (RP, PR, RK, VW, IG), as well as for the incompressible equations (IC). As shown in figure~\ref{Fig6}(a), at $\Rey=10000$, $\alpha=1$ and $\beta=0$, the A-, P- and S- branches \citep[originally named by][]{Mack_1976} are reproduced by incompressible equations. Comparing the results using different equations of state, the eigenvalues fall on top of the incompressible counterparts, verifying the correct behaviour of the compressible models at low Eckert (Mach) numbers. One of the modes (highlighted in red) is exclusively unstable. Despite being solved with different thermodynamic models, this mode is shown to be a dynamic mode, which leads to identical neutral curve and eigenfunctions as shown in figure~\ref{Fig6}(b,c). The contour lines in figure \ref{Fig6}(b) show the growth rate $\omega_i$ (RP model). In fact, inspecting the stability equations \eqref{Stability} (see Appendix~\ref{appA}), it can be shown that the thermodynamic and transport properties do not influence the dynamic modes in the isothermal limit. For instance, gradients of properties, which vary among different models, are multiplied with thermodynamic components of the perturbations.
On the other hand, more stable modes emerge when the compressible equations are solved. By looking into the corresponding eigenfunctions (not shown), thermodynamic components become important in these modes, and as such dependent on the non-ideal gas properties. We plot one of the stable modes in figure \ref{Fig6}(d), where density perturbations are captured by compressible equations (indicated by blue ellipses).
\subsection{Compressible flows}\label{Sec4-2}
To achieve a first impression of the non-ideal gas effects, the problem is first studied with the RP model, where thermodynamic and transport properties are taken from the REFPROP library. Figure~\ref{Fig7} shows the neutral curves (a-c) as well as eigenfunctions (d-f) at representative parameters. As discussed in \S~\ref{Sec3-2}, the temperature is subcritical, transcritical and supercritical with $T^*_w=$ 290K, 300K and 310K respectively. The results are compared with ideal gas (IG).
By increasing $\PrEc$, the base flow of the ideal gas becomes more stable as the critical Reynolds number increases, regardless of $T_w$ specified. In fact, despite the difference in wall temperature, the dimensionless thermodynamic and transport properties (scaled with wall values) remain much the same. On the other hand, the behaviour of the non-ideal cases is different for the three cases investigated. In the subcritical case, the flow becomes more unstable when $\PrEc$ is increased. This is manifested by the enlargement of the neutral curve. Similarly, the transcritical case becomes more unstable as $\PrEc$ increases. However, once $\PrEc$ reaches the critical value (in this case $\PrEc=0.05115$), the base flow becomes inflectional. The flow is thus inviscid unstable and the critical Reynolds number is substantially reduced. For instance, the flow is unstable for $\Rey<1000$ and $\PrEc=0.06$.
In the supercritical case, the increase of $\PrEc$ stabilizes the base flow and the non-ideal gas is even more stable than the ideal gas. In this case, when $\PrEc$ reaches 0.03, the modal instability is found after $\Rey>8000$. Interestingly, a weak influence of $\PrEc$ on the velocity perturbations is observed (see figure~\ref{Fig7} (d-f)), while the amplitudes of density and temperature perturbations are considerably larger if $\PrEc$ increases. For the transcritical case, when the flow enters the triangular zone, the density perturbation are the most dominant.
\begin{figure}
\begin{center}
\includegraphics[width=0.85\linewidth]{Figure_7a}\\
\includegraphics[width=0.85\linewidth]{Figure_7b}\\
\includegraphics[width=0.85\linewidth]{Figure_7c}\\
\vspace{0.2cm}
\includegraphics[width=0.85\linewidth]{Figure_7d}\\
\includegraphics[width=0.85\linewidth]{Figure_7e}\\
\includegraphics[width=0.85\linewidth]{Figure_7f}
\end{center}
\caption{Neutral curves and profiles of perturbations for the non-ideal gas (RP model) and ideal gas (IG model). (a,d) $T_w^*=290$ K, (b,e) $T_w^*=300$ K, (c,f) $T_w^*=310$ K. The neutral curves are obtained for 2-D perturbations ($\beta=0$). The profiles shown are subject to $\alpha=1$, $\beta=0$ and $\Rey=10000$, and they are normalized with $|u^\prime|_{\max}$. The left and right half shows the non-ideal and ideal gas respectively.}
\label{Fig7}
\end{figure}
Below, we compare the fidelity of the cubic EoS models with the EoS model from REFPROP. The solutions for the ideal EoS are also shown to highlight the difference with respect to the results obtained with the non-ideal EoS models. Figure~\ref{Fig8} shows the growth rate of the unstable modes for all EoS models. Recall the discussion in \S \ref{Sec4-1}, all these curves collapse under the isothermal limit. As can be inferred from each row of figure~\ref{Fig8}, the differences between these models magnify when $\PrEc$ is increased. In all three cases, the cubic EoS models predict the correct trend that the flow becomes more unstable in sub-/transcritical cases, and more stable in supercritical cases as $\PrEc$ increases.
Specifically, the van der Waals EoS shows a good agreement with the RP EoS model in the subcritical case, while both Peng-Robinson and Redlich-Kwong EoS predict a lower growth rate (shown in figure~\ref{Fig8}). In the transcritical case, the van der Waals and Redlich-Kwong give acceptable growth rates if compared to RP. When the base flow becomes inflectional ($\PrEc=0.06$), the Peng-Robinson EoS shows the best approximation. In the supercritical case, Redlich-Kwong produces the best results, while the van der Waals EoS gives a much lower growth rate. Given these observations, it can be concluded that all non-ideal EoS models give the same trends. However, it is not possible to state the fidelity of the cubic EoS models in terms of the growth rate.
\begin{figure}
\includegraphics[scale=0.46,clip]{Figure_8a}
\includegraphics[scale=0.46,clip]{Figure_8b}
\includegraphics[scale=0.46,clip]{Figure_8c}
\caption{Growth rates of the perturbation for different gas models. Results shown are at $\Rey=10000$, $\beta=0$ for the subcritical case (a-d, $\PrEc=$ 0.01, 0.03, 0.05 and 0.07), transcritical case (e-h, $\PrEc=$ 0.01, 0.03, 0.05 and 0.06) and supercritical case (i-l, $\PrEc=$ 0.01, 0.015, 0.02 and 0.03). Note that the y-coordinate of subplot (h) is different from the others.}
\label{Fig8}
\end{figure}
\subsection{\red{The kinetic energy budget}}
\red{To further understand the instability mechanism of the non-ideal fluids, we perform a kinetic energy budget analysis for the 2D perturbation. The energy balance equation is the sum of the x-momentum perturbation equation, multiplied with $\hat{u}^\dagger$, and the y- equation, multiplied with $\hat{v}^\dagger$. Here, dagger stands for the complex conjugate. The continuity equation is used to substitute the temporal growth of density, which appears in the x-momentum equation. This gives the following kinetic energy balance equation:
\begin{equation}\label{budget}
K=\Theta+P+T+V,
\end{equation}
where
\begin{equation}
K=-i\omega\int\rho_{0}\left(\hat{u}\hat{u}^{\dagger}+\hat{v}\hat{v}^{\dagger}\right)\dd y,
\end{equation}
\begin{equation}
\Theta=-i\alpha\int\rho_{0}u_{0}\left(\hat{u}\hat{u}^{\dagger}+\hat{v}\hat{v}^{\dagger}\right)\dd y,
\end{equation}
\begin{equation}
P=-\int\rho_{0}\frac{\partial u_{0}}{\partial y}\hat{v}\hat{u}^{\dagger}\dd y,
\end{equation}
\begin{equation}
\begin{alignedat}{1}T=-\int\left[i\alpha\frac{\partial p_{0}}{\partial\rho_{0}}\hat{\rho}\hat{u}^{\dagger}+i\alpha\frac{\partial p_{0}}{\partial T_{0}}\hat{T}\hat{u}^{\dagger}+\frac{\partial p_{0}}{\partial\rho_{0}}\frac{\partial\hat{\rho}}{\partial y}\hat{v}^{\dagger}+\frac{\partial p_{0}}{\partial T_{0}}\frac{\partial\hat{T}}{\partial y}\hat{v}^{\dagger}+\right.\\
\left.\left(\frac{\partial^{2}p_{0}}{\partial\rho_{0}^{2}}\frac{\partial\rho_{0}}{\partial y}+\frac{\partial^{2}p_{0}}{\partial\rho_{0}\partial T_{0}}\frac{\partial T_{0}}{\partial y}\right)\hat{\rho}\hat{v}^{\dagger}+\left(\frac{\partial^{2}p_{0}}{\partial T_{0}^{2}}\frac{\partial T_{0}}{\partial y}+\frac{\partial^{2}p_{0}}{\partial\rho_{0}\partial T_{0}}\frac{\partial\rho_{0}}{\partial y}\right)\hat{T}\hat{v}^{\dagger}\right]\dd y,
\end{alignedat}
\end{equation}
\begin{equation}
\begin{alignedat}{1}V=\frac{1}{Re}\int\left[-\alpha^{2}\left(2\mu_{0}+\lambda_{0}\right)\hat{u}\hat{u}^{\dagger}+\mu_{0}\frac{\partial^{2}\hat{u}}{\partial y^{2}}\hat{u}^{\dagger}+i\alpha\left(\mu_{0}+\lambda_{0}\right)\frac{\partial\hat{v}}{\partial y}\hat{u}^{\dagger}\right.\\
+i\alpha\frac{\partial\mu_{0}}{\partial y}\hat{v}\hat{u}^{\dagger}+\frac{\partial\mu_{0}}{\partial\rho_{0}}\frac{\partial u_{0}}{\partial y}\frac{\partial\hat{\rho}}{\partial y}\hat{u}^{\dagger}+\frac{\partial\mu_{0}}{\partial y}\frac{\partial\hat{u}}{\partial y}\hat{u}^{\dagger}+\frac{\partial\mu_{0}}{\partial T_{0}}\frac{\partial u_{0}}{\partial y}\frac{\partial\hat{T}}{\partial y}\hat{u}^{\dagger}\\
+\frac{\partial\mu_{0}}{\partial\rho_{0}}\frac{\partial^{2}u_{0}}{\partial y^{2}}\hat{\rho}\hat{u}^{\dagger}+\frac{\partial u_{0}}{\partial y}\left(\frac{\partial^{2}\mu_{0}}{\partial\rho_{0}^{2}}\frac{\partial\rho_{0}}{\partial y}+\frac{\partial^{2}\mu_{0}}{\partial\rho_{0}\partial T_{0}}\frac{\partial T_{0}}{\partial y}\right)\hat{\rho}\hat{u}^{\dagger}\\
+\frac{\partial\mu_{0}}{\partial T_{0}}\frac{\partial^{2}u_{0}}{\partial y^{2}}\hat{T}\hat{u}^{\dagger}+\frac{\partial u_{0}}{\partial y}\left(\frac{\partial^{2}\mu_{0}}{\partial T_{0}^{2}}\frac{\partial T_{0}}{\partial y}+\frac{\partial^{2}\mu_{0}}{\partial T_{0}\partial\rho_{0}}\frac{\partial\rho_{0}}{\partial y}\right)\hat{T}\hat{u}^{\dagger}\\
-\alpha^{2}\mu_{0}\hat{v}\hat{v}^{\dagger}+\left(2\mu_{0}+\lambda_{0}\right)\frac{\partial^{2}\hat{v}}{\partial y^{2}}\hat{v}^{\dagger}+i\alpha\left(\mu_{0}+\lambda_{0}\right)\frac{\partial\hat{u}}{\partial y}\hat{v}^{\dagger}\\
\left.+i\alpha\frac{\partial\mu_{0}}{\partial\rho_{0}}\frac{\partial u_{0}}{\partial y}\hat{\rho}\hat{v}^{\dagger}+i\alpha\frac{\partial\lambda_{0}}{\partial y}\hat{u}\hat{v}^{\dagger}+i\alpha\frac{\partial\mu_{0}}{\partial T_{0}}\frac{\partial u_{0}}{\partial y}\hat{T}\hat{v}^{\dagger}+\left(2\frac{\partial\mu_{0}}{\partial y}+\frac{\partial\lambda_{0}}{\partial y}\right)\frac{\partial\hat{v}}{\partial y}\hat{v}^{\dagger}\right]\dd y.
\end{alignedat}
\end{equation}
The real part of the equation \eqref{budget} describes the balance of the kinetic energy growth. In particular, $K_r$ is the temporal growth of the kinetic energy, $\Theta$ is purely imaginary and does therefore not contribute to the temporal growth, $P_r$ is the production term, $T_r$ is the thermodynamic term, and $V_r$ is the viscous dissipation.}
\red{The results of the kinetic energy budget analysis are summarized in table~\ref{Table3}. The analysis is performed for all three cases at $\alpha=1$ and $\Rey=10000$. It clearly shows that for all the cases, the energy growth $K_r$ originates from the production term $P_r$. The thermodynamic term $T_r$ slightly reduces the growth. The viscous dissipation $V_r$ is not sensitive to the parameters and remains almost constant, except in the transcritical case ($T_w^*=300$ K, $\PrEc=0.06$), which has a considerably larger growth rate (as also shown in figure~\ref{Fig7} and \ref{Fig8}). The reason for this lies in a much larger production and a smaller viscous dissipation. Figure~\ref{Fig9} compares the production of the two cases with $\PrEc=0.05$ and 0.06 at $T_w^*=300$ K. It can be inferred that the inflectional velocity profile ($\PrEc=0.06$) has caused a larger $\rho_0\pp u_0/\pp y$ near both walls, the amplitude of the velocity perturbation $\hat{v}\hat{u}^\dagger$ is larger as well. Therefore, a large production term $-\int\rho_{0}\frac{\partial u_{0}}{\partial y}\hat{v}\hat{u}^{\dagger}\dd y$ and accordingly the large growth rate can be explained. }
\begin{table}
\begin{center}
\begin{tabular}{llrrrr}
\multicolumn{2}{l}{Cases} & \multicolumn{4}{l}{Budgets ($\times10^{-3}$)} \vspace{5pt} \\ \cline{1-2} \cline{3-6}
\\
$T_w^*$ &\PrEc & $K_r$ & $P_r$ & $T_r$ & $V_r$ \vspace{8pt}\\
\multirow{4}{*}{290 K} & 0.01 & 3.4 & 8.2 & 0.0 & -4.8 \\
& 0.03 & 4.5 & 9.4 & 0.0 & -4.9 \\
& 0.05 & 5.6 & 10.7 & 0.0 & -5.1 \\
& 0.07 & 6.8 & 12.1 & -0.1 & -5.2\vspace{5pt} \\
\multirow{4}{*}{300 K} & 0.01 & 3.5 & 8.4 & -0.1 & -4.8 \\
& 0.03 & 5.0 & 10.1 & -0.1 & -5.0 \\
& 0.05 & 6.2 & 11.6 & -0.2 & -5.2 \\
& 0.06 & 17.5 & 21.1 & -2.4 & -1.2\vspace{5pt} \\
\multirow{4}{*}{310 K} & 0.01 & 1.8 & 6.6 & 0.0 & -4.8 \\
& 0.015 & 1.2 & 6.1 & -0.2 & -4.7 \\
& 0.02 & 0.6 & 5.5 & -0.2 & -4.7 \\
& 0.03 &-0.7 & 4.3 & -0.3 & -4.7 \\
\end{tabular}
\caption{\red{Kinetic energy budget analysis for two-dimensional perturbations. $\alpha=1$, $\Rey=10000$.}}
\label{Table3}
\end{center}
\end{table}
\begin{figure}
\includegraphics[width=0.45\linewidth]{Figure_9a}
\includegraphics[width=0.45\linewidth]{Figure_9b}
\caption{Production of the kinetic perturbation energy with $T_w^*=300$ K, $\alpha=1$, $\Rey=10000$. (a) $\PrEc=0.05$, (b) $\PrEc=0.06$. }
\label{Fig9}
\end{figure}
\section{Algebraic growth}\label{Sec5}
\subsection{Choice of the energy norm}
The Mack's energy norm \citep{Mack1969, Hanifi1996} has been extensively used in compressible flows. The norm is designed under the ideal gas assumption, therefore the pressure-related energy transfer terms can be eliminated by choosing suitable coefficients for each components. \red{In fact, Mack's norm is equivalent to Chu's norm \citep{Chu1965,George2011}}. In the current non-ideal gas flows, the equation of states can be different (PR, RK, VW, IG), or even implicit (look-up table) as in the case of the RP EoS model. \red{Therefore, we choose a general form of the norm:}
\begin{equation}\label{Norm}
E\left(\boldsymbol{q}\right)=\int \left(u^{\prime\dagger}u^{\prime}+v^{\prime\dagger}v^{\prime}+w^{\prime\dagger}w^{\prime}\right)+m_{\rho}\rho^{\prime\dagger}\rho^{\prime}+m_{T}T^{\prime\dagger}T^{\prime}~{\rm d}V,
\end{equation}
where $\dagger$ denotes the complex conjugate. This norm has been tested for the compressible ideal/non-ideal gas flows at various conditions. Figure \ref{Fig10} shows the optimal energy growth $G_{\max}$ (the maximum of $G(t)$ over time $t$) as a function of $m_{T}$ and $m_{\rho}$, for $\PrEc=0.05$ (thermodynamic components become important) and a wall temperature of $T^*_w=290$~K. When $m_{\rho}$ is set to 0, $G_{\max}$ converges to a constant value when $m_T$ is large enough. On the other hand, the energy norm is shown to be rather robust when the density component is properly accounted for, \eg $m_{\rho}=1$. Therefore, the results presented in this section are mainly obtained for $m_{\rho}=m_{T}=1$. \red{A comparison with Mack's energy norm ($m_{\rho}=T_0/(\rho_0^2\gamma\Ma^2)$, $m_T=1/(\gamma(\gamma-1)T_0\Ma^2)$) is proivded at the end of this section}.
\begin{figure}
\begin{center}
\includegraphics[width=0.55\linewidth,clip]{Figure_10}
\end{center}
\caption{Maximum energy growth $G_{\max}$ versus $m_{T}$ using the energy norm \eqref{Norm}. $m_{\rho}=0$, 1, 2 and 10. The fluid (with RP model) is at $\PrEc=0.05$, $\Rey=2000$, $\alpha=1.0$, $\beta=0.25$ and $T^*_w=290$ K.}
\label{Fig10}
\end{figure}
\subsection{The isothermal limit}
Although all EoS considered in this work give the same most unstable mode in the isothermal limit (discussed in \S \ref{Sec4-1}), their eigenvalue spectrum can be rather different (see figure \ref{Fig6}(a)). Their corresponding eigenfunctions form the basis of the optimal perturbation and the algebraic growth.
We show the contour plot of $G_{\max}$ in $\alpha-\beta$ diagram in figure~\ref{Fig11}(a). Lines and circles show results of RP and IG models, respectively. It is evident that they fall on top of each other. In fact, all five models (RP, PR, RK, VW, IG) show the same results, and correspond to the results using incompressible equations. The largest transient growth occurs at $\alpha=0$ and $2\leq\beta\leq2.1$, which is well-known for ideal gas. The optimal perturbation and the corresponding output are shown in figure \ref{Fig11}(b,c) for $\alpha=0$, $\beta=2$. The classic streamwise vortices (the optimal perturbation) and streaks (the corresponding output) are recovered. There is no discernible difference between the non-ideal and ideal gases under the isothermal limit.
\begin{figure}
\begin{center}
\includegraphics[width=0.55\linewidth,clip]{Figure_11a}\vspace{4pt}
\includegraphics[width=0.75\linewidth,clip]{Figure_11bc}
\end{center}
\caption{Transient growth in the isothermal limit. $\PrEc\rightarrow0$, $\Rey=2000$, $T_w^*=290$ K. (a) Contour plot of $G_{\max}$. (b) The optimal perturbation (input) for $\alpha=0$, $\beta=2$. (c) The corresponding output. Lines and circle symbols show results of non-ideal (RP) and ideal gas (IG) respectively.}
\label{Fig11}
\end{figure}
\subsection{Compressible flows}
The algebraic growth has been studied for the subcritical, transcritical and supercritical cases at $\Rey=1000$ and $\PrEc=$ 0.01, 0.03, 0.05, 0.07. The optimal energy growth $G_{max}$ for RP model is compared with IG model in figure \ref{Fig12}, \ref{Fig13} and \ref{Fig14}, respectively. The three cases actually start from the same results at the isothermal limit (figure~\ref{Fig11}a). Regardless of the wall temperature and $\PrEc$, the largest transient growth occurs at $\alpha=0$ and $2\leq\beta\leq2.1$ for both ideal and non-ideal gases. In the subcritical and transcritical cases (figure \ref{Fig12} and \ref{Fig13}), when $\PrEc$ is increased, the ideal gas tends to be slightly more stable, while the non-ideal gas becomes more unstable. In fact, due to the Power/Sutherland law (for the transport properties), the results for the ideal gas are weakly dependent on the wall temperature. Notably in figure \ref{Fig13}(d), where $\PrEc=0.07$, an area of $G_{\max}\rightarrow\infty$ stands out. Recall the discussion in \S \ref{Sec4}, the base flow has entered the triangular zone (see figure \ref{Fig5}) and becomes inflectional. Hence, the flow is inviscid unstable and the critical $\Rey$ is reduced considerably (see figure \ref{Fig7}(c)). As a result, a sub-zone of modal growth (near $\beta=0$) in the $\alpha-\beta$ diagram is observed (where $G_{\max}\rightarrow\infty$). For better display of the results, we have limited the color band to $G_{\max}=450$ in figure~\ref{Fig13}. In the supercritical case (figure \ref{Fig14}), the plots are almost symmetrical, indicating the non-ideal gas effects are rather insignificant. The non-ideal gas is only slightly more unstable than the ideal gas. Table \ref{Table4} summarizes the above maximum transient growth $G_{\max}$. With the increase in $\PrEc$, a similar trend as for the modal growth can be observed. Namely, the ideal gas becomes more stable, while the non-ideal gas tends to be more unstable for the subcritical and transcritical case, and more stable for the supercritical case. \red{On the whole, the non-ideal gas effects increase the algebraic instability in all regimes, most prominently in the transcritical regime}.
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth,clip]{Figure_12}
\end{center}
\caption{Contour plot of $G_{\max}$ at $T_w^*=290$ K. $\Rey=1000$. On the left and right side of each subplot, we show the results for non-ideal (RP) and ideal (IG) gases respectively. (a) $\PrEc=0.01$, (b) $\PrEc=0.03$, (c) $\PrEc=0.05$, (d) $\PrEc=0.07$.}
\label{Fig12}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth,clip]{Figure_13}
\end{center}
\caption{Same as figure \ref{Fig12} but for $T_w^*=300$ K.}
\label{Fig13}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth,clip]{Figure_14}
\end{center}
\caption{Same as figure \ref{Fig12} but for $T_w^*=310$ K.}
\label{Fig14}
\end{figure}
\begin{table}
\begin{center}\def~{\hphantom{0}}
\begin{tabular}{lc|ccc}
& ideal gas (IG) & & non-ideal gas (RP) & \\[3pt]
& $T_w^*=290/300/310$ K \hspace{10pt} & \hspace{10pt} $T_w^*=290$ K & $T_w^*=300$ K & $T_w^*=310$ K\\[3pt]
$\PrEc=0.01$\hspace{10pt} & 193.3 & \hspace{10pt}206.5 & 212.2 &201.4\\
$\PrEc=0.03$\hspace{10pt} & 187.9 & \hspace{10pt}231.6 & 262.7 &204.4\\
$\PrEc=0.05$\hspace{10pt} & 182.8 & \hspace{10pt}265.6 & 472.3 &199.5\\
$\PrEc=0.07$\hspace{10pt} & 178.1 & \hspace{10pt}316.7 & $\infty$ & 190.3\\
\end{tabular}
\caption{Maximum transient growth $G_{\max}$ of the perturbations at $\Rey=1000$.}\label{Table4}
\end{center}\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth,clip]{Figure_15}
\end{center}
\caption{Optimal perturbations (a) and the resulting output (b). $\PrEc=0.07$, $\alpha=0$ and $\beta=2$. Only significant components are plotted, namely in (a) $|v^\prime|$, $|w^\prime|$, (b) $|u^\prime|$, $|\rho^\prime|$ and $|T^\prime|$. }
\label{Fig15}
\end{figure}
The typical optimal perturbation and the resulting output are shown in figure \ref{Fig15} at $\PrEc=0.07$, $\alpha=0$ and $\beta=2$. Similar to an incompressible flow, the streamwise vortices and velocity streaks are recovered as the optimal perturbation and the output, respectively. For compressible flows, thermal streaks ($\rho^\prime$ and $T^\prime$) also become significant. Considering the non-ideal gas effects, the subcritical and supercritical cases share similar optimal perturbations as the ideal gas. In the transcritical case, the profiles are strongly influenced by the inflectional base flow and the strong property variations. On the other hand, the output perturbations are almost the same with regard to the $u^\prime$ component, indicating similar dynamic streaks are being generated. The amplitude of the thermal streak is much larger in the transcritical case close to the wall.
We have shown in \S \ref{Sec4-2} that cubic EoS cannot guarantee accurate results for the growth rate if compared to results obtained with the accurate REFPROP EoS. This is also true for the algebraic instability as shown in figure~\ref{Fig16}, depicting $G-t$ curves of the three cases with different EoS at $\PrEc=0.07$, $\alpha=0$ and $\beta=2$. For example, the van der Waals EoS over-predicts $G_{max}$ by 270\% for the transcritical case. In the supercritical case, the non-ideal gas effects are less significant, and the results of all considered EoS are close to each other.
\begin{figure}
\begin{center}
\includegraphics[width=0.65\linewidth]{Figure_16}
\end{center}
\caption{The transient amplification curve $G(t)$ at $\PrEc=0.07$, $\alpha=0$ and $\beta=2$. (a) $T_w^*=290$ K, (b) $T_w^*=300$ K, (c) $T_w^*=310$ K.}
\label{Fig16}
\end{figure}
\red{The main results presented in this section are based on the energy norm: $m_\rho=m_T=1$. When Mack's energy norm is used, figure~\ref{Fig17} provides a comparison for all three regimes with highly non-ideal gas effects ($\PrEc=0.07$, $\alpha=0$). Indeed, the non-ideal gas has a larger algebraic growth in all three cases with Mack's energy norm, while on the other hand, the ideal gas are rather insensitive to different norms. As a result, the conclusion on algebraic growth will not change.}
\begin{figure}
\begin{center}
\includegraphics[width=0.32\linewidth]{Figure_17a}
\includegraphics[width=0.32\linewidth]{Figure_17b}
\includegraphics[width=0.32\linewidth]{Figure_17c}
\end{center}
\caption{Comparison of maximum algebraic growth using Mack's energy norm. Subplots show cases with (a) $T^*_w=290$ K, (b) $T^*_w=300$ K and (c) $T^*_w=310$ K. The other parameters are kept constant: $\alpha=0$, $\Rey=1000$, $\PrEc=0.07$.}
\label{Fig17}
\end{figure}
|
\section{Introduction}
Pushing the limits of sensing technologies is one of the main challenges in modern physics, opening the door to high-precision measurements of fundamental constants as well as applications in many different areas of science. Specifically, the development of highly-sensitive compact magnetic field sensors enables from detecting extremely weak biologically relevant signals to localize geological structures or archaeological sites \cite{review_magnetometers}. In this context, superconducting quantum interference devices (SQUIDs) \cite{review_SQUIDS1,review_SQUIDS2} and atomic \cite{atomic_magnetometers,Kominis2003,Sheng2013,Baumgart2016,MitchellCold,MitchellHot,Kitching,Polzik} and nitrogen-vacancy diamonds \cite{NV1, NV2} magnetometers are the three main approaches that allow achieving, in a non-invasive way, unprecedented sensitivity to extremely small magnetic fields.
In particular, the extraordinary degree of control of ultracold atomic systems \cite{Bloch2008,llibreveronica} makes them ideal platforms for precision measurements \cite{Zhang2016}. There are basically two types of ultracold atomic magnetometers depending on whether the magnetic field drives the internal or the external degrees of freedom of the atoms. The former are typically based on the detection of the Larmor spin precession of optically pumped atoms while the latter encode the magnetic field information in the spatial density profile of the matter wave. Atomic magnetometers with Bose--Einstein condensates (BECs) have been investigated, for instance, by using stimulated Raman transitions \cite{magnetometryRaman}, probing separately the different internal states of a spinor BEC after free fall \cite{magnetometryfreefall}, or measuring the Larmor precession in a spinor BEC \cite{magnetometryLarmor1,magnetometryLarmor2,magnetometryLarmor3,magnetometryLarmor4,magnetometryLarmor5}. In the latter case, sensitivity can be increased by probing spin-squeezed states \cite{spinsqueezed}. In \cite{FeshbachMiscibility}, the possibility of taking profit of Feshbach resonances to use a two-component BEC as a magnetometer was also outlined. Ultracold atomic magnetometers based on detecting density fluctuations in a BEC due to the deformation of the trapping potential have also been demonstrated \cite{magnetometrydensity1,magnetometrydensity2,magnetometrydensity3}.
Ring-shaped potentials for ultracold atoms are a particularly interesting trapping geometry for quantum sensing and atomtronics \cite{atomtronics1,atomtronics2}. Ring potentials are currently implemented by means of a variety of techniques, such as optically plugged magnetic traps \cite{ring1}, static Laguerre-Gauss Beams \cite{ring2}, painting \cite{ring3,ring4} and time-averaged potentials \cite{ring5,ring6,ringTaver} or conical refraction \cite{ring7}. In fact, persistent currents have been observed in BECs confined in annular traps \cite{persistent,QuenchSupercurrent} and it has also been shown that their physical behavior is in close analogy to that of SQUIDs \cite{squid1,squid2,squid3,squid4,squid5,squid6b,squid6,squid7,squid8,squid9}. It has also been suggested \cite{ring5,squid5} that BECs in this trapping geometry could be used as rotation sensors, which have already been realized with superfluid $^3$He \cite{rotationHe} and have been proposed for matter waves based on the Sagnac effect \cite{Sagnac1, Sagnac2, Sagnac3, Sagnac4}.
In this article, we propose to use a BEC trapped in a two-dimensional (2D) ring potential for measuring with high sensitivity non-linear interactions, scalar magnetic fields and rotations. We consider an imbalanced superposition of counter-rotating Orbital Angular Momentum (OAM) modes, whose spatial density distribution presents a minimal line. A weak two-body interaction between the atoms of the BEC leads to a rotation of the minimal atomic density line whose angular frequency is directly related to the strength of such interactions. This phenomenon is somehow reminiscent of the propagation of gray solitons, which originate in repulsively interacting BECs due to a compensation between the kinetic and mean field interaction energies. In this case, however, the minimal density line appears for attractive, repulsive or even non-interacting BECs, and is a consequence of the interference between the counter-propagating modes that takes place due to the circular geometry of the system.
The rest of the paper is organized as follows. In section \ref{sec1} we describe the physical system and we derive an analytical expression that accounts for the rotation of the line of minimal density. In section \ref{sec2}, we take profit of this expression to propose a full experimental protocol to measure the interaction strength, which is proportional to the $s$-wave scattering length. Far from the resonant field or with a dilute enough BEC, the relation between the scattering length and the applied magnetic field given by Feshbach resonances could be exploited to use the system as a novel type of scalar magnetometer. We also outline the possibility of using the system as a rotation sensor. Finally, in section \ref{conclusions} we summarize the main conclusions. In appendix \ref{CoupledEquations} we derive the general equations that govern the dynamics of a BEC carrying OAM in a ring potential, and in appendix \ref{imaging} we give further details about the experimental implementation of the measurement protocol.
\section{Quantum sensing device}
\label{sec1}
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\linewidth]{figure1.pdf}
\caption{Sketch of the physical system under consideration. A BEC formed by $N$ atoms is loaded in an annular trap, with a $p_{1+}$ population of the state $\ket{1,+}$ and $p_{1-}$ population of $\ket{1,-}$. The interference between these two counter-rotating modes yields a minimum line in the probability density. $R$ is the radius of the annulus and $\sigma$ is the width of the radial harmonic potential.}
\label{physicalsystem}
\end{figure}
\subsection{Physical system}
We consider a BEC formed by $N$ atoms of mass $m$ confined in the $z$ direction by a harmonic potential of frequency $\omega_z$ and in the perpendicular plane by an annular trap of radial frequency $\omega$ and radius $R$. We study the system in the limit of strong confinement along the $z$ direction; $\omega_z\gg \omega$. Under this assumption, in the limit $a_za_sn_2\ll 1$, where $a_s$ is the $s$-wave scattering length, $n_2$ the two-dimensional density of the BEC and $a_z=\sqrt{\hbar/(m\omega_z)}$ the harmonic oscillator length along the $z$ direction, the three-dimensional (3D) Gross--Pitaevskii equation (GPE) can be restricted to the $x-y$ plane by considering the profile for the BEC order parameter along the $z$ direction as a Gaussian of width $a_z$, which corresponds to its ground state along this direction \cite{BECbook}. In doing so, the 3D two-body interaction parameter $g_3=(N4\pi\hbar^2a_s)/m$ is transformed to its two-dimensional (2D) form $g_2=(N\sqrt{8\pi}\hbar^2a_s)/(ma_z)$ (note that in these expressions we have taken the BEC wave function to be normalized to 1). Thus, the 2D GPE that we will use to describe the system reads
\begin{equation}
i\hbar\frac{\partial \Psi}{\partial t}=\left[-\frac{\nabla^2}{{{2m}}}+V(r)+g_{2}|\Psi|^2\right]\Psi,
\label{GPE1}
\end{equation}
where $V(r)=\frac{1}{2}m\omega^2(r-R)^2$ is the potential created by the ring. Furthermore, by expressing the distances in units of $\sigma=\sqrt{\frac{\hbar}{m\omega}}$, the energies in units of $\hbar \omega$ and time in units of $1/\omega$, we arrive at the following dimensionless form of the 2D GPE, which is the one that we will use throughout the paper
\begin{equation}
i\frac{\partial \Psi}{\partial t}=H\Psi=\left[-\frac{\nabla^2}{2}+\frac{1}{2}(r-R)^2+g_{2d}|\Psi|^2\right]\Psi,
\label{GPE}
\end{equation}
where all quantities are now expressed in terms of the above defined units and the dimensionless non-linear interaction parameter is given by
\begin{equation}
g_{2d}=Na_s\sqrt{\frac{8\pi m \omega_{z}}{\hbar}}.
\label{g2d}
\end{equation}
The system supports stationary states with a well-defined total OAM $l$ and positive or negative winding number, which we denote as $\ket{l,\pm}$. The OAM eigenstates have the wave functions
\begin{equation}
\braket{\vec{r}|l,\pm}=\phi_{l\pm}(\vec{r})=\phi_{l\pm}(r,\varphi)=f_{l}(r)e^{\pm i l\varphi},
\label{OAMstates}
\end{equation}
where $f_l(r)$ is the corresponding radial part of the wave function.
\subsection{Dynamics in the weakly interacting regime}
Let us consider as initial state an imbalanced superposition of the $\ket{1,+}$ and $\ket{1,-}$ states, with
$n_{1\pm} \equiv p_{1+}-p_{1-}$ being the population imbalance.
Such state could be realized for instance by preparing the BEC in the ground state of the ring, imprinting a $2\pi$ round phase and momentarily breaking the cylindrical symmetry of the potential to induce a coupling between the degenerate states of positive and negative circulation \cite{reversingcirculation,geometricallyinduced} or by directly transferring OAM with a laser beam \cite{OAMLightAtoms}.
Due to parity reasons, the non-linear term in the GPE can only couple OAM states with odd total OAM $l$, see appendix \ref{CoupledEquations} for a more detailed justification. Thus, we can write the total wave function at any time $t$ as
\begin{align}
\Psi(\vec{r},t)&=\sum_{l\,{\rm odd}}\sum_{\beta=\pm} a_{l\beta}(t)\phi_{l\beta}(\vec{r}).
\label{FSM}
\end{align}
Since we focus on the weakly interacting regime, we consider that the only higher energetic states with a relevant role in the dynamics are $\ket{3,+}$ and $\ket{3,-}$. In order to simplify the forthcoming analytical expressions, we assume that the radial part of the wave functions are the ones of the ring potential ground state, i.e. we take $f_l(r)=f_0(r)$ in Eq.~\eqref{OAMstates}. This is an excellent approximation as long as the width of the the density profile of the BEC along the radial direction is much smaller than the radius of the ring, which is always the case in the weakly interacting regime.
The time evolution of the probability amplitudes $a_{l\pm}(t)$ ($l=1,3$) is obtained by substituting \eqref{FSM} into the GPE \eqref{GPE} (see appendix \ref{CoupledEquations} for details)
\begin{equation}
i\frac{d}{dt}
\begin{pmatrix}
a_{1+}\\
a_{1-}\\
a_{3+}\\
a_{3-}
\end{pmatrix}
=H_{\text{FSM}}
\begin{pmatrix}
a_{1+}\\
a_{1-}\\
a_{3+}\\
a_{3-}
\end{pmatrix}.
\label{dynamicsFSM}
\end{equation}
where the four-state model (FSM) Hamiltonian reads
\begin{widetext}
\begin{equation}
H_{\text{FSM}}
=U
\begin{pmatrix}
\mu_1/U&\rho_{1+1-}+\rho^*_{1+3+}+\rho_{1-3-}&\rho^*_{1+1-}+\rho_{1+3+}+\rho^*_{1-3-}
&\rho_{1+3-}+\rho^*_{1-3+}\\
\rho^*_{1+1-}+\rho_{1+3+}+\rho^*_{1-3-}&\mu_1/U&\rho^*_{1+3-}+\rho_{1-3+}
&\rho_{1+1-}+\rho^*_{1+3+}+\rho_{1-3-}\\
\rho_{1+1-}+\rho^*_{1+3+}+\rho_{1-3-}&\rho_{1+3-}+\rho^*_{1-3+}&\mu_3/U&\rho_{3+3-}\\
\rho^*_{1+3-}+\rho_{1-3+}&\rho^*_{1+1-}+\rho_{1+3+}+\rho^*_{1-3-}&\rho^*_{3+3-}&\mu_3/U\\
\end{pmatrix},
\label{hamFSM}
\end{equation}
\end{widetext}
where $\rho_{i\pm j\pm}\equiv a_{i\pm}a^*_{j\pm}$ with $i,j=1,3$ are the density matrix elements, $\mu_{l}$ the chemical potential of the ${l=1,3}$ OAM states, $H\phi_{l\pm}=\mu_{l}\phi_{l\pm}$, and ${U=g_{2d}\int d^2r |f_0(r)|^4}$. From these parameter definitions, the validity condition of the weakly interacting regime reads ${(\mu_3-\mu_1)\equiv \Delta \gg U}$.
Within this regime, Fig.~\ref{dynamics_g1_p+07_v3}(a) shows a typical temporal evolution of the populations of all the OAM states involved in the dynamics considering as initial state an imbalanced superposition of the $\ket{1,+}$ and $\ket{1,-}$ states. The continuous lines have been obtained by solving with a high order Runge-Kutta method the FSM, Eq.~\eqref{dynamicsFSM}, and
the insets show the comparison with the results obtained by a full numerical integration of the 2D GPE (points). We have performed this integration using a standard Crank-Nicolson algorithm in a space-splitting scheme \cite{SpaceSplitting}, i.e., we have introduced the Trotter decomposition $e^{iH(x,y)\Delta t} \approx e^{iH(x)\Delta t} e^{iH(x)\Delta t}$, where $\Delta t$ is the discrete time step, that we have taken to be $\Delta t=10^{-3}$. The grid used for the simulations has a spatial discretization width $\Delta x=2.4\times 10^{-3}$ and a total of 1000 points in each dimension. For all the populations, we find an excellent agreement between the results obtained with the two different methods, with relative discrepancies typically on the order of $10^{-2}$. Despite the fact that the populations of the different OAM states present only very small fluctuations, the initial state is not in general a stationary state of the system because the minimum appearing in the density profile rotates at a constant speed. This fact can be appreciated in Fig.~\ref{dynamics_g1_p+07_v3}(b), where the density profile is shown for different times. At $t=0$, the density profile has a minimum density line at $x=0$, and as time marches on this line rotates in the $x-y$ plane. The fact that the minimum density line rotates means that there is a time-dependent relative phase $\alpha(t)$ between the $a_{1+}(t)$ and $a_{1-}(t)$ coefficients, so that the state of the system evolves in time as $\Psi({\vec{r},t})\approx a_{1+}(0)\phi_{1+}(\vec{r})+a_{1-}(0)e^{i\alpha (t)}\phi_{1-}(\vec{r})$. This phase difference is due to the non-linear interaction, and can be understood as a consequence of the presence of off-diagonal terms in the FSM Hamiltonian \eqref{hamFSM}. In order to determine the time dependence of $\alpha$, in Fig.~\ref{dynamics_g1_p+07_v3}(c) we plot the temporal evolution of the real part of the coherence $\rho_{1+1-}=a_{1+}(t)a^*_{1-}(t)$. We observe that it oscillates harmonically, which means that $\alpha$ evolves linearly with time. The oscillation frequency of the coherence corresponds to the rotation frequency of the minimum density line.
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\linewidth]{figure2.pdf}
\caption{(a) Time evolution of the population of the states involved in the dynamics. (b) Snapshots of the density profile for different instants of the dynamical evolution. (c) Time evolution of the real part of the coherence between the $\ket{1,+}$ and $\ket{1,-}$ states. The points correspond to the numerical simulation of the GPE, while the continuous lines are obtained by solving the FSM equations. The considered parameter values are $R=5$, $g_{2d}=1$, for which $U=0.0128$, ${\mu_1=0.529}$ and $\mu_3=0.699$, $a_{1+}(0)=\sqrt{p_{1+}(0)}=\sqrt{0.7}$ and $a_{1-}(0)=\sqrt{p_{1-}(0)}=\sqrt{0.3}$.}
\label{dynamics_g1_p+07_v3}
\end{figure}
From the FSM, we can obtain the oscillation frequency of $\rho_{1+1-}$ by solving the {von Neumann} equation $i\dot{\rho}=[H_{\text{FSM}},\rho]$. After assuming $\rho_{1+1+}=p_{1+}$ and $\rho_{1-1-}=p_{1-}$ to be constant and neglecting all terms $\mathcal{O}(a^2_{3\pm}(t))$, we arrive at a linear system of three coupled differential equations
\begin{subequations}
\begin{align}
i\frac{d\rho_{1+1-}}{dt}&=Up_{1-}(2\rho^*_{1+3+}+\rho_{1+1-}+\rho_{1-3-})\nonumber\\
&-Up_{1+}(\rho^*_{1+3+}+\rho_{1+1-}+2\rho_{1-3-})\label{eqcoherences1}\\
i\frac{d\rho^*_{1+3+}}{dt}&=Up_{1+}(\rho^*_{1+3+}+\rho_{1+1-}+2\rho_{1-3-})+\Delta\rho^*_{1+3+}\label{eqcoherences2}\\
i\frac{d\rho_{1-3-}}{dt}&=-Up_{1-}(2\rho^*_{1+3+}+\rho_{1+1-}+\rho_{1-3-})-\Delta\rho_{1-3-}.\label{eqcoherences3}
\end{align}
\label{eqscoherences}
\end{subequations}
The characteristic frequencies $k$ of the system of equations \eqref{eqscoherences} are obtained by solving the eigenvalue equation
\begin{equation}
ik^3+ik(U\Delta+\Delta^2-p_{1+}p_{1-}U^2)+U\Delta^2(p_{1+}-p_{1-})=0.
\label{eqfreqs}
\end{equation}
Since $U \ll \Delta$ in the weakly interacting regime, the term proportional to $p_{1+}p_{1-}U^2$ can be neglected in front of the others. The three eigenvalues that are obtained after solving Eq.~\eqref{eqfreqs} are imaginary. The eigenmode associated to the eigenvalue of lowest modulus $k_0$ has a predominant component of $\rho_{1+1-}(t)$, allowing us to write ${\rho_{1+1-}(t)\approx \rho_{1+1-}(0)e^{k_0t}}$. Thus, the rotation frequency of the nodal line is $\Omega_{\text{FSM}}=-\frac{i}{2}k_0$, where the subscript indicates that the rotation frequency has been obtained in the context of the FSM. In the limit ${\Delta \gg \Omega_{FSM}}$, the rotation frequency of the nodal line is given by
\begin{equation}
\Omega_{\text{FSM}}=\frac{U n_{1\pm}}{2(1+\frac{U}{\Delta})}.
\label{omeganode}
\end{equation}
Note that, although the $l=3$ states are nearly not populated during the dynamical evolution, the parameter $\Delta$, which contains the chemical potential $\mu_3$, plays a significant role in the expression of the rotation frequency \eqref{omeganode}. Thus, these states must be taken into account for an accurate description of the dynamics of the system.
\section{Quantum sensing protocol}
\label{sec2}
\subsection{Sensing of two-body interactions}
Recalling that the parameter $U$ of the FSM Hamiltonian \eqref{hamFSM} is given by ${U=g_{2d}\int d^2r |f_0(r)|^4\equiv g_{2d}I}$ and assuming that we are in the regime of validity of the FSM, Eq.~\eqref{omeganode} allows us to express the interaction parameter $g_{2d}$ as
\begin{equation}
g_{2d}=\frac{1}{I}\frac{2\Omega}{n_{1\pm}-2\frac{\Omega}{\Delta}},
\label{eqg2d}
\end{equation}
where $\Omega$ is the observed frequency of rotation of the nodal line. The relation \eqref{eqg2d} constitutes the basis to use the physical system under consideration as a quantum sensing device. By determining the parameters appearing on the right hand side, one can infer the value of $g_{2d}$ and thus, from Eq.~\eqref{g2d}, either the $s$-wave scattering length or the number of atoms forming the BEC.
In Fig.~\ref{plotfreqcomplet}(a), we plot $\Omega$ as a function of $g_{2d}$ for different values of $n_{1\pm}$, computed using \eqref{omeganode} (continuous lines) and the full numerical integration of the 2D GPE (points), showing an excellent agreement between the two methods for low non-linearities and population imbalances. For $g_{2d}<4$, Fig.~\ref{plotfreqcomplet}(b) shows the relative error $\frac{\delta\Omega}{\Omega_{\text{GPE}}}$, where $\Omega_{\text{GPE}}$ is the rotation frequency of the nodal line obtained from the GPE and ${\delta\Omega=|\Omega_{\text{FSM}}-\Omega_{\text{GPE}}|}$, as a function of the \textit{ab initio} values of $n_{1\pm}$ and $g_{2d}$ in the numerical simulation, finding a maximum relative error of $10^{-2}$.
Since all the treatment developed so far is valid for low values of $g_{2d}$, this sensing device could be used for dilute BECs.
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\linewidth]{figure3.pdf}
\caption{(a) Rotation frequency of the nodal line $\Omega$ as a function of $g_{2d}$ for different values of $n_{1\pm}$ obtained with the FSM (continuous lines) and full integration of the GPE (points) (b) Relative error committed in the determination of $\Omega$ using Eq.~\eqref{omeganode} as a function of the \textit{ab initio} values of $g_{2d}$ and $n_{1\pm}$ in the simulation.}
\label{plotfreqcomplet}
\end{figure}
The rotation frequency of the minimum density line, $\Omega$, can be measured by direct imaging in real time of the density distribution of the BEC. If the coherence time of the BEC is $\tau$, in order for this measurement to be possible the condition $\Omega \omega \gtrsim 1/\tau$ must be fulfilled, since otherwise the rotation would be so slow that it could not be appreciated during the time that the experiment lasts. The upper limit of observable relevant values of $\Omega$ is imposed by the regime of validity of the model. If the interaction is too large, the assumptions of the FSM model are no longer valid and it is thus not possible to relate the rotation frequency of the nodal line to the non-linear interaction parameter using \eqref{eqg2d}.
The rest of parameters appearing on the right hand side of \eqref{eqg2d} can be determined experimentally from fluorescence images of the BEC. In Appendix A we design a specific protocol to measure the population imbalance $n_{1\pm}$, the integral of the radial wave function $I$, and the chemical potential difference $\Delta$. Note that currently there are different approaches to measure the $s-$wave scattering length of ultracold atoms \cite{PethickandSmith} such as those based on photoassociation spectroscopy, ballistic expansion, and collective excitations. Our proposal constitutes an alternative to these approaches where all the unknowns can be directly inferred from fluorescence images of the BEC. However, the limit $g_{2d}<4$ obtained for the configuration discussed in Fig.~\ref{plotfreqcomplet} implies that for a BEC of, e.g., $10^4$ atoms of $^{23}$Na, with a trapping frequency $\omega_z$ of a few hundreds of Hz, the maximum $s-$wave scattering length that could be measured with high precision, e.g., with a relative error of $10^{-2}$, would be few times the Bohr radius.
\subsection{Sensing of magnetic fields}
Assuming that the total number of atoms of the BEC $N$ and the trapping frequency in the $z$ direction $\omega_z$ are precisely known quantities, Eqs.~\eqref{eqg2d} and~\eqref{g2d} together with the protocols to measure $n_{1\pm}$, $I$ and $\Delta$ allow to determine the scattering length $a_S$ at zero magnetic field. Alternatively, if the scattering length is a known quantity, the measurements of $\Omega$, $I$ and $\Delta$ can be used to determine $n_{1\pm}$ through the aforementioned relations.
If the scattering length depends somehow on the modulus of the external magnetic field $B$, turning on the field will be translated into a variation of $\Omega$. Thus, the system could be used as a scalar magnetometer by relating changes on the frequency of rotation of the minimal line to variations of the modulus of the magnetic field. Taking into account that $I$ and $\Delta$ are almost independent of $g_{2d}$ and thus of $B$ in the regime of interaction strengths for which the model is valid, combining Eqs. ~\eqref{g2d} and ~\eqref{omeganode} we can evaluate the sensitivity that this magnetometer would have as
\begin{equation}
\frac{d\Omega_{\text{FSM}}}{dB}=\frac{n_{1\pm} I N\sqrt{\frac{8\pi m\omega_z}{\hbar}}}{2(1+\frac{U(B)}{\Delta})^2}\frac{da_S}{dB}.
\label{dOmegadB}
\end{equation}
Since we must have $U\ll \Delta$ in order for the model to be valid, we can define a threshold limit for the sensitivity by taking $U/\Delta=1$ in \eqref{dOmegadB}. Defining the aspect ratio $\Lambda\equiv \omega_z/\omega$ and changing the differentials in \eqref{dOmegadB} by finite increments, we find the following upper threshold for the sensitivity in magnetic field variations $\Delta B_{\text{th}}$ as a function of the change in the rotation frequency of the nodal line
\begin{equation}
\Delta B_{\text{th}} = \frac{8\sigma}{n_{1\pm} I N\sqrt{8\pi \Lambda}} \frac{1}{\frac{da_S}{dB}}\Delta \Omega.
\label{sensitivity}
\end{equation}
From Eq. ~\eqref{sensitivity}, we observe that the sensitivity is improved by having a large number of condensed particles and a strong dependence of the scattering length on the magnetic field modulus. However, since the parameter $g_{2d}\propto Na_s$ needs to be small in order for the model to be valid, it is also required that the scattering length takes small values.
In the presence of a Feshbach resonance, the scattering length depends of the magnetic field modulus as
\begin{equation}
a_S(B)=\tilde{a}_S\left(1-\frac{\delta}{B-B_0}\right),
\label{Feshbach}
\end{equation}
where $\tilde{a}_S$ is the background scattering length, $B_0$ is the value of $B$ at resonance and $\delta$ is the width of the resonance. Thus, by placing the magnetic field close to the resonant value $B_0$, one could in principle meet both the requirement that the scattering length is small and that it depends strongly on the magnetic field modulus. However, in most cases this procedure would have the inconvenience that close to a Feshbach resonance the three-body losses are greatly enhanced, limiting the lifetime of the BEC and hindering the measurement procedure. Nevertheless, some atomic species such as $^{85}$Rb \cite{BECRb85}, $^{133}$Cs \cite{BECCs133}, $^{39}$K \cite{BECK39} or $^7$Li \cite{BECLi7} have been reported to form BECs that are stable across Feshbach resonances, so they could be potential candidates for using the system as a magnetometer. Additionally, the BECs formed by these species have lifetimes on the order of a few seconds. Taking into account that the trapping frequency $\omega$, in units of which $\Omega$ is expressed, is typically of the order of a few hundreds of Hz for ring-shaped traps, and considering typical values of $\Omega$ shown in figure \ref{plotfreqcomplet} (a), in International System units $\Omega\sim 1$Hz. This means that in the typical time that an experiment would last, $\tau\sim 1$s, the minimum density line would perform some complete laps. Under the reasonable assumption that the fluorescence imaging system could resolve angular differences on the order of $\sim 0.1$ rad, incrementals in the rotation frequency on the order of $10^{-2}$Hz could be measured. Thus, in the dimensionless units of Eq.~\eqref{sensitivity}, sensitivites on the order of $\Delta\Omega\sim 10^{-4}$ could be achieved. These atomic species have, however, the drawback that they typically form BECs with a low number of particles, which limits the sensitivity to magnetic fields. Although it is outside of the scope of this paper to give accurate values of the sensitivities that could be achieved with this apparatus, making use of Eqs.~\eqref{sensitivity} and \eqref{Feshbach}, and considering the experimental parameters reported in \cite{squid7}, we have estimated that, in principle, this magnetometer would allow to measure changes in the magnetic field on the order of a few pT at a bandwidth of 1 Hz.
As a last remark, we point out that after measuring the scattering length, far from the resonant field $B_0$, if the line of minimal density rotates at a constant speed the relation \eqref{Feshbach} can be inverted to infer the absolute value of the magnetic field.
\subsection{Sensing of rotations}
Let us consider the case when the BEC is placed in a reference frame rotating at an angular frequency $\Omega_{\text{ext}}$, which is positive (negative) if the rotation is clockwise (counter-clockwise). Now the dynamics is governed by the modified GPE
\begin{equation}
i\frac{\partial \Psi}{\partial t}=\left[-\frac{\nabla^2}{2}+V(r)+g_{2d}|\Psi|^2-\Omega_{\text{ext}} L_z\right]\Psi,
\label{GPErotation}
\end{equation}
where $L_z=-i\frac{\partial}{\partial\varphi}$ is the $z$ component of the angular momentum operator. The ideal instance for using the system under study as a sensor of rotations is the non-interacting limit $g_{2d}=0$. In that case, it can be easily shown that the effect of the external rotation is to make the line of minimal density rotate at an angular speed $\Omega_{\text{ext}}$, which can be directly measured in experiments.
In the weakly interacting regime, the system under study can still be used as a sensor of external rotations. In that case, we find that the only difference in the dynamics with respect to the case when there is no external rotation is that the rotation frequency of the nodal line is shifted precisely by a quantity $\Omega_{\text{ext}}$. Thus, if $g_{2d}$ is known and $I$, $n_{1\pm}$ and $\Delta$ are measured using the protocol provied in the appendix A, the system under consideration can be used as a sensing device for external rotations by computing the external rotation as $\Omega_{\text{ext}}=\Omega-\Omega_{\text{FSM}}$, where $\Omega$ is the rotation frequency of the nodal line observed in the experiment and $\Omega_{\text{FSM}}$ is given by \eqref{omeganode}.
The proposed setup constitutes an alternative to the two main lines of development of rotation sensors using ultracold atoms: the atomic-gas analogues of superconducting quantum interference devices (SQUIDs)
\cite{squid1,squid2,squid3,squid4,squid5,squid6b} and the Sagnac interferometers, for a review see \cite{Sagnac1}. Gyroscopes based on the Sagnac effect measure a rotation rate relative to an inertial reference frame, based on a rotationally induced phase shift between two paths of an interferometer and the low available atomic fluxes and low effective areas are the main limiting factors of their sensitivity.
\section{Conclusions}
\label{conclusions}
We have studied the dynamics of an imbalanced superposition of the two degenerate counter-rotating $l=1$ OAM modes of a weakly interacting BEC trapped in a 2D ring potential. We have found that the non-linear interaction induces a time-dependent phase difference between these two modes which leads to a rotation of the line of minimal atomic density of the BEC. The derived few state model provides a simple analytical dependence between the rotation frequency and the non-linear parameter which, for low non-linearities, perfectly matches with the \textit{ab initio} numerical simulations. The measurement of the rotation frequency allows to use the system as a quantum sensor of two-body interactions, scalar magnetic fields and rotations. The theoretical treatment exposed in this work can also be extended to a regime of higher interactions, where higher OAM modes are excited and a myriad of new physical scenarios opens up.
\acknowledgements
We thank M. W. Mitchell for fruitful and stimulating discussions. We acknowledge support from the Spanish Ministry of
Economy and Competitiveness under Contract No. FIS2014-57460-P and FIS2017-86530-P, and from the Catalan Government under Contract No. SGR2014-1639 and SGR2017-1646. G.P. also acknowledges financial support from the FPI Grant No. BES-2015-073772.
|
\section{Introduction and useful informations}
\subsection{Background}
The classical fractional integral (The classical fractional integral
operator is also known as Riesz potential.) was introduced by Riesz in 1949
\cite{Riesz}, defined by
\begin{eqnarray*}
I_{\alpha }f(x) &=&\left( -\Delta \right) ^{-\frac{\alpha }{2}}f(x)\qquad
0<\alpha <n, \\
&=&\frac{1}{\gamma \left( \alpha \right) }\int \limits_{{\mathbb{R}^{n}}
\frac{f(y)}{|x-y|^{n-\alpha }}dy
\end{eqnarray*
with
\begin{equation*}
\gamma \left( \alpha \right) =\frac{\pi ^{\frac{n}{2}}2^{\alpha }\Gamma
\left( \frac{\alpha }{2}\right) }{\Gamma \left( \frac{n}{2}-\frac{\alpha }{2
\right) },
\end{equation*
where $\Gamma \left( \cdot \right) $ is the standard gamma function and
I_{\alpha }$ plays an important role in partial diferential equation as the
inverse of power of Laplace operator. Especially, Its most significant
feature is that $I_{\alpha }$ maps $L_{p}({\mathbb{R}^{n}})$ continuously
into $L_{q}({\mathbb{R}^{n}})$, with $\frac{1}{q}=\frac{1}{p}-\frac{\alpha }
n}$ and $1<p<\frac{n}{\alpha }$, through the well known
Hardy-Littlewood-Sobolev imbedding theorem (see pp. 119-121,Theorem 1 and
its proof in \cite{Stein}) for $I_{\alpha }$.
Let $\Omega \in L_{s}(S^{n-1})$, $1<s\leq \infty $, $\Omega (\mu x)=\Omega
(x)~$for any$~\mu >0$, $x\in {\mathbb{R}^{n}}\setminus \{0\}$ and satisfy
the cancellation condition
\begin{equation*}
\int \limits_{S^{n-1}}\Omega (x^{\prime })d\sigma (x^{\prime })=0,
\end{equation*
where $x^{\prime }=\frac{x}{|x|}$ for any $x\neq 0$.
We first recall the definitions of rough fractional integral operator
\overline{T}_{\Omega ,\alpha }$ and a related rough fractional maximal
operator $M_{\Omega ,\alpha }$.
\begin{definition}
Defin
\begin{equation*}
I_{\Omega ,\alpha }f(x)=\int \limits_{{\mathbb{R}^{n}}}\frac{\Omega (x-y)}
|x-y|^{n-\alpha }}f(y)dy\qquad 0<\alpha <n,
\end{equation*
\begin{equation*}
M_{\Omega ,\alpha }f\left( x\right) =\sup_{r>0}\frac{1}{r^{n-\alpha }}\dint
\limits_{|x-y|<r}\left \vert \Omega (x-y)\right \vert \left \vert f(y)\right
\vert dy\qquad 0<\alpha <n.
\end{equation*}
\end{definition}
Next, we give the definition of weighted Lebesgue spaces as follows:
\begin{definition}
$\left( \text{\textbf{Weighted Lebesgue space}}\right) $ Let $1\leq p\leq
\infty $ and given a weight $w\left( x\right) \in A_{p}\left( {{\mathbb{R
^{n}}}\right) $, we shall define weighted Lebesgue spaces as
\begin{eqnarray*}
L_{p}(w) &\equiv &L_{p}({{\mathbb{R}^{n}}},w)=\left \{ f:\Vert f\Vert
_{L_{p,w}}=\left( \dint \limits_{{{\mathbb{R}^{n}}}}|f(x)|^{p}w(x)dx\right)
^{\frac{1}{p}}<\infty \right \} ,\qquad 1\leq p<\infty . \\
L_{\infty ,w} &\equiv &L_{\infty }({{\mathbb{R}^{n}}},w)=\left \{ f:\Vert
f\Vert _{L_{\infty ,w}}=\limfunc{esssup}\limits_{x\in {\mathbb{R}^{n}
}|f(x)|w(x)<\infty \right \} .
\end{eqnarray*}
\end{definition}
Here and later, $A_{p}$ denotes the Muckenhoupt classes (see \cite{Gurbuz}).
Now, let us consider the Muckenhoupt-Wheeden class $A\left( p,q\right) $ in
\cite{Muckenhoupt}. One says that $w\left( x\right) \in A\left( p,q\right) $
for $1<p<q<\infty $ if and only if
\begin{equation}
\lbrack w]_{A\left( p,q\right) }:=\sup \limits_{B}\left( |B|^{-1}\dint
\limits_{B}w(x)^{q}dx\right) ^{\frac{1}{q}}\left( |B|^{-1}\dint
\limits_{B}w(x)^{-p^{\prime }}dx\right) ^{\frac{1}{p^{\prime }}}<\infty ,
\label{13}
\end{equation
where the supremum is taken over all the balls $B$. Note that, by H\"{o
lder's inequality, for all balls $B$ we have
\begin{equation}
\lbrack w]_{A\left( p,q\right) }\geq \lbrack w]_{A\left( p,q\right)
(B)}=|B|^{\frac{1}{p}-\frac{1}{q}-1}\Vert w\Vert _{L_{q}(B)}\Vert
w^{-1}\Vert _{L_{p^{\prime }}(B)}\geq 1. \label{14}
\end{equation}
By (\ref{13}), we hav
\begin{equation}
\left( \dint \limits_{B}w(x)^{q}dx\right) ^{\frac{1}{q}}\left( \dint
\limits_{B}w(x)^{-p^{\prime }}dx\right) ^{\frac{1}{p^{\prime }}}\lesssim
\left \vert B\right \vert ^{\frac{1}{q}+\frac{1}{p^{\prime }}}. \label{16}
\end{equation
On the other hand, let $\mu \left( x\right) =w\left( x\right) ^{s^{\prime }}
, $\tilde{p}=\frac{p}{s^{\prime }}$ and $\tilde{q}=\frac{q}{s^{\prime }}$.
If $w\left( x\right) ^{s^{\prime }}\in A\left( \frac{p}{s^{\prime }},\frac{
}{s^{\prime }}\right) $, then we get $\mu \left( x\right) \in A\left( \tilde
p},\tilde{q}\right) $. By (\ref{14}) and (\ref{16}),
\begin{equation}
\Vert \mu \Vert _{L_{\tilde{q}}(B)}\Vert \mu ^{-1}\Vert _{L_{\tilde{p
^{\prime }}(B)}\approx |B|^{1+\frac{1}{\tilde{q}}-\frac{1}{\tilde{p}}}
\label{fg}
\end{equation
is valid.
Now, we introduce some spaces which play important roles in PDE. Except the
weighted Lebesgue space $L_{p}(w)$, the weighted Morrey space $L_{p,\kappa
}(w)$, which is a natural generalization of $L_{p}(w)$ is another important
function space. Then, the definition of generalized weighted Morrey spaces
M_{p,\varphi }\left( w\right) $ which could be viewed as extension of
L_{p,\kappa }(w)$ has been given as follows:
For $1\leq p<\infty $, positive measurable function $\varphi (x,r)$ on $
\mathbb{R}^{n}}\times (0,\infty )$ and nonnegative measurable function $w$
on ${\mathbb{R}^{n}}$, $f\in M_{p,\varphi }(w)\equiv M_{p,\varphi }({\mathbb
R}^{n}},w)$ if $f\in L_{p,w}^{loc}({\mathbb{R}^{n}})$ and
\begin{equation*}
\Vert f\Vert _{M_{p,\varphi }(w)}=\sup \limits_{x\in {\mathbb{R}^{n}},r>0
\frac{1}{\varphi (x,r)}\Vert f\Vert _{L_{p}(B(x,r),w)}<\infty .
\end{equation*
is finite. Note that for $\varphi (x,r)\equiv w(B(x,r))^{\frac{\kappa }{p}}
, $0<\kappa <1$ and $\varphi (x,r)\equiv 1$, we have $M_{p,\varphi
}(w)=L_{p,\kappa }(w)$ and $M_{p,\varphi }(w)=L_{p}(w)$, respectively.
Extending the definition of vanishing generalized Morrey spaces in \cit
{Gurbuz1} to the case of generalized weighted Morrey spaces defined above,
we introduce the following definition.
\begin{definition}
\textbf{(Vanishing generalized weighted Morrey spaces) }For\textbf{\ }$1\leq
p<\infty $, $\varphi (x,r)$ is a positive measurable function on ${\mathbb{R
^{n}}\times (0,\infty )$ and nonnegative measurable function $w$ on $
\mathbb{R}^{n}}$, $f\in VM_{p,\varphi }\left( w\right) \equiv VM_{p,\varphi
}({\mathbb{R}^{n},w})$ if $f\in L_{p,w}^{loc}({\mathbb{R}^{n}})$ an
\begin{equation}
\lim \limits_{r\rightarrow 0}\sup \limits_{x\in {\mathbb{R}^{n}}}\frac{1}
\varphi (x,r)}\Vert f\Vert _{L_{p}(B(x,r),w)}=0. \label{1*}
\end{equation}
\end{definition}
Inherently, it is appropriate to impose on $\varphi (x,t)$ with the
following circumstances:
\begin{equation}
\lim_{t\rightarrow 0}\sup \limits_{x\in {\mathbb{R}^{n}}}\frac{\left(
w(B(x,t))\right) ^{^{\frac{1}{p}}}}{\varphi (x,t)}=0, \label{2}
\end{equation
an
\begin{equation}
\inf_{t>1}\sup \limits_{x\in {\mathbb{R}^{n}}}\frac{\left( w(B(x,t))\right)
^{^{\frac{1}{p}}}}{\varphi (x,t)}>0. \label{3}
\end{equation
From (\ref{2}) and (\ref{3}), we easily know that the bounded functions with
compact support belong to $VM_{p,\varphi }\left( w\right) $. On the other
hand, the space $VM_{p,\varphi }(w)$ is Banach space with respect to the
following finite quasi-nor
\begin{equation*}
\Vert f\Vert _{VM_{p,\varphi }(w)}=\sup \limits_{x\in {\mathbb{R}^{n}},r>0
\frac{1}{\varphi (x,r)}\Vert f\Vert _{L_{p}(B(x,r),w)},
\end{equation*
such tha
\begin{equation*}
\lim \limits_{r\rightarrow 0}\sup \limits_{x\in {\mathbb{R}^{n}}}\frac{1}
\varphi (x,r)}\Vert f\Vert _{L_{p}(B(x,r),w)}=0,
\end{equation*
we omit the details. Moreover, we have the following embeddings
\begin{equation*}
VM_{p,\varphi }\left( w\right) \subset M_{p,\varphi }\left( w\right) ,\qquad
\Vert f\Vert _{M_{p,\varphi }\left( w\right) }\leq \Vert f\Vert
_{VM_{p,\varphi }\left( w\right) }.
\end{equation*
Henceforth, we denote by $\varphi \in \mathcal{B}\left( w\right) $ if
\varphi (x,r)$ is a positive measurable function on ${\mathbb{R}^{n}}\times
(0,\infty )$ and positive for all $(x,r)\in {\mathbb{R}^{n}}\times (0,\infty
)$ and satisfies (\ref{2}) and (\ref{3}).
The purpose of this paper is to consider the mapping properties for the
rough fractional type sublinear operators $T_{\Omega ,\alpha }$ satisfying
the following conditio
\begin{equation}
|T_{\Omega ,\alpha }f(x)|\lesssim \int \limits_{{\mathbb{R}^{n}}}\frac
|\Omega (x-y)|}{|x-y|^{n-\alpha }}\,|f(y)|\,dy,\qquad x\notin \text{supp }f
\label{e1}
\end{equation
on vanishing generalized weighted Morrey spaces. Similar results still hold
for the operators $I_{\Omega ,\alpha }$ and $M_{\Omega ,\alpha }$,
respectively. On the other hand, these operators have not also been studied
so far on vanishing generalized weighted Morrey spaces and this paper seems
to be the first in this direction.
At last, here and henceforth, $F\approx G$ means $F\gtrsim G\gtrsim F$;
while $F\gtrsim G$ means $F\geq CG$ for a constant $C>0$; and $p^{\prime }$
and $s^{\prime }$ always denote the conjugate index of any $p>1$ and $s>1$,
that is, $\frac{1}{p^{\prime }}:=1-\frac{1}{p}$ and $\frac{1}{s^{\prime }
:=1-\frac{1}{s}$ and also $C$ stands for a positive constant that can change
its value in each statement without explicit mention. Throughout the paper
we assume that $x\in {\mathbb{R}^{n}}$ and $r>0$ and also let $B(x,r)$
denotes $x$-centred Euclidean ball with radius $r$, $B^{C}(x,r)$ denotes its
complement. For any set $E$, $\chi _{_{E}}$ denotes its characteristic
function, if $E$ is also measurable and $w$ is a weight, $w(E):=\dint
\limits_{E}w(x)dx$.
\section{Main Results}
Our result can be stated as follows.
\begin{theorem}
\label{teo1}Suppose that $0<\alpha <n$, $1\leq s^{\prime }<p<\frac{n}{\alpha
}$, $\frac{1}{q}=\frac{1}{p}-\frac{\alpha }{n}$, $1<q<\infty $, $\Omega \in
L_{s}(S^{n-1})$, $1<s\leq \infty $, $\Omega (\mu x)=\Omega (x)~$for any$~\mu
>0$, $x\in {\mathbb{R}^{n}}\setminus \{0\}$ such that $T_{\Omega ,\alpha }$
is rough fractional type sublinear operator satisfying (\ref{e1}). For $p>1
, $w\left( x\right) ^{s^{\prime }}\in A\left( \frac{p}{s^{\prime }},\frac{q}
s^{\prime }}\right) $ and $s^{\prime }<p$, the following pointwise estimat
\begin{equation}
\left \Vert T_{\Omega ,\alpha }f\right \Vert _{L_{q}\left( B\left(
x_{0},r\right) ,w^{q}\right) }\lesssim \left( w^{q}\left( B\left(
x_{0},r\right) \right) \right) ^{\frac{1}{q}}\int \limits_{2r}^{\infty
}\left \Vert f\right \Vert _{L_{p}\left( B\left( x_{0},t\right)
,w^{p}\right) }\left( w^{q}\left( B\left( x_{0},t\right) \right) \right) ^{
\frac{1}{q}}\frac{dt}{t} \label{5}
\end{equation
holds for any ball $B\left( x_{0},r\right) $ and for all $f\in
L_{p,w}^{loc}\left( {\mathbb{R}^{n}}\right) $. If $\varphi _{1}\in \mathcal{
}\left( w^{p}\right) $, $\varphi _{2}\in \mathcal{B}\left( w^{q}\right) $
and the pair $\left( \varphi _{1},\varphi _{2}\right) $ satisfies the
following conditions
\begin{equation}
c_{\delta }:=\dint \limits_{\delta }^{\infty }\sup_{x\in {\mathbb{R}^{n}}
\frac{\varphi _{1}\left( x,t\right) }{\left( w^{q}\left( B\left( x,t\right)
\right) \right) ^{\frac{1}{q}}}\frac{1}{t}dt<\infty \label{6}
\end{equation
for every $\delta >0$, and
\begin{equation}
\int \limits_{r}^{\infty }\frac{\varphi _{1}\left( x,t\right) }{\left(
w^{q}\left( B\left( x,t\right) \right) \right) ^{\frac{1}{q}}}\frac{1}{t
dt\lesssim \frac{\varphi _{2}(x,r)}{\left( w^{q}\left( B\left( x,t\right)
\right) \right) ^{\frac{1}{q}}}, \label{7}
\end{equation}
then for $p>1$, $w\left( x\right) ^{s^{\prime }}\in A\left( \frac{p}
s^{\prime }},\frac{q}{s^{\prime }}\right) $ and $s^{\prime }<p$, the
operator $T_{\Omega ,\alpha }$ is bounded from $VM_{p,\varphi _{1}}\left(
w^{p}\right) $ to $VM_{q,\varphi _{2}}\left( w^{q}\right) $. Moreover
\begin{equation}
\left \Vert T_{\Omega ,\alpha }f\right \Vert _{VM_{q,\varphi _{2}}\left(
w^{q}\right) }\lesssim \left \Vert f\right \Vert _{VM_{p,\varphi _{1}}\left(
w^{p}\right) }\cdot \label{8}
\end{equation}
\end{theorem}
\begin{proof}
Since inequality (\ref{5}) is the heart of the proof of (\ref{8}), we first
prove (\ref{5}).
For any $x_{0}\in {\mathbb{R}^{n}}$, we write as $f=f_{1}+f_{2}$, where
f_{1}\left( y\right) =f\left( y\right) \chi _{B\left( x_{0},2r\right)
}\left( y\right) $, $f_{2}\left( y\right) =f\left( y\right) \chi _{\left(
B\left( x_{0},2r\right) \right) ^{C}}\left( y\right) $, $r>0$ and $\chi
_{B\left( x_{0},2r\right) }$ denotes the characteristic function of $B\left(
x_{0},2r\right) $. The
\begin{equation*}
\left \Vert T_{\Omega ,\alpha }f\right \Vert _{L_{q}\left( w^{q},B\left(
x_{0},r\right) \right) }\leq \left \Vert T_{\Omega ,\alpha }f_{1}\right
\Vert _{L_{q}\left( w^{q},B\left( x_{0},r\right) \right) }+\left \Vert
T_{\Omega ,\alpha }f_{2}\right \Vert _{L_{q}\left( w^{q},B\left(
x_{0},r\right) \right) }.
\end{equation*
Let us estimate $\left \Vert T_{\Omega ,\alpha }f_{1}\right \Vert
_{L_{q}\left( w^{q},B\left( x_{0},r\right) \right) }$ and $\left \Vert
T_{\Omega ,\alpha }f_{2}\right \Vert _{L_{q}\left( w^{q},B\left(
x_{0},r\right) \right) }$, respectively.
Since $f_{1}\in L_{p}\left( w^{p},{\mathbb{R}^{n}}\right) $, by the
boundedness of $T_{\Omega ,\alpha }$ from $L_{p}\left( w^{p},{\mathbb{R}^{n}
\right) $ to $L_{q}\left( w^{q},{\mathbb{R}^{n}}\right) $ (see Theorem 3.4.2
in \cite{Lu}), (\ref{fg}) and since $1\leq s^{\prime }<p<q$ we get
\begin{eqnarray*}
\left \Vert T_{\Omega ,\alpha }f_{1}\right \Vert _{L_{q}\left( w^{q},B\left(
x_{0},r\right) \right) } &\leq &\left \Vert T_{\Omega ,\alpha }f_{1}\right
\Vert _{L_{q}\left( w^{q},{\mathbb{R}^{n}}\right) } \\
&\lesssim &\left \Vert f_{1}\right \Vert _{L_{p}\left( w^{p},{\mathbb{R}^{n}
\right) } \\
&=&\left \Vert f\right \Vert _{L_{p}\left( w^{p},B\left( x_{0},2r\right)
\right) } \\
&\lesssim &r^{n-\alpha s^{\prime }}\left \Vert f\right \Vert _{L_{p}\left(
w^{p},B\left( x_{0},2r\right) \right) }\dint \limits_{2r}^{\infty }\frac{dt}
t^{n-\alpha s^{\prime }+1}} \\
&\approx &\Vert w^{s^{\prime }}\Vert _{L_{\frac{q}{s^{\prime }}}(B\left(
x_{0},r\right) )}\Vert w^{-s^{\prime }}\Vert _{L_{\left( \frac{p}{s^{\prime
}\right) ^{\prime }}(B\left( x_{0},r\right) )}\dint \limits_{2r}^{\infty
}\left \Vert f\right \Vert _{L_{p}\left( w^{p},B\left( x_{0},t\right)
\right) }\frac{dt}{t^{n-\alpha s^{\prime }+1}} \\
&\lesssim &\left( w^{q}\left( B(x_{0},r)\right) \right) ^{\frac{1}{q}}\dint
\limits_{2r}^{\infty }\left \Vert f\right \Vert _{L_{p}\left( w^{p},B\left(
x_{0},t\right) \right) }\Vert w^{-s^{\prime }}\Vert _{L_{\left( \frac{p}
s^{\prime }}\right) ^{\prime }}(B\left( x_{0},t\right) )}\frac{dt}
t^{n-\alpha s^{\prime }+1}} \\
&\lesssim &\left( w^{q}\left( B(x_{0},r)\right) \right) ^{\frac{1}{q}}\dint
\limits_{2r}^{\infty }\left \Vert f\right \Vert _{L_{p}\left( w^{p},B\left(
x_{0},t\right) \right) }\left[ \Vert w^{s^{\prime }}\Vert _{L_{\left( \frac{
}{s^{\prime }}\right) }(B\left( x_{0},t\right) )}\right] ^{-1}\frac{1}{t}dt
\\
&\lesssim &\left( w^{q}\left( B(x_{0},r)\right) \right) ^{\frac{1}{q}} \\
&&\times \dint \limits_{2r}^{\infty }\left \Vert f\right \Vert _{L_{p}\left(
w^{p},B(x_{0},t)\right) }\left( w^{q}\left( B(x_{0},t)\right) \right) ^{
\frac{1}{q}}\frac{1}{t}dt.
\end{eqnarray*
Now, let's estimate the second part ($=\left \Vert T_{\Omega ,\alpha
}f_{2}\right \Vert _{L_{q}\left( w^{q},B\left( x_{0},r\right) \right) }$).
For the estimate used in $\left \Vert T_{\Omega ,\alpha }f_{2}\right \Vert
_{L_{q}\left( w^{q},B\left( x_{0},r\right) \right) }$, we first have to
prove the below inequality
\begin{equation}
\left \vert T_{\Omega ,\alpha }f_{2}\left( x\right) \right \vert \lesssim
\dint \limits_{2r}^{\infty }\left \Vert f\right \Vert _{L_{p}\left(
w^{p},B(x_{0},t)\right) }\left( w^{q}\left( B(x_{0},t)\right) \right) ^{
\frac{1}{q}}\frac{1}{t}dt. \label{11}
\end{equation}
By \cite{Balak} (see pp. 7 in the proof of Lemma2:), we ge
\begin{equation}
\left \vert T_{\Omega ,\alpha }f_{2}\left( x\right) \right \vert \lesssim
\int \limits_{2r}^{\infty }\left \Vert \Omega \left( x-\cdot \right) \right
\Vert _{L_{s}\left( B\left( x_{0},t\right) \right) }\left \Vert f\right
\Vert _{L_{s^{\prime }}\left( B\left( x_{0},t\right) \right) }\frac{dt}
t^{n+1-\alpha }}. \label{310}
\end{equation
On the other hand, by H\"{o}lder's inequality we hav
\begin{eqnarray}
\left \Vert f\right \Vert _{L_{s^{\prime }}\left( B\left( x_{0},t\right)
\right) } &=&\left( \dint \limits_{B\left( x_{0},t\right) }\left \vert
f\left( y\right) \right \vert ^{s^{\prime }}dy\right) ^{\frac{1}{s^{\prime }
} \notag \\
&\leq &\left( \dint \limits_{B\left( x_{0},t\right) }\left \vert f\left(
y\right) \right \vert ^{p}\left \vert \mu \left( y\right) \right \vert ^
\widetilde{p}}dy\right) ^{\frac{1}{p}}\left( \dint \limits_{B\left(
x_{0},t\right) }\left \vert \mu \left( y\right) \right \vert ^{-\widetilde{p
^{\prime }}dy\right) ^{\frac{1}{\widetilde{p}^{\prime }s^{\prime }}} \notag
\\
&\leq &\left( \dint \limits_{B\left( x_{0},t\right) }\left \vert f\left(
y\right) \right \vert ^{p}\left \vert \mu \left( y\right) \right \vert ^
\widetilde{p}}dy\right) ^{\frac{1}{p}}\left( w^{q}\left( B(x_{0},t)\right)
\right) ^{-\frac{1}{q}}|B(x_{0},t)|^{\frac{1}{s^{\prime }}+\frac{1}{q}-\frac
1}{p}} \notag \\
&=&\left \Vert f\right \Vert _{L_{p}\left( w^{p},B(x_{0},t)\right) }\left(
w^{q}\left( B(x_{0},t)\right) \right) ^{-\frac{1}{q}}|B(x_{0},t)|^{\frac{1}
s^{\prime }}+\frac{1}{q}-\frac{1}{p}}, \label{36}
\end{eqnarray
where in the second inequality we have used the following fact:
By (\ref{fg}), we get the following
\begin{eqnarray}
\left( \dint\limits_{B\left( x_{0},t\right) }\left\vert \mu \left( y\right)
\right\vert ^{-\widetilde{p}^{\prime }}dy\right) ^{\frac{1}{\widetilde{p
^{\prime }s^{\prime }}} &\approx &\left[ \Vert \mu \Vert _{L_{\tilde{q
}(B\left( x_{0},t\right) )}\right] ^{-\frac{1}{s^{\prime }}}\left[ |B\left(
x_{0},t\right) |^{1+\frac{1}{\tilde{q}}-\frac{1}{\tilde{p}}}\right] ^{\frac{
}{s^{\prime }}} \notag \\
&=&\left[ \left( \Vert w^{s^{\prime }}\Vert _{L_{\tilde{q}}(B\left(
x_{0},t\right) )}\right) ^{-1}|B\left( x_{0},t\right) |^{1+\frac{1}{\tilde{q
}-\frac{1}{\tilde{p}}}\right] ^{\frac{1}{s^{\prime }}} \notag \\
&=&\left[ \left( \dint\limits_{B\left( x_{0},t\right) }\left\vert w\left(
y\right) \right\vert ^{q}dy\right) ^{-\frac{s^{\prime }}{q}}|B\left(
x_{0},t\right) |^{1+\frac{s^{\prime }}{q}-\frac{s^{\prime }}{p}}\right] ^
\frac{1}{s^{\prime }}} \notag \\
&=&\left( w^{q}\left( B(x_{0},t)\right) \right) ^{-\frac{1}{q}}|B(x_{0},t)|^
\frac{1}{s^{\prime }}+\frac{1}{q}-\frac{1}{p}}. \label{100}
\end{eqnarray
At last, substituting (3.10) in \cite{Balak} and (\ref{36}) into (\ref{310
), the proof of (\ref{11}) is completed. Thus, by (\ref{11}) we get
\begin{eqnarray*}
\left\Vert T_{\Omega ,\alpha }f_{2}\right\Vert _{L_{q}\left( w^{q},B\left(
x_{0},r\right) \right) } &\lesssim &\left( w^{q}\left( B(x_{0},r)\right)
\right) ^{\frac{1}{q}} \\
&&\times \dint\limits_{2r}^{\infty }\left\Vert f\right\Vert _{L_{p}\left(
w^{p},B(x_{0},t)\right) }\left( w^{q}\left( B(x_{0},t)\right) \right) ^{
\frac{1}{q}}\frac{1}{t}dt.
\end{eqnarray*
Combining all the estimates for $\left\Vert T_{\Omega ,\alpha
}f_{1}\right\Vert _{L_{q}\left( w^{q},B\left( x_{0},r\right) \right) }$ and
\left\Vert T_{\Omega ,\alpha }f_{2}\right\Vert _{L_{q}\left( w^{q},B\left(
x_{0},r\right) \right) }$, we get (\ref{5}).
Now, let's estimate the second part (\ref{8}) of Theorem \ref{teo1}. Indeed,
by the definition of vanishing generalized weighted Morrey spaces, (\ref{5})
and (\ref{7}), we hav
\begin{eqnarray*}
\Vert T_{\Omega ,\alpha }f\Vert _{VM_{q,\varphi _{2}}\left( w^{q}\right) }
&=&\sup \limits_{x\in {\mathbb{R}^{n}},r>0}\frac{1}{\varphi _{2}(x,r)}\Vert
T_{\Omega ,\alpha }f\Vert _{L_{q}\left( w^{q},B\left( x_{0},r\right) \right)
} \\
&\lesssim &\sup \limits_{x\in {\mathbb{R}^{n}},r>0}\frac{1}{\varphi _{2}(x,r
}\left( w^{q}\left( B\left( x_{0},r\right) \right) \right) ^{\frac{1}{q}} \\
&&\times \int \limits_{r}^{\infty }\left \Vert f\right \Vert _{L_{p}\left(
B\left( x_{0},t\right) ,w^{p}\right) }\left( w^{q}\left( B\left(
x_{0},t\right) \right) \right) ^{-\frac{1}{q}}\frac{dt}{t} \\
&\lesssim &\sup \limits_{x\in {\mathbb{R}^{n}},r>0}\frac{1}{\varphi _{2}(x,r
}\left( w^{q}\left( B\left( x_{0},r\right) \right) \right) ^{\frac{1}{q}} \\
&&\times \int \limits_{r}^{\infty }\left( w^{q}\left( B\left( x_{0},t\right)
\right) \right) ^{-\frac{1}{q}}\varphi _{1}\left( x,t\right) \left[ \varphi
_{1}\left( x,t\right) ^{-1}\left \Vert f\right \Vert _{L_{p}\left( B\left(
x_{0},t\right) ,w^{p}\right) }\right] \frac{dt}{t} \\
&\lesssim &\left \Vert f\right \Vert _{VM_{p,\varphi _{1}}\left(
w^{p}\right) }\sup \limits_{x\in {\mathbb{R}^{n}},r>0}\frac{1}{\varphi
_{2}(x,r)}\left( w^{q}\left( B\left( x_{0},r\right) \right) \right) ^{\frac{
}{q}} \\
&&\times \int \limits_{r}^{\infty }\left( w^{q}\left( B\left( x_{0},t\right)
\right) \right) ^{-\frac{1}{q}}\varphi _{1}\left( x,t\right) \frac{dt}{t} \\
&\lesssim &\left \Vert f\right \Vert _{VM_{p,\varphi _{1}}\left(
w^{p}\right) }.
\end{eqnarray*
At last, we need to prove tha
\begin{equation*}
\lim \limits_{r\rightarrow 0}\sup \limits_{x\in {\mathbb{R}^{n}}}\frac{1}
\varphi _{2}(x,r)}\Vert T_{\Omega ,\alpha }f\Vert _{L_{q}\left(
w^{q},B\left( x_{0},r\right) \right) }\lesssim \lim \limits_{r\rightarrow
0}\sup \limits_{x\in {\mathbb{R}^{n}}}\frac{1}{\varphi _{1}(x,r)}\Vert
f\Vert _{L_{p}\left( w^{p},B\left( x_{0},r\right) \right) }=0.
\end{equation*
But, because the proof of above inequality is similar to Theorem 2 in \cit
{Gurbuz1}, we omit the details, which completes the proof.
\end{proof}
\begin{corollary}
Under the conditions of Theorem \ref{teo1}, the operators $M_{\Omega ,\alpha
}$ and $I_{\Omega ,\alpha }$ are bounded from $VM_{p,\varphi _{1}}\left(
w^{p}\right) $ to $VM_{q,\varphi _{2}}\left( w^{q}\right) $.
\end{corollary}
\begin{corollary}
For $w\equiv 1$, under the conditions of Theorem \ref{teo1}, we get the
Theorem 2 in \cite{Gurbuz1}.
\end{corollary}
|
\section{Introduction}
\label{sec:intro}
The proposed Iron Calorimeter (ICAL) at the India-based Neutrino Observatory (INO) will consist of three 17\,kton modules each having 150 layers of iron interleaved with Resistive Plate Chamber (RPC) detectors \citep{akumar}. The major goal of the ICAL is to determine the mass hierarchy in the neutrino sector \citep{ghosh2013}. INO will be situated at Pottipuram in Bodi West hills of Theni, India under a mountain cover of about 1\,km from all the sides. The rock overburden above the ICAL significantly reduces the background arising due to cosmic ray muons (by a factor of 10$^{6}$). A background free, or a very low background, environment is indispensable for searching rare physics processes. To reduce the background level one needs to first understand all the sources of background and then design a technique to reduce it. One such technique is the use of a cosmic muon veto detector (CMVD), the use of which, in the context of shallow depth the ICAL has been discussed in \citep{npanchal}. The reduction in cosmic ray muon flux at a depth of 100\,m is about 10$^{2}$. Therefore, the veto efficiency required to achieve the background level of the Theni site is 99.99\%, a goal which appears to be feasible \citep{npanchal}. At shallow depths, the most significant background is due to cosmic ray muons which can be vetoed with a high veto efficiency. However, the secondaries generated due to muon-nucleus interaction with the rock material can be a serious concern which needs to be carefully investigated. High energy cosmic muons can undergo an inelastic interaction with the nucleus or could be absorbed. The secondary neutral particles produced in such interactions could escape detection in the veto detector and give rise to events in the ICAL that could mimic a neutrino interaction. The SICAL is a worthwhile proposition if the number of such false positives are reduced to a level that is significantly lower than the true positives (i.e. genuine neutrino events). It may not, therefore, be out of place to mention some of possible advantages of the SICAL detector viz.
\\
(1) a much larger choice of sites,\\
(2) much larger caverns,\\
(3) monitoring of RPCs using the much larger muon flux at 100\,m depth as compared to that at 1\,km depth,\\
(4) if shown to be feasible, using muon spin rotation to get additional information on the internal magnetic field \citep{muSR} and\\
(5) enhancing the sensitivity for exotic searches including probing the origin of the anomalous KGF events \citep{dash1} and cosmogenic magnetic monopoles \citep{dash2} using the CMVD and the ICAL.
In this paper, we present results of a Monte Carlo study to estimate the contribution of this background and compare it with the event rate expected from atmospheric neutrinos in the SICAL detector. We first present the details of the simulation framework in Section 2. The outcome of the simulation by propagating the neutrals through the ICAL detector using the INO-ICAL simulation code~\cite{ical_update} is discussed in the Section 3. In Section 4, we discuss the estimation of the false positive event rate at the SICAL. The summary of the major results of the simulation is presented in Section 5.
\section{Simulation framework}
A schematic of the SICAL detector at a depth of 103\,m with the CMVD is shown in Fig.~\ref{fig:SCICAL}. The cavern $\rm(80\,m\,\times\,26\,m\,\times\,26\,m )$ is surrounded by a rock coverage of 2\,km from both the horizons and the CMVD, made up of 3\,cm thick scintillators, is placed up against the 4 walls and the ceiling of the cavern.
\begin{figure}[ht]
\centering
\includegraphics[width=100mm]{ICAL_big.png}
\caption{A schematic of the SICAL detector at a depth of 103\,m along with the CMVD}
\label{fig:SCICAL}
\end{figure}
A full simulation of the neutrino-like events in the ICAL involves propagating the cosmic muons at the surface in the intervening rock corresponding to the chosen location. This would include keeping track of the secondary particles produced in muon-nucleus interactions anywhere along their path towards the ICAL detector at a specified depth. We performed the simulation for a depth of 103\,m and the results are presented in this paper. The propagation of low energy muons and secondary hadrons, produced in high energy muon interactions in the upper part of the 103\,m rock overburden, increases the computation time although these particles will not survive the remaining rock thickness. Therefore, to decrease the computation time, the simulation was done in two parts with the corresponding geometry using GEANT4 \citep{GEANT4}. In the first part, the cosmic muon flux is generated at a depth of 100\,m. Using this flux, muon-nuclear interactions are simulated in the last 3\,m of rock and the neutrals produced are propagated to the ICAL detector in the second part.
A schematic of the geometry used for the first part of the simulation is shown in Fig.~\ref{fig:schem_part1}.
\begin{figure}[ht]
\centering
\begin{tabular}{ccc}
\includegraphics[width=55mm]{EM.png} &
\includegraphics[width=45mm]{EM_side.png} &
\includegraphics[width=45mm]{EM_top.png}
\end{tabular}
\caption{A schematic representation for the geometry used for the first part of the simulation. 3D cartoon picture (\textit{left}), 2D side view (\textit{center}) and top view (\textit{right}).}
\label{fig:schem_part1}
\end{figure}
In this part, $N_0~(6.48 \times 10^{13})$ cosmic muons of energy E and polar angles ($\theta, \phi$) are generated uniformly over the surface of the rock of area 2\,km $\times$ 2\,km at the sea level using CORSIKA~\citep{corsika} software with the sibyll model~\citep{sybill}. Ideally, a rock with infinite surface area would be required to generate the full cosmic muon spectrum for all possible angles. In the simulation, the area of the 100\,m thick rock is chosen to have transverse dimensions of 2\,km $\times$ 2\,km which mimics an infinite plane as compared to the ICAL surface (48\,m $\times$ 16\,m) and covers $\sim$\,99.9\% of the complete cosmic muon zenith angle spectrum. Muons having low energy would not be able to come out of the rock due to ionization energy loss. Furthermore, muons having an initial direction that doesn't intersect with any side planes of the `volume V', a cuboid of size $\rm(100\,m\,\times\,40\,m\,\times\,40\,m)$ (see Fig.~\ref{fig:schem_part1}), are not expected to contribute to the muon-induced neutral events at the ICAL. The dimension of `volume V' is chosen to be more than the CMVD to take into consideration the change in the direction of muons due to multiple scattering in the rock. Hence, to save computation time, only $N_1 (9\times10^8)$ muons out of $N_0$, having energy $E>\frac{E_{th}}{cos\theta}$ ($E_{th} = 48$\,GeV) and direction ($\theta, \phi$) which intersects any two planes of `volume V' are selected and allowed to propagate through the 100\,m of rock. The ratio between $N_0$ and $N_1$ obtained from the simulation is $7.2\times10^{4}$. This involves a factor due to the solid angle coverage of volume V at the surface of the rock of $6\times10^2$ and a factor of $1.2\times10^2$ from the energy loss of muons in 100\,m of rock. It should be mentioned here that, if a stopping power of $\frac{1}{\rho}\frac{dE}{dx} = 2\,\rm MeVg^{-1}cm^{2}$ is assumed for a MIP (minimum ionizing particle), a muon loses 46\,GeV energy after traversing 100\,m of rock. However, non-MIP muons lose more than $2 \rm MeVg^{-1}cm^{2}$ as they propagate, hence, only $N_2$ muons out of $N_1$ come out of the rock. It is observed that, $N_2$ is $\sim 0.65 N_1$ for $E_{th} = 46$\,GeV and $\sim 0.82 N_1$ for $E_{th} = 48$\,GeV. We have used $E_{th} = 48$\,GeV in this simulation. Throughout the simulation, $\rm SiO_2$ of density 2.32 $\rm g cm^{-3}$ is used as a rock material. The muon-nuclear interaction is not important in this part of the simulation, hence, it was excluded in the physics list of GEANT4. The (x,y) position at the rock surface for $N_0$, $N_1$ and $N_2$ is shown in Fig.~\ref{fig:xy_pos_N0N1N2} and the corresponding $E, \theta$ and $\phi$ distribution is shown in Fig.~\ref{fig:E_theta_phi_N0N1N2} for a subset of the data.
\begin{figure}[ht]
\centering
\begin{tabular}{ccc}
\includegraphics[width=45mm]{generated_xy_08jan19.png} &
\includegraphics[width=45mm]{selected_xy_08jan19.png}&
\includegraphics[width=45mm]{100mdepth_xy_08jan19.png}\\
\end{tabular}
\caption{The position distribution, (x,y) of events at the rock surface for $N_0$ (\textit{left}), $N_1$ (\textit{center}) and $N_2$ (\textit{right}).}
\label{fig:xy_pos_N0N1N2}
\end{figure}
\begin{figure}[ht]
\centering
\begin{tabular}{ccc}
\includegraphics[width=45mm]{energy_N0.png} &
\includegraphics[width=45mm]{energy_N1.png} &
\includegraphics[width=45mm]{energy_N2.png}\\
\includegraphics[width=45mm]{theta_N0.png} &
\includegraphics[width=45mm]{theta_N1.png} &
\includegraphics[width=45mm]{theta_N2.png} \\
\includegraphics[width=45mm]{phi_N0.png}&
\includegraphics[width=45mm]{phi_N1.png}&
\includegraphics[width=45mm]{phi_N2.png}\\
\end{tabular}
\caption {\label{fig:E_theta_phi_N0N1N2} $E$ (\textit{top}), $\theta$ (\textit{center}) and $\phi$ (\textit{bottom}) distribution for $N_0$ (\textit{left}), $N_1$ (\textit{center}) and $N_2$ (\textit{right}) for event selections $N_0$, $N_1$, $N_2$ as described in the text.}
\end{figure}
The detector geometry used for the second part of the simulation is shown in Fig.~\ref{fig:schem_part2}.
\begin{figure*}[ht]
$$\begin{array}{cc}
\includegraphics[width=0.25\textwidth]{ICAL_top.png} &
\includegraphics[width=0.6\textwidth]{ICAL_side.png}\\
\multicolumn{2}{c}{\includegraphics[width=0.5\textwidth]{ICAL.png}}
\end{array}$$
\caption{A schematic representation for the detector geometry used in the second part of the simulation.}
\label{fig:schem_part2}
\end{figure*}
In this part, the muon-induced neutral background for the SICAL detector is studied. For this, $N_2(7.38\times 10^8)$ muons with $E, \theta$ and $\phi$ distribution as shown in Fig.~\ref{fig:E_theta_phi_N0N1N2} (\textit{bottom}) are propagated from the top surface of the 3\,m rock having the same (x,y) coordinates as shown in Fig.~\ref{fig:xy_pos_N0N1N2} (\textit{right}) obtained from the first part. It is legitimate to consider that the neutrals produced in muon-nuclear interactions, mostly from the last part of the 3\,m depth of rock, could exit the rock. The choice of 3\,m was guided by the hadronic interaction length $\lambda$ for rock which is 36\,cm \citep{PDG} i.e. $\sim$10 times smaller. This was verified by performing the simulation for 5\,m and 10\,m of rock which produced, within error, the same number of outgoing neutral particles as with 3\,m rock. Following this argument, only $N_3 (3.69 \times 10^8)$ muons out of $N_2$, that are expected to pass through rock material of 3 m surrounding the cavern, are allowed to propagate. The Kokoulin model \citep{K_model} is used to simulate muon-nuclear interactions. The cross-section for this process was increased by a factor 100 to reduce the computation time. The hadronic interactions of the secondaries are also considered. All the particles (both neutral and charged) that are coming out of the rock and entering through the cavern are recorded in the scintillator of the CMVD. As the muon-nuclear cross-section is increased by a factor of 100, the charged particles (predominantly muons) will create extraneous interactions. Hence, they were not propagated beyond the CMVD once detected in the scintillator. In contrast, the neutrals were allowed to propagate through the cavern towards the ICAL detector.
\section{Results and discussions}
The total number of muon-nuclear interactions in the rock are found to be $\,5.58\, \times$\,10$^{8}$ and the vertices of these interactions are shown in Fig.~\ref{fig:vertices_intetactn}
\begin{figure}[ht]
\centering
\begin{tabular}{ccc}
\includegraphics[width=40mm]{muNucXY_07jan19.png} &
\includegraphics[width=40mm]{muNucXZ_07jan19.png}&
\includegraphics[width=40mm]{muNucYZ_07jan19.png}\\
\end{tabular}
\caption{The vertices of all the muon-nuclear interaction in different views.}
\label{fig:vertices_intetactn}
\end{figure}
It can be seen that the number of interactions increases with the distance traversed by the muon in the rock as expected. The number of secondaries produced due to muon-nuclear interactions is 4.95\,$\times$\,10$^{9}$ and out of this 2.7\,$\times$\,10$^{9}$ could come out of the rock. Although a neutron cannot produce a muon through its decay, it can interact with the Fe in the ICAL and produce charged pions and kaons which can decay producing muons. The K$^{0}_{L}$ has a decay mode leading to muons and kaons with a branching ratio of 27 \% which makes it important for the present study. The $\pi^{0}$s produced are not of major concern as they have a very short lifetime ($\sim$\,10$^{-16}$ sec) and decay into 2 $\gamma$-rays. Each $\gamma$-ray leads to an electromagnetic shower in the ICAL which can be very well distinguished from the muon trajectories. Consequently, the most important muon-induced neutral background are neutrons and K$^{0}_{L}$. The total number of neutrons and K$^{0}_{L}$ that are produced due to muon-nuclear interaction is $1.08\times 10^9$ and $4.23\times 10^6$, respectively. A typical ($E,\theta$) distribution for the neutrons and the $K^0_L$ at the entrance of the CMVD is shown in Fig.~\ref{fig:muNuc_E_theta_phi}.
\begin{figure}[ht]
\centering
\begin{tabular}{cc}
\includegraphics[width=60mm]{sci_E_neutron_07jan19.png} &
\includegraphics[width=60mm]{sci_E_kl0_07jan19.png}\\
\includegraphics[width=60mm]{sci_th_neutron_07jan19.png}&
\includegraphics[width=60mm]{sci_th_kl0_07jan19.png}\\
\end{tabular}
\caption {\label{fig:muNuc_E_theta_phi} Energy and theta spectra of neutrons (left top and left bottom) and $K^{0}_{L}$s (right top and right bottom).}
\end{figure}
It should be pointed out that, the neutral particles should have more than 1\,GeV of energy to produce a charged particle in nuclear interaction, which can pass through five layers of RPC detector in the ICAL. Hence, only those events in which neutrals have energy $>$1\,GeV are considered for further estimation of muon-induced neutral background in the ICAL. The number of such events relevant for this study are\\
a) Events in which neutrals are accompanied by no charged particle\\
b) Events in which neutrals are accompanied by charged particles having kinetic energy less than 10\,MeV as below this energy the charged particle may not give any signal in the CMVD.\\
These events, that are generated as a result of interaction of particles different from a neutrino but are likely to be classified as neutrino induced events in the ICAL detector are called false positive signals. All these events are then reconstructed in the ICAL detector using the Kalman filter technique~\citep{kolahal_kalman,ical_update}. The trajectory of a charged particle due to a false positive signal is considered to be $\nu$-induced muon signal, if it has hits in a minimum 5 layers with $\chi^{2}/ndf < 10$ and is contained within the fiducial volume of the ICAL detector, where the fiducial volume excludes the region of the top 4 layers and 30\,cm from all the four sides of the ICAL detector. Consequently, 2 out of $9\times10^8$ simulated events have resulted in false positive events. These are shown in Fig.~\ref{fig:tracks} and the relevant parameters are listed in Table~\ref{Tab:table1}. This also provides an upper bound of 6.3 false positive events at ICAL with a confidence level~\cite{PDG} (C.L.) of 95\%. Due to almost 100\% efficiency of the veto detector, a large fraction of all the primary muons coming out of the rock would be vetoed. From an earlier measurement with a small Cosmic Muon Veto detector \citep{npanchal} the veto efficiency achieved was 99.987\% which is equivalent to a reduction in muon flux by about 10$^{4}$. Nevertheless, due to the small inefficiency, a part of the total primary cosmic muons will leak through the veto detector undetected. It should be emphasized that, the number of such muons will be comparable to the muon background level in the ICAL detector placed at a depth of 1\,km. This residual primary muon background will be identified and removed in the same way as in the original plan of the ICAL detector with about 1\,km rock over-burden by using the algorithm to detect events in the fiducial volume of the ICAL detector.
\begin{figure}[ht]
\centering
\begin{tabular}{rr}
\includegraphics[width=150mm]{655343_02jan19.png}\\
\includegraphics[width=150mm]{411601_02jan19.png}
\end{tabular}
\caption {\label{fig:tracks} Reconstructed tracks of the two false positive events in the ICAL detector obtained from the simulation. }
\end{figure}
\begin{table}[htbp]\label{Tab:table1}
\centering
\caption{\label{table:tab_4}{Characteristics of the two events that are false positive signal at SICAL. $V_{x}$ and $V_{y}$ denotes the x and y position of the vertices of the reconstructed track.}}
\label{Tab:table1}
\smallskip
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
No. of hits & $\chi^{2}/ndf$ & $\theta$ (rad) & $\phi$ (rad) & $V_{x}$ (m) & $V_{y}$ (m) & Layer No.\\
\hline
5 & 0.37 & 2.34 & 2.33 & 6.63 & 7.65 & 101\\
5 & 0.24 & 2.17 & -1.44 & -21.8 & -7.48 & 137\\
\hline
\end{tabular}
\end{table}
\section{Estimation of false positive event rate in the SICAL due to muon-induced neutrals.}
In the present study, the muon-induced neutral background at the SICAL is simulated for $N_0 = 6.48 \times 10^{13}$ incident muons at sea level. As mentioned earlier, the cross-section of muon-nuclear interaction is increased by factor of 100 in the simulation in the last 3 m of rock. Therefore, $N_0$ would correspond to an effective number of muons at the sea level $N = 6.48 \times$ 10$^{15}$. The primary cosmic ray muon flux at the sea level is 70\,m$^{-2}$\,sec$^{-1}$\,sr$^{-1}$ \citep{PDG} which integrates to $\sim$\,4.8$\times$10$^{13}$\,/day for a surface area of 2\,km $\times$ 2\,km. Consequently, $N$ muons amounts to a simulation of 135 days. Hence, the false positive rate at the SICAL equates to 0.015\,/day with an upper bound of 0.05\,/day at 95\% confidence level.
The neutrino event rate at INO-ICAL using a Monte Carlo simulation is reported by A. Kumar \textit{et al.}~\citep{white_paper}, where the ICAL geometry and Honda flux \cite{neutrinoflux} was used as inputs in the nuance \citep{nuance} event generator and the neutrino event rate is estimated to be $\sim$\,4\,/day. The false positive to signal for the SICAL detector is therefore, about $\sim$~0.4\% with an upper bound of $\sim$\,1.2\% at 95\% C.L. which makes the SICAL a feasible proposition. It may be relevant to examine an issue of false vetoes, discussed in the context in the IceCube \citep{yuan}, that may potentially affect the ICAL. The cosmic ray interaction with the upper atmosphere leading to neutrinos is also associated with muons. However, the tracking capability of the ICAL detector in combination with the CMVD should be able to clearly identify the event due to muon neutrinos as opposed to the those originating from the upper atmosphere and giving a track at a physically different location.
While the energy and angle dependence of the muon spectrum is pretty robust there could be some leeway in the deep inelastic partial differential cross sections used in GEANT4. A more foolproof test of this idea would be to place a reasonable sized ICAL prototype detector at a shallow depth of $\sim$\,30\,m, enclose it in a CMVD and quantify the false positive which could mimic neutrino events in the ICAL, at the same time comparing with the simulations at that depth.
\section{Summary}
In summary, we have presented results of simulations which support the possibility of locating the INO-ICAL detector at a shallow location with a rock overburden of $\sim$\,100\,m when used with an efficient cosmic ray shield with an efficiency of about 99.99\% for detecting charged particles. The main background is due to neutral, long lived and energetic particles produced in the last few metres of rock either unaccompanied by or associated charged particles which go undetected. This fraction has been estimated to be much smaller ($\sim$\,0.4\% with an upper bound of $\sim$\,1.2\% at 95\% C.L.) than the signal due to atmospheric neutrinos. Therefore, the proposal of a the SICAL detector opens up the possibility of having a much larger choice of locations, saving time due to the shorter tunnel and allowing for much larger caverns. However, it must be mentioned that it is perhaps necessary, and prudent that a proof-of-principle detector be built at a shallow depth, perhaps even at 30\,m rock overburden, together with the Cosmic Veto Detector to validate the simulation.
\acknowledgments
We would like thank M.V.N. Murthy, Naba K. Mondal, D. Indumathi, Paschal Coyle, John G. Learned, Morihiro Honda, Thomas K. Gaisser, Shashi R. Dugad, Deepak Samuel and Satyajit Saha for useful comments and suggestions and Apoorva Bhatt and P. Nagaraj for help in simulation related issues. Finally, we would also like to thank the anonymous referee for critical but useful comments.
|
\section{Introduction}
\label{sec:S1}
Recent theoretical investigations
predicted that the hydrogen Ly-$\alpha$ line of the solar disk radiation
should be linearly polarized by the scattering of anisotropic radiation, with measurable polarization signals in both the core and wings of the line (Trujillo Bueno et al. 2011; Belluzzi et al. 2012; \v{S}t\v{e}p\'an et al. 2015). Moreover, such radiative transfer investigations pointed out that the line-center polarization is modified by the presence of magnetic fields in the chromosphere-corona TR via the Hanle effect.
These theoretical investigations motivated the development of the Chromospheric Lyman-Alpha SpectroPolarimeter (CLASP), an international sounding rocket experiment that on 3rd September 2015 successfully measured the wavelength variation of the intensity and linear polarization of the Ly-$\alpha$ line in quiet regions of the solar disk (see Kano et al. 2017, and references therein to the papers describing the instrument).
CLASP observed the Ly-$\alpha$ Stokes profiles $I$, $Q$, and $U$ with a spatial resolution of about 3 arcsec and a temporal resolution of 5 minutes. The resulting $Q/I$ and $U/I$ linear polarization signals are of order 0.1\% in the line center and up to a few percent in the nearby wings, both showing conspicuous spatial variations with scales of ${\sim}10$ arcsec, in agreement with the theoretical predictions. However, the observed $Q/I$ line-center signals do not show any clear center-to-limb variation (CLV; see Fig. \ref{figure-1}), in contrast with the results of the above-mentioned radiative transfer investigations in several models of the solar atmosphere.
The observed lack of a clear CLV in the $Q/I$ line-center signal came as an interesting surprise because many other solar spectral lines show such a variation (e.g., Stenflo et al. 1997), including the K line of Ca {\sc ii} whose line-center $Q/I$ and $U/I$ spatial variations are sensitive to the magnetic and thermal structure of the chromosphere at heights a few hundred kilometers below the TR (Holzreuter \& Stenflo 2007).
The plasma of the upper solar chromosphere is highly structured and dynamic, departing radically from a horizontally homogeneous model atmosphere. In 3D models resulting from magneto-hydrodynamic (MHD) simulations (e.g., Carlsson et al. 2016), the TR,
where the temperature suddenly increases from about $10^4$ K to the $10^6$ K coronal values, is not a thin horizontal layer, but it delineates a highly corrugated surface.
In this work, such corrugated surface is characterized by specifying, at each point, the vector $\vec n$ indicating the direction of the local temperature gradient (hereafter, the local TR normal vector).
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.9]{fig-1a}
\includegraphics[scale=0.9]{fig-1b}
\includegraphics[scale=0.9]{fig-1c}
\caption{CLASP observations.
The variation with the cosine of the heliocentric angle ($\mu$)
of the Stokes $I$, $Q/I$ and $U/I$ line-center signals observed by CLASP in the hydrogen Ly-$\alpha$ line.
The reference direction for the positive Stokes $Q$ is the parallel to the nearest limb.}
\label{figure-1}
\end{center}
\end{figure}
We begin by illustrating that the spatial variations of the line-center $Q/I$ and $U/I$ signals of the Ly-$\alpha$ line are very sensitive to the geometric complexity of the corrugated surface that delineates the TR. Secondly, we demonstrate that the significant CLV of the $Q/I$ line-center signal of the Ly-$\alpha$ line
calculated in Carlsson et al's (2016) 3D radiation MHD model of the solar atmosphere can be reduced by increasing the magnetic field strength and/or the geometrical complexity of the model's TR. We then show how this can be exploited to constrain, from the CLASP line-center data, the magnetic strength and geometric complexity of the solar TR by applying the statistical inference method discussed in \v{S}t\v{e}p\'an et al. (2018). To this end, we confront the statistical properties of the CLASP line-center data with those of the polarization signals calculated in a grid of 3D model atmospheres characterized by different degrees of geometrical complexity and magnetization.
\section{An analytical corrugated transition region model}
\label{sec:S2}
In a 1D model atmosphere, static or with radial velocities, unmagnetized or with
a magnetic field having a random azimuth at sub-resolution scales,
the radiation field at each point within the medium has axial symmetry around the vertical.
Under such circumstances, taking a reference system with the Z-axis (i.e., the quantization axis) along the vertical, and choosing the reference direction for linear polarization perpendicular to the plane formed by the vertical and the LOS (i.e., the parallel to the limb), then $U/I=0$ and the following approximate formula can be applied to estimate the line-center $Q/I$ scattering polarization signal of the hydrogen Ly-$\alpha$ line (Trujillo Bueno et al. 2011):
\begin{equation}
\frac QI\,\approx\,{1\over{2\sqrt{2}}}(1-\mu^2)\,{\cal H}\,{{{\bar J}^2_0}\over{{\bar J}^0_0}}\,,
\label{eq:qieddb}
\end{equation}
where $\mu={\rm cos}\,{\theta}$ (with $\theta$ the heliocentric angle), ${\cal H}$ is the Hanle depolarization factor (${\cal H}=1$ for $B=0$ G and ${\cal H}{<}1$
in the presence of a magnetic field), and ${{\bar J}^2_0}/{{\bar J}^0_0}$ is the degree of anisotropy of the spectral line radiation at the height in the model atmosphere where the line-center optical depth is unity along the line of sight (see equations 2 and 3 of Trujillo Bueno et al. 2011). As seen in figure 1 of Trujillo Bueno et al. (2011), the anisotropy factor of the hydrogen Ly-$\alpha$ line suddenly becomes significant at the atmospheric height where the line-center optical depth is unity, which practically coincides with the location of the model's chromosphere-corona TR. Therefore, under such assumptions, we should expect a clear CLV in the $Q/I$ line-center signal, established by the $(1-\mu^2)$ factor of Eq. (1), and modified by the height-dependence of the anisotropy factor (see the height range between the two solid-line arrows in figure 1 of Trujillo Bueno et al. 2011). Because of our choice for the reference direction for linear polarization, the only way to have $U/I{\ne}0$ in a static 1D model atmosphere is by means of the Hanle effect of a magnetic field inclined with respect to the local vertical. However, in such a 1D model atmosphere there is no way to destroy the CLV of the $Q/I$ line-center signal.
The line-center photons of the hydrogen Ly-$\alpha$ line originate just at the boundary of the model's TR.
In a 1D model atmosphere (e.g., Fontenla et al. 1993), the vector normal to the TR lies along the vertical, which coincides with the symmetry axis of the incident radiation field. In a three-dimensional model atmosphere (e.g., Carlsson et al. 2016), the model's TR delineates a corrugated surface, so that the vector normal to the model's TR changes its direction from point to point (see figure 7 of \v{S}t\v{e}p\'an et al. 2015). On the other hand, at each point on such corrugated surface the stratification of the physical quantities along the local normal vector $\vec n$ is much more important than along the perpendicular direction. In order to estimate how the line-center fractional polarization of the hydrogen Ly-$\alpha$ line is at each point of the field of view, it is reasonable to assume that the incident radiation field has axial symmetry around the direction of the normal vector $\vec n$ corresponding to the spatial point under consideration (see also \v{S}t\v{e}p\'an et al. 2018). Taking a reference system with the Z-axis directed along the normal vector $\vec n$, and recalling Eq. (1), at each point of the corrugated TR surface we can estimate the line-center fractional polarization signals through the following formula
\begin{equation}
{\Big{[}}\frac QI{\Big{]}}_{\vec n}\,\approx\,{1\over{2\sqrt{2}}}(1-{\mu}_{\vec n}^2)\,{\cal H}\,{\Big{[}}{{{\bar J}^2_0}\over{{\bar J}^0_0}}{\Big{]}}_{\vec n} \, ,
\label{eq:qieddb}
\end{equation}
where the positive Stokes $Q$ reference direction is now the perpendicular to the plane formed by $\vec n$ and the LOS, ${\mu}_{\vec n}$ is the cosine of the angle between $\vec n$ and the LOS, and ${{[}}{{\bar J}^2_0}/{\bar J}^0_0{{]}}_{\vec n}$ is the anisotropy factor calculated in the new reference system.
Clearly, since the direction of $\vec n$ changes as we move through the corrugated surface that delineates the TR, the $Q/I$ signals estimated with Eq. (2) do not share the same reference direction for the quantification of the linear polarization. In order to arrive at equations for $Q/I$ and $U/I$ having the parallel to the nearest limb as the positive Stokes $Q$ reference direction, we have applied suitable rotations of the reference system. Specifying the direction of the local TR normal vector through its inclination $\theta_{\vec n}$ with respect to the vertical (with $\theta_{\vec n}$ between $0^{\circ}$ and $90^{\circ}$) and azimuth $\chi_{\vec n}$ (with $\chi_{\vec n}$ between $0^{\circ}$ and $360^{\circ}$), our analytical calculations show that the line-center $Q/I$ and $U/I$ signals of the hydrogen Ly-$\alpha$ line can be estimated by the following formulas:
\begin{equation}
\frac QI\,\approx\,{\cal H}{1\over{\sqrt{2}}}\,{[}{1\over{4}}{\sin}^2{\theta}(3{\cos}^2{\theta_{\vec n}}-1)-{\sin}{\theta}{\cos}{\theta}{\sin}{\theta_{\vec n}}{\cos}{\theta_{\vec n}}{\cos}{\chi_{\vec n}}+{1\over{4}}(1+{\cos}^2{\theta})\,{\sin}^2{\theta_{\vec n}}{\cos}{(2\chi_{\vec n})}\,{]}\,[{{{\bar J}^2_0}\over{{\bar J}^0_0}}]_{\vec n},
\end{equation}
\begin{equation}
\frac UI\,\approx\,{\cal H}{1\over{\sqrt{2}}}\,[{1\over{2}}{\cos}{\theta}{\sin}^2{\theta_{\vec n}}{\sin}{(2\chi_{\vec n})}-{\sin}{\theta}{\sin}{\theta_{\vec n}}{\cos}{\theta_{\vec n}}{\sin}{\chi_{\vec n}}]\,[{{{\bar J}^2_0}\over{{\bar J}^0_0}}]_{\vec n}.
\end{equation}
Figure 2 shows examples of CLV of the $Q/I$ and $U/I$ Ly-$\alpha$ line-center signals for several topologies of the model's TR, with random azimuth $\chi_{\vec n}$. The first three rows of the figure correspond to the indicated fixed inclinations ${\theta_{\vec n}}$, while the bottom row panels
show the case in which also ${\theta_{\vec n}}$ has random values (i.e., the probability distribution $p(\cos{\theta}_{\vec n})=1$ for $0{\le}\cos{\theta}_{\vec n}{\le}1$ and $p(\cos{\theta}_{\vec n})=0$ for $-1{\le}\cos{\theta}_{\vec n}{<}0$).
Note that the CLV of the fractional linear polarization line-center signals is very sensitive to the geometry of the model's TR, and that there is no CLV at all when the normal vector $\vec n$ has random orientations.
Moreover, it is also very important to point out that the $Q/I$ and $U/I$ amplitudes are sensitive to the magnetic field strength of the chromosphere-corona TR, through the Hanle factor $\cal H$ in Eqs. (3) and (4).
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.7]{fig-2a}
\includegraphics[scale=0.7]{fig-2b}
\includegraphics[scale=0.7]{fig-2c}
\includegraphics[scale=0.7]{fig-2d}
\caption{The CLV of the Stokes $Q/I$ (left) and $U/I$ (right) line-center signals of the hydrogen Ly-$\alpha$ line calculated
according to Eqs. (3) and (4) with ${\cal H}=1$ (no magnetic field) and ${\Big{[}{{{\bar J}^2_0}/{{\bar J}^0_0}}\Big{]}}_{\vec n}=-0.02$,
for the fixed inclinations (${\theta_{\vec n}}$) of the TR normal vector $\vec n$ indicated in the panel insets and assuming random azimuths $\chi_{\vec n}$.
Note that the bottom panel corresponds to the case in which ${\theta_{\vec n}}$ also has random values (see the text for details).
The reference direction for the positive Stokes $Q$ is the parallel to the nearest limb.
}
\label{figure-2}
\end{center}
\end{figure}
\section{The impact of the magnetization and geometrical complexity in 3D models}
\label{sec:S3}
The analytical TR model described in the previous section is very useful to understand
why the spatial variations of the $Q/I$ and $U/I$ line-center signals of the Ly-$\alpha$ line are very sensitive to the
geometry of the corrugated surface that delineates the solar TR. The magnetic
field of such analytical model is characterized by a random azimuth at sub-resolution scales, which implies that
the impact of the model's magnetic field strength on the CLV is just the scaling factor $\cal H$ that appears in Eqs. (3) and (4). We now show theoretical results for the 3D radiation magneto-hydrodynamical model of the chromosphere-corona TR of Carlsson et al. (2016), which is representative of an enhanced network region and has magnetic field lines reaching chromospheric and coronal heights.
The first two rows of Fig. 3 show examples of the CLV of the $Q/I$ and $U/I$ line-center signals calculated
with the radiative transfer code PORTA (\v{S}t\v{e}p\'an \& Trujillo Bueno 2013) in the 3D model of Carlsson et al. (2016)
ignoring (first row) and taking into account (second row) the CLASP instrument degradation (see Giono et al. 2016 for information on the point spread function).
Clearly, the peak-to-peak amplitudes of the $Q/I$ and $U/I$ variations are significantly smaller when the degradation produced by the instrument is accounted for. Note that the $Q/I$ line-center signals calculated in such a 3D model show a clear CLV.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.8]{fig-3a}
\includegraphics[scale=0.8]{fig-3b}
\includegraphics[scale=0.8]{fig-3c}
\includegraphics[scale=0.8]{fig-3d}
\includegraphics[scale=0.8]{fig-3e}
\includegraphics[scale=0.8]{fig-3f}
\includegraphics[scale=0.8]{fig-3g}
\includegraphics[scale=0.8]{fig-3h}
\caption{The CLV of the Stokes $Q/I$ (left) and $U/I$ (right) line-center signals of the hydrogen Ly-$\alpha$ line calculated in the 3D model of Carlsson et al. (2016), ignoring (topmost panels) and accounting for (rest of the panels) the degradation produced by the CLASP instrument.
The panels of the first and second rows show the results corresponding to the magnetic field and TR corrugation of the original 3D model. Those of the third row
show the results obtained assuming an enhanced magnetization characterized by a scaling factor $\beta=8$ (which implies a mean field strength of 120 G in the model's TR). The panels of the fourth row show the results obtained assuming a compression factor $\gamma=1/4$ (which quantifies the geometrical complexity of the model's TR). The red curves in the left panels are best fits quadratic in $\mu$ to the fluctuating spatial variations.
The reference direction for the positive Stokes $Q$ is the parallel to the nearest limb.
}
\label{figure-2}
\end{center}
\end{figure}
The panels in the third row of Fig. 3 show what happens when the magnetic field strength at each point in the 3D model is increased by a scaling factor $\beta=8$, so that the mean field strength at the model's TR is 120 G instead of 15 G. Clearly, an enhanced magnetization in the TR plasma has a significant impact on the linear polarization amplitudes, but it does not destroy the CLV of the $Q/I$ line-center signal.
We show now what happens when the geometric complexity of the TR of the 3D model is modified. To this end, we simply compress the horizontal extension of the 3D model by a factor $\gamma$ along the $X$ and $Y$ directions, so that the divergence of the model's magnetic
field vector at each grid point remains equal to zero and its modulus unaltered.
The panels in the bottom row of Fig. 3 shows an example for $\beta=1$ (i.e., the model's magnetic field), but with
$\gamma=1/4$. Clearly, the degree of corrugation of the model's TR surface has an important impact on the CLV of the line-center $Q/I$ signals.
\section{Constraining the degree of magnetization and corrugation of the solar TR}
\label{sec:S4}
We aim at constraining the magnetic field strength and degree of corrugation of the TR corresponding to the
solar disk regions observed by CLASP. To this end, we confront the $Q/I$ and $U/I$ line-center signals measured
by CLASP (hereafter, $\vec S^{\rm obs}$), with the theoretical line-center signals ($\vec S$) calculated in a grid of 3D models characterized by
the scaling factors $\beta$ and $\gamma$, with $\beta=\gamma=1$ corresponding to the 3D model described in Carlsson et al. (2016).
Each pixel along the 400 arcsec spectrograph slit image corresponds to a particular LOS, and we have considered those having $\mu$ values between 0.1 and 0.8, with the exception of the faint filament region located around $\mu=0.7$ in Fig. 1. The theoretical $I$, $Q$, and $U$ profiles have been degraded to mimic the CLASP resolution according to the laboratory measurements described in Giono et al. (2016).
To determine the 3D model whose emergent $Q/I$ and $U/I$ line-center signals are the closest to those observed by CLASP, we apply the statistical inference approach discussed in Section 3.1 of \v{S}t\v{e}p\'an et al. (2018), which gives the following expression for the posterior of the hyper-parameters $\vec\theta$ (i.e., the corrugation parameter $\gamma$ and the magnetization parameter $\beta$) based on the line-center data at all the CLASP spatial pixels $i$:
\begin{equation}
p(\vec\theta|\vec S^{\rm obs})=\prod_i \mathcal{L}_i(\vec\theta|\vec S^{\rm obs}) p(\vec\theta)=\mathcal{L}(\vec\theta|\vec S^{\rm obs})p(\vec\theta)\,,
\label{eq:finpost}
\end{equation}
where $p(\vec\theta)$ is the prior of the hyper-parameters. The likelihood $\mathcal{L}_i(\vec\theta|\vec S^{\rm obs})$ is
\begin{equation}
\mathcal{L}_i(\vec\theta|\vec S^{\rm obs})\equiv\int d\vec S\;\frac{p^{\mu_i}(\vec S|\vec\theta)}{(\sqrt{2\pi}\sigma)^2}
\exp\left\{ -\frac{\| \vec S-\vec S^{\rm obs} \|^2}{2\sigma^2} \right\}\,,
\label{eq:llii}
\end{equation}
with $\sigma^2$ the variance of the uncorrelated Gaussian noise.
We point out that $p^{\mu_i}(\vec S|\vec\theta)$ are the hierarchical priors of $\vec S$ that depend on the hyper-parameters $\vec\theta$.
This is derived from the histograms of the $\vec S$ signals calculated for each $\mu_i$ value in each particular 3D model characterized by its scaling factors $\beta$ and $\gamma$.
In the CLASP observations the standard deviation of the noise
is $\sigma=0.05\%$ after averaging three pixels in wavelength around
the line center.
Our prior for the corrugation of the TR is that any compression of the 3D model is allowed, while
expansions are much less likely (i.e., $p(\gamma)=1/2$ for $\gamma\in(0,1]$ and $p(\gamma)={\rm exp}(-(\gamma-1))/2$ for $\gamma>1$).
For the magnetic field strength of the TR of the quiet Sun we have deemed reasonable
to use $p(B)={\rm exp}(-B/130)/130$, based on the argument that the mean field strength of the quiet regions of the solar photosphere is of the order of 100 G and that
it decreases with height in the atmosphere (Trujillo Bueno et al. 2004). This is our first step for determining the parametrization $\vec{\theta}$ of the above-mentioned 3D statistical model of the solar atmosphere that maximizes the marginal posterior of Eq. (5). Our aim is to estimate some global properties of the quiet Sun atmosphere observed by CLASP, i.e., the mean field strength and the degree of corrugation of the chromosphere-corona TR.
Figure 4 shows the result.
It suggests that 3D models with less magnetization in the TR than in the model of Carlsson et al. (2016) produce scattering polarization signals in better agreement with those observed by CLASP. A more robust conclusion is that, among the magnetized models needed to explain the CLASP observations, 3D models with a significantly larger (i.e., $\gamma{\approx}1/3$) degree of corrugation of the TR would yield a better agreement with the observations.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=1.2]{fig-4}
\caption{The variation with $\beta$ and $\gamma$ of the marginal posterior of Eq. (5),
which quantifies the ability of the parametrized 3D models discussed in the text to explain the Ly-$\alpha$ line-center data.
$B_H=53$ G is the line's Hanle critical field. Each value of $\beta$ implies a TR mean field strength ${\langle B \rangle}$.
The white circle corresponds to the 3D model of Carlsson et al. (2016).
Note that the posterior peaks for ${\langle B \rangle}\,{\approx}\,7$ G and a compression factor $\gamma\approx {1/3}$.
}
\label{figure-5}
\end{center}
\end{figure}
\section{Conclusions}
\label{sec:concl}
Our statistical approach to interpret the spectropolarimetric observations of the hydrogen Ly-$\alpha$ line achieved by CLASP suggests that the mean field strength of the TR of the observed quiet Sun regions is significantly lower than the 15 G of Carlsson et al's (2016) 3D model, perhaps not surprisingly since it is for an enhanced network region, and that a TR plasma with a substantially larger degree of geometrical complexity is needed to explain the CLASP observations.
We think that our conclusion on the geometric complexity of the TR plasma is more robust than the one on its degree of magnetization because in order to cancel completely the CLV of the $Q/I$ line-center signals we need to increase the degree of corrugation of the model's TR, while the information on its magnetic strength is mainly encoded in the peak-to-peak amplitudes of the $Q/I$ and $U/I$ spatial radial variations, which are sensitive to the CLASP point spread function used (based on laboratory measurements) and to the 3D model atmosphere chosen for the statistical inference. In any case, a key point to emphasize is that the linear polarization produced by scattering processes in the core of the hydrogen Ly-$\alpha$ line encodes valuable information on the magnetic field and geometrical complexity of the chromosphere-corona TR, which can be revealed with the help of statistical inference methods such as the approach applied in this paper.
Clearly, having simultaneous spectropolarimetric observations in two or more spectral lines would facilitate the determination of the plasma magnetization, especially if the main difference between the spectral lines used lies within their sensitivity regime to the Hanle effect.
Evidently, we need more realistic 3D numerical models of the upper chromosphere of the quiet Sun.
Chromospheric spicules are ubiquitous in subarcsecond resolution Ly-$\alpha$ filtergrams (Vourlidas et al. 2010),
but they are not present in Carlsson et al's (2016) model. Our investigation suggests that 3D models with such needle-like plasma structures
all over the field of view would be a much better representation of the geometric complexity of the TR plasma observed by CLASP. Finally, it is also important to emphasize that a suitable way to validate or refute numerical models of the chromosphere-corona TR is by confronting calculations and observations of the scattering polarization in ultraviolet lines sensitive to the Hanle effect. We plan to pursue further this line of research by also exploiting the polarization observed by CLASP in the resonance line of Si {\sc iii} at 1206 \AA\ (see Ishikawa et al. 2017) and, of course, the observations of the Mg {\sc ii} $h$ \& $k$ lines that the future flight of CLASP-2 will provide.
\acknowledgements
The CLASP team is an international partnership between NASA Marshall Space Flight Center,
National Astronomical Observatory of Japan (NAOJ), Japan Aerospace Exploration Agency
(JAXA), Instituto de Astrof\'{i}sica de Canarias (IAC) and Institut d'Astrophysique Spatiale;
additional partners are the Astronomical Institute ASCR, Istituto Ricerche Solari Locarno (IRSOL),
Lockheed Martin and University of Oslo. The US participation was funded by NASA Low Cost Access to Space (Award
Number 12-SHP 12/2-0283). The Japanese participation was funded by the basic research
program of ISAS/JAXA, internal research funding of NAOJ, and JSPS KAKENHI Grant
Numbers 23340052, 24740134, 24340040, and 25220703. The Spanish participation was
funded by the Ministry of Economy and Competitiveness through project AYA2010-18029
(Solar Magnetism and Astrophysical Spectropolarimetry). The French hardware participation
was funded by Centre National d'Etudes Spatiales (CNES). Moreover, we acknowledge the grants provided by the Barcelona Supercomputing Center (National Supercomputing Center, Barcelona, Spain), as well as the funding received through grant \mbox{16--16861S} of the Grant Agency of the Czech Republic, project \mbox{RVO:67985815} of the Czech Academy of Sciences, and projects \mbox{AYA2014-60476-P} and \mbox{AYA2014-55078-P} of the Spanish Ministry of Economy and Competitiveness. JS is grateful to the Severo Ochoa Visiting Researchers Programme of the IAC for financing a few months working visit at the IAC.
JTB acknowledges the funding received from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (ERC Advanced Grant agreement No 742265).
\software{PORTA (\v{S}t\v{e}p\'an \& Trujillo Bueno 2013)}
|
\section{Introduction}
Within this paper, we present a new abstract formulation of parabolic initial-boundary value problems with dynamic boundary conditions. This includes locally reacting {\em Wentzell boundary conditions} as well as non-local {\em kinetic boundary conditions} modeling a diffusion on the surface of the computational domain. In both cases, we derive a formulation as partial differential-algebraic equation (PDAE) where the boundary condition is treated in form of a coupling condition. This constraint is included to the system as an additional equation and enforced with the help of a Lagrange multiplier. The resulting structure is then similar to the one we would obtain for a parabolic problem with pure Dirichlet boundary conditions. \smallskip
In general, dynamic boundary conditions appear in applications where the momentum on the boundary should not be neglected. For hyperbolic systems, such boundary conditions are of particular interest in fluid-structure interaction, where one component can be modelled in form of a boundary layer instead of using a full model~\cite{Hip17}. For the heat equation, dynamic boundary conditions can be found in~\cite[Ch.~2]{DuvL76} and enable a proper way to model a heat source on the boundary~\cite{Gol06}. On the other hand, dynamic boundary conditions may be used as stabilizing feedback control on the boundary~\cite{KomZ90}.
An abstract framework for the weak formulation of parabolic systems with dynamic boundary conditions was recently presented in~\cite{KovL17}. Therein, it was shown that the problem fits into the standard formulation of parabolic problems if the inner products are adapted accordingly.
Formulations including semigroups have received much more attention in the literature.
Here, the boundary conditions are prescribed within the domain of the differential operator. The proof of the analytic semigroup property in~\cite{EngF05} considers a decoupling similar to the idea presented in the present paper but without the inclusion of a Lagrange multiplier. \smallskip
In this paper, we interpret such parabolic initial-boundary value problems as a coupled system including dynamics in the bulk and on the boundary. As we consider the weak formulation of the problem, we have a direct connection to spatial discretization schemes. This may lead to a new class of robust numerical methods and allows independent meshes in the domain and on the boundary.
\section{Preliminaries}\label{sec:prelim}
This section is devoted to the introduction of needed functions spaces and the trace operator. Further, we discuss parabolic systems with Dirichlet boundary conditions and their formulation as constrained system. It turns out that this is helpful for the formulation presented in Section~\ref{sec:dynBC}.
\subsection{Function spaces and the trace operator}\label{sec:prelim:spaces}
Throughout this paper, $\Omega \subset \mathbb{R}^d$ denotes a bounded Lipschitz domain with boundary $\Gamma:= \partial\Omega$ and $[0, T]$ the bounded time interval of interest. We consider the standard Sobolev spaces
\[
V := H^1(\Omega), \qquad
V_0 := H^1_0(\Omega), \qquad
Q := H^{-1/2}(\Gamma).
\]
The corresponding dual spaces are given by~$V^*$,~$V_0^*=H^{-1}(\Omega)$, and~$Q^*= H^{1/2}(\Gamma)$, respectively. With the pivot space~$H:=L^2(\Omega)$ this leads to the Gelfand triples $V, H, V^*$ and $V_0, H, V^*_0$, see~\cite[Ch.~23.4]{Zei90a}. Also the trace spaces form a Gelfand triple, namely with the pivot space $L^2(\Gamma)$.
In the bulk as well as on the boundary, we write $(\, \cdot\, , \cdot\, )$ for the inner product of the corresponding pivot space, i.e., for $u, v \in H$ and $p, q \in L^2(\Gamma)$ we have
\[
(u, v)
:= (u, v)_{L^2(\Omega)}
= \int_\Omega u(x)\, v(x) \,\text{d}x, \qquad
(p, q)
:= (p, q)_{L^2(\Gamma)}
= \int_\Gamma p(x)\, q(x) \,\text{d}x.
\]
Further, we use duality pairings, i.e., for $u\in V$ and $f\in V^*$ we write $\langle f, u\rangle := \langle f, u\rangle_{V^*,V}$. If we have~$f\in L^2(\Omega)$, then the embedding given by the Gelfand triple implies~$\langle f, u\rangle = (f, u)$. Since we deal with time-dependent problems, the solution spaces will be Sobolev-Bochner spaces, cf.~\cite[Ch.~25]{Wlo87}. This includes spaces of the form~$L^2(0,T;V)$, i.e, abstract functions~$u\colon [0,T] \to V$ with $\int_0^T \Vert u(t)\Vert^2_V \,\text{d}t < \infty$.
A decisive role in this paper plays the trace operator, which we denote by~$\ensuremath{\mathcal{D}} \colon V \to Q^*$. Recall that this operator is defined through~$\ensuremath{\mathcal{D}} u := u|_\Gamma$ for continuous functions $u \in C^0(\overline \Omega) \cap V$, which are dense in~$V$. For prescribed boundary data~$g\in Q^*$, Dirichlet boundary conditions then simply read~$\ensuremath{\mathcal{D}} u = g$. The trace operator~$\ensuremath{\mathcal{D}} $ is surjective~\cite[Th.~2.21]{Ste08} and satisfies an inf-sup condition \cite[Lem.~4.7]{Ste08}. The corresponding dual operator~$\ensuremath{\mathcal{D}} ^*$ maps from~$Q$ to~$V^*$ and is defined by~$\langle \ensuremath{\mathcal{D}} ^* \lambda, u \rangle := \langle \lambda, \ensuremath{\mathcal{D}} u \rangle_{Q,Q^*}$ for all~$u\in V$ and~$\lambda \in Q$.
\subsection{Homogeneous Dirichlet boundary conditions}\label{sec:prelim:homDirichlet}
In order to get used to the abstract setting, we first consider the homogeneous case, i.e., we consider the parabolic equation
\begin{align}
\label{eq:homDirichletBC}
\dot u - \nabla\cdot(\kappa \nabla u) = f \quad\text{in } \Omega, \qquad
u = 0 \quad\text{on } \Gamma
\end{align}
with initial condition $u(0) = \ensuremath{u_0}$. For the diffusion parameter~$\kappa \in L^\infty(\Omega)$ we assume~$\kappa(x) \ge c_\kappa > 0$ for almost all~$x\in \Omega$. The corresponding weak formulation, which we obtain by an application of the integration by parts formula, can be written as an operator equation, namely
\begin{align}
\label{eq:PDAE:homDirichlet}
\dot u + \ensuremath{\mathcal{K}} u = f \qquad \text{in } V_0^*.
\end{align}
This equation is stated in~$V_0^*$, meaning that we consider test functions in $V_0$. The linear operator~$\ensuremath{\mathcal{K}} \colon V\to V^*$ is defined for~$u, v\in V$ by
\begin{align}
\label{def:calK}
\langle \ensuremath{\mathcal{K}} u, v\rangl
:= \int_{\Omega} (\kappa \nabla u) \cdot \nabla v \,\text{d}x.
\end{align}
This operator is continuous and, due to the assumptions on~$\kappa$, elliptic on the subspace~$V_0$. We emphasize that we may replace~$\ensuremath{\mathcal{K}} $ by any other continuous operator mapping from~$V$ to~$V^*$ and satisfying a G\aa{}rding inequality on~$V_0$. This then leads to the well-known result that~$f\in L^2(0,T;V_0^*)$ and~$\ensuremath{u_0} \in H$ imply the existence of a unique solution of~\eqref{eq:PDAE:homDirichlet} satisfying~$u \in L^2(0,T;V_0)$ and~$\dot u \in L^2(0,T;V_0^*)$, cf.~\cite[Ch.~26]{Wlo87}.
\subsection{Inhomogeneous Dirichlet boundary conditions}\label{sec:prelim:Dirichlet}
We now turn to general time-dependent Dirichlet boundary conditions. For this, we replace the homogeneous boundary condition in~\eqref{eq:homDirichletBC} by~$u = g$ on~$\Gamma$ for given data~$g\colon [0,T] \to Q^*$.
The standard procedure is to construct a function~$u_g$ on the domain~$\Omega$ with~${u_g}|_\Gamma = g$ in each time point. Then, one considers~$\tilde u := u - u_g$ as the new unknown. Assuming sufficient regularity, this leads to an operator equation for~$\tilde u$ similar to~\eqref{eq:PDAE:homDirichlet} but with an adapted right-hand side and initial condition.
As preparation for the following section, we rather consider the formulation as constrained operator equation, where the boundary condition is given as an explicit constraint, cf.~\cite{Sim00,Alt15}. This then leads to a PDAE. In contrast to the homogeneous case, we consider test functions in~$V$ rather than~$V_0$ and introduce a Lagrange multiplier~$\lambda\colon [0,T] \to Q$ in order to enforce the constraint~$\ensuremath{\mathcal{D}} u = g$. This then leads to the constrained system
\begin{subequations}
\label{eq:PDAE:Dirichlet}
\begin{align}
\dot u + \ensuremath{\mathcal{K}} u + \ensuremath{\mathcal{D}} ^*\lambda &= f\qquad\text{in } V^*, \label{eq:PDAE:Dirichlet:a} \\
\ensuremath{\mathcal{D}} u \phantom{j + \ensuremath{\mathcal{D}} \lambda} &= g\qquad\text{in } Q^*. \label{eq:PDAE:Dirichlet:b}
\end{align}
\end{subequations}
The solution $u$ of~\eqref{eq:PDAE:Dirichlet} obviously satisfies the boundary conditions. Further, restricting the test functions in~\eqref{eq:PDAE:Dirichlet:a} to $V_0$, we obtain~$\dot u + \ensuremath{\mathcal{K}} u = f$. In this particular case, the Lagrange multiplier allows the physical interpretation of the normal derivative of $u$ along the boundary~\cite{Sim00}.
\begin{theorem}
Assume right-hand sides~$f\in L^2(0,T;V^*)$, $g\in H^1(0,T;Q^*)$ and a consistent initial condition $u(0) = \ensuremath{u_0} \in V$, i.e., $\ensuremath{\mathcal{D}} \ensuremath{u_0} = g(0)$. Then, system~\eqref{eq:PDAE:Dirichlet} has a unique distributional solution satisfying~$u \in L^2(0,T;V)$ and~$\dot u + \ensuremath{\mathcal{D}} ^* \lambda \in L^2(0,T;V^*)$.
\end{theorem}
\begin{proof}
The result follows by the positivity of~$\ensuremath{\mathcal{K}} $ and the inf-sup stability of~$\ensuremath{\mathcal{D}} $, cf.~\cite[Th.~3.3]{EmmM13}.
\end{proof}
\section{Dynamic boundary conditions}\label{sec:dynBC}
Finally, we address linear parabolic problems with dynamic boundary conditions, i.e.,
\begin{subequations}
\label{eq:dynamicBC}
\begin{align}
\dot u - \nabla\cdot(\kappa \nabla u) &= f \qquad\text{in } \Omega, \label{eq:dynamicBC:a} \\
\dot u - \beta\, \Delta_\Gamma u + \ensuremath{\partial_{\kappa,n}} u + \alpha\, u &= g \qquad \text{on } \Gamma \label{eq:dynamicBC:b}
\end{align}
\end{subequations}
with initial condition $u(0) = \ensuremath{u_0}$. We denote the unit normal vector by~$n$ and the corresponding normal derivative by~$\partial_{\kappa,n} u := n\cdot(\kappa \nabla u)$. The Laplace-Beltrami operator is denoted by~$\Delta_\Gamma$, see~\cite[Ch.~16.1]{GilT01}.
For the parameters we assume~$\kappa$ to be positive as before, $\alpha \in L^\infty(\Omega)$, and~$\beta\ge 0$ constant. The boundary condition~\eqref{eq:dynamicBC:b} is called \emph{locally reacting} for $\beta= 0$ and \emph{non-local} otherwise. The aim of this section is to derive an operator formulation of~\eqref{eq:dynamicBC} as a coupled system.
\subsection{Locally reacting dynamic boundary conditions}\label{sec:dynBC:local}
We consider dynamic boundary conditions without the Laplace-Beltrami operator, i.e., we assume $\beta= 0$. These so-called \emph{Wentzell boundary conditions} then read
\[
\dot u + \ensuremath{\partial_{\kappa,n}} u + \alpha\, u
= g \qquad \text{on } \Gamma.
\]
We now introduce a new variable, namely $p:=u|_\Gamma$, such that system~\eqref{eq:dynamicBC} with $\beta=0$ is equivalent to
\begin{subequations}
\label{eq:dynBC:local:newVar}
\begin{align}
\dot u - \nabla\cdot(\kappa \nabla u) &= f \qquad\text{in } \Omega, \label{eq:dynBC:local:newVar:a} \\
\dot p + \ensuremath{\partial_{\kappa,n}} u + \alpha\, p &= g \qquad \text{on } \Gamma, \label{eq:dynBC:local:newVar:b} \\
p - u &= 0 \qquad\text{on } \Gamma. \label{eq:dynBC:local:newVar:c}
\end{align}
\end{subequations}
Note that~\eqref{eq:dynBC:local:newVar:a} and~\eqref{eq:dynBC:local:newVar:b} are dynamic equations for~$u$ and~$p$, respectively, whereas equation~\eqref{eq:dynBC:local:newVar:c} marks the coupling condition. For the weak formulation we define the spaces
\[
\ensuremath{\mathcal{V}} := H^1(\Omega) \times H^{1/2}(\Gamma), \qquad
\ensuremath{\mathcal{H}} := L^2(\Omega) \times L^{2}(\Gamma), \qquad
\ensuremath{\mathcal{Q}} := H^{-1/2}(\Gamma).
\]
On~$\ensuremath{\mathcal{V}}$ and~$\ensuremath{\mathcal{H}}$ we define the norms~$\Vert \big[ \substack{u\\ p} \big] \Vert^2_\ensuremath{\mathcal{V}} := \Vert u \Vert^2_{H^1(\Omega)} + \Vert p \Vert^2_{H^{1/2}(\Gamma)}$ and $\Vert \big[ \substack{u\\ p} \big] \Vert^2_\ensuremath{\mathcal{H}} := \Vert u \Vert^2_{L^2(\Omega)} + \Vert p \Vert^2_{L^{2}(\Gamma)}$, respectively. Further, we define the coupling operator $\ensuremath{\mathcal{B}} \colon \ensuremath{\mathcal{V}} \to \ensuremath{\mathcal{Q}} ^*$ as
\[
\ensuremath{\mathcal{B}} \, \big[ \substack{u\\ p} \big]
:= p - \ensuremath{\mathcal{D}} u \in \ensuremath{\mathcal{Q}} ^*.
\]
The dual operator $\ensuremath{\mathcal{B}} ^*\colon \ensuremath{\mathcal{Q}} \to \ensuremath{\mathcal{V}}^*$ is defined accordingly via~$\langle \ensuremath{\mathcal{B}} ^* \lambda, \big[ \substack{u\\ p} \big] \rangle := \langle \lambda, p - \ensuremath{\mathcal{D}} u \rangle_{\ensuremath{\mathcal{Q}} , \ensuremath{\mathcal{Q}} ^*}$. We now consider a test function $\big[ \substack{v\\ q} \big] \in\ensuremath{\mathcal{V}}$ and multiply equation~\eqref{eq:dynBC:local:newVar:a} by~$v$ and~\eqref{eq:dynBC:local:newVar:b} by~$q$. With~$\ensuremath{\mathcal{K}} $ as defined in~\eqref{def:calK}, integration by parts and taking the sum leads to
\begin{align*}
\langle f, v\rangle + (g, q)
&= (\dot u, v) + \langle \ensuremath{\mathcal{K}} u, v \rangle - \langle \ensuremath{\partial_{\kappa,n}} u, \ensuremath{\mathcal{D}} v \rangle
+ (\dot p, q) + \langle \ensuremath{\partial_{\kappa,n}} u, q \rangle + (\alpha p,q) \\
&= (\dot u, v) + (\dot p, q) + \langle \ensuremath{\mathcal{K}} u, v \rangle
+ \langle \ensuremath{\partial_{\kappa,n}} u, \ensuremath{\mathcal{B}} \big[ \substack{v\\ q} \big]\, \rangle + (\alpha p,q).
\end{align*}
Proceeding similarly as in Section~\ref{sec:prelim:Dirichlet}, we introduce a Lagrange multiplier $\lambda\colon [0,T] \to \ensuremath{\mathcal{Q}} $ in place of the normal trace $\ensuremath{\partial_{\kappa,n}} u$. This then leads to the PDAE
\begin{subequations}
\label{eq:PDAE:local}
\begin{align}
\begin{bmatrix} \dot u \\ \dot p \end{bmatrix}
+ \begin{bmatrix} \ensuremath{\mathcal{K}} & \\ & \alpha \end{bmatrix}
\begin{bmatrix} u \\ p \end{bmatrix}
+ \ensuremath{\mathcal{B}} ^*\lambda
&= \begin{bmatrix} f \\ g \end{bmatrix} \qquad \text{in } \ensuremath{\mathcal{V}}^*, \\
\ensuremath{\mathcal{B}} \, \begin{bmatrix} u \\ p \end{bmatrix} \phantom{i + \ensuremath{\mathcal{B}} \lambda}
&= \phantom{[]} 0\hspace{2.751em} \text{in } \ensuremath{\mathcal{Q}} ^*.
\end{align}
\end{subequations}
with initial conditions~$u(0) = \ensuremath{u_0}$ and~$p(0) = \ensuremath{p_0}$. Note that the structure of this operator equation is similar to the pure Dirichlet case in~\eqref{eq:PDAE:Dirichlet}. For the well-posedness one needs to discuss the inf-sup stability of~$\ensuremath{\mathcal{B}} $ as well as the positivity of the differential operator~$\diag(\ensuremath{\mathcal{K}} , \alpha)$.
\begin{lemma}
\label{lem:infsup:local}
The coupling operator~$\ensuremath{\mathcal{B}} \colon \ensuremath{\mathcal{V}} \to \ensuremath{\mathcal{Q}} ^*$ is inf-sup stable with constant~$1$, i.e.,
\[
\adjustlimits \inf_{q\in \ensuremath{\mathcal{Q}} } \sup_{\big[ \substack{u\\ p} \big] \in \ensuremath{\mathcal{V}}}
\frac{\big\langle \ensuremath{\mathcal{B}} \big[ \substack{u\\ p} \big],\, q \big\rangle_{\ensuremath{\mathcal{Q}} ^*, \ensuremath{\mathcal{Q}} }}{\Vert \big[ \substack{u\\ p} \big] \Vert_\ensuremath{\mathcal{V}}\, \Vert q\Vert_\ensuremath{\mathcal{Q}} }
\ge 1.
\]
Further, the operator~$\diag(\ensuremath{\mathcal{K}} , \alpha)\colon \ensuremath{\mathcal{V}}\to\ensuremath{\mathcal{V}}^*$ is continuous and satisfies a G\aa{}rding inequality on~$\ensuremath{\mathcal{V}}_0 := \ker \ensuremath{\mathcal{B}} $.
\end{lemma}
\begin{proof}
For the inf-sup stability let~$q\in \ensuremath{\mathcal{Q}} = H^{-1/2}(\Gamma)$ be arbitrary with norm~$1$. We set $\tilde u = 0$ and $\tilde p$ as the Riesz representation of $q$ in the Hilbert space $\ensuremath{\mathcal{Q}} ^*= H^{1/2}(\Gamma)$ such that
\[
\sup_{\big[ \substack{u\\ p} \big] \in \ensuremath{\mathcal{V}}}
\frac{\big\langle \ensuremath{\mathcal{B}} \big[ \substack{u\\ p} \big],\, q \big\rangle}{\Vert \big[ \substack{u\\ p} \big] \Vert_\ensuremath{\mathcal{V}} }
\ge \frac{\langle \tilde p - \ensuremath{\mathcal{D}} \tilde u,\, q \rangle}{\Vert \big[ \substack{\tilde u\\ \tilde p} \big] \Vert_\ensuremath{\mathcal{V}}}
= \frac{\langle \tilde p,\, q \rangle}{\Vert \tilde p \Vert_{H^{1/2}(\Gamma)}}
= \frac{\Vert \tilde p \Vert^2_{\ensuremath{\mathcal{Q}} ^*}}{\Vert \tilde p \Vert_{\ensuremath{\mathcal{Q}} ^*}}
= \Vert \tilde p \Vert_{\ensuremath{\mathcal{Q}} ^*}
= \Vert q \Vert_{\ensuremath{\mathcal{Q}} }
= 1.
\]
The continuity of~$\diag(\ensuremath{\mathcal{K}} , \alpha)$ follows directly form the continuity of~$\ensuremath{\mathcal{K}} $ and the continuity of the embedding~$L^{2}(\Gamma) \hookrightarrow H^{-1/2}(\Gamma)$. For the G\aa{}rding inequality, we consider~$\big[\substack{u\\ p} \big] \in \ensuremath{\mathcal{V}}_0 = \{ \big[ \substack{u\\ p} \big] \in \ensuremath{\mathcal{V}}\, |\, \ensuremath{\mathcal{D}} u = p \}$ and estimate with the continuity constant of the trace operator~$C_\text{tr}$ and $C_\alpha := \Vert \alpha\Vert_{L^\infty(\Omega)}$,
\begin{align*}
\Big\langle \begin{bmatrix} \ensuremath{\mathcal{K}} & \\ & \alpha \end{bmatrix}
\begin{bmatrix} u \\ p \end{bmatrix}, \begin{bmatrix} u \\ p \end{bmatrix} \Big\rangle
&\ge c_\kappa \Vert u \Vert^2_{H^1(\Omega)} - c_\kappa\Vert u \Vert^2_{L^2(\Omega)} + (\alpha p, p) \\
&\ge \frac{c_\kappa}{2} \Vert u \Vert^2_{H^1(\Omega)} + \frac{c_\kappa}{2 C^2_\text{tr}} \Vert p \Vert^2_{H^{1/2}(\Gamma)} - c_\kappa\Vert u \Vert^2_{L^2(\Omega)} - C_\alpha \Vert p \Vert^2_{L^2(\Gamma)} \\
&\gtrsim \Vert \big[ \substack{u\\ p} \big] \Vert^2_{\ensuremath{\mathcal{V}}} - c\, \Vert \big[ \substack{u\\ p} \big] \Vert^2_{\ensuremath{\mathcal{H}}}.
\end{align*}
Note that we have applied the fact that~$u$ and~$p$ can be replaced by each other on the boundary.
\end{proof}
Using again~\cite[Th.~3.3]{EmmM13}, we obtain by the previous lemma the following well-posedness result.
\begin{theorem}
Given a bounded Lipschitz domain~$\Omega$, we assume~$\big[ \substack{f\\ g} \big] \in L^2(0,T;\ensuremath{\mathcal{V}}^*)$ and consistent initial data~$\big[ \substack{\ensuremath{u_0}\\ \ensuremath{p_0}} \big] \in \ensuremath{\mathcal{V}}_0$. Then, system~\eqref{eq:PDAE:local} has a unique distributional solution~$\big[ \substack{u\\ p} \big] \in L^2(0,T; \ensuremath{\mathcal{V}})$, i.e, $u \in L^2(0,T; H^1(\Omega))$~and~$p \in L^2(0,T; H^{1/2}(\Gamma))$.
\end{theorem}
\begin{remark}
We emphasize that the condition on the initial value may be relaxed to~$\big[ \substack{\ensuremath{u_0}\\ \ensuremath{p_0}} \big] \in \ensuremath{\mathcal{H}}$, cf.~\cite[Rem.~6.9]{Alt15}. Further, we like to mention the connection to the abstract setting introduced in~\cite{KovL17}. Therein, the ansatz space equals~$\ensuremath{\mathcal{V}}_0=\ker\ensuremath{\mathcal{B}} $, i.e., the connection of~$u$ and the boundary variable~$p$ is given a priori.
\end{remark}
\subsection{Nonlocal dynamic boundary conditions}\label{sec:dynBC:nonlocal}
Finally, we consider the case~$\beta > 0$. Due to the Laplace-Beltrami operator the boundary condition is then non-local. As in the previous subsection, we introduce~$p:=u|_\Gamma$ such that~\eqref{eq:dynamicBC} is equivalent to
\begin{subequations}
\label{eq:dynBC:nonLocal:newVar}
\begin{align}
\dot u - \nabla\cdot(\kappa \nabla u) &= f \qquad\text{in } \Omega, \label{eq:dynBC:nonLocal:newVar:a} \\
\dot p - \beta\, \Delta_\Gamma p + \ensuremath{\partial_{\kappa,n}} u + \alpha\, p &= g \qquad \text{on } \Gamma, \label{eq:dynBC:nonLocal:newVar:b} \\
p - u &= 0 \qquad\text{on } \Gamma. \label{eq:dynBC:nonLocal:newVar:c}
\end{align}
\end{subequations}
The presence of the Laplace-Beltrami operator requires an adjustment of the ansatz spaces to
\[
\ensuremath{\mathcal{V}} := H^{1}(\Omega) \times H^{1}(\Gamma), \qquad
\ensuremath{\mathcal{H}} := L^2(\Omega) \times L^{2}(\Gamma), \qquad
\ensuremath{\mathcal{Q}} := H^{-1/2}(\Gamma)
\]
with the corresponding norm $\Vert \big[ \substack{u\\ p} \big] \Vert^2_\ensuremath{\mathcal{V}} := \Vert u \Vert^2_{H^1(\Omega)} + \Vert p \Vert^2_{H^{1}(\Gamma)}$. The coupling operator $\ensuremath{\mathcal{B}} \colon \ensuremath{\mathcal{V}} \to \ensuremath{\mathcal{Q}} ^*$ remains unchanged.
In order to obtain an operator formulation of~\eqref{eq:dynBC:nonLocal:newVar} we again derive the weak formulation via integration by parts. Here, we also have to integrate by parts on the boundary leading to the weak form of the Laplace-Beltrami operator, which we denote by~$\ensuremath{\mathcal{K}} _\Gamma\colon H^{1}(\Gamma) \to H^{-1}(\Gamma)$. With a Lagrange multiplier~$\lambda\colon [0,T] \to \ensuremath{\mathcal{Q}} $ this then leads to
\begin{subequations}
\label{eq:PDAE:nonlocal}
\begin{align}
\begin{bmatrix} \dot u \\ \dot p \end{bmatrix}
+ \begin{bmatrix} \ensuremath{\mathcal{K}} & \\ & (\ensuremath{\mathcal{K}} _\Gamma + \alpha) \end{bmatrix}
\begin{bmatrix} u \\ p \end{bmatrix}
+ \ensuremath{\mathcal{B}} ^*\lambda
&= \begin{bmatrix} f \\ g \end{bmatrix} \qquad \text{in } \ensuremath{\mathcal{V}}^*, \\
\ensuremath{\mathcal{B}} \, \begin{bmatrix} u \\ p \end{bmatrix} \phantom{i + \ensuremath{\mathcal{B}} \lambda} &= \phantom{[]} 0\hspace{2.751em} \text{in } \ensuremath{\mathcal{Q}} ^*.
\end{align}
\end{subequations}
For the well-posedness, we consider the following lemma.
\begin{lemma}
\label{lem:infsup:nonLocal}
Let~$C_\text{invTr}$ denote the continuity constant of the inverse trace theorem, cf.~\cite[Th.~2.22]{Ste08}. Then, the operator $\ensuremath{\mathcal{B}} \colon \ensuremath{\mathcal{V}} \to \ensuremath{\mathcal{Q}} $ is inf-sup stable with constant~$1/C_\text{invTr}$.
Further, the differential operator~$\diag(\ensuremath{\mathcal{K}} , \ensuremath{\mathcal{K}} _\Gamma+\alpha)\colon \ensuremath{\mathcal{V}}\to\ensuremath{\mathcal{V}}^*$ is continuous and satisfies a G\aa{}rding inequality on~$\ensuremath{\mathcal{V}}$.
\end{lemma}
\begin{proof}
Let $q\in \ensuremath{\mathcal{Q}} $ be arbitrary with norm~$1$ and $\tilde q \in \ensuremath{\mathcal{Q}} ^*$ its Riesz representation. Further, set~$\tilde p = 0$ and $\tilde u \in V$ as the weak solution of $-\Delta \tilde u = 0$ with Dirichlet boundary condition $\ensuremath{\mathcal{D}} \tilde u = -\tilde q$. The latter implies~$\Vert \tilde u \Vert_V \le C_\text{invTr} \Vert \tilde q\Vert_{\ensuremath{\mathcal{Q}} ^*}$, which leads to the estimate
\[
\sup_{\big[ \substack{u\\ p} \big] \in \ensuremath{\mathcal{V}}}
\frac{\big\langle \ensuremath{\mathcal{B}} \big[ \substack{u\\ p} \big],\, q \big\rangle}{\Vert \big[ \substack{u\\ p} \big] \Vert_\ensuremath{\mathcal{V}} }
\ge \frac{\langle \tilde p - \ensuremath{\mathcal{D}} \tilde u,\, q \rangle}{\Vert \big[ \substack{\tilde u\\ \tilde p} \big] \Vert_\ensuremath{\mathcal{V}}}
= \frac{\langle \tilde q,\, q \rangle}{\Vert \tilde u \Vert_V}
= \frac{\Vert \tilde q \Vert^2_{H^{1/2}(\Gamma)}}{\Vert \tilde u \Vert_{H^1(\Omega)}}
\ge \frac{\Vert \tilde q \Vert_{\ensuremath{\mathcal{Q}} ^*}}{C_\text{invTr}}
= \frac{1}{C_\text{invTr}}.
\]
The remaining part of the proof is similar to the one of~Lemma~\ref{lem:infsup:local}, using the fact that the Laplace-Beltrami operator satisfies a G\aa{}rding inequality on~$H^1(\Gamma)$, cf.~\cite[Ch.~16.1]{GilT01}.
\end{proof}
As a direct consequence of Lemma~\ref{lem:infsup:nonLocal} we conclude the unique solvability of the PDAE~\eqref{eq:PDAE:nonlocal}.
\begin{theorem}
Given a bounded Lipschitz domain~$\Omega$, a right-hand side~$\big[ \substack{f\\ g} \big] \in L^2(0,T;\ensuremath{\mathcal{V}}^*)$, and consistent initial data~$\big[ \substack{\ensuremath{u_0}\\ \ensuremath{p_0}} \big] \in \ensuremath{\mathcal{V}}_0$, there exists a unique distributional solution of~\eqref{eq:PDAE:nonlocal} with~$\big[ \substack{u\\ p} \big] \in L^2(0,T; \ensuremath{\mathcal{V}})$, i.e, $u \in L^2(0,T; H^1(\Omega))$~and~$p \in L^2(0,T; H^{1}(\Gamma))$.
\end{theorem}
\section{Conclusion and discussion}
We have presented an alternative abstract formulation of parabolic systems with dynamic boundary conditions, namely as constrained operator system. This novel approach aims to provide a new perspective on these type of systems. In particular, this is of interest for the construction of new robust discretization methods. Since we include the coupling of~$u$ and~$p$ in form of an equation rather than restricting the ansatz space as in~\cite{KovL17}, we allow different spatial discretizations in the bulk and on the boundary. Further, the obtained structure in~\eqref{eq:PDAE:local} and~\eqref{eq:PDAE:nonlocal} enables a direct application of algebraically stable Runge-Kutta schemes~\cite{AltZ18}.
\section*{Acknowledgement}
The author thanks David Hipp (KIT Karlsruhe) and Christoph Zimmer (TU Berlin) for valuable discussions on all sorts of boundary conditions.
|
\section{Introduction}
Recent years have seen a flurry of activity aimed at understanding
the motion of fluid particles in free-surface flows. The problem
has been studied from various point of view, including field
campaigns, wave-flume experiments, asymptotics, and rigorous mathematical
analysis. One of the most well known results about particle trajectories in
free-surface flows is Stokes's finding that the particle motion
associated to small-amplitude (linear) periodic waves features
a net forward drift which is attributed to the decrease of the
Eulerian particle velocity with increasing depth.
Recent advances in laboratory technology, in particular those relating to
particle image velocimitry (PIV) and particle tracking velocimitry (PTV)
have led to rapid improvements of our understanding of the motion of
fluid particles in free-surface flow.
The influence of the boundary layers
on the particle drift in a regular wave train was studied in the seminal paper \cite{LH1953}
which was in part inspired by experiments reported in \cite{Bagnold}.
In essence, particle drift will be positive near the bottom and
near the free surface but negative in an intermediate region.
In recent works, experimental measurements are coupled with high-performance data analysis
techniques to paint a rather complete picture of particle motion
in highly nonlinear waves created in a wave flume \cite{GK1,GK2}.
The findings presented in these works are also related to the importance
of the effect of the boundary layer both at the bed and at the free surface.
Indeed, it is observed that the Stokes drift may
take a very different form than what was originally found by Stokes \cite{Stokes}.
In particular, these recent studies confirm and extend
the results of \cite{LH1953}.
However, depending on the motion of the wave maker, these results may vary.
For example, the experiments reported in \cite{ChHsCh} seem to confirm
the essence of Stokes's original work. Indeed,
using both experiments and high-order asymptotics, a strong
case is made in \cite{ChHsCh} that there is a net forward drift throughout the fluid column.
These results are also in line with
mathematical advances in the understanding
of particle motion in free surface flow (see \cite{EV} for a review).
In particular, a firm mathematical
proof was given that particle trajectories in Stokes waves are not closed \cite{Co}.
In the case of finite depth, even if viscosity is not taken into
account, mass conservation during the creation phase may lead
to zero drift when averaged over the fluid column \cite{LH1953,RamsdenNath1988}.
On the other hand, in the current analysis, the case of (infinitely) deep water
is investigated and the bottom boundary layer is ignored.
In some studies, in particular in the presence of a background
current in deep water, it has been observed that there is no (Lagrangian) particle
drift either in the average, or even in the pointwise sense \cite{Monismith}.
Some reports of field campaigns also point to the absence of the Stokes
drift in wave trains in the open ocean \cite{Smith}. With these partially conflicting
pieces of evidence, it appears that there is a need for being able to describe
particle paths as accurately as possible using theoretical models.
In the current contribution, we study particle trajectories due to wave motion
described in the narrow-banded spectrum approximation in an idealized infinitely deep fluid.
We focus on irrotational flow which is described in terms of a potential function,
$\phi(x,z,t)$, satisfying Laplace's equation, $\phi_{xx} + \phi_{zz} = 0$, in the fluid domain.
In general terms, if $(\xi(t),\zeta(t))$ denotes the position of a fluid particle in the $(x,z)$-plane
at a time $t$, then the particle motion is described by the ordinary differential equations
\begin{subequations}
\begin{equation}
\frac{d\xi}{dt}=\frac{\partial \phi}{\partial
x}\Big{(}\xi(t),\zeta(t),t\Big{)},
\end{equation}
\begin{equation}
\frac{d\zeta}{dt}=\frac{\partial \phi}{\partial
z}\Big{(}\xi(t),\zeta(t),t\Big{)}.
\end{equation}
\label{LODEs}
\end{subequations}
The corresponding initial conditions are given by $(\xi(0),\zeta(0))=(\xi_0,\zeta_0)$
where $(\xi_0,\zeta_0)$ represents the initial coordinates of the particle in the $(x,z)$-plane.
In this paper, all variables are dimensional with standard SI units.
As stated above, the focus of this paper is to examine particle trajectories
associated with wave motion which can be approximately described by the nonlinear Schr\"{o}dinger (NLS)
equation and a few of its generalizations. This equation arises in the case where a carrier wave of a certain frequency is modulated slowly. In terms of the dimensional slow variables $X = \epsilon x$ and $T = \epsilon t$,
the NLS equation is written as
\begin{equation}
2i \omega_0\Big( B_T+\frac{g}{2\omega_0}B_X\Big{)}+\epsilon\Big{(}\frac{g}{4k_0}B_{XX}+4gk_0^3|B|^2B\Big) = 0,
\label{NLS}
\end{equation}
where $g=9.8 m/s^2$ is the gravitational acceleration,
$B$ describes the dimensional amplitude envelope of the oscillations of the carrier wave,
$k_0>0$ represents the dimensional wave number of the carrier wave,
$\omega_0 = \sqrt{g k_0}$ represents the dimensional frequency of the carrier wave,
and $\epsilon=2|a_0|k_0$ represents the (dimensionless) wave steepness.
This equation has been well studied mathematically
(see, for example, \cite{SulemSulem})
and has been shown to favorably predict experimental measurements
of modulated wave trains when $\epsilon<0.1$ (see for example \cite{LoMei}).
Given a solution, $B(X,T)$, the free surface is reconstructed by
\begin{equation}\label{etaIntro}
\eta(x,t)=\epsilon^3\bar{\eta}+\epsilon
B\mbox{e}^{i\omega_0t-ik_0x}+\epsilon^2B_2\mbox{e}^{2(i\omega_0t-ik_0x)} + \dots + c.c.,
\end{equation}
where $\bar{\eta}$ is a dimensional average term and $B_2$ is the dimensional amplitude of the first harmonic. These quantities are defined in terms of $B$ and are examined in detail in Section 2.1. The abbreviation $c.c.$ stands for complex conjugate. Additionally, in the derivation of the NLS equation, the following ansatz is used for the velocity potential
\begin{equation}\label{phiIntro}
\phi(x,z,t)=\epsilon^2\bar{\phi}+\epsilon A \mbox{e}^{k_0z+i\omega_0t-ik_0x}
+ \dots + c.c.,
\end{equation}
where $\bar{\phi}=\bar\phi(X,Z,T)$, $A=A(X,Z,T)$, and $Z=\epsilon z$ is the dimensional slow vertical variable. Substituting this expression into the Laplace equation leads to the following transport equation with imaginary speed
\begin{equation}\label{transport}
A_{Z} - iA_{X}=0,
\end{equation}
at leading order in $\epsilon$. In addition, at leading order in $\epsilon$ the Bernoulli condition gives
\begin{equation}\label{transportData}
=\frac{i\omega_0}{k_0}B,\hspace*{1cm}\text{at $Z=0$.}
\end{equation}
Solving this boundary-value problem for $A$ yields the first-order approximation to the velocity field everywhere in the fluid. It is then possible to compute the particle trajectories associated to any given surface profile numerically. A number of examples for particle paths that result from NLS solutions are given in the remainder of this section. Examples of particle paths that result from solutions to higher-order and dissipative generalizations of NLS are examined in the following sections. Similar considerations regarding particle paths and the potential function have been introduced for a variety of long-wave equations in \cite{AK2,AK4}. In particular, particle paths and streamlines for waves in the KdV regime were described in detail in \cite{BoKa,He,He3}. Somewhat different procedures have also recently been used to understand properties of particle motion in the context of the narrow-banded spectrum approximation in the presence of shear flows \cite{CCK} and in the presence of point vortices \cite{CurtisKalisch}.
\subsection{NLS plane-wave solutions}
The plane-wave solutions of NLS are given by
\begin{equation}
B(X,T)=B_0\mbox{e}^{ikX-i\lambda_{_{NLS}} T},
\label{NLSPWSoln}
\end{equation}
where
\begin{equation}
\lambda_{_{NLS}}=\Big{(}\frac{k}{2k_0}+\frac{\epsilon
k^2}{8k_0^2}-2\epsilon B_0^2k_0^2\Big{)}\omega_0,
\end{equation}
and $B_0$ and $k$ are real constants. Closed-form expressions for the corresponding surface displacement and velocity potential are included in Appendix \ref{etasPWNLS}. The surface displacement corresponding to this solution has a (temporal) period of
\begin{equation}
t_{_{NLS}}^*=\frac{2\pi}{\omega_0-\epsilon\lambda_{_{NLS}}}=\frac{16k_0^2\pi}{(8k_0^2-4\epsilon kk_0-\epsilon^2
k^2+16\epsilon^2B_0^2 k_0^4)\omega_0}.
\label{NLSPeriod}
\end{equation}
For demonstrative purposes, we select $\epsilon=0.1$, $B_0=1$, $k_0=1$ and $k=0$ as parameter values for the plane-wave solution. For these parameter values, the period of the surface is given by
$t_{_{NLS}}^*=100\pi\sqrt{5}/357\dot{=}1.968$,
the crest height is $0.223$, the trough height is $-0.183$, and the wave profile has mean zero.
The crest and trough heights are not symmetric about $z=0$ even though the mean term is zero because $\eta(x,t)$ is comprised of six (complex) Fourier modes. See Figure \ref{NLSPWPlots} for time-series plots of one period of this NLS plane-wave solution and one period of the corresponding surface displacement.
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{NLSPWPlots.eps}
\caption{\small Time-series plots of (a) the real (--) and imaginary (- -) parts of the NLS plane-wave solution at $X=0$ and (b) the corresponding surface displacement at $x=0$ for one period of the carrier wave for $\epsilon=0.1$, $k_0=1$, $B_0=1$, and $k=0$.}
\label{NLSPWPlots}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{NLS3Motion.eps}
\caption{\small Evolution of the free surface and a particle located in the surface as described
by a plane-wave solution of the NLS equation with $\epsilon=0.1$ , $k_0 = 1$, $B_0 = 1$, $k = 0$, and
$(\xi_0 , \zeta_0 ) = (0, 0.223)$.
}
\label{NLS3Motion}
\end{center}
\end{figure}
The paths of NLS plane-wave particles are determined by using \eqref{NLSPWSoln} to determine asymptotic expressions for $\phi$ and $\eta$ that are valid up to $\mathcal{O}(\epsilon^3)$ (the details of this process are described in detail in Section \ref{Construction}) and then numerically integrating the system of ODEs given in equation \eqref{LODEs}. Figure \ref{NLS3Motion} includes a waterfall plot showing the path of a particle that starts on the surface. Figure \ref{NLS3LagPlots}(a) shows the fluid surface and the position of a NLS plane-wave particle that starts on the surface at four different $t$ values. Figure \ref{NLS3LagPlots}(b) shows the paths of three NLS plane-wave particles corresponding to different $\zeta_0$. The top particle starts on the fluid surface while the other two start (and remain) inside the fluid.
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{NLS3LagPP.eps}
\caption{\small (a) The fluid surfaces (four solid curves) and positions (dots) of a NLS plane-wave particle that starts on the surface at four equally spaced times in $t\in[0,\frac{1}{2}t_{_{NLS}}(0.223)]$. The dashed curve represents the path of a particle over an entire period. The black circle and star represent the initial and final positions of the particle respectively. (b) The paths of three NLS plane-wave particles with initial positions $(0,0.223)$, $(0,-0.5)$, and $(0,-1)$ on the interval $t\in[0,5t_{_{NLS}}(0.223)]$.}
\label{NLS3LagPlots}
\end{center}
\end{figure}
Let $t_{_{NLS}}(\zeta_0)$ represent the period of the vertical motion of a particle
with initial $z$ coordinate $\zeta_0$.
Values for $t_{_{NLS}}(\zeta_0)$ were computed by numerically solving equation \eqref{LODEs}
and determining the period of the resulting solution for a range of $\zeta_0$ values.
A plot showing how $t_{_{NLS}}(\zeta_0)$ and $\zeta_0$ are related is included in Figure \ref{NLS3LagData}(a).
Note that
\begin{subequations}
\begin{equation}
t_{_{NLS}}(\zeta_0)>t_{_{NLS}}^*,\hspace*{1cm}\text{when $\zeta_0>-\infty$},
\end{equation}
\begin{equation}
\lim_{\zeta_0\rightarrow-\infty}t_{_{NLS}}(\zeta_0)=t_{_{NLS}}^*.
\end{equation}
\end{subequations}
The horizontal motion of the particles is not periodic because $\xi(t_{_{NLS}}(\zeta_0))>\xi_0$ regardless of the initial position of the particle. (Note that this does not mean the particle always moves to the right.) The difference between the final and initial horizontal positions during one period of vertical motion,
$\xi(t_{_{NLS}}(\zeta_0))-\xi_0$, is known as the horizontal Lagrangian drift. Figure \ref{NLS3LagData}(b) contains a plot relating the average horizontal Lagrangian velocity,
\begin{equation}
u_{_{NLS}}(\zeta_0)=\frac{\xi(t_{_{NLS}}(\zeta_0))-\xi_0}{t_{_{NLS}}(\zeta_0)},
\label{LagVel}
\end{equation}
and $\zeta_0$. As expected from equation \eqref{BottomBC},
both $u_{_{NLS}}(\zeta_0)$ and the horizontal Lagrangian drift limit to zero as $\zeta_0\rightarrow-\infty$.
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{NLS3LagData.eps}
\caption{\small Plots showing how (a) $t_{_{NLS}}$ and (b) $u_{_{NLS}}$
change as $\zeta_0$ changes for NLS plane-wave particles with $\epsilon=0.1$, $k_0=1$, $B_0=1$, and $k=0$.}
\label{NLS3LagData}
\end{center}
\end{figure}
Additional simulations (not shown) establish that increasing $B_0$ increases $t_{_{NLS}}$, $u_{_{NLS}}$ and the horizontal Lagrangian drift in roughly exponential manners. Increasing the wave-number of the solution, $k$, increases $t_{_{NLS}}$, and decreases $u_{_{NLS}}$ and the drift in approximately linear manners.
Equation \eqref{BC1}, the kinematic boundary condition, implies that a particle that starts on the surface stays on the surface. This equation is only approximately satisfied by NLS since NLS is an asymptotic approximation to the Euler equations. Therefore NLS particles that start on the surface of the fluid do not necessarily remain on the surface. A plot of the difference between the particle's vertical position and the fluid surface,
\begin{equation}
\mathcal{D}(t)=\big{|}\zeta(t)-\eta\big{(}\xi(t),t\big{)}\big{|},
\label{Diff}
\end{equation}
is included in Figure \ref{NLS3Diff}. As an additional check on our
work, we computed
\begin{equation}
\mathcal{E}(\epsilon)=\int_0^{t_{_{NLS}}}\mathcal{D}(t)~dt,
\label{error}
\end{equation}
for a variety of values of $\epsilon\in(0,\frac{1}{4}]$. Table
\ref{NLSErrors} contains a summary of these results that establish $\mathcal{E}(\epsilon)\sim\mathcal{O}(\epsilon^4)$ for plane-wave solutions of NLS (as expected because NLS is an $\mathcal{O}(\epsilon^3)$ approximation).
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{NLS3Diff.eps}
\caption{\small A plot of $\mathcal{D}(t)$ for a NLS plane-wave solution with $\epsilon=0.1$, $B_0=1$, $k_0=1$ and $k=0$ on $t\in[0,5t_{_{NLS}}(0.223)]$.}
\label{NLS3Diff}
\end{center}
\end{figure}
\begin{table}
\begin{center}
\begin{tabular}{|c c c c|}
\hline
$j$ & $\epsilon$ & $\mathcal{E}_j$ &
$\frac{\mathcal{E}_j}{\mathcal{E}_{j-1}}$ \\
\hline
1 & $1/4$ & $6.05*10^{-2}$ & ---\\
2 & $1/8$ & $2.89*10^{-3}$ & 20.9\\
3 & $1/16$ & $1.51*10^{-4}$ & 19.1\\
4 & $1/32$ & $8.78*10^{-6}$ & 17.2\\
6 & $1/64$ & $5.32*10^{-7}$ & 16.5\\
7 & $1/128$ & $3.27*10^{-8}$ & 16.2\\
8 & $1/256$ & $2.03*10^{-9}$ & 16.1\\
\hline
\end{tabular}
\caption{\small Table of $\mathcal{E}(\epsilon)$ for eight values of
$\epsilon = 1/2^{j+1}$ for the plane-wave solutions of NLS with $k_0=1$, $B_0=1$, and $k=0$.}
\label{NLSErrors}
\end{center}
\end{table}
\subsection{Cnoidal-wave solutions of NLS}
The cnoidal-wave solutions of NLS are given by
\begin{equation}
B(X,T)=B_0~\mbox{cn}\Big{(}\frac{2\sqrt{2}~B_0k_0^2}{\kappa}\big{(}X-\frac{\omega_0}{2k_0}T\big{)},\kappa\Big{)}\mbox{e}^{i\epsilon
B_0^2k_0^2\omega_0(2\kappa^2-1)T/\kappa^2},
\label{cnsoln}
\end{equation}
where $B_0$ and $\kappa\in[0,1)$ are real parameters. Here $\mbox{cn}(\cdot,\kappa)$ is a Jacobi elliptic function with elliptic modulus $\kappa$ and period $4K(\kappa)$ where $K$ is the complete elliptic integral of the first kind~\cite{bf}. Formulas for the corresponding surface displacement and velocity potential are included in Appendix \ref{etasCNNLS}. There are two situations in which $B$ is periodic in $T$. First, if $\kappa=1/\sqrt{2}$, then $B$ is real for all $X$ and $T$ and has a $T$-period of $\frac{2K(1/\sqrt{2})}{B_0k_0\omega_0}$. In this case, if $B_0=\frac{K(1/\sqrt{2})}{\pi k_0}$, then the solution corresponds to a periodic surface displacement with $t$-period
\begin{equation}
t_{_{CN}}^*=\frac{2\pi}{\epsilon\omega_0}.
\end{equation}
Plots of the NLS cnoidal-wave solution and the corresponding surface displacement for $\kappa=1/\sqrt{2}$, $\epsilon=0.1$, $k_0=1$, and $B_0=K(1/\sqrt{2})/\pi$ are included in Figure \ref{NLSCN1Plots}. Plots of the paths of three particles are included in Figures \ref{CNLagPP1} and \ref{CNxPlot1}. Unlike plane-wave solutions, cnoidal-wave solutions do not have constant magnitude. This means that the vertical motion of the particles (except in rare cases) is quasiperiodic rather than periodic. For example, the particle that starts at a peak on the surface has an initial elevation of $\zeta_0=0.1248$. The vertical motion of this particle has a quasiperiod of $t=20.2047$. After one quasiperiod, the particle reaches an elevation of $\zeta(20.2047)=0.1246$, slightly lower than its initial height. This is caused by the fact that the particle experienced a horizontal drift of 0.4140 during this interval. Due to this quasiperiodicity, the Lagrangian drift and average Lagrangian velocity are not well defined.
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{NLSCN1Plots.eps}
\caption{\small Time-series plots of one period of the NLS cnoidal-wave solution and (b) one period of the corresponding surface displacement for $\kappa=1/\sqrt{2}$, $\epsilon=0.1$, $k_0=1$, and $B_0=K(1/\sqrt{2})/\pi$. Note that the imaginary part of $B$ is zero in this case.}
\label{NLSCN1Plots}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{CNLagPP1.eps}
\caption{\small A plot showing the paths of three particles at different levels on the interval $t\in [0, 30]$ during the propagation of a NLS cnoidal-wave solution with $\kappa=1/\sqrt{2}$, $\epsilon=0.1$, $k_0=1$, and
$B_0=K(1/\sqrt{2})/\pi$. The initial particle positions are $(0,0.1248)$, $(0,-0.25)$, and $(0,-0.5)$.}
\label{CNLagPP1}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{CNxPlot1.eps}
\caption{\small Plots of (a) $\xi(t)$ and (b) $\zeta(t)$ for the three NLS
cnoidal-wave particles in Figure \ref{CNLagPP1}.}
\label{CNxPlot1}
\end{center}
\end{figure}
The second situation in which a $T$-periodic solution is obtained is if $B_0$ and $\kappa\ne\frac{1}{\sqrt{2}}$ satisfy
\begin{equation}
B_0=\frac{\sqrt{2}~\kappa\pi}{\epsilon(2\kappa^2-1)k_0K(\kappa)}.
\end{equation}
In order for this solution to correspond to a $t$-periodic surface displacement, the period of this solution must align with the period of the carrier wave. Enforcing this restriction leads to large-amplitude surface displacements that are well outside of the regime where NLS is expected to be valid. Therefore, for demonstrative purposes, we select $\epsilon=0.1$, $k_0=1$, $\kappa=0.999$ and $B_0=1$ (parameters that lead to a non-periodic surface displacement in the NLS regime). Plots of this solution and the corresponding surface displacement are included in Figure \ref{NLSCN2Plots}. Figures \ref{CNLagPP2} and \ref{CNxPlot2} show how three of these NLS cnoidal-wave particles move in $t$. Just as in the previous case, the vertical motion of the particles is quasiperiodic even though the solution is periodic.
For the cnoidal-wave solutions of NLS, increasing $B_0$ increases the horizontal drift for particles that start on the surface.
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{NLSCN2Plots.eps}
\caption{\small Time-series plots of (a) the real (--) and imaginary (- -) parts of the cnoidal-wave solution to NLS and (b) the corresponding surface displacement corresponding to $\kappa=0.999$, $\epsilon=0.1$, $k_0=1$, and $B_0=1$.}
\label{NLSCN2Plots}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{CNLagPP2.eps}
\caption{\small A plot showing how three NLS cnoidal-wave solutions with
$\epsilon=0.1$, $\kappa=0.999$, $k_0=1$, and $B_0=1$ evolve
over $t\in[0,50]$. The initial positions are $(0,0.220)$,
$(0,-0.333)$, and $(0,-0.666)$.}
\label{CNLagPP2}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{CNxPlot2.eps}
\caption{\small Plots of (a) $\xi(t)$ and (b) $\zeta(t)$ for the three NLS
cnoidal-wave particles in Figure \ref{CNLagPP2}.}
\label{CNxPlot2}
\end{center}
\end{figure}
\subsection{Solitary-wave solutions of NLS}
The solitary-wave solutions of NLS,
\begin{equation}
B(X,T)=B_0~\mbox{sech}\Big{(}2\sqrt{2}~B_0k_0^2\big{(}X-\frac{\omega_0}{2k_0}T\big{)}\Big{)}\mbox{e}^{i\epsilon
B_0^2k_0^2\omega_0T},
\label{sechsoln}
\end{equation}
are obtained from the cnoidal-wave solutions via the limit $\kappa\rightarrow1^-$. Formulas for the corresponding surface displacement and velocity potential are included in Appendix \ref{etasSechNLS}. This solution is not
periodic in $X$ or $T$ and therefore does not lead to a periodic surface displacement. We select the parameter values $\epsilon=0.1$, $k_0=1$, and $B_0=1$ for demonstrative purposes.
Figure \ref{NLSSechPlots} contains plots of this solution and the corresponding surface displacement.
Figures \ref{SechLagPP} and \ref{SechxPlot} show the paths of three NLS solitary-wave particles on the interval $t\in[0,50]$. During $t\in[0,50]$, the horizontal drift for the particle that starts on the surface is approximately $0.6170$.
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{NLSSechPlots.eps}
\caption{\small Time-series plots of (a) the real (--) and imaginary (- -) parts of the solitary-wave solution to NLS and (b) the corresponding surface displacement for $\epsilon=0.1$, $k_0=1$, and $B_0=1$.}
\label{NLSSechPlots}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{SechLagPP.eps}
\caption{\small A plot showing the paths of three particles at different levels on the interval $t\in [0, 50]$ during the propagation of a NLS solitary-wave solution with $\epsilon=0.1$, $k_0=1$, and $B_0=1$. The initial particle positions are $(20,6.33*10^{-4})$ (a point on the surface), $(20,-0.333)$, and $(20,-0.666)$.}
\label{SechLagPP}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{SechxPlot.eps}
\caption{\small Plots of (a) $\xi(t)$ and (b) $\zeta(t)$ for the three NLS solitary-wave particles in Figure \ref{SechLagPP}.}
\label{SechxPlot}
\end{center}
\end{figure}
In the next section, we present the full derivation of the velocity potential
in the more general case of the higher-order NLS (Dysthe) equation
and the viscous models. Then in Section 3, we describe particle paths
for a number of examples.
\section{Construction of the velocity potential}
\label{Construction}
We now give the full details of the reconstruction of the velocity potential
in terms of the unknown $B(X,T)$ of the nonlinear Schr\"{o}dinger equations.
In fact, we will take a slightly more general view by including viscous effects,
and higher-order terms. In order to explain how to find the potential $\phi$,
it is convenient to first review the derivation of the evolution
equation.
\subsection{Derivation of the Viscous Dysthe System}
Wu {\emph{et al.}}~\cite{WuLiuYue} proposed the following system for a two-dimensional, infinitely-deep, weakly-dissipative fluid
\begin{subequations}\label{DDZ}
\begin{equation}
\phi_{xx}+\phi_{zz}=0, \hspace*{1cm}\mbox{for } -\infty<z<\eta,
\label{Laplace}
\end{equation}
\begin{equation}
\phi_t+\frac{1}{2}\big{(}\phi_{x}^2+\phi_{z}^2\big{)}+g\eta=-4\bar{\alpha}\phi_{zz},\hspace*{1cm}\mbox{at
} z=\eta,
\label{BC2}
\end{equation}
\begin{equation}
\eta_t+\eta_x\phi_x=\phi_z,\hspace*{1cm}\mbox{at } z=\eta,
\label{BC1}
\end{equation}
\begin{equation}
\phi_x,~\phi_z\rightarrow0, \hspace*{1cm}\mbox{as } z\rightarrow-\infty.
\label{BottomBC}
\end{equation}
\end{subequations}
Here $\phi=\phi(x,z,t)$ represents the velocity potential of the fluid, $\eta=\eta(x,t) $ represents the free-surface displacement, $g$ represents the acceleration due to gravity, and $\bar{\alpha}>0$ represents dissipation from all sources. The classical Euler equations (see, for example, Debnath~\cite{Debnath}) are recovered from this system by setting $\bar{\alpha}=0$. Following the work of Zakharov~\cite{zak68} and Dysthe~\cite{Dysthe},
we make the following modulated wave-train ansatz
\begin{subequations
\begin{equation}
\eta(x,t)=\epsilon^3\bar{\eta}+\epsilon
B\mbox{e}^{i\omega_0t-ik_0x}+\epsilon^2B_2\mbox{e}^{2(i\omega_0t-ik_0x)}
+\dots+c.c.,
\label{etaansatz}
\end{equation}
\begin{equation}
\phi(x,z,t)=\epsilon^2\bar{\phi}+\epsilon
A_1\mbox{e}^{k_0z+i\omega_0t-ik_0x}+\epsilon^2
A_2\mbox{e}^{2(k_0z+i\omega_0t-ik_0x)}
+\dots+c.c.,
\label{phiansatz}
\end{equation}
\end{subequations}
where $\omega_0$ represents the frequency of the carrier wave, $k_0>0$ represents the wave number of the carrier wave, $\epsilon=2|a_0|k_0$ represents the (dimensionless) wave steepness, $a_0$ represents a typical amplitude, and $c.c.$ represents complex conjugate. Further assume
\begin{subequations}
\begin{equation}
\bar{\eta}=\bar{\eta}(X,T)=\bar{\eta}_0(X,T)+\epsilon \bar{\eta}_1(X,T)+\epsilon^2
\bar{\eta}_2(X,T)+\cdots,
\end{equation}
\begin{equation}
B=B(X,T),
\end{equation}
\begin{equation}
B_j=B_j(X,T)=B_{j0}(X,T)+\epsilon B_{j1}(X,T)+\epsilon^2
B_{j2}(X,T)+\cdots,\hspace*{0.5cm}\text{for $j=2,3,\dots$},
\end{equation}
\begin{equation}
\bar{\phi}=\bar{\phi}(X,Z,T)=\bar{\phi}_0(X,Z,T)+\epsilon \bar{\phi}_1(X,Z,T)+\epsilon^2
\bar{\phi}_2(X,Z,T)+\cdots.
\end{equation}
\begin{equation}
A_j=A_j(X,Z,T)=A_{j0}(X,Z,T)+\epsilon A_{j1}(X,Z,T)+\epsilon^2
A_{j2}(X,Z,T)+\cdots,\hspace*{0.5cm}\text{for $j=1,2,\dots$}.
\end{equation}
\end{subequations}
The slow space and time variables are defined by $X=\epsilon x$, $Z=\epsilon z$, and $T=\epsilon t$.
Assuming that $\bar{\alpha}=\epsilon^2\alpha$ and carrying out the perturbation analysis
through $\mathcal{O}(\epsilon^4)$ leads to (see \cite{FD,FD2} for details)
\begin{subequations}\label{vDysthe1}
\begin{equation}
\begin{split}
2i&\omega_0\Big{(}B_T+\frac{g}{2\omega_0}B_X\Big{)}+\epsilon\Big{(}\frac{g}{4k_0}B_{XX}+4gk_0^3|B|^2+4ik_0^2\omega_0\alpha
B\Big{)}
+\epsilon^2 \Big{(}-\frac{ig}{8k_0^2}B_{XXX}
\\ &
+2igk_0^2B^2B_X^*+12igk_0^2|B|^2B_X+2k_0\omega_0B\bar{\phi}_{0X}-8k_0\omega_0\alpha
B_X\Big{)}=0,\hspace*{0.5cm}\text{at $Z=0$},
\end{split}
\end{equation}
\begin{equation}
\bar{\phi}_{0Z}=2\omega_0\big{(}|B|^2\big{)}_X,\hspace*{1cm}\text{at }Z=0,
\end{equation}
\begin{equation}
\bar{\phi}_{0XX}+\bar{\phi}_{0ZZ}=0,\hspace*{1cm}\text{for }-\infty<Z<0,
\end{equation}
\begin{equation}
\bar{\phi}_{0X},~\bar{\phi}_{0Z}\rightarrow0,\hspace*{1cm}\text{as }Z\rightarrow-\infty,
\end{equation}
\end{subequations}
where $\omega_0^2=gk_0$ and $B^*$ represents the complex conjugate of $B$. If the Hilbert transform, $\mathcal{H}$, is defined by
\begin{equation*}
\mathcal{H}\big{(}f(x)\big{)}=\mathcal{F}^{-1}\big{(}-i\mbox{sgn}(k)\hat{f}(k)\big{)},
\end{equation*}
where the Fourier transform and its inverse are defined by
\begin{equation*}
\hat{f}(k)=\mathcal{F}\big{(}f(x)\big{)}=\int_\mathbb{R} f(x)\mbox{e}^{-ikx}dx,\hspace*{1cm}f(x)=\mathcal{F}^{-1}\big{(}\hat{f}(k)\big{)}=\frac{1}{2\pi}\int_\mathbb{R} \hat{f}(k)\mbox{e}^{ikx}dk,
\end{equation*}
then the system \eqref{vDysthe1} can be rewritten as
\begin{equation}
\begin{split}
&2i\omega_0\Big{(}B_T+\frac{g}{2\omega_0}B_X\Big{)}+\epsilon\Big{(}\frac{g}{4k_0}B_{XX}+4gk_0^3|B|^2B+4ik_0^2\omega_0\alpha
B\Big{)}\\ & +\epsilon^2 \Big{(} \!- \!\frac{ig}{8k_0^2}B_{XXX} \!+ \! 2igk_0^2B^2B_X^*+12igk_0^2|B|^2B_X \!- \!
4gk_0^2B\big{(}\mathcal{H}(|B|^2)\big{)}_X \!- \!8k_0\omega_0\alpha
B_X\Big{)}=0,
\end{split}
\label{vDysthe}
\end{equation}
with
\begin {equation*}
\bar{\phi}_0(X,Z,T)=\frac{\omega_0}{\pi}\int_\mathbb{R}i\mbox{sgn}(k)\mathcal{F}\Big{(}|B|^2\Big{)}\mbox{e}^{ikx+|k|z}dk.
\end {equation*}
We refer to this system as the viscous Dysthe (vDysthe) system.
The NLS equation is obtained from this system by setting $\alpha=0$
and disregarding terms of order $\epsilon^2$ and higher.
The Dysthe system, also known as the modified NLS system, is obtained from Equation \eqref{vDysthe}
by setting $\alpha=0$. The dissipative NLS equation (dNLS) is obtained by setting $\epsilon^2=0$.
Given a solution to Equation \eqref{vDysthe}, obtaining the corresponding surface displacement is straightforward but tedious. Determining the velocity potential is more complicated (and more tedious) because partial differential equations (PDEs) must be solved to determine each $A_{jk}$. To this order,
the surface displacement and velocity potential are given by (see Section \ref{AsymptoticDetail} for more detail)
\begin{subequations}\label{surfvelpot}
\begin{equation}
\begin{split}
\eta(x,t)
= & \, \epsilon BE+\epsilon^2k_0B^2E^2 + \epsilon^3 \Big(\frac{3}{2}k_0^2B^3E^3+iBB_XE^2+\frac{2\omega_0}{g}\big{(}\mathcal{H}(|B|^2)\big{)}_T\Big)\\
& + \epsilon^4 \Big(\frac{4ik_0^2\omega_0\alpha}{g}B^2E^2+\frac{17k_0^3}{3}|B|^2B^2E^2+\frac{8k_0^3}{3}B^4E^4+3ik_0B^2B_XE^3-\frac{1}{4k_0}|B_X|^2\\
&-\frac{1}{4k_0}BB^*_{XX}-\frac{1}{g}\bar{\phi}_{1T}\Big)+\mathcal{O}(\epsilon^5)+c.c.,
\end{split}
\label{eta}
\end{equation}
\begin{equation}
\begin{split}
\phi(x,z,t)=&\bigg\{\frac{i\epsilon\omega_0}{k_0}\check{B}+\frac{\epsilon^2\omega_0}{2k_0^2}\check{B}_X+\epsilon^3\Big(-\frac{ik_0\omega_0}{2}|\check{B}|^2\check{B}-\frac{3i\omega_0}{8k_0^3}\check{B}_{XX}\Big)\\&
+\epsilon^4\Big(-\frac{\omega_0}{4}\check{B}^2\check{B}^*_X+\frac{\omega_0}{2}|\check{B}|^2\check{B}_X-\frac{5\omega_0}{16k_0^4}\check{B}_{XXX}-2ik_0\check{B}\big{(}\mathcal{H}(|\check{B}|^2)\big{)}_T\Big)\bigg\}E\mbox{e}^{k_0z}\\& +\epsilon^4\Big{(}-2k_0^2\alpha\check{B}^2+4ik_0^2\omega_0|\check{B}|^2\check{B}^2\Big{)}E^2\mbox{e}^{2k_0z}+\mathcal{O}(\epsilon^5)+c.c.,
\label{phi}
\end{split}
\end{equation}
\end{subequations}
\noindent
where $\check{B}=B(X+iZ,T)$, $\check{B^*}=B^*(X+iZ,T)$, $E=\mbox{e}^{i\omega_0t-ik_0x}$,
and $c.c.$ represents the complex conjugate.
Unfortunately, a simple, general formula for $\bar{\phi}_1$ does not exist.
However, we include formulas for $\bar{\phi}_1$ for the particular solutions we examine below.
\subsection{Derivation of the velocity potential}
\label{AsymptoticDetail}
In order to determine the $Z$ dependence of each of the terms in the expansion for the velocity potential, PDEs must be solved at each order. At $\mathcal{O}(\epsilon)$ the PDE and surface boundary condition are
\begin{subequations}
\begin{equation}
A_{10Z} - iA_{10X}=0,
\end{equation}
\begin{equation}
A_{10}=\frac{i\omega_0}{k_0}B,\hspace*{1cm}\text{at $Z=0$.}
\end{equation}
\end{subequations}
At $\mathcal{O}(\epsilon^2)$ the PDEs and surface boundary conditions are
\begin{subequations}
\begin{equation}
A_{20Z}-iA_{20X}=0,
\end{equation}
\begin{equation}
A_{11Z}-iA_{11X}=-\frac{1}{2k_0}\Delta A_{10},
\end{equation}
\begin{equation}
A_{20}=0,\hspace*{1cm}\text{at $Z=0$,}
\end{equation}
\begin{equation}
A_{11}=\frac{\omega_0}{2k_0^2}B_X,\hspace*{1cm}\text{at $Z=0$.}
\end{equation}
\end{subequations}
At $\mathcal{O}(\epsilon^3)$ the PDEs and surface boundary conditions are
\begin{subequations}
\begin{equation}
A_{30Z}-iA_{30X}=0,
\end{equation}
\begin{equation}
A_{21Z}-iA_{21X}=-\frac{1}{4k_0}\Delta A_{20},
\end{equation}
\begin{equation}
A_{12Z}-iA_{12X}=-\frac{1}{2k_0}\Delta A_{11},
\end{equation}
\begin{equation}
A_{30}=0,\hspace*{1cm}\text{at $Z=0$,}
\end{equation}
\begin{equation}
A_{21}=0,\hspace*{1cm}\text{at $Z=0$,}
\end{equation}
\begin{equation}
A_{12}=-2k_0\alpha B+\frac{3}{2}ik_0\omega_0|B|^2B-\frac{i\omega_0}{4k_0^3}B_{XX},\hspace*{1cm}\text{at $Z=0$.}
\end{equation}
\end{subequations}
At $\mathcal{O}(\epsilon^4)$
\begin{subequations}
\begin{equation}
A_{40Z}-iA_{40X}=0
\end{equation}
\begin{equation}
A_{31Z}-iA_{31X}=-\frac{1}{6k_0}\Delta A_{30}
\end{equation}
\begin{equation}
A_{22Z}-iA_{22X}=-\frac{1}{4k_0}\Delta A_{21}
\end{equation}
\begin{equation}
A_{13Z}-iA_{13X}=-\frac{1}{2k_0}\Delta A_{12}
\end{equation}
\begin{equation}
A_{40}=0,\hspace*{1cm}\text{at $Z=0$,}
\end{equation}
\begin{equation}
A_{31}=0,\hspace*{1cm}\text{at $Z=0$,}
\end{equation}
\begin{equation}
A_{22}=-6k_0^2\alpha B^2+8ik_0^2\omega_0|B|^2B^2+\frac{i\omega_0}{4k_0^2}BB_{XX},\hspace*{1cm}\text{at $Z=0$,}
\end{equation}
\begin{equation}
A_{13}=\frac{3}{4}\omega_0B^2B^*_X-2i\alpha B_X-\frac{3\omega_0}{2}|B|^2B_X-\frac{\omega_0}{8k_0^4}B_{XXX}+\frac{i\omega_0}{g}B\bar{\phi}_{0T}+iB\bar{\phi}_{0X},\hspace*{1cm}\text{at $Z=0$.}
\end{equation}
\end{subequations}
Here $\Delta$ is the two-dimensional Laplacian operator, $\Delta F=F_{XX}+F_{ZZ}$.
The first of the PDEs listed at each order can be solved as an advection equation with a complex velocity.
The remaining PDEs are complex advection equations with nonhomogeneous terms
that can be solved using variation of parameters.
For brevity, the details have been omitted,
but the solutions to these systems are included in the equations \eqref{surfvelpot}.
\section{Particle paths for generalizations of NLS}
We now present particle paths associated with three generalizations of the NLS equation.
We start with the higher-order non-viscous model, the Dysthe equation.
Then we continue with the dissipative nonlinear Schr\"{o}dinger equation (dNLS)
and finally with the viscous Dysthe equation.
\subsection{The Dysthe system}
\label{SectionMNLS}
The Dysthe system is obtained from Equation \eqref{vDysthe} by setting $\alpha=0$.
This system has been shown to provide accurate predictions for the evolution of modulated wave trains
for a wider range of $\epsilon$ values than NLS (see, for example, Lo \& Mei~\cite{LoMei}).
Its plane-wave solutions are given by
\begin{equation}
B(X,T)=B_0\mbox{e}^{ikX-i\lambda_{_{Dys}}T},
\label{MNLSPWSoln}
\end{equation}
where
\begin{equation}
\lambda_{_{Dys}}=\omega_0\Big{(}\frac{k}{2k_0}+\frac{\epsilon
k^2}{8k_0^2}-2\epsilon B_0^2k_0^2+\frac{\epsilon^2k^3}{16k_0^3}+5\epsilon^2B_0^2kk_0\Big{)},
\end{equation}
and $B_0$, $k$ are real constants. The formulas for the corresponding surface displacement and velocity potential are included in Appendix \ref{etasMNLS}. For demonstrative purposes, we select $\epsilon=0.1$, $B_0=1$, $k_0=1$ and $k=0$. These parameters lead to a surface displacement with crest height of $0.2247$, trough height of $-0.1813$, and a $t$-period of
\begin{equation}
t_{_{Dys}}^*=\frac{2\pi}{\omega_0-\epsilon\lambda_{Dys}}=\frac{100\pi\sqrt{5}}{357}\dot{=}1.968.
\end{equation}
This particular period is the same as in the NLS example examined above because we selected $k=0$ in both cases. Figure \ref{DysthePlots} contains time-series plots of one period of the real and imaginary parts of the Dysthe plane-wave solution and one period of the corresponding surface displacement. Figure \ref{NLS4LagPlots} shows the paths followed by three Dysthe plane-wave particles. One starts on the surface and two start inside the fluid. Figure \ref{NLS4LagData} contains plots demonstrating how the period and mean horizontal velocity depend on $\zeta_0$ for this Dysthe plane-wave solution.
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{DysthePlots.eps}
\caption{\small Time series plots of (a) the real (--) and imaginary (- -) parts of the plane-wave solution to the Dysthe system, $B(0,T)$, and (b) the corresponding surface displacement $\eta(0,t)$ for one period of the carrier wave for
$\epsilon=0.1$, $k_0=1$, $B_0=1$, and $k=0$.}
\label{DysthePlots}
\end{center}
\end{figure}
For a given surface amplitude and $\zeta_0$, the periods, horizontal drifts,
and average horizontal velocities of the Dysthe particles
are all greater than the corresponding NLS quantities.
The differences between the NLS and Dysthe results are small due to the orders of the equations,
and these differences go to zero as $z\rightarrow\infty$.
Finally, for these Dysthe plane-wave solutions, the difference between the particle's vertical position
and the fluid surface, as defined in Equation \eqref{error}, decreases like $\mathcal{O}(\epsilon^5)$.
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{NLS4LagPP.eps}
\caption{\small The paths of three Dysthe plane-wave particles with $\epsilon=0.1$, $k_0=1$, $B_0=1$, and $k=0$ on $t\in[0,5T^*_{_{Dys}}]$. The initial positions are $(0,0.2247)$, $(0,-0.5)$, and $(0,-1)$.}
\label{NLS4LagPlots}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{NLS4LagData.eps}
\caption{\small Plots showing how (a) $t_{_{Dys}}$ and (b) $u_{_{Dys}}$ change for Dysthe plane-wave solutions as $\zeta_0$ changes.}
\label{NLS4LagData}
\end{center}
\end{figure}
\subsection{The dNLS equation}
\label{SectiondNLS}
The dissipative nonlinear Schr\"odinger equation is obtained from \eqref{vDysthe} by setting $\epsilon^2=0$.
The dNLS equation was first derived as a model of water waves by Dias {\emph{et al.}}~\cite{ddz08}.
However, Lo \& Mei~\cite{LoMei} and Segur {\emph{et al.}}~\cite{bf05} had previously studied it
and had shown that it accurately models the evolution of modulated wave trains.
The plane-wave solutions of dNLS are given by
\begin{equation}
B(X,T)=B_0\mbox{exp}\Big{(}ikX+\omega_r(T)+i\omega_i(T)\Big{)},
\label{dNLSPWSoln}
\end{equation}
where
\begin{subequations}\label{dNLSPWs}
\begin{equation}
\omega_r(T)=-2\epsilon\alpha k_0^2T,
\end{equation}
\begin{equation}
\omega_i(T)=\omega_0\Big{(}-\frac{kT}{2k_0}-\frac{\epsilon
k^2T}{8k_0^2}+\frac{B_0^2}{2\alpha}(1-\mbox{e}^{-4\epsilon\alpha k_0^2T})\Big{)},
\end{equation}
\end{subequations}
and $B_0$ and $k$ are real constants. These solutions were chosen so that they limit to the plane-wave solutions of NLS as $\alpha\rightarrow0$. Formulas for the corresponding surface displacement and velocity potential are included in Appendix \ref{etasMNLS}. The plane-wave solutions of dNLS are not periodic in $T$, so the corresponding surface displacements are not periodic in $t$. For demonstrative purposes, we selected $\epsilon=0.10$, $B_0=1$, $k_0=1$, $k=0$, and $\alpha=4$. This value of $\alpha$ is roughly an order of magnitude larger than the experimentally determined parameters used by Segur {\emph{et al.}}~\cite{bf05} and Carter \& Govan~\cite{FD}. We used a relatively large value for $\alpha$ so that viscous effects could be seen on relatively short time scales. Figures \ref{dNLS3LagPlots} and \ref{dNLS3xPlot} show the paths of three dNLS plane-wave particles on the interval $t\in[0,75]$. The top particle starts on the surface and the other two start and stay inside the fluid. Although the motion of the particles is not periodic, a number of comments can still be made. Due to dissipative effects, the motion, both horizontal and vertical, of all particles decreases as time increases. Each particle eventually spirals in toward a fixed point. The ``period'', the average horizontal Lagrangian velocity, and the horizontal drift all decrease as $t$ increases. This is consistent with the NLS results where smaller-amplitude solutions lead to smaller periods, velocities, and drifts.
Figure \ref{dNLS3Diff} contains a plot of $\mathcal{D}(t)$ versus $t$ for this plane-wave solution of dNLS. Note that $\mathcal{D}(t)$ does not limit to zero as $t\rightarrow\infty$. This is because the particle that starts on the surface ends up inside the fluid. This is due to two facts: (i) dNLS is an asymptotic model and (ii) the weakly viscous Euler equations are only an approximation to the true viscous system. However, dNLS remains valid in the small viscosity, $\bar{\alpha}\rightarrow0$ limit. Both $\mathcal{D}(t)$ and $\mathcal{E}(\epsilon)$ limit to zero as $\epsilon\rightarrow0$.
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{dNLS3LagPP.eps}
\caption{\small The paths of three dNLS plane-wave particles with
$\epsilon=0.10$, $B_0=1$, $k_0=1$, $k=0$, and $\alpha=4$. The initial positions are $(0,0.223)$, $(0,-0.333)$, and
$(0,-0.666)$.}
\label{dNLS3LagPlots}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{dNLS3xPlot.eps}
\caption{\small Plots of (a) $\xi(t)$ and (b) $\zeta(t)$ for the three dNLS
plane-wave particles in Figure \ref{dNLS3LagPlots}.}
\label{dNLS3xPlot}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{dNLS3Diff.eps}
\caption{\small A plot of $\mathcal{D}(t)$ versus $t$ corresponding to a dNLS
plane-wave solution with $\epsilon=0.1$, $k_0=1$, $B_0=1$, $k=0$, and $(\xi_0,\zeta_0)=(0,0.223)$ on the interval
$t\in[0,75]$.}
\label{dNLS3Diff}
\end{center}
\end{figure}
\subsection{The viscous Dysthe equation}
\label{SectionvDysthe}
The viscous Dysthe system is given in Equation \eqref{vDysthe}.
The plane-wave solutions of this system are given by
\begin{equation}
B(X,T)=B_0\mbox{exp}\Big{(}ikX+\omega_r(T)+i\omega_i(T)\Big{)},
\label{vDysthePWSoln}
\end{equation}
where
\begin{subequations}\label{vDysthePWs}
\begin{equation}
\omega_r(T)=-2\epsilon\alpha k_0(k0-2\epsilon k)T,
\end{equation}
\begin{equation}
\omega_i(T)=\omega_0\Big{(}\frac{kT}{2k_0}+\frac{\epsilon
k^2T}{8k_0^2}+\frac{\epsilon^2k^3T}{16k_0^3}-\frac{B_0^2(2k_0-5\epsilon k)}{4\alpha(k_0-2\epsilon k)}\big{(}1-\mbox{e}^{2\omega_r(T)}\big{)}\Big{)},
\end{equation}
\end{subequations}
where $B_0$ and $k$ are real constants. These solutions limit to the plane-wave solutions of the Dysthe system as $\alpha\rightarrow0$. The plane-wave solutions of vDysthe are not periodic in $T$. For demonstrative purposes, we selected $\epsilon=0.1$, $B_0=1$, $k_0=1$, $k=0$, and $\alpha=4$. This value of $\alpha$ is roughly an order of magnitude larger than the experimentally determined parameters used by Segur {\emph{et al.}}~\cite{bf05} and Carter \& Govan~\cite{FD}. We used a relatively large value for $\alpha$ so that viscous effects could be seen on relatively short time scales. Figures \ref{dNLS4LagPlots} and \ref{dNLS4xPlot} show the paths of three vDysthe plane-wave particles on the interval $t\in[0,75]$. There are many similarities with the dNLS case. As $t$ increases, the ``period'', the average horizontal Lagrangian velocity, and the horizontal drift all decrease. The horizontal drift and average horizontal velocity limit to zero as $t\rightarrow\infty$. The particles spiral in toward a fixed point.
Figure \ref{dNLS4Diff} shows that the error term $\mathcal{D}(t)$
is not smaller in the viscous Dysthe context than in the dNLS context.
This observation further demonstrates the limitation that Equation \eqref{BC1}
is only valid in the small $\alpha$ limit.
Using a higher-order approximation to Equation \eqref{DDZ} does not provide a more accurate approximation
to Equation \eqref{BC1} because Equation \eqref{BC1}
with viscosity is only an approximation to the true viscous system.
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{dNLS4LagPP.eps}
\caption{\small The paths of three vDysthe plane-wave particles with $\epsilon=0.10$, $B_0=1$, $k_0=1$, $k=0$, and
$\alpha=4$. The initial positions are $(0,0.223)$, $(0,-0.333)$, and $(0,-0.666)$ on the interval $t\in[0,75]$.}
\label{dNLS4LagPlots}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{dNLS4xPlot.eps}
\caption{\small Plots of (a) $\xi(t)$ and (b) $\zeta(t)$ for the three
viscous Dysthe plane-wave particles in Figure \ref{dNLS4LagPlots}.}
\label{dNLS4xPlot}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{dNLS4Diff.eps}
\caption{\small A plot of $\mathcal{D}(t)$ versus $t$ corresponding to a vDysthe
plane-wave solution with $\epsilon=0.1$, $k_0=1$,
$B_0=1$, $k=0$, $\alpha=4$, and $(\xi_0,\zeta_0)=(0,0.223)$ over
$t\in[0,75]$.}
\label{dNLS4Diff}
\end{center}
\end{figure}
\section{Conclusion}
In this paper, we used the classical derivation of the NLS equation to derive formulas for the velocity potential throughout the fluid. Using these formulas, we numerically computed and examined the paths of fluid particles corresponding to plane-, cnoidal, and solitary-wave solutions to the NLS equation. Following a similar procedure, we examined the paths of particles subject to the motion of plane-wave solutions to the Dysthe, dissipative NLS, and viscous Dysthe equations. We showed that dissipative/viscous effects decrease particle speed and displacement. Finally, we showed that the boundary conditions of the full water wave problem are only asymptotically satisfied by the solutions to these equations.
\section*{Acknowledgments}
This research was supported by the Research Council of Norway under grant numbers 213474/F20 and 239300/F20, by the U.S.~National Science Foundation under grant number DMS-1716120, and by a Fulbright Core Scholar Award that allowed JDC to spend a semester visiting HK at the University of Bergen. HK and JDC thank CIRM in Marseilles, France for hospitality during a two-week stay. The authors also thank John Grue for helpful discussions.
|
\section{Introduction}
\subsection{Fractional Sobolev spaces and the quest for a fractional gradient}
In the last decades, fractional Sobolev spaces have been given an increasing attention, see~\cite{DiNPV12}*{Section~1} for a detailed list of references in many directions. If $p\in[1,+\infty)$ and $\alpha\in(0,1)$, the fractional Sobolev space $W^{\alpha,p}(\mathbb{R}^n)$ is the space
\begin{equation*}
W^{\alpha,p}(\mathbb{R}^n):=\set*{u\in L^p(\mathbb{R}^n) : [u]_{W^{\alpha,p}(\mathbb{R}^n)}:=\left(\int_{\mathbb{R}^{n}} \int_{\mathbb{R}^{n}} \frac{|u(x)-u(y)|^p}{|x-y|^{n+p\alpha}}\,dx\,dy\right)^{\frac{1}{p}}<+\infty}
\end{equation*}
endowed with the natural norm
\begin{equation*}
\|u\|_{W^{\alpha,p}(\mathbb{R}^n)}:=\|u\|_{L^p(\mathbb{R}^n)}+[u]_{W^{\alpha,p}(\mathbb{R}^n)}.
\end{equation*}
Differently from the standard Sobolev space $W^{1,p}(\mathbb{R}^n)$, the space $W^{\alpha,p}(\mathbb{R}^n)$ does not have an evident distributional nature, in the sense that the seminorm $[\,\cdot\,]_{W^{\alpha,p}(\mathbb{R}^n)}$ does not seem to be the $L^p$-norm of some kind of weakly-defined gradient of fractional order.
Recently, the search for a good notion of differential operator in this fractional setting has led several authors to consider the following \emph{fractional gradient}
\begin{equation}\label{intro_eq:def_D_alpha}
\nabla^\alpha u(x):=\mu_{n,\alpha}\int_{\mathbb{R}^n}\frac{(y-x)(u(y)-u(x))}{|y-x|^{n+\alpha+1}}\,dy,
\end{equation}
where $\mu_{n,\alpha}$ is a multiplicative normalising constant controlling the behaviour of~$\nabla^{\alpha}$ as $\alpha \to 1^{-}$. For a detailed account on the existing literature on this operator, see~\cite{SS18}*{Section~1}. Here we only refer to~\cites{SSS15,SSVanS17,SSS18,SS15,SS18,S18} for the articles tightly connected to the present work. According to~\cite{SS18}*{Section~1}, it is interesting to notice that~\cite{H59} seems to be the earliest reference for the operator defined in~\eqref{intro_eq:def_D_alpha}.
From its very definition, it is not difficult to see that the fractional gradient~$\nabla^\alpha$ is well defined in $L^1(\mathbb{R}^n; \mathbb{R}^{n})$ for functions in $W^{\alpha,1}(\mathbb{R}^n)$, since we have the simple estimate
\begin{equation}\label{intro_eq:frac_Sobolev_enough}
\int_{\mathbb{R}^n}|\nabla^\alpha u|\, dx\le \mu_{n,\alpha}\,[u]_{W^{\alpha,1}(\mathbb{R}^n)}.
\end{equation}
In analogy with the standard Sobolev space $W^{1,p}(\mathbb{R}^n)$, this observation leads to consider the space
\begin{equation*}\label{intro_eq:def_distrib_frac_Sobolev_p}
S^{\alpha,p}_0(\mathbb{R}^n):=\closure[-1]{C^\infty_c(\mathbb{R}^n)}^{\|\cdot\|_{S^{\alpha,p}(\mathbb{R}^n)}},
\end{equation*}
where
\begin{equation*}\label{intro_eq:def_distrib_frac_Sobolev_norm_p}
\|u\|_{S^{\alpha,p}(\mathbb{R}^n)}:= \|u\|_{L^p(\mathbb{R}^n)}+\|\nabla^\alpha u\|_{L^p(\mathbb{R}^n;\mathbb{R}^n)}
\end{equation*}
for all $u\in C^\infty_c(\mathbb{R}^n)$. This is essentially the line followed in~\cite{SS15}, where the space $S^{\alpha,p}_0(\mathbb{R}^n)$ has been introduced (with a different, but equivalent, norm). By~\cite{SS15}*{Theorem~2.2}, it is known that
\begin{equation}\label{intro_eq:S_space_rel_1}
S^{\alpha+\varepsilon,p}_0(\mathbb{R}^n)\subset W^{\alpha, p}(\mathbb{R}^{n})\subset S^{\alpha-\varepsilon,p}_0(\mathbb{R}^n)
\end{equation}
with continuous embeddings for all $\alpha\in(0,1)$, $p\in(1,+\infty)$ and $0<\varepsilon<\min\set*{\alpha,1-\alpha}$. In the particular case $p=2$, by~\cite{SS15}*{Theorem 2.2} we actually have that
\begin{equation}\label{intro_eq:S_space_rel_2}
S^{\alpha, 2}_0(\mathbb{R}^{n})=W^{\alpha, 2}(\mathbb{R}^{n})
\end{equation}
for all $\alpha \in (0, 1)$. In addition, as observed in~\cite{S70}*{Chapter~V, Section~5.3}, we have
\begin{equation}\label{intro_eq:S_space_rel_3}
W^{\alpha,p}(\mathbb{R}^n)\subset S^{\alpha,p}_0(\mathbb{R}^n)
\end{equation}
with continuous embedding for all $\alpha\in(0,1)$ and $p\in(1,2]$. For further properties of the space $S^{\alpha,p}_0(\mathbb{R}^n)$, we refer to \cref{subsec:space_S_alpha_p} below.
The inclusions~\eqref{intro_eq:S_space_rel_1} and~\eqref{intro_eq:S_space_rel_3}, and the identification~\eqref{intro_eq:S_space_rel_2} may suggest that the spaces $S^{\alpha,p}_0(\mathbb{R}^n)$ can be considered as an interesting distributional-type alternative of the usual fractional Sobolev spaces $W^{\alpha,p}(\mathbb{R}^n)$ and thus as a natural setting for the development of a general theory for solutions to PDEs involving the fractional gradient in~\eqref{intro_eq:def_D_alpha}, proceeding similarly as in the classical Sobolev framework. This is the point of view pursued in~\cites{SS15,SS18}.
Another important aspect of the fractional gradient in~\eqref{intro_eq:def_D_alpha} is that it satisfies three natural `qualitative' requirements as a fractional operator: \emph{invariance} under translations and rotations, \emph{homogeneity} of order~$\alpha$ under dilations and some \emph{continuity} properties in an appropriate functional space, e.g.\ Schwartz's space~$\mathcal{S}(\mathbb{R}^n)$. A fundamental result of~\cite{S18} is that these three requirements actually characterise the fractional gradient in~\eqref{intro_eq:def_D_alpha} (up to multiplicative constants), see~\cite{S18}*{Theorem~2.2}. This shows that the definition in~\eqref{intro_eq:def_D_alpha} is well posed not only from a \emph{mathematical} point of view, but also from a \emph{physical} point of view.
Besides, the same characterisation holds for the following \emph{fractional divergence}
\begin{equation}\label{intro_eq:def_div_alpha}
\div^\alpha\phi(x):= \mu_{n,\alpha}\int_{\mathbb{R}^{n}} \frac{(y - x) \cdot (\phi(y) - \phi(x)) }{|y - x|^{n + \alpha + 1}} \, dy,
\end{equation}
see~\cite{S18}*{Theorem~2.4}. Moreover, as it is observed in~\cite{S18}*{Section~6}, the operators $\nabla^\alpha$ and $\div^\alpha$ are \emph{dual}, in the sense that
\begin{equation}\label{intro_eq:duality}
\int_{\mathbb{R}^{n}} u\,\div^{\alpha} \varphi \, dx = - \int_{\mathbb{R}^{n}} \varphi \cdot \nabla^{\alpha} u \, dx
\end{equation}
for all $u\in C^\infty_c(\mathbb{R}^n)$ and $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$. The fractional integration by parts formula in~\eqref{intro_eq:duality} can be thus taken as the starting point for the development of a general theory of fractional differential operators on the space of Schwartz's distributions. This is the direction of research pursued in~\cite{S18}.
\subsection{De Giorgi's distributional approach to perimeter}
In the classical framework, the Sobolev space $W^{1,1}(\mathbb{R}^n)$ is naturally continuously embedded in $BV(\mathbb{R}^n)$, the space of functions with \emph{bounded variation}, i.e.\
\begin{equation*}
BV(\mathbb{R}^n):=\set*{u\in L^1(\mathbb{R}^n) : |D u|(\mathbb{R}^n)<+\infty},
\end{equation*}
endowed with the norm
\begin{equation*}
\|u\|_{BV(\mathbb{R}^n)}=\|u\|_{L^1(\mathbb{R}^n)}+|D u|(\mathbb{R}^n),
\end{equation*}
where
\begin{equation*}
|D u|(\mathbb{R}^n)=\sup\set*{\int_{\mathbb{R}^n}u\,\div\phi\,dx : \phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n),\ \|\phi\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}\le1}
\end{equation*}
is the \emph{total variation} of the function $u\in BV(\mathbb{R}^n)$. Thanks to Riesz's Representation Theorem, one can see that a function $u\in L^1(\mathbb{R}^n)$ belongs to $BV(\mathbb{R}^n)$ if and only if there exists a finite vector valued Radon measure $Du\in\mathcal{M}(\mathbb{R}^n;\mathbb{R}^n)$ such that
\begin{equation*}
\int_{\mathbb{R}^n} u\,\div\phi\,dx=-\int_{\mathbb{R}^n}\phi\cdot dDu
\end{equation*}
for all $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$.
Functions of bounded variation have revealed to be the perfect tool for the development of a deep geometric analysis of sets with finite perimeter, starting directly from the seminal and profound works of R.~Caccioppoli and E.~De~Giorgi. For a modern exposition of this vast subject and a detailed list of references, see~\cites{AFP00,EG15,M12}.
A measurable set $E\subset\mathbb{R}^n$ has finite Caccioppoli perimeter if
\begin{equation}\label{intro_eq:def_perimeter}
P(E):=|D\chi_E|(\mathbb{R}^n)<+\infty.
\end{equation}
The perimeter functional in~\eqref{intro_eq:def_perimeter} coincides with the classical surface measure when~$E$ has a sufficiently smooth (topological) boundary and, precisely, one can prove that
\begin{equation}\label{intro_eq:perimeter_is_Hausdorff}
P(E)=\Haus{n-1}(\partial E)
\end{equation}
for all sets~$E$ with Lipschitz boundary, where $\Haus{s}$ denotes the $s$-dimensional Hausdorff measure for all $s\ge0$.
One of the finest De Giorgi's intuitions is that, for a finite perimeter set~$E$ with non-smooth boundary, the right `boundary object' to keep the validity of~\eqref{intro_eq:perimeter_is_Hausdorff} is a special subset of the topological boundary, the so-called \emph{reduced boundary}~$\redb E$. With this notion in hand, a measurable set $E\subset\mathbb{R}^n$ has finite Caccioppoli perimeter if (and only if) $\Haus{n-1}(\redb E)<+\infty$, in which case we have
\begin{equation}\label{intro_eq:perimeter_is_Hausdorff_redb}
P(E)=\Haus{n-1}(\redb E).
\end{equation}
Besides the validity of~\eqref{intro_eq:perimeter_is_Hausdorff_redb}, an essential feature of De~Giorgi's reduced boundary is the following \emph{blow-up property}: if $x\in\redb E$, then
\begin{equation}\label{intro_eq:blow-up}
\chi_{\frac{E-x}{r}}\to\chi_{H_{\nu_{E}(x)}}
\quad
\text{in } L^1_{\loc}(\mathbb{R}^n)
\end{equation}
as $r\to0$, where
\begin{equation*}
H_{\nu_{E}(x)}:=\set*{y\in\mathbb{R}^n : y \cdot\nu_E(x) \ge 0},
\qquad
\nu_E(x)=\lim_{r\to0}\frac{D\chi_E(B_r(x))}{|D\chi_E|(B_r(x))}.
\end{equation*}
The function $\nu_E\colon\redb E\to\mathbb{S}^{n-1}$ denotes the so-called \emph{measure theoretic inner unit normal} of~$E$ and coincides with the usual inner unit normal of~$E$ when the boundary of~$E$ is sufficiently smooth. In other words, the blow-up property in~\eqref{intro_eq:blow-up} shows that, in a neighbourhood of a point $x\in\redb E$, the finite perimeter set~$E$ is infinitesimally close to $x + H_{\nu_{E}(x)} = \{y \in \mathbb{R}^{n} : (y - x) \cdot \nu_{E}(x) \ge 0 \}$.
\subsection{Fractional variation and perimeter: a new distributional approach}
In the fractional framework, an analogue of the space $BV(\mathbb{R}^n)$ is completely missing, since no distributional definition of the space $W^{\alpha,1}(\mathbb{R}^n)$ is available.
Nevertheless, a theory for sets with finite fractional perimeter has been developed in recent years, with a strong interest on minimal fractional surfaces. We refer to~\cite{CF17}*{Section~7} for a detailed exposition of the most recent results in this direction.
A measurable set $E\subset\mathbb{R}^n$ has finite fractional perimeter if
\begin{equation}\label{intro_eq:def_frac_perimeter}
P_\alpha(E):= [\chi_E]_{W^{\alpha,1}(\mathbb{R}^n)}
= 2 \int_{\mathbb{R}^n \setminus E}\int_{E}\frac{1}{|x-y|^{n+\alpha}}\,dx\,dy<+\infty.
\end{equation}
The fractional perimeter functional in~\eqref{intro_eq:def_frac_perimeter} has a strong \emph{non-local nature} in the sense that its value depends also on points which are very far from the boundary of the set~$E$. For this reason, it is not clear if such a perimeter measure may be linked with some kind of fractional analogue of De~Giorgi's reduced boundary (which, a posteriori, cannot be expected to be a special subset of the topological boundary of~$E$).
In this paper, we want to combine the functional approach of~\cites{SS15,SS18} with the distributional point of view of~\cite{S18} to develop a satisfactory theory extending De~Giorgi's approach to variation and perimeter in the fractional setting.
The natural starting point is the duality relation~\eqref{intro_eq:duality}, which motivates the definition of the space
\begin{equation}\label{intro_eq:BV_alpha}
BV^\alpha(\mathbb{R}^n):=\set*{u\in L^1(\mathbb{R}^n) : |D^\alpha u|(\mathbb{R}^n)<+\infty},
\end{equation}
where
\begin{equation}\label{intro_eq:def_frac_variation}
|D^\alpha u|(\mathbb{R}^n)=\sup\set*{\int_{\mathbb{R}^n}u\,\div^\alpha\phi\,dx : \phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n),\ \|\phi\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}\le1}
\end{equation}
stands for the \emph{fractional variation} of the function $u\in BV^\alpha(\mathbb{R}^n)$. Note that the fractional variation in~\eqref{intro_eq:def_frac_variation} is well defined, since one can show that $\div^\alpha\phi\in L^\infty(\mathbb{R}^n)$ for all $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$ (see \cref{prop:frac_div_repr}).
A different approach to fractional variation was developed in~\cite{Z18}. We do not know if the fractional variation defined in~\eqref{intro_eq:def_frac_variation} is linked to the one introduced in~\cite{Z18} and it would be very interesting to establish a connection between the two.
With definition~\eqref{intro_eq:BV_alpha}, we are able to show that
\begin{equation*}
W^{\alpha,1}(\mathbb{R}^n)\subset BV^\alpha(\mathbb{R}^n)
\end{equation*}
with continuous embedding, in perfect analogy with the classical framework.
Thus, emulating the classical definition in~\eqref{intro_eq:def_perimeter}, it is very natural to define the fractional analogue of the Caccioppoli perimeter using the total variation in~\eqref{intro_eq:def_frac_variation}. Note that this definition is well posed, since $\div^\alpha\phi\in L^1(\mathbb{R}^n)$ for all $\phi\in W^{\alpha, 1}(\mathbb{R}^{n}; \mathbb{R}^{n})$ arguing similarly as in~\eqref{intro_eq:frac_Sobolev_enough}. With this notion, we are able to show that
\begin{equation} \label{eq:D_alpha_chi_E_P_alpha_intro}
|D^\alpha\chi_E|(\mathbb{R}^n)\le \mu_{n,\alpha} P_\alpha(E)
\end{equation}
for all measurable sets~$E$ with finite fractional perimeter, so that our approach naturally incorporates the current notion of fractional perimeter.
Following the classical framework, the main results concerning the space $BV^\alpha(\mathbb{R}^n)$ we are able to prove are the following:
\begin{itemize}
\item $BV^\alpha(\mathbb{R}^n)$ is a Banach space and its norm is lower semicontinuous with respect to $L^1$-convergence;
\item the inclusion $W^{\alpha,1}(\mathbb{R}^n)\subset BV^\alpha(\mathbb{R}^n)$ is continuous and strict;
\item the sets $C^\infty(\mathbb{R}^n)\cap BV^\alpha(\mathbb{R}^n)$ and $C^\infty_c(\mathbb{R}^n)$ are dense in $BV^\alpha(\mathbb{R}^n)$ with respect to the distance
\begin{equation*}
d(u,v):=\|u-v\|_{L^1(\mathbb{R}^n)} + \abs*{|D^\alpha u|(\mathbb{R}^n)-|D^\alpha v|(\mathbb{R}^n)};
\end{equation*}
\item a fractional analogue of Gagliardo--Nirenberg--Sobolev inequality holds, i.e.\ for all $n\ge2$ the embedding
\begin{equation*}
BV^\alpha(\mathbb{R}^n)\subset L^{\frac{n}{n-\alpha}}(\mathbb{R}^n)
\end{equation*}
is continuous;
\item the natural analogue of the coarea formula does not hold in $BV^{\alpha}(\mathbb{R}^{n})$, since there are functions $u \in BV^{\alpha}(\mathbb{R}^{n})$ such that $\int_{\mathbb{R}}|D^\alpha\chi_{\set{u > t}}|(\mathbb{R}^n)\,dt=+\infty$;
\item any uniformly bounded sequence in $BV^\alpha(\mathbb{R}^n)$ admits limit points in $L^1(\mathbb{R}^n)$ with respect the $L^1_{\loc}$-convergence.
\end{itemize}
Concerning sets with finite fractional Caccioppoli $\alpha$-perimeter, the main results we are able to prove are the following:
\begin{itemize}
\item fractional Caccioppoli $\alpha$-perimeter is lower semicontinuous with respect to $L^1_{\loc}$-convergence;
\item a fractional isoperimetric inequality holds, i.e.\
\begin{equation*}
|E|^{\frac{n-\alpha}{n}} \le c_{n, \alpha} |D^{\alpha} \chi_E|(\mathbb{R}^{n});
\end{equation*}
\item any sequence of sets with uniformly bounded fractional Caccioppoli $\alpha$-perimeter confined in a fixed ball admits limit points with respect to $L^1$-convergence;
\item a natural analogue of De Giorgi's reduced boundary, that we call \emph{fractional reduced boundary} $\redb^\alpha E$, is well posed for any set~$E$ with finite fractional Caccioppoli $\alpha$-perimeter;
\item if~$E$ has finite fractional Caccioppoli $\alpha$-perimeter, then its fractional Caccioppoli $\alpha$-perimeter measure satisfies $|D^\alpha\chi_E|\ll\Haus{n-\alpha}\res\redb^\alpha E$;
\item if~$E$ has finite fractional Caccioppoli $\alpha$-perimeter and $x\in\redb^\alpha E$, then the family $(\frac{E-x}{r})_{r>0}$ admits limit points in the $L^1_{\loc}$-topology and any such limit point satisfies a rigidity condition.
\end{itemize}
Some of the results listed above are proved similarly as in the classical framework. Since we believe that our approach might be interesting also to researchers that may be not familiar with the theory of functions of bounded variation, we tried to keep the exposition the most self-contained as possible.
\subsection{Future developments}
Thanks to this new approach, a large variety of classical results might be extended to the context of functions with bounded fractional variation. Here we just list some of the most intriguing open problems:
\begin{itemize}
\item investigate the case of equality in~\eqref{eq:D_alpha_chi_E_P_alpha_intro};
\item achieve a better characterisation of the blow-ups (possibly, their uniqueness);
\item prove a Structure Theorem for $\redb^{\alpha} E$ in the spirit of De Giorgi's Theorem;
\item study the fractional isoperimetric inequality and its stability, possibly also for its relative version;
\item develop a \emph{calibration theory} for sets of finite fractional Caccioppoli $\alpha$-perimeter as a useful tool for the study of fractional minimal surfaces;
\item consider the asymptotics as $\alpha \to 1^{-}$ and investigate the $\Gamma$-convergence to the classical perimeter;
\item extend the Gauss--Green and integration by parts formulas to sets of finite fractional Caccioppoli $\alpha$-perimeter;
\item give a good definition of $BV^\alpha$ functions on a general open set.
\end{itemize}
Some of these open problems will be the subject of a forthcoming paper, see~\cite{CS18}.
\subsection{Organisation of the paper}
The paper is organised as follows. In \cref{sec:Silhavy_calculus}, we introduce and study the fractional gradient and divergence, proving generalised Leibniz's rules and representation formulas. In \cref{sec:frac_BV_func}, we define the space $BV^{\alpha}(\mathbb{R}^{n})$ and we study approximation by smooth functions, embeddings and compactness exploiting a fractional version of the Fundamental Theorem of Calculus. In \cref{sec:frac_Caccioppoli_sets}, we define sets of (locally) finite fractional Caccioppoli $\alpha$-perimeter, we prove some compactness results and we introduce the notion of fractional reduced boundary. Finally, in \cref{sec:blow-ups}, we prove existence of blow-ups of sets with locally finite fractional Caccioppoli $\alpha$-perimeter.
\section{\v{S}ilhav\'{y}'s fractional calculus}
\label{sec:Silhavy_calculus}
\subsection{General notation}
We start with a brief description of the main notation used in this paper.
Given an open set $\Omega$, we say that a set $E$ is compactly contained in $\Omega$, and we write $E \Subset \Omega$, if the $\overline{E}$ is compact and contained in $\Omega$.
We denote by $\Leb{n}$ and $\Haus{\alpha}$ the Lebesgue measure and the $\alpha$-dimensional Hausdorff measure on $\mathbb{R}^n$ respectively, with $\alpha \ge 0$. Unless otherwise stated, a measurable set is a $\Leb{n}$-measurable set. We also use the notation $|E|=\Leb{n}(E)$. All functions we consider in this paper are Lebesgue measurable, unless otherwise stated. We denote by $B_r(x)$ the standard open Euclidean ball with center $x\in\mathbb{R}^n$ and radius $r>0$. We let $B_r=B_r(0)$. Recall that $\omega_{n} := |B_1|=\pi^{\frac{n}{2}}/\Gamma\left(\frac{n+2}{2}\right)$ and $\Haus{n-1}(\partial B_{1}) = n \omega_n$, where $\Gamma$ is Euler's \emph{Gamma function}, see~\cite{A64}.
For $k \in \mathbb{N}_{0} \cup \set{+ \infty}$ and $m \in \mathbb{N}$ we denote by $C^{k}_{c}(\Omega ; \mathbb{R}^{m})$ and, respectively, by $\Lip_c(\Omega; \mathbb{R}^{m})$, the space of $C^{k}$-regular, respectively, Lipschitz regular, $m$-vector valued functions defined on~$\mathbb{R}^n$ with compact support in~$\Omega$.
For any exponent $p\in[1,+\infty]$, we denote by
\begin{equation*}
L^p(\Omega;\mathbb{R}^m):=\set*{u\colon\Omega\to\mathbb{R}^m : \|u\|_{L^p(\Omega;\,\mathbb{R}^m)}<+\infty}
\end{equation*}
the space of $m$-vector valued Lebesgue $p$-integrable functions on~$\Omega$. We denote by
\begin{equation*}
W^{1,p}(\Omega;\mathbb{R}^m):=\set*{u\in L^p(\Omega;\,\mathbb{R}^m) : [u]_{W^{1,p}(\Omega;\,\mathbb{R}^m)}:=\|\nabla u\|_{L^p(\Omega;\,\mathbb{R}^{n+m})}<+\infty}
\end{equation*}
the space of $m$-vector valued Sobolev functions on~$\Omega$, see for instance~\cite{L09}*{Chapter~10} for its precise definition and main properties. We also let
\begin{equation*}
w^{1,p}(\Omega;\mathbb{R}^m):=\set*{u\in L^1_{\loc}(\Omega;\mathbb{R}^m) : [u]_{W^{1,p}(\Omega;\,\mathbb{R}^m)}<+\infty}.
\end{equation*}
We denote by
\begin{equation*}
BV(\Omega;\mathbb{R}^m):=\set*{u\in L^1(\Omega;\mathbb{R}^m) : [u]_{BV(\Omega;\mathbb{R}^m)}:=|Du|(\Omega)<+\infty}
\end{equation*}
the space of $m$-vector valued functions of bounded variation on~$\Omega$, see for instance~\cite{AFP00}*{Chapter~3} or~\cite{EG15}*{Chapter~5} for its precise definition and main properties. We also let
\begin{equation*}
bv(\Omega;\mathbb{R}^m):=\set*{u\in L^1_{\loc}(\Omega;\mathbb{R}^m) : [u]_{BV(\Omega;\,\mathbb{R}^m)}<+\infty}.
\end{equation*}
For $\alpha\in(0,1)$ and $p\in[1,+\infty)$, we denote by
\begin{equation*}
W^{\alpha,p}(\Omega;\mathbb{R}^m):=\set*{u\in L^p(\Omega;\mathbb{R}^m) : [u]_{W^{\alpha,p}(\Omega;\,\mathbb{R}^m)}\!:=\left(\int_\Omega\int_\Omega\frac{|u(x)-u(y)|^p}{|x-y|^{n+p\alpha}}\,dx\,dy\right)^{\frac{1}{p}}\!<+\infty}
\end{equation*}
the space of $m$-vector valued fractional Sobolev functions on~$\Omega$, see~\cite{DiNPV12} for its precise definition and main properties. We also let
\begin{equation*}
w^{\alpha,p}(\Omega;\mathbb{R}^m):=\set*{u\in L^1_{\loc}(\Omega;\mathbb{R}^m) : [u]_{W^{\alpha,p}(\Omega;\,\mathbb{R}^m)}<+\infty}.
\end{equation*}
For $\alpha\in(0,1)$ and $p=+\infty$, we simply let
\begin{equation*}
W^{\alpha,\infty}(\Omega;\mathbb{R}^m):=\set*{u\in L^\infty(\Omega;\mathbb{R}^m) : \sup_{x,y\in \Omega,\, x\neq y}\frac{|u(x)-u(y)|}{|x-y|^\alpha}<+\infty},
\end{equation*}
so that $W^{\alpha,\infty}(\Omega;\mathbb{R}^m)=C^{0,\alpha}_b(\Omega;\mathbb{R}^m)$, the space of $m$-vector valued bounded $\alpha$-H\"older continuous functions on~$\Omega$.
\subsection{Definition of \texorpdfstring{$\nabla^\alpha$}{nablaˆalpha} and \texorpdfstring{$\diverg^\alpha$}{divergenceˆalpha}}
We now recall and study the non-local operators~$\nabla^\alpha$ and~$\diverg^\alpha$ introduced by \v{S}ilhav\'{y} in~\cite{S18}.
Let $\alpha\in(0,1)$ and set
\begin{equation}\label{eq:def_mu_alpha}
\mu_{n, \alpha} := 2^{\alpha} \pi^{- \frac{n}{2}} \frac{\Gamma\left ( \frac{n + \alpha + 1}{2} \right )}{\Gamma\left ( \frac{1 - \alpha}{2} \right )}.
\end{equation}
We let
\begin{equation} \label{eq:def_frac_grad}
\nabla^{\alpha} f(x) := \mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_{\{ |z| > \varepsilon \}} \frac{z f(x + z)}{|z|^{n + \alpha + 1}} \, dz
\end{equation}
be the \emph{$\alpha$-gradient} of $f\in C^\infty_c(\mathbb{R}^n)$ at $x\in\mathbb{R}^n$. We also let
\begin{equation}\label{eq:def_frac_div}
\div^{\alpha} \varphi(x) := \mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_{\{ |z| > \varepsilon \}} \frac{z \cdot \varphi(x + z)}{|z|^{n + \alpha + 1}} \, dz
\end{equation}
be the \emph{$\alpha$-divergence} of $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$ at $x\in\mathbb{R}^n$. The non-local operators~$\nabla^\alpha$ and~$\diverg^\alpha$ are well defined in the sense that the involved integrals converge and the limits exist, see~\cite{S18}*{Section~7}.
Since
\begin{equation}\label{eq:cancellation_kernel}
\int_{\set*{|z| > \varepsilon}} \frac{z}{|z|^{n + \alpha + 1}} \, dz=0,
\qquad
\forall\varepsilon>0,
\end{equation}
it is immediate to check that $\nabla^{\alpha}c=0$ for all $c\in\mathbb{R}$. Moreover, the cancellation in~\eqref{eq:cancellation_kernel} yields
\begin{subequations}
\begin{align}
\nabla^{\alpha} f(x)
\label{eq:def_frac_grad_1}
&=\mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_{\{ |y -x| > \varepsilon \}} \frac{(y - x)}{|y - x|^{n + \alpha + 1}} f(y) \, dy\\
\label{eq:def_frac_grad_2}
&= \mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_{\{ |x - y| > \varepsilon \}} \frac{(y - x) (f(y) - f(x)) }{|y - x|^{n + \alpha + 1}} \, dy\\
\label{eq:def_frac_grad_3}
&=\mu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{(y - x) (f(y) - f(x)) }{|y - x|^{n + \alpha + 1}} \, dy,
\qquad
\forall x\in\mathbb{R}^n,
\end{align}
\end{subequations}
for all $f\in C^\infty_c(\mathbb{R}^n)$. Indeed, \eqref{eq:def_frac_grad_1} follows by a simple change of variables and \eqref{eq:def_frac_grad_2} is a consequence of~\eqref{eq:cancellation_kernel}. To prove~\eqref{eq:def_frac_grad_3} it is enough to apply Lebesgue's Dominated Convergence Theorem. Indeed, we can estimate
\begin{equation}\label{eq:Lip_nabla_estim_1}
\int_{\{|y-x|\le1\}} \abs*{\frac{(y - x) (f(y) - f(x)) }{|y - x|^{n + \alpha + 1}} }\, dy\le\Lip(f)\int_0^1r^{-\alpha}\, dr
\end{equation}
and
\begin{equation}\label{eq:Lip_nabla_estim_2}
\int_{\{|y-x|>1\}} \abs*{\frac{(y - x) (f(y) - f(x)) }{|y - x|^{n + \alpha + 1}} }\, dy\le2\|f\|_{L^\infty(\mathbb{R}^n)}\int_1^{+\infty}r^{-(1+\alpha)}\, dr.
\end{equation}
As a consequence, the operator $\nabla^\alpha f$ defined by~\eqref{eq:def_frac_grad_3} is well defined for all $f\in\Lip_c(\mathbb{R}^n)$ and satisfies~\eqref{eq:def_frac_grad}, \eqref{eq:def_frac_grad_1} and~\eqref{eq:def_frac_grad_2}.
By~\cite{S18}*{Theorem 4.3}, $\nabla^{\alpha}$ is invariant by translations and rotations and is $\alpha$-homoge\-neous. Moreover, for all $f\in C^\infty_c(\mathbb{R}^n)$ and $\lambda \in \mathbb{R}$, we have
\begin{equation}\label{eq:alpha_homogeneous}
(\nabla^{\alpha} f(\lambda \cdot))(x) = |\lambda|^{\alpha} \sgn{(\lambda)} (\nabla^{\alpha} f)(\lambda x),
\qquad
x\in\mathbb{R}^n.
\end{equation}
Arguing similarly as above, we can write
\begin{subequations}
\begin{align}
\div^{\alpha} \varphi(x)
\label{eq:def_frac_div_1}
&= \mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_{\{ |x - y| > \varepsilon \}} \frac{(y - x) \cdot \varphi(y) }{|y - x|^{n + \alpha + 1}} \, dy,\\
\label{eq:def_frac_div_2}
&= \mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_{\{ |x - y| > \varepsilon \}} \frac{(y - x) \cdot (\varphi(y) - \varphi(x)) }{|y - x|^{n + \alpha + 1}} \, dy,\\
\label{eq:def_frac_div_3}
&= \mu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{(y - x) \cdot (\varphi(y) - \varphi(x)) }{|y - x|^{n + \alpha + 1}} \, dy,
\qquad
\forall x\in\mathbb{R}^n,
\end{align}
\end{subequations}
for all $\phi\in\Lip_c(\mathbb{R}^n; \mathbb{R}^{n})$.
Exploiting~\eqref{eq:def_frac_grad_3} and~\eqref{eq:def_frac_div_3}, we can extend the operators $\nabla^\alpha$ and $\diverg^\alpha$ to functions with $w^{\alpha,1}$-regularity.
\begin{lemma}[Extension of $\nabla^\alpha$ and $\diverg^\alpha$ to $w^{\alpha,1}$]\label{result:Sobolev_frac_enough}
Let $\alpha\in(0,1)$. If $f\in w^{\alpha,1}(\mathbb{R}^n)$ and $\phi\in w^{\alpha,1}(\mathbb{R}^n;\mathbb{R}^n)$, then the functions $\nabla^\alpha f(x)$ and $\diverg^\alpha f(x)$ given by~\eqref{eq:def_frac_grad_3} and~\eqref{eq:def_frac_div_3} respectively are well defined for $\Leb{n}$-a.e.\ $x\in\mathbb{R}^n$. As a consequence, $\nabla^\alpha f(x)$ and $\diverg^\alpha f(x)$ satisfy~\eqref{eq:def_frac_grad}, \eqref{eq:def_frac_grad_1}, \eqref{eq:def_frac_grad_2} and~\eqref{eq:def_frac_div}, \eqref{eq:def_frac_div_1}, \eqref{eq:def_frac_div_2} respectively for $\Leb{n}$-a.e.\ $x\in\mathbb{R}^n$.
\end{lemma}
\begin{proof}
Let $f\in w^{\alpha,1}(\mathbb{R}^n)$. Then
\begin{equation*}
\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}\left|\frac{(y - x) (f(y) - f(x)) }{|y - x|^{n + \alpha + 1}} \right|\, dy\, dx
\le
[f]_{W^{\alpha,1}(\mathbb{R}^n)}
\end{equation*}
and thus the function $\nabla^\alpha f(x)$ given by~\eqref{eq:def_frac_grad_3} is well defined for $\Leb{n}$-a.e.\ $x\in\mathbb{R}^n$ and satisfies~\eqref{eq:def_frac_grad}, \eqref{eq:def_frac_grad_1} and \eqref{eq:def_frac_grad_2} by~\eqref{eq:cancellation_kernel} and by Lebesgue's Dominated Convergence Theorem. A similar argument proves the result for any $\phi\in w^{\alpha,1}(\mathbb{R}^n;\mathbb{R}^n)$.
\end{proof}
\subsection{Equivalent definition of \texorpdfstring{$\nabla^\alpha$}{nablaˆalpha} and \texorpdfstring{$\div^\alpha$}{divergenceˆalpha} via Riesz potential}
We let
\begin{equation} \label{eq:Riesz_potential_def}
I_{\alpha} f(x) :=
\frac{\Gamma \left ( \frac{n - \alpha}{2} \right )}{2^{\alpha} \pi^{\frac{n}{2}} \Gamma \left ( \frac{\alpha}{2} \right )} \int_{\mathbb{R}^{n}} \frac{f(y)}{|x - y|^{n - \alpha}} \, dy,
\qquad
x\in\mathbb{R}^n,
\end{equation}
be the \emph{Riesz potential} of order $\alpha\in(0,n)$ of a function $f\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^m)$.
Let $\alpha\in(0,1)$. Note that $I_{1-\alpha} f\in C^\infty(\mathbb{R}^n;\mathbb{R}^m)$. Recalling~\eqref{eq:def_mu_alpha}, one easily sees that
\begin{equation*}
I_{1 - \alpha} f (x)
= \frac{\mu_{n, \alpha}}{n + \alpha - 1} \int_{\mathbb{R}^{n}} \frac{f(x+ y)}{|y|^{n + \alpha - 1}} \, dy
\end{equation*}
and
\begin{equation*}
\nabla I_{1 - \alpha} f(x)
= \frac{\mu_{n, \alpha}}{n + \alpha - 1} \int_{\mathbb{R}^{n}} \frac{\nabla_{x} f(x + y)}{|y|^{n + \alpha - 1}} \, dy
= \frac{\mu_{n, \alpha}}{n + \alpha - 1} \int_{\mathbb{R}^{n}} \frac{\nabla_{y} f(x + y)}{|y|^{n + \alpha - 1}} \, dy,
\end{equation*}
so that
\begin{equation*}
\nabla I_{1-\alpha}f=I_{1-\alpha}\nabla f
\end{equation*}
for all $f\in C^\infty_c(\mathbb{R}^n)$. A similar argument proves that
\begin{equation*}
\div I_{1-\alpha}\phi=I_{1-\alpha}\div\phi
\end{equation*}
for all $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$.
Thus, accordingly to the approach developed in~\cites{H59,SS18,SSS18,SSS15,SS15,SSVanS17}, we can consider the operators
\begin{equation*}
\widetilde{\nabla^\alpha}:=\nabla I_{1-\alpha}\colon C^\infty_c(\mathbb{R}^n)\to C^\infty(\mathbb{R}^n;\mathbb{R}^n)
\end{equation*}
and
\begin{equation*}
\widetilde{\div^\alpha}:=\div I_{1-\alpha}\colon C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)\to C^\infty(\mathbb{R}^n).
\end{equation*}
We can prove that these two operators coincide with the operators defined in~\eqref{eq:def_frac_grad} and~\eqref{eq:def_frac_div}. See also~\cite{SS15}*{Theorem~1.2}.
\begin{proposition}[Equivalence]\label{rem:equiv_def_Riesz_potential}
Let $\alpha\in(0,1)$. We have $\widetilde{\nabla^\alpha}=\nabla^\alpha$ on~$\Lip_c(\mathbb{R}^n)$ and $\widetilde{\div^\alpha}=\div^\alpha$ on~$\Lip_c(\mathbb{R}^n;\mathbb{R}^n)$.
\end{proposition}
\begin{proof}
Let $f\in\Lip_c(\mathbb{R}^n)$ and fix $x\in\mathbb{R}^n$. Integrating by parts, we can compute
\begin{align*}
\widetilde{\nabla^\alpha}f(x)
&=\frac{\mu_{n, \alpha}}{n + \alpha - 1} \lim_{\varepsilon\to0}\int_{\set*{|y|>\varepsilon}} \frac{\nabla_{y} f(x + y)}{|y|^{n + \alpha - 1}} \, dy \\
& = \mu_{n, \alpha} \lim_{\varepsilon\to0}\int_{\set*{|y|>\varepsilon}} \frac{y f(y + x)}{|y |^{n + \alpha + 1}} \, dy
= \nabla^\alpha f(x),
\end{align*}
since we can estimate
\begin{equation*}
\begin{split}
\left | \int_{\set*{|y|=\varepsilon}} \frac{f(x + y)}{|y|^{n + \alpha - 1}}\frac{y}{|y|}\, d \Haus{n - 1}(y) \right |
& = \left | \int_{\set*{|y|=\varepsilon}} \frac{(f(x + y) - f(x))}{|y|^{n + \alpha - 1}}\frac{y}{|y|}\, d \Haus{n - 1}(y) \right | \\
&\le n \omega_{n} \|\nabla f\|_{L^{\infty}(\mathbb{R}^{n}; \mathbb{R}^{n})} \varepsilon^{1 - \alpha}.
\end{split}
\end{equation*}
The proof of $\widetilde{\div^\alpha}\phi=\div^\alpha\phi$ for all $\phi\in\Lip_c(\mathbb{R}^n;\mathbb{R}^n)$ follows similarly.
\end{proof}
A useful consequence of the equivalence proved in \cref{rem:equiv_def_Riesz_potential} above is the following result.
\begin{corollary}[Representation formula for $\diverg^\alpha$ and $\nabla^\alpha$] \label{prop:frac_div_repr}
Let $\alpha\in(0,1)$. If $\phi \in\Lip_c(\mathbb{R}^{n}; \mathbb{R}^{n})$ then $\div^{\alpha} \phi \in L^1(\mathbb{R}^n)\cap L^{\infty}(\mathbb{R}^{n})$ with
\begin{equation}\label{eq:frac_div_repres}
\div^{\alpha} \phi(x)
= \frac{\mu_{n, \alpha}}{n + \alpha - 1} \int_{\mathbb{R}^{n}} \frac{\div \phi(y)}{|y - x|^{n +\alpha-1}} \, dy
\end{equation}
for all $x\in\mathbb{R}^n$,
\begin{equation}\label{eq:frac_div_repr_Lip_L1_estimate}
\| \div^{\alpha} \phi \|_{L^1(\mathbb{R}^{n})} \le \mu_{n, \alpha} [\phi]_{W^{\alpha,1}(\mathbb{R}^{n};\mathbb{R}^n)}
\end{equation}
and
\begin{equation} \label{eq:frac_div_repr_Lip_estimate}
\| \div^{\alpha} \phi \|_{L^{\infty}(\mathbb{R}^{n})} \le C_{n, \alpha, U} \|\div \phi \|_{L^{\infty}(\mathbb{R}^{n})}
\end{equation}
for any bounded open set $U\subset\mathbb{R}^n$ such that $\supp(\phi) \subset U$, where
\begin{equation} \label{eq:constant_estimate}
C_{n, \alpha, U} :=
\frac{n \mu_{n, \alpha}}{(1 - \alpha)(n + \alpha - 1)}
\left(
\omega_n\diam(U)^{1 - \alpha}
+\left(
\frac{n \omega_{n}}{n+\alpha-1}
\right)^\frac{n + \alpha - 1}{n}|U|^\frac{1 - \alpha}{n}
\right).
\end{equation}
Analogously, if $f\in\Lip_c(\mathbb{R}^{n})$, then $\nabla^{\alpha}f \in L^1(\mathbb{R}^n; \mathbb{R}^{n})\cap L^{\infty}(\mathbb{R}^{n}; \mathbb{R}^{n})$ with
\begin{equation}\label{eq:frac_nabla_repres}
\nabla^{\alpha} f(x)
= \frac{\mu_{n, \alpha}}{n + \alpha - 1} \int_{\mathbb{R}^{n}} \frac{\nabla f(y)}{|y - x|^{n +\alpha-1}} \, dy
\end{equation}
for all $x\in\mathbb{R}^n$,
\begin{equation}\label{eq:frac_nabla_repr_Lip_L1_estimate}
\| \nabla^{\alpha} f \|_{L^1(\mathbb{R}^{n};\mathbb{R}^n)} \le \mu_{n, \alpha} [f]_{W^{\alpha,1}(\mathbb{R}^{n})}
\end{equation}
and
\begin{equation} \label{eq:frac_nabla_repr_Lip_estimate}
\| \nabla^{\alpha} f \|_{L^{\infty}(\mathbb{R}^{n};\mathbb{R}^n)} \le C_{n, \alpha, U} \|\nabla f\|_{L^{\infty}(\mathbb{R}^{n};\mathbb{R}^n)}
\end{equation}
for any bounded open set $U\subset\mathbb{R}^n$ such that $\supp(f) \subset U$, where $C_{n, \alpha, U}$ is as in~\eqref{eq:constant_estimate}.
\end{corollary}
\begin{proof}
The representation formula~\eqref{eq:frac_div_repres} follows directly from \cref{rem:equiv_def_Riesz_potential}. The estimate in~\eqref{eq:frac_div_repr_Lip_L1_estimate} is a consequence of \cref{result:Sobolev_frac_enough}. Finally, if $U\subset\mathbb{R}^n$ is a bounded open set such that $\supp(\phi)\subset U$, then
\begin{equation*}
\begin{split}
|\div^{\alpha} \phi(x)|
&\le \frac{\mu_{n, \alpha}}{n + \alpha - 1} \int_{\mathbb{R}^{n}} |y - x|^{1 - n - \alpha} \, |\div \phi(y)| \, dy\\
&\le \frac{\mu_{n, \alpha}\|\div \phi\|_{L^\infty(\mathbb{R}^n)}}{n + \alpha - 1} \int_U |y - x|^{1 - n - \alpha} \, dy
\end{split}
\end{equation*}
and~\eqref{eq:frac_div_repr_Lip_estimate} follows by \cref{result:riesz_kernel_estimate} below. The proof of~\eqref{eq:frac_nabla_repres}, \eqref{eq:frac_nabla_repr_Lip_L1_estimate} and~\eqref{eq:frac_nabla_repr_Lip_estimate} is similar and is left to the reader.
\end{proof}
\begin{lemma} \label{result:riesz_kernel_estimate}
Let $\alpha \in (0, 1)$ and let $U\subset\mathbb{R}^n$ be a bounded open set. For all $x \in \mathbb{R}^{n}$, we have
\begin{equation} \label{eq:riesz_kernel_estimate}
\int_{U} |y - x|^{1 - n - \alpha} \, d y
\le
\frac{n}{1 - \alpha}
\left(
\omega_n\diam(U)^{1 - \alpha}
+\left(
\frac{n \omega_{n}}{n+\alpha-1}
\right)^\frac{n + \alpha - 1}{n}|U|^\frac{1 - \alpha}{n}
\right).
\end{equation}
\end{lemma}
\begin{proof}
For $\delta > 0$, set $U^{\delta} := \{ x \in \mathbb{R}^{n} : \dist(x, U) < \delta \}$. Since clearly
\begin{equation*}
x \in U^{\delta}
\implies
B_{(\diam{(U)} + \delta)}(x) \supset U,
\end{equation*}
for all $x\in U^\delta$ we can estimate
\begin{align*}
\int_{U} |y - x|^{1 - n - \alpha} \, d y
&\le \int_{B_{(\diam (U) + \delta)}(x)} |y - x|^{1 - n - \alpha} \, d y \\
& = n \omega_{n} \int_{0}^{\diam(U) + \delta} r^{- \alpha} \, dr \\
& = \frac{n \omega_{n}}{1 - \alpha} \left (\diam{(U)} + \delta \right)^{1 - \alpha}.
\end{align*}
On the other hand, it is plain that
\begin{equation*}
x \notin U^{\delta},\ y \in U
\implies
|y - x| > \delta,
\end{equation*}
so that for all $x\notin U^\delta$ we can estimate
\begin{align*}
\int_{U} |y - x|^{1 - n - \alpha} \, d y \le \delta^{1 - n - \alpha} |U|.
\end{align*}
Thus, for all $\delta > 0$ and $x \in \mathbb{R}^{n}$, we can estimate
\begin{align*}
\int_{U} |y - x|^{1 - n - \alpha} \, d y
& \le \frac{n \omega_{n}}{1 - \alpha} \left (\diam{(U)} + \delta \right)^{1 - \alpha} + \delta^{1 - n - \alpha} |U| \\
& \le \frac{n \omega_{n}}{1 - \alpha} \left (\diam{(U)}^{1 - \alpha} + \delta^{1 - \alpha} \right) + \delta^{1 - n - \alpha} |U|
\end{align*}
since the function $s \mapsto s^{1 - \alpha}$ is subadditive for all $s>0$. Thus~\eqref{eq:riesz_kernel_estimate} follows minimising in~$\delta>0$ the right-hand side.
\end{proof}
\subsection{Duality and Leibniz's rules}
We now study the properties of the operators $\nabla^\alpha$ and $\diverg^\alpha$. We begin with the following duality relation, see~\cite{S18}*{Section~6}.
\begin{lemma}[Duality]\label{result:duality}
Let $\alpha\in(0,1)$. For all $f\in\Lip_c(\mathbb{R}^n)$ and $\phi\in\Lip_c(\mathbb{R}^n;\mathbb{R}^n)$ it holds
\begin{equation}\label{eq:duality_smooth}
\int_{\mathbb{R}^{n}} f\,\div^{\alpha} \varphi \, dx = - \int_{\mathbb{R}^{n}} \varphi \cdot \nabla^{\alpha} f \, dx.
\end{equation}
\end{lemma}
\begin{proof}
Recalling \cref{result:Sobolev_frac_enough} and exploiting~\eqref{eq:def_frac_grad_1} and~\eqref{eq:def_frac_div_1}, we can write
\begin{align*}
\int_{\mathbb{R}^{n}} f\,\div^{\alpha} \varphi \, dx
& = \mu_{n, \alpha}\int_{\mathbb{R}^{n}} f(x) \lim_{\varepsilon \to 0} \int_{ \{ |x - y| > \varepsilon \}} \frac{(y - x) \cdot \varphi(y) }{|y - x|^{n + \alpha + 1}} \, dy \, dx \\
& = \mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_{\mathbb{R}^{n}} \int_{\{ |x - y| > \varepsilon \}} f(x) \, \frac{(y - x) \cdot \varphi(y) }{|y - x|^{n + \alpha + 1}} \, dy \, dx \\
& = - \mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_{\mathbb{R}^{n}} \int_{\{ |x - y| > \varepsilon \}} \phi(y)\cdot \frac{(x - y)\,f(x) }{|x - y|^{n + \alpha + 1}} \, dx \, d y \\
& = - \int_{\mathbb{R}^{n}} \varphi(y) \cdot \nabla^{\alpha} f(y) \, dy
\end{align*}
by the Lebesgue's Dominated Convergence Theorem and Fubini's Theorem.
\end{proof}
We now prove two Leibniz-type rules for the operators $\nabla^\alpha$ and $\diverg^\alpha$, which in particular show the strong non-local nature of these two operators.
\begin{lemma}[Leibniz's rule for $\nabla^\alpha$] \label{lem:Leibniz_frac_grad}
Let $\alpha\in(0,1)$. For all $f,g\in\Lip_c(\mathbb{R}^n)$ it holds
\begin{equation*}
\nabla^{\alpha}(f g)=f \nabla^{\alpha} g + g \nabla^{\alpha} f+\nabla^{\alpha}_{\rm NL}(f,g),
\end{equation*}
where
\begin{equation*}
\nabla^{\alpha}_{\rm NL}(f, g)(x):=\mu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{(y - x) (f(y) - f(x)) ( g(y) - g(x) ) }{|y - x|^{n + \alpha + 1}} \, dy,
\quad
\forall x\in\mathbb{R}^n,
\end{equation*}
with~$\mu_{n, \alpha}$ as in~\eqref{eq:def_mu_alpha}. Moreover, it holds
\begin{equation*}
\|\nabla^{\alpha}_{\rm NL} (f, g)\|_{L^{1}(\mathbb{R}^{n};\mathbb{R}^{n})}\le \mu_{n, \alpha} [f]_{W^{\frac{\alpha}{p}, p}(\mathbb{R}^{n})} [g]_{W^{\frac{\alpha}{q}, q}(\mathbb{R}^{n})}
\end{equation*}
with $p,q\in(1,\infty)$ such that $\frac{1}{p} +\frac{1}{q}=1$ and similarly
\begin{equation*}
\| \nabla^{\alpha}_{\rm NL} (f, g)\|_{L^{1}(\mathbb{R}^{n}; \mathbb{R}^{n})}\le 2 \mu_{n, \alpha} \|f\|_{L^{\infty}(\mathbb{R}^{n})} [g]_{W^{\alpha, 1}(\mathbb{R}^{n})}.
\end{equation*}
\end{lemma}
\begin{proof}
Given $f, g \in\Lip_c(\mathbb{R}^{n})$, by \cref{result:Sobolev_frac_enough} and by \eqref{eq:def_frac_grad_3} we have
\begin{align*}
\nabla^{\alpha}(f g)(x) & = \mu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{(y - x) (f(y) g(y) - f(x) g(x)) }{|y - x|^{n + \alpha + 1}} \, dy \\
& = \mu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{(y - x) (f(y) g(y) - f(y) g(x) + f(y) g(x) - f(x) g(x) ) }{|y - x|^{n + \alpha + 1}} \, dy \\
& = \mu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{(y - x) f(y) ( g(y) - g(x) ) }{|y - x|^{n + \alpha + 1}} \, dy + g(x) \nabla^{\alpha} f(x) \\
& = \mu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{(y - x) (f(y) - f(x)) ( g(y) - g(x) ) }{|y - x|^{n + \alpha + 1}} \, dy + f(x) \nabla^{\alpha} g(x) + g(x) \nabla^{\alpha} f(x).
\end{align*}
We also have that
\begin{align*}
\| \nabla^{\alpha}_{\rm NL} (f, g) &\|_{L^{1}(\mathbb{R}^{n}; \mathbb{R}^{n})}
\le \mu_{n, \alpha} \int_{\mathbb{R}^{n}} \int_{\mathbb{R}^{n}} \frac{|f(y) - f(x)|}{|x - y|^{\frac{n + \alpha}{p}}} \frac{|g(y) - g(x)|}{|y - x|^{\frac{n + \alpha}{q}}} \, dy \, dx, \\
& \le \mu_{n, \alpha} \left ( \int_{\mathbb{R}^{n}} \int_{\mathbb{R}^{n}} \frac{|f(y) - f(x)|^{p}}{|x - y|^{n + \alpha}} \, dy \, dx \right )^{\frac{1}{p}} \left ( \int_{\mathbb{R}^{n}} \int_{\mathbb{R}^{n}} \frac{|g(y) - g(x)|^{q}}{|x - y|^{n + \alpha}} \, dy \, dx \right )^{\frac{1}{q}}
\end{align*}
for any $p, q \in (1, \infty)$ such that $\frac{1}{p} + \frac{1}{q} = 1$. The case $p = \infty$, $q = 1$ follows similarly.
\end{proof}
\begin{lemma}[Leibniz's rule for $\div^\alpha$] \label{lem:Leibniz_frac_div}
Let $\alpha\in(0,1)$. For all $f \in\Lip_c(\mathbb{R}^n)$ and $\phi\in\Lip_c(\mathbb{R}^{n};\mathbb{R}^{n})$ it holds
\begin{equation*}
\div^{\alpha}(f \phi)=f \div^{\alpha} \phi + \phi \cdot \nabla^{\alpha} f+\div^{\alpha}_{\rm NL}(f, \phi),
\end{equation*}
where
\begin{equation} \label{eq:div_NL_term}
\div^{\alpha}_{\rm NL}(f, \phi)(x):=\mu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{(y - x) \cdot ( \phi(y) - \phi(x) ) (f(y) - f(x)) }{|y - x|^{n + \alpha + 1}} \, dy,
\quad
\forall x\in\mathbb{R}^n,
\end{equation}
with~$\mu_{n, \alpha}$ as in~\eqref{eq:def_mu_alpha}. Moreover, it holds
\begin{equation*}
\|\div^{\alpha}_{\rm NL} (f, \phi)\|_{L^{1}(\mathbb{R}^{n})}\le \mu_{n, \alpha} [f]_{W^{\frac{\alpha}{p}, p}(\mathbb{R}^{n})} [\phi]_{W^{\frac{\alpha}{q}, q}(\mathbb{R}^{n}; \mathbb{R}^{n})}
\end{equation*}
with $p,q\in(1,\infty)$ such that $\frac{1}{p}+\frac{1}{q}=1$ and similarly
\begin{align*}
\| \div^{\alpha}_{\rm NL} (f, \phi)\|_{L^{1}(\mathbb{R}^{n})} & \le 2 \mu_{n, \alpha} \|f\|_{L^{\infty}(\mathbb{R}^{n})} [\phi]_{W^{\alpha, 1}(\mathbb{R}^{n}; \mathbb{R}^{n})}, \\
\| \div^{\alpha}_{\rm NL} (f, \phi)\|_{L^{1}(\mathbb{R}^{n})} & \le 2 \mu_{n, \alpha} \|\phi \|_{L^{\infty}(\mathbb{R}^{n}; \mathbb{R}^{n})} [f]_{W^{\alpha, 1}(\mathbb{R}^{n})}.
\end{align*}
\end{lemma}
\begin{proof}
Given $f \in\Lip_c(\mathbb{R}^n)$ and $\phi\in\Lip_c(\mathbb{R}^n;\mathbb{R}^n)$, by \cref{result:Sobolev_frac_enough} and by \eqref{eq:def_frac_grad_3} we have
\begin{align*}
\div^{\alpha}(f \phi)(x) & = \mu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{(y - x) \cdot (f(y) \phi(y) - f(x) \phi(x)) }{|y - x|^{n + \alpha + 1}} \, dy \\
& = \mu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{(y - x) \cdot (f(y) \phi(y) - f(y) \phi(x) + f(y) \phi(x) - f(x) \phi(x) ) }{|y - x|^{n + \alpha + 1}} \, dy \\
& = \mu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{(y - x) \cdot ( \phi(y) - \phi(x) ) f(y) }{|y - x|^{n + \alpha + 1}} \, dy + \phi(x) \cdot \nabla^{\alpha} f(x) \\
& = \mu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{(y - x) \cdot ( \phi(y) - \phi(x) ) (f(y) - f(x))}{|y - x|^{n + \alpha + 1}} \, dy + f(x) \div^{\alpha} \phi(x) + \\
& \quad + \phi(x) \cdot \nabla^{\alpha} f(x).
\end{align*}
We also have that
\begin{align*}
\| \div^{\alpha}_{\rm NL} (f, \phi) &\|_{L^{1}(\mathbb{R}^{n})}
\le \mu_{n, \alpha} \int_{\mathbb{R}^{n}} \int_{\mathbb{R}^{n}} \frac{|f(y) - f(x)|}{|x - y|^{\frac{n + \alpha}{p}}} \frac{|\phi(y) - \phi(x)|}{|y - x|^{\frac{n + \alpha}{q}}} \, dy \, dx, \\
& \le \mu_{n, \alpha} \left ( \int_{\mathbb{R}^{n}} \int_{\mathbb{R}^{n}} \frac{|f(y) - f(x)|^{p}}{|x - y|^{n + \alpha}} \, dy \, dx \right )^{\frac{1}{p}} \left ( \int_{\mathbb{R}^{n}} \int_{\mathbb{R}^{n}} \frac{|\phi(y) - \phi(x)|^{q}}{|x - y|^{n + \alpha}} \, dy \, dx \right )^{\frac{1}{q}}
\end{align*}
for any $p, q \in (1, \infty)$ such that $\frac{1}{p} + \frac{1}{q} = 1$. The case $p = \infty$, $q = 1$ follows similarly.
\end{proof}
\begin{remark}[Extension of $\nabla^\alpha_{\mathrm{NL}}$ and $\diverg^\alpha_{\mathrm{NL}}$ to fractional Sobolev spaces] \label{rem:div_nabla_NL_extension}
Thanks to the estimates in \cref{lem:Leibniz_frac_grad}, for all $\alpha \in (0, 1)$ the bilinear operator
\begin{equation*}
\nabla^{\alpha}_{\rm NL} \colon
\Lip_c(\mathbb{R}^{n}) \times \Lip_c(\mathbb{R}^{n})
\to
L^{1}(\mathbb{R}^{n}; \mathbb{R}^{n})
\end{equation*}
can be continuously extended to a bilinear operator
\begin{equation*}
\nabla^{\alpha}_{\rm NL} \colon
w^{\frac{\alpha}{p}, p}(\mathbb{R}^{n})\times w^{\frac{\alpha}{q}, q}(\mathbb{R}^{n})
\to
L^{1}(\mathbb{R}^{n}; \mathbb{R}^{n})
\end{equation*}
for any $p, q \in[1,\infty]$ such that $\frac{1}{p} + \frac{1}{q} = 1$, for which we retain the same notation (we tacitly adopt the convention $w^{\frac{\alpha}{\infty},\infty}=L^\infty$). Analogously, because of the estimates in \cref{lem:Leibniz_frac_div}, the bilinear operator
\begin{equation*}
\div^{\alpha}_{\rm NL} \colon
\Lip_c(\mathbb{R}^{n}) \times \Lip_c(\mathbb{R}^{n}; \mathbb{R}^{n})
\to
L^{1}(\mathbb{R}^{n})
\end{equation*}
can be continuously extended to a bilinear operator
\begin{equation*}
\div^{\alpha}_{\rm NL} \colon
w^{\frac{\alpha}{p}, p}(\mathbb{R}^{n}) \times w^{\frac{\alpha}{q}, q}(\mathbb{R}^{n}; \mathbb{R}^{n})
\to
L^{1}(\mathbb{R}^{n})
\end{equation*}
for any $p, q \in [1, \infty]$ such that $\frac{1}{p} + \frac{1}{q} = 1$, for which we retain the same notation.
\end{remark}
\section{Fractional \texorpdfstring{$BV$}{BV} functions}
\label{sec:frac_BV_func}
In this section we introduce and study the fractional $BV$ space naturally induced by the operators $\nabla^\alpha$ and $\diverg^\alpha$ defined in \cref{sec:Silhavy_calculus} following De Giorgi's distributional approach. In the presentation of the results, we will frequently refer to~\cite{EG15}*{Chapter~5}.
\subsection{Definition of \texorpdfstring{$BV^\alpha(\mathbb{R}^n)$}{BVˆalpha(Rˆn)} and Structure Theorem}
In analogy with the classical case (see~\cite{EG15}*{Definition~5.1} for instance), we start with the following definition.
\begin{definition}[$BV^\alpha(\mathbb{R}^n)$ space]\label{def:BV_alpha_space}
Let $\alpha\in(0,1)$. A function $f\in L^1(\mathbb{R}^n)$ belongs to the space $BV^\alpha(\mathbb{R}^n)$ if
\begin{equation*}
\sup\set*{\int_{\mathbb{R}^n} f\,\div^\alpha\phi\ dx : \phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n),\ \|\phi\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}\le1}<+\infty.
\end{equation*}
\end{definition}
We can now state the following fundamental result relating non-local distributional gradients of $BV^\alpha$ functions to vector valued Radon measures.
\begin{theorem}[Structure Theorem for $BV^\alpha$ functions]\label{th:structure_BV_alpha}
Let $\alpha\in(0,1)$ and $f \in L^{1}(\mathbb{R}^{n})$. Then, $f \in BV^\alpha(\mathbb{R}^n)$ if and only if there exists a finite vector valued Radon measure $D^{\alpha} f \in \M(\mathbb{R}^n; \mathbb{R}^{n})$ such that
\begin{equation}\label{eq:BV_alpha_duality}
\int_{\mathbb{R}^{n}} f\, \div^{\alpha} \varphi \, dx = - \int_{\mathbb{R}^n} \varphi \cdot d D^{\alpha} f
\end{equation}
for all $\varphi \in C^{\infty}_{c}(\mathbb{R}^n; \mathbb{R}^{n})$. In addition, for any open set $U\subset\mathbb{R}^n$ it holds
\begin{equation}\label{eq:BV_alpha_weak_grad_tot_var}
|D^{\alpha} f|(U) = \sup\set*{\int_{\mathbb{R}^n}f\,\div^\alpha\phi\ dx : \phi\in C^\infty_c(U;\mathbb{R}^n),\ \|\phi\|_{L^\infty(U;\mathbb{R}^n)}\le1}.
\end{equation}
\end{theorem}
\begin{proof}
If $f \in L^{1}(\mathbb{R}^{n})$ and if there exists a finite vector valued Radon measure $D^{\alpha} f \in \M(\mathbb{R}^n; \mathbb{R}^{n})$ such that \eqref{eq:BV_alpha_duality} holds, then $f \in BV^{\alpha}(\mathbb{R}^{n})$ by \cref{def:BV_alpha_space}.
If $f \in BV^{\alpha}(\mathbb{R}^{n})$, then the proof is identical to the one of~\cite{EG15}*{Theorem~5.1}, with minor modifications. Define the linear functional $L\colon C_{c}^\infty(\mathbb{R}^n; \mathbb{R}^{n})\to\mathbb{R}$ setting
\begin{equation*}
L(\phi) := - \int_{\mathbb{R}^{n}} f\, \div^{\alpha}\phi \, dx
\qquad
\forall\phi \in C^{\infty}_{c}(\mathbb{R}^n; \mathbb{R}^{n}).
\end{equation*}
Note that $L$ is well defined thanks to \cref{prop:frac_div_repr}. Since $f\in BV^\alpha(\mathbb{R}^n)$, we have
\begin{equation*}
C(U):=\sup\set*{L(\phi) : \phi\in C^\infty_c(U;\mathbb{R}^n),\ \|\phi\|_{L^\infty(U;\mathbb{R}^n)}\le 1}<+\infty
\end{equation*}
for each open set $U\subset\mathbb{R}^n$, so that
\begin{equation*}
\left | L(\phi) \right | \le C(U) \|\phi\|_{L^\infty(U;\mathbb{R}^n)}
\qquad
\forall \phi \in C^{\infty}_{c}(U; \mathbb{R}^{n}).
\end{equation*}
Thus, by the density of $C^{\infty}_{c}(\mathbb{R}^n; \mathbb{R}^{n})$ in $C_{c}(\mathbb{R}^n; \mathbb{R}^{n})$, the functional $L$ can be uniquely extended to a continuous linear functional $\tilde{L}\colon C_{c}(\mathbb{R}^n; \mathbb{R}^{n})\to\mathbb{R}$ and the conclusion follows by Riesz's Representation Theorem.
\end{proof}
\subsection{Lower semicontinuity of fractional variation}
Similarly to the classical case, the \emph{fractional variation measure} given by \cref{th:structure_BV_alpha} in~\eqref{eq:BV_alpha_weak_grad_tot_var} is lower semicontinuous with respect to $L^1$-convergence.
\begin{proposition}[Lower semicontinuity of fractional variation measure]\label{result:frac_var_meas_is_lsc}
Let $\alpha\in(0,1)$. If $(f_k)_{k\in\mathbb{N}}\subset BV^\alpha(\mathbb{R}^n)$ and $f_k\to f$ in $L^1(\mathbb{R}^n)$ as $k\to+\infty$, then $f\in BV^\alpha(\mathbb{R}^n)$ with
\begin{equation*}
|D^\alpha f|(U)\le\liminf_{k\to+\infty}|D^\alpha f_k|(U)
\end{equation*}
for any open set $U\subset\mathbb{R}^n$.
\end{proposition}
\begin{proof}
Let $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$ with $\|\phi\|_{L^\infty(\mathbb{R}^n; \mathbb{R}^{n})}\le1$. Then $\div^\alpha\phi\in L^\infty(\mathbb{R}^n)$ by \cref{prop:frac_div_repr} and so we can estimate
\begin{equation*}
\begin{split}
\int_{\mathbb{R}^n}f\,\div^\alpha\phi\ dx
=\lim_{k\to+\infty}\int_{\mathbb{R}^n}f_k\,\div^\alpha\phi\ dx
=-\lim_{k\to+\infty}\int_{\mathbb{R}^n}\phi\ dD^\alpha f_k
\le\liminf_{k\to+\infty} |D^\alpha f_k|(\mathbb{R}^n).
\end{split}
\end{equation*}
This shows that
\begin{equation*}
|D^{\alpha} f|(\mathbb{R}^{n}) \le \liminf_{k \to + \infty} |D^\alpha f_k|(\mathbb{R}^n),
\end{equation*}
thanks to \cref{th:structure_BV_alpha}. Finally, if $U$ is an open set in $\mathbb{R}^{n}$, it is enough to take $\phi \in C^\infty_c(U;\mathbb{R}^n)$ and to argue as above, applying \eqref{eq:BV_alpha_weak_grad_tot_var}.
\end{proof}
From \cref{result:frac_var_meas_is_lsc} we immediately deduce the following result, whose standard proof is left to the reader.
\begin{corollary}[$BV^\alpha$ is a Banach space]
Let $\alpha\in(0,1)$. The linear space $BV^{\alpha}(\mathbb{R}^n)$ equipped with the norm
\begin{equation*}
\|f\|_{BV^\alpha(\mathbb{R}^n)}:=\|f\|_{L^1(\mathbb{R}^n)}+|D^\alpha f|(\mathbb{R}^n),
\qquad
f\in BV^\alpha(\mathbb{R}^n),
\end{equation*}
where $D^\alpha f$ is given by \cref{th:structure_BV_alpha}, is a Banach space.
\end{corollary}
\subsection{Approximation by smooth functions}
Here and in the following, we let $\rho\in C^\infty_c(\mathbb{R}^n)$ be a function such that
\begin{equation}\label{eq:def_rho}
\supp\rho\subset B_1,
\qquad
\rho\ge0,
\qquad
\int_{\mathbb{R}^n}\rho(x)\ dx=1,
\end{equation}
see~\cite{EG15}*{Section~4.2.1} for an example. We thus let $(\rho_\varepsilon)_{\varepsilon>0}\subset C^\infty_c(\mathbb{R}^n)$ be defined as
\begin{equation}\label{eq:def_rho_eps}
\rho_\varepsilon(x):=\frac{1}{\varepsilon^n}\rho\left(\frac{x}{\varepsilon}\right)
\quad
\forall x\in\mathbb{R}^n.
\end{equation}
We call $(\rho_\varepsilon)_{\varepsilon>0}$ a family of \emph{standard mollifiers}. We have the following result.
\begin{lemma}[Convolution with standard mollifiers] \label{result:commutation_mollifier}
Let $\alpha\in(0,1)$ and let $(\rho_\varepsilon)_{\varepsilon>0}$ as in~\eqref{eq:def_rho_eps}. If $\phi \in \Lip_{c}(\mathbb{R}^{n}; \mathbb{R}^{n})$, then
\begin{equation} \label{eq:commutation_mollifier_frac_div}
\div^{\alpha} (\rho_{\varepsilon} \ast \phi) = \rho_{\varepsilon} \ast \div^{\alpha} \phi
\end{equation}
for any $\varepsilon>0$. Thus, if $f \in BV^{\alpha}(\mathbb{R}^n)$, then
\begin{equation} \label{eq:commutation_mollifier_frac_grad}
D^{\alpha} (\rho_{\varepsilon} \ast f) = (\rho_{\varepsilon} \ast D^{\alpha} f) \Leb{n}
\end{equation}
for any $\varepsilon>0$, and
\begin{equation} \label{eq:commutation_mollifier_weak_conv}
D^{\alpha} (\rho_{\varepsilon} \ast f) \rightharpoonup D^{\alpha} f
\end{equation}
in $\M(\mathbb{R}^n; \mathbb{R}^{n})$ as $\varepsilon\to0$.
\end{lemma}
\begin{proof}
Let $\phi \in \Lip_{c}(\mathbb{R}^n; \mathbb{R}^{n})$ and $x \in \mathbb{R}^{n}$. Recalling~\eqref{eq:frac_div_repres}, we can write
\begin{equation*}
\div^\alpha\phi=K_{n,\alpha}*\div\phi,
\end{equation*}
where
\begin{equation*}
K_{n,\alpha}(x)=\frac{\mu_{n,\alpha}}{n+\alpha-1}\,|x|^{1-n-\alpha},
\quad
x\in\mathbb{R}^n \setminus \{0\}.
\end{equation*}
Since $\rho_{\varepsilon}*\phi\in \Lip_{c}(\mathbb{R}^n; \mathbb{R}^{n})$, we can compute
\begin{equation*}
\begin{split}
\div^\alpha(\rho_{\varepsilon}*\phi)
&=K_{n,\alpha}*\div(\rho_{\varepsilon}*\phi)\\
&=K_{n,\alpha}*(\rho_{\varepsilon}*\div\phi)\\
&=\rho_{\varepsilon}*(K_{n,\alpha}*\div\phi)\\
&=\rho_{\varepsilon}*\div^\alpha\phi
\end{split}
\end{equation*}
and~\eqref{eq:commutation_mollifier_frac_div} follows.
Now let $f \in BV^{\alpha}(\mathbb{R}^n)$ and $\phi \in C^{\infty}_{c}(\mathbb{R}^n; \mathbb{R}^{n})$. By~\eqref{eq:BV_alpha_duality} and~\eqref{eq:commutation_mollifier_frac_div}, for all $\varepsilon>0$ we can compute
\begin{align*}
- \int_{\mathbb{R}^{n}} (\rho_{\varepsilon} \ast f)\, \div^{\alpha} \phi \, dx
&= - \int_{\mathbb{R}^{n}} f\, (\rho_{\varepsilon} \ast \div^{\alpha} \phi) \, dx \\
& = - \int_{\mathbb{R}^{n}} f \,\div^{\alpha} (\rho_{\varepsilon} \ast \phi) \, dx \\
&= \int_{\mathbb{R}^{n}} (\rho_{\varepsilon} \ast \phi) \, d D^{\alpha} f \\
& = \int_{\mathbb{R}^{n}} \phi \cdot (\rho_{\varepsilon} \ast D^{\alpha} f) \, dx,
\end{align*}
proving~\eqref{eq:commutation_mollifier_frac_grad}. The convergence in~\eqref{eq:commutation_mollifier_weak_conv} thus follows from standard properties of the mollification of Radon measures, see~\cite{AFP00}*{Theorem 2.2} for instance.
\end{proof}
As an immediate application of \cref{result:commutation_mollifier}, we can prove that a function in $BV^\alpha(\mathbb{R}^n)$ can be tested against the fractional divergence of any $\Lip_c$-regular vector field.
\begin{proposition}[$\Lip_c$-regular test] \label{result:Lip_test}
Let $\alpha \in (0, 1)$. If $f \in BV^{\alpha}(\mathbb{R}^{n})$, then~\eqref{eq:BV_alpha_duality} holds for all $\phi \in \Lip_{c}(\mathbb{R}^{n}; \mathbb{R}^{n})$.
\end{proposition}
\begin{proof}
Fix $\phi\in\Lip_c(\mathbb{R}^n;\mathbb{R}^n)$ and let $(\rho_\varepsilon)_{\varepsilon>0}\subset C^\infty_c(\mathbb{R}^n)$ be as in~\eqref{eq:def_rho_eps}. Then $\rho_\varepsilon*\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$ and so, by~\cref{result:commutation_mollifier} and~\eqref{eq:BV_alpha_duality}, we have
\begin{equation}\label{eq:duality_Lip_eps}
\int_{\mathbb{R}^{n}} (\rho_{\varepsilon} \ast f)\, \div^{\alpha} \phi \, dx
=\int_{\mathbb{R}^{n}} f\, \div^{\alpha} (\rho_{\varepsilon} \ast \phi) \, dx
= - \int_{\mathbb{R}^{n}} (\rho_{\varepsilon} \ast \phi) \cdot d D^{\alpha} f.
\end{equation}
Since $\rho_{\varepsilon} *\phi \to \phi$ uniformly and $\rho_{\varepsilon} \ast f\to f$ in $L^1(\mathbb{R}^n)$ as $\varepsilon\to0$, and $\diverg^\alpha\phi\in L^\infty(\mathbb{R}^n)$ by \cref{prop:frac_div_repr}, we can pass to the limit as $\varepsilon\to0$ in~\eqref{eq:duality_Lip_eps} getting
\begin{equation*}
\int_{\mathbb{R}^{n}} f\, \div^{\alpha} \phi \, dx
= - \int_{\mathbb{R}^{n}} \phi \cdot d D^{\alpha} f
\end{equation*}
for any $\phi\in\Lip_c(\mathbb{R}^n;\mathbb{R}^n)$.
\end{proof}
As in the classical case, we can prove the density of $C^\infty(\mathbb{R}^n)\cap BV^\alpha(\mathbb{R}^n)$ in $BV^\alpha(\mathbb{R}^n)$.
\begin{theorem}[Approximation by $C^\infty\cap BV^\alpha$ functions]\label{result:approx_by_smooth_BV}
Let $\alpha\in(0,1)$. If $f\in BV^\alpha(\mathbb{R}^n)$, then there exists $(f_k)_{k\in\mathbb{N}}\subset BV^\alpha(\mathbb{R}^n)\cap C^\infty(\mathbb{R}^n)$ such that
\begin{enumerate}[\indent(i)]
\item $f_{k} \to f$ in $L^{1}(\mathbb{R}^{n})$;
\item $|D^{\alpha} f_{k}|(\mathbb{R}^{n}) \to |D^{\alpha} f|(\mathbb{R}^{n})$.
\end{enumerate}
\end{theorem}
\begin{proof}
Let $(\rho_\varepsilon)_{\varepsilon>0}\subset C^\infty_c(\mathbb{R}^n)$ be as in~\eqref{eq:def_rho_eps}. Fix $f\in BV^\alpha(\mathbb{R}^n)$ and consider $f_\varepsilon:=f*\rho_\varepsilon$ for all $\varepsilon>0$. Since $f_\varepsilon\to f$ in $L^1(\mathbb{R}^n)$, by \cref{result:frac_var_meas_is_lsc} we get that
\begin{equation*}
|D^\alpha f|(\mathbb{R}^n)\le\liminf_{\varepsilon\to0}|D^\alpha f_\varepsilon|(\mathbb{R}^n).
\end{equation*}
By \cref{result:commutation_mollifier} we also have that
\begin{equation*}
|D^\alpha f_\varepsilon|(\mathbb{R}^n)
=\int_{\mathbb{R}^n}|\rho_\varepsilon*D^\alpha f|\,dx
\le|D^\alpha f|(\mathbb{R}^n)
\end{equation*}
and the proof is complete.
\end{proof}
Let $(\eta_R)_{R>0}\subset C^\infty_c(\mathbb{R}^n)$ be such that
\begin{equation}\label{eq:def_cut-off}
0\le\eta_R\le 1,
\qquad
\eta_R=1\text{ on $B_R$},
\qquad
\supp(\eta_R)\subset B_{R+1},
\qquad
\Lip(\eta_R)\le 2.
\end{equation}
We call $\eta_R$ a \emph{cut-off function}. As in the classical case, we can prove the density of $C^\infty_c(\mathbb{R}^n)$ in $BV^\alpha(\mathbb{R}^n)$.
\begin{theorem}[Approximation by $C^\infty_c$ functions]\label{result:approx_by_smooth_c_BV}
Let $\alpha\in(0,1)$. If $f\in BV^\alpha(\mathbb{R}^n)$, then there exists $(f_k)_{k\in\mathbb{N}}\subset C^\infty_c(\mathbb{R}^n)$ such that
\begin{enumerate}[\indent(i)]
\item $f_{k} \to f$ in $L^{1}(\mathbb{R}^{n})$;
\item $|D^{\alpha} f_{k}|(\mathbb{R}^{n}) \to |D^{\alpha} f|(\mathbb{R}^{n})$.
\end{enumerate}
\end{theorem}
\begin{proof}
Let $(\eta_R)_{R>0}\subset C^\infty_c(\mathbb{R}^n)$ be as in~\eqref{eq:def_cut-off}. Thanks to \cref{result:approx_by_smooth_BV}, it is enough to prove that $f\eta_R\to f$ in $BV^\alpha(\mathbb{R}^n)$ as $R\to+\infty$ for all $f\in C^\infty(\mathbb{R}^n)\cap BV^\alpha(\mathbb{R}^n)$. Clearly, $f\eta_R\to f$ in $L^1(\mathbb{R}^n)$ as $R\to+\infty$. Thus, by \cref{result:frac_var_meas_is_lsc}, we just need to prove that
\begin{equation}\label{eq:approx_by_smooth_c_BV_limsup}
\limsup_{R\to+\infty}|D^\alpha (f\eta_R)|(\mathbb{R}^n)\le |D^\alpha f|(\mathbb{R}^n).
\end{equation}
Fix $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$. Then, by \cref{lem:Leibniz_frac_div}, we get
\begin{equation*}
\begin{split}
\int_{\mathbb{R}^n}f\eta_R\,\div^\alpha\phi\,dx
&=\int_{\mathbb{R}^n}f\,\div^\alpha(\eta_R\phi)\,dx
-\int_{\mathbb{R}^n}f\,\phi\cdot\nabla^\alpha\eta_R\,dx
-\int_{\mathbb{R}^n}f\,\div^\alpha_{\mathrm{NL}}(\eta_R, \phi)\,dx.
\end{split}
\end{equation*}
Since $f\in BV^\alpha(\mathbb{R}^n)$ and $0\le\eta_R\le 1$, we have
\begin{equation*}
\abs*{\int_{\mathbb{R}^n}f\,\div^\alpha(\eta_R\phi)\,dx}
\le\|\phi\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}|D^\alpha f|(\mathbb{R}^n).
\end{equation*}
Moreover, we have
\begin{equation*}
\abs*{\int_{\mathbb{R}^n}f\,\phi\cdot\nabla^\alpha\eta_R\,dx}
\le
\mu_{n, \alpha} \|\phi\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}\int_{\mathbb{R}^n}|f(x)|\int_{\mathbb{R}^n}\frac{|\eta_R(y)-\eta_R(x)|}{|y-x|^{n+\alpha}}\,dy\,dx
\end{equation*}
and, similarly,
\begin{equation*}
\abs*{\int_{\mathbb{R}^n}f\,\div^\alpha_{\mathrm{NL}}(\eta_R,\phi)\,dx}
\le
2\mu_{n, \alpha}\|\phi\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}\int_{\mathbb{R}^n}|f(x)|\int_{\mathbb{R}^n}\frac{|\eta_R(y)-\eta_R(x)|}{|y-x|^{n+\alpha}}\,dy\,dx.
\end{equation*}
Combining these three estimates, we conclude that
\begin{equation*}
\begin{split}
\left | \int_{\mathbb{R}^n}f\eta_R\,\div^\alpha\phi\,dx \right |
&\le
\|\phi\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}|D^\alpha f|(\mathbb{R}^n)\\
&\quad+3\mu_{n, \alpha}\|\phi\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}\int_{\mathbb{R}^n}|f(x)|\int_{\mathbb{R}^n}\frac{|\eta_R(y)-\eta_R(x)|}{|y-x|^{n+\alpha}}\,dy\,dx
\end{split}
\end{equation*}
and~\eqref{eq:approx_by_smooth_c_BV_limsup} follows by \cref{th:structure_BV_alpha}. Indeed, we have
\begin{equation*}
\lim_{R\to+\infty}\int_{\mathbb{R}^n}|f(x)|\int_{\mathbb{R}^n}\frac{|\eta_R(y)-\eta_R(x)|}{|y-x|^{n+\alpha}}\,dy\,dx=0
\end{equation*}
combining~\eqref{eq:Lip_nabla_estim_1}, \eqref{eq:Lip_nabla_estim_2} and~\eqref{eq:def_cut-off} with Lebesgue's Dominated Convergence Theorem.
\end{proof}
\subsection{Gagliardo--Nirenberg--Sobolev inequality}
Thanks to \cref{result:approx_by_smooth_c_BV}, we are able to prove the analogous of the Gagliardo--Nirenberg--Sobolev inequality for the space $BV^{\alpha}(\mathbb{R}^{n})$.
\begin{theorem}[Gagliardo--Nirenberg--Sobolev inequality] \label{thm:GNS_immersion}
Let $\alpha \in (0, 1)$ and $n \ge 2$. There exists a constant $c_{n, \alpha} > 0$ such that
\begin{equation} \label{eq:GNS_inequality}
\|f\|_{L^{\frac{n}{n - \alpha}}(\mathbb{R}^{n})} \le c_{n, \alpha} |D^{\alpha} f|(\mathbb{R}^{n})
\end{equation}
for any $f \in BV^{\alpha}(\mathbb{R}^{n})$. As a consequence, $BV^{\alpha}(\mathbb{R}^{n})$ is continuously embedded in $L^q(\mathbb{R}^n)$ for any $q\in[1,\frac{n}{n-\alpha}]$.
\end{theorem}
\begin{proof}
By \cite{SSVanS17}*{Theorem A'}, we know that~\eqref{eq:GNS_inequality} holds for any $f \in C^{\infty}_{c}(\mathbb{R}^{n})$. So let $f \in BV^{\alpha}(\mathbb{R}^{n})$ and let $(f_k)_{k\in\mathbb{N}}\subset C^\infty_c(\mathbb{R}^n)$ be as in \cref{result:approx_by_smooth_c_BV}. By Fatou's Lemma and \cref{result:frac_var_meas_is_lsc}, we thus obtain
\begin{equation*}
\|f\|_{L^{\frac{n}{n - \alpha}}(\mathbb{R}^{n})} \le \liminf_{k \to + \infty} \|f_{k}\|_{L^{\frac{n}{n - \alpha}}(\mathbb{R}^{n})} \le c_{n, \alpha} \lim_{k \to + \infty} |D^{\alpha} f_{k}|(\mathbb{R}^{n}) = c_{n, \alpha} |D^{\alpha} f|(\mathbb{R}^{n})
\end{equation*}
and the proof is complete.
\end{proof}
\begin{remark}
We stress the fact that \cref{thm:GNS_immersion} does not hold for $n = 1$, as will be shown in \cref{rem:no_GNS_one_dim} below. It is worth to notice that an analogous restriction holds for~\cite{SSVanS17}*{Theorem A}, for which the authors provide a counterexample in the case $n = 1$ (see~\cite{SSVanS17}*{Counterexample 3.2}). The authors then derive~\cite{SSVanS17}*{Theorem A'} as a consequence of~\cite{SSVanS17}*{Theorem A}, without proving the necessity of the restriction to $n \ge 2$ in this second case, as we do in \cref{rem:no_GNS_one_dim}.
\end{remark}
\subsection{Coarea inequality}
In analogy with the classical case, we can prove a coarea inequality formula for functions in $BV^{\alpha}(\mathbb{R}^{n})$.
\begin{theorem}[Coarea inequality] \label{th:coarea_inequality}
Let $\alpha \in (0, 1)$. If $f \in BV^{\alpha}(\mathbb{R}^{n})$ is such that
\begin{equation}\label{eq:coarea_int_finite}
\int_{\mathbb{R}}|D^\alpha\chi_{\set{f>t}}|(\mathbb{R}^n)\,dt<+\infty,
\end{equation}
then
\begin{equation} \label{eq:weak_coarea_formula}
D^{\alpha} f = \int_{\mathbb{R}} D^{\alpha} \chi_{\{ f > t \}} \, dt
\end{equation}
and
\begin{equation} \label{eq:tot_var_coarea_inequality}
|D^{\alpha} f| \le \int_{\mathbb{R}} |D^{\alpha} \chi_{\{f > t \}}| \, d t.
\end{equation}
\end{theorem}
\begin{proof}
Let $\phi \in C^{\infty}_{c}(\mathbb{R}^{n}; \mathbb{R}^{n})$. By~\eqref{eq:coarea_int_finite} and applying Fubini's Theorem twice, we can compute
\begin{align*}
\int_{\mathbb{R}^{n}} \phi \cdot \, d D^{\alpha} f
& = - \int_{\mathbb{R}^{n}} f\, \div^{\alpha} \phi(x) \, dx \\
& = - \int_{\mathbb{R}^{n}} \div^{\alpha}\phi(x) \left ( \int_\mathbb{R} \chi_{(- \infty, f(x))}(t) - \chi_{(- \infty, 0)}(t) \, dt \right ) dx \\
& = - \int_{\mathbb{R}} \int_{\mathbb{R}^{n}} \div^{\alpha} \phi(x) \left ( \chi_{\{f > t \}}(x) - \chi_{(- \infty, 0)}(t) \right ) \, dx \, dt \\
& = \int_{\mathbb{R}} \int_{\mathbb{R}^{n}} \phi\cdot \, d D^{\alpha} \chi_{\{f > t \}} \, dt \\
& = \int_{\mathbb{R}^{n}}\phi \cdot d\left(\int_{\mathbb{R}} D^{\alpha} \chi_{\{ f > t \}} \, dt\right)
\end{align*}
proving~\eqref{eq:weak_coarea_formula}. Thus
\begin{equation*}
|D^{\alpha} f |
=\abs*{\int_{\mathbb{R}} D^{\alpha} \chi_{\{f > t \}} \, d t}
\le\int_{\mathbb{R}} |D^{\alpha} \chi_{\{f > t \}}| \, d t
\end{equation*}
and the proof is complete.
\end{proof}
\subsection{A fractional version of the Fundamental Theorem of Calculus}
Let $\alpha\in(0,1)$ and let $\mu_{n, -\alpha}$ be given by~\eqref{eq:def_mu_alpha} (note that the expression in~\eqref{eq:def_mu_alpha} makes sense for all $\alpha\in(-1,1)$). We let
\begin{equation}\label{eq:def_T_space}
\mathcal{T}(\mathbb{R}^n):=\set*{f\in C^\infty(\mathbb{R}^n): D^\mathsf{a} f\in L^1(\mathbb{R}^n)\cap C_0(\mathbb{R}^n)\text{ for all multi-indices } \mathsf{a}\in\mathbb{N}^n_0}
\end{equation}
and
\begin{equation*}
\mathcal{T}(\mathbb{R}^n;\mathbb{R}^n):=\set*{\phi\in C^\infty(\mathbb{R}^n;\mathbb{R}^n) : \phi_i\in\mathcal{T}(\mathbb{R}^n),\ i=1,\dots, n}.
\end{equation*}
By~\cite{S18}*{Section~5}, the operator
\begin{equation} \label{eq:div_negative_alpha_def}
\div^{- \alpha} \phi(x) := \mu_{n, - \alpha} \int_{\mathbb{R}^{n}} \frac{z\cdot \phi(x + z)}{|z|^{n + 1 - \alpha}} \, dz
\end{equation}
is well defined for any $\phi \in \mathcal{T}(\mathbb{R}^{n}; \mathbb{R}^{n})$. Moreover, by \cite{S18}*{Theorem 5.3}, we have the following \emph{inversion formula}
\begin{equation} \label{eq:div_-alpha_nabla_alpha_id}
- \div^{- \alpha} \nabla^{\alpha} =\mathrm{id}_{\mathcal{T}(\mathbb{R}^n)}.
\end{equation}
Exploiting~\eqref{eq:div_negative_alpha_def} and~\eqref{eq:div_-alpha_nabla_alpha_id} we can prove the following fractional version of the Fundamental Theorem of Calculus. See~\cite{SS15}*{Theorem~2.1} for a similar approach.
\begin{theorem}[Fractional Fundamental Theorem of Calculus] \label{thm:fund_theorem_calculus_frac}
Let $\alpha \in (0, 1)$. If $f \in C^{\infty}_{c}(\mathbb{R}^{n})$, then
\begin{equation} \label{eq:fund_theorem_calculus_frac}
f(y) - f(x) = \mu_{n, - \alpha} \int_{\mathbb{R}^{n}} \left ( \frac{z-x}{|z-x|^{n + 1 - \alpha}} - \frac{z - y}{|z - y|^{n + 1 - \alpha}} \right ) \cdot \nabla^{\alpha} f(z) \, dz
\end{equation}
for any $x, y \in \mathbb{R}^{n}$.
\end{theorem}
\begin{proof}
Since clearly $C^\infty_c(\mathbb{R}^n)\subset\mathcal{T}(\mathbb{R}^n)$, we have $\nabla^\alpha f\in\mathcal{T}(\mathbb{R}^n;\mathbb{R}^n)$ by~\cite{S18}*{Theorem~4.3}. Applying~\eqref{eq:div_-alpha_nabla_alpha_id}, we have
\begin{align*}
f(y) - f(x) & = (- \div^{- \alpha} \nabla^{\alpha} f)( y) - (- \div^{- \alpha} \nabla^{\alpha} f)(x) \\
& = \mu_{n, - \alpha} \int_{\mathbb{R}^{n}} \frac{z}{|z|^{n + 1 - \alpha}} \cdot \big( \nabla^{\alpha} f(x + z) - \nabla^{\alpha} f(y + z) \big ) \, dz
\end{align*}
for all $x, y \in \mathbb{R}^{n}$. Then~\eqref{eq:fund_theorem_calculus_frac} follows splitting the integral and changing variables.
\end{proof}
An easy consequence of \cref{thm:fund_theorem_calculus_frac} is that the distributional $\alpha$-divergence of the kernel appearing in~\eqref{eq:fund_theorem_calculus_frac} is a difference of Dirac deltas.
\begin{proposition} \label{prop:div_alpha_delta}
Let $\alpha \in (0, 1)$. If $x, y \in \mathbb{R}^{n}$, then
\begin{equation} \label{eq:div_alpha_delta}
\mu_{n, - \alpha} \div^{\alpha} \left ( \frac{\cdot - y}{|\cdot - y|^{n + 1 - \alpha}} - \frac{\cdot - x}{|\cdot - x|^{n + 1 - \alpha}} \right ) = \delta_{y} - \delta_{x}
\end{equation}
in the sense of Radon measures.
\end{proposition}
\begin{proof}
It follows immediately from~\eqref{eq:fund_theorem_calculus_frac}.
\end{proof}
\subsection{Compactness}
We start with the following H\"older estimate on the $L^1$-norm of translations of functions in $C^\infty_c(\mathbb{R}^n)$.
\begin{proposition}[$L^1$-estimate on translations] \label{prop:Holder_estimate}
Let $\alpha \in (0,1)$. If $f \in C^{\infty}_{c}(\mathbb{R}^{n})$, then
\begin{equation} \label{eq:Holder_estimate}
\int_{\mathbb{R}^n}|f(x + y) - f(x)|\,dx \le \gamma_{n, \alpha}\, |y|^{\alpha}\, \|\nabla^{\alpha} f\|_{L^{1}(\mathbb{R}^{n}; \mathbb{R}^{n})}
\end{equation}
for all $y \in \mathbb{R}^{n}$, where
\begin{equation}\label{eq:def_gamma_n,alpha}
\gamma_{n, \alpha} := \mu_{n, - \alpha} \int_{\mathbb{R}^{n}} \left | \frac{z}{|z|^{n + 1 - \alpha}} - \frac{z - \mathrm{e}_{1}}{|z - \mathrm{e}_{1}|^{n + 1- \alpha}} \right | \, dz.
\end{equation}
\end{proposition}
\begin{proof}
By \eqref{eq:fund_theorem_calculus_frac}, we have
\begin{align*}
\int_{\mathbb{R}^n}|f(x + y) - f(x)|\,dx
& \le \mu_{n, - \alpha} \int_{\mathbb{R}^{n}} \int_{\mathbb{R}^{n}} \left | \frac{z}{|z|^{n + 1 - \alpha}} - \frac{z - y}{|z - y|^{n + 1- \alpha}} \right | |\nabla^{\alpha} f(x + z)| \, dz \, dx \\
& = \mu_{n, - \alpha} \|\nabla^{\alpha} f\|_{L^{1}(\mathbb{R}^{n}; \mathbb{R}^{n})} \int_{\mathbb{R}^{n}} \left | \frac{z}{|z|^{n + 1 - \alpha}} - \frac{z - y}{|z - y|^{n + 1- \alpha}} \right | \, dz.
\end{align*}
Now we notice that the integral appearing in the last term is actually a radial function of~$y$. Indeed, let $\mathrm{R} \in {\rm SO}(n)$ be such that $\mathrm{R} y = |y| \nu$, for some $\nu \in \mathbb{S}^{n - 1}$. Making the change of variable $z = |y|\transp{\mathrm{R}}w$, we obtain
\begin{align*}
\int_{\mathbb{R}^{n}} \left | \frac{z}{|z|^{n + 1 - \alpha}} - \frac{z - y}{|z - y|^{n + 1- \alpha}} \right | \, dz
& = |y|^{\alpha} \int_{\mathbb{R}^{n}} \left | \frac{\transp{\mathrm{R}} w}{|w|^{n + 1 - \alpha}} - \frac{\transp{\mathrm{R}}(w - \nu)}{|w - \nu|^{n + 1- \alpha}} \right | \, dw \\
& = |y|^{\alpha} \int_{\mathbb{R}^{n}} \left | \frac{w}{|w|^{n + 1 - \alpha}} - \frac{(w - \nu)}{|w - \nu|^{n + 1- \alpha}} \right | \, dw.
\end{align*}
Since $\nu$ is arbitrary, we may choose $\nu = \mathrm{e}_{1}$. We now prove that
\begin{equation*}
\int_{\mathbb{R}^{n}} \left | \frac{z}{|z|^{n + 1 - \alpha}} - \frac{z - \mathrm{e}_{1}}{|z - \mathrm{e}_{1}|^{n + 1- \alpha}} \right | \, dz < + \infty.
\end{equation*}
To this purpose, we notice that
\begin{align*}
\int_{B_{2}} \left | \frac{z}{|z|^{n + 1 - \alpha}} - \frac{z - \mathrm{e}_{1}}{|z - \mathrm{e}_{1}|^{n + 1- \alpha}} \right | \, dz & \le \int_{B_{2}} \frac{1}{|z|^{n - \alpha}} \, dz + \int_{B_{2}} \frac{1}{|z - \mathrm{e}_{1}|^{n - \alpha}} \, dz \\
& \le 2 \int_{B_{3}} \frac{1}{|z|^{n - \alpha}}\, dz = 2 n \omega_{n} \frac{3^{\alpha}}{\alpha}.
\end{align*}
On the other hand, for all $z\in\mathbb{R}^n\setminus B_2$ we have
\begin{align*}
\frac{z - \mathrm{e}_{1}}{|z - \mathrm{e}_{1}|^{n + 1- \alpha}} - \frac{z}{|z|^{n + 1 - \alpha}} & = \int_{0}^{1} \frac{d}{dt} \left ( \frac{(z - t \mathrm{e}_{1})}{|z - t \mathrm{e}_{1}|^{n + 1 - \alpha}} \right) \, dt \\
& = \int_{0}^{1} - \frac{\mathrm{e}_{1}}{|z - t \mathrm{e}_{1}|^{n + 1 - \alpha}} + (n + 1 - \alpha) (z_{1} - t) \frac{(z - t \mathrm{e}_{1})}{|z - t \mathrm{e}_{1}|^{n + 3 - \alpha}} \, dt
\end{align*}
so that
\begin{align*}
\int_{\mathbb{R}^{n} \setminus B_{2}} \left | \frac{z}{|z|^{n + 1 - \alpha}} - \frac{z - \mathrm{e}_{1}}{|z - \mathrm{e}_{1}|^{n + 1- \alpha}} \right | \, dz & \le \int_{\mathbb{R}^{n} \setminus B_{2}} \int_{0}^{1} \frac{|z - t \mathrm{e}_{1}|+ (n - \alpha +1)|z_{1} - t|}{|z - t \mathrm{e}_{1}|^{n + 2 - \alpha}} \, dt \, dz \\
& \le (n - \alpha + 2) \int_{0}^{1} \int_{\mathbb{R}^{n} \setminus B_{2}} \frac{1}{|z - t \mathrm{e}_{1}|^{n + 1 - \alpha}} \, dz \, dt \\
& \le (n - \alpha + 2) \int_{0}^{1} \int_{\mathbb{R}^{n} \setminus B_{1}} \frac{1}{|z|^{n + 1 - \alpha}} \, dz \, dt \\
& = (n - \alpha + 2) \frac{n \omega_{n}}{1 - \alpha}.
\end{align*}
We conclude that
\begin{equation*}
\int_{\mathbb{R}^{n}} \left | \frac{z}{|z|^{n + 1 - \alpha}} - \frac{z - \mathrm{e}_{1}}{|z - \mathrm{e}_{1}|^{n + 1- \alpha}} \right | \, dz \le n \omega_{n} \left ( 2 \frac{3^{\alpha}}{\alpha} + \frac{(n - \alpha + 2)}{1 - \alpha} \right ) < + \infty.
\end{equation*}
Thus, the proof is complete.
\end{proof}
Similarly to the classical case, as a consequence of the previous result we can prove the following key estimate of the $L^1$-distance of a function in $BV^\alpha(\mathbb{R}^n)$ and its convolution with a mollifier.
\begin{corollary}[$L^1$-distance with convolution] \label{result:mollifier_conv_estimate_Holder}
Let $\alpha \in (0, 1)$. If $f \in BV^{\alpha}(\mathbb{R}^{n})$, then
\begin{equation} \label{eq:mollifier_conv_estimate_Holder}
\|\rho_{\varepsilon} \ast f - f\|_{L^{1}(\mathbb{R}^{n})}
\le \gamma_{n, \alpha}\, \varepsilon^{\alpha} |D^{\alpha} f|(\mathbb{R}^{n})
\end{equation}
for all $\varepsilon > 0$, where $(\rho_\varepsilon)_{\varepsilon>0}\subset C^\infty_c(\mathbb{R}^n)$ is as in~\eqref{eq:def_rho_eps} and $\gamma_{n, \alpha}$ as in \cref{prop:Holder_estimate}.
\end{corollary}
\begin{proof}
By \cref{result:approx_by_smooth_c_BV}, it is enough to prove~\eqref{eq:mollifier_conv_estimate_Holder} for $f \in C^{\infty}_{c}(\mathbb{R}^{n})$. By~\eqref{eq:Holder_estimate}, we get
\begin{align*}
\|\rho_{\varepsilon} \ast f - f\|_{L^{1}(\mathbb{R}^{n})}
& \le \int_{\mathbb{R}^{n}} \int_{\mathbb{R}^{n}} \rho(y) |f(x - \varepsilon y) - f(x)| \, dy \, dx \\
& = \int_{\mathbb{R}^{n}} \rho(y)\int_{\mathbb{R}^{n}} |f(x - \varepsilon y) - f(x)| \, dx \, dy \\
& \le \gamma_{n, \alpha}\, \varepsilon^{\alpha} \|\nabla^{\alpha}f\|_{L^{1}(\mathbb{R}^{n}; \mathbb{R}^{n})} \int_{B_1} \rho(y) |y|^{\alpha} \, dy\\
& \le \gamma_{n, \alpha}\, \varepsilon^{\alpha} \|\nabla^{\alpha}f\|_{L^{1}(\mathbb{R}^{n}; \mathbb{R}^{n})}
\end{align*}
and the proof is complete.
\end{proof}
We are now ready to prove following compactness result for the space $BV^\alpha(\mathbb{R}^n)$.
\begin{theorem}[Compactness for $BV^\alpha(\mathbb{R}^n)$]\label{result:compactness_BV_alpha}
Let $\alpha \in (0, 1)$. If $(f_k)_{k\in\mathbb{N}}\subset BV^\alpha(\mathbb{R}^n)$ satisfies
\begin{equation*}
\sup_{k\in\mathbb{N}}\|f_k\|_{BV^\alpha(\mathbb{R}^n)}<+\infty,
\end{equation*}
then there exists a subsequence $(f_{k_j})_{j\in\mathbb{N}}\subset BV^\alpha(\mathbb{R}^n)$ and a function $f\in L^1(\mathbb{R}^n)$ such that
\begin{equation*}
f_{k_j}\to f
\text{ in } L^1_{\loc}(\mathbb{R}^n)
\end{equation*}
as $j\to+\infty$.
\end{theorem}
\begin{proof}
We follow the line of the proof of~\cite{AFP00}*{Theorem~3.23}. Let $(\rho_\varepsilon)_{\varepsilon>0}\subset C^\infty_c(\mathbb{R}^n)$ be as in~\eqref{eq:def_rho_eps} and set $f_{k,\varepsilon}:=\rho_\varepsilon*f_k$. Clearly $f_{k,\varepsilon}\in C^\infty(\mathbb{R}^n)$ and
\begin{equation*}
\|f_{k,\varepsilon}\|_{L^\infty(U)}\le\|\rho_\varepsilon\|_{L^\infty(\mathbb{R}^n)} \|f_k\|_{L^1(\mathbb{R}^n)},
\qquad
\|\nabla f_{k,\varepsilon}\|_{L^\infty(U; \mathbb{R}^{n})}\le\|\nabla\rho_{\varepsilon}\|_{L^\infty(\mathbb{R}^n; \mathbb{R}^{n})}\|f_k\|_{L^1(\mathbb{R}^n)}
\end{equation*}
for any open set $U\Subset\mathbb{R}^n$. Thus $(f_{k,\varepsilon})_{k\in\mathbb{N}}$ is locally equibounded and locally equicontinuous for each $\varepsilon>0$ fixed. By a diagonal argument, we can find a sequence $(k_j)_{j\in\mathbb{N}}$ such that $(f_{k_j,\varepsilon})_{j\in\mathbb{N}}$ converges in $C(U)$ for any open set $U\Subset\mathbb{R}^n$ with $\varepsilon=1/p$ for all $p\in\mathbb{N}$. By \cref{result:mollifier_conv_estimate_Holder}, we thus get
\begin{equation*}
\begin{split}
\limsup_{h,j\to+\infty}\int_U|f_{k_h}-f_{k_j}|\,dx
&=\limsup_{h,j\to+\infty}\int_U|f_{{k_h},1/p}-f_{{k_j},1/p}|\,dx\\
&\quad+\limsup_{h,j\to+\infty}\int_U|f_{k_h}-f_{{k_h},1/p}|+|f_{k_j}-f_{{k_j},1/p}|\,dx\\
&\le\frac{2\gamma_{n,\alpha}}{p^\alpha}\,\sup_{k\in\mathbb{N}}|D^\alpha f_k|(\mathbb{R}^n)
\end{split}
\end{equation*}
for all open set $U\Subset\mathbb{R}^n$. Since $p\in\mathbb{N}$ is arbitrary and $L^1(U)$ is a Banach space, this shows that $(f_{k_j})_{j\in\mathbb{N}}$ converges in $L^1(U)$ for all open set $U\Subset\mathbb{R}^n$. Up to extract a further subsequence (which we do not relabel for simplicity), we also have that $f_{k_j}(x)\to f(x)$ for $\Leb{n}$-a.e.\ $x\in\mathbb{R}^n$. By Fatou's Lemma, we can thus infer that
\begin{equation*}
\|f\|_{L^1(\mathbb{R}^n)}
\le\liminf_{j\to+\infty}\|f_{k_j}\|_{L^1(\mathbb{R}^n)}
\le\sup_{k\in\mathbb{N}}\|f_k\|_{BV^\alpha(\mathbb{R}^n)}.
\end{equation*}
Hence $f\in L^1(\mathbb{R}^n)$ and the proof is complete.
\end{proof}
\begin{remark}[Improvement of~\cite{SS15}*{Theorem~2.1}]
The argument presented above can be used to extend the validity of~\cite{SS15}*{Theorem~2.1} to all exponents $p\in[1,\frac{n}{\alpha})$, since our strategy does not rely on the boundedness of Riesz's transform but only on the inversion formula~\eqref{eq:div_-alpha_nabla_alpha_id}. We leave the details of the proof of this improvement of~\cite{SS15}*{Theorem~2.1} to the interested reader.
\end{remark}
\subsection{The inclusion \texorpdfstring{$W^{\alpha,1}(\mathbb{R}^n)\subset BV^\alpha(\mathbb{R}^n)$}{Wˆ{alpha,1}(Rˆn) in BVˆalpha(Rˆn)}}
As in the classical case, fractional $BV$ functions naturally include fractional Sobolev functions.
\begin{theorem}[$W^{\alpha,1}(\mathbb{R}^n)\subset BV^\alpha(\mathbb{R}^n)$]\label{result:Sobolev_subset_BV}
Let $\alpha\in(0,1)$. If $f\in W^{\alpha,1}(\mathbb{R}^n)$ then $f\in BV^\alpha(\mathbb{R}^n)$, with
\begin{equation}\label{eq:BV_subset_Sobolev}
|D^{\alpha} f|(\mathbb{R}^n) \le \mu_{n, \alpha} [ f ]_{W^{\alpha, 1}(\mathbb{R}^n)}
\end{equation}
and
\begin{equation}\label{eq:duality_Sobolev_frac}
\int_{\mathbb{R}^n}f\,\div^\alpha\phi\, dx=-\int_{\mathbb{R}^n}\phi\cdot\nabla^\alpha f\, dx
\end{equation}
for all $\phi\in\Lip_c(\mathbb{R}^n;\mathbb{R}^n)$, so that $D^\alpha f=\nabla^\alpha f\,\Leb{n}$.
Moreover, if $f \in BV(\mathbb{R}^{n})$, then $f \in W^{\alpha, 1}(\mathbb{R}^{n})$ for any $\alpha \in (0, 1)$, with
\begin{equation} \label{eq:W_alpha_norm_bound_BV}
\|f\|_{W^{\alpha,1}(\mathbb{R}^n)} \le c_{n,\alpha}\|f\|_{BV(\mathbb{R}^n)}
\end{equation}
for some $c_{n, \alpha} > 0$ and
\begin{equation} \label{eq:weak_frac_grad_repr}
\nabla^{\alpha} f(x)
= \frac{\mu_{n, \alpha}}{n + \alpha - 1} \int_{\mathbb{R}^n} \frac{d D f(y)}{|y - x|^{n + \alpha - 1}}
\end{equation}
for $\Leb{n}$-a.e.\ $x\in\mathbb{R}^n$.
\end{theorem}
\begin{proof}
Let $f\in W^{\alpha,1}(\mathbb{R}^n)$. For any $\phi\in\Lip_{c}(\mathbb{R}^n;\mathbb{R}^{n})$, by Lebesgue's Dominated Convergence Theorem, Fubini's Theorem and \cref{result:Sobolev_frac_enough}, and recalling~\eqref{eq:cancellation_kernel}, we can compute
\begin{align*}
\int_{\mathbb{R}^{n}} f \div^{\alpha} \varphi \, dx
& = \mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_{\mathbb{R}^{n}} \int_{\{|x -y| > \varepsilon \}} f(x) \frac{(y - x) \cdot \varphi(y) }{|y - x|^{n + \alpha + 1}} \, dy \, dx \\
& = \mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_{\mathbb{R}^{n}} \int_{\{|x -y| > \varepsilon \}} \varphi(y) \cdot \frac{(y - x) f(x) }{|y - x|^{n + \alpha + 1}} \, dx \, dy\\
& = \mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_{\mathbb{R}^{n}} \int_{\{|x -y| > \varepsilon \}} \varphi(y) \cdot \frac{(y - x) (f(x) - f(y)) }{|y - x|^{n + \alpha + 1}} \, dx \, dy\\
& = - \int_{\mathbb{R}^{n}} \varphi(y) \cdot\nabla^\alpha f(y) \, dy.
\end{align*}
This proves~\eqref{eq:duality_Sobolev_frac}, so that $f\in BV^\alpha(\mathbb{R}^n)$. Inequality~\eqref{eq:BV_subset_Sobolev} follows as in \cref{result:Sobolev_frac_enough}.
Now let $f \in BV(\mathbb{R}^{n})$. We claim that $f \in W^{\alpha, 1}(\mathbb{R}^{n})$. Indeed, take $(f_k)_{k\in\mathbb{N}}\subset C^\infty(\mathbb{R}^n)\cap BV(\mathbb{R}^n)$ such that $f_k\to f$ in $L^1(\mathbb{R}^n)$ and $\|\nabla f_k\|_{L^1(\mathbb{R}^n; \mathbb{R}^{n})}\to|Df|(\mathbb{R}^n)$ as $k\to+\infty$ (for instance, see~\cite{EG15}*{Theorem~5.3}). Since $W^{1,1}(\mathbb{R}^n) \subset W^{\alpha,1}(\mathbb{R}^n)$ (the proof of this inclusion is similar to the one of~\cite{DiNPV12}*{Proposition~2.2}, for example), by Fatou's Lemma we get that
\begin{equation*}
\begin{split}
\|f\|_{W^{\alpha,1}(\mathbb{R}^n)}
&\le\liminf_{k\to+\infty}\|f_k\|_{W^{\alpha,1}(\mathbb{R}^n)}\\
&\le c_{n,\alpha}\liminf_{k\to+\infty}\|f_k\|_{W^{1,1}(\mathbb{R}^n)}\\
&=c_{n,\alpha}\lim_{k\to+\infty}(\|f_k\|_{L^1(\mathbb{R}^n)}+|Df_k|(\mathbb{R}^n))\\
&=c_{n,\alpha}\|f\|_{BV(\mathbb{R}^n)}.
\end{split}
\end{equation*}
Since $|Df|(\mathbb{R}^n)<+\infty$, by \cref{result:riesz_kernel_estimate} the function in~\eqref{eq:weak_frac_grad_repr} is well defined in~$L^1_{\loc}(\mathbb{R}^n)$. Fix $\phi \in C^{\infty}_{c}(\mathbb{R}^n; \mathbb{R}^{n})$. By \cref{prop:frac_div_repr}, we can write
\begin{equation*}
\int_{\mathbb{R}^{n}} f(x) \, \div^{\alpha} \phi(x) \, dx
= \frac{\mu_{n, \alpha}}{n + \alpha - 1} \int_{\mathbb{R}^{n}} \int_{\mathbb{R}^n} f(x)\,\frac{\div \phi(y)}{|y - x|^{n + \alpha-1}} \, dy \, dx.
\end{equation*}
Recalling \cref{result:riesz_kernel_estimate}, applying Fubini's Theorem twice and integrating by parts, we obtain
\begin{align*}
\int_{\mathbb{R}^{n}} \int_{\mathbb{R}^n} f(x)\,\frac{\div \phi(y)}{|y - x|^{n + \alpha-1}} \, dy \, dx
& = \int_{\mathbb{R}^{n}} \int_{\mathbb{R}^n} f(x)\,\frac{\div_y \phi(x+y)}{|y|^{n + \alpha-1}} \, dy \, dx \\
& = \int_{\mathbb{R}^{n}} \int_{\mathbb{R}^n} f(x)\,\frac{\div_x \phi(x+y)}{|y|^{n + \alpha-1}} \, dy \, dx \\
& = \int_{\mathbb{R}^{n}} |y|^{1-n - \alpha}\int_{\mathbb{R}^n} f(x)\,\div\phi(x+y) \, dx \, dy \\
& = - \int_{\mathbb{R}^n} |y|^{1 - n - \alpha} \int_{\mathbb{R}^{n}} \phi(y + x) \cdot \, d Df(x) \, dy \\
& = - \int_{\mathbb{R}^{n}} \int_{\mathbb{R}^n} \frac{\phi(y)}{|y - x|^{n + \alpha -1}} \, dy \cdot \, d Df(x) \\
& = - \int_{\mathbb{R}^{n}} \phi(y) \cdot \int_{\mathbb{R}^n} \frac{d Df(x)}{|x-y|^{n + \alpha -1}} \, dy.
\end{align*}
Thus we conclude that
\begin{equation*}
\int_{\mathbb{R}^{n}} f(x) \, \div^{\alpha} \phi(x) \, dx
=
- \frac{\mu_{n, \alpha}}{n + \alpha - 1} \int_{\mathbb{R}^{n}} \phi(y) \cdot \int_{\mathbb{R}^n} \frac{d Df(x)}{|x-y|^{n + \alpha -1}} \, dy.
\end{equation*}
Recalling~\eqref{eq:duality_Sobolev_frac}, this proves~\eqref{eq:weak_frac_grad_repr} and the proof is complete.
\end{proof}
\subsection{The space \texorpdfstring{$S^{\alpha,p}(\mathbb{R}^n)$}{Sˆ{alpha,p}(Rˆn)} and the inclusion \texorpdfstring{$S^{\alpha,1}(\mathbb{R}^n)\subset BV^\alpha(\mathbb{R}^n)$}{Sˆ{alpha,1}(Rˆn) in BVˆalpha(Rˆn)}}
\label{subsec:space_S_alpha_p}
It is now tempting to approach fractional Sobolev spaces from a distributional point of view. Recalling \cref{prop:frac_div_repr}, we can give the following definition.
\begin{definition}[Weak $\alpha$-gradient]
Let $\alpha\in(0,1)$, $p\in[1,+\infty]$, $f\in L^p(\mathbb{R}^n)$. We say that $g\in L^1_{\loc}(\mathbb{R}^n;\mathbb{R}^n)$ is a \emph{weak $\alpha$-gradient} of $f$, and we write $g=\nabla^\alpha_w f$, if
\begin{equation*}
\int_{\mathbb{R}^n}f\,\div^\alpha\phi\, dx
=-\int_{\mathbb{R}^n}g\cdot\phi\, dx
\end{equation*}
for all $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$.
\end{definition}
For $\alpha\in(0,1)$ and $p\in[1,+\infty]$, we can thus introduce the \emph{distributional fractional Sobolev space} $(S^{\alpha,p}(\mathbb{R}^n),\|\cdot\|_{S^{\alpha,p}(\mathbb{R}^n)})$ letting
\begin{equation}\label{eq:def_distrib_frac_Sobolev}
S^{\alpha,p}(\mathbb{R}^n):=\set*{f\in L^p(\mathbb{R}^n) : \exists\, \nabla^\alpha_w f \in L^p(\mathbb{R}^n;\mathbb{R}^n)}
\end{equation}
and
\begin{equation}\label{eq:def_distrib_frac_Sobolev_norm}
\|f\|_{S^{\alpha,p}(\mathbb{R}^n)}:=\|f\|_{L^p(\mathbb{R}^n)}+\|\nabla^\alpha_w f\|_{L^p(\mathbb{R}^n; \mathbb{R}^{n})},
\qquad
\forall f\in S^{\alpha,p}(\mathbb{R}^n).
\end{equation}
We omit the standard proof of the following result.
\begin{proposition}[$S^{\alpha,p}$ is a Banach space]
Let $\alpha\in(0,1)$ and $p\in[1,+\infty]$. The space $(S^{\alpha,p}(\mathbb{R}^n),\|\cdot\|_{S^{\alpha,p}(\mathbb{R}^n)})$ is a Banach space.
\end{proposition}
We leave the proof of the following interpolation result to the reader.
\begin{lemma}[Interpolation]
Let $\alpha\in(0,1)$ and $p_1,p_2\in[1,+\infty]$, with $p_1\le p_2$. Then
\begin{equation*}
S^{\alpha,p_1}(\mathbb{R}^n)\cap S^{\alpha,p_2}(\mathbb{R}^n)\subset S^{\alpha,q}(\mathbb{R}^n)
\end{equation*}
with continuous embedding for all $q\in[p_1,p_2]$.
\end{lemma}
Taking advantage of the techniques developed in the study of the space $BV^\alpha(\mathbb{R}^n)$ above, we are able to prove the following approximation result.
\begin{theorem}[Approximation by $C^\infty\cap S^{\alpha,p}$ functions]\label{result:approx_by_smooth_S_alpha_p}
Let $\alpha\in(0,1)$ and $p\in[1,+\infty)$. The set $C^\infty(\mathbb{R}^n)\cap S^{\alpha,p}(\mathbb{R}^n)$ is dense in $S^{\alpha,p}(\mathbb{R}^n)$.
\end{theorem}
\begin{proof}
Let $(\rho_\varepsilon)_{\varepsilon>0}\subset C^\infty_c(\mathbb{R}^n)$ be as in~\eqref{eq:def_rho_eps}. Fix $f\in S^{\alpha,p}(\mathbb{R}^n)$ and consider $f_\varepsilon:=f*\rho_\varepsilon$ for all $\varepsilon>0$. By \cref{result:commutation_mollifier}, it is easy to check that $f_\varepsilon\in C^\infty(\mathbb{R}^n)\cap S^{\alpha,p}(\mathbb{R}^n)$ with $\nabla^\alpha_w f_\varepsilon=\rho_\varepsilon*\nabla^\alpha_w f$ for all $\varepsilon>0$, so that the conclusion follows by standard properties of the convolution.
\end{proof}
Given $\alpha\in(0,1)$ and $p\in[1,+\infty]$, it is easy to see that, if $f\in C^\infty_c(\mathbb{R}^n)$, then, by \cref{result:duality}, $f\in S^{\alpha,p}(\mathbb{R}^n)$ with $\nabla^\alpha_w f=\nabla^\alpha f$. In the case $p=1$, we can prove that $C^\infty_c(\mathbb{R}^n)$ is also a dense subset of~$S^{\alpha,1}(\mathbb{R}^n)$.
\begin{theorem}[Approximation by $C^\infty_c$ functions]\label{result:approx_by_smooth_c_S_alpha_1}
Let $\alpha\in(0,1)$. The set $C^\infty_c(\mathbb{R}^n)$ is dense in $S^{\alpha,1}(\mathbb{R}^n)$.
\end{theorem}
\begin{proof}
Let $(\eta_R)_{R>0}\subset C^\infty_c(\mathbb{R}^n)$ be as in~\eqref{eq:def_cut-off}. Thanks to \cref{result:approx_by_smooth_S_alpha_p}, it is enough to prove that $f\eta_R\to f$ in $S^{\alpha,1}(\mathbb{R}^n)$ as $R\to+\infty$ for all $f\in C^\infty(\mathbb{R}^n)\cap S^{\alpha,1}(\mathbb{R}^n)$. Clearly, $f\eta_R\to f$ in $L^1(\mathbb{R}^n)$ as $R\to+\infty$. We now argue as in the proof of \cref{result:approx_by_smooth_c_BV}. Fix $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$. Then, by \cref{lem:Leibniz_frac_div}, we get
\begin{equation*}
\begin{split}
\int_{\mathbb{R}^n}f\eta_R\,\div^\alpha\phi\,dx
&=\int_{\mathbb{R}^n}f\,\div^\alpha(\eta_R\phi)\,dx
-\int_{\mathbb{R}^n}f\,\phi\cdot\nabla^\alpha\eta_R\,dx
-\int_{\mathbb{R}^n}f\,\div^\alpha_{\mathrm{NL}}(\eta_R, \phi)\,dx.
\end{split}
\end{equation*}
Since $f\in S^{\alpha,1}(\mathbb{R}^n)$, we have
\begin{equation*}
\int_{\mathbb{R}^n}f\,\div^\alpha(\eta_R\phi)\,dx
=-\int_{\mathbb{R}^n}\eta_R\phi\cdot\nabla^\alpha_w f\,dx.
\end{equation*}
Since $f\eta_R\in C^\infty_c(\mathbb{R}^n)$, we also have
\begin{equation*}
\int_{\mathbb{R}^n}f\eta_R\,\div^\alpha\phi\,dx
=-\int_{\mathbb{R}^n}\phi\cdot\nabla^\alpha(\eta_R f)\,dx.
\end{equation*}
Thus we can write
\begin{equation*}
\begin{split}
\int_{\mathbb{R}^n}(\nabla^\alpha_w f-\nabla^\alpha(\eta_R f))\cdot\phi\,dx
& = \int_{\mathbb{R}^n}(1-\eta_R)\phi\cdot\nabla^\alpha_w f\,dx\\
&\quad-\int_{\mathbb{R}^n}f\,\phi\cdot\nabla^\alpha\eta_R\,dx
-\int_{\mathbb{R}^n}f\,\div^\alpha_{\mathrm{NL}}(\eta_R, \phi)\,dx.
\end{split}
\end{equation*}
Moreover, we have
\begin{equation*}
\abs*{\int_{\mathbb{R}^n}f\,\phi\cdot\nabla^\alpha\eta_R\,dx}
\le
\mu_{n, \alpha} \|\phi\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}\int_{\mathbb{R}^n}|f(x)|\int_{\mathbb{R}^n}\frac{|\eta_R(y)-\eta_R(x)|}{|y-x|^{n+\alpha}}\,dy\, dx
\end{equation*}
and, similarly,
\begin{equation*}
\abs*{\int_{\mathbb{R}^n}f\,\div^\alpha_{\mathrm{NL}}(\eta_R, \phi)\,dx}
\le
2\mu_{n, \alpha}\|\phi\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}\int_{\mathbb{R}^n}|f(x)|\int_{\mathbb{R}^n}\frac{|\eta_R(y)-\eta_R(x)|}{|y-x|^{n+\alpha}}\,dy\,dx.
\end{equation*}
Combining these two estimates, we get that
\begin{equation*}
\begin{split}
\left | \int_{\mathbb{R}^n}(\nabla^\alpha_w f-\nabla^\alpha(\eta_R f))\cdot\phi\,dx\right |
&\le
\|\phi\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}\int_{\mathbb{R}^n}(1-\eta_R)|\nabla^\alpha_w f|\,dx\\
&\quad+3\mu_{n, \alpha}\|\phi\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}\int_{\mathbb{R}^n}|f(x)|\int_{\mathbb{R}^n}\frac{|\eta_R(y)-\eta_R(x)|}{|y-x|^{n+\alpha}}\,dy\,dx.
\end{split}
\end{equation*}
We thus conclude that
\begin{align*}
\|\nabla^\alpha_w f-\nabla^\alpha(\eta_R f)\|_{L^1(\mathbb{R}^n; \mathbb{R}^n)}
\le & \int_{\mathbb{R}^n}(1-\eta_R)|\nabla^\alpha_w f|\,dx \\
& +3\mu_{n, \alpha}\int_{\mathbb{R}^n}|f(x)|\int_{\mathbb{R}^n}\frac{|\eta_R(y)-\eta_R(x)|}{|y-x|^{n+\alpha}}\,dy\,dx.
\end{align*}
Therefore $\nabla^\alpha(\eta_R f)\to\nabla^\alpha_w f$ in $L^1(\mathbb{R}^n; \mathbb{R}^{n})$ as $R\to+\infty$. Indeed, we have
\begin{equation*}
\lim_{R\to+\infty}\int_{\mathbb{R}^n}(1-\eta_R)|\nabla^\alpha_w f|\,dx=0
\end{equation*}
combining~\eqref{eq:def_cut-off} with Lebesgue's Dominated Convergence Theorem and
\begin{equation*}
\lim_{R\to+\infty}\int_{\mathbb{R}^n}|f(x)|\int_{\mathbb{R}^n}\frac{|\eta_R(y)-\eta_R(x)|}{|y-x|^{n+\alpha}}\,dy\,dx=0
\end{equation*}
combining~\eqref{eq:Lip_nabla_estim_1}, \eqref{eq:Lip_nabla_estim_2} and~\eqref{eq:def_cut-off} with Lebesgue's Dominated Convergence Theorem.
\end{proof}
We do not know if $C^\infty_c(\mathbb{R}^n)$ is a dense subset of $S^{\alpha,p}(\mathbb{R}^n)$ for $\alpha\in(0,1)$ and $p\in(1,+\infty)$. In other words, defining
\begin{equation*}
S^{\alpha,p}_0(\mathbb{R}^n):=\closure[-1]{C^\infty_c(\mathbb{R}^n)}^{\|\cdot\|_{S^{\alpha,p}(\mathbb{R}^n)}},
\end{equation*}
we do not know if the (continuous) inclusion $S^{\alpha,p}_0(\mathbb{R}^n)\subset S^{\alpha,p}(\mathbb{R}^n)$ is strict.
The space $(S^{\alpha,p}_0(\mathbb{R}^n),\|\cdot\|_{S^{\alpha,p}(\mathbb{R}^n)})$ was introduced in~\cite{SS15} (with a different, but equivalent, norm). Thanks to~\cite{SS15}*{Theorem 1.7}, for all $\alpha \in (0, 1)$ and $p \in (1, + \infty)$ we have $S^{\alpha,p}_0(\mathbb{R}^n) = L^{\alpha, p}(\mathbb{R}^{n})$, where $L^{\alpha, p}(\mathbb{R}^{n})$ is the \emph{Bessel potential space}, see~\cite{SS15}*{Definition~2.1}. It is known that $L^{\alpha+\varepsilon,p}(\mathbb{R}^n)\subset W^{\alpha, p}(\mathbb{R}^{n})\subset L^{\alpha-\varepsilon,p}(\mathbb{R}^n)$ with continuous embeddings for all $\alpha\in(0,1)$, $p\in(1,+\infty)$ and $0<\varepsilon<\min\set*{\alpha,1-\alpha}$, see~\cite{SS15}*{Theorem~2.2}. In the particular case $p=2$, it holds that $L^{\alpha, 2}(\mathbb{R}^{n})=W^{\alpha, 2}(\mathbb{R}^{n})$ for all $\alpha \in (0, 1)$, see~\cite{SS15}*{Theorem 2.2}. In addition, $W^{\alpha,p}(\mathbb{R}^n)\subset L^{\alpha,p}(\mathbb{R}^n)$ with continuous embedding for all $\alpha\in(0,1)$ and $p\in(1,2]$, see~\cite{S70}*{Chapter~V, Section~5.3}.
\begin{proposition}[Relation between $W^{\alpha,p}$ and $S^{\alpha,p}$]
The following properties hold.
\begin{enumerate}[\indent (i)]
\item\label{item:embedding_W_S_1} If $\alpha\in(0,1)$ and $p\in[1,2]$, then $W^{\alpha,p}(\mathbb{R}^n)\subset S^{\alpha,p}(\mathbb{R}^n)$ with continuous embedding.
\item\label{item:embedding_W_S_2} If $0<\alpha<\beta<1$ and $p\in(2,+\infty]$, then $W^{\beta,p}(\mathbb{R}^n)\subset S^{\alpha,p}(\mathbb{R}^n)$ with continuous embedding.
\end{enumerate}
\end{proposition}
\begin{proof}
Property~\eqref{item:embedding_W_S_1} follows from the discussion above for the case $p\in(1,2]$ and from \cref{result:Sobolev_subset_BV} for the case $p=1$. Property~\eqref{item:embedding_W_S_2} follows from the discussion above for the case $p\in(2,+\infty)$, while for the case $p=+\infty$ it is enough to observe that
\begin{equation*}
\begin{split}
\|\nabla^\alpha f\|_{L^\infty(\mathbb{R}^n; \mathbb{R}^{n})}
&\le \mu_{n, \alpha} \sup_{x\in\mathbb{R}^n}\int_{\mathbb{R}^n}\frac{|f(y)-f(x)|}{|y-x|^{n+\alpha}}\,dy\\
&\le2 \mu_{n, \alpha} \|f\|_{L^\infty(\mathbb{R}^n)}\int_{\set{|y|>1}}\frac{dy}{|y|^{n+\alpha}}
+ \mu_{n, \alpha} [f]_{W^{\beta,\infty}(\mathbb{R}^n)}\int_{\set{|y|\le1}}\frac{dy}{|y|^{n+\alpha-\beta}}\\
&\le c_{n,\alpha,\beta}\|f\|_{W^{\beta,\infty}(\mathbb{R}^n)}
\end{split}
\end{equation*}
for all $f\in W^{\beta,\infty}(\mathbb{R}^n)$.
\end{proof}
As in the classical case, we have $S^{\alpha,1}(\mathbb{R}^n)\subset BV^\alpha(\mathbb{R}^n)$ with continuous embedding.
\begin{theorem}[$S^{\alpha,1}(\mathbb{R}^n)\subset BV^\alpha(\mathbb{R}^n)$]\label{result:S_alpha_1_in_BV_alpha}
Let $\alpha\in(0,1)$. If $f\in BV^\alpha(\mathbb{R}^n)$, then $f\in S^{\alpha,1}(\mathbb{R}^n)$ if and only if $|D^\alpha f|\ll\Leb{n}$, in which case
\begin{equation*}
D^\alpha f=\nabla^\alpha_w f\,\Leb{n}
\quad
\text{in $\M(\mathbb{R}^n;\mathbb{R}^n)$}.
\end{equation*}
\end{theorem}
\begin{proof}
Let $f\in BV^\alpha(\mathbb{R}^n)$ and assume that $|D^\alpha f|\ll\Leb{n}$. Then $D^\alpha f=g\,\Leb{n}$ for some $g\in L^1(\mathbb{R}^n;\mathbb{R}^n)$. But then, by \cref{th:structure_BV_alpha}, we must have
\begin{equation*}
\int_{\mathbb{R}^n}f\,\div^\alpha\phi\,dx=-\int_{\mathbb{R}^n} g\cdot\phi\,dx
\end{equation*}
for all $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$, so that $f\in S^{\alpha,1}(\mathbb{R}^n)$ with $\nabla^\alpha_w f=g$. Viceversa, if $f\in S^{\alpha,1}(\mathbb{R}^n)$ then
\begin{equation*}
\int_{\mathbb{R}^n}f\,\div^\alpha\phi\,dx=-\int_{\mathbb{R}^n} \phi\cdot\nabla^\alpha_w f\,dx
\end{equation*}
for all $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$, so that $f\in BV^\alpha(\mathbb{R}^n)$ with $D^\alpha f=\nabla^\alpha_w f\,\Leb{n}$ in $\M(\mathbb{R}^n;\mathbb{R}^n)$.
\end{proof}
\subsection{The inclusion \texorpdfstring{$S^{\alpha,1}(\mathbb{R}^n)\subset BV^\alpha(\mathbb{R}^n)$}{Sˆ{alpha,1}(Rˆn) in BVˆalpha(Rˆn)} is strict}
It seems natural to ask whether the inclusion $S^{\alpha,1}(\mathbb{R}^n)\subset BV^\alpha(\mathbb{R}^n)$ is strict as in the classical case. We start to solve this problem in the case $n = 1$.
\begin{theorem}[$BV^{\alpha}(\mathbb{R})\setminus S^{\alpha, 1}(\mathbb{R})\neq\varnothing$] \label{prop:BV_alpha_W_alpha_inclusion_R}
Let $\alpha \in (0, 1)$. The inclusion $S^{\alpha, 1}(\mathbb{R}) \subset BV^{\alpha}(\mathbb{R})$ is strict, since for any $a, b \in \mathbb{R}$, with $a \neq b$, the function
\begin{equation*}
f_{a, b, \alpha}(x) := |x - b|^{\alpha - 1} \sgn(x - b) - |x - a|^{\alpha - 1} \sgn(x - a)
\end{equation*}
satisfies $f_{a, b, \alpha} \in BV^{\alpha}(\mathbb{R})$ with
\begin{equation} \label{eq:strict_inclusion_example_nabla_1}
D^{\alpha} f_{a, b, \alpha} = \frac{\delta_{b} - \delta_{a}}{\mu_{1, - \alpha}}
\end{equation}
in the sense of finite Radon measures.
\end{theorem}
\begin{proof}
Let $a, b \in \mathbb{R}$ be fixed with $a \neq b$. One can easily check that $f_{a, b, \alpha} \in L^{1}(\mathbb{R})$. Since $n=1$, we have $\nabla^{\alpha}=\div^{\alpha}$. Thus, \eqref{eq:strict_inclusion_example_nabla_1} follows from~\eqref{eq:div_alpha_delta}, proving that $f\in BV^\alpha(\mathbb{R})$. But $|D^\alpha f_{a, b, \alpha}|\perp\Leb{1}$, so that $f_{a, b, \alpha}\notin S^{\alpha,1}(\mathbb{R})$ by \cref{result:S_alpha_1_in_BV_alpha}.
\end{proof}
\begin{remark} \label{rem:no_GNS_one_dim}
Note that $f_{a, b, \alpha} \in BV^{\alpha}(\mathbb{R})\setminus L^{\frac{1}{1 - \alpha}}(\mathbb{R})$, since
\begin{equation*}
|f_{a, b, \alpha}(x)|^{\frac{1}{1 - \alpha}}
\sim
\begin{cases}
|x - a|^{-1} & \text{as $x \to a$},\\[3mm]
|x - b|^{-1} & \text{as $x \to b$}.
\end{cases}
\end{equation*}
Thus, \cref{thm:GNS_immersion} cannot hold for $n = 1$.
\end{remark}
For the case $n>1$, we need to recall the definition of the \emph{fractional Laplacian operator~$(- \Delta)^{\frac{\alpha}{2}}$} and some of its properties.
Following~\cite{S18}, for any $f \in C^{\infty}_{c}(\mathbb{R}^{n})$ we set
\begin{equation} \label{eq:fract_Laplacian}
(- \Delta)^{\frac{\alpha}{2}} f(x) :=
\begin{cases}
\displaystyle \nu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{f(x + h)}{|h|^{n + \alpha}} \, dh & \text{if} \ \alpha \in (- 1, 0), \\[5mm]
f(x) & \text{if} \ \alpha = 0, \\[5mm]
\displaystyle \nu_{n, \alpha} \int_{\mathbb{R}^{n}} \frac{f(x + h) - f(x)}{|h|^{n + \alpha}}\,dh & \text{if} \ \alpha \in (0, 1), \\[5mm]
\displaystyle \nu_{n, \alpha} \lim_{\varepsilon \to 0} \int_{\{ |h| > \varepsilon \}} \frac{f(x + h) - f(x)}{|h|^{n + \alpha}}\,dh & \text{if} \ \alpha \in [1, 2),
\end{cases}
\end{equation}
where
\begin{equation} \label{eq:nu_alpha_constant}
\nu_{n, \alpha} := 2^{\alpha} \pi^{- \frac{n}{2}} \frac{\Gamma \left ( \frac{n + \alpha}{2} \right )}{\Gamma \left ( - \frac{ \alpha}{2} \right )}.
\end{equation}
We stress the fact that this definition is consistent with the previous definitions of fractional gradient and divergence in the sense that
\begin{equation*}
-\div^{\alpha} \nabla^{\beta} = (- \Delta)^{\frac{\alpha + \beta}{2}}
\end{equation*}
for any $\alpha \in (-1, 1)$ and $\beta \in (0, 1)$ (see \cite{S18}*{Theorem 5.3}), so that, in particular,
\begin{equation*}
-\div^{\alpha} \nabla^{\alpha} = (- \Delta)^{\alpha}
\end{equation*}
for any $\alpha \in (0, 1)$.
In the case $\alpha\in(-1,0)$,
we have
\begin{equation*}
(- \Delta)^{\frac{\alpha}{2}}=I_{-\alpha}
\quad
\text{on $C^\infty_c(\mathbb{R}^n)$},
\end{equation*}
where $I_\alpha$ is as in~\eqref{eq:Riesz_potential_def}.
In the case $\alpha\in(0,1)$, notice that
\begin{equation} \label{eq:fract_Laplacian_L_1}
\| (- \Delta)^{\frac{\alpha}{2}} f \|_{L^{1}(\mathbb{R}^{n})} \le \nu_{n, \alpha} [f]_{W^{\alpha, 1}(\mathbb{R}^{n})}
\end{equation}
for all $f \in C^{\infty}_{c}(\mathbb{R}^{n})$. Thus the linear operator
\begin{equation*}
(- \Delta)^{\frac{\alpha}{2}}\colon C^{\infty}_{c}(\mathbb{R}^{n}) \to L^1(\mathbb{R}^n)
\end{equation*}
can be continuously extended to a linear operator
\begin{equation*}
(- \Delta)^{\frac{\alpha}{2}} \colon W^{\alpha, 1}(\mathbb{R}^{n}) \to L^{1}(\mathbb{R}^{n}),
\end{equation*}
for which we retain the same notation.
Given $\alpha\in(0,1)$ and $\varepsilon>0$, for all $f\in W^{\alpha,1}(\mathbb{R}^n)$ we also set
\begin{equation*}
(- \Delta)^{\alpha/2}_\varepsilon f(x):=
\nu_{n, \alpha} \int_{\set{|h|>\varepsilon}} \frac{f(x + h) - f(x)}{|h|^{n + \alpha}}\,dh.
\end{equation*}
By Lebesgue's Dominate Convergence Theorem, we have that
\begin{equation*}
\lim_{\varepsilon\to0}\|(- \Delta)^{\alpha/2}_\varepsilon f-(- \Delta)^{\frac{\alpha}{2}}f\|_{L^1(\mathbb{R}^n)}=0
\end{equation*}
for all $f\in W^{\alpha,1}(\mathbb{R}^n)$. Thus, arguing as in the proof of~\cite{S98}*{Lemma~2.4} (see also~\cite{SKM93}*{Section~25.1}), for all $f\in W^{\alpha,1}(\mathbb{R}^n)$ we have
\begin{equation}\label{eq:riesz_inversion_laplacian_sobolev}
I_{\alpha}(- \Delta)^{\frac{\alpha}{2}}f= f
\quad
\text{in $L^1(\mathbb{R}^n)$}.
\end{equation}
Taking advantage of the identity in~\eqref{eq:riesz_inversion_laplacian_sobolev}, we can prove the following result.
\begin{lemma}[Relation between $BV^\alpha(\mathbb{R}^n)$ and $bv(\mathbb{R}^n)$]\label{result:correspondence_BV_alpha_bv}
Let $\alpha\in(0,1)$. The following properties hold.
\begin{enumerate}[\indent (i)]
\item\label{item:BV_alpha_bv_1} If $f\in BV^\alpha(\mathbb{R}^n)$, then $u:=I_{1-\alpha}f\in bv(\mathbb{R}^n)$ with $Du=D^\alpha f$ in $\M(\mathbb{R}^{n}; \mathbb{R}^{n})$.
\item\label{item:BV_alpha_bv_2} If $u\in BV(\mathbb{R}^n)$, then $f:= (-\Delta)^{\frac{1-\alpha}{2}}u\in BV^\alpha(\mathbb{R}^n)$ with
\begin{equation*}
\|f\|_{L^1(\mathbb{R}^n)}\le c_{n,\alpha}\|u\|_{BV(\mathbb{R}^n)}
\quad\text{and}\quad
D^\alpha f=D u
\quad\text{in $\M(\mathbb{R}^{n}; \mathbb{R}^{n})$}.
\end{equation*}
As a consequence, the operator $(-\Delta)^{\frac{1-\alpha}{2}}\colon BV(\mathbb{R}^n)\to BV^\alpha(\mathbb{R}^n)$ is continuous.
\end{enumerate}
\end{lemma}
\begin{proof}
We prove the two properties separately.
\smallskip
\textit{Proof of~\eqref{item:BV_alpha_bv_1}}. Let $f\in BV^\alpha(\mathbb{R}^n)$. Since $f \in L^{1}(\mathbb{R}^{n})$, we have $I_{1 - \alpha} f \in L^{1}_{\rm loc}(\mathbb{R}^{n})$. By Fubini's Theorem, for any $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$ we have
\begin{equation} \label{eq:proof_correspondence_BV_alpha_bv}
\int_{\mathbb{R}^n}f\,\div^\alpha\phi\,dx
=\int_{\mathbb{R}^n}f\,I_{1-\alpha}\div\phi\,dx
=\int_{\mathbb{R}^n}u\,\div\phi\,dx,
\end{equation}
proving that $u:=I_{1-\alpha}f\in bv(\mathbb{R}^n)$ with $Du=D^\alpha f$ in $\M(\mathbb{R}^{n}; \mathbb{R}^{n})$.
\smallskip
\textit{Proof of~\eqref{item:BV_alpha_bv_2}}. Let $u\in BV(\mathbb{R}^n)$. By \cref{result:Sobolev_subset_BV}, we know that $u\in W^{1-\alpha,1}(\mathbb{R}^n)$, so that $f:= (-\Delta)^{\frac{1-\alpha}{2}}u\in L^1(\mathbb{R}^n)$ with $\|f\|_{L^1(\mathbb{R}^n)}\le c_{n,\alpha}\|u\|_{BV(\mathbb{R}^n)}$ by~\eqref{eq:W_alpha_norm_bound_BV} and~\eqref{eq:fract_Laplacian_L_1}. Then, arguing as before, for any $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$ we get \eqref{eq:proof_correspondence_BV_alpha_bv}, since we have $I_{1-\alpha}f=u$ in~$L^1(\mathbb{R}^n)$ by \eqref{eq:riesz_inversion_laplacian_sobolev}. The proof is complete.
\end{proof}
\begin{remark}[Integrability issues]
Note that the inclusion $I_{1-\alpha}(BV^\alpha(\mathbb{R}^n))\subset L^1_{\loc}(\mathbb{R}^n)$ in \cref{result:correspondence_BV_alpha_bv} above is sharp. Indeed, by Tonelli's Theorem it is easily seen that $I_{1-\alpha}\chi_E\notin L^1(\mathbb{R}^n)$ whenever $\chi_E\in W^{\alpha,1}(\mathbb{R}^n)$. However, when $n\ge2$, by \cref{thm:GNS_immersion} and by Hardy--Littlewood--Sobolev inequality (see~\cite{S70}*{Chapter~V, Section~1.2} for instance), the map $I_{1-\alpha}\colon BV^\alpha\to L^p(\mathbb{R}^n)$ is continuous for each $p\in\left(\frac{n}{n-1+\alpha},\frac{n}{n-1}\right]$.
\end{remark}
As a consequence of \cref{result:correspondence_BV_alpha_bv}, we can prove that the inclusion $S^{\alpha, 1}(\mathbb{R}^{n})\subset BV^{\alpha}(\mathbb{R}^{n})$ is strict for all $\alpha\in(0,1)$ and $n\ge1$.
\begin{theorem}[$BV^{\alpha}(\mathbb{R}^{n}) \setminus S^{\alpha, 1}(\mathbb{R}^{n}) \neq \varnothing$] \label{thm:strict_inclusion_W_BV_alpha}
Let $\alpha \in (0, 1)$. The inclusion $S^{\alpha, 1}(\mathbb{R}^{n}) \subset BV^{\alpha}(\mathbb{R}^{n})$ is strict.
\end{theorem}
\begin{proof}
Let $u \in BV(\mathbb{R}^n)\setminus W^{1,1}(\mathbb{R}^n)$. By \cref{result:correspondence_BV_alpha_bv}, we know that $f:=(-\Delta)^{\frac{1-\alpha}{2}}u\in BV^\alpha(\mathbb{R}^n)$ with $Du=D^\alpha f$ in $\M(\mathbb{R}^{n}; \mathbb{R}^{n})$. But then $|D^\alpha f|$ is not absolutely continuous with respect to $\Leb{n}$, so that $f\notin S^{\alpha, 1}(\mathbb{R}^{n})$ by \cref{result:S_alpha_1_in_BV_alpha}.
\end{proof}
\subsection{The inclusion \texorpdfstring{$W^{\alpha, 1}(\mathbb{R}^{n}) \subset S^{\alpha,1}(\mathbb{R}^{n})$}{Wˆ{alpha,1}(Rˆn) in Sˆ{alpha,1}(Rˆn)} is strict}
By \cref{thm:strict_inclusion_W_BV_alpha}, we know that the inclusion $W^{\alpha, 1}(\mathbb{R}^{n}) \subset BV^{\alpha}(\mathbb{R}^{n})$ is strict. In the following result we prove that also the inclusion $W^{\alpha, 1}(\mathbb{R}^{n}) \subset S^{\alpha,1}(\mathbb{R}^{n})$ is strict.
\begin{theorem}[$S^{\alpha,1}(\mathbb{R}^{n}) \setminus W^{\alpha, 1}(\mathbb{R}^{n}) \neq \varnothing$] \label{thm:strict_inclusion_W_S_alpha_1}
Let $\alpha \in (0, 1)$. The inclusion $W^{\alpha, 1}(\mathbb{R}^{n}) \subset S^{\alpha,1}(\mathbb{R}^{n})$ is strict.
\end{theorem}
\begin{proof}
We argue by contradiction. If $W^{\alpha, 1}(\mathbb{R}^{n}) = S^{\alpha,1}(\mathbb{R}^{n})$, then the inclusion map $W^{\alpha, 1}(\mathbb{R}^{n})\hookrightarrow S^{\alpha,1}(\mathbb{R}^{n})$ is a linear and continuous bijection. Thus, by the Inverse Mapping Theorem, there must exist a constant $C>0$ such that
\begin{equation}\label{eq:contradiction_S_W}
[g]_{W^{\alpha,1}(\mathbb{R}^n)}\le C\|g\|_{S^{\alpha,1}(\mathbb{R}^n)}
\end{equation}
for all $g \in S^{\alpha, 1}(\mathbb{R}^{n})$. Now let $f\in BV^{\alpha}(\mathbb{R}^{n}) \setminus S^{\alpha, 1}(\mathbb{R}^{n})$ be given by \cref{thm:strict_inclusion_W_BV_alpha}. By \cref{result:approx_by_smooth_c_BV}, there exists $(f_k)_{k\in\mathbb{N}}\subset C^\infty_c(\mathbb{R}^n)$ such that $f_k\to f$ in $L^1(\mathbb{R}^n)$ and $|D^\alpha f_k|(\mathbb{R}^n)\to|D^\alpha f|(\mathbb{R}^n)$ as $k\to+\infty$. Up to extract a subsequence (which we do not relabel for simplicity), we can assume that $f_k(x)\to f(x)$ as $k\to+\infty$ for $\Leb{n}$-a.e.\ $x\in\mathbb{R}^n$. By~\eqref{eq:contradiction_S_W} and Fatou's Lemma, we have that
\begin{equation*}
\begin{split}
[f]_{W^{\alpha,1}(\mathbb{R}^n)}
&\le\liminf_{k\to+\infty}[f_k]_{W^{\alpha,1}(\mathbb{R}^n)}\\
&\le C\liminf_{k\to+\infty}\|f_k\|_{S^{\alpha,1}(\mathbb{R}^n)}\\
&= C\lim_{k\to+\infty}\|f_k\|_{BV^{\alpha}(\mathbb{R}^n)}\\
&=C\,\|f\|_{BV^{\alpha}(\mathbb{R}^n)}<+\infty.
\end{split}
\end{equation*}
Therefore $f\in W^{\alpha,1}(\mathbb{R}^n)$, in contradiction with \cref{thm:strict_inclusion_W_BV_alpha}. We thus must have that the inclusion map $W^{\alpha, 1}(\mathbb{R}^{n})\hookrightarrow S^{\alpha,1}(\mathbb{R}^{n})$ cannot be surjective.
\end{proof}
\subsection{The inclusion \texorpdfstring{$BV^{\alpha}(\mathbb{R}^{n})\subset W^{\beta, 1}(\mathbb{R}^{n})$}{BVˆalpha(Rˆn) in Wˆ{beta,1}(Rˆn)} for \texorpdfstring{$\beta<\alpha$}{beta<alpha}}
Even though the inclusion $W^{\alpha, 1}(\mathbb{R}^{n}) \subset BV^{\alpha}(\mathbb{R}^{n})$ is strict, it is interesting to notice that $BV^{\alpha}(\mathbb{R}^{n})\subset W^{\beta, 1}(\mathbb{R}^{n})$ for all $0<\beta<\alpha < 1$ with continuous embedding.
\begin{theorem}[$BV^{\alpha}(\mathbb{R}^{n}) \subset W^{\beta, 1}(\mathbb{R}^{n})$ for $\beta < \alpha$] \label{result:BV_alpha_W_beta_embedding}
Let $\alpha,\beta\in(0,1)$ with $\beta<\alpha$. Then $BV^{\alpha}(\mathbb{R}^{n}) \subset W^{\beta, 1}(\mathbb{R}^{n})$, with
\begin{equation} \label{eq:BV_alpha_W_beta_embedding}
[f]_{W^{\beta, 1}(\mathbb{R}^{n})} \le C_{n, \alpha, \beta} \, \|f\|_{BV^{\alpha}(\mathbb{R}^{n})},
\end{equation}
for all $f\in BV^{\alpha}(\mathbb{R}^{n})$, where
\begin{equation} \label{eq:BV_alpha_W_beta_embedding_constant}
C_{n, \alpha, \beta} := n \omega_{n} \frac{\alpha2^{\frac{\alpha - \beta}{\beta}} \gamma_{n, \alpha}^{\beta/\alpha}}{\beta (\alpha - \beta)}
\end{equation}
and $\gamma_{n,\alpha}$ is as in~\eqref{eq:def_gamma_n,alpha}.
\end{theorem}
\begin{proof}
Let $f \in C^{\infty}_{c}(\mathbb{R}^{n})$ and $r>0$. By~\eqref{eq:Holder_estimate}, we get
\begin{align*}
[f]_{W^{\beta, 1}(\mathbb{R}^{n})}
& = \int_{\mathbb{R}^{n}} \int_{\mathbb{R}^{n}} \frac{|f(x + y) - f(x)|}{|y|^{n + \beta}} \, dx \, dy\\
& \le \int_{\mathbb{R}^{n}} \frac{1}{|y|^{n + \beta}} \left ( 2 \|f\|_{L^{1}(\mathbb{R}^{n})} \chi_{\mathbb{R}^{n} \setminus B_{r}}(y) + \gamma_{n, \alpha} |y|^{\alpha} \|\nabla^{\alpha} f\|_{L^{1}(\mathbb{R}^{n}; \mathbb{R}^{n})} \chi_{B_{r}}(y) \right ) \, dy \\
& = 2 \frac{n \omega_{n}}{\beta} r^{- \beta} \|f\|_{L^{1}(\mathbb{R}^{n})} + \frac{n \omega_{n}}{\alpha - \beta} \gamma_{n, \alpha} r^{\alpha - \beta} \|\nabla^{\alpha} f\|_{L^{1}(\mathbb{R}^{n}; \mathbb{R}^{n})} \\
& \le \left ( 2 \frac{n \omega_{n}}{\beta} r^{- \beta} + \frac{n \omega_{n}}{\alpha - \beta} \gamma_{n, \alpha} r^{\alpha - \beta} \right ) \| f\|_{BV^{\alpha}(\mathbb{R}^{n})},
\end{align*}
so that both~\eqref{eq:BV_alpha_W_beta_embedding} and~\eqref{eq:BV_alpha_W_beta_embedding_constant} are proved by minimising in $r > 0$ for all $f\in C^\infty_c(\mathbb{R}^n)$. Now let $f\in BV^\alpha(\mathbb{R}^n)$. By \cref{result:approx_by_smooth_c_BV}, there exists $(f_k)_{k\in\mathbb{N}}\subset C^\infty_c(\mathbb{R}^n)$ such that $\|f_k\|_{BV^\alpha(\mathbb{R}^n)}\to\|f\|_{BV^\alpha(\mathbb{R}^n)}$ and $f_k\to f$ a.e.\ as $k\to+\infty$. Thus, by Fatou's Lemma, we get that
\begin{equation*}
[f]_{W^{\beta, 1}(\mathbb{R}^{n})}
\le\liminf_{k\to+\infty}\, [f_k]_{W^{\beta, 1}(\mathbb{R}^{n})}
\le\lim_{k\to+\infty} C_{n, \alpha, \beta}\, \|f_k\|_{BV^{\alpha}(\mathbb{R}^{n})}
=C_{n, \alpha, \beta}\, \|f\|_{BV^{\alpha}(\mathbb{R}^{n})}
\end{equation*}
and the conclusion follows.
\end{proof}
\noindent
Note that the constant in~\eqref{eq:BV_alpha_W_beta_embedding_constant} satisfies
\begin{equation*}
\lim_{\beta\to\alpha^-}C_{n, \alpha, \beta}=+\infty,
\end{equation*}
accordingly to the strict inclusion $W^{\alpha, 1}(\mathbb{R}^{n}) \subset BV^{\alpha}(\mathbb{R}^{n})$. In particular, the function in \cref{prop:BV_alpha_W_alpha_inclusion_R} is such that $f_{a, b, \alpha} \in W^{\beta, 1}(\mathbb{R})$ for all $\beta \in (0, \alpha)$.
As an immediate consequence of \cref{result:BV_alpha_W_beta_embedding}, we have the following result.
\begin{corollary}
Let $0<\beta<\alpha<1$. Then $BV^\alpha(\mathbb{R}^n)\subset BV^\beta(\mathbb{R}^n)$ and $S^{\alpha,1}(\mathbb{R}^n)\subset S^{\beta,1}(\mathbb{R}^n)$ with continuous embeddings.
\end{corollary}
\section{Fractional Caccioppoli sets}
\label{sec:frac_Caccioppoli_sets}
\subsection{Definition of fractional Caccioppoli sets and the Gauss--Green formula}
As in the classical case (see~\cite{AFP00}*{Definition~3.3.5} for instance), we start with the following definition.
\begin{definition}[Fractional Caccioppoli set]\label{def:frac_Caccioppoli_set}
Let $\alpha\in(0,1)$ and let $E\subset\mathbb{R}^n$ be a measurable set. For any open set $\Omega\subset\mathbb{R}^n$, the \emph{fractional Caccioppoli $\alpha$-perimeter in $\Omega$} is the \emph{fractional variation} of $\chi_E$ in $\Omega$, i.e.\
\begin{equation*}
|D^\alpha\chi_E|(\Omega)=\sup\set*{\int_E\div^\alpha\phi\,dx : \phi\in C^\infty_c(\Omega;\mathbb{R}^n),\ \|\phi\|_{L^\infty(\Omega;\mathbb{R}^n)}\le1}.
\end{equation*}
We say that $E$ is a set with \emph{finite fractional Caccioppoli $\alpha$-perimeter in $\Omega$} if $|D^\alpha\chi_E|(\Omega)<+\infty$. We say that $E$ is a set with \emph{locally finite fractional Caccioppoli $\alpha$-perimeter in $\Omega$} if $|D^\alpha\chi_E|(U)<+\infty$ for any $U\Subset\Omega$.
\end{definition}
We can now state the following fundamental result relating non-local distributional gradients of characteristic functions of fractional Caccioppoli sets and vector valued Radon measures.
\begin{theorem}[Gauss--Green formula for fractional Caccioppoli sets]\label{result:Gauss-Green}
Let $\alpha\in(0,1)$ and let $\Omega\subset\mathbb{R}^n$ be an open set. A measurable set $E \subset \mathbb{R}^{n}$ is a set with finite fractional Caccioppoli $\alpha$-perimeter in $\Omega$ if and only if $D^{\alpha} \chi_E \in \M(\Omega; \mathbb{R}^{n})$ and
\begin{equation}\label{eq:Gauss-Green}
\int_E \div^{\alpha} \phi \, dx = - \int_{\Omega} \varphi \cdot d D^{\alpha} \chi_E
\end{equation}
for all $\phi \in C^{\infty}_{c}(\Omega; \mathbb{R}^{n})$. In addition, for any open set $U\subset\Omega$ it holds
\begin{equation}\label{eq:Caccioppoli-measure}
|D^{\alpha} \chi_{E}|(U) = \sup\set*{\int_E\,\div^\alpha\phi\ dx : \phi\in C^\infty_c(U;\mathbb{R}^n),\ \|\phi\|_{L^\infty(U;\mathbb{R}^n)}\le1}.
\end{equation}
\end{theorem}
\begin{proof}
The proof is similar to the one of \cref{th:structure_BV_alpha}. If $D^{\alpha} \chi_E \in \M(\Omega; \mathbb{R}^{n})$ and~\eqref{eq:Gauss-Green} holds, then~$E$ has finite fractional Caccioppoli $\alpha$-perimeter in~$\Omega$ by \cref{def:frac_Caccioppoli_set}.
If~$E$ is a set with finite fractional Caccioppoli $\alpha$-perimeter in $\Omega$, then define the linear functional $L\colon C_{c}^\infty(\Omega; \mathbb{R}^{n})\to\mathbb{R}$ setting
\begin{equation*}
L(\phi) := - \int_E \div^{\alpha}\phi \, dx
\qquad
\forall\phi \in C^{\infty}_{c}(\Omega; \mathbb{R}^{n}).
\end{equation*}
Note that $L$ is well defined thanks to \cref{prop:frac_div_repr}. Since $E$ has finite fractional Caccioppoli $\alpha$-perimeter in $\Omega$, we have
\begin{equation*}
C(U):=\sup\set*{L(\phi) : \phi\in C^\infty_c(U;\mathbb{R}^n),\ \|\phi\|_{L^\infty(U;\mathbb{R}^n)}\le 1}<+\infty
\end{equation*}
for each open set $U\subset\Omega$, so that
\begin{equation*}
\left | L(\phi) \right | \le C(U) \|\phi\|_{L^\infty(U;\mathbb{R}^n)}
\qquad
\forall \phi \in C^{\infty}_{c}(U; \mathbb{R}^{n}).
\end{equation*}
Thus, by the density of $C^{\infty}_{c}(\Omega; \mathbb{R}^{n})$ in $C_{c}(\Omega; \mathbb{R}^{n})$, the functional $L$ can be uniquely extended to a continuous linear functional $\tilde{L}\colon C_{c}(\Omega; \mathbb{R}^{n})\to\mathbb{R}$ and the conclusion follows by Riesz's Representation Theorem.
\end{proof}
\subsection{Lower semicontinuity of fractional variation}
As in the classical case, the variation measure of a set with finite fractional Caccioppoli $\alpha$-perimeter is lower semicontinuous with respect to the local convergence in measure. We also achieve a weak convergence result.
\begin{proposition}[Lower semicontinuity of fractional variation measure]
\label{result:frac_Caccioppoli_perimeter_is_lsc}
Let $\alpha\in(0,1)$ and let $\Omega\subset\mathbb{R}^n$ be an open set. If $(E_k)_{k\in\mathbb{N}}$ is a sequence of sets with finite fractional Caccioppoli $\alpha$-perimeter in $\Omega$ and $\chi_{E_k}\to\chi_E$ in $L^1_{\loc}(\mathbb{R}^n)$, then
\begin{equation} \label{eq:weak_conv}
D^{\alpha} \chi_{E_{k}} \rightharpoonup D^{\alpha} \chi_{E} \ \ \text{in} \ \mathcal{M}(\Omega; \mathbb{R}^{n}),
\end{equation}
and
\begin{equation}\label{eq:lsc_frac_Caccioppoli_perim}
|D^\alpha\chi_E|(\Omega)\le\liminf_{k\to+\infty}|D^\alpha\chi_{E_k}|(\Omega).
\end{equation}
\end{proposition}
\begin{proof}
Up to extract a further subsequence, we can assume that $\chi_{E_k}(x)\to\chi_E(x)$ as $k\to+\infty$ for $\Leb{n}$-a.e.\ $x\in\mathbb{R}^n$. Now let $\phi\in C^\infty_c(\Omega;\mathbb{R}^n)$ be such that $\|\phi\|_{L^{\infty}(\Omega; \mathbb{R}^{n})}\le1$. Then $\div^\alpha\phi\in L^1(\mathbb{R}^n)$ by \cref{prop:frac_div_repr} and so, by Lebesgue's Dominated Convergence Theorem, we have
\begin{align*}
\int_E\div^\alpha\phi\ dx
=\lim_{k\to+\infty}\int_{E_k}\div^\alpha\phi\ dx
=-\lim_{k\to+\infty}\int_\Omega\phi \cdot \ dD^\alpha\chi_{E_k}
\le \liminf_{k\to+\infty} |D^\alpha\chi_{E_k}|(\Omega).
\end{align*}
By \cref{result:Gauss-Green}, we get~\eqref{eq:lsc_frac_Caccioppoli_perim}. The convergence in~\eqref{eq:weak_conv} easily follows.
\end{proof}
\subsection{Fractional isoperimetric inequality}
As a simple application of \cref{thm:GNS_immersion}, we can prove the following fractional isoperimetric inequality.
\begin{theorem}[Fractional isoperimetric inequality] \label{result:isoperimetric_ineq}
Let $\alpha \in (0, 1)$ and $n \ge 2$. There exists a constant $c_{n, \alpha} > 0$ such that
\begin{equation}\label{eq:insoperimetric_ineq}
|E|^{\frac{n-\alpha}{n}} \le c_{n, \alpha} |D^{\alpha} \chi_E|(\mathbb{R}^{n})
\end{equation}
for any set $E\subset\mathbb{R}^n$ such that $|E|<+\infty$ and $|D^{\alpha} \chi_E|(\mathbb{R}^{n})<+\infty$.
\end{theorem}
\begin{proof}
Since $\chi_E\in BV^\alpha(\mathbb{R}^n)$, the result follows directly by \cref{thm:GNS_immersion}.
\end{proof}
\subsection{Compactness}
As an application of \cref{result:compactness_BV_alpha}, we can prove the following compactness result for sets with finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$ (see for instance~\cite{M12}*{Theorem~12.26} for the analogous result in the classical case).
\begin{theorem}[Compactness for sets with finite fractional Caccioppoli $\alpha$-perimeter]\label{result:compactness_Caccioppoli}
Let $\alpha\in(0,1)$ and $R>0$. If $(E_k)_{k\in\mathbb{N}}$ is a sequence of sets with finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$ such that
\begin{equation*}
\sup_{k\in\mathbb{N}}|D^\alpha\chi_{E_k}|(\mathbb{R}^n)<+\infty
\quad\text{and}\quad
E_k\subset B_R\quad \forall k\in\mathbb{N},
\end{equation*}
then there exist a subsequence $(E_{k_j})_{j\in\mathbb{N}}$ and a set $E\subset B_R$ with finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$ such that
\begin{equation*}
\chi_{E_{k_j}}\to\chi_E\text{ in~$L^1(\mathbb{R}^n)$}
\end{equation*}
as $j\to+\infty$.
\end{theorem}
\begin{proof}
Since $E_k\subset B_R$ for all $k\in\mathbb{N}$, we clearly have that $(\chi_{E_k})_{k\in\mathbb{R}^n}\subset BV^\alpha(\mathbb{R}^n)$. By \cref{result:compactness_BV_alpha}, there exist a subsequence $(E_{k_j})_{j\in\mathbb{N}}$ and a function $f\in L^1(\mathbb{R}^n)$ such that $\chi_{E_{k_j}}\to f$ in $L^1_{\loc}(\mathbb{R}^n)$ as $j\to+\infty$. Since again $E_{k_j}\subset B_R$ for all $j\in\mathbb{N}$, we have that $\chi_{E_{k_j}}\to f$ in $L^1(\mathbb{R}^n)$ as $j\to+\infty$. Up to extract a further subsequence (which we do not relabel for simplicity), we can assume that $\chi_{E_{k_j}}(x)\to f(x)$ for $\Leb{n}$-a.e.\ $x\in\mathbb{R}^n$ as $j\to+\infty$, so that $f=\chi_E$ for some $E\subset B_R$. By \cref{result:frac_Caccioppoli_perimeter_is_lsc} we conclude that $E$ has finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$.
\end{proof}
\cref{result:compactness_Caccioppoli} can be applied to prove the following compactness result for sets with locally finite fractional Caccioppoli $\alpha$-perimeter.
\begin{corollary}[Compactness for locally finite fractional Caccioppoli $\alpha$-perimeter sets] \label{result:compactness_Caccioppoli_local}
Let $\alpha \in (0,1)$. If $(E_k)_{k\in\mathbb{N}}$ is a sequence of sets with locally finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$ such that
\begin{equation}\label{eq:uniform_bound_compactness_loc_Caccioppoli}
\sup_{k\in\mathbb{N}}|D^\alpha\chi_{E_k}|(B_R)<+\infty
\quad \forall R>0,
\end{equation}
then there exist a subsequence $(E_{k_j})_{j\in\mathbb{N}}$ and a set $E$ with locally finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$ such that
\begin{equation*}
\chi_{E_{k_j}}\to\chi_E\text{ in~$L^1_{\loc}(\mathbb{R}^n)$}
\end{equation*}
as $j\to+\infty$.
\end{corollary}
\begin{proof}
We divide the proof into two steps, essentially following the strategy presented in the proof of~\cite{M12}*{Corollary~12.27}.
\smallskip
\textit{Step~1}. Let $F\subset\mathbb{R}^n$ be a set with locally finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$. We claim that
\begin{equation}\label{eq:claim_perim_intersection_ball}
|D^\alpha\chi_{F\cap B_R}|(\mathbb{R}^n)\le|D^\alpha\chi_F|(B_R)+3\mu_{n, \alpha} P_\alpha(B_R)
\quad \forall R>0.
\end{equation}
Indeed, let $R'<R$ and, recalling \cref{remark:density_test_in_frac_Sobolev}, let $(u_k)_{k\in\mathbb{N}}\subset C^\infty_c(\mathbb{R}^n)$ be such that $\supp(u_k)\Subset B_R$ and $0\le u_k\le 1$ for all $k\in\mathbb{N}$ and also $u_k\to\chi_{B_{R'}}$ in $W^{\alpha,1}(\mathbb{R}^n)$ as $k\to+\infty$. If $\phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n)$ with $\|\phi\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}\le1$, then
\begin{equation*}
\begin{split}
\int_F u_k\,\div^\alpha\phi\,dx
&=\int_F \div^\alpha(u_k\phi)\,dx
-\int_F \phi\cdot\nabla^\alpha u_k\,dx
-\int_F \div^\alpha_{\mathrm{NL}}(u_k,\phi)\,dx\\
&\le\int_F \div^\alpha(u_k\phi)\,dx
+3\mu_{n, \alpha} [u_k]_{W^{\alpha,1}(\mathbb{R}^n)}\\
&\le|D^\alpha\chi_F|(B_{R'})
+3\mu_{n, \alpha} [u_k]_{W^{\alpha,1}(\mathbb{R}^n)}\\
&\le|D^\alpha\chi_F|(B_R)
+3\mu_{n, \alpha} [u_k]_{W^{\alpha,1}(\mathbb{R}^n)}
\end{split}
\end{equation*}
by \cref{lem:Leibniz_frac_div}. Passing to the limit as $k\to+\infty$, we conclude that
\begin{equation*}
\int_{F\cap B_{R'}} \div^\alpha\phi\,dx
\le|D^\alpha\chi_F|(B_R)
+3\mu_{n, \alpha} P_\alpha(B_{R'})
\end{equation*}
and thus
\begin{equation*}
|D^\alpha\chi_{F\cap B_{R'}}|(\mathbb{R}^n)\le|D^\alpha\chi_F|(B_R)+3\mu_{n, \alpha} P_\alpha(B_R)
\end{equation*}
by \cref{result:Gauss-Green}. Since $\chi_{F\cap B_{R'}}\to\chi_{F\cap B_R}$ in $L^1(\mathbb{R}^n)$ as $R'\to R$, the claim in~\eqref{eq:claim_perim_intersection_ball} follows by \cref{result:frac_Caccioppoli_perimeter_is_lsc}.
\smallskip
\textit{Step~2}. By~\eqref{eq:uniform_bound_compactness_loc_Caccioppoli} and~\eqref{eq:claim_perim_intersection_ball}, we can apply \cref{result:compactness_Caccioppoli} to $(E_k\cap B_j)_{k\in\mathbb{N}}$ for each fixed $j\in\mathbb{N}$. By a standard diagonal argument, we find a subsequence $(E_{k_h})_{h\in\mathbb{N}}$ and a sequence $(F_j)_{j\in\mathbb{N}}$ of sets with finite fractional Caccioppoli $\alpha$-perimeter such that $\chi_{E_{k_h}\cap B_j}\to\chi_{F_j}$ in $L^1(\mathbb{R}^n)$ as $h\to+\infty$ for each $j\in\mathbb{N}$. Up to null sets, we have $F_j\subset F_{j+1}$, so that $\chi_{E_{k_h}}\to\chi_E$ in $L^1_{\loc}(\mathbb{R}^n)$ with $E:=\bigcup_{j\in\mathbb{N}}F_j$. The conclusion thus follows by \cref{result:frac_Caccioppoli_perimeter_is_lsc}.
\end{proof}
\subsection{Fractional reduced boundary}
Thanks to the scaling property of the fractional divergence, we have
\begin{equation} \label{scaling_prop_eq}
D^{\alpha} \chi_{\lambda E} = \lambda^{n - \alpha} (\delta_{\lambda})_{\#}D^{\alpha} \chi_{E}
\quad
\text{on}\ \lambda\Omega,
\end{equation}
where $\delta_{\lambda}(x) = \lambda x$ for all $x\in\mathbb{R}^n$ and $\lambda>0$. Indeed, we can compute
\begin{equation*}
\int_{\lambda E} \div^{\alpha} \phi \, dx
= \lambda^{n} \int_E (\div^{\alpha} \varphi)\circ\delta_\lambda\, dx
= \lambda^{n - \alpha} \int_E \div^{\alpha} (\varphi\circ\delta_\lambda) \, dx
\end{equation*}
for all $\phi\in C^{\infty}_{c}(\Omega; \mathbb{R}^{n})$. In analogy with the classical case, we are thus led to the following definition.
\begin{definition}[Fractional reduced boundary] \label{def:redb_alpha}
Let $\alpha\in (0, 1)$ and let $\Omega\subset\mathbb{R}^n$ be an open set. If $E\subset\mathbb{R}^n$ is a set with finite fractional Caccioppoli $\alpha$-perimeter in $\Omega$, then we say that a point $x\in\Omega$ belongs to the \emph{fractional reduced boundary} of $E$ (inside $\Omega$), and we write $x\in\redb^\alpha E$, if
\begin{equation*}
x\in\supp(D^\alpha \chi_{E})
\qquad\text{and}\qquad
\exists\lim_{r\to0}\frac{D^\alpha\chi_E(B_r(x))}{|D^\alpha\chi_E|(B_r(x))}\in\mathbb{S}^{n-1}.
\end{equation*}
We thus let
\begin{equation*}
\nu_E^\alpha\colon\Omega\cap\redb^\alpha E\to\mathbb{S}^{n-1},
\qquad
\nu_E^\alpha(x):=\lim_{r\to0}\frac{D^\alpha\chi_E(B_r(x))}{|D^\alpha\chi_E|(B_r(x))},
\quad
x\in\Omega\cap\redb^\alpha E,
\end{equation*}
be the (\emph{measure theoretic}) \emph{inner unit fractional normal} to~$E$ (inside $\Omega$).
\end{definition}
As a consequence of \cref{def:redb_alpha} and arguing similarly as in the proof of \cref{result:Lip_test}, if $E\subset\mathbb{R}^n$ is a set with finite fractional Caccioppoli $\alpha$-perimeter in $\Omega$, then the following Gauss--Green formula
\begin{equation}\label{eq:Caccioppoli_Gauss-Green}
\int_E\div^\alpha\phi\, dx
=-\int_{\Omega\cap\redb^\alpha E}\phi\cdot\nu^\alpha_E\ d|D^\alpha\chi_E|,
\end{equation}
holds for any $\varphi \in \Lip_{c}(\Omega; \mathbb{R}^{n})$.
\subsection{Sets of finite fractional perimeter are fractional Caccioppoli sets}
In analogy with the classical case and with the inclusion $W^{\alpha,1}(\mathbb{R}^n)\subset BV^\alpha(\mathbb{R}^n)$, we can show that sets with finite fractional $\alpha$-perimeter have finite fractional Caccioppoli $\alpha$-perimeter. Recall that the \emph{fractional $\alpha$-perimeter} of a set~$E\subset\mathbb{R}$ in an open set $\Omega\subset\mathbb{R}^n$ is defined as
\begin{equation*}
P_{\alpha}(E; \Omega)
:= \int_{\Omega} \int_{\Omega} \frac{|\chi_{E}(x) - \chi_{E}(y)|}{|x - y|^{n + \alpha}} \, dx \, dy + 2 \int_{\Omega} \int_{\mathbb{R}^{n} \setminus \Omega} \frac{|\chi_{E}(x) - \chi_{E}(y)|}{|x - y|^{n + \alpha}} \, dx \, dy,
\end{equation*}
see~\cite{CF17} for an account on this subject.
\begin{proposition}[Sets of finite fractional perimeter are fractional Caccioppoli sets]\label{result:Sobolev_is_Caccioppoli}
Let $\alpha\in(0,1)$ and let $\Omega\subset\mathbb{R}^n$ be an open set. If $E\subset\mathbb{R}^n$ satisfies $P_\alpha(E;\Omega)<+\infty$, then $E$ is a set with finite fractional Caccioppoli $\alpha$-perimeter in $\Omega$ with
\begin{equation}\label{eq:FFPS_is_Caccioppoli_estim}
|D^{\alpha}\chi_E|(\Omega) \le \mu_{n, \alpha} P_\alpha(E;\Omega)
\end{equation}
and
\begin{equation}\label{eq:FFPS_is_Caccioppoli_Gauss-Green}
\int_E\div^\alpha\phi\, dx = - \int_\Omega\phi\cdot\nabla^\alpha\chi_E\, dx
\end{equation}
for all $\phi\in\Lip_c(\Omega;\mathbb{R}^n)$, so that $D^\alpha\chi_E=\nu^\alpha_E\,|D^\alpha\chi_E|=\nabla^\alpha\chi_E\,\Leb{n}$. Moreover, if $E$ is such that $|E|<+\infty$ and $P(E)<+\infty$, then $\chi_{E} \in W^{\alpha, 1}(\mathbb{R}^{n})$ for any $\alpha \in (0, 1)$, and
\begin{equation}\label{eq:FFPS_is_Caccioppoli_repres_formula}
\nabla^{\alpha} \chi_{E}(x) = \frac{\mu_{n, \alpha}}{n + \alpha - 1} \int_{\mathbb{R}^n} \frac{\nu_{E}(y)}{|y - x|^{n + \alpha - 1}} \, d |D \chi_{E}|(y)
\end{equation}
for $\Leb{n}$-a.e.\ $x\in\mathbb{R}^n$.
\end{proposition}
\begin{proof}
Note that $\nabla^\alpha\chi_E\in L^1(\Omega;\mathbb{R}^n)$, because
\begin{equation*}
\begin{split}
\int_\Omega|\nabla^\alpha\chi_E|\,dx
&\le\mu_{n, \alpha} \int_\Omega\int_{\mathbb{R}^n}\frac{|\chi_E(y)-\chi_E(x)|}{|y-x|^{n+\alpha}}\,dy\,dx\\
&\le\mu_{n, \alpha} \int_\Omega\int_\Omega\frac{|\chi_E(y)-\chi_E(x)|}{|y-x|^{n+\alpha}}\,dy\,dx
+\mu_{n, \alpha} \int_\Omega\int_{\mathbb{R}^{n} \setminus \Omega}\frac{|\chi_E(y)-\chi_E(x)|}{|y-x|^{n+\alpha}}\,dy\,dx\\
&\le\mu_{n, \alpha} P_\alpha(E;\Omega).
\end{split}
\end{equation*}
Now let $\phi \in \Lip_c(\Omega;\mathbb{R}^{n})$ be fixed. By Lebesgue's Dominated Convergence Theorem, by~\eqref{eq:cancellation_kernel} and by Fubini's Theorem (applied for each fixed $\varepsilon > 0$), we can compute
\begin{align*}
\int_E \div^{\alpha} \varphi \, dx
& = \mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_E \int_{\{|x -y| > \varepsilon \}} \frac{(y - x) \cdot \varphi(y) }{|y - x|^{n + \alpha + 1}} \, dy \, dx \\
& = \mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_{\Omega} \int_{\{|x -y| > \varepsilon \}} \varphi(y) \cdot \frac{(y - x)\, \chi_E(x) }{|y - x|^{n + \alpha + 1}} \, dx \, dy\\
& = -\mu_{n, \alpha} \lim_{\varepsilon \to 0} \int_\Omega \int_{\{|x -y| > \varepsilon \}} \varphi(y) \cdot \frac{(y - x) (\chi_E(y) - \chi_E(x)) }{|y - x|^{n + \alpha + 1}} \, dx \, dy\\
&=- \int_\Omega \varphi \cdot \nabla^\alpha\chi_E\, dy.
\end{align*}
Thus~\eqref{eq:FFPS_is_Caccioppoli_estim} and~\eqref{eq:FFPS_is_Caccioppoli_Gauss-Green} follow by \cref{result:Gauss-Green} and \cref{def:redb_alpha}. Finally, \eqref{eq:FFPS_is_Caccioppoli_repres_formula} follows from~\eqref{eq:weak_frac_grad_repr}, since $\chi_E\in BV(\mathbb{R}^n)$.
\end{proof}
\noindent
At the present moment, we do not know if $|D^{\alpha}\chi_E|(\Omega)<+\infty$ implies that $P_\alpha(E;\Omega)<+\infty$.
\begin{remark}[$\redb^\alpha E$ is not $\Leb{n}$-negligible in general]
It is important to notice that, by \cref{result:Sobolev_is_Caccioppoli}, we have
\begin{equation*}
P_\alpha(E;\Omega)<+\infty
\implies
\Leb{n}(\Omega\cap\redb^\alpha E)>0
\end{equation*}
including even the case $\chi_E\in BV(\mathbb{R}^n)$. This shows a substantial difference between the standard \emph{local} De Giorgi's perimeter measure $|D\chi_E|$ and the \emph{non-local} fractional De Giorgi's perimeter measure $|D^\alpha\chi_E|$: the former is supported on a $\Leb{n}$-negligible set contained in the topological boundary of~$E$, while the latter, in general, can be supported on a set of positive Lebesgue measure and, for this reason, cannot be expected to be contained in the topological boundary of~$E$.
\end{remark}
\begin{remark}[Fractional reduced boundary and precise representative]
We let
\begin{equation*}
u^{*}(x) :=
\begin{cases}
\displaystyle \lim_{r \to 0} \frac{1}{|B_{r}(x)|} \int_{B_{r}(x)} u(y) \, dy & \text{if the limit exists and is finite}, \\[5mm]
0 & \text{otherwise},
\end{cases}
\end{equation*}
be the \emph{precise representative} of a function $u \in L^{1}_{\loc}(\mathbb{R}^{n}; \mathbb{R}^{m})$. Note that $u^{*}$ is well defined at any Lebesgue point of~$u$.
By \cref{result:Sobolev_is_Caccioppoli}, if $P_{\alpha}(E; \Omega) < +\infty$ then $D^{\alpha} \chi_{E} = \nabla^{\alpha} \chi_{E} \Leb{n}$ with $\nabla^{\alpha} \chi_{E} \in L^{1}(\Omega; \mathbb{R}^{n})$. Therefore the set
\begin{equation*}
\mathcal{R}^\alpha _\Omega E :=\set*{x \in \Omega : |(\nabla^{\alpha} \chi_{E})^{*}(x)| = |\nabla^{\alpha} \chi_{E}|^{*}(x) \neq 0}
\end{equation*}
is such that
\begin{equation} \label{eq:redb_alpha_Lebesgue_inclusion}
\mathcal{R}^\alpha _\Omega E \subset \Omega \cap \redb^{\alpha} E
\end{equation}
and
\begin{equation*}
\nu_{E}^{\alpha}(x) = \frac{(\nabla^{\alpha} \chi_{E})^{*}(x)}{|\nabla^{\alpha} \chi_{E}|^{*}(x)}
\qquad
\text{for all $x\in \mathcal{R}^\alpha _\Omega E$}.
\end{equation*}
\end{remark}
The following simple example shows that the inclusion in~\eqref{eq:redb_alpha_Lebesgue_inclusion} and the inequality in \eqref{eq:FFPS_is_Caccioppoli_estim} can be strict.
\begin{example}\label{example:interval_a_b}
Let $n = 1$, $\alpha \in (0, 1)$ and $a, b \in \mathbb{R}$, with $a < b$. It is easy to see that $\chi_{(a, b)} \in W^{\alpha, 1}(\mathbb{R})$. By~\eqref{eq:FFPS_is_Caccioppoli_repres_formula}, for any $x \neq a, b$ we have that
\begin{align*}
\nabla^{\alpha} \chi_{(a, b)}(x) & = \frac{\mu_{1, \alpha}}{\alpha} \int_{\mathbb{R}} \frac{1}{|x - y|^{\alpha}} \, d \left (\delta_{a} - \delta_{b} \right )(y) \\
& = \frac{2^{\alpha}} {\alpha \sqrt{\pi}} \frac{\Gamma\left ( 1 + \frac{\alpha}{2} \right )}{\Gamma\left ( \frac{1 - \alpha}{2} \right )} \left ( \frac{1}{|x - a|^{\alpha}} - \frac{1}{|x - b|^{\alpha}} \right ).
\end{align*}
We claim that
\begin{equation}\label{eq:frac_redb_a_b}
\redb^{\alpha} (a, b) = \mathbb{R} \setminus\set*{\frac{a + b}{2}}
\end{equation}
while
\begin{equation}\label{eq:special_frac_bound_a_b}
\mathcal{R}^\alpha_{\mathbb{R}} (a,b) = \mathbb{R} \setminus \set*{ a, \frac{a + b}{2}, b},
\end{equation}
so that inclusion~\eqref{eq:redb_alpha_Lebesgue_inclusion} is strict. Finally, we also claim that
\begin{equation}\label{eq:strict_ineq_one_dim}
\|\nabla^{\alpha} \chi_{(a, b)}\|_{L^{1}(\mathbb{R})} < \mu_{1, \alpha} P_{\alpha}((a, b)).
\end{equation}
Indeed, notice that
\begin{equation*}
\nabla^{\alpha} \chi_{(a, b)}(x) \ge 0
\end{equation*}
if and only if $x \le \frac{a + b}{2}$, so that
\begin{equation*}
\lim_{r \to 0} \frac{\displaystyle\int_{x - r}^{x + r} \nabla^{\alpha} \chi_{(a, b)}(y) \, dy}{\displaystyle\int_{x - r}^{x + r} |\nabla^{\alpha} \chi_{(a, b)}(y)| \, dy}
=
\begin{cases}
1 & \text{if $x < \displaystyle\frac{a + b}{2}$},\\[5mm]
-1 & \text{if $x > \displaystyle\frac{a + b}{2}$}.
\end{cases}
\end{equation*}
If $x = \frac{a + b}{2}$, then
\begin{equation*}
\int_{\frac{a + b}{2} - r}^{\frac{a + b}{2} + r} \nabla^{\alpha} \chi_{(a, b)}(y) \, dy = 0
\qquad
\forall r>0,
\end{equation*}
and claim~\eqref{eq:frac_redb_a_b} follows. In particular, we have
\begin{equation*}
\nu_{(a, b)}^{\alpha}(x) =
\begin{cases}
1 & \text{if} \ \ x < \displaystyle\frac{a + b}{2}, \\[5mm]
- 1 & \text{if} \ \ x > \displaystyle\frac{a + b}{2}.
\end{cases}
\end{equation*}
On the other hand, it is clear that
\begin{equation*}
\lim_{r \to 0} \frac{1}{2r} \int_{a - r}^{a + r} \nabla^{\alpha} \chi_{(a, b)}(y) \, dy = + \infty
\end{equation*}
and
\begin{equation*}
\lim_{r \to 0} \frac{1}{2r} \int_{b - r}^{b + r} \nabla^{\alpha} \chi_{(a, b)}(y) \, dy = - \infty,
\end{equation*}
so that claim~\eqref{eq:special_frac_bound_a_b} follows. To prove~\eqref{eq:strict_ineq_one_dim}, note that
\begin{equation} \label{eq:frac_per_a_b}
P_{\alpha}((a, b)) = \frac{4}{\alpha (1 - \alpha)} (b - a)^{1 - \alpha}
\end{equation}
since $P_{\alpha}((a, b)) = (b - a)^{1 - \alpha} P_{\alpha}((0, 1))$ by the scaling property of the fractional perimeter and
\begin{align*}
P_{\alpha}((0, 1)) & = 2 \int_{\mathbb{R} \setminus (0, 1)} \int_{0}^{1} \frac{1}{|y - x|^{1 + \alpha}} \, dy \, dx \\
& = \frac{2}{\alpha} \int_{\mathbb{R} \setminus (0, 1)} \left [ \frac{\sgn(x - y)}{|y - x|^\alpha} \right ]_{y=0}^{y=1} \, dx \\
& = \frac{2}{\alpha} \int_{\mathbb{R} \setminus (0, 1)} \frac{\sgn(x - 1)}{|1 - x|^{\alpha}} - \frac{\sgn(x)}{|x|^{\alpha}} \, dx \\
& = \frac{2}{\alpha} \int_{1}^{\infty} \frac{1}{(x - 1)^{\alpha}} - \frac{1}{x^{\alpha}} \, dx + \frac{2}{\alpha}\int_{- \infty}^{0} \frac{1}{(- x)^{\alpha}} - \frac{1}{(1 - x)^{\alpha}} \, dx \\
& = \frac{4}{\alpha} \int_{0}^{\infty} \frac{1}{x^{\alpha}} - \frac{1}{(1 + x)^{\alpha}} \, dx = \frac{4}{\alpha (1 - \alpha)}.
\end{align*}
On the other hand, we have
\begin{equation} \label{eq:L_1_nabla_alpha_chi_a_b}
\|\nabla^{\alpha} \chi_{(a, b)}\|_{L^{1}(\mathbb{R})}
= \frac{2^{1 + \alpha} \mu_{1, \alpha}}{\alpha (1 - \alpha)} (b - a)^{1 - \alpha}.
\end{equation}
Indeed, $\|\nabla^{\alpha} \chi_{(a, b)}\|_{L^{1}(\mathbb{R})} = (b - a)^{1 - \alpha} \|\nabla^{\alpha} \chi_{(0, 1)}\|_{L^{1}(\mathbb{R})}$ by~\eqref{scaling_prop_eq} and
\begin{align*}
\frac{\alpha}{\mu_{1, \alpha}} \|\nabla^{\alpha} \chi_{(0, 1)}\|_{L^{1}(\mathbb{R})} & = \int_{\mathbb{R}} \left | \frac{1}{|x|^{\alpha}} - \frac{1}{|x - 1|^{\alpha}} \right | \, dx \\
& = \int_{1}^{\infty} \left | \frac{1}{x^{\alpha}} - \frac{1}{(x - 1)^{\alpha}} \right | \, dx + \int_{0}^{1} \left | \frac{1}{x^{\alpha}} - \frac{1}{(1 - x)^{\alpha}} \right | \, dx \\
&\quad + \int_{- \infty}^{0} \left | \frac{1}{(-x)^{\alpha}} - \frac{1}{(1 - x)^{\alpha}} \right | \, dx \\
& = \int_{1}^{\infty} \frac{1}{(x - 1)^{\alpha}} - \frac{1}{x^{\alpha}} \, dx + \int_{\frac{1}{2}}^{1} \frac{1}{(1 - x)^{\alpha}} - \frac{1}{x^{\alpha}} \, dx\\
& \quad + \int_{0}^{\frac{1}{2}} \frac{1}{x^{\alpha}} - \frac{1}{(1 - x)^{\alpha}} \, dx + \int_{- \infty}^{0} \frac{1}{(-x)^{\alpha}} - \frac{1}{(1 - x)^{\alpha}} \, dx \\
& = 2 \int_{0}^{\infty} \frac{1}{x^{\alpha}} - \frac{1}{(1 + x)^{\alpha}} \, dx + 2 \int_{0}^{\frac{1}{2}} \frac{1}{x^{\alpha}} - \frac{1}{(1 - x)^{\alpha}} \, dx \\
& = \frac{2}{1 - \alpha} \left ( 1 + 2^{\alpha - 1} + 2^{\alpha - 1} - 1 \right ) = \frac{2^{1 + \alpha}}{1 - \alpha}.
\end{align*}
Combining~\eqref{eq:frac_per_a_b} and~\eqref{eq:L_1_nabla_alpha_chi_a_b}, we get \eqref{eq:strict_ineq_one_dim}.
\end{example}
Thanks to \cref{example:interval_a_b} above, we know that inequality~\eqref{eq:D_alpha_chi_E_P_alpha_intro} is strict for $E=(a,b)$ with $a,b\in\mathbb{R}$, $a<b$. We conclude this section proving that this fact holds for all sets $E\subset\mathbb{R}$ such that $\chi_E\in W^{\alpha,1}(\mathbb{R})$.
\begin{proposition}
Let $\alpha\in(0,1)$. If $\chi_E\in W^{\alpha,1}(\mathbb{R})$, then $|D^\alpha\chi_E|(\mathbb{R})<\mu_{1,\alpha}P_\alpha(E)$.
\end{proposition}
\begin{proof}
We argue by contradiction. Assume $\chi_E\in W^{\alpha,1}(\mathbb{R})$ is such that $|D^\alpha\chi_E|(\mathbb{R})=\mu_{1,\alpha}P_\alpha(E)$. Then
\begin{equation}\label{eq:equality_absurd_strict_ineq}
\int_{\mathbb{R}}\int_{\mathbb{R}} \abs*{f_E(x,y)}\,dy\,dx
=\int_{\mathbb{R}}\abs*{\int_{\mathbb{R}} f_E(x,y)\sgn(y-x)\,dy}\,dx
\end{equation}
where
\begin{equation*}
f_E(x,y):=\frac{\chi_E(y)-\chi_E(x)}{|y-x|^{1+\alpha}} \qquad\forall x,y\in\mathbb{R},\ x\ne y.
\end{equation*}
From~\eqref{eq:equality_absurd_strict_ineq} we deduce that
\begin{equation*}
\int_{\mathbb{R}} \abs*{f_E(x,y)}\,dy
=\abs*{\int_{\mathbb{R}} f_E(x,y)\sgn(y-x)\,dy}
\end{equation*}
for a.e.\ $x\in\mathbb{R}$. If $x\in E$, then $f_E(x,y)\le0$ for all $y\in\mathbb{R}$, $y\ne x$, and thus
\begin{equation*}
\int_x^{+\infty}|f_E(x,y)|\,dy
+\int_{-\infty}^x|f_E(x,y)|\,dy
=
\abs*{\int_x^{+\infty}|f_E(x,y)|\,dy
-\int_{-\infty}^x |f_E(x,y)|\,dy}
\end{equation*}
for a.e.\ $x\in E$. Squaring both sides and simplifying, we get that
\begin{equation*}
\left(\int_x^{+\infty}|f_E(x,y)|\,dy\right)
\left(\int_{-\infty}^x|f_E(x,y)|\,dy\right)=0,
\end{equation*}
so that either $|E^c\cap(x,+\infty)|=0$ or $|E^c\cap(-\infty,x)|=0$ for a.e.\ $x\in E$, contradicting the fact that $|E|<+\infty$.
\end{proof}
\section{Existence of blow-ups for fractional Caccioppoli sets}
\label{sec:blow-ups}
In this section we prove existence of blow-ups for sets with locally finite fractional Caccioppoli $\alpha$-perimeter. We follow the approach presented in~\cite{EG15}*{Section~5.7}.
We start with the following technical preliminary result.
\begin{lemma}\label{lem:frac_grad_cutoff_ball}
Let $\alpha \in (0, 1)$. For all $\varepsilon, r > 0$ and $x \in \mathbb{R}^{n}$ we define
\begin{equation*}
h_{\varepsilon, r, x}(y) :=
\begin{cases} 1 & \text{if} \ \ 0 \le |y - x| \le r, \\[3mm]
\dfrac{r + \varepsilon - |y - x|}{\varepsilon} & \text{if} \ \ r < |y - x| < r + \varepsilon, \\[4mm]
0 & \text{if} \ \ |y -x| \ge r + \varepsilon.
\end{cases}
\end{equation*}
Then $\nabla^\alpha h_{\varepsilon,r,x}\in L^1(\mathbb{R}^n;\mathbb{R}^n)$ with
\begin{equation} \label{eq:frac_grad_cutoff_ball}
\nabla^{\alpha} h_{\varepsilon, r, x}(y)
= \frac{\mu_{n, \alpha}}{\varepsilon (n + \alpha - 1)} \int_{B_{r + \varepsilon}(x) \setminus B_r(x)} \frac{x - z}{|x - z|}|z - y|^{1 - n - \alpha} \, dz
\end{equation}
for $\Leb{n}$-a.e. $y \in \mathbb{R}^{n}$.
\end{lemma}
\begin{proof}
Clearly $h_{\varepsilon, r, x} \in\Lip_c(\mathbb{R}^{n})$ and
\begin{equation*}
\nabla h_{\varepsilon, r, x}(y) = - \frac{1}{\varepsilon} \frac{y - x}{|y - x|}\, \chi_{B_{r + \varepsilon}(x) \setminus B_r(x)}(y).
\end{equation*}
Therefore by~\eqref{eq:weak_frac_grad_repr} we get
\begin{equation*}
\nabla^{\alpha} h_{\varepsilon, r, x}(y)
= - \frac{1}{\varepsilon}\frac{\mu_{n, \alpha}}{n + \alpha - 1} \int_{\mathbb{R}^{n}} \frac{1}{|z - y|^{n + \alpha - 1}} \frac{z - x}{|z - x|} \chi_{B_{r+\varepsilon}(x) \setminus B_r(x)}(z) \, dz
\end{equation*}
for $\Leb{n}$-a.e.\ $y \in \mathbb{R}^{n}$. By \cref{result:Sobolev_subset_BV}, we get $\nabla^\alpha h_{\varepsilon,r,x}\in L^1(\mathbb{R}^n;\mathbb{R}^n)$.
\end{proof}
We now proceed with the following formula for integration by parts on balls, see~\cite{EG15}*{Lemma~5.2} for the analogous result in the classical setting.
\begin{theorem}[Integration by parts on balls] \label{th:int_by_parts_E_ball}
Let $\alpha \in (0, 1)$. If $E\subset\mathbb{R}^n$ is a set with locally finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$, then
\begin{equation} \label{eq:int_by_parts_E_ball}
\int_{E \cap B_r(x)} \div^{\alpha} \phi \, dy
+ \int_{E} \phi\, \cdot \nabla^{\alpha} \chi_{B_r(x)}\, dy
+ \int_{E} \div^{\alpha}_{\rm NL} (\chi_{B_r(x)}, \phi) \, dy
=-\int_{B_r(x)} \phi \cdot \, d D^{\alpha} \chi_{E}
\end{equation}
for all $\phi \in\Lip_c(\mathbb{R}^{n}; \mathbb{R}^{n})$, $x \in \redb^{\alpha} E$ and for $\Leb{1}$-a.e.\ $r > 0$.
\end{theorem}
\begin{proof}
Fix $\varepsilon, r > 0$, $x \in \redb^{\alpha} E$ and $\phi \in \Lip_{c}(\mathbb{R}^{n}; \mathbb{R}^{n})$ and let $h_{\varepsilon, r, x}$ be as in \cref{lem:frac_grad_cutoff_ball}. On the one hand, by~\eqref{eq:Caccioppoli_Gauss-Green} we have
\begin{equation} \label{eq:IBP_E_1}
\int_{E} \div^{\alpha}(\phi\, h_{\varepsilon, r, x}) \, dy
= - \int_{\redb^{\alpha} E} (h_{\varepsilon, r, x}\,\phi) \cdot \, d D^{\alpha} \chi_{E}.
\end{equation}
Since $h_{\varepsilon, r, x}(y) \to \chi_{\closure{B_r(x)}}(y)$ as $\varepsilon \to 0$ for any $y \in \mathbb{R}^{n}$ and $|D^{\alpha} \chi_{E}|(\partial B_r(x)) = 0$ for $\Leb{1}$-a.e.\ $r > 0$, we can compute
\begin{equation*}
\lim_{\varepsilon\to0}\int_{\redb^{\alpha} E} (h_{\varepsilon, r, x}\,\phi) \cdot \, d D^{\alpha} \chi_{E}
=\int_{B_r(x)} \phi \cdot \, d D^{\alpha} \chi_{E}.
\end{equation*}
On the other hand, by \cref{result:Sobolev_frac_enough} and \cref{lem:Leibniz_frac_div}, we have
\begin{equation} \label{eq:Leibniz_rule_h_phi}
\div^{\alpha}(\phi \, h_{\varepsilon, r, x})
= h_{\varepsilon, r, x}\, \div^{\alpha}\phi
+ \phi \cdot \nabla^{\alpha} h_{\varepsilon, r, x}
+ \div_{\rm NL}^{\alpha}(h_{\varepsilon, r, x}, \phi).
\end{equation}
We deal with each term of the right-hand side of~\eqref{eq:Leibniz_rule_h_phi} separately. For the first term, since $0\le h_{\varepsilon, r, x}\le\chi_{B_{r+1}(x)}$ for all $\varepsilon\in(0,1)$ and $h_{\varepsilon, r, x} \to \chi_{B_r(x)}$ in $L^{1}(\mathbb{R}^{n})$ as $\varepsilon \to 0$, by \cref{prop:frac_div_repr} and Lebesgue's Dominated Convergence Theorem we can compute
\begin{equation} \label{eq:IBP_conv_1}
\lim_{\varepsilon\to0}\int_{E} h_{\varepsilon, r, x} \,\div^{\alpha}\phi \, dy
=\int_{E \cap B_r(x)} \div^{\alpha}\phi \, dy.
\end{equation}
For the second term, by~\eqref{eq:frac_grad_cutoff_ball} we have
\begin{equation*}
\int_{E} \phi(y) \cdot \nabla^{\alpha} h_{\varepsilon, r, x}(y) \, dy
= \frac{\mu_{n, \alpha}}{\varepsilon (n + \alpha - 1)} \int_{E} \phi(y) \cdot \int_{B_{r+\varepsilon}(x) \setminus B_r(x)} \frac{x - z}{|x - z|} |z - y|^{1 - n - \alpha}\, dz \, dy.
\end{equation*}
By Fubini's Theorem, we can compute
\begin{align*}
\int_E & \phi(y)\cdot \int_{B_{r+\varepsilon}(x) \setminus B_r(x)} \frac{x - z}{|x - z|} |z - y|^{1 - n - \alpha}\, dz \, dy \\
& = \int_{B_{r+\varepsilon}(x) \setminus B_r(x)}\frac{x - z}{|x - z|}\cdot \int_E \phi(y)\,|z - y|^{1 - n - \alpha} \, dy \, dz \\
& = \int_{r}^{r + \varepsilon} \int_{\partial B_{\rho}(x)} \frac{x - z}{|x - z|}\cdot \int_E \phi(y)\,|z - y|^{1 - n - \alpha} \, dy \, d \Haus{n - 1}(z) \, d \rho.
\end{align*}
By Lebesgue's Differentiation Theorem, we have
\begin{equation*}
\begin{split}
\lim_{\varepsilon\to0}\frac{1}{\varepsilon}\int_E & \phi(y)\cdot \int_{B_{r+\varepsilon}(x) \setminus B_r(x)} \frac{x - z}{|x - z|}\, |z - y|^{1 - n - \alpha}\, dz \, dy\\
&=\lim_{\varepsilon\to0}\frac{1}{\varepsilon}\int_{r}^{r + \varepsilon} \int_{\partial B_{\rho}(x)} \frac{x - z}{|x - z|}\cdot \int_E \phi(y)\,|z - y|^{1 - n - \alpha} \, dy \, d \Haus{n - 1}(z) \, d \rho\\
&=\int_{\partial B_r(x)} \frac{x - z}{|x - z|}\cdot \int_E \phi(y)\,|z - y|^{1 - n - \alpha} \, dy \, d \Haus{n - 1}(z)\\
&=\int_E \phi(y)\cdot\int_{\partial B_r(x)} \frac{x - z}{|x - z|}\,|z - y|^{1 - n - \alpha}\, d \Haus{n - 1}(z)\, dy\\
&=\int_E \phi(y)\cdot\int_{\mathbb{R}^n} |z - y|^{1 - n - \alpha}\, dD\chi_{B_r(x)}(z)\, dy
\end{split}
\end{equation*}
for $\Leb{1}$-a.e.\ $r > 0$. Therefore, by~\eqref{eq:weak_frac_grad_repr}, we get that
\begin{equation} \label{eq:IBP_conv_3}
\begin{split}
\lim_{\varepsilon\to0}\int_{E} & \phi \cdot \nabla^{\alpha} h_{\varepsilon, r, x} \, dy \\
& = \frac{\mu_{n, \alpha}}{n + \alpha - 1} \int_E\phi(y) \cdot\int_{\mathbb{R}^{n}} |z - y|^{1 - n - \alpha} \, d D \chi_{B_{r}(x)}(z) \, dy \\
& = \int_E \,\phi \cdot \nabla^{\alpha} \chi_{B_{r}(x)} \, dy
\end{split}
\end{equation}
for $\Leb{1}$-a.e.\ $r > 0$.
Finally, for the third term, note that
\begin{equation*}
\abs*{\frac{(z - y) \cdot (\phi(z) - \phi(y)) (h_{\varepsilon, r, x}(z) - h_{\varepsilon, r, x}(y))}{|z - y|^{n + \alpha + 1}}}
\le 2 \frac{|\phi(z) - \phi(y)|}{|z - y|^{n + \alpha}} \in L^{1}_{z}(\mathbb{R}^{n})
\end{equation*}
for all $y\in\mathbb{R}^n$, so that
\begin{equation*}
\lim_{\varepsilon\to0}\div_{\rm NL}^{\alpha}(h_{\varepsilon, r, x}, \phi)(y)
= \div^{\alpha}_{\rm NL}(\chi_{B_r(x)}, \phi)(y)
\end{equation*}
for $\Leb{n}$-a.e.\ $y\in\mathbb{R}^n$ by Lebesgue's Dominated Convergence Theorem. Since
\begin{equation*}
\abs*{\div_{\rm NL}^{\alpha}(h_{\varepsilon, r, x}, \phi)(y)}
\le 2 \int_{\mathbb{R}^{n}}\frac{|\phi(z) - \phi(y)|}{|z - y|^{n + \alpha}}\,dz \in L^{1}_y(\mathbb{R}^{n}),
\end{equation*}
again by Lebesgue's Dominated Convergence Theorem we can compute
\begin{equation} \label{eq:IBP_conv_2}
\lim_{\varepsilon\to0}\int_{E} \div_{\rm NL}^{\alpha}(h_{\varepsilon, r, x}, \phi) \, dy
=\int_{E} \div^{\alpha}_{\rm NL}(\chi_{B_{r}(x)}, \phi) \, dy.
\end{equation}
Combining~\eqref{eq:IBP_E_1}, \eqref{eq:Leibniz_rule_h_phi}, \eqref{eq:IBP_conv_1}, \eqref{eq:IBP_conv_3} and~\eqref{eq:IBP_conv_2}, we obtain~\eqref{eq:int_by_parts_E_ball}.
\end{proof}
We can now deduce the following decay estimates for the fractional De Giorgi's perimeter measure, see~\cite{EG15}*{Lemma~5.3} for the analogous result in the classical setting.
\begin{theorem}[Decay estimates]\label{th:decay_D_alpha_E_B}
Let $\alpha \in (0, 1)$. There exist $A_{n, \alpha}, B_{n, \alpha} > 0$ with the following property. Let $E\subset\mathbb{R}^n$ be a set with locally finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$. For any $x \in \redb^{\alpha} E$, there exists $r_x > 0$ such that
\begin{equation} \label{eq:decay_D_alpha_E_B_1}
|D^{\alpha} \chi_{E}|(B_r(x)) \le A_{n, \alpha} r^{n - \alpha}
\end{equation}
and
\begin{equation} \label{eq:decay_D_alpha_E_B_2}
|D^{\alpha} \chi_{E \cap B_r(x)}|(\mathbb{R}^{n}) \le B_{n, \alpha} r^{n - \alpha}
\end{equation}
for all $r \in (0, r_x)$.
\end{theorem}
\begin{proof}
We divide the proof in two steps, dealing with the two estimates separately.
\smallskip
\textit{Step~1: proof of~\eqref{eq:decay_D_alpha_E_B_1}}.
Fix $x \in \redb^{\alpha} E$ and choose $\phi \in \Lip_{c}(\mathbb{R}^{n}; \mathbb{R}^{n})$ such that $\phi \equiv \nu^{\alpha}_{E}(x)$ in $B_1(x)$ and $\|\phi\|_{L^{\infty}(\mathbb{R}^{n}; \mathbb{R}^{n})} \le 1$. On the one hand, by \cref{def:redb_alpha}, there exists $r_x\in(0,1)$ such that
\begin{equation} \label{eq:decay_estimate_1}
\int_{B_r(x)} \phi\cdot\, d D^{\alpha} \chi_{E} \ge \frac{1}{2} |D^{\alpha} \chi_{E}|(B_r(x))
\end{equation}
for all $r\in(0,r_x)$. On the other hand, by~\eqref{eq:int_by_parts_E_ball} we have
\begin{equation}\label{eq:decay_estimate_1.5}
\begin{split}
\int_{B_r(x)} \phi\cdot\, d D^{\alpha} \chi_{E}
&\le \abs*{\int_{E \cap B_{r}(x)} \div^{\alpha} \phi \, dy}
+\abs*{ \int_{E} \phi\cdot \, d D^{\alpha} \chi_{B_r(x)} }\\
&\quad+\abs*{ \int_{E} \div^{\alpha}_{\rm NL} (\chi_{B_r(x)}, \phi) \, dy }
\end{split}
\end{equation}
for $\Leb{1}$-a.e.\ $r\in(0,r_x)$. We now estimate the three terms in the right-hand side separately. For the first one, since $\phi(y) \equiv \nu_{E}^{\alpha}(x)$ in $B_r(x)$, we can estimate
\begin{align*}
\left | \int_{E \cap B_r(x)} \div^{\alpha} \phi(y) \, dy \right |
& \le \mu_{n, \alpha} \int_{E \cap B_r(x)} \int_{\mathbb{R}^{n}} \frac{|\phi(z) - \phi(y)|}{|z - y|^{n + \alpha}} \, dz \, d y \\
& = \mu_{n, \alpha} \int_{E \cap B_r(x)} \int_{\mathbb{R}^{n} \setminus B_r(x)} \frac{|\phi(z) - \nu_{E}^{\alpha}(x)|}{|z - y|^{n + \alpha}} \, dz \, d y \\
& \le 2 \mu_{n, \alpha} \int_{B_r(x)} \int_{\mathbb{R}^{n} \setminus B_r(x)} \frac{1}{|z - y|^{n + \alpha}} \, dz \, d y \\
& = 2 \mu_{n, \alpha} P_{\alpha}(B_r(x))
\end{align*}
so that
\begin{equation} \label{eq:decay_estimate_2}
\left | \int_{E \cap B_r(x)} \div^{\alpha} \phi(y) \, dy \right |
\le 2 \mu_{n, \alpha} P_{\alpha}(B_1)\, r^{n - \alpha}.
\end{equation}
For the second term, by \cref{result:Sobolev_is_Caccioppoli} we can estimate
\begin{equation} \label{eq:decay_estimate_3}
\left | \int_{E} \phi \cdot \, d D^{\alpha} \chi_{B_{r}(x)} \right |
\le |D^{\alpha} \chi_{B_r(x)}|(\mathbb{R}^{n})
\le \mu_{n, \alpha} P_{\alpha}(B_r(x))
= \mu_{n, \alpha} P_{\alpha}(B_1)\, r^{n - \alpha}.
\end{equation}
Finally, by \cref{lem:Leibniz_frac_div}, we can estimate
\begin{equation*}
\begin{split}
\abs*{\int_{E} \div^{\alpha}_{\rm NL}(\chi_{B_r(x)}, \phi) \, dy}
&\le\| \div^{\alpha}_{\rm NL}(\chi_{B_r(x)}, \phi)\|_{L^{1}(\mathbb{R}^{n})} \\
&\le 2 \mu_{n, \alpha} [\chi_{B_r(x)}]_{W^{\alpha, 1}(\mathbb{R}^{n})}\\
&=2 \mu_{n, \alpha}P_{\alpha}(B_r(x))
\end{split}
\end{equation*}
so that
\begin{equation} \label{eq:decay_estimate_4}
\abs*{\int_{E} \div^{\alpha}_{\rm NL}(\chi_{B_r(x)}, \phi) \, dy}
\le 2 \mu_{n, \alpha} P_{\alpha}(B_1)\, r^{n - \alpha}.
\end{equation}
Combining~\eqref{eq:decay_estimate_1}, \eqref{eq:decay_estimate_1.5}, \eqref{eq:decay_estimate_2}, \eqref{eq:decay_estimate_3} and \eqref{eq:decay_estimate_4}, we conclude that
\begin{equation} \label{eq:decay_estimate_5}
|D^{\alpha} \chi_{E}|(B_r(x))
\le 10 \mu_{n, \alpha} P_{\alpha}(B_1) \, r^{n - \alpha}
\end{equation}
for $\Leb{1}$-a.e.\ $r\in(0,r_x)$. Hence~\eqref{eq:decay_D_alpha_E_B_1} follows with $A_{n, \alpha} = 10 \mu_{n, \alpha} P_{\alpha}(B_1)$ for all $r\in(0,r_x)$ by a simple continuity argument.
\smallskip
\textit{Step~2: proof of~\eqref{eq:decay_D_alpha_E_B_2}}.
Fix $x \in \redb^{\alpha} E$ and $\phi \in \Lip_{c}(\mathbb{R}^{n}; \mathbb{R}^{n})$ with $\|\phi\|_{L^{\infty}(\mathbb{R}^{n}; \mathbb{R}^{n})} \le 1$. Again by~\eqref{eq:int_by_parts_E_ball} we can estimate
\begin{equation*}
\abs*{\int_{E \cap B_r(x)} \div^{\alpha} \phi \, dy}
\le |D^{\alpha} \chi_{E}|(B_r(x)) + |D^{\alpha} \chi_{B_r(x)}|(\mathbb{R}^n)
+ \int_{E} |\div^{\alpha}_{\rm NL}(\chi_{B_r(x)}, \phi)| \, dy
\end{equation*}
for $\Leb{1}$-a.e.\ $r\in(0,r_x)$. Using~\eqref{eq:decay_estimate_3}, \eqref{eq:decay_estimate_4} and~\eqref{eq:decay_estimate_5}, we conclude that
\begin{equation*}
|D^{\alpha} \chi_{E \cap B_r(x)}|(\mathbb{R}^{n}) \le 13 \mu_{n, \alpha} P_{\alpha}(B_1) \, r^{n - \alpha}
\end{equation*}
for $\Leb{1}$-a.e.\ $r\in(0,r_x)$. Hence~\eqref{eq:decay_D_alpha_E_B_2} follows with $B_{n, \alpha} = 13 \mu_{n, \alpha} P_{\alpha}(B_1)$ for all $r\in(0,r_x)$ by a simple continuity argument. This concludes the proof.
\end{proof}
As an easy consequence of \cref{th:decay_D_alpha_E_B}, we can prove that
\begin{equation*}
|D^{\alpha} \chi_{E}|\ll\Haus{n-\alpha}\res\redb^\alpha E
\end{equation*}
for any set~$E$ with locally finite fractional Caccioppoli $\alpha$-perimeter in $\mathbb{R}^{n}$.
\begin{corollary}[$|D^{\alpha} \chi_{E}|\ll\Haus{n-\alpha}\res\redb^\alpha E$] \label{cor:abs_continuity}
Let $\alpha \in (0, 1)$. If $E$ is a set with locally finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^{n}$, then
\begin{equation} \label{eq:abs_cont_estimate}
|D^{\alpha} \chi_{E}| \le 2^{n - \alpha} \frac{A_{n, \alpha}}{\omega_{n - \alpha}} \Haus{n - \alpha} \res \redb^{\alpha} E,
\end{equation}
where $A_{n, \alpha}$ is as in~\eqref{eq:decay_D_alpha_E_B_1}.
\end{corollary}
\begin{proof}
By \eqref{eq:decay_D_alpha_E_B_1}, we have that
\begin{equation*}
\Theta^{*}_{n - \alpha}(|D^{\alpha} \chi_{E}|, x) := \limsup_{r \to 0} \frac{|D^{\alpha} \chi_{E}|(B_{r}(x))}{\omega_{n - \alpha} r^{n - \alpha}} \le \frac{A_{n, \alpha}}{\omega_{n - \alpha}}
\end{equation*}
for any $x \in \redb^{\alpha} E$.
Therefore, \eqref{eq:abs_cont_estimate} is a simple application of \cite{AFP00}*{Theorem 2.56}.
\end{proof}
For any set~$E$ of locally finite fractional Caccioppoli $\alpha$-perimeter, \cref{cor:abs_continuity} enables us to obtain a lower bound on the Hausdorff dimension of~$\redb^{\alpha}E$.
\begin{proposition}
Let $\alpha \in (0, 1)$. If~$E$ is a set with locally finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^{n}$, then
\begin{equation} \label{eq:redb_alpha_dim_estimate}
\dim_{\Haus{}}(\redb^{\alpha} E) \ge n - \alpha.
\end{equation}
\end{proposition}
\begin{proof}
Since $|D^{\alpha} \chi_{E}|(\redb^{\alpha} E) > 0$ by \cref{def:redb_alpha}, by \cref{cor:abs_continuity} we conclude that $\Haus{n - \alpha}(\redb^{\alpha} E) > 0$, proving~\eqref{eq:redb_alpha_dim_estimate}.
\end{proof}
As another interesting consequence of \cref{cor:abs_continuity}, we are able to prove that assumption~\eqref{eq:coarea_int_finite} in~\cref{th:coarea_inequality} cannot be dropped.
\begin{corollary}[No coarea formula in $BV^{\alpha}(\mathbb{R})$]
Let $\alpha\in(0,1)$. There exist $f\in BV^\alpha(\mathbb{R}^n)$ such that
\begin{equation}\label{eq:coarea_explosion}
\int_{\mathbb{R}} |D^{\alpha} \chi_{\{ f > t \}}|(\mathbb{R}^n) \, dt=+\infty.
\end{equation}
\end{corollary}
\begin{proof}
Let $E\subset \mathbb{R}^n$ be such that $\chi_E\in BV(\mathbb{R}^n)$ and consider $f:=(-\Delta)^{\frac{1-\alpha}{2}}\chi_E$. By \cref{result:correspondence_BV_alpha_bv}, we know that $f\in BV^\alpha(\mathbb{R}^n)$ with $|D^\alpha f|=|D\chi_E|=\Haus{n-1}\res\redb E$. If
\begin{equation*}
\int_{\mathbb{R}} |D^{\alpha} \chi_{\{ f > t \}}|(\mathbb{R}^n) \, dt<+\infty
\end{equation*}
then
\begin{equation*}
|D^\alpha f|\le \int_{\mathbb{R}} |D^{\alpha} \chi_{\{ f > t \}}| \, dt
\end{equation*}
by \cref{th:coarea_inequality}. Thus $|D^{\alpha} f| \ll \Haus{n - \alpha}$ by \cref{cor:abs_continuity}, so that $\Haus{n-1}(\redb E)=0$, which is clearly absurd.
\end{proof}
\begin{remark}
If $f\in W^{\alpha,1}(\mathbb{R}^n)$, then
\begin{equation*}
\int_{\mathbb{R}}|D^\alpha\chi_{\set{f>t}}|(\mathbb{R}^n)\,dt
\le\mu_{n,\alpha}\int_{\mathbb{R}}P_\alpha(\set{f>t})\,dt
=\mu_{n,\alpha}[f]_{W^{\alpha,1}(\mathbb{R}^n)}<+\infty
\end{equation*}
by \cref{result:Sobolev_is_Caccioppoli} and Tonelli's Theorem, so that~\eqref{eq:coarea_explosion} does not hold for all $f\in BV^\alpha(\mathbb{R}^n)$. We do not know if~\eqref{eq:tot_var_coarea_inequality} is an equality for some functions $f\in BV^\alpha(\mathbb{R}^n)$.
\end{remark}
We can now prove the existence of blow-ups for sets with locally finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$, see~\cite{EG15}*{Theorem~5.13} for the analogous result in the classical setting. Here and in the following, given a set~$E$ with locally finite fractional Caccioppoli $\alpha$-perimeter and $x\in\redb^\alpha E$, we let $\tang(E,x)$ be the set of all \emph{tangent sets of~$E$ at~$x$}, i.e.\ the set of all limit points in $L^1_{\loc}(\mathbb{R}^n)$-topology of the family $\set*{\frac{E - x}{r} : r>0}$ as $r\to0$.
\begin{theorem}[Existence of blow-up] \label{th:blow_up}
Let $\alpha \in (0, 1)$. Let $E$ be a set with locally finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$. For any $x \in \redb^{\alpha} E$ we have $\tang(E,x)\ne\varnothing$.
\end{theorem}
\begin{proof}
Fix $x \in \redb^{\alpha} E$. Up to a translation, we can assume $x = 0$. We set $E_{r} :=E/r= \set*{ y \in \mathbb{R}^{n} : r y \in E }$ for all $r>0$. We divide the proof in two steps.
\smallskip
\textit{Step~1}. For each $p\in\mathbb{N}$, we define $D_{r}^p := E_{r} \cap B_p$. By the $\alpha$-homogeneity of $\div^\alpha$, we have
\begin{align*}
\int_{D_{r}^p} \div^{\alpha} \phi \, dy
= r^{-n}\int_{E \cap B_{rp}} (\div^{\alpha} \phi)( r^{-1}z) \,dz
= r^{\alpha - n}\int_{E \cap B_{rp}} \div^{\alpha} (\phi(r^{-1}\cdot)) \, dz
\end{align*}
for all $\phi \in C^{\infty}_{c}(\mathbb{R}^{n}; \mathbb{R}^{n})$. By~\eqref{eq:decay_D_alpha_E_B_2}, we thus get
\begin{equation*}
|D^\alpha\chi_{D_r^p}|(\mathbb{R}^n)
=r^{\alpha - n}|D^\alpha\chi_{E\cap B_{rp}}|(\mathbb{R}^n)
\le B_{n,\alpha}p^{n-\alpha}
\end{equation*}
for all $r>0$ such that $rp<r_0$. Hence, for each fixed $p\in\mathbb{N}$, we have
\begin{equation*}
\sup_{r<r_0/p}|D^{\alpha} \chi_{D_{r}^p}|(\mathbb{R}^{n}) \le B_{n,\alpha}p^{n-\alpha}.
\end{equation*}
\smallskip
\textit{Step~2}. Let $(r_k)_{k\in\mathbb{N}}$ be such that $r_k\to0$ as $k\to+\infty$ and let $E_k:=E_{r_k}$ and $D_k^p:=D_{r_k}^p$ for simplicity. By \textit{Step~1}, for each $p\in\mathbb{N}$ we know that
\begin{equation*}
\sup_{r_k<r_0/p}|D^{\alpha} \chi_{D_k^p}|(\mathbb{R}^{n}) \le B_{n,\alpha}p^{n-\alpha}
\quad\text{and}\quad
D_k^p\subset B_p
\quad \forall k\in\mathbb{N}.
\end{equation*}
Thanks to \cref{result:compactness_Caccioppoli}, by a standard diagonal argument we find a subsequence $(D_{k_j}^p)_{j\in\mathbb{N}}$ and a sequence $(F_p)_{p\in\mathbb{N}}$ of sets with finite fractional Caccioppoli $\alpha$-perimeter such that $\chi_{D_{k_j}^p}\to\chi_{F_p}$ in $L^1(\mathbb{R}^n)$ as $j\to+\infty$ for each $p\in\mathbb{N}$. Up to null sets, we have $F_p\subset F_{p+1}$, so that $\chi_{E_{k_j}}\to\chi_F$ in $L^1_{\loc}(\mathbb{R}^n)$, where $F:=\bigcup_{p\in\mathbb{N}} F_p$. We thus conclude that $F\in\tang(E,x)$.
\end{proof}
We now give a characterisation of the blow-ups of sets with locally finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$, see Claim~\#1 in the proof of~\cite{EG15}*{Theorem~5.13} for the result in the classical setting.
\begin{proposition}[Characterisation of blow-ups] \label{prop:char_blow_up}
Let $\alpha \in (0, 1)$. Let $E$ be a set with locally finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$ and let $x \in \redb^{\alpha} E$. If $F \in \tang(E,x)$, then $F$~is a set of locally finite fractional Caccioppoli $\alpha$-perimeter such that $\nu^{\alpha}_{F}(y)=\nu^\alpha_E(x)$ for $|D^\alpha\chi_F|$-a.e.\ $y\in\redb^\alpha F$.
\end{proposition}
\begin{proof}
As in the proof of \cref{th:blow_up}, we assume $x = 0$ and we set $E_{r} = E/r$. By \cref{th:blow_up}, there exists $(r_k)_{k\in\mathbb{N}}$ such that $r_{k} \to 0$ as $k\to+\infty$ and $\chi_{E_{r_{k}}} \to \chi_{F}$ in $L^{1}_{\loc}(\mathbb{R}^{n})$. By \cref{result:frac_var_meas_is_lsc}, it is clear that $F$ has locally finite fractional Caccioppoli $\alpha$-perimeter in~$\mathbb{R}^n$. By \eqref{eq:weak_conv}, we get
\begin{equation*}
D^{\alpha} \chi_{E_{r_{k}}} \rightharpoonup D^{\alpha} \chi_{F}
\quad
\text{in $\mathcal{M}_{\rm loc}(\mathbb{R}^{n}; \mathbb{R}^{n})$}
\end{equation*}
as $k\to+\infty$. Thus, for $\Leb{1}$-a.e.\ $L > 0$, we have
\begin{equation} \label{eq:weak_conv_blow_up_ball}
D^{\alpha} \chi_{E_{r_{k}}}(B_{L}) \to D^{\alpha} \chi_{F}(B_{L})
\quad
\text{as $k\to+\infty$}.
\end{equation}
Since
\begin{equation*}
D^{\alpha} \chi_{E_{r}}
= r^{\alpha - n} (\delta_{\frac{1}{r}})_{\#} D^{\alpha} \chi_{E}
\qquad
\forall r>0,
\end{equation*}
we have that
\begin{equation*}
|D^{\alpha} \chi_{E_{r_{k}}}| (B_{L})
= r_{k}^{\alpha - n} |D^{\alpha} \chi_{E}|(B_{r_{k} L})
\end{equation*}
and
\begin{equation*}
D^{\alpha} \chi_{E_{r_{k}}}(B_{L})
= r_{k}^{\alpha - n} D^{\alpha} \chi_{E}(B_{r_{k} L}).
\end{equation*}
Since $0\in\redb^\alpha E$, we thus get
\begin{equation}\label{eq:asymptotic_blow_up}
\lim_{k\to+\infty}\frac{D^{\alpha} \chi_{E_{r_{k}}} (B_{L})}{|D^{\alpha} \chi_{E_{r_{k}}}| (B_{L})}
=\lim_{k\to+\infty}\frac{D^{\alpha} \chi_{E}(B_{r_{k} L})}{|D^{\alpha} \chi_{E}|(B_{r_{k} L})}
=\nu_{E}^{\alpha}(0).
\end{equation}
Therefore, by \cref{result:frac_var_meas_is_lsc}, \eqref{eq:weak_conv_blow_up_ball} and~\eqref{eq:asymptotic_blow_up}, we obtain that
\begin{align*}
|D^{\alpha} \chi_{F}|(B_{L})
& \le \liminf_{k \to + \infty} |D^{\alpha} \chi_{E_{r_{k}}}|(B_{L}) \\
& = \lim_{k \to + \infty} \int_{B_{L}} \nu_{E}^{\alpha}(0) \cdot d D^{\alpha} \chi_{E_{r_{k}}} \\
& = \int_{B_{L}} \nu_{E}^{\alpha}(0) \cdot d D^{\alpha} \chi_{F} \\
& = \int_{B_{L}} \nu_{E}^{\alpha}(0) \cdot \nu_{F}^{\alpha} \ d |D^{\alpha} \chi_{F}| \\
&\le |D^{\alpha} \chi_{F}|(B_{L})
\end{align*}
for $\Leb{1}$-a.e.\ $L > 0$. We thus get that $\nu_{F}^{\alpha}(y) = \nu_{E}^{\alpha}(0)$ for $|D^\alpha\chi_F|$-a.e.\ $y \in B_{L} \cap \redb^{\alpha} F$ and $\Leb{1}$-a.e.\ $L > 0$, so that the conclusion follows.
\end{proof}
|
\section{Introduction}
\subsection{Brief overview}
The Alexander module of a knot $K$ supports a non-singular Hermitian~$\mathbb{Q}(t)/\mathbb{Z}[t^{\pm 1}]$-valued pairing $\operatorname{Bl}(K)$ known as the \emph{Blanchfield form}.
This pairing, which was introduced by Blanchfield~\cite{Blanchfield} in 1957, has since then been generalized in several different directions.
It has been studied for links~\cite{BorodzikFriedlPowell, Hillman, ConwayFriedlToffoli, ConwayBlanchfield, CochranOrr}, 3-manifolds~\cite{FriedlPowell, Powell} and with various twisted coefficients~\cite{CochranOrrTeichner, Leidy, CochranHarveyLeidy, Powell, MillerPowell, Nosaka}.
Arguably, a challenge when dealing with any such Blanchfield pairing consists in extracting tractable invariants from it.
The search for such invariants naturally leads us to study arbitrary linking forms over $\L{\F}$, where~$\mathbb{F}$ is either $\mathbb{R}$ or $\mathbb{C}$.
Here, given an integral domain $R$ with involution $x \mapsto x^\#$ and field of fractions $Q$, a \emph{linking form} is a sesquilinear Hermitian form $\lambda\colon M\times M\to Q/R$ on a torsion $R$-module $M$.
The goal of this article is twofold.
Firstly, we develop an algebraic machinery to extract signatures invariants from linking forms over~$R=\L{\F}$.
Secondly, we apply this theory to twisted Blanchfield pairings, with a view towards knot concordance.
Before providing a detailed overview of this paper, we briefly state some of our main results:
\begin{itemize}
\item Sections~\ref{sec:LinkingFormClassification} to~\ref{sec:Signatures} provide a concrete framework for working with linking forms over $\L{\F}$, including various definitions of signature functions.
Algebraic applications of this framework include classifications of linking forms over $\L{\F}$ up to isometry~(Theorem \ref{thm:MainLinkingForm}) and Witt equivalence (Theorems~\ref{thm:WittClassificationReal} and~\ref{thm:WittClassificationComplex}).
\item Sections~\ref{sec:TwistedHomology} and~\ref{sec:TwistedSign3ManifKnot} define twisted Blanchfield pairings and twisted signature functions for knots, generalizing the Levine-Tristram signature.
The definition is 3-dimensional and explicit computations can be made.
Basic properties of this function can be found in Proposition~\ref{prop:LocallyConstant}.
Obstructions to concordance invariance are provided in Proposition~\ref{prop:witt-equiv-concordance}.
Furthermore, the behavior of these twisted signatures under satellite operations is investigated in Theorem~\ref{thm:CablingTheorem}.
\item Section~\ref{sec:CassonGordon} focuses on metabelian representations.
Theorem~\ref{thm:BlanchfieldCG} relates our twisted signature functions to the Casson-Gordon invariants, while Theorem~\ref{thm:metabelian-cabling-formula} studies the behavior of these metabelian signatures under satellite operations; the resulting formulas are reminiscent of work due to Litherland~\cite{Litherland}.
\item Section~\ref{sec:TorusKnots} is concerned with explicit computations of metabelian signatures.
Proposition~\ref{prop:DecompoTwistedBlanchfield}
calculates these signatures for $(2,2k+1)$-torus knots.
Theorem~\ref{thm:HKLNotSlice} uses this result and our satellite formulas to obstruct the sliceness of the algebraic knot $T_{2,13}\#T_{2,3;2,15}\#-T_{2,15}\#-T_{2,3;2,13}$, reproving a theorem of Hedden, Kirk and Livingston~\cite{HeddenKirkLivingston}.
\end{itemize}
\subsection{Detailed overview}
The first part of this work aims to extract signature invariants from linking forms over~$\L{\F}$, with $\mathbb{F}=\mathbb{R},\mathbb{C}$.
As an initial step, Section~\ref{sec:LinkingFormClassification} describes a classification of non-singular $\L{\F}$-linking forms up to isometry.
In the real case, such classifications are implicit in~\cite{Ne95, BorodzikFriedl2}.
On the other hand, the complex case appears to have remained more elusive.
While the methods require further work than in the real case, our approach is explicit and concrete: Theorem~\ref{thm:MainLinkingForm} proves that every non-singular linking form over $\L{\F}$ can be decomposed into explicit ``basic pairings".
In fact, Theorem~\ref{thm:MainLinkingForm} also deals with singular pairings under an additional technical assumption.
In Section~\ref{sec:FurtherProperties}, we apply our classification results to study linking forms over some local rings and investigate whether linking forms are representable by a Hermitian matrix (the classification is also used in Subsection~\ref{sec:uniqueness} to define analogues of the Hodge numbers of N\'emethi~\cite{Ne95}).
Here, we call a non-singular linking form $\lambda \colon M \times M \to Q/R$ \emph{representable} if there is a non-degenerate Hermitian matrix~$A$ over $R$ such that $\lambda$ is isometric to
\begin{align*}
\lambda_{A} \colon R^n /A^TR^n \times R^n/A^T R^n &\to Q/R \\
([x],[y]) & \mapsto x^TA^{-1} y^\#.
\end{align*}
While linking forms over $\L{\R}$ are known to be representable~\cite{BorodzikFriedl2}, the methods of Section~\ref{sec:LinkingFormClassification} allow us to provide explicit obstructions to the representability of linking forms over $\L{\C}$, see for example Proposition~\ref{prop:notsostupidexample}.
In Section~\ref{sec:Witt}, we use our explicit classification results to re-derive the classification of $\L{\F}$-linking forms up to Witt equivalence~\cite{LevineMetabolicHyperbolic, Litherland, RanickiExact, KeartonSignature, MilnorInfiniteCyclic}.
In Section~\ref{sec:Signatures}, we associate various signatures to $\L{\F}$-linking forms.
A glance at the existing literature~\cite{MilnorInfiniteCyclic, Matumoto, Neumann, Litherland,LevineMetabolicHyperbolic} suggests that two approaches are possible: the first relies on ``signature jumps" (associated to primary summands of the linking form), while the second involves signatures of a representing matrix.
As a linking form need not be representable, one attractive feature of our approach is that we are able to associate to \emph{any} non-singular $\L{\F}$-linking form $(M,\lambda)$ a signature function
$$ \sigma_{(M,\lambda)} \colon S^1 \to \mathbb{Z}.$$
When $(M,\lambda)$ is representable, we relate $\sigma_{(M,\lambda)}$ to signatures of a representing matrix (Corollary~\ref{cor:sigissig}) and to signature jumps; see Proposition~\ref{prop:JumpIsJump}.
These results can also be seen as generalizations of work due to Matumoto~\cite{Matumoto} (see also~\cite{Litherland, LevineMetabolicHyperbolic}).
Furthermore, Corollary~\ref{cor:WittRepresentability} shows that (jumps of) the signature function also provides an obstruction to the representability of a linking form.
Finally, while $\sigma_{(M,\lambda)}$ is not invariant under Witt equivalence, an averaging process yields a signature function $\sigma_{(M,\lambda)}^{\operatorname{av}}$ which is well defined on the Witt group of non-singular linking forms over $\L{\F}$; see Proposition~\ref{prop:avsig}.
\medbreak
The second part of this paper provides topological applications of our algebraic machinery.
Namely, relying on twisted Blanchfield pairings (which are reviewed in Section~\ref{sec:TwistedHomology}), we define twisted signature invariants for closed 3-manifolds and knots.
Let us expand on the case of knots.
Let $K \subset S^3$ be a knot with 0-framed surgery $M_K$ and let $\beta \colon \pi_1(M_K) \to GL_d(\L{\F})$ be a unitary representation such that the twisted Alexander module $H_1(M_K;\L{\F}^d_\beta)$ is $\L{\F}$-torsion.
Associated to this data, in Section~\ref{sec:TwistedSign3ManifKnot}, we define a \emph{twisted signature function}
$$ \sigma_{(K,\beta)} \colon S^1 \to \mathbb{Z}. $$
This function generalizes the classical Levine-Tristram signature~\cite{LevineMetabolicHyperbolic, Tristram} and shares many of its properties: it is locally constant on the complement of the zeroes of the twisted Alexander polynomial, its behavior under mirror images and orientation reversal can be described (Proposition~\ref{prop:LocallyConstant}), under favorable assumptions, it obstructs concordance~(Proposition~\ref{prop:witt-equiv-concordance}), it is additive under connected sums (Corollary~\ref{cor:SignatureSatellite}) and its behavior under satellites is well understood (Theorem~\ref{thm:CablingTheorem} and Corollary~\ref{cor:SignatureSatellite}).
In Section~\ref{sec:CassonGordon}, we focus on signature invariants arising from certain metabelian representations.
These representations (which stem from the seminal work of Casson and Gordon~\cite{CassonGordon1,CassonGordon2}) have been applied successfully to twisted Alexander polynomials~\cite{KirkLivingston, KirkLivingstonMutation, HeraldKirkLivingston, ChaFriedl, Miller} and more recently to twisted Blanchfield pairings~\cite{MillerPowell}.
Applying the latter work of Miller and Powell, Theorem~\ref{thm:MillerPowellSignature} shows that our metabelian signature invariants provide obstructions to sliceness.
Furthermore, in Theorem~\ref{thm:BlanchfieldCG}, we relate our signatures to (signatures of) the Witt classes of Casson-Gordon~\cite{CassonGordon1, CassonGordon2}.
Finally, in Section~\ref{sec:TorusKnots}, we compute the aforementioned metabelian signature invariants of some specific knots.
First, in ~Proposition~\ref{prop:DecompoTwistedBlanchfield} we use Powell's algorithm~\cite{PowellThesis} to determine the isometry type (and thus the signatures) of the twisted Blanchfield pairing of the 2-stranded torus knots~$T_{2,2k+1}$ for some metabelian representations.
To the best of our knowledge this is the first time that metabelian signatures have been computed this explicitly.
The calculations are rather tedious, but very explicit, owing once again to the concrete nature of our classification results.
In particular, since the behavior of our signatures invariants under cabling is understood, we are able to compute our metabelian signature invariants for the algebraic knot $T_{2,3;2,15}\# T_{2,13}\# -T_{2,3;2,13}\# -T_{2,15}$, obstructing its sliceness and recovering a result of Hedden, Kirk and Livingston~\cite{HeddenKirkLivingston}.
\medbreak
It is worth noting that the relationship between (signature jumps of) the Levine-Tristram signature, the Milnor signatures and the untwisted Blanchfield pairing (as a form over $\L{\R}$) is well known~\cite{KeartonSignature, Matumoto}. The same remark holds for the behavior of each of these classical invariants under satellite operations~\cite{LivingstonMelvin, Shinohara, LitherlandIterated, KeartonCompound}. On the other hand, it appears that the algebraic background needed to investigate similar questions for twisted Blanchfield forms (in particular as forms over~$\L{\C}$) is lacking. We believe that this issue, coupled with the unavailability of explicit representing matrices and algorithms to calculate signatures of linking forms, explains the absence of computable twisted generalizations of classical signature invariants.
\medbreak
We conclude this introduction with some remarks concerning other twisted signature invariants which have appeared in the literature.
Levine used the Atiyah-Patodi-Singer eta invariant to define signature invariants for links~\cite{LevineEta}.
These invariants, which take a unitary representation $\pi_1(M_K) \to U(n)$ as an additional input, have been thoroughly studied by Friedl~\cite{FriedlEta, FriedlEtaLink,FriedlApproximation}.
We strongly believe that for certain choices of $\beta$, both invariants are related.
Finally, we have little doubt that our invariants are also closely related to the \emph{twisted Milnor signatures} introduced by Kirk-Livingston~\cite{KirkLivingston} and further studied by Nosaka~\cite{Nosaka}.
\subsection*{Acknowledgments}
We are indebted to Andrew Ranicki for his guidance and invaluable help while we were writing the paper.
We have benefited from discussions with several people, including Jae Choon Cha, Stefan Friedl, Patrick Orson and Mark Powell.
We also wish to thank the University of Geneva and the University of Warsaw at which part of this work was conducted.
The first author is supported by the National Science Center grant 2016/22/E/ST1/00040.
The second author thanks Durham University for its hospitality and was supported by an early Postdoc.Mobility fellowship funded by the Swiss FNS.
The third author is supported by the National Science Center grant 2016/20/S/ST1/00369.
\begin{convention}
If $R$ is a commutative ring and $f,g\in R$, we write $f\doteq g$ if there exists a unit $u\in R$ such that $f=ug$.
For a ring $R$ with involution, we denote this involution by $x\mapsto x^\#$; the symbol~$\ol{x}$ is reserved for the complex conjugation.
In particular, for $R=\L{\C}$ the involution $(-)^\#$ is the composition of the complex conjugation with the map $t\mapsto t^{-1}$.
For example, if $p(t)=t-i$, then~$p^\#(i)=0$, but $\ol{p}(i)=2i$.
Given an $R$-module $M$, we denote by $M^\#$ the $R$-module that has the same underlying additive group as $M$, but for which the action by $R$ on $M$ is precomposed with the involution on $R$.
For a matrix $A$ over $R$, we write $\makeithashT{A}$ for the transpose followed by the conjugation.
\end{convention}
\section{Classification of split linking forms up to isometry}\label{sec:LinkingFormClassification}
This section is organized as follows.
In Subsection~\ref{sub:Terminology}, we introduce some basic terminology on linking forms.
Subsection~\ref{sec:basic_linking_forms} describes the building blocks for split linking forms. In Subsection~\ref{sec:classification-theorem}, we classify split linking forms over $\mathbb{F}[t^{\pm 1}]$ up to isometry.
The proof of the classification result occupies Subsections~\ref{sub:sign_of_four}, \ref{sub:cyclic_pairing} and~\ref{sec:uniqueness}, with some technical parts deferred to Sections~\ref{sec:sym_real} and~\ref{sec:sym_complex} in the Appendix.
\subsection{Linking forms}
\label{sub:Terminology}
In this subsection, we first introduce some terminology on linking forms over an integral domain $R$ with involution.
We then restrict ourselves to the case where $R=\L{\F}$ where the field $\mathbb{F}$ is either $\mathbb{R}$ or $\mathbb{C}$. References include~\cite{Litherland, RanickiExact, RanickiLocalization}.
\medbreak
Let $R$ be an integral domain with a (possibly trivial) involution $x \mapsto x^\#$.
Let $Q$ be the field of fractions of $R$ which inherits an involution from $R$.
A \emph{linking pairing} consists of a finitely generated torsion $R$-module $M$ and an $R$-linear map
\[\lambda^{\bullet} \colon M \to \operatorname{Hom}_R(M,Q/R)^\#.\]
In practice, we shall often think of a linking pairing as a pair $(M,\lambda)$, where $\lambda$ is a pairing
\[\lambda \colon M \times M \to Q/R, \text{ where } \lambda(x,y) = \lambda^{\bullet}(x)(y).\]
Using this point of view, a linking pairing is \emph{Hermitian} if $\lambda(x,y)=\makeithash{\lambda(y,x)}$ for all $x,y \in M$.
\begin{definition}
\label{def:LinkingForm}
A Hermitian linking pairing $(M,\operatorname{\lambda})$ is called a \emph{linking form}.
\end{definition}
A linking form $(M,\lambda)$ is \emph{non-degenerate} if $\lambda^{\bullet}_M$ is injective and \emph{non-singular} if $\lambda^{\bullet}_M$ is an isomorphism.
In the literature, linking forms are usually assumed to be either non-degenerate or non-singular.
Note that if the ring $R$ is a PID, then non-degenerate linking forms over a torsion $R$-module are in fact non-singular~\cite[Lemma 3.24]{Hillman}.
A \emph{morphism} of linking forms $(M,\lambda_M)$ and~$(N,\lambda_{M'})$ consists of a $R$-linear homomorphism $\psi \colon M \to N$ which intertwines $\lambda_M$ and $\lambda_N$, i.e.\ such that $\lambda_N(\psi(x),\psi(y))=\lambda_M(x,y)$.
A morphism $\psi$ from $(M,\lambda_M)$ to $(N,\lambda_N)$ is an \emph{isometry} if~$\psi \colon M \to N$ is an isomorphism.
Let $\mathbb{F}$ be a field with involution which we will denote by \(a \mapsto \bar{a}\).
We will study linking forms over the ring of Laurent polynomials $\mathbb{F}[t^{\pm 1}]$.
As usual, we denote by~$\mathbb{F}(t)$ the field of fractions of~$\L{\F}$.
The involution on $\mathbb{F}$ can be extended to an involution on $\L{\F}$ defined by
\[p=\sum a_it^i\mapsto p^\#=\sum\ol{a}_it^{-i}.\]
A Laurent polynomial $p\in\mathbb{F}[t^{\pm 1}]$ is called \emph{symmetric} if $p=p^\#$.
Next comes the main definition of this subsection.
\begin{definition}
\label{def:Split}
A linking form \((M,\lambda)\) over $\L{\F}$ is \emph{weakly split} if \(M\) decomposes into a direct sum $M_1 \oplus \cdots \oplus M_n$ of cyclic modules in such a way that
the decomposition is orthogonal with respect to $\lambda$. A linking form is \emph{split} if it decomposes into an orthogonal sum of cyclic modules $M_1\oplus\dots\oplus M_n$
and each of the $M_i$ is of the form $\L{\F}/f_i$ and $f_i^\#\doteq f_i$.
\end{definition}
The goal of this section is to classify weakly split linking forms up to isometry.
Note that this task reduces to the classification of linking pairings on cyclic modules.
\begin{remark}
\label{rem:SplitInvolution}
Note that if $M$ is endowed with a non-singular linking form, then the order of $M$ necessarily satisfies $\operatorname{ord}(M)^\#\doteq \operatorname{ord}(M)$,
see~\cite[Lemma 3.25]{Hillman}. Hence for non-singular linking forms, the condition $\ord M_i\doteq \ord M_i$ is automatically satisfied.
\end{remark}
\begin{remark}
\label{rem:NeedSplit}
There is a reason why we study split forms and not just non-singular forms: their topological applications appear in Subsection~\ref{sub:Satellite}. Technical results
leading to weakly split forms appear in Subsection~\ref{sec:isometricembedding}.
\end{remark}
Since our interest in classifying linking forms stems from the study of twisted Blanchfield pairings (which are non-singular if one works with closed 3-manifolds and PID's~\cite{Powell}), it is worth pointing out that non-singular pairings are indeed split.
\begin{theorem}[\cite{BorodzikFriedl2}]
\label{thm:non-degenerate-split}
For \(\mathbb{F} = \mathbb{R}\) and \( \mathbb{F}= \mathbb{C}\), non-singular linking forms over $\L{\F}$ are split.
\end{theorem}
\begin{proof}
The result is proved in \cite[Lemma 4.3]{BorodzikFriedl2} in the real case, but the proof carries through to the complex case.
\end{proof}
Not all linking forms are weakly split: an example of a non weakly split linking form is given in~Subsection~\ref{sub:NonSplit}.
\subsection{Basic linking forms}\label{sec:basic_linking_forms}
The classification of weakly split linking forms essentially boils down to the classification of linking forms defined on cyclic modules.
This section provides the key examples in this setting.
Following the approach of~\cite{Ne95}, we first describe simple ``basic'' linking forms that will be the building blocks of all split linking forms.
\medbreak
Let us start with a discussion of the real case.
\begin{example}
\label{BasicPairingXiS1Real}
Fix a complex number $\xi \in S^1$ with $\re (\xi)>0$, a positive integer $n$, a non-negative integer $k \leq n$, and $\epsilon= \pm 1$.
Define ${}^\R\! B_\xi(t) = (t-\xi)(1-\ol{\xi}t^{-1})$ and consider the linking form~$\mathfrak{e}(n,k,\epsilon,\xi,\mathbb{R})$ defined on $\L{\R}/{}^\R\! B_\xi(t)^n$ by the formula
\begin{equation}\label{eq:e_n_k_form_real}
\begin{split}
\L{\R} / {}^\R\! B_\xi(t)^n \times\L{\R} / {}^\R\! B_\xi(t)^n& \to\O{\R}/\L{\R}, \\
(x,y) &\mapsto \frac{\epsilon x y^{\#}}{{}^\R\! B_\xi(t)^{n-k}}.
\end{split}
\end{equation}
Observe that this pairing is Hermitian since ${}^\R\! B_\xi(t)$ is symmetric.
Moreover, \(\mathfrak{e}(n,k,\epsilon,\xi,\mathbb{R})\) is non-singular if and only if $k=0$.
Indeed, if \(k>0\), then $\lambda({}^\R\! B_\xi(t)^{n-k}x,y) = 0$ for any \(x,y\).
To prove the converse, assume that $x$ is such that $\frac{\epsilon x y^{\#}}{{}^\R\! B_\xi(t)^{n}}$ vanishes for all $y$.
This implies that $x$ lies in the submodule generated by ${}^\R\! B_\xi(t)^n$.
\end{example}
The case \(\xi=\pm 1\) is treated separately.
\begin{example}
\label{BasicPairingXipm1Real}
Let $\xi = \pm 1$.
Fix a positive integer $n$ and a non-negative integer~$k \leq n$ such that~$n-k$ is even.
Define \({}^\R\! B_\xi(t) = (t-\xi)\) and consider the real linking form $\mathfrak{e}(n,k,\epsilon, \pm 1,\mathbb{R})$:
\begin{equation}\label{eq:e_n_k_form_real_xi_one}
\begin{split}
\L{\R}/ {}^\R\! B_\xi(t)^n \times\L{\R} / {}^\R\! B_\xi(t)^n &\to\O{\R} /\L{\R},\\
(x,y) &\mapsto \frac{\epsilon x y^{\#}}{{}^\R\! B_\xi(t)^{(n-k)/2}{}^\R\! B_\xi(t^{-1})^{(n-k)/2}}.
\end{split}
\end{equation}
This pairing is clearly Hermitian since ${}^\R\! B_\xi(t){}^\R\! B_\xi(t^{-1})$ is symmetric.
It is non-singular if and only if \(k=0\): the proof follows the same steps as in Example~\ref{BasicPairingXiS1Real} by noting that since $\xi=\pm 1$, the modules
$\L{\R}/ {}^\R\! B_\xi(t)^n$ and $\L{\R}/ {}^\R\! B_\xi(t)^{(n-k)/2}{}^\R\! B_\xi(t^{-1})^{(n-k)/2}$ are isomorphic.
\end{example}
Finally, we deal with the case where $\xi$ does not belong to $S^1$.
\begin{example}
\label{ex:RealBasicXiNotReal}
For a complex number $\xi$ such that $|\xi| \in (0,1)$, define ${}^\R\! B_\xi(t)={}^\R\! B_{\xi+}(t){}^\R\! B_{\xi-}(t)$, where
\begin{align*}
{}^\R\! B_{\xi+}(t)&=(t-\xi), & {}^\R\! B_{\xi-}(t)&=(t^{-1}-\ol{\xi})&\textrm{ if $\xi\in\mathbb{R}$,}\\
{}^\R\! B_{\xi+}(t)&=(t-\xi)(t-\ol{\xi}), & {}^\R\! B_{\xi-}(t)&=(t^{-1}-\ol{\xi})(t^{-1}-\xi)&\textrm{otherwise.}
\end{align*}
Fix non-negative integers~\(a,b\) such that $a+b>0$, a non-negative integer \(k \leq \min(a,b)\) and consider the linking form
$\mathfrak{f}^w((a,b),k,\xi,\mathbb{R})$:
\begin{equation}\label{eq:f_n_k_form_real_weak}
\begin{split}
\L{\R} /({}^\R\! B_{\xi+}(t)^a {}^\R\! B_{\xi-}(t)^b) \times \L{\R} / ({}^\R\! B_{\xi+}(t)^a {}^\R\! B_{\xi-}(t)^b) &\to \O{\R}/\L{\R}, \\
(x,y) &\mapsto \frac{x y^{\#}}{{}^\R\! B_\xi(t)^{\min(a,b)-k}}.
\end{split}
\end{equation}
By definition, this pairing is Hermitian and weakly split. Furthermore, it is split if and only if~$a=b$ and, in this case, the pairing takes the form
\begin{equation}\label{eq:f_n_k_form_real}
\begin{split}
\L{\R} / {}^\R\! B_\xi(t)^a \times \L{\R} / {}^\R\! B_\xi(t)^a& \to \O{\R}/\L{\R}, \\
(x,y) &\mapsto \frac{x y^{\#}}{{}^\R\! B_\xi(t)^{a-k}},
\end{split}
\end{equation}
and is denoted $\mathfrak{f}(a,k,\xi,\mathbb{R})$. Finally, note that $\mathfrak{f}^w((a,b),k,\xi,\mathbb{R})$ is non-singular if and only if $a=b$ and $k=0$: the argument is exactly the same as in Example~\ref{BasicPairingXiS1Real}.
\end{example}
In the complex case, elementary pairings are defined in a similar fashion, but there are only two cases to consider depending on the norm of \(\xi\).
In order to deal with a technical issue which appears in the case $|\xi|=1$, we introduce some terminology.
\begin{definition}\label{def:positive}
Given $\xi\in S^1$, a complex polynomial $r(t)$ is called \emph{$\xi$-positive} if $(t^{-1}-\ol{\xi})r(t)$ is a complex symmetric polynomial and the function
$\theta\mapsto (e^{-i\theta}-\ol{\xi})r(e^{i\theta})$ changes sign from positive to negative as $\theta$ crosses the value $\theta_0$ for which $e^{i\theta_0}=\xi$.
\end{definition}
While one could work with explicit $\xi$-positive polynomials, it is often useful to use the abstract notion instead.
To build up some intuition, we provide some concrete examples of $\xi$-positive polynomials:
\begin{example}
\label{ex:XiPositive}
If $\xi\neq\pm 1$ then $r(t)=1-\xi t$ is $\xi$-positive if $\iim(\xi)>0$ and $-(1-\xi t)$ is $\xi$-positive if $\iim(\xi)<0$.
Indeed, if $\iim(\xi)>0$, then $(t^{-1}-\ol{\xi})r(t)=t-2\re\xi +t^{-1}$
and so $(e^{-i\theta}-\xi)r(e^{i\theta})<0$ if $\re(\theta)<\re(\xi)$, that is if $\theta<\theta_0$.
\end{example}
We leave the construction of a linear $\xi$-positive polynomial with $\xi=\pm 1$ to the reader (this involves Lemma~\ref{lem:quadratic}).
\begin{remark}\label{rem:positive}
The involution $r(t)\mapsto r(t)^\#$ can be extended to the case where $r(t)$ is a complex analytic function near $\xi$: if $r(t)=\sum a_j(t-\xi)^j$ near $\xi$, then one sets $r(t)^\#=\sum \ol{a}_j(t^{-1}-\ol{\xi})^j$.
Definition~\ref{def:positive} can also be extended to the case where $r(t)$ is complex analytic.
Namely, $r(t)$ is called $\xi$-positive if $(t^{-1}-\ol{\xi})r(t)$ is symmetric and changes sign from positive to negative when crossing the value $\xi$, as in Definition~\ref{def:positive}.
This extension is used in Section~\ref{sub:LocalizationDiagonalization}.
\end{remark}
For later use, we record the following characterisation of $\xi$-positivity.
\begin{lemma}\label{lem:residue}
Let $r(t)$ be a $\xi$-positive function (a polynomial or an analytic function) and $r(\xi)\neq~0$.
Set $r_0=\res_{t=\xi}\frac{r(t)}{t-\xi}\in\mathbb{C}$.
Then $\re (\ol{\xi}r_0)=0$ and $\iim(\ol{\xi}r_0)<0$.
\end{lemma}
The proof is elementary but technical and it is deferred to Section~\ref{sec:residue}.
We can now list the basic linking forms in the complex case.
\begin{example}
\label{BasicPairingXiS1Complex}
Fix a complex number $\xi\in S^1$, a positive integer $n$, a non-negative integer $k \leq n$.
Define ${}^\C\! B_{\xi}(t)= t-\xi$ and consider the linking form $\mathfrak{e}(n,k,\epsilon,\xi,\mathbb{C})$:
\begin{align}
\L{\C} / {}^\C\! B_{\xi}(t)^n \times \L{\C} / {}^\C\! B_{\xi}(t)^n &\to \O{\C}/\L{\C}, \nonumber\\
(x,y) &\mapsto \frac{\epsilon x y^{\#}}{{}^\C\! B_{\xi}(t)^{\frac{n-k}{2}}{}^\C\! B_{\bar{\xi}}(t^{-1})^{\frac{n-k}{2}}}, \quad \ \ \ \ \ \text{if $n-k$ is even}, \label{eq:e_n_k_form_complex_even}\\
(x,y) &\mapsto \frac{\epsilon r(t) x y^{\#}}{{}^\C\! B_{\xi}(t)^{\frac{n-k+1}{2}}{}^\C\! B_{\bar{\xi}}(t^{-1})^{\frac{n-k-1}{2}}}, \quad \text{if $n-k$ is odd}, \label{eq:e_n_k_form_complex_odd}
\end{align}
where the polynomial $r(t)$ appearing in the numerator of \eqref{eq:e_n_k_form_complex_odd} is linear and $\xi$-positive (note that the existence of such polynomials follows from Example~\ref{ex:XiPositive}).
In the odd case, it will be shown in Lemma~\ref{lem:qlinear} below that the isometry class of $\mathfrak{e}(n,k,\epsilon,\xi,\mathbb{C})$ does not depend on the particular choice of the linear $\xi$-positive polynomial $r(t)$.
Regardless of the parity of $n-k$, these pairings are Hermitian: the case $n-k$ even is obvious while the case $n-k$ odd follows from the definition of $\xi$-positivity; for this, note that $C_{\ol{\xi}}(t^{-1})=C_\xi(t)^\#$.
Following the same steps as in the real case, we easily show that these pairings are non-singular if and only if $k=0$.
\end{example}
As in the real case, we conclude with the case where $\xi$ does not belong to $S^1$.
\begin{example}
For a
complex number $\xi$ such that $|\xi| \in (0,1)$, set ${}^\C\! B_\xi(t) = {}^\C\! B_{\xi+}(t){}^\C\! B_{\xi-}(t)$, where ${}^\C\! B_{\xi+}(t)=(t-\xi)$ and ${}^\C\! B_{\xi-}(t)=(t^{-1}-\bar{\xi})$.
Fix non-negative integers $(a,b)$ such that $a+b>0$ and a non-negative integer $k\le \min(a,b)$ and consider the linking form
$\mathfrak{f}^w((a,b),k,\xi,\mathbb{C})$:
\begin{equation}\label{eq:f_n_k_form_complex_weak}
\begin{split}
\L{\C} / ({}^\C\! B_{\xi+}(t)^a {}^\C\! B_{\xi-}(t)^b) \times \L{\C} / ({}^\C\! B_{\xi+}(t)^a {}^\C\! B_{\xi-}(t)^b)& \to \O{\C}/\L{\C}, \nonumber \\
(x,y) &\mapsto \frac{x y^{\#}}{{}^\C\! B_{\xi}(t)^{\min(a,b)-k}}.
\end{split}
\end{equation}
The pairing is Hermitian and weakly split, but it is split if and only if $a=b$. In the latter case we write it as $\mathfrak{f}(a,k,\xi,\mathbb{C})$:
\begin{align}
\L{\C} / {}^\C\! B_{\xi}(t)^a \times \L{\C} / {}^\C\! B_{\xi}(t)^a &\to \O{\C}/\L{\C}, \nonumber \\
(x,y) &\mapsto \frac{x y^{\#}}{{}^\C\! B_{\xi}(t)^{a-k}}\label{eq:f_n_k_form_complex}
\end{align}
Once again this pairing is Hermitian and it is non-singular if and only if $k=0$.
\end{example}
\subsection{The classification theorem}
\label{sec:classification-theorem}
The goal of this subsection is to classify weakly split linking forms over $\L{\F}$ up to isometry.
More precisely, the following theorem is the main result of this section; its proof will be given in the next three subsections.
\begin{theorem}
\label{thm:MainLinkingForm}
Suppose $\mathbb{F}$ is either $\mathbb{R}$ or $\mathbb{C}$.
Every weakly split linking form $(M,\lambda)$ over $\L{\F}$ is isometric to one of the type
\begin{equation}\label{eq:splitting}
\bigoplus_{\substack{ n_i,k_i,\epsilon_i,\xi_i\\ i\in I}}\mathfrak{e}(n_i,k_i,\epsilon_i,\xi_i,\mathbb{F})\oplus
\bigoplus_{\substack{a_j,b_j,k_j,\xi_j\\ j\in J}}\mathfrak{f}^w((a_j,b_j),k_j,\xi_j,\mathbb{F}).
\end{equation}
Here $I$ and $J$ are finite sets of indices.
Moreover, {if the linking form is non-singular}, then the decomposition~\eqref{eq:splitting} is unique up to permuting summands.
\end{theorem}
\begin{remark}
\label{rem:OrderAndClassification}
In the decomposition~\eqref{eq:splitting}, there can only be finitely many non-trivial summands.
$\mathfrak{e}(n,k,\epsilon,\xi,\mathbb{F})$ enters the decomposition.
Note furthermore that since we are working over a PID, the order of $M$ is equal to the product of the basic polynomials which enter the decomposition.
\end{remark}
\begin{remark}
{Although we believe that the decomposition~\eqref{eq:splitting} is unique for all weakly split linking forms, we do not have a rigorous proof.}
\end{remark}
The core of proof of the first part of~Theorem~\ref{thm:MainLinkingForm} will be given in Subsection~\ref{sub:cyclic_pairing},
while the uniqueness is proved in Subsection~\ref{sec:uniqueness}.
The proofs of several technical lemmas are deferred to Subsections~\ref{sec:sym_real} and~\ref{sec:sym_complex} in the Appendix.
We now resume the notations ${}^\R\! B_\xi(t)$ and ${}^\C\! B_\xi(t)$ that were introduced in Section~\ref{sec:basic_linking_forms}.
Namely, in the real case, we define
\[\Xi:=\{\xi\in\mathbb{C}\setminus\{0\} \ | \ \iim (\xi)\ge 0,\ |\xi|\le 1\}\]
and, for $\xi\in\Xi$, we shall refer to the following real polynomials as the (real) \emph{basic polynomials}:
\[
{}^\R\! B_\xi(t):=\begin{cases}
(t-\xi), & \text{for } \xi=\pm1,\\
(t-\xi)(1-\ol{\xi}t^{-1}), & \text{for } |\xi|=1,\;\xi\neq\pm 1,\\
(t-\xi)(t-\ol{\xi})(1-\ol{\xi}^{-1}t^{-1})(1-t^{-1}\xi^{-1}), & \text{for } |\xi|<1,\;\xi\notin\mathbb{R},\\
(t-\xi)(1-\xi^{-1}t^{-1}), & \text{for } \xi\in (-1,1).
\end{cases}
\]
Recall also that for $|\xi|<1$ we have ${}^\R\! B_\xi(t)={}^\R\! B_{\xi+}(t){}^\R\! B_{\xi-}(t)$, where we recall that
\begin{align*}
{}^\R\! B_{\xi+}(t)&=(t-\xi), & {}^\R\! B_{\xi-}(t)&=(t^{-1}-\ol{\xi})&\textrm{ if $\xi\in\mathbb{R}$,}\\
{}^\R\! B_{\xi+}(t)&=(t-\xi)(t-\ol{\xi}), & {}^\R\! B_{\xi-}(t)&=(t^{-1}-\ol{\xi})(t^{-1}-\xi)&\textrm{otherwise.}
\end{align*}
The purpose of these basic polynomials should be clear: any real symmetric polynomial $f$ can be decomposed as $p_1^{n_1} \cdots p_{r}^{n_r}$, where the $p_i$ are basic.
Moreover, a \emph{basic module} is one of the form~\(\L{\R}/{}^\R\! B_\xi(t)^n\) or $\L{\R}/({}^\R\! B_{\xi+}(t)^a{}^\R\! B_{\xi-}(t)^b)$, where $n,a,b$ are non-negative integers.
Similarly, in the complex case, we define the \emph{basic polynomials} to be
\[{}^\C\! B_{\xi}(t) =
\begin{cases}
(t-\xi), & \text{if } |\xi|=1,\\
(t-\xi)(t^{-1}-\xi), & \text{if } |\xi| \in (0,1).
\end{cases}\]
If $|\xi|<1$ we have ${}^\C\! B_{\xi}(t)={}^\C\! B_{\xi+}(t){}^\C\! B_{\xi-}(t)$ with ${}^\C\! B_{\xi+}(t)=(t-\xi)$, ${}^\C\! B_{\xi-}(t)=(t^{-1}-\ol{\xi})$.
By analogy with the real case, we define the \emph{basic modules} to be \(\L{\C}/{}^\C\! B_{\xi}(t)^n\), for some \(n>0\) or $\L{\C}/({}^\C\! B_{\xi+}(t)^a{}^\C\! B_{\xi-}(t)^b)$ for $a,b\ge 0$ and
$a+b>0$.
If the field $\mathbb{F}$ is not specified, then we sometimes write ${}^\F\! B_\xi(t)$ with ${}^\F\! B_\xi(t)={}^\R\! B_\xi(t)$ if $\mathbb{F}=\mathbb{R}$, and ${}^\F\! B_\xi(t)={}^\C\! B_\xi(t)$ if $\mathbb{F}=\mathbb{C}$.
\begin{remark}
The polynomials
${}^\F\! B_\xi(t)$ are characterised as the minimal degree polynomials over~$\L{\F}$ that vanish on $\xi$ and satisfy ${}^\F\! B_\xi\doteq{}^\F\! B_\xi(t)^\#$. The difference between split and weakly split linking forms lies in the fact that the polynomials ${}^\F\! B_\xi(t)$ are decomposable if $|\xi|<1$.
\end{remark}
\subsection{Four technical lemmas}\label{sub:sign_of_four}
In order to prove Theorem~\ref{thm:MainLinkingForm}, we first focus on the study of linking forms over modules $\L{\F}/{}^\F\! B_\xi(t)^n$ and $\L{\F}/({}^\F\! B_{\xi+}(t)^a{}^\F\! B_{\xi-}(t)^b)$.
Namely, we classify these forms in the following order:
\begin{itemize}
\item[(a)] $\mathbb{F}=\mathbb{R}$, $\xi \in S^1 \setminus \lbrace \pm 1 \rbrace $;
\item[(b)] $\mathbb{F}=\mathbb{R}$, $\xi=\pm 1$;
\item[(c)] $\mathbb{F}=\mathbb{C}$, $\xi \in S^1$;
\item[(d)] $\mathbb{F}=\mathbb{R}$, or $\mathbb{F}=\mathbb{C}$ and $|\xi|<1$.
\end{itemize}
The first two cases correspond to the basic pairing $\mathfrak{e}(n,k,\epsilon,\xi,\mathbb{R})$, the third case corresponds to the basic pairing $\mathfrak{e}(n,k,\epsilon,\xi,\mathbb{C})$ and the last case comprises both the basic pairings $\mathfrak{f}^w((a,b),k,\xi,\mathbb{R})$ and $\mathfrak{f}^w((a,b),k,\xi,\mathbb{C})$.
The following lemma deals with case~(a).
\begin{lemma}
\label{lem:structure_lemma}
Fix a complex number $\xi$ in $S^1 \setminus \lbrace \pm 1 \rbrace$ and a positive integer $n$.
Any nontrivial linking form \((\L{\R}/{}^\R\! B_\xi(t)^n,\lambda)\) is isometric to $\mathfrak{e}(n,k,\epsilon,\xi,\mathbb{R})$ for some $\epsilon=\pm 1$ and some~$k < n$.
\end{lemma}
\begin{proof}
Set $\zeta=\lambda(1,1)$.
Formally, $\zeta$ belongs to $\mathbb{R}(t)/\mathbb{R}[t^{\pm 1}]$, but we will think of $\zeta$ as an element in~$\mathbb{R}(t)$ defined up to an element in $\mathbb{R}[t^{\pm 1}]$.
As $\operatorname{\lambda}({}^\R\! B_\xi(t)^n,1)=0$ we can assume that $\lambda(1,1)=\frac{q}{{}^\R\! B_\xi(t)^n}$
for some~$q$ in $\L{\R}$ that is well defined modulo ${}^\R\! B_\xi(t)^n$.
Since ${}^\R\! B_\xi(t)^n$ is symmetric and~$\lambda$ is Hermitian,~$q$ must be symmetric.
Given $x,y$ in $\L{\R}/{}^\R\! B_\xi(t)^n$, since \(\lambda\) is sesquilinear, we deduce~that
\begin{equation}
\label{eq:DisplayedPairing}
\lambda(x,y)=\frac{qxy^{\#}}{{}^\R\! B_\xi(t)^n}.
\end{equation}
Write $q=r{}^\R\! B_\xi(t)^k$, where $r$ is coprime with ${}^\R\! B_\xi(t)$.
Since both $q$ and ${}^\R\! B_\xi(t)^k$ are symmetric, so is $r$.
Furthermore, note that if $k$ is greater or equal to $n$, then \(\lambda\) is zero.
Consequently, we can assume that \(k < n\).
To conclude the proof of the lemma, we apply an isometry of the basic module~$\L{\R}/{}^\R\! B_\xi(t)^n$ to change $r$ into $\epsilon=\pm 1$.
More precisely, we use Proposition~\ref{prop:BFSummary}(c) to show that if~$\epsilon$ is the sign of $r(\xi)$, then $r=\epsilon gg^\#\bmod {}^\R\! B_\xi(t)^n$ for some $g \in \mathbb{R}[t^{\pm 1}]$.
Multiplication by $g$ now induces an isometry between $\mathfrak{e}(n,k,\epsilon,\xi,\mathbb{R})$ and the pairing displayed in~(\ref{eq:DisplayedPairing}).
\end{proof}
The next result covers case~(b).
\begin{lemma}\label{lem:structure1}
Let $\xi = \pm1$ and fix a integer $n>0$.
Any nontrivial linking form~\((\L{\R}/(t-~\xi)^n,\lambda)\) is isometric to $\mathfrak{e}(n,k,\epsilon,\pm1,\mathbb{R})$ for some $\epsilon=\pm 1$ and some
$k$ such that $n-k$ is even.
\end{lemma}
\begin{proof}
Proceeding as in the proof of the previous lemma, we have $\operatorname{\lambda}(1,1)=\frac{q}{(t-\xi)^n}$ for some polynomial $q$.
Writing $q=(t-\xi)^{k}r$, where \(r\) is coprime with \({}^\R\! B_\xi(t)\), we deduce that
\[\operatorname{\lambda}(1,1)=\frac{r}{(t-\xi)^{n-k}}.\]
We will first show that \(n-k\) is even.
Suppose on the contrary that \(n-k\) is odd.
Consider the polynomial $\wt{r}=r t^{-(n-k-1)/2}$ and observe that
\[\frac{r}{(t-\xi)^{n-k}}=\frac{\wt{r}}{(t-\xi)((t-\xi)(t^{-1}-\ol{\xi}))^{(n-k-1)/2}}.\]
Since both $\lambda(1,1)$ and $((t-\xi)(t^{-1}-\ol{\xi}))^{(n-k-1)/2}$ are symmetric, it follows that $\wt{r}/(t-\xi)$ must also be symmetric.
From this, we deduce that
$-\xi t \wt{r}(t^{-1})=\wt{r}(t).$
Putting \(t = \xi\) in the latter equality yields $\wt{r}(\xi) = -\wt{r}(\xi)$.
Thus \(\xi\) is a root of \(r\), contradicting the assumption that $r$ is coprime to~\({}^\R\! B_\xi(t)\).
Since we now know that $n-k$ is even, we set $\wt{r}=r t^{-(n-k)/2}$.
It follows that
\[\lambda(1,1)=\frac{\wt{r}}{((t-\xi)(t^{-1}-\ol{\xi}))^{(n-k)/2}}.\]
Since it is easy to check that \(\wt{r}\) is symmetric, Lemma~\ref{lem:simplify_pairing_for_pm1} implies that there exists $\epsilon=\pm 1$ and some~$g\in\L{\R}$ such that $\widetilde{r}=\epsilon gg^\# \bmod {}^\R\! B_\xi(t)^n$.
Multiplication by $g$ now induces an isometry between \(\lambda\) and \(\mathfrak{e}(n,k,\epsilon,\pm1,\mathbb{R})\).
This concludes the proof of the lemma.
\end{proof}
The following result takes care of case~(c).
\begin{lemma}\label{lem:evenpairing}
Let $\xi\in S^1$ and let $n$ be a positive integer.
Any nontrivial linking form \((\L{\C}/{}^\C\! B_{\xi}(t)^n, \lambda)\) is isometric to $\mathfrak{e}(n,k,\epsilon,\pm 1,\mathbb{C})$ for some $\epsilon=\pm 1$ and some $k \leq n$.
\end{lemma}
\begin{proof}
Proceed as in the previous lemmas to obtain polynomials \(\wt{q}, q \in \L{\C}\) coprime with~\({}^\C\! B_{\xi}(t)\) and an integer \(1 \leq k < n\) which satisfy
\begin{equation}\label{eq:pairingcomplex}
\operatorname{\lambda}(1,1)=\frac{\wt{q}}{(t-\xi)^{n-k}} =
\begin{cases}
\frac{q}{((t-\xi)(t^{-1}-\ol{\xi}))^{(n-k)/2}}, & \text{if $n-k$ is even},\\
\frac{q}{(t-\xi)^{(n-k+1)/2}(t^{-1}-\ol{\xi})^{(n-k-1)/2}}, & \text{if $n-k$ is odd}.
\end{cases}
\end{equation}
If $n-k$ is even, then \(q\) is complex symmetric.
Since $q$ is coprime to ${}^\C\! B_{\xi}(t)$, Lemma~\ref{lem:simplifycomplex_part1} guarantees that there exists $\epsilon=\pm 1$ and $g\in\L{\C}$ such that
$q=\epsilon gg^\#\bmod {}^\C\! B_{\xi}(t)^{n}$.
A suitable change of basis then transforms \(\lambda\) to \(\mathfrak{e}(k,n,\epsilon,\xi,\mathbb{C})\).
If \(n-k\) is odd, then Lemma~\ref{lem:complexodd} guarantees the existence of a degree one Laurent polynomial~$q_0$ which is coprime with $(t-\xi)$, such that $(t^{-1}-\ol{\xi})q_0$ is complex symmetric and such that $q=q_0gg^\#\bmod (t-\xi)^n$ for some $g \in \mathbb{C}[t^{\pm 1}]$.
It follows that \(\lambda\) is isometric to the linking form
\[\lambda' \colon (x,y) \mapsto \frac{q_0 xy^\#}{(t-\xi)^{(n-k+1)/2}(t^{-1}-\ol{\xi})^{(n-k-1)/2}}.\]
Using Lemma~\ref{lem:qlinear}, we can then find an isometry between \(\lambda'\) and \(\mathfrak{e}(n,k,\epsilon,\xi,\mathbb{C})\), for some \(\epsilon = \pm1\).
The sign $\epsilon$ is $+1$ if $q_0$ is $\xi$-positive and $-1$ if $-q_0$ is $\xi$-positive.
\end{proof}
The following lemma deals with case~(d).
\begin{lemma}\label{prop:notins1}
Let $a,b$ be positive integers and fix a nonzero complex number $\xi$ satisfying $|\xi| \in (0,1)$.
Any nontrivial linking form \((\L{\F}/({}^\F\! B_{\xi+}(t)^a{}^\F\! B_{\xi-}(t)^b), \lambda)\) is isometric to $\mathfrak{f}^w((a,b),k,\xi, \mathbb{F})$ for some $k\le \min(a,b)$.
\end{lemma}
\begin{proof}
We write $1_\xi$ for the image of $1$ under the quotient map $\L{\F}\to\L{\F}/({}^\F\! B_{\xi+}(t)^a{}^\F\! B_{\xi-}(t)^b)$.
Write $R={}^\F\! B_{\xi+}(t)^a{}^\F\! B_{\xi-}(t)^b$. As $\operatorname{\lambda}(R\cdot1_\xi,1_\xi)=0$ we infer that
$\operatorname{\lambda}(1_\xi,1_\xi)=\frac{\wt{q}}{R}$ for some $q\in\L{\F}$. Write the last fraction as
\[\operatorname{\lambda}(1_\xi,1_\xi)=\frac{q}{{}^\F\! B_{\xi+}(t)^c{}^\F\! B_{\xi-}(t)^d},\]
where $q$ is coprime with ${}^\F\! B_{\xi+}(t),{}^\F\! B_{\xi-}(t)$. Note that $c\le a$ and $d\le b$. Since the pairing is Hermitian, we have $\operatorname{\lambda}(1_\xi,1_\xi)=\operatorname{\lambda}(1_\xi,1_\xi)^\#$ and it follows that
\begin{equation}\label{eq:fracab}\frac{q^\#}{{}^\F\! B_{\xi-}(t)^c{}^\F\! B_{\xi+}(t)^d}=\frac{q}{{}^\F\! B_{\xi+}(t)^c{}^\F\! B_{\xi-}(t)^d}\in\O{\F}/\L{\F}.\end{equation}
From the fact that $q^\#$ is also coprime with ${}^\F\! B_{\xi+}(t),{}^\F\! B_{\xi-}(t)$ we infer that $c=d$. A possible way to see this is to multiply both
sides of \eqref{eq:fracab} by ${}^\F\! B_{\xi+}(t)^c{}^\F\! B_{\xi-}(t)^d$. Then the right hand side becomes zero in $\O{\F}/\L{\F}$, but the left hand side is not, unless $c=d$.
This shows that $c=d$. The
conditions $c\le a$ and $d\le b$ imply that $c\le \min(a,b)$. Set $k=\min(a,b)-c$. We have shown that the pairing is given by
\[\operatorname{\lambda}(1_\xi,1_\xi)=\frac{q}{{}^\F\! B_\xi(t)^{\min(a,b)-k}}.\]
It follows that $q=q^\#$. Next, we relate this pairing to $\mathfrak{f}^w((a,b),k,\xi, \mathbb{F})$. Note that ${}^\F\! B_\xi(t)$, by definition, is positive on $S^1$.
Therefore, by the second point of Proposition~\ref{prop:BFSummary} or Lemma~\ref{lem:coprimepositive} (depending on whether $\mathbb{F}=\mathbb{R}$ or $\mathbb{F}=\mathbb{C}$), we can find $P,Q\in\L{\F}$ taking positive values on $S^1$ such that
\[P{}^\F\! B_\xi(t)^{\min(a,b)}+Qq=1.\]
Note that the first point of Proposition~\ref{prop:BFSummary} or Lemma~\ref{lem:posiissquare} (once more depending on the field) imply that $Q=gg^\#$ for $g\in\L{\F}$.
Thus, we deduce that
$q=q(P{}^\F\! B_\xi(t)^{\min(a,b)}+Qq)\equiv q^2gg^\#\bmod {}^\F\! B_\xi(t)^{\min(a,b)}.$
Since $g$ is coprime with ${}^\F\! B_\xi(t)$, we infer that $\lambda$ is isometric to the linking form
$$\operatorname{\lambda}'(x,y)=\frac{q^2xy^\#}{{}^\F\! B_\xi(t)^{\min(a,b)-k}}.$$
Recalling that $q$ satisfies $q=q^\#$, we can replace the $q^2$ in the numerator by $qq^\#$.
From this, it follows that $\lambda'$ (and thus $\lambda$) is isometric to the pairing $\mathfrak{f}^w((a,b),k,\xi,\mathbb{F})$.
\end{proof}
We can now prove Theorem~\ref{thm:MainLinkingForm} which states that every weakly split linking form is isometric to some linking form of the type \eqref{eq:splitting}.
\subsection{Proof of Theorem~\ref{thm:MainLinkingForm}}
\label{sub:cyclic_pairing}
Let $(M,\lambda)$ be a split linking form.
We can write $M$ as a direct sum $M=M_1\oplus \ldots\oplus M_m$ for some $m$ in such a way that each of the modules $M_i$ is cyclic and the decomposition is orthogonal.
We now claim that each cyclic module $M_i$ decomposes into an orthogonal sum of basic modules (that is, modules of form $\L{\F}/{}^\F\! B_\xi(t)^n$ or $\L{\F}/({}^\F\! B_{\xi+}(t)^a{}^\F\! B_{\xi-}(t)^b)$).
To see this, suppose that $M_i$ is cyclic with $f$ as an annihilator.
Next, factor $f$ as a product of polynomials: $f=u p_1\dots p_r$, where $p_i$ is either ${}^\F\! B_{\xi_i}^{n_i}$ or $p_i={}^\F\! B_{\xi_i+}^{a_i}{}^\F\! B_{\xi_i-}^{b_i}$.
Now for each $i$, set \(q_i = f/p_{i}\).
By the Chinese Remainder Theorem, there are units $u_i$ in~\( \L{\F}/p_{i}\) such that the following map is a ring isomorphism
\begin{align}
\label{eq:Chinese}
\nu &\colon \L{\F} / p_{1} \oplus \L{\F}/p_2 \oplus \ldots \oplus \L{\F}/p_r \xrightarrow{\cong} \L{\F}/f \nonumber \\
& (a_1, a_2, \ldots, a_r) \mapsto a_1 q_1 u_1 + a_2 q_2 u_2+ \ldots + a_r q_r u_r.
\end{align}
Use $\lambda_i$ to denote the restriction of $\lambda$ to $M_i$.
To conclude the proof of the claim, we must show that~(\ref{eq:Chinese}) is orthogonal with respect to the pairing $\lambda_i$.
Since $f$ annihilates $M_i$, we deduce that $\lambda_i(1,1)=Q/f$ for some $Q$ in $\L{\F}$.
Next, use $1_k$ to denote the class of $1$ in $\L{\F}/p_k$.
Using this notation, we compute that
$$\lambda_i(\nu(1_k),\nu(1_m))=\lambda_i(q_k, q_m) = \frac{Q q_k q_m^{\#}}{f} = \frac{Q q_m^{\#}}{p_k}.$$
But since, for $i \neq j$, we know that $p_i$ divides $q_j$, we deduce that~$\lambda_i(\nu(1_k),\nu(1_m))$ vanishes.
Identifying the left hand side cyclic terms in~(\ref{eq:Chinese}) with their isomorphic image in $M_i$, we see that the decomposition is orthogonal.
This concludes the proof of the claim.
Using the claim, we can decompose each summand of $M \cong M_1 \oplus \cdots \oplus M_m$ into an orthogonal sum of basic modules.
The theorem now follows from Lemmas~\ref{lem:structure_lemma},~\ref{lem:structure1},~\ref{lem:evenpairing}, and~\ref{prop:notins1} which classify the basic pairings.
\begin{remark}
Using the notations of Example~\ref{ex:RealBasicXiNotReal}, we note that if the linking form $(M,\operatorname{\lambda})$ is split, then the decomposition
\eqref{eq:splitting} simplifies to the decomposition
\begin{equation}\label{eq:splitting_nonsing}
\bigoplus_{\substack{ n_i,k_i,\epsilon_i,\xi_i\\ i\in I}}\mathfrak{e}(n_i,k_i,\epsilon_i,\xi_i,\mathbb{F})\oplus
\bigoplus_{\substack{n_j,k_j,\xi_j\\ j\in J}}\mathfrak{f}(n_j,k_j,\xi_j,\mathbb{F}).
\end{equation}
\end{remark}
\subsection{Uniqueness of the decomposition}\label{sec:uniqueness}
We now prove the uniqueness statement of Theorem~\ref{thm:MainLinkingForm}.
In the proofs we mostly focus on the case $\mathbb{F}=\mathbb{C}$, because the case $\mathbb{F}=\mathbb{R}$ is known to the experts, however the methods we present can also be used if $\mathbb{F}=\mathbb{R}$.
\medbreak
We start by presenting the key lemma of the proof of uniqueness. We state it for forms that are possibly degenerate, but later we use it only in the non-singular case.
\begin{lemma}\label{lem:linear_of_basic}
For any $n,k$, $\xi\in S^1$, and for any two sequences $\epsilon_1,\ldots,\epsilon_\ell$, $\epsilon'_1,\ldots,\epsilon'_\ell$ of
integers $\epsilon_i,\epsilon'_i=\pm 1$ the existence of an isomorphism
\begin{equation}\label{eq:required_isometry}
\bigoplus_{i=1}^\ell \mathfrak{e}(n,k,\epsilon_i,\xi,\mathbb{F})\simeq\bigoplus_{i=1}^\ell \mathfrak{e}(n,k,\epsilon'_i,\xi,\mathbb{F})
\end{equation}
implies that the sequences $\epsilon_i$ and $\epsilon'_i$ differ by a permutation.
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{lem:linear_of_basic}]
We will first show that in the form~(\ref{eq:myformM}) below, the term $\sum_{i=1}^\ell \epsilon_i$ can be computed as a signature of a certain Hermitian form~$\lambda_{res}$. First of all, we set
\begin{equation}
\label{eq:myformM}
(M,\lambda)=\bigoplus_{i=1}^\ell \mathfrak{e}(n,k,\epsilon_i,\xi,\mathbb{F}).
\end{equation}
We only prove the case $\mathbb{F}=\mathbb{C}$ and $p=t-\xi$ for $|\xi|=1$, leaving the other cases to the reader.
Note that $M$ has a structure of a finite dimensional vector space over~$\mathbb{C}$ with $\dim_\mathbb{C} M=n\ell$.
To describe a $\mathbb{C}$-basis of $M$, note that the $i$-th summand of $M$ is isomorphic to~$\L{\C}/p^n$ and for~$j=0,1,\ldots, n-1$, we define $\nu_{ij}$ to be the class of $p^j$ in this quotient (note in particular that~$(t-\xi)\nu_{ij}=\nu_{i,j+1}$ for~$j<n-1$).
A $\mathbb{C}$-basis for $M$ is now given by the $\nu_{ij}$ for $i=1,\ldots,\ell$ and~$j=0,1,\ldots,n-1$.
Still considering $M$ as $\mathbb{C}$-vector space, we consider the following sesquilinear form on $M$:
\begin{align*}
\wt{\lambda}_{\res}\colon M\times M&\to\mathbb{C}\\
(x,y)\mapsto&
\begin{cases}
((t-\xi)(t^{-1}-\ol{\xi}))^{(n-k)/2}\lambda(x,y)|_{t=\xi} & \textrm{ if } n-k \textrm{ is even,}\\
i\ol{\xi}\res\limits_{t=\xi}((t-\xi)(t^{-1}-\ol{\xi}))^{(n-k-1)/2}\lambda(x,y) & \textrm{ if } n-k \textrm{ is odd.}
\end{cases}
\end{align*}
The form $\wt{\lambda}_{\res}$ is obviously Hermitian if $n-k$ is even.
If $n-k$ is odd, then $\wt{\lambda}_{\res}$ is also Hermitian: the proof of this fact follows essentially the same pattern as the proof of Lemma~\ref{lem:residue}.
Let $M_0$ be the subspace $(t-\xi)M$ which is equal to the linear subspace of $M$ spanned by the $\nu_{ij}$ with $i=1,\ldots,\ell$ and $j>0$.
Set $M_{\res}:=M/M_0$.
As the form $\wt{\lambda}_{\res}$ vanishes on $M_0$, it descends to a Hermitian form $\lambda_{\res}\colon M_{\res}\times M_{\res}\to\mathbb{C}$.
We claim that the signature of $\lambda_{\res}$ is equal to $\sum_{i=1}^\ell \epsilon_i$.
First, observe that the vectors $\nu_{i0}$, $i=1,\ldots,\ell$ give rise to a basis of $M_{\res}$.
Therefore, using $\delta$ for the Kronecker delta, the claim will follow if we manage to show that for each $i=1,\ldots, \ell$, there is a real positive constant~$c_i$ such that
\begin{equation}\label{eq:lambda_res_bas}
\lambda_{\res}(\nu_{i0},\nu_{i'0})=c_i\epsilon_i\delta_{ii'}.
\end{equation}
To see this, we start by observing that if $i\neq i'$, then $\lambda_{\res}(\nu_{i0},\nu_{i'0})=0$: indeed $\nu_{i0}$ and $\nu_{i'0}$
belong to different orthogonal components of the sum \eqref{eq:myformM}.
The announced value of $\lambda_{\res}(\nu_{i0},\nu_{i'0})$ for $i=i'$ follows from the definition of $\lambda$ (recall~\eqref{eq:myformM} and Example~\ref{BasicPairingXiS1Complex}).
In more details, if $n-k$ is even, then \eqref{eq:e_n_k_form_complex_even} implies that $c_i=1$, while if $n-k$ is odd, \eqref{eq:e_n_k_form_complex_odd} implies that
\[\lambda_{\res}(\nu_{i0},\nu_{i0})=i\ol{\xi}\res_{t=\xi}\left( \frac{r(t)}{(t-\xi)} \right) \epsilon_i ,\]
where $r(t)$ is a $\xi$-positive linear polynomial.
By Lemma~\ref{lem:residue}, we know that $\ol{\xi}\res_{t=\xi} \left( \frac{r(t)}{(t-\xi)} \right)$ is
a real negative multiple of $i$.
As a consequence, $i\ol{\xi}\res_{t=\xi} \left( \frac{r(t)}{(t-\xi)} \right)$ is positive.
This concludes the proof of the claim.
The pairing $\lambda_{\res}$ and its signature were defined intrinsically, without referring to any choice of a presentation,
so they are invariants of the isometry type of $(M,\lambda)$.
Using this fact and our hypothesis (recall~\eqref{eq:required_isometry}), we deduce that $\sum\epsilon_i=\sum\epsilon_i'$.
Thus the sequences $\epsilon_i$ and $\epsilon'_i$ differ by a permutation, concluding the proof of the proposition.
\end{proof}
We can now prove the uniqueness of the decomposition~\eqref{eq:splitting}.
\begin{proposition}\label{prop:uniqueness}
If $(M,\operatorname{\lambda})$ is a non-singular linking form over $\L{\R}$ or $\L{\C}$, then the decomposition \eqref{eq:splitting} is unique up to permuting factors.
\end{proposition}
\begin{proof}
Clearly, it is enough to prove the uniqueness of decomposition \eqref{eq:splitting} if $M$ is $p$-primary for some basic polynomial $p$.
We have yet another simplification, namely if $p$ has roots outside of unit circle, the linking
form is determined up to isometry by the underlying module structure.
In that case, uniqueness of the decomposition of a linking form boils down to the uniqueness of the underlying module.
Therefore the main difficulty arises in the case where $p$ has roots on the unit circle. In this proof, we focus on the case $\mathbb{F}=\mathbb{C}$ (so that~$p=t-\xi$) leaving the case $\mathbb{F}=\mathbb{R}$ to the reader.
Assume that $\phi \colon (M,\lambda) \to (M',\lambda')$ is an isometry of non-singular linking forms over~$\L{\F}$, with $\mathbb{F}$ either $\mathbb{R}$ or $\mathbb{C}$.
Our goal is to show that the decompositions of $M$ and $M'$ described in \eqref{eq:splitting} are the same up to permuting factors.
Fix a decomposition \eqref{eq:splitting} of $M$ and let $m=\max n_i$ be the maximal order of an element in $M$.
Note that since the pairings are non-singular, we have $k_i=0$ for all $i$.
Write $M=M_1\oplus\dots\oplus M_m$, where
\[M_s=\bigoplus_{n_i=s}\mathfrak{e}(n_i,0,\epsilon_i,\xi,\mathbb{F}),\]
and decompose $M'$ in an analogous way.
For each $s$, pick a $\L{\C}$-basis $(x_{s1},\dots,x_{sl_s})$ (resp. $y_{s1},\dots,y_{sl_s}$) for $M_s$ (resp. $M'_s$) such that on each of the generators, the pairing $\lambda$ (resp. $\lambda'$) has the form \eqref{eq:e_n_k_form_complex_even} or \eqref{eq:e_n_k_form_complex_odd}.
We insist that in the latter case, the $\xi$-positive polynomial $r(t)$ is common for all the $x_{si}$ and $y_{sj}$.
Reformulating, each of the $x_{si}$ (resp. $y_{sj}$) is a generator of one of the basic modules entering the decomposition of $M_s$ (resp.$M'_s$).
Note furthermore that the bases of~$M_s$ and~$M'_s$ have the same length because the ranks of $M_s$ and~$M'_s$ are equal.
Using these bases, the isometry $\phi$ is represented by a block matrix of maps $\phi_{qs}\colon M_q\to M'_s$.
The map $\phi_{qq}$ is then written as
\[\phi_{qq}=C_q+\ldots,\]
where $C_q$ is a constant matrix
and the dots denote higher order terms in~$(t-\xi)$.
It now suffices to show that $C_q$ is an isometry between $M_q$ and $M'_q$: indeed the proof of Proposition~\ref{prop:uniqueness} would then be concluded by invoking Lemma~\ref{lem:linear_of_basic}.
To see that $C_q$ is an isometry, we use $U_q$ (resp. $V_q$) to denote $M_q/pM_q$ (resp. $M'_q/pM'_q$) and let $\pi \colon M_q \to U_q$ (resp. $\pi' \colon M_q' \to V_q$) denote the corresponding projection. Observe that $U_q$ and $V_q$ can be regarded as $\mathbb{C}$-vector spaces with respective bases given by the $e_{qi}=\pi(x_{qi})$ and $f_{qi}=\pi'(y_{qi})$ for $i=1,\ldots,l_q$.
Define a $\mathbb{C}$-valued Hermitian pairing on $U_q$ by $\langle e_{qi},e_{qj}\rangle =\epsilon\delta_{ij}$, where $\epsilon$ is the sign of the summand of~$\mathfrak{e}(n_i,0,\xi,\epsilon,\mathbb{C})$ generated by $x_{qi}$. We define a Hermitian pairing on $V_q$ in the exact same way and note that these Hermitian pairings can be regarded as approximations of $\lambda$ and $\lambda'$ in $U_q$ and $V_q$.
More precisely, assuming that $q$ is even, using the definition of $\lambda$ and of the basic pairings (recall \eqref{eq:e_n_k_form_complex_even} and \eqref{eq:e_n_k_form_complex_odd}), we have
\begin{equation}\label{eq:lambdaxqi}\lambda(x_{qi},x_{qj})=\frac{\langle e_{qi},e_{qj}\rangle_{U_q}}{(t-\xi)^{q/2}(t^{-1}-\ol{\xi})^{q/2}}+\dots,\end{equation}
where the dots denote higher order terms in $t-\xi$. Analogously, for $q$ odd, we have
\[\lambda(x_{qi},x_{qj})=\frac{\langle e_{qi},e_{qj}\rangle_{U_q} r(t)}{(t-\xi)^{(q+1)/2}(t^{-1}-\ol{\xi})^{(q-1)/2}}+\dots.\]
Using the $y_{qi}$ and $f_{qi}$ instead of the $x_{qi}$ and $e_{qi}$, the Hermitian pairing on $V_q$ has the same property.
We claim that if $s\neq q$, then the largest possible order of a pole of $\lambda'(\phi_{qs}x_{qi},\phi_{qs}x_{qj})$ at $t=\xi$ is~$q-1$, i.e. the $(t-\xi)^j$ terms in the Laurent expansion of $\lambda'(\phi_{qs}x_{qi},\phi_{qs}x_{qj})$ satisfy $j\ge 1-q$.
If $s<q$, the claim is obvious, because the pairing of any two elements on $M_s$ has a pole of order at most~$s$ at $\xi$; reformulating, for any $\mu_1,\mu_2\in M'_s$, the expansion of $\lambda'(\mu_1,\mu_2)$ in Laurent series in $(t-\xi)$ has terms $(t-\xi)^j$ with $j\ge -s$.
Suppose that $s>q$. Observe that since $\phi_{qs}x_{qi}$ and $\phi_{qs}x_{qj}$ are annihilated by $(t-\xi)^q$ in~$M'_s$, each of these elements is a $p^{s-q}$ multiple of some $\nu_{si},\nu_{sj}\in M'_s$, i.e. $\phi_{qs}x_{qi}=p^{s-q}\nu_{si}$ and $\phi_{qs}x_{qj}=p^{s-q}\nu_{sj}$. This means that
$\lambda'(\phi_{qs}x_{qi},\phi_{qs}x_{qj})=p^{2(s-q)}\lambda'(\nu_{si},\nu_{sj})$. Now the order of the pole of~$\lambda'(\nu_{si},\nu_{sj})$ is smaller or equal to $s$
(i.e. its expansion in
Laurent power series at $t=\xi$ contains terms $(t-\xi)^j$ with $j\ge -s$). Thus, the terms $(t-\xi)^j$ in the Laurent expansion of $\lambda'(\phi_{qs}x_{qi},\phi_{qs}x_{qj})$ satisfy $j\ge 2(s-q)-s=(s-q)-q>-q$. This proves the claim.
We now return to the proof that $C_q$ is an isometry. Using the definition of the $\phi_{qs}$ and the orthogonality of $\lambda'$ on the decomposition of $M'$, we have
\[\lambda'(\phi x_{qi},\phi x_{qj})=\sum_{s}\lambda'(\phi_{qs} x_{qi},\phi_{qs} x_{qj}).\]
For $s \neq q$, the claim implies that the terms in this sum have poles of order less than $q$. Assume $q$ is even (the odd case is completely analogous). Using dots to denote higher order terms in $(t-\xi)$
(terms with $(t-\xi)^j$ for $j>-q$), we have
\begin{equation}\label{eq:lambdaxip}\lambda'(\phi x_{qi},\phi x_{qj})=\lambda'(\phi_{qq} x_{qi},\phi_{qq}x_{qj})+\ldots=
\frac{\langle C_q e_{qi},C_qe_{qj}\rangle_{V_q}}{(t-\xi)^{q/2}(t^{-1}-\ol{\xi})^{q/2}}+\dots,
\end{equation}
where the second equality follows from the definition of $C_q$ as the linearization of $\phi_{qq}$ and \eqref{eq:lambdaxqi} applied for $\lambda'$.
As $\lambda'(\phi x_{qi},\phi x_{qj})=\lambda(x_{qi},x_{qj})$, applying \eqref{eq:lambdaxqi} to \eqref{eq:lambdaxip}
shows that
$C_q$ is an isometry between $U_q$ and $V_q$. We can now extend $C_q$ to a $\L{\C}$-linear map
$\Phi\colon M_q\to M'_q$ by putting $\Phi(x_{qi})=\sum c_{qij}x_{qi}$ on the generators and extending $\L{\C}$-linearly to the whole of $M_p$, where
$C_q=\{c_{qij}\}_{i,j=1}^{l_q}$.
This map is an isometry, because $C_q\colon U_q\to V_q$ was an isometry. In particular, $M_q$ and $M'_q$ are isometric and
Lemma~\ref{lem:linear_of_basic} concludes the proof.
\end{proof}
Given the uniqueness result we can make the following definition.
\begin{definition}\label{def:hodge_number}
Let $n$ be a positive integer, and let $\epsilon$ be~$\pm 1$. If $|\xi|=1$ (resp. $|\xi|\in(0,1)$), the \emph{Hodge number}
$\mathcal{P}(n,\epsilon,\xi)$ (resp. $\mathcal{Q}(n,\xi)$)
of a non-singular linking form $(M,\operatorname{\lambda})$
is the number of times $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{F})$ (resp. $\mathfrak{f}(n,0,\xi,\mathbb{F})$) enters the sum \eqref{eq:splitting}.
\end{definition}
From Proposition~\ref{prop:uniqueness} we deduce the following result.
\begin{corollary}\label{prop:hodge_are_inv}
The Hodge numbers are invariants of non-singular linking forms.
\end{corollary}
The definition of the Hodge numbers is essentially the same as in \cite{Ne95} (in that paper, one takes $\mathbb{F}=\mathbb{R}$) with one exception, namely $\xi=1$.
The definition in \cite{Ne95} of the Hodge numbers for $\xi=1$ is different and,
with N\'emethi's definition, it might happen that $\mathcal{P}(n,\epsilon,1,\mathbb{R})\neq 0$ even if $n$ is odd.
The calculation of N\'emethi's Hodge numbers for $\xi=1$ is usually more involved. Often the value $\xi=1$ deserves a special
treatment both in singularity theory and in knot theory.
\section{Further properties of linking forms}
\label{sec:FurtherProperties}
In this section, we gather some additional results on linking forms. First, in Subsection~\ref{sub:NonSplit} we provide an example of a non weakly split linking form.
Subsection~\ref{sec:forms_cyclic} contains an explicit way of finding a decomposition of a linking form over a cyclic module into basic forms. Applications of this result appear in Section~\ref{sec:blanchf-forms-twist}: it it is used to calculate twisted signatures of some linear combinations of torus knots.
Subsection~\ref{sub:LocalizationDiagonalization} proves a classification of forms over the local ring $\mathcal{O}_\xi$. Together with results of Subsection~\ref{sec:loc2},
theses techniques are a key tool to prove Proposition~\ref{prop:JumpIsJump} relating two types of signature invariants. Finally, Sections~\ref{sec:Representability} and~\ref{sec:rep2} introduce and discuss the notion of a representable linking form.
\subsection{An example of a non weakly split linking form}
\label{sub:NonSplit}
As we stated in Theorem~\ref{thm:non-degenerate-split}, non-singular linking forms are split. The case of singular pairings is more subtle. In this subsection, we give an example indicating that not all singular linking forms are weakly split.
\begin{example}
\label{ex:NonSplit}
Consider $\L{\F}$ for $\mathbb{F}=\mathbb{R}$ or $\mathbb{C}$ and let $p$ be an irreducible symmetric polynomial, for instance, one could take $\mathbb{F}=\mathbb{R}$ and $p={}^\R\! B_\xi(t)$ for $\xi\in S^1$ with $\iim(\xi)>0$.
We let
$M=\L{\F}/p^5\L{\F}\oplus\L{\F}/p^4\L{\F}$, use $\alpha $ and $\beta$ to denote the generators of the two summands and consider the linking form given by
\begin{equation}\label{eq:example_pairing}
\operatorname{\lambda}(\alpha,\alpha)=\frac{1}{p},\ \operatorname{\lambda}(\beta,\beta)=\frac{1}{p^2},\ \operatorname{\lambda}(\alpha,\beta)=\frac{1}{p^3}.
\end{equation}
The remainder of this example is devoted to showing that $\lambda$ is not split. First, observe that $p^3\operatorname{\lambda}(x,y)=0$ for all $x,y\in M$. Next, note that taking $x=\alpha$, $y=\beta$ guarantees the existence of a pair of elements $x,y\in M$ such that
\begin{equation}\label{eq:psquarexy}
p^2\operatorname{\lambda}(x,y)\neq 0.
\end{equation}
Assume that the pairing is weakly split. Then we can find a presentation $M=\L{\F}/p^5\L{\F}\oplus\L{\F}/p^4\L{\F}$ with generators $\alpha'$ and $\beta'$
such that $\operatorname{\lambda}(\alpha',\beta')=0$. By Theorem~\ref{thm:MainLinkingForm} we may assume that
$\operatorname{\lambda}(\alpha',\alpha')=\frac{\epsilon_1}{p^k}$ and $\operatorname{\lambda}(\beta',\beta')=\frac{\epsilon_2}{p^{k'}}$ for some $k$ and $k'$ and
$\epsilon_1,\epsilon_2=\pm 1$. Clearly~$k,k'\le 3$.
Write the elements $x$ and $y$ satisfying \eqref{eq:psquarexy} in the basis $\alpha',\beta'$, so that $x=x_1+x_2$, $y=y_1+y_2$,
where $x_1,y_1$ are multiples of $\alpha'$ and $x_2,y_2$ are multiples of $\beta'$. By orthogonality of $\alpha'$ and $\beta'$, we have $\operatorname{\lambda}(x,y)=
\operatorname{\lambda}(x_1,y_1)+\operatorname{\lambda}(x_2,y_2)$. Since \eqref{eq:psquarexy} holds for $x$ and $y$, it must either hold for the pair~$x_1,y_1$, or for
the pair $x_2,y_2$. Then at least one of the $k,k'$ must be equal to $3$. Without losing generality we assume that $k=3$.
Consequently writing $\alpha'=a\alpha+b\beta$ and $\beta'=a'\alpha+b'\beta$, we obtain
\begin{equation}
\label{eq:ContradictionNonSplit}
\operatorname{\lambda}(\alpha',\beta')=\frac{a{a'}^\#}{p}+\frac{b{b'}^\#}{p^2}+\frac{a{b'}^\#+{a'}^\#b}{p^3}.
\end{equation}
We now claim that $p$ divides $a'$ and $p$ does not divide $a,b,b'$. Before proving this claim, observe that this implies that $p$ does not divide the sum $a{b'}^\#+{a'}^\#b$. In particular, using~(\ref{eq:ContradictionNonSplit}), we deduce that $\operatorname{\lambda}(\alpha',\beta')$ cannot be zero, contradicting the orthogonality of $\alpha'$ and $\beta'$.
We now prove the claim. First, observe that since we showed that $k=3$, we have $p^2\operatorname{\lambda}(\alpha',\alpha')\neq~0$. Via \eqref{eq:example_pairing}, this translates into $p\nmid (ab^\#+a^\# b)$. In particular,~$p$ divides neither $a$
nor~$b$. Next, we show that $p|a'$. As $p^4\alpha\neq 0$ and $p\nmid a$, we have $p^4\alpha'\neq 0$. Observe that $p^4\beta'=0$: by the definition of $M$, we have $p^4M=\L{\F}/p \L{\F}$, so if $p^4\beta'\neq 0$, then
it would be linearly dependent with $p^4\alpha'$, contradicting the fact that~$\alpha'$ and~$\beta'$ are generators of $M$. Consequently, since $p^4 \alpha \neq 0$, while $p^4 \beta =0$ and $p^4 \beta'=0$, the definition of $\beta'$ implies that $p|a'$. Finally, we show that $p\nmid b'$. If we had $p|b'$, the fact that $p|a' $ and the definition of $\beta'$ would imply that $\beta'=pz$ for
some $z\in M$. This contradicts the fact that $\beta'$ is a generator. We have therefore shown that $\lambda$ is not a weakly split linking form.
\end{example}
\subsection{Forms over cyclic modules}\label{sec:forms_cyclic}
In this subsection, we state a consequence of Theorem~\ref{thm:MainLinkingForm} which has practical applications
in finding the decomposition~\eqref{eq:splitting}.
\medbreak
Suppose $M$ is a cyclic $\L{\F}$-module, that is $M=\L{\F}/f$ for some polynomial $f$.
We use~$1_M$ to denote the generator of $M$ which is obtained as the image of $1\in\L{\F}$ under the projection map $\L{\F}\to\L{\F}/f$. Given a linking form $\lambda\colon M\times M\to\O{\F}/\L{\F}$, we observe that
\[\lambda(1_M,1_M)=\frac{h}{f}\in\O{\F}/\L{\F}\]
for some $h$.
It is convenient to think of $h$ as an element in $\L{\F}$ defined modulo $f$. Also, note that $\frac{h}{f}$ is a symmetric rational function.
The following result gives a precise description of the isometry type of $\lambda$ in terms of $h$ and $f$.
\begin{proposition}\label{prop:cyclic_classif}
Let $\xi\in S^1$ be a root of $f$ of order $n>0$ and let ${}^\F\! B_\xi(t)$ be a basic polynomial having root at $\xi$. Consider the restriction $\lambda_\xi$ of $\lambda$
to the direct summand~$M_\xi=\L{\F}/{}^\F\! B_\xi(t)^n$ of~$M$. Then~$(M_\xi,\lambda_\xi)$ is isometric to $\mathfrak{e}(n,k,\xi,\epsilon,\mathbb{F})$, where
$k$ is the order of $h$ at $t-\xi$ and
\begin{itemize}
\item if $n-k$ is even, then $\epsilon$ is equal to $+1$ if $q=h({}^\F\! B_\xi(t){}^\F\! B_\xi(t)^\#)^{(n-k)/2}/f$ is positive near $\xi$, $\epsilon=-1$ if $q$ is negative near $\xi$;
\item if $n-k$ is odd and $\mathbb{F}=\mathbb{R}$, then $\epsilon=+1$ if $q'=h{}^\F\! B_\xi(t)^n/f$ is positive near $\xi$, otherwise $\epsilon=-1$;
\item if $n-k$ is odd and $\mathbb{F}=\mathbb{C}$, then $\epsilon=+1$ if $q''=h(t-\xi)^{(n-k+1)/2}(t^{-1}-\ol{\xi})^{(n-k-1)/2}/f$ is $\xi$-positive, otherwise $\epsilon=-1$.
\end{itemize}
\end{proposition}
\begin{proof}
Since the three items are proved in a very similar way, we only give the proof of the first one in the case where $\mathbb{F}=\mathbb{R}$ (so that ${}^\F\! B_\xi(t)^\#={}^\F\! B_\xi(t)$).
Write $q=\frac{h{}^\F\! B_\xi(t)^{n-k}}{f}$. So that $q(\xi)\neq 0$ and we have $\operatorname{\lambda}(1_M,1_M)=\frac{q}{{}^\F\! B_\xi(t)^{n-k}}$.
The natural injection $M_\xi\to M$ is given
by $1_\xi\mapsto g\cdot 1_M$, where $g=f/{}^\F\! B_\xi(t)^n$. Let $1_\xi$ be the image of $1$ under the projection $\L{\F}\to M_\xi$. The restricted pairing $\lambda_\xi$ is now isometric to the following pairing:
\[ (1_\xi,1_\xi) \mapsto gg^\#\frac{h}{f}=\frac{qgg^\#}{{}^\F\! B_\xi(t)^{n-k}}.\]
Now $g$ is invertible in $\L{\F}/{}^\F\! B_\xi(t)^{n-k}$, so the pairing $\lambda_\xi$ is isometric to the pairing
\[ (1_\xi,1_\xi) \mapsto \frac{q}{{}^\F\! B_\xi(t)^{n-k}}.\]
It follows from the classification results in Subsection~\ref{sub:sign_of_four} that $(M_\xi,\lambda_\xi)$ is isometric to $\mathfrak{e}(n,k,\xi,\epsilon,\mathbb{F})$
where $\epsilon$ is the sign of $q(\xi)$. But this is precisely the statement of the proposition.
\end{proof}
\subsection{Localization}
\label{sub:LocalizationDiagonalization}
The techniques used in the proof of Theorem~\ref{thm:MainLinkingForm} can also be applied to study forms on the local ring of germs of holomorphic functions
near $\xi$,
that is, forms over the ring $\mathcal{O}_\xi$ of
analytic functions near $\xi$ for $\xi\in\mathbb{C}$, i.e. over the ring of convergent power series $\sum_{i \geq 0} a_i(t-\xi)^i$.
\medbreak
Throughout this subsection, we assume that $\mathbb{F}=\mathbb{C}$ and $\xi\in S^1$. Write $\Omega_\xi$ for the field of fractions of $\mathcal{O}_\xi$. As alluded to in Remark~\ref{rem:positive}, the ring $\mathcal{O}_\xi$ (and therefore $\Omega_\xi$) has an involution~$r\mapsto~r^\#$.
\begin{proposition}\label{prop:classify_cyclic}
If $\wh{M}$ is a cyclic $\mathcal{O}_\xi$--module, then any linking form $\wh{\lambda}\colon\wh{M}\times\wh{M}\to\Omega_\xi/\mathcal{O}_\xi$ is isometric to
\begin{align*}
\mathcal{O}_\xi/(t-\xi)^n\times\mathcal{O}_\xi/(t-\xi)^n&\to \Omega_\xi/\mathcal{O}_\xi\\
x,y&\mapsto \frac{\epsilon r(t) xy^\#}{(t-\xi)^{n-k}},
\end{align*}
where $n,k$ are non-negative integers, $\epsilon=\pm 1$ and $r(t)$ is such that
\begin{itemize}
\item[(a)] if $n-k$ is even, then $\frac{r(t)}{(t-\xi)^{n-k}}=\frac{1}{((t-\xi)(t^{-1}-\ol{\xi}))^{(n-k)/2}}$;
\item[(b)] if $n-k$ is odd, then $\frac{r(t)}{(t-\xi)^{n-k}}=\frac{\wt{r}(t)}{(t-\xi)((t-\xi)(t^{-1}-\ol{\xi}))^{(n-k-1)/2}}$, where $\wt{r}(t)$ is linear and $\xi$-positive.
\end{itemize}
\end{proposition}
\begin{proof}
The statement resembles Lemma~\ref{lem:evenpairing}, but the proof is simpler, because $\mathcal{O}_\xi$ is a local ring.
We invoke the following well known fact from algebra. We include a short proof for completeness.
\begin{lemma}
Any torsion cyclic module over $\mathcal{O}_\xi$ is of the form $\mathcal{O}_\xi/(t-\xi)^n$ for some $n>0$.
\end{lemma}
\begin{proof}
If $f$ is an annihilator of a cyclic module $N$ over $\mathcal{O}_\xi$, then we write $f=(t-\xi)^n\wt{f}$, where~$\wt{f}$ is an analytic function such that $\wt{f}(\xi)\neq 0$. Then $\wt{f}$ is
invertible in $\mathcal{O}_\xi$ and so the module is isomorphic to $\mathcal{O}_\xi/(t-\xi)^n$.
\end{proof}
Continuing with the proof of Proposition~\ref{prop:classify_cyclic}, we suppose that $n-k$ is odd (the even case is analogous). Identifying $\wh{M}$ with $\mathcal{O}_\xi/(t-\xi)^n$, there are
polynomials~$q$ (with $q(\xi) \neq 0$) and~$\widetilde{q}$ which satisfy
$$\wh{\lambda}(1,1)=\frac{q}{(t-\xi)^{n-k}}=\frac{\wt{q}}{(t-\xi)((t-\xi)(t^{-1}-\ol{\xi}))^{(n-k-1)/2}}.$$
Arguing as in the proof of Lemma~\ref{lem:qlinear}, we can choose a $\xi$-positive linear polynomial $\wt{r}$ such that $p:=\wt{q}\wt{r}^\#$ is symmetric and $p(\xi)\neq 0$. Using this fact, we pick $\epsilon=\pm 1$ such that $\epsilon p(\xi)>0$.
We claim that there is a rational function $g$ such that $ \epsilon p=g g^\#$. As $\epsilon p$ is an analytic function which is positive at $\xi$, there exists a rational function $g\in\mathcal{O}_\xi$ such that $\epsilon p=g^2$. As~$\epsilon p=(\epsilon p)^{\#}$, we also have $\epsilon p={g^\#}^2$, and so $g=\pm g^\#$. But since $g(\xi)=g^{\#}(\xi)$, we deduce that $g=g^\#$ and hence~$\epsilon p=gg^\#$, concluding the proof of the claim.
Using the claim, we see that $\epsilon \widetilde{q}\widetilde{r}^\#\widetilde{r} =\epsilon p \widetilde{r} =gg^\#\widetilde{r} $. It follows that multiplication by~$g$ and~$\widetilde{r}$ in~$\wh{M}$ induces an isometry between $\widehat{\lambda}$ and the pairing
$(x,y)\mapsto \frac{\epsilon\wt{r}}{(t-\xi)((t-\xi)(t^{-1}-\ol{\xi}))^{(n-k-1)/2}}$ (note that~$\widetilde{r}$ and $g$ are invertible in $\widehat{M}$).
This concludes the proof of the proposition.
\end{proof}
From now on, we use $\wh{\mathfrak{e}}(n,k,\epsilon,\xi,\mathbb{C})$ to denote the linking form described in Proposition~\ref{prop:classify_cyclic}. Furthemore, given a linking form $(M,\lambda)$ over $\L{\C}$, we let $(\wh{M}_\xi,\wh{\lambda}_\xi)$ denote the linking form
$$ (\wh{M}_\xi,\wh{\lambda}_\xi)=(M,\lambda)\otimes_{\L{\C}}\mathcal{O}_\xi.$$
In other words, the linking form $\widehat{\lambda}$ takes values in $\Omega_\xi/\mathcal{O}_\xi$. We conclude this subsection with the following observation which we shall use in Subsection~\ref{sec:JumpIsJump}.
\begin{proposition}\label{prop:tensor}
Suppose that $(M,\lambda)$ is a split linking form over $\L{\C}$ and choose $\xi\in S^1$. The~$(t-\xi)$-primary summand of~$(M,\lambda)$ is isometric to $\oplus_{i\in I} \mathfrak{e}(n_i,k_i,\xi,\epsilon_i,\mathbb{C})$ if and only if we have the following~isometry:
\begin{equation}
\label{eq:LocalizedGuy}
(\wh{M}_\xi,\wh{\lambda}_\xi) \cong \bigoplus_{i\in I} \wh{\mathfrak{e}}(n_i,k_i,\xi,\epsilon_i,\mathbb{C}).
\end{equation}
\end{proposition}
\begin{proof}
The proposition follows from the following observation: if $\eta\neq \xi$, then multiplication by~$t-\eta$ is an isomorphism of $\mathcal{O}_\xi$
and, consequently, tensoring by $\mathcal{O}_\xi$ kills the $p$-primary part of $M$ for any basic polynomial $p$ coprime with $t-\xi$.
\end{proof}
\subsection{Representable linking forms}
\label{sec:Representability}
A special class of linking forms is formed by so-called representable linking forms. These linking forms will be frequently used in Section~\ref{sec:Signatures}.
\begin{definition}
\label{def:RepresentBlanchfield}
A non-singular linking form $(M,\lambda)$ over $R$ is \emph{representable} if there exists a non-degenerate Hermitian matrix $A$ over $R$ such that
$(M,\lambda)$ is isometric to $(R^n/A^T R^n,\lambda_A)$, where the latter linking form is defined by
\begin{align*}
\lambda_{A} \colon R^n /A^TR^n \times R^n/A^T R^n &\to Q/R \\
([x],[y]) & \mapsto x^TA^{-1}\makeithash{y}.
\end{align*}
In this case, we say that the Hermitian matrix $A$ \emph{represents} the linking form $(M,\lambda)$.
\end{definition}
While the terminology appears to be novel, we observe that the concept of representability has already frequently appeared in the literature~\cite{FriedlThesis, BorodzikFriedl0, BorodzikFriedl, BorodzikFriedl2, ConwayFriedlToffoli,ConwayBlanchfield}. However since the conventions sometimes vary, we take a moment for some remarks on the pairing $\lambda_A$.
\begin{remark}
\label{rem:RepresentableWellDefined}
The pairing $\lambda_A$ is well defined since we used $A^T$ in the presentation of the module, since our pairings are anti-linear in the \emph{second} variable and since, by definition, a matrix $A$ is Hermitian if $\makeithashT{A}=A$, see also~\cite[Proposition 4.2]{ConwayBlanchfield}. Note that different conventions were used in~\cite{BorodzikFriedl}, the module was given by $M=R^n/AR^n$, but the pairing was anti-linear in the \emph{first} variable.
\end{remark}
We will now address the question of which linking forms are representable.
First of all, note that there is a significant difference between representability of real and complex linking forms. Indeed, while in the real case it is known that any non-singular linking form is representable~\cite{BorodzikFriedl2}, we shall see that this does not hold in the complex case. Before expanding on the latter case, we state the former result for future reference.
\begin{proposition}\label{prop:diagonalreal}
For every non-singular linking form over $\L{\R}$, there exists a diagonal Hermitian matrix $A$ over $\L{\R}$ representing this form.
\end{proposition}
\begin{remark}\label{rem:diagonalreal}
Proposition~\ref{prop:diagonalreal} is proved in \cite{BorodzikFriedl2} with an extra assumption: the multiplication by~$t\pm 1$ is required to be an isomorphism i.e. the modules have no~$(t\pm 1)$-torsion. However, this assumption is easily overcome because the form $\mathfrak{e}(n,0,\pm 1,\epsilon,\mathbb{R})$ is representable by a $1\times 1$ matrix $((t\mp 2+t^{-1})^{n/2})$; here note that $n$ must be even by Lemma~\ref{lem:structure1}.
\end{remark}
In order to build some intuition in the complex case, let us start with some examples of representable linking forms.
\begin{proposition}\label{prop:EasyRepresent}
The following linking forms are representable:
\begin{enumerate}
\item the basic linking form $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{C})$ for $n$ even,
\item the basic linking form $\mathfrak{f}(n,0,\xi,\mathbb{C})$,
\item the direct sum $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{C})\oplus \mathfrak{e}(n',0,-\epsilon,\xi',\mathbb{C})$, for $n,n'$ odd and distinct $\xi,\xi'\in S^1$.
\end{enumerate}
\end{proposition}
\begin{proof}
We begin with the representability of the first two linking forms. For $n$ even, the linking form $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{C})$ is represented by a $1\times 1$ matrix $\epsilon((t-\xi)(t^{-1}-\ol{\xi}))^{n/2}$. The pairing $\mathfrak{f}(n,0,\xi,\mathbb{C})$ is represented by $((t-\xi)(t^{-1}-\ol{\xi}))^n$.
We now prove the representability of the third linking form.
Without loss of generality, suppose that~$n>n'$.
Let $p$ be an elementary symmetric polynomial (see Definition~\ref{def:quadratic}) with roots at $\xi$ and~$\xi'$.
Set $\wt{p}=((t-\xi)(t^{-1}-\ol{\xi}))^{(n-n')/2}p^{n'}$. The $1\times 1$ matrix $(\wt{p})$ represents a non-singular pairing
over
the module $M=\L{\C}/(t-\xi)^n\L{\C} \oplus \L{\C}/(t-\xi')^{n'}\L{\C}$.
Using Proposition~\ref{prop:cyclic_classif} (or Proposition~\ref{prop:sumofjumps} below), we see that the pairing is
isometric to $\mathfrak{e}(n,0,\wt{\epsilon},\xi)\oplus \mathfrak{e}(n',0,-\wt{\epsilon},\xi')$ for some~$\wt{\epsilon}=\pm 1$. If $\epsilon\neq\wt{\epsilon}$, then we replace the matrix $(\wt{p})$ by~$(-\wt{p})$.
\end{proof}
Next, we give an example of non-representable linking form.
\begin{proposition}\label{prop:verystupidexample}
Let $\xi$ be a complex number satisfying $|\xi|=1$.
For each positive $n$, the basic linking form $\mathfrak{e}(2n+1,0,\epsilon,\xi,\mathbb{C})$ is not representable.
\end{proposition}
\begin{proof}
Suppose $A(t)$ is a matrix representing $\mathfrak{e}(2n+1,0,\epsilon,\xi,\mathbb{C})$. Then the determinant of $A(t)$ is the order of the module $\L{\C}/(t-\xi)^{2n+1}$, so $\det A(t)=u(t-\xi)^{2n+1}$ for some unit $u\in\L{\C}$. On the other hand, since $A(t)$ is Hermitian, its determinant is symmetric. The result follows by noting that there is no symmetric polynomial in $\L{\C}$ that is equal to $(t-\xi)^{2n+1}u$ for some unit~$u\in\L{\C}$.
\end{proof}
We end this section with an example which will be used in the proof of Proposition~\ref{prop:inversejumps} below.
\begin{proposition}\label{prop:notsostupidexample}
Given $\xi \in S^1$, the linking form $\mathfrak{e}(1,0,+1,\xi,\mathbb{C})\oplus\mathfrak{e}(1,0,-1,\xi,\mathbb{C})$ is representable, but not by a diagonal matrix.
\end{proposition}
\begin{proof}
First of all, if a diagonal matrix $A$ represents a form $(M,\lambda)$, and $A$ has $b_1(t),\ldots,b_n(t)$ on its diagonal, then the underlying module structure is $M=\L{\C}/b_1(t)\oplus\ldots\oplus\L{\C}/b_n(t)$ and each of the $b_i(t)$ is a symmetric polynomial.
Assume by way of contradiction that $\mathfrak{e}(1,0,+1,\xi,\mathbb{C})\oplus\mathfrak{e}(1,0,-1,\xi,\mathbb{C})$ is representable by a diagonal matrix and let $M$ be its underlying module. We know that $M \cong \L{\C}/(t-\xi)\oplus \L{\C}/(t-\xi)$. This indicates that two of the $b_i(t)$ must be divisible by $(t-\xi)$. Suppose that~$(t-\xi)$ divides $b_1(t)$. Since $b_1(t)$ is symmetric, Proposition~\ref{prop:factorcomplex} implies that $b_1(t)$ is also divisible by $(t-\eta)$ for some $\eta\in S^1$. But then, either~$M$ has a summand $\L{\C}/(t-\eta)^k$ for some $k>0$ (if $\eta\neq \xi$), or~$M$ has a summand $\L{\C}/(t-\xi)^k$, $k>2$ (if $\eta=\xi$). Both cases yield a contradiction.
Next, we show that $\mathfrak{e}(1,0,+1,\xi,\mathbb{C})\oplus\mathfrak{e}(1,0,-1,\xi,\mathbb{C})$ is representable. Choose four complex numbers $a,b,c,d$ and consider the matrix
\begin{equation}\label{eq:trickymatrix}
A=\begin{pmatrix}
at^{-1}-\ol{a\xi} & dt^{-1}+c\\
-\ol{c\xi}t^{-1}-\ol{d\xi} & bt^{-1}-\ol{b\xi}
\end{pmatrix}.
\end{equation}
A direct computation shows that $(t-\xi)A$ is Hermitian.
Now we calculate the determinant of $A$:
\begin{align*}
\det A&=(at^{-1}-\ol{a\xi})(bt^{-1}-\ol{b\xi})+(dt^{-1}+c)(\ol{c\xi}t^{-1}+\ol{d\xi})\\
&=t^{-2}(ab+d\ol{c\xi})-t^{-1}\ol{\xi}(a\ol{b}+b\ol{a}-d\ol{d}-c\ol{c})+(\ol{ab\xi^2}+c\ol{d\xi}).
\end{align*}
Suppose $a,b,c,d$ are such that
\begin{equation}\label{eq:neweq}
ab=-d\ol{c\xi} \textrm{ and }2\re(a\ol{b})\neq|c|^2+|d|^2.
\end{equation}
It is clear that such $a,b,c,d\in\mathbb{C}$ exist. With these conditions, the determinant of $A$ is equal to the non-zero complex number~$\ol{\xi}(2\re(a\ol{b})-|c|^2-|d|^2)t^{-1}$. Since this is a unit in $\L{\C}$, the matrix $A$ is invertible, and we now set
\[B=(t-\xi)A.\]
We claim that $B$ represents $\mathfrak{e}(1,0,+1,\xi,\mathbb{C})\oplus\mathfrak{e}(1,0,-1,\xi,\mathbb{C})$. To see this, we first note that the Smith normal form of $B$ is a diagonal matrix with $((t-\xi),(t-\xi))$
on the diagonal. In fact, as~$A$ is invertible, the Smith normal form of $A$ is the identity matrix, and multiplying a matrix by a polynomial amounts to multiplying the Smith Normal Form by the same polynomial.
Given this observation, we deduce that $\L{\C}^2/B\L{\C}^2$ is isomorphic $\L{\C}/(t-\xi)\oplus\L{\C}/(t-\xi)$. Using Lemma~\ref{lem:evenpairing}, the form represented by~$B$ is therefore isometric to $\mathfrak{e}(1,0,\epsilon_1,\xi,\mathbb{C})\oplus\mathfrak{e}(1,0,\epsilon_2,\xi,\mathbb{C})$ for some $\epsilon_1,\epsilon_2=\pm1$.
Although the signs $\epsilon_1,\epsilon_2$ can be explicitly calculated, Proposition~\ref{prop:sumofjumps} below immediately implies that $\epsilon_1+\epsilon_2=0$. This concludes the proof of the proposition.
\end{proof}
\subsection{Representability over $\L{\C}$}\label{sec:rep2}
Generalizing the example given in Proposition~\ref{prop:notsostupidexample}, we can show the following result (whose converse will be proved in Proposition~\ref{prop:sumofjumps}).
\begin{proposition}\label{prop:inversejumps}
A non-singular linking form $(M,\lambda)$ over $\L{\C}$ with decomposition as in~\eqref{eq:splitting} is representable if
it is a sum of $\mathfrak{e}(n_i,0,\epsilon_i,\xi_i,\mathbb{C})$ and $\mathfrak{f}(n_j,0,\xi_j,\mathbb{C})$, where the signs $\epsilon_i$ satisfy
\begin{equation}\label{eq:signatureswithoutsignatures}
\sum_{i\colon n_i\textrm{ odd}} \epsilon_i=0.
\end{equation}
\end{proposition}
\begin{proof}
Decompose $(M,\lambda)$ into basic summands as in \eqref{eq:splitting}. We proceed by induction over the number of summands in this decomposition. Suppose that $(M,\lambda)$ has $N$ summands and that the result is proved for all forms with fewer than $N$ summands, that satisfy \eqref{eq:signatureswithoutsignatures}.
First, suppose $(M,\lambda)$ contains a basic summand $(M',\lambda')$ of the type $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{C})$ for $n$ even, or $\mathfrak{f}(n,0,\epsilon,\mathbb{C})$. In this case, $(M',\lambda')$ is representable by the first two items of Proposition~\ref{prop:EasyRepresent} and we can write $$(M,\lambda)=(M',\lambda')\oplus(M'',\lambda'')$$
where $(M'',\lambda'')$ is the sum of other summands. Since both $(M,\lambda)$ and $(M',\lambda')$ satisfy~\eqref{eq:signatureswithoutsignatures}, so does $(M'',\lambda'')$. We now conclude this case by applying the induction hypothesis to $(M'',\lambda'')$.
Next, suppose there is a form $\mathfrak{e}(2n+1,0,\xi,+1,\mathbb{C})$ entering the decomposition~\eqref{eq:splitting} of $(M,\lambda)$. By \eqref{eq:signatureswithoutsignatures},
there must be another form $\mathfrak{e}(2n'+1,0,\xi',-1,\mathbb{C})$ entering the decomposition. Write
$$(M',\lambda')=\mathfrak{e}(2n+1,0,\xi,+1,\mathbb{C})\oplus\mathfrak{e}(2n'+1,0,\xi',-1,\mathbb{C})$$ and decompose once again $(M,\lambda)$ as $(M',\lambda')\oplus (M'',\lambda'').$ Since
both $(M',\lambda')$ and $(M,\lambda)$ satisfy~\eqref{eq:signatureswithoutsignatures}, so does $(M'',\lambda'')$. As $(M'',\lambda'')$ has $N-2$ basic summands, the induction assumption applies. Therefore, in order to apply the induction hypothesis, we need only argue that~$(M',\lambda')$ is representable. If $\xi\neq \xi'$, then the third item of Proposition~\ref{prop:EasyRepresent} ensures that~$(M',\lambda')$ is representable. We can therefore assume that $\xi = \xi'$.
If $n=n'$ we take the matrix $A$ of \eqref{eq:trickymatrix} and multiply it by $(t-\xi)((t-\xi)(t^{-1}-\ol{\xi}))^{n}$ and it represents the desired pairing by the same arguments as in the proof of Proposition~\ref{prop:notsostupidexample}.
Suppose $n>n'$ (the other case is analogous). Set $k=n-n'$ and consider the matrix
\begin{equation}\label{eq:trickymatrix2}
A'=\begin{pmatrix}
((t-\xi)(t^{-1}-\ol{\xi}))^k(at^{-1}-\ol{a\xi}) & (t-\xi)^k(dt^{-1}+c)\\
-(\ol{c\xi}t^{-1}+\ol{d\xi})(t^{-1}-\ol{\xi})^k & bt^{-1}-\ol{b\xi},
\end{pmatrix},
\end{equation}
where $a,b,c,d$ are as in \eqref{eq:neweq} and, additionally, $b$ is not a real number (so that $\xi$ is not a root of~$(bt^{-1}-\ol{b\xi})$).
The matrix $A'$ is such that $(t-\xi)A'$ is Hermitian and the determinant of $A'$ is equal to the product of $((t-\xi)(t^{-1}-\ol{\xi}))^k$ with a unit of $\L{\C}$.
We compute the module presented by~$A'$.
The Smith normal form algorithm allows us to write $A'$ as
\begin{equation}\label{eq:aprimsmith}
A'=X\begin{pmatrix} b_1(t) & 0 \\ 0 & b_2(t) \end{pmatrix} Y,
\end{equation}
where $X$ and $Y$ are invertible matrices over $\L{\C}$, and $b_1(t)b_2(t)=\det A'$ up to a unit of~$\L{\C}$. As a shorthand, we write $\stackrel{.}{=}$ for equality up to multiplication by a unit of $\L{\C}$. Since we have $\det A'\stackrel{.}{=}((t-\xi)(t^{-1}-\ol{\xi}))^k$, we see that $b_1\stackrel{.}{=}(t-\xi)^{k_1}$ and $b_2\stackrel{.}{=}(t-\xi)^{k_2}$ with $k_1+k_2=2k$.
From~\eqref{eq:aprimsmith} we deduce that $(t-\xi)^{\min(k_1,k_2)}$ divides $v^TA'w$ for any two vectors $v,w\in\L{\C}^2$. But
if $v=w=(0,1)$, then~\eqref{eq:trickymatrix2} implies that $v^TA'w=bt^{-1}-\ol{b\xi}$, and the latter is coprime with~$(t-\xi)$ by our choice of $b$. It follows that $\min(k_1,k_2)=0$. Thus, we may readjust $X$ and $Y$ in such a way that $b_1=(t-\xi)^{2k}$ and $b_2=1$.
Since $(t-\xi)A'$ is Hermitian, $A''=(t-\xi)((t-\xi)(t^{-1}-\ol{\xi}))^{n'}A'$ is also Hermitian. Thus, using \eqref{eq:aprimsmith} and the same arguments as in the proof of Proposition~\ref{prop:notsostupidexample}, we obtain the following isomorphism of $\L{\C}$-modules:
\[\L{\C}^2/A''\L{\C}^2\cong\L{\C}/(t-\xi)^{2n+1}\oplus\L{\C}/(t-\xi)^{2n'+1}.\]
It follows from our classification Theorem~\ref{thm:MainLinkingForm} that $A''$ represents the linking form
\[\mathfrak{e}(2n+1,0,\epsilon_1,\xi,\mathbb{C})\oplus\mathfrak{e}(2n'+1,0,\epsilon_2,\xi,\mathbb{C}),\]
where $\epsilon_1,\epsilon_2=\pm 1$. By a direct calculation (or using Proposition~\ref{prop:sumofjumps} below), we deduce that $\epsilon_1+\epsilon_2=0$. If $\epsilon_1=-1$, we replace $A''$ by $-A''$, which changes the sign of $\epsilon_1$ and $\epsilon_2$. We conclude that $A''$ represents $(M',\lambda')$, so the induction step is accomplished and the proof is concluded.
\end{proof}
\subsection{Diagonalization over local rings}\label{sec:loc2}
Suppose that $(M,\lambda)$ is a form represented by a matrix~$A(t)$. It is often much easier to deduce information about the pairing if $A(t)$ is diagonal. While all non-singular forms over $\L{\R}$ are representable by a diagonal matrix (Proposition~\ref{prop:diagonalreal}), this is not the case in general: over $\L{\C}$ not all representable forms are represented by a diagonal matrix (Proposition~\ref{prop:notsostupidexample}). This subsection shows that diagonalizability can be achieved if we pass to a local ring.
\medbreak
The following lemma
bears a strong resemblance to
classical results \cite{Adkins,Wimmer}, but since the precise formulation is slightly different in our paper, we provide a proof.
\begin{lemma}\label{lem:analytic_change}
Let $A(t)$ be a Hermitian matrix over $\L{\C}$ and let $\xi \in S^1$.
There exists a matrix~$P(t)$ with entries analytic functions near $\xi$ such that
$B(t)=P(t)A(t){P(t)^\#}^T$ is diagonal.
\end{lemma}
\begin{proof}
It is not a surprise that the proof relies on the Gram-Schmidt orthogonalization process.
There is an extra difficulty stemming from the fact that we perform the process over the ring $\mathcal{O}_\xi$ and not over a field.
This fact adds some more technicalities. We rely on the fact that $\mathcal{O}_\xi$ is a discrete valuation ring.
To give the proof of Lemma~\ref{lem:analytic_change}, it is convenient to think about the matrix $A(t)$ as of an intersection matrix
of a Hermitian pairing on $\mathcal{O}_\xi^n$ given by $\langle e,f\rangle=e A (f^\#)^T$.
Write $A(t)=\{a_{ij}(t)\}_{i,j=1}^n$ and let $e_1,\ldots,e_n$ be a basis of $\mathcal{O}_\xi^n$.
Write $\alpha_{ij}$ for the order of $a_{ij}$ at $t=\xi$.
\begin{definition}
A basis $e_1,\ldots,e_n$ for $\mathcal{O}_\xi^n$ is \emph{adapted} if $\alpha_{ij}\ge\alpha_{jj}$ for all $i>j$ and if we additionally have $\alpha_{11}\ge\alpha_{22}\ge \ldots\ge~\alpha_{nn}$.
\end{definition}
\begin{lemma}\label{lem:basechange}
There is a base change $(e_1,\ldots,e_n)\to(\wt{e}_1,\ldots,\wt{e}_n)$ such that $(\wt{e}_1,\dots,\wt{e}_n)$ is adapted.
\end{lemma}
We temporarily postpone the proof of Lemma~\ref{lem:basechange} in order to see how it implies Lemma~\ref{lem:analytic_change}. We argue by induction, where the induction assumption is that $e_1,\ldots,e_k$ are pairwise orthogonal. Since the case $k=1$ is trivial, we assume $k\ge 2$. First, we replace $e_{k+1}$ by
\begin{equation}\label{eq:wt0}\wt{e}_{k+1}=e_{k+1}-\sum_{j=1}^k\frac{a_{k+1,j}}{a_{jj}}e_j.\end{equation}
If the basis $e_1,\ldots,e_n$ is adapted, then, as $k+1>j$, the function $\frac{a_{k+1,j}}{a_{jj}}$ is analytic near~$\xi$, so the base change is legitimate in $\mathcal{O}_\xi$. As in the classical Gram--Schmidt process, we have that $\langle\wt{e}_{k+1},e_j\rangle=0$ for $j\le k$.
To complete the induction step we need to show that the basis $e_1,\ldots,e_k,\wt{e}_{k+1},e_{k+2},\dots,e_n$ is still adapted.
Write $\wt{a}_{k+1,j}=\langle \wt{e}_{k+1},e_\ell\rangle$ and let $\wt{\alpha}_{k+1,\ell}$ be the order at $t=\xi$ of $\wt{a}_{k+1,\ell}$.
We aim to show that for $\ell>k+1$, we have
\begin{align}
\wt{\alpha}_{k+1,\ell}&\ge \alpha_{k+1,k+1}\label{eq:wt1},\\
\wt{\alpha}_{k+1,k+1}&=\alpha_{k+1,k+1}\label{eq:wt2}.
\end{align}
The first of the two equations holds on the nose, while the second might require an additional base change from the basis $e_1,\ldots,e_k,\wt{e}_{k+1},e_{k+2},\dots,e_n$ to a basis $e_1,\dots,e_{k-1},e'_k,e'_{k+1},e_{k+2},\dots,e_n$ (which still satisfies~\ref{eq:wt1}).
We first show \eqref{eq:wt1}. By \eqref{eq:wt0} we have $\wt{a}_{k+1,\ell}=a_{k+1,\ell}-\sum_{j=1}^k\frac{a_{k+1,j}}{a_{jj}}a_{j,\ell}.$
We compare the orders. Since the basis $e_1,\ldots,e_n$ is adapted, we have $\alpha_{j,\ell}\ge \alpha_{jj}\ge\alpha_{k+1,k+1}$ and the order of~$\frac{a_{k+1,j}}{a_{jj}}$ is non-negative. Thus all the summands have order at least $\alpha_{k+1,k+1}$. Since adaptedness also implies that $\alpha_{k+1,\ell}\ge\alpha_{k+1,k+1}$, we conclude that $\wt{\alpha}_{k+1,\ell}\ge \alpha_{k+1,k+1}$, proving~\eqref{eq:wt1}.
To prove \eqref{eq:wt2}, we use the definition of $\widetilde{a}_{k+1,k+1}$, the definition of $\widetilde{e}_{k+1}$ and orthogonality to~get
\begin{equation}\label{eq:wt33}
\begin{split}
\wt{a}_{k+1,k+1}&=\langle\wt{e}_{k+1},\wt{e}_{k+1}\rangle=\langle\wt{e}_{k+1},e_{k+1}-\sum_{j=1}^k\frac{a_{k+1,j}}{a_{jj}}e_j\rangle \\
&=\langle \wt{e}_{k+1},e_{k+1}\rangle=\langle e_{k+1}-\sum_{j=1}^k\frac{a_{k+1,j}}{a_{jj}}e_j,e_{k+1}\rangle=a_{k+1,k+1}-\sum_{j=1}^k\frac{a_{k+1,j}}{a_{jj}}a_{j,k+1}.
\end{split}
\end{equation}
In the latter formula, all terms have order at least $\alpha_{k+1,k+1}$, so as in the proof of \eqref{eq:wt1} we conclude that $\wt{\alpha}_{k+1,k+1}\ge\alpha_{k+1,k+1}$.
If we have an equality, we are done with the induction step.
Unfortunately it might happen that the order of $\wt{a}_{k+1,k+1}$ becomes strictly greater
than the order of $a_{k+1,k+1}$, so \eqref{eq:wt2} does not hold. This can only be the case if in the
last sum of \eqref{eq:wt33} there appears a term of order $\alpha_{k+1,k+1}$, cancelling the term of $a_{k+1,k+1}$.
We claim that if this happens, then $\alpha_{k,k}=\alpha_{k+1,k+1}$. Note that for $j\le k$ we have $\alpha_{j,k+1}\ge\alpha_{jj}\ge\alpha_{k+1,k+1}$ and, if at least one of the inequalities
is strict, then
the order of $\frac{a_{k+1,j}}{a_{jj}}a_{j,k+1}$ is strictly greater than~$\alpha_{k+1,k+1}$. Thus, if the terms of order $\alpha_{k+1,k+1}$ cancel out in right-hand side
of~\eqref{eq:wt33}, then some term in the sum on the right-hand side of \eqref{eq:wt33} has
order equal to~$\alpha_{k+1,k+1}$, and therefore
we must have $\alpha_{jj}=\alpha_{k+1,k+1}$ for some $j$. As $\alpha_{jj}\ge \alpha_{kk}\ge\alpha_{k+1,k+1}$,
we infer that $\alpha_{kk}=\alpha_{k+1,k+1}$. This concludes the proof of the claim.
Replace now $e_{k}$ and $\wt{e}_{k+1}$ by $e'_k=\frac12(e_k+\wt{e}_{k+1})$ and $e'_{k+1}=\frac12(e_k-\wt{e}_{k+1})$. This is an orthogonal base change
that, by the claim, makes the order of $\langle e'_{k},e'_{k}\rangle$ and $\langle e'_{k+1,k+1},e'_{k+1,k+1}\rangle$ equal to~$\alpha_{k+1,k+1}$. Moreover, for $m=k,k+1$, $\ell>k+1$, the order of
$\langle e'_{m},e_{\ell}\rangle$ is greater or equal than the minimum of the orders of $\langle e_k,e_{\ell}\rangle$ and $\langle e_{k+1},e_{\ell}\rangle$, so it is at least $\alpha_{k+1,k+1}$ by \eqref{eq:wt1}.
Summarizing, we replace the basis $e_1,\ldots,e_{n}$ with either $e_1,\ldots,e_{k},\wt{e}_{k+1},e_{k+2},\dots,e_n$ (if $\wt{\alpha}_{k+1,k+1}=\alpha_{k+1,k+1}$ from the beginning)
or $e_1,\dots,e_{k-1},e'_k,e'_{k+1},e_{k+2},\dots,e_n$ otherwise, and proceed with the induction step. This concludes the proof of the lemma.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{lem:basechange}]
We will show that there is a base change such that all the $\alpha_{ij}$ are equal. To this end, let $m\ge 0$ be the minimum
of all the $\alpha_{ij}$ for $1\le i,j\le m$. Let $C=\{c_{ij}\}$ be the complex matrix such that $c_{ij}$ is the term at $(t-\xi)^m$ of $a_{ij}$, that is, $c_{ij}=\frac{a_{ij}(t)}{(t-\xi)^m}|_{t=\xi}$. By definition of $m$, note that $C$ is not a zero matrix.
We claim that there exist a complex $n\times n$ matrix $B$, such that $\det B\neq 0$ and $BC\ol{B}^T$ has all terms different than zero. Writing $B=\{b_{ij}\}$, the $k\ell$-entry of $BC\ol{B}^T$ is given by
\[F_{k\ell}=\sum_{i,j}b_{ki}c_{ij}\ol{b}_{\ell j},\]
and $F_{k\ell}$, regarded as a map from $\mathbb{C}^{n\times n}\to\mathbb{C}$ is a non-trivial polynomial in $b_{ij},\ol{b}_{ij}$,
because some of the $c_{ij}$ are non-zero by the assumption.
The set of zeros of $F_{k\ell}$ is a codimension one subset of~$\mathbb{C}^{n\times n}$, in particular it is a boundary closed set, that is, a set whose complement is open and dense.
The union
\[ U=\bigcup_{k,\ell=1}^n\{b_{ij} \ | \ F_{k\ell}=0\}\cup\{B \ | \ \det B=0\}\]
is again a boundary closed set. It is enough to take the matrix $B$ in $\mathbb{C}^{n\times n}\setminus U$.
Given such a $B$, we replace $A(t)$ by $\wt{A}(t)=BA(t)\ol{B}^T$ (as $B$ is a constant matrix, we have $\ol{B}=B^\#$). In the expansion of $BA\ol{B}^T$,
the lowest order terms are given by $BC\ol{B}^T$. Using the claim, we deduce that each entry of $\widetilde{A}(t)$ has non-zero coefficient at $(t-\xi)^m$. Clearly, none of the entries has non-zero term at $(t-\xi)^{m'}$ for any $m'<m$.
This shows that all the orders of the entries of $\wt{A}(t)$ are equal to $m$. In particular $\wt{A}(t)$ is adapted.
\end{proof}
\section{Classification of forms up to Witt equivalence}
\label{sec:Witt}
While Section~\ref{sec:LinkingFormClassification} dealt with the classification of linking forms over $\L{\F}$ up to isometry, this section is concerned with the classification up to Witt equivalence. Namely, given a PID with involution, we recall the definition of the Witt group of linking forms over $R$ (Subsection~\ref{sub:Witt}), deal with polynomial rings (Subsection~\ref{sub:Devissage}) and then specialize to the real and complex cases (Subsections~\ref{sub:WittReal} and~\ref{sub:WittComplex}) all of which are fairly well known.
Finally we study morphisms of linking forms in Subsection~\ref{sec:isometricembedding}. The main result of that section, Theorem~\ref{thm:isoprojection}, is used
in the proof of the satellite formula in Section~\ref{sub:Satellite}.
\subsection{Witt groups of rings with involution}
\label{sub:Witt}
In this subsection, we review some generalities about Witt groups on rings with involution.
References include~\cite{Bourrigan, BargeLannesLatourVogel, Litherland, OrsonThesis, RanickiLocalization, LevineMetabolicHyperbolic}.
\medbreak
Let $R$ be a PID with involution and let $Q$ denote its field of fractions. Given a non-degenerate linking form $(M,\operatorname{\lambda})$ over $R$, a submodule $L \subset M$ is \emph{isotropic} (or \emph{sublagrangian}) if $L \subset L^\perp$ and \emph{metabolic} if $L=L^\perp$.
\begin{definition}
\label{def:metabolic_and_so_on}
The \emph{Witt group of linking forms}, denoted $W(Q,R)$, consists of the monoid of non-degenerate linking forms modulo the submonoid of metabolic linking forms. Two non-degenerate linking forms $(M,\operatorname{\lambda})$ and $(M',\operatorname{\lambda}')$ are called \emph{Witt equivalent} if they represent the same element in~$W(Q,R)$.
\end{definition}
The Witt group of linking forms is known to be an abelian group under direct sum, where the inverse of the class $[(M,\lambda)]$ is represented by $(M,-\lambda)$. Recall that since the ring $R$ is a PID, non-degenerate linking forms over a torsion $R$-module are in fact non-singular~\cite[Lemma 3.24]{Hillman}. Regarding the notation $W(Q,R)$ we chose to follow \cite[Appendix A.1]{BargeLannesLatourVogel}.
Given a non-singular linking form $(M,\operatorname{\lambda})$ and an isotropic submodule $L \subset M$, we define a form $\operatorname{\lambda}_L$ on $L^\perp/L$ by the formula
$\operatorname{\lambda}_L([x],[y])=\operatorname{\lambda}(x,y)$, where $[x],[y]\in L^\perp/L$ are classes of elements $x,y\in L^\perp$. It can be checked that $\operatorname{\lambda}_L$
is well-defined and in fact, gives a non-singular linking form $(L^\perp/L,\operatorname{\lambda}_L)$.
\begin{definition}\label{def:sublagrangian_reduction}
The non-singular linking form~$(L^\perp / L,\operatorname{\lambda}_L)$ is said to be obtained by \emph{sublagrangian reduction} on $L$.
\end{definition}
We refer to~\cite[Corollary A.12]{BargeLannesLatourVogel} and~\cite[Section II.A.4]{Bourrigan} for the proof of the following proposition.
\begin{proposition}
\label{prop:Reduction}
Given an isotropic subspace $L$ of a non-singular linking form $(M,\operatorname{\lambda})$, the non-singular linking forms $(M,\operatorname{\lambda})$ and $(L^\perp/L,\operatorname{\lambda}_L)$ are Witt equivalent.
\end{proposition}
From now on, we shall restrict ourselves to our ring of interest, namely to the ring $\L{\F}$ of Laurent polynomials, where $\mathbb{F}$ is either $\mathbb{R}$ or $\mathbb{C}$.
\subsection{Witt group of Laurent polynomial rings and d\'evissage}
\label{sub:Devissage}
This subsection has two goals. Firstly, we recall how the computation of Witt groups is simplified by considering the primary decomposition. Secondly, we recall the process known as \emph{d\'evissage}. References for this section include \cite{MilnorHusemoller, Lam, Bourrigan, OrsonThesis, LevineMetabolicHyperbolic}.
\medbreak
We start by introducing some additional terminology. A Laurent polynomial $p$ is \emph{weakly symmetric} if $p \stackrel{.}{=} p^\#$. Given an irreducible polynomial $p$ in $\L{\F}$, we denote by $M_p$ the $p$-primary part of $M$ and by $W(\O{\F},\L{\F},p)$ the Witt group of linking forms defined over $p$-primary modules.
\begin{proposition}
\label{prop:Primary}
Let $\mathcal{S}$ denote the set of irreducible
polynomials over $\L{\F}$. The primary decomposition theorem induces a canonical isomorphism
\[ W(\O{\F},\L{\F})\cong\bigoplus_{p \in \mathcal{S}} W(\O{\F},\L{\F},p).\]
\end{proposition}
\begin{proof}
It follows immediately from the definition that if $p$ is not equivalent to $q^\#$, then $M_p$ and~$M_q$ are orthogonal. Likewise, if $p$ is not weakly
symmetric, then the linking form restricted to $M_p\oplus M_{p^\#}$ is metabolic, in fact $M_p$ and $M_{p^\#}$ are two metabolizers.
\end{proof}
In order to describe d\'evissage, we recall some generalities on Witt groups of Hermitian forms. Let $R$ be a ring with involution and let $u$ be an element of $R$ which satisfies $u\makeithash{u}=1$. Given a projective $R$-module $H$, a sesquilinear form $\alpha \colon H \times H \to R$ is \emph{$u$-Hermitian} if $\alpha(y,x)=u \alpha(x,y)^\#$. We use~$W_u(R)$ to denote the Witt group of non-singular $u$-Hermitian forms. If $u=1$, then we write~$W(R)$ instead of~$W_1(R)$.
\begin{remark}
\label{rem:ElementuWitt}
Assume that $\mathbb{F}$ is a field whose characteristic is different from $2$. It can be checked that~$W_u(\mathbb{F})$ is trivial if $u=-1$ and the involution is trivial, and is isomorphic to $W(\mathbb{F})$ otherwise~\cite[Proposition A.1]{Litherland} and~\cite[Chapter II.A.3]{Bourrigan}. Nevertheless, as we shall see below, keeping track of this extra unit often proves useful.
\end{remark}
The next proposition (which is mentioned in~\cite[top of page 349]{Litherland}) is an instance of \emph{d\'evissage}; a proof can be found in~\cite[Chapter II.B.2]{Bourrigan} or~\cite[Proposition 2.3.14]{OrsonThesis}.
\begin{proposition}
\label{prop:Devissage}
Given an irreducible weakly symmetric polynomial $p$, there is an isomorphism
\[ W(\O{\F},\L{\F},p) \cong W_{\wt{u}}(\L{\F}/p),\]
where $\wt{u} \in \L{\F}/p$ is the reduction mod $p$ of the element $u\in\L{\F}$ such that $p=u\makeithash{p}$
\end{proposition}
Next, we specialize our study to the case where $\mathbb{F}$ is either $\mathbb{R}$ or $\mathbb{C}$.
\subsection{Real linking forms up to Witt equivalence.}
\label{sub:WittReal}
This subsection has two goals: firstly to discuss the structure of the group $W(\O{\R},\L{\R})$ and secondly to prove the following result.
\begin{theorem}
\label{thm:WittClassificationReal}
Set $S^1_+=\lbrace z \in S^1 \ | \ \iim(z)>0 \rbrace$.
Any real non-singular linking form is Witt equivalent to a direct sum of the type
$$ \bigoplus_{\substack{ n_i \text{ odd, }\epsilon_i=\pm 1
\\ \xi_i \in S^1_+, \ i\in I
}}\mathfrak{e}(n_i,0,\epsilon_i,\xi_i,\mathbb{R}).$$
\end{theorem}
Here recall that
~$\mathfrak{e}(n_i,0,\epsilon_i,\xi_i,\mathbb{R})$ denotes the basic linking form described in Example~\ref{BasicPairingXiS1Real}. The proof of Theorem~\ref{thm:WittClassificationReal} is based on the two following lemmas which show that the basic linking forms which do not appear in the statement of Theorem~\ref{thm:WittClassificationReal} are metabolic.
\begin{lemma}\label{lem:xiless1}
Given $\xi \in \Xi$, the following statements holds:
\begin{enumerate}
\item if $|\xi|\in(0,1)$, then the Witt group $W(\O{\R},\L{\R},{}^\R\! B_\xi(t))$ vanishes;
\item if $\xi=\pm 1$, then the Witt group $W(\O{\R},\L{\R},{}^\R\! B_\xi(t))$ vanishes.
\end{enumerate}
\end{lemma}
Note that we separate the lemma into two assertions since the definition of the basic polynomial corresponding to $\xi= \pm 1$ differs from the one associated to $\xi$ with $|\xi|$ lying in $(0,1)$.
\begin{proof}
We start with the first assertion. If $|\xi|<1$, then ${}^\R\! B_\xi(t)$ decomposes as a product of two weakly symmetric polynomials: ${}^\R\! B_\xi(t)=pp^\#$ where $p=(t-\xi)(t-\ol{\xi})$. If $M$ is a ${}^\R\! B_\xi(t)$-torsion module then $M_p$ and $M_{p^\#}$ are two metabolizers for the linking form.
To prove the second assertion, we use $\xi$ to denote $\pm 1$ and we claim that if the basic pairing $\mathfrak{e}(n,k,\epsilon,\xi,\mathbb{R})$ is non-singular, then it must be metabolic. First, as we mentioned in Example~\ref{BasicPairingXipm1Real}, the basic linking form $\mathfrak{e}(n,k,\epsilon,\xi,\mathbb{R})$ is non-singular if and only if $k=0$ (recall that in this case,~$n$ must be even). Since $\xi=\pm 1$, assuming that the aforementioned conditions hold, one can check that the submodule generated by $(t-\xi)^j\cdot [1]$ (where $j=n/2+1,\ldots,n$ and $[1]$ is the class of~$1$ in $\L{\R}/(t-\xi)^n$) is a metabolizer for $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{R})$. This concludes the proof of the claim. The second assertion of the lemma now follows: indeed, by Theorem~\ref{thm:non-degenerate-split}, non-singular forms are split and by Theorem~\ref{thm:MainLinkingForm}, any split pairing over a $(t-\xi)$-primary module decomposes into a direct sum of $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{R})$'s.
\end{proof}
From a more abstract point of view, the second point of Lemma~\ref{lem:xiless1} can be proved as follows: Thanks to Proposition~\ref{prop:Devissage}, we know that $W(\O{\R},\L{\R},{}^\R\! B_\xi(t))$ is isomorphic to $W_{-1}(\L{\R}/{}^\R\! B_\xi(t))$. Since the latter group is isomorphic to $W_{-1}(\mathbb{R})$ (with the trivial involution), the result follows from Remark~\ref{rem:ElementuWitt}.
\begin{lemma}\label{lem:LRxi}
If $|\xi|=1$ and $\xi\neq\pm 1$, then there is an isomorphism $W(\O{\R},\L{\R},{}^\R\! B_\xi(t))\cong \mathbb{Z}$ under which the basic pairing $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{R})$ is mapped to $\epsilon$ if $n$ is odd and to $0$ if $n$ is even.
\end{lemma}
\begin{proof}
As $\xi$ is not real, we deduce that $\L{\R}/{}^\R\! B_\xi(t)$ is isomorphic to $\mathbb{C}$. Using Proposition~\ref{prop:Devissage} and Remark~\ref{rem:ElementuWitt}, it follows that $W(\O{\R},\L{\R},{}^\R\! B_\xi(t))$ is isomorphic to $ W(\mathbb{C}) \cong\mathbb{Z}$ (where this latter isomorphism is given by the signature).
We now move on to the second statement. First, we assume that $n=2m$ is even and we set $H:=\L{\R}/{}^\R\! B_\xi(t)^{2m}$. Considering $L={}^\R\! B_\xi(t)^m H$, using the symmetry of ${}^\R\! B_\xi(t)$ and the non-singularity of the pairing, we observe that $L=L^\perp$ is a metabolizer for $\mathfrak{e}(2m,0,\epsilon,\xi,\mathbb{R})$. Next assume that $n=2m+1$ is odd, set $H:=\L{\R}/{}^\R\! B_\xi(t)^{2m+1}$ and $\lambda=e(2m+1,0,\epsilon,\xi,\mathbb{R})$. Considering $L={}^\R\! B_\xi(t)^{m+1}H$, one sees that $L^\perp$ is equal to ${}^\R\! B_\xi(t)^mH$. It follows that multiplication by ${}^\R\! B_\xi(t)^m$ induces an isometry between the sublagrangian reduction $(L^\perp/L,\lambda_L)$ and $e(1,0,\epsilon,\xi,\mathbb{R})$. Proposition~\ref{prop:Reduction} now implies that $\mathfrak{e}(2m+1,0,\epsilon,\xi,\mathbb{R})$ is Witt equivalent to $\mathfrak{e}(1,0,\epsilon,\xi,\mathbb{R})$.
To conclude the proof, it only remains to show that the basic linking form~$\mathfrak{e}(1,0,\epsilon,\xi,\mathbb{R})$ is mapped to $\epsilon$ under the aforementioned isomorphism. This is fairly immediate since under the isomorphisms $W(\O{\R},\L{\R},{}^\R\! B_\xi(t)) \cong W_{\widetilde{u}}(\L{\R}/{}^\R\! B_\xi(t)) \cong W(\mathbb{C})$, the image of the linking form~$\mathfrak{e}(1,0,\epsilon,\xi,\mathbb{R})$ is congruent to the Hermitian form $(x,y) \mapsto \epsilon xy^\#$.
\end{proof}
Combining the previous lemmas, we can prove Theorem~\ref{thm:WittClassificationReal} which states that any real non-singular linking form is Witt equivalent to a direct sum of $\mathfrak{e}(n_i,0,\epsilon_i,\xi_i,\mathbb{R})$, where each $n_i$ is odd and each $\xi_i$ lies in $S^1_+$.
\begin{proof}[Proof of Theorem~\ref{thm:WittClassificationReal}]
Let $(M,\lambda)$ be a real non-singular linking form. Using Theorem~\ref{thm:non-degenerate-split}, we deduce that $(M,\lambda)$ is split and thus Theorem~\ref{thm:MainLinkingForm} implies that $(M,\lambda)$ must be isometric to a direct sum of basic forms:
$$\bigoplus_{\substack{ n_i,\epsilon_i,\xi_i\\ i\in I}}\mathfrak{e}(n_i,0,\epsilon_i,\xi_i,\mathbb{R})\oplus
\bigoplus_{\substack{n_j,\xi_j\\ j\in J}}\mathfrak{f}(n_j,0,\xi_j,\mathbb{R}).$$
Using Lemma~\ref{lem:xiless1}, we know that the basic forms $\mathfrak{f}(n,0,\xi,\mathbb{R})$ and $\mathfrak{e}(n,0,\epsilon,\pm 1,\mathbb{R})$ are Witt trivial Furthermore, Lemma~\ref{lem:LRxi} implies that $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{R})$ is metabolic if $n$ is even. This concludes the proof of the theorem.
\end{proof}
We conclude this subsection by discussing the structure of the group $W(\O{\R},\L{\R})$, see~\cite[Section 5]{MilnorInfiniteCyclic},~\cite[Section 2.10]{Hillman} and~\cite[Example 2.3.24]{OrsonThesis} for closely related discussions.
Combining the decomposition described in Proposition~\ref{prop:Primary} with the results of Lemmas~\ref{lem:xiless1} and~\ref{lem:LRxi}, we obtain
\begin{equation}
\label{eq:WittReal}
W(\O{\R},\L{\R}) \cong \bigoplus_{\xi \in S^1_+} W(\O{\R},\L{\R},{}^\R\! B_\xi(t)) \cong \bigoplus_{\xi \in S^1_+} \mathbb{Z}.
\end{equation}
In more detail, Proposition~\ref{prop:Primary} implies that $W(\O{\R},\L{\R})$ is isomorphic to the direct sum of the $W(\O{\R},\L{\R},{}^\R\! B_\xi(t))$, where ${}^\R\! B_\xi(t)$ ranges over the basic polynomials. Lemmas~\ref{lem:xiless1} and~\ref{lem:LRxi} compute each of these groups, whence the announced isomorphism.
\subsection{Complex linking forms up to Witt equivalence}
\label{sub:WittComplex}
Just as in the real case, this subsection has two goals: firstly to discuss the structure of the group $W(\O{\C},\L{\C})$ and secondly to prove the following result:
\begin{theorem}
\label{thm:WittClassificationComplex}
Any complex non-singular linking form is Witt equivalent to a direct of the type:
$$ \bigoplus_{\substack{ n_i \text{ odd, }\epsilon_i=\pm 1 \\ \xi_i \in S^1, \ i\in I}}\mathfrak{e}(n_i,0,\epsilon_i,\xi_i,\mathbb{C}).$$
\end{theorem}
The proof of Theorem~\ref{thm:WittClassificationComplex} proceeds as in the real case. Namely, we deal with one basic linking form at a time.
\begin{lemma}\label{lem:xiless1a}
The following statements hold:
\begin{enumerate}
\item If $|\xi| \in (0,1)$, then the Witt group $W(\O{\C},\L{\C},(t-\xi)(t^{-1}-\ol{\xi}))$ vanishes.
\item If $\xi$ lies in $S^1$, then there is an isomorphism $W(\O{\C},\L{\C},(t-\xi))\cong \mathbb{Z}$. Furthermore, this isomorphism maps the basic pairing $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{C})$ to $\epsilon$ if $n$ is odd and to $0$ if $n$ is even.
\end{enumerate}
\end{lemma}
\begin{proof}
The proof of the first assertion is the same as in the real case, see Lemma~\ref{lem:xiless1}. The proof the second assertion is analogous to the proof of Lemma~\ref{lem:LRxi}. Namely, using d\'evissage (recall Proposition~\ref{prop:Devissage}), we see that $W(\O{\C},\L{\C},t-\xi)$ is isomorphic to $W_{-\xi^2}(\L{\C}/(t-\xi))$. But now, since~$\L{\C}/(t-\xi)$ is isomorphic to $\mathbb{C}$, Remark~\ref{rem:ElementuWitt} implies that the latter group is isomorphic to~$\mathbb{Z}$, as desired. Finally, as in the proof of Lemma~\ref{lem:LRxi} we show that under this isomorphism $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{C})$ is mapped to $0$ if $n$ is even and to $\epsilon$ if $n$ is odd.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:WittClassificationComplex}]
Let $(M,\lambda)$ be a complex non-singular linking form. Using Theorem~\ref{thm:non-degenerate-split}, we deduce that $(M,\lambda)$ is split and thus Theorem~\ref{thm:MainLinkingForm} implies that $(M,\lambda)$ form must be isometric to a direct sum of basic forms:
$$\bigoplus_{\substack{ n_i,\epsilon_i,\xi_i\\ i\in I}}\mathfrak{e}(n_i,0,\epsilon_i,\xi_i,\mathbb{C})\oplus
\bigoplus_{\substack{n_j,\xi_j\\ j\in J}}\mathfrak{f}(n_j,0,\xi_j,\mathbb{C}).$$
Using the first point of Lemma~\ref{lem:xiless1a}, we know that the basic forms $\mathfrak{f}(n,0,\xi,\mathbb{C})$ are metabolic. The second point of Lemma~\ref{lem:xiless1a} implies that $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{C})$ is metabolic if $n$ is even. This concludes the proof of the theorem.
\end{proof}
We conclude this subsection by discussing the structure of the group $W(\O{\C},\L{\C})$ (note that this result has appeared in~\cite[Appendix A]{Litherland},~\cite[Chapter II.C]{Bourrigan} and~\cite[Example 2.3.24]{OrsonThesis}). Namely, combining the decomposition described in Proposition~\ref{prop:Primary} with Lemma~\ref{lem:xiless1a}, we obtain
\begin{equation}
\label{eq:WittComplex}
W(\O{\C},\L{\C}) \cong \bigoplus_{\xi \in S^1} W(\O{\C},\L{\C},t-\xi) \cong \bigoplus_{\xi \in S^1} \mathbb{Z}.
\end{equation}
In more details, Proposition~\ref{prop:Primary} implies that $W(\O{\C},\L{\C})$ is isomorphic to the direct sum of the $W(\O{\C},\L{\C},t-\xi)$, where $\xi$ ranges over $S^1$. Lemma~\ref{lem:xiless1a} computes each of these groups, whence the announced isomorphism.
\subsection{Witt equivalence and representability}\
In this subsection, we briefly outline how representability (see Section~\ref{sec:Representability})
fits into the general framework of Witt groups.
\medbreak
Instead of describing the integrality of the localization exact sequence of Witt groups/symmetric L-groups (see e.g.~\cite{RanickiExact, RanickiLocalization}), we focus on the part of the theory that is relevant to representability. Namely, we describe the map
$$\partial \colon W(\mathbb{F}(t)) \to W(\mathbb{F}(t),\L{\F}).$$
A \emph{lattice} for a $\mathbb{F}(t)$-Hermitian form $(E,\alpha)$ is a free $\L{\F}$-submodule $P \subset E$ which satisfies $\alpha(x,y) \in \L{\F}$ for all $x,y$ in $P$, i.e. such that $\alpha$ restricts to a well defined $\L{\F}$-homomorphism $P \to \operatorname{Hom}_{\L{\F}}(P,\L{\F})^\#$.
Given such a lattice, one can consider the \emph{dual lattice}
$$ P_d:=\lbrace x \in E \ | \ \alpha(x,p) \in P \text{ for all } p \in P \rbrace. $$
Observe that the $\mathbb{F}(t)$-isomorphism $x \mapsto \alpha(x,-)$ restricts to a well defined $\L{\F}$-isomorphism $P_d \to \operatorname{Hom}_{\L{\F}}(P,\L{\F})^\#$. Since $P$ is free, $\operatorname{Hom}_{\L{\F}}(P,\L{\F})^\#$ is torsion-free. Since $\L{\F}$ is a PID, this implies that $P_d$ is in fact free. Since $P$ is a lattice, the canonical inclusion $P \subset E$ induces a well defined inclusion $P \subset P_d$, and it can be checked that $P_d / P$ is $\L{\F}$-torsion. One can then verify that the following assignment gives rise to a well defined linking form on $P_d /P$:
$$ \partial \alpha ([x],[y]) :=\alpha(x,y).$$
It can also be checked that $\partial$ descends to a well defined map on the level of Witt groups. In fact, as a Witt class $\alpha$ in $W(\mathbb{F}(t))$ can always be represented by a Hermitian matrix $A$ with coefficients in~$\L{\F}$ and non-zero determinant, a computation shows that $\partial \alpha$ is isometric to the linking form represented by $A$. In particular, a linking form $(M,\lambda)$ is Witt equivalent to a representable linking form if and only if $[(M,\lambda)]$ lies in the image of $\partial$.
\begin{remark}
\label{rem:Litherland}
Using Proposition~\ref{prop:Primary}, one can decompose $W(\mathbb{F}(t),\L{\F})$ into the direct sum of $W^0(\mathbb{F}(t),\L{\F}):=\bigoplus_{p \in \mathcal{S} \setminus \lbrace t-1 \rbrace} W(\mathbb{F}(t),\L{\F},p)$ and $W(\mathbb{F}(t),\L{\F},t-1)$. It has been claimed that the map $\partial$ restricts a surjection $W(\mathbb{F}(t)) \to W^0(\mathbb{F}(t),\L{\F})$~\cite[Theorem A.2]{Litherland}. While this result is correct for $\mathbb{F}=\mathbb{R}$ (since linking forms over $\mathbb{R}[t^{\pm 1}]$ are always representable) it is incorrect over $\mathbb{C}[t^{\pm 1}]$ since for any $\xi \in \mathbb{C}$, the basic linking form $\mathfrak{e}(1,0,1,\xi,\mathbb{C})$ is not Witt equivalent to a representable form. This can either be seen by reasoning as in Proposition~\ref{prop:verystupidexample} (and using that if~$(M,\lambda)$ is metabolic, then the order of~$M$ is a norm) or by using Proposition~\ref{prop:sumofjumps} below.
\end{remark}
\subsection{Forms restricted to submodules and Witt equivalence}\label{sec:isometricembedding}
In this subsection, we study submodules of basic forms as well as the restriction of morphisms of linking forms to such submodules. Some of the results in this section might appear technical, but they are
used in Subsection~\ref{sub:Satellite} while studying cabling formulae.
\medbreak
As a warm-up to more complicated results, we prove the following proposition.
\begin{proposition}\label{prop:missing_in_action}
Suppose that $(M,\operatorname{\lambda})$ is a non-singular linking form over $\L{\F}$ and let $(M',\operatorname{\lambda}')$ be another linking form. Let $\iota\colon M\to M'$ be a morphism of linking forms.
Then $M'$ splits as an orthogonal sum $M'_{im}\oplus M'_{ort}$ and $\iota\colon M\to M'_{im}$ is an isometry.
\end{proposition}
\begin{proof}
We shall first prove the result when $(M,\operatorname{\lambda})$ is a basic form. We will then deduce the general statement by induction on the number of summands of $(M,\operatorname{\lambda})$ in the decomposition \eqref{eq:splitting}.
Suppose that $(M,\operatorname{\lambda})=\mathfrak{e}(n,0,\xi,\epsilon,\mathbb{F})$ (the case of $\mathfrak{f}$-forms is analogous). Choose a generator~$x$ of~$M$ and note that we have $\operatorname{\lambda}(x,x)=\frac{q}{{}^\F\! B_\xi(t)^n}$
for some $q\in\L{\F}$ coprime with ${}^\F\! B_\xi(t)$.
Set $x'=\iota(x)$. Since $\iota$ is a morphism of linking forms, we have $\operatorname{\lambda}(x,x)=\operatorname{\lambda}'(x',x')$. Consequently, since the pairing~$\lambda'$ is non-singular, we infer that the order of $x'$ is ${}^\F\! B_\xi(t)^n$, i.e. no smaller power of~${}^\F\! B_\xi(t)$
annihilates~$x'$.
We claim that $M'=\iota(M)\oplus M'_0$ for some $M_0'$. Suppose that $y'\in M'$ satisfies ${}^\F\! B_\xi(t)^s y'=x'$ for some $s\ge 0$. Write $M'_{y'}$ for the submodule
of $M'$ generated by $y'$ and let $\lambda'_{y'}$ be the pairing $\operatorname{\lambda}'$ restricted to $M'_{y'}$. Using the definition of $y'$ and the symmetry of ${}^\F\! B_\xi(t)$, we obtain
\[\lambda'_{y'}(x',x')=\lambda'_{y'}({}^\F\! B_\xi(t)^s y',{}^\F\! B_\xi(t)^sy')={}^\F\! B_\xi(t)^{2s}\lambda'_{y'}(y',y').\]
On the other hand, using the first paragraph of the proof, we know that $\lambda'_{y'}(x',x')=\frac{q}{{}^\F\! B_\xi(t)^n}$. Combining these observations, we deduce that
\begin{equation}
\label{eq:y'}
\lambda'_{y'}(y',y')=\frac{q}{{}^\F\! B_\xi(t)^{n+2s}}.
\end{equation}
Since ${}^\F\! B_\xi(t)^nx'=0$, we know that ${}^\F\! B_\xi(t)^{n+s}y'=0$. This shows that $s=0$ (multiply both sides of~\eqref{eq:y'} by ${}^\F\! B_\xi(t)^{n+s}$
and use that $q$ is coprime with ${}^\F\! B_\xi(t)$).
In particular,~$x'$ is not a multiple by~${}^\F\! B_\xi(t)$ of any element~$y'$. That is, $\iota(M)$ splits
off as a direct summand in $M'$. This concludes the proof of the claim.
Next, we modifiy the summand $M_0'$ from the claim in order to obtain a summand that is orthogonal with respect to the linking form. To do this, we set
\begin{equation}\label{eq:GS_pairing}
M'_{ort}=\left\{u-\frac{\operatorname{\lambda}'(u,x')}{\operatorname{\lambda}'(x',x')}x'\,|\,u\in M'_0\right\}.
\end{equation}
To show that the definition of $M'_{ort}$ makes sense, we must show that $\frac{\operatorname{\lambda}'(u,x')}{\operatorname{\lambda}'(x',x')}x'$ is a well defined element of $M'$. Recall that $\operatorname{\lambda}'(x',x')=\frac{q}{{}^\F\! B_\xi(t)^n}$, where $q$ is coprime with ${}^\F\! B_\xi(t)$. Since $q$ coprime with~${}^\F\! B_\xi(t)$, we can write $qz'=x'$ for some~$z'$ in~$M'$. Note that the order of~$z'$ is also~${}^\F\! B_\xi(t)^n$. Furthermore~$\operatorname{\lambda}'(u,x')$ can be written as $\frac{r}{{}^\F\! B_\xi(t)^k}$ for some $r \in \L{\F}$. Since $\lambda'$ is non-singular, $n \geq k$. Combining these observations, we obtain
\begin{equation}
\label{eq:WellDefOrt}
\frac{\operatorname{\lambda}'(u,x')}{\operatorname{\lambda}'(x',x')}x'=(F_\xi(t)^{n-k}(t)\cdot r) \cdot z' \in M'.
\end{equation}
Now $r$ (resp. $q$) is only defined up to a multiple of ${}^\F\! B_\xi(t)^k$ (resp. ${}^\F\! B_\xi(t)^n$), so we must show that~\eqref{eq:WellDefOrt} is independent of these choices. As the order of~$z'$ is~${}^\F\! B_\xi(t)^n$, we deduce~\eqref{eq:WellDefOrt} is independent of the choice of $r$ and that $z'$ is independent of the choice of $q$. Thus the definition of $M_{ort}$ makes~sense.
Now $M'_{ort}$ is a submodule of $M'$, and using the claim, we see that $M'_{ort}\cap\iota(M)=~0$ and $M'=\iota(M)+M'_{ort}$. By construction, $M'_{ort}$ is orthogonal to $\iota(M)$ with respect to the linking form.
\smallskip
Having proved the proposition for basic forms, we pass to the general statement. We proceed by induction. Use the decomposition Theorem~\ref{thm:MainLinkingForm}
to write $(M,\operatorname{\lambda})$ as $(M_0,\operatorname{\lambda}_0)\oplus (M_1,\operatorname{\lambda}_1)$, where~$(M_0,\operatorname{\lambda}_0)$ is basic and~$(M_1,\operatorname{\lambda}_1)$ is non-singular and has fewer summands in its decomposition into basic forms. From the first part of the proof applied to $(M_0,\operatorname{\lambda}_0)$, we conclude that $M'$ splits as an orthogonal sum
$\iota(M_0)\oplus M'_{ort}$. As $M_1$ is orthogonal to $M_0$, we have that $\iota(M_1)$ is orthogonal to $\iota(M_0)$. Consequently, $\iota$ maps $M_1$ into $M'_{ort}$. Since $M_1$ has few basic summands than~$M$, we can apply the induction hypothesis to $\iota|_{M_1} \colon M_1 \to M'_{ort}$. This concludes the proof of the proposition.
\end{proof}
Before moving on to next result, we make some remarks.
\begin{remark}\
\begin{itemize}
\item[(a)] In the statement of Proposition~\ref{prop:missing_in_action}, we make no assumpotions on $(M',\operatorname{\lambda}')$: this form need not even be weakly split. Furthermore, the proof essentially shows that the non-degenerate
part of a linking form splits off as a direct summand.
\item[(b)] Proposition~\ref{prop:missing_in_action} is false if we do not assume that $(M,\operatorname{\lambda})$ is non-degenerate. A counterexample is given by the form in Subsection~\ref{sub:NonSplit}: for $(M',\lambda')$, take the form described in Example~\ref{ex:NonSplit} and for $(M,\lambda)$, take one of its summands.
\end{itemize}
\end{remark}
Next, we study morphisms between basic linking forms where neither the source nor the target is assumed to be non-singular.
\begin{lemma}\label{lem:submodule}
Let $(M',\lambda')$ be a basic linking form over $\L{\F}$ and let $M$ be an $\L{\F}$-module. Suppose there exists a monomorphism $\iota\colon M\to M'$
and define a sesquilinear pairing $\lambda\colon M\times M\to\mathbb{F}(t)/\L{\F}$ by setting $\lambda(x,y):=\lambda'(\iota(x),\iota(y))$.
\begin{itemize}
\item[(a)] If $(M',\lambda')= \mathfrak{e}(n',k',\epsilon,\xi,\mathbb{F})$, then for some $\ell\in\{0,1,\dots,n'-1\}$ we have
$(M,\lambda)=\mathfrak{e}(n,k,\epsilon,\xi,\mathbb{F})$, where $n=n'-\ell$ and $k=\min(k'+\ell,n'-\ell)$.
\item[(b)] If $(M',\lambda')=\mathfrak{f}(n',k',\xi,\mathbb{F})$,
then~$(M,\lambda)=\mathfrak{f}^w((n'-a,n'-b),n'-k'-a-b,\xi,\mathbb{F})$.
Here $a,b$ are integers such that $0\le a,b\le n'$.
\end{itemize}
\end{lemma}
\begin{proof}
\emph{Item (a).}
For simplicity, we restrict ourselves to the case $\mathbb{F}=\mathbb{R}$. The case $\mathbb{F}=\mathbb{C}$ leads to slightly different formulae, depending on the parity of $n'-k'$, but the proof is completely
analogous. Let $x'$ be a generator of $M'$ such that \[\lambda'(x',x')=\frac{\epsilon}{{}^\F\! B_\xi(t)^{n'-k'}}.\] As $\iota(M)$ is a submodule of $M'$ and $\L{\F}$ is a PID, we infer
that $M$ is a cyclic module, which is ${}^\F\! B_\xi(t)$-primary. Let $x_0$ be a generator of $M$. We have $\iota(x_0)=g_0x'$ for some $g_0\in\L{\F}$. As ${}^\F\! B_\xi(t)$
is irreducible over $\L{\F}$, we can write $g_0=g{}^\F\! B_\xi(t)^{\ell}$ for some $\ell\ge 0$ and some $g$ coprime with ${}^\F\! B_\xi(t)$. Choose $g'\in\L{\F}$ such that $g'g\equiv 1\bmod{}^\F\! B_\xi(t)^{n'}$ and set $x=g'x_0$. It follows that $\iota(x)={}^\F\! B_\xi(t)^\ell x'$. Moreover $x$ is a generator of $M$.
By definition of the sesquilinear form $\lambda$, we deduce that
\[\lambda(x,x)=\lambda'(\iota(x),\iota(x))=\frac{\epsilon}{{}^\F\! B_\xi(t)^{n'-k'-2\ell}}.\]
Note that if $n'-k'-2\ell\le 0$, then the pairing $\lambda$
is trivial. Combining these observations, we obtain that $(M,\lambda)=\mathfrak{e}(n'-\ell,\min(k'+\ell,n'-\ell),\epsilon,\xi,\mathbb{F})$.
\smallskip
\emph{Item (b)}. As above, we choose a generator $x'$ of $M'$ such that \[\lambda'(x',x')=\frac{1}{{}^\F\! B_\xi(t)^{n'-k'}}.\] Again, we also choose a generator $x_0$ of $M$ and write
$\iota(x_0)=g_0x$. The difference is that since ${}^\F\! B_\xi(t)={}^\F\! B_{\xi+}(t){}^\F\! B_{\xi-}(t)$ is not irreducible over $\L{\F}$, we have $g_0=g{}^\F\! B_{\xi+}(t)^a{}^\F\! B_{\xi-}(t)^b$
for some $a,b\ge 0$, $a,b\le n'$ and $g$ coprime with ${}^\F\! B_\xi(t)$.
Take $g'\in\L{\F}$ such that $gg'\equiv 1\bmod{}^\F\! B_\xi(t)^{n'}$ and set $x=g'x_0$. Then $x$ is a generator of $M$ and \[\lambda(x,x)=\lambda(\iota(x),\iota(x))=\frac{1}{{}^\F\! B_\xi(t)^{n'-k'-a-b}}.\]
This concludes the proof of the lemma.
\end{proof}
Note that Lemma~\ref{lem:submodule} also holds when $(M',\lambda')$ is a $\mathfrak{f}^w$-form although the proof requires a few additional few lines.
The following result is the main result of this section; it will play an important role in the proof of the satellite formula of Theorem~\ref{thm:CablingTheorem}.
\begin{theorem}\label{thm:isoprojection}
Let $(M',\lambda')$ and $(M'',\lambda'')$ be two non-singular linking forms over $\L{\F}$, let $M$ be a $\L{\F}$-module and let
$\iota\colon M\to M'$ be a monomorphism. Write $\lambda$ for the form on~$M$ induced by~$\iota$ as in Lemma~\ref{lem:submodule}.
Suppose that $\pi\colon(M,\lambda)\to(M'',\lambda'')$ is a surjective morphism of linking forms and
set $L:=\ker(\pi)$ and $L':=\iota(L)$. Then
$\ord(M)\ord(M)^\#$ divides $\ord(M')\ord(M'')$. Moreover, if
\begin{equation}\label{eq:order}\ord(M)\ord(M)^\#\doteq \ord(M'')\ord(M'),
\end{equation}
then the following statements hold:
\begin{enumerate}
\item $\iota(M)={L'}^{\perp}$;
\item The linking form $(M'',\lambda'')$ is isometric to the sublagrangian reduction of $M'$ with respect to~$L'$. In particular $(M',\lambda')$ and~$(M'',\lambda'')$ are Witt equivalent.
\end{enumerate}
\end{theorem}
The proof of Theorem~\ref{thm:isoprojection} relies on the following special case.
\begin{proposition}\label{prop:isoprojection_basic}
Let $(M',\lambda')$ be a basic form, let $M$ be a $\L{\F}$-module and let $\iota\colon M\to M'$ be a monomorphism. Write $\lambda$ for the pairing induced on~$M$ by $\iota$. Let $(M'',\lambda'')$ be a non-trivial linking form and suppose that $\pi\colon(M,\lambda)\to (M'',\lambda'')$ is
a surjective morphism of linking forms.
If~$(M',\lambda')$ and $(M'',\lambda'')$ are non-singular, then the following statements hold:
\begin{itemize}
\item[(a)] We have the equality $\ord(M)\ord(M)^\#=\ord(M')\ord(M'')$.
\item[(b)] The module $L'=\iota(\ker(\pi))$ is sublagrangian. We have $\iota(M)={L'}^{\perp}$ and $M''\cong{L'}^{\perp}/{L'}$. Moreover $(M'',\lambda'')$ is isometric to the sublagrangian reduction of $M'$ with respect to $L'$.
\end{itemize}
\end{proposition}
\begin{proof}
\emph{Case I: $M'$ is a type $\mathfrak{e}$-form.}
We first consider the case when $(M',\lambda')=\mathfrak{e}(n',0,\xi,\epsilon,\mathbb{F})$ (we have $k'=0$, because $(M',\lambda')$ is non-singular).
We start with item (a). As in the proof of Lemma~\ref{lem:submodule}, we restrict ourselves to the case $\mathbb{F}=\mathbb{R}$.
By Lemma~\ref{lem:submodule}(a), we have $(M,\lambda)=\mathfrak{e}(n'-\ell,\min(\ell,n'-\ell),\xi,\epsilon,\mathbb{F})$.
Let $x\in M$ be a generator as in Lemma~\ref{lem:submodule}.
As $\pi$ is surjective, the image $x''_0:=\pi(x)$ is a generator of~$M''$, and in particular~$M''$ is a cyclic module.
Since we assumed that $(M'',\lambda'')$ is non-singular, it follows that $(M'',\lambda'')=\mathfrak{e}(n'',0,\xi,\epsilon,\mathbb{F})$ for some $n''$.
On the other hand, since $\pi$ is a morphism of linking forms, we get
\[\lambda''(\pi(x),\pi(x))=\lambda(x,x)=\frac{\epsilon}{{}^\F\! B_\xi(t)^{n'-2\ell}}.\]
Since we saw that $(M'',\lambda'')=\mathfrak{e}(n'',0,\xi,\epsilon,\mathbb{F}),$ we infer that $n''=n'-2\ell$. In particular
$n'-2\ell>0$.
We now have $\ord(M)={}^\F\! B_\xi(t)^{n'-\ell}$, $\ord(M')={}^\F\! B_\xi(t)^{n'}$ and $\ord(M'')={}^\F\! B_\xi(t)^{n'-2\ell'}$. But since ${}^\F\! B_\xi(t)\doteq{}^\F\! B_\xi(t)^\#$, we obtain item (a) of the proposition.
To prove (b), write $L=\ker(\pi)$ and $L':=\iota(L)$. To check that $L'$ is a sublagrangian, pick $y,z \in L$ and observe that by definition of $\lambda$ and since $\pi$ is a morphism of linking forms, we have
$$ \lambda'(\iota(y),\iota(z))=\lambda(y,z)=\lambda''(\pi(y),\pi(z))=0. $$
We first show that $\iota(M)={L'}^\perp$. Keeping the same notations as in Lemma~\ref{lem:submodule}, we can pick generators $x$ and $x'$ of $M$ and $M'$ such that $\iota(x)={}^\F\! B_\xi(t)^\ell x'$. Note that $L=\ker(\pi)$ is generated by~${}^\F\! B_\xi(t)^{n'-2\ell}x$,
so $L'=\iota(L)$ is generated by ${}^\F\! B_\xi(t)^{n'-\ell}x'$.
As $M$ is generated by $x$, the image $\iota(M)$ is generated by ${}^\F\! B_\xi(t)^{\ell}x'$, and it follows that $\iota(M)={L'}^\perp$, as desired.
Next, we study the sublagrangian reduction of $M'$ with respect to $L'$. Note that $M''\cong M/L$ (here we used that $\pi$ is surjective and $L=\ker(\pi)$) can be identified with ${L'}^\perp/L'$ via the map
induced by $\iota$. Let
$\lambda_{L'}$ be the form on ${L'}^\perp/L'$ obtained by the sublagrangian reduction.
It remains to see that the forms $({L'}^\perp/L',\lambda_{L'})$ is isometric to $(M'',\lambda'')$. The modules ${L'}^\perp/L'$ and $M''$ are isomorphic to
$\L{\F}/{}^\F\! B_\xi(t)^{n'-2\ell}$, so using the classification theorem (Theorem~\ref{thm:MainLinkingForm}) it is enough to show that
there are elements $x_{L'}\in {L'}^\perp/L'$ and $x''\in M'$ such that
\[\lambda_{L'}(x_{L'},x_{L'})=\lambda''(x'',x'')=\frac{\epsilon}{{}^\F\! B_\xi(t)^{n'-2\ell}}.\] For this,
we take $x_{L'}$ to be the class of $\iota(x)\in{L'}^\perp$ in the quotient ${L'}^\perp/L'$ and $x''=\pi(x)$.
\smallskip
\emph{Case II, $M'$ is a type $\mathfrak{f}$-form.}
We now pass to the case when $(M',\lambda')=\mathfrak{f}(n',0,\xi,\mathbb{F})$. The strategy is analogous, but additional difficulties arise because ${}^\F\! B_\xi(t)$ is not irreducible in $\L{\F}$.
We start with item (a) and write ${}^\F\! B_\xi(t)={}^\F\! B_{\xi+}(t){}^\F\! B_{\xi-}(t)$.
Let $x'$ be a generator of $(M',\lambda')$
such that $\lambda'(x',x')=\frac{1}{{}^\F\! B_\xi(t)^{n'}}$. As in the proof of Lemma~\ref{lem:submodule}(b), we choose a generator $x$ of $M$ in such a way that $\iota(x)={}^\F\! B_{\xi+}(t)^a{}^\F\! B_{\xi-}(t)^bx'$. By definition of $\lambda$, we therefore obtain
$$\lambda(x,x)=\lambda'(\iota(x),\iota(x))=\frac{1}{{}^\F\! B_\xi(t)^{n'-a-b}}.$$
The map $\pi$ is a surjection, so it takes a generator of $M$ to a generator of $M''$. Thus, if we set~$x''=~\pi(x)$, then $\lambda''(x'',x'')=\frac{1}{{}^\F\! B_\xi(t)^{n'-a-b}}$ and so $(M'',\lambda'')=\mathfrak{f}(n'-a-b,0,\xi,\mathbb{F})$. In particular we have $a+b<n'$. We now have $\ord(M')={}^\F\! B_\xi(t)^{n'}$, $\ord(M)={}^\F\! B_{\xi+}(t)^{n'-a}{}^\F\! B_{\xi-}(t)^{n'-b}$ as well~as~$\ord(M'')=~{}^\F\! B_\xi(t)^{n'-a-b}$,
so item~(a) of the proposition follows readily.
We now prove item $(b)$. We first show that $\iota(M)={L'}^\perp$. Using (a), note that $\pi({}^\F\! B_\xi(t)^{n'-a-b}x)=~0$ and it follows that $L=\ker(\pi)$ is generated by ${}^\F\! B_\xi(t)^{n'-a-b}x$. This implies that $L'=\iota(L)$ is
generated by $z:={}^\F\! B_{\xi+}(t)^{n'-b}{}^\F\! B_{\xi-}(t)^{n'-a}x'$. To calculate ${L'}^\perp$, we observe that if $y:={}^\F\! B_{\xi+}(t)^c{}^\F\! B_{\xi-}(t)^dx'$,~then
\[\lambda'(z,y)=\frac{{}^\F\! B_{\xi+}(t)^{n'-b+d}{{}^\F\! B_{\xi-}(t)^{n'-a+c}}}{{}^\F\! B_\xi(t)^{n'}}.\]
In particular $\lambda'(z,y)=0$ if and only if $c\ge a$ and $d\ge b$. This shows that ${L'}^\perp$ is generated by~${}^\F\! B_{\xi+}(t)^a{}^\F\! B_{\xi-}(t)^bx'$. It follows that is $\iota(M)={L'}^\perp$, as desired.
We conclude by studying the sublagrangian reduction of $M'$ with respect to $L'$.
Write $\lambda_{L'}$ for the form on ${L'}^{\perp}/L'$ obtained by the sublagrangian reduction. The linking forms $(M'',\lambda'')$ and~$({L'}^\perp/L',\lambda_L')$
have the same underlying module and are non-singular $\mathfrak{f}$-type forms, hence they are isometric. This concludes the proof of the lemma.
\end{proof}
We now deduce Theorem~\ref{thm:isoprojection} from Proposition~\ref{prop:isoprojection_basic}.
\begin{proof}[Proof of Theorem~\ref{thm:isoprojection}]
As $(M',\lambda')$ is non-singular, Theorem~\ref{thm:non-degenerate-split} implies that it is split. We can therefore write $(M',\lambda')$ as a direct sum
\[(M',\lambda')=\bigoplus_{i\in I} (M_i',\lambda_i'),\]
where the summands $(M'_i,\lambda'_i)$ are basic and $I$ is a finite set of indices. Take $i\in I$, set $M_i=\iota^{-1}(M'_i)$ and $M''_i=\pi(M_i)$.
We have the following three cases:
\begin{itemize}
\item[(i)] $M_i$ is trivial;
\item[(ii)] $M_i$ is non-trivial, but $M''_i$ is trivial;
\item[(iii)] $M''_i$ is non-trivial.
\end{itemize}
\begin{lemma}\label{lem:divides}
The product $\ord(M_i)\ord(M_i)^\#$ divides $\ord(M''_i)\ord(M'_i)$. The equality \[\ord(M_i)\ord(M_i)^\#\doteq \ord(M''_i)\ord(M'_i)\] never holds in case (i), always holds in case (iii), and holds in case (ii) if and only if $\iota(M_i)$ is a lagrangian in $M'_i$.
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{lem:divides}]
In case (i), we have $\ord(M_i)=\ord(M''_i)=1$ and $\ord(M'_i)\neq 1$, so the conclusion is immediate. Case (iii) follows from Proposition~\ref{prop:isoprojection_basic}(a).
In order to treat case~(ii), first note that since $\pi$ is a morphism of linking forms, $\lambda$ must be trivial.
We will now use Lemma~\ref{lem:submodule}. Suppose that $M_i'=\mathfrak{f}(n',0,\xi,\mathbb{F})$ (that is, case (b) of Lemma~\ref{lem:submodule},
case (a) is analogous). As $\lambda$ is trivial, we must have $a+b\ge n'$ and we can write
\[\ord(M_i)\ord(M_i)^\#={}^\F\! B_{\xi+}(t)^{2n'-a-b}{}^\F\! B_{\xi-}(t)^{2n'-a-b}={}^\F\! B_\xi(t)^{2n'-a-b}.\]
On the other hand, $\ord(M_i'')=1$ and $\ord(M_i')={}^\F\! B_\xi(t)^{n'}$. Thus $\ord(M_i)\ord(M_i)^\#$ always divides $\ord(M'_i)\ord(M''_i)$ and the equality holds
if and only if $a+b=n'$. Looking carefully at the proof of Case II item (b) of Proposition~\ref{prop:isoprojection_basic}, the latter condition means that~$\iota(M_i)$ is equal to its orthogonal complement in $M_i'$.
\end{proof}
\begin{corollary}\label{cor:divides}
The product $\ord(M)\ord(M)^\#$ divides $\ord(M')\ord(M'')$ and \eqref{eq:order} holds if and only if case (i) is excluded (i.e. none of the $M_i$ is trivial)
and, if for some $i\in I$ the module $M''_i=\pi(M_i)$ is trivial, then $\iota(M_i)$ is lagrangian in $M_i'$.
\end{corollary}
\begin{proof}
We have $M'=\oplus M_i'$ and, since $M_i:=\iota^{-1}(M_i')$ and $\iota$ is injective, we also have $M=\oplus M_i$. As $\pi$ is a surjection we infer that $M''=\oplus M''_i$. The corollary follows from Lemma~\ref{lem:divides} and the multiplicativity of the orders.
\end{proof}
\emph{Conclusion of the proof of Theorem~\ref{thm:isoprojection}.}
We first show that $\iota(M)=L^\perp$. Equation~\eqref{eq:order} and Corollary~\ref{cor:divides} imply that for any~$i\in~I$, the summand $M_i$ is non-trivial and, if for some $i \in I$ the module $M_i''=\pi(M_i)$ is trivial, then $\iota(M_i)$ is trivial. Set $L_i:=\ker(\pi|_{M_i})$ and $L_i':=\iota(L_i)$, and note that Proposition~\ref{prop:isoprojection_basic} gives $\iota(M_i)={L_i'}^{\perp_i}$,
where $\perp_i$ means that we take the orthogonal complement only in $M'_i$ and not in the whole of $M'$. Recalling that $L=\ker(\pi)$, we have now $L=\bigoplus L_i$ and $L'=\iota(L)=\bigoplus L_i'$. We can now conclude the proof of the first assertion:
$$ \iota(M)=\iota(\bigoplus M_i)=\bigoplus \iota(M_i)=\bigoplus {L_i'}^{\perp_i}=L^\perp. $$
We must now show that $(M'',\lambda'')$ is isometric to the sublagrangian reduction of $M'$ along $L'$. First note that this assumptions holds for the summands $M_i'',M_i'$ and $L_i'$:
in case (ii), $M_i''$ is trivial, and since we saw above that $\iota(M_i)=L_i^{\perp_i}$, we deduce that $L_i^{\perp_i}/L_i$ is also trivial. In case (iii), this follows from the second statement in Proposition~\ref{prop:isoprojection_basic}. The statement on $M''$ now follows since we have $M''=\pi(M)=\bigoplus \pi(M_i)=\bigoplus M_i''$ (by definition of $M_i''$) as well as ${L'}^{\perp}/L'=\bigoplus {L'_i}^{\perp_i}/L'_i$ (where the sum is orthogonal).
This concludes the proof of the theorem.
\end{proof}
\section{Signatures of linking forms}
\label{sec:Signatures}
One of the most important invariants of a linking form are signature invariants. Since there are several non-equivalent definitions of such invariants, our aim is to gather the different definitions and to explain the relation between them. In Subsection~\ref{sub:SignatureJump} we give a definition of the local signature jump, which echoes Milnor's definition~\cite{MilnorInfiniteCyclic}.
These signature jumps are used in Subsection~\ref{sub:sig_function} to define the signature function. This function is reminiscent of the classical Levine--Tristram signature of a knot~\cite{Tristram, LevineMetabolicHyperbolic}: it is locally constant and only becomes Witt invariant after replacing its value at discontinuity points by the average of the one-side limits. In Subsection~\ref{sec:sigrep} we study the signatures of forms that are
representable and give a formula for the signature function of a representable form in terms of the signatures of a matrix that represents it. As a consequence, we give an
explicit obstruction to representability. Subsection~\ref{sec:JumpIsJump} contains the proof of Proposition~\ref{prop:JumpIsJump}, which is the main technical result in this section.
\subsection{Signature jumps}
\label{sub:SignatureJump}
In this subsection, we introduce signature jumps and explain how they obstruct a linking form from being metabolic.
\medbreak
Recall from Definition~\ref{def:hodge_number} that the \emph{Hodge number} $\mathcal{P}(n,0,\epsilon,\xi,\mathbb{F})$ of a non-singular linking form $(M,\operatorname{\lambda})$ is the number of times $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{F})$ enters its decomposition. Recall furthermore from Theorems~\ref{thm:WittClassificationReal} and~\ref{thm:WittClassificationComplex} that every non-singular linking form is Witt equivalent to a direct sum of the forms $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{F})$ in which each $n$ is odd. Motivated by this result, we introduce the following terminology.
\begin{definition}\label{def:sigjump}
Let $(M,\operatorname{\lambda})$ be a non-singular linking form over $\L{\F}$. If the linking form is real, then the \emph{signature jump} of $(M,\operatorname{\lambda})$ at $\xi \in S^1_+$ is defined as
\[\delta\sigma_{(M,\lambda)}(\xi)=\sum_{\substack{n \textrm{ odd}\\ \epsilon=\pm 1}} \epsilon \mathcal{P}(n,\epsilon,\xi,\mathbb{R}).\]
The signature jump at $\xi \in S^1_-$ is defined as $\delta\sigma_{(M,\lambda)}(\xi)=-\delta\sigma_{(M,\lambda)}(\ol{\xi})$.
If the linking form is complex, then the \emph{signature jump} of $(M,\operatorname{\lambda})$ at $\xi \in S^1$
is defined as
\[\delta\sigma_{(M,\lambda)}(\xi)=-\sum_{\substack{n \textrm{ odd}\\ \epsilon=\pm 1}} \epsilon \mathcal{P}(n,\epsilon,\xi,\mathbb{C}).\]
\end{definition}
Note that a linking form only admits a finite number of non-zero signature jumps. On the other hand, the appearance of the minus sign in Definition~\ref{def:sigjump} will become clear in Proposition~\ref{prop:JumpIsJump}.
Here is an example of Definition~\ref{def:sigjump}. We refer to Subsection~\ref{sec:blanchf-forms-twist} for more involved computations.
\begin{example}
\label{ex:Trefoil}
Consider the non-singular linking form given by $(x,y) \mapsto \frac{xy^\#}{t-1+t^{-1}}$. Observe that~$\omega=e^{2\pi i/6}$ and its complex conjugate are the only roots of $t-1+t^{-1}$. It follows that, as a real linking form, $(M,\lambda)$ is isometric to $\mathfrak{e}(1,0,1,\omega,\mathbb{R})$. We therefore deduce that
$$ \delta\sigma_{(M,\lambda)}(\xi)
=\begin{cases}
-1 & \text{ if } \xi=\overline{\omega}, \\
1 & \text{ if } \xi=\omega, \\
0 &\text{ otherwise.}
\end{cases}
$$
\end{example}
Next, we observe that the signature jumps obstruct a linking form from being metabolic.
\begin{theorem}
\label{thm:metabolic_signature}
When $\mathbb{F}=\mathbb{R},\mathbb{C}$, the signature jumps define homomorphisms from~$W(\O{\F},\L{\F})$ to $\mathbb{Z}$. Moreover, a linking form over $\L{\F}$ is metabolic if and only if all its signature jumps vanish.
\end{theorem}
\begin{proof}
The first assertion follows by combining Definition~\ref{sub:SignatureJump} with Theorems~\ref{thm:WittClassificationReal} and~\ref{thm:WittClassificationComplex}: these results show that a basic linking form contributes to the signature jump if and only if it is not Witt trivial. To prove the second assertion, we use Theorems~\ref{thm:WittClassificationReal} and~\ref{thm:WittClassificationComplex} to assume without loss of generality that the linking form is (Witt equivalent to) a direct sum of $\mathfrak{e}(n,0,\epsilon,\xi,\mathbb{F})$, where each~$n$ is odd. Under the isomorphism
\begin{equation}\label{eq:displayed_isomorphism}
W(\O{\F},\L{\F},{}^\R\! B_\xi(t)) \cong \mathbb{Z}
\end{equation}
described in Lemmas~\ref{lem:xiless1} and~\ref{lem:xiless1a}, Definition~\ref{def:sigjump} implies that the aforementioned direct sum of basic forms is Witt trivial if and only if the corresponding signature jumps vanish.
\end{proof}
\subsection{The signature function and the average signature function}\label{sub:sig_function}
In this subsection, we use the signature jumps in order to define the signature function of a non-singular linking form.
\medbreak
In a nutshell, a signature function of a linking form $(M,\lambda)$ can be defined as follows: fix the value of the function at a point $\xi_0\in S^1$ and define the function at $\xi \in S^1 \setminus \lbrace \xi_0 \rbrace$ by adding up the signature jumps between $\xi_0$ and $\xi$ (going anticlockwise). For reasons discussed below, we shall fix the value of the signature function to be $\delta\sigma_{(M,\lambda)}(1)$ as $\xi$ approaches $1$ going clockwise along $S^1$.
With this intuition in mind, we now give the precise definition of the signature function.
\begin{definition}\label{def:sig_func}
Suppose $(M,\lambda)$ is a non-singular linking form. The \emph{signature function} of~$(M,\lambda)$ is the map $\sigma_{(M,\lambda)} \colon S^1 \to \mathbb{Z}$ whose value at $\xi_1=e^{2\pi i\theta_1}$ is defined as
\begin{equation}\label{eq:def_sig_func}
\sigma_{(M,\lambda)}(\xi_1)=\sum_{\tau\in(0,\theta_1)} 2\delta\sigma_{(M,\lambda)}(e^{2\pi i \tau})-\sum_{\substack{\epsilon=\pm 1\\n\textrm{ even}}} \epsilon \mathcal{P}(n,\epsilon,\xi_1,\mathbb{F})+\delta\sigma_{(M,\lambda)}(\xi_1)+\delta\sigma_{(M,\lambda)}(1).
\end{equation}
\end{definition}
Note that this signature function differs from the fomula in~\cite[Proposition 4.14]{BorodzikNemethi} by an overall constant. In our case, we have $\lim_{\theta\to 0^+}\sigma_{(M,\lambda)}(e^{i\theta})=\delta\sigma_{(M,\lambda)}(1)$ while in~\cite{BorodzikNemethi}, this is not always the case.
We now describe some basic properties of the signature function.
\begin{proposition}
\label{prop:BasicProperties}
Let $(M,\lambda)$ be a non-singular linking form over $\L{\F}$ and let $\Delta_M(t)$ denote the order of the $\L{\F}$-module $M$.
\begin{enumerate}
\item The signature function is constant on $S^1 \setminus \lbrace \xi \in S^1 \ | \ \Delta_M(\xi)=0 \rbrace$.
\item If $(M,\lambda)$ is a real linking form, then we have $\sigma_{(M,\lambda)}(\overline{\xi})=\sigma_{(M,\lambda)}(\xi)$.
\end{enumerate}
\end{proposition}
\begin{proof}
By Remark~\ref{rem:OrderAndClassification}, the zeros of $\Delta_M(t)$ are precisely the values of $\xi$ which enter the decomposition of~$(M,\lambda)$ into basic forms. The first property now follows from the definition of~$\sigma_{(M,\lambda)}$. The second assertion is an immediate consequence of the definition of the signature jump over~$\L{\R}$.
\end{proof}
Next, we move on to the behavior of the signature function under Witt equivalence. Clearly~$\sigma_{(M,\lambda)}$ is not well defined on Witt groups: for instance, $\mathfrak{e}(2,0,1,\xi,\mathbb{F})$ is metabolic (see Lemmas~\ref{lem:LRxi} and~\ref{lem:xiless1a}) but its signature function at $\xi$ is equal to $1$. More generally, this behavior of the signature function can be traced back to the following quantity which is not Witt invariant:
\begin{equation}\label{eq:sigma_loc}
\sigma^{loc}_{(M,\lambda)}(\xi):=-\sum_{\substack{\epsilon=\pm 1\\n\textrm{ even}}} \epsilon \mathcal{P}(n,\epsilon,\xi,\mathbb{F}).
\end{equation}
Since we saw in Theorem~\ref{thm:metabolic_signature} that the signature jumps $\delta \sigma_{(M,\lambda)}$ are invariant under Witt equivalence, we deduce that the signature function defines a function on the Witt group for all but finitely many values of $\xi$. In order to obtain a function which is Witt invariant on the whole circle, we use a well known construction which consists in taking averages at each point.
\begin{definition}\label{def:average_signature}
The \emph{averaged signature} $\sigma^{av}_{(M,\lambda)}(\xi)$ is defined as
\begin{equation}\label{eq:average_signature}
\sigma^{av}_{(M,\lambda)}(\xi)=\frac12\left(\lim_{\theta\to 0^+}\sigma_{(M,\lambda)}(e^{i\theta}\xi)+\lim_{\theta\to 0^-}\sigma_{(M,\lambda)}(e^{i\theta}\xi)\right).\end{equation}
\end{definition}
The next lemma relates the averaged signature to the signature jumps and proves its invariance under Witt equivalence. Briefly, averaging gets rid of the $\sigma_{(M,\lambda)}^{\text{loc}}$ term in the definition of~$\sigma_{(M,\lambda)}$.
\begin{proposition}
\label{prop:avsig}
Let $(M,\lambda)$ be a non-singular linking form. The averaged signature function of~$(M,\lambda)$ at $\xi_1=e^{2\pi i\theta_1}$ can be described as follows:
\begin{equation}
\label{eq:jump_av}
\sigma^{av}_{(M,\lambda)}(\xi_1)=
\begin{cases}
\sigma_{(M,\lambda)}(\xi_1)-\sigma^{loc}_{(M,\lambda)}(\xi_1) & \text{if } \xi_1 \neq 1, \\
\sum \limits_{\xi\in S^1}\delta\sigma_{(M,\lambda)}(\xi) & \text{if } \xi_1=1.
\end{cases}
\end{equation}
In particular, the averaged signature function $\sigma^{av}_{(M,\lambda)}$ is invariant under Witt equivalence. In the real case, we have $\sigma_{(M,\lambda)}^{\text{av}}(1)=\delta\sigma_{(M,\lambda)}(1)$.
\end{proposition}
\begin{proof}
We first assume that $\xi_1 \neq 1$. Let $\theta_0< 1$ be small enough so that for all $\theta\in[-\theta_0,\theta_0]\setminus\{0\}$ and all $n,\epsilon$, one has
$\mathcal{P}(n,\epsilon,e^{2\pi i \theta}\xi,\mathbb{F})=0$. Note that such a $\theta_0$ always exists, because $\mathcal{P}(n,\epsilon,\xi',\mathbb{F})$ can only be non-zero only for finitely many $\xi'$.
Using successively the definition of the averaged signature function and of the signature function (and assuming that $\theta\in (0,\theta_0)$ at each of these two steps),
we obtain
\begin{align*}
\sigma_{(M,\lambda)}^{av}(\xi_1)
&=\frac12(\sigma_{(M,\lambda)}(e^{2\pi i\theta}\xi)+\sigma_{(M,\lambda)}(e^{-2\pi i\theta}\xi)) \\
&=\frac12\sum_{\tau\in(0,\theta_1+\theta_0)}2\delta\sigma_{(M,\lambda)}(e^{2\pi i\tau})+\frac12\sum_{\tau\in(0,\theta_1-\theta_0)}2\delta\sigma_{(M,\lambda)}(e^{2\pi i\tau})+\delta\sigma_{(M,\lambda)}(1).
\end{align*}
Next, grouping part of the second sum into the first, using once again our particular choice of~$\theta$ and recalling the definitions of
$\sigma_{(M,\lambda)}(\xi_1)$ and $\sigma^{loc}_{(M,\lambda)}(\xi_1)$, we get the desired equation:
\begin{align*}
\sigma_{(M,\lambda)}^{av}(\xi_1)
&=\sum_{\tau\in(0,\theta_1-\theta_0)}2\delta\sigma_{(M,\lambda)}(e^{2\pi i\tau})+\frac12\sum_{\tau\in(\theta_1-\theta_0,\theta_1+\theta_0)}2\delta\sigma_{(M,\lambda)}(e^{2\pi i\tau})+\delta\sigma_{(M,\lambda)}(1) \\
&=\sum_{\tau\in(0,\theta_1)}2\delta\sigma_{(M,\lambda)}(e^{2\pi i\tau})+\delta\sigma_{(M,\lambda)}(\xi_1)+\delta\sigma_{(M,\lambda)}(1) \\
&= \sigma_{(M,\lambda)}(\xi_1)-\sigma^{loc}_{(M,\lambda)}(\xi_1).
\end{align*}
Next, we assume that $\xi_1=1$. Using the definition of the averaged signature function and the definition of the signature function, we immediately see that
\begin{equation}
\label{eq:AveragedAtOne}
\sigma^{\text{av}}_{(M,\lambda)}(1)=\frac{1}{2}\left( \lim_{\theta\to 0^+} \sigma_{(M,\lambda)}(e^{2\pi i\theta})+\lim_{\theta\to 1^-} \sigma_{(M,\lambda)}(e^{2\pi i\theta}) \right)=\frac{1}{2}\left(\delta\sigma_{(M,\lambda)}(1)+\lim_{\theta\to 1^-}
\sigma_{(M,\lambda)}(e^{2\pi i\theta}) \right).
\end{equation}
It therefore only remains to deal with the $\lim_{\theta\to 1^-} \sigma_{(M,\lambda)}(e^{2\pi i\theta})$ term. Since the signature function is obtained by summing the signature jumps along the circle,
we deduce that $\lim_{\theta\to 1^-} \sigma_{(M,\lambda)}(e^{2\pi i\theta})=\sum_{\tau\in(0,1)}2\delta\sigma_{(M,\lambda)}(e^{2\pi i \tau})+\delta\sigma_{(M,\lambda)}(1)$. Plugging this back into~(\ref{eq:AveragedAtOne}) immediately concludes the proof of~(\ref{eq:jump_av}).
The Witt invariance of the averaged signature now follows from Theorem~\ref{thm:metabolic_signature}: the signature jumps are already known to be invariant under Witt equivalence. Finally, the last statement is immediate: in the real case, the definition of the signature jump implies that~$\sum_{\xi \in S^1 \setminus \lbrace 1 \rbrace} \delta \sigma_{(M,\lambda)}(\xi)=0$ and therefore the equation $\sigma_{(M,\lambda)}^{\text{av}}(1)=\delta\sigma_{(M,\lambda)}(1)$ follows from~(\ref{eq:jump_av}).
\end{proof}
The statement of Theorem~\ref{thm:metabolic_signature} can now be rephrased in terms of the signature function.
\begin{corollary}\label{cor:metabolic_signature}
Given a non-singular linking form $(M,\lambda)$ over $\L{\F}$, the following assertions are equivalent:
\begin{enumerate}
\item The linking form $(M,\lambda)$ is metabolic.
\item The signature function $\sigma_{(M,\lambda)}$ is zero for all but finitely many values of $\xi$.
\item The averaged signature function $\sigma^{av}_{(M,\lambda)}$ is zero.
\end{enumerate}
\end{corollary}
\subsection{Signatures of representable forms}\label{sec:sigrep}
We study the signature function of representable linking forms. Namely, if $(M,\lambda)$ is a linking form represented by a matrix $A(t)$, then we will relate the signatures of (evaluations of) $A(t)$ to the signatures of the pairing.
\medbreak
Given a Hermitian matrix $A(t)$ over $\L{\F}$ and $\xi\in S^1$, we write $\operatorname{sign} A(\xi)$ for the signature of the complex Hermitian matrix $A(\xi)$. If $A(t)$ represents a linking form $(M,\lambda)$, then the signature~$\operatorname{sign} A(\xi)$ is not an invariant of
the linking form: for instance $A(t) \oplus (1)$ also represents~$(M,\lambda)$ but $\operatorname{sign} (A(t) \oplus (1))(\xi)=\operatorname{sign} A(\xi) +1$. On the other hand, the following result shows that~$\operatorname{sign} A(\xi)-\operatorname{sign} A(\xi')$ is an invariant of $(M,\lambda)$ for any $\xi,\xi' \in S^1$ (see also~\cite[Lemma 3.2]{BorodzikFriedl}).
\begin{proposition}\label{prop:signature_well_defined}
Let $(M,\lambda)$ be a non-singular linking form over $\L{\F}$. If $(M,\lambda)$ is represented by~$A(t)$, then for any $\xi,\xi'$ in $S^1$, the difference $\operatorname{sign} A(\xi)-\operatorname{sign} A(\xi')$ does not depend on~$A(t)$.
\end{proposition}
\begin{proof}
We begin the proof by recalling
a result which is implicit in~\cite{RanickiExact} and which is stated in~\cite[Proposition 2.2]{BorodzikFriedl2} (see also~\cite[Proposition 3.1]{BorodzikFriedl}).
\begin{proposition}
\label{prop:Ranicki}
Let $A$ and $B$ be Hermitian matrices over $\L{\F}$ with non-zero determinant. The linking forms $\lambda_A$ and $\lambda_B$ are isometric if and only if $A$ and $B$ are related by a sequence of the three following moves:
\begin{enumerate}
\item Replace $C$ by $PC{P^\#}^T$, where $P$ is a matrix over $\L{\F}$ with $\det(P)$ a unit of $\L{\F}$.
\item Replace $C$ by $C \oplus D$, where $D$ is a Hermitian matrix over $\L{\F}$ with $\det(D)$ a unit.
\item The inverse of $(2)$.
\end{enumerate}
\end{proposition}
From Proposition~\ref{prop:Ranicki}, we quicky conclude the proof of Proposition~\ref{prop:signature_well_defined} because the operation (1) does not change the signature of $A(\xi)$
and $A(\xi')$, while the operation (2) changes both signatures by the same number, compare \cite[proof of Lemma 3.2]{BorodzikFriedl}.
\end{proof}
The goal of the next few results is to show how the signature function of a representable linking form can be computed from the (difference of) signatures of its representing matrices. The following proposition takes the first step in this process by studying signature jumps.
\begin{proposition}
\label{prop:JumpIsJump}
Let $\xi\in S^1$. If a non-singular linking form $(M,\lambda)$ is representable by a matrix~$A(t)$, then the following equation holds:
\begin{equation}\label{eq:jumpisjump}
\lim_{\theta\to 0^+} \operatorname{sign} A(e^{i \theta}\xi)-\lim_{\theta\to 0^-}\operatorname{sign} A(e^{i \theta}\xi)=2\delta\sigma_{(M,\lambda)}(\xi).
\end{equation}
Moreover, one also has
\begin{equation}\label{eq:jumpisjump_half}
\operatorname{sign} A(\xi)-\lim_{\theta\to 0^-}\operatorname{sign} A(e^{i \theta}\xi)=\delta\sigma_{(M,\lambda)}(\xi)+\sigma^{loc}_{(M,\lambda)}(\xi).
\end{equation}
\end{proposition}
We delay the proof of Proposition~\ref{prop:JumpIsJump} to Section~\ref{sec:JumpIsJump} in order to describe its consequences. The first corollary shows that, up to the signature jump of $(M,\lambda)$ at $1$, one can compute the signature function of $(M,\lambda)$ from the signature of a representing matrix.
\begin{corollary}\label{cor:sigissig}
If $(M,\lambda)$ is represented by a matrix $A(t)$, then for any $\xi_0\in S^1$ one has
\begin{equation}\label{eq:sigissig}
\sigma_{(M,\lambda)}(\xi_0)=\operatorname{sign} A(\xi_0)-\lim_{\theta\to 0^+} \operatorname{sign} A(e^{i\theta})+\delta\sigma_{(M,\lambda)}(1).
\end{equation}
\end{corollary}
\begin{proof}
First, suppose that $\det A(\xi_0)\neq 0$. Let $\theta_0 \in (0,1]$ be such that $e^{2\pi i\theta_0}=\xi_0$ and let
$\xi_1,\ldots,\xi_n$ be the elements in $S^1$ for which $\det A(\xi_j)=0$ and $\xi_j=e^{2\pi i\theta_j}$ for some $\theta_j\in(0,\theta_0)$.
Assuming that $\xi_1,\ldots,\xi_n$ are ordered by increasing arguments, we claim that both sides of~\eqref{eq:sigissig} are equal to
\[2\sum_{j=1}^n \delta\sigma_{(M,\lambda)}(\xi_j)+\delta\sigma_{(M,\lambda)}(1).\]
For the left-hand side of \eqref{eq:sigissig}, this is the definition of the signature, and so we deal with the right-hand side.
As the signature function $\xi\mapsto \operatorname{sign} A(\xi)$ is constant on the subset of $S^1$ on which~$A(t)$ is invertible, the definition of the $\xi_j$ implies that
\[\operatorname{sign} A(\xi_0)-\lim_{\theta\to 0^+}\operatorname{sign} A(e^{i\theta})=\sum_{j=1}^n \left( \lim_{\theta\to 0^+}\operatorname{sign} A(e^{i\theta}\xi_j)-\lim_{\theta\to 0^-}\operatorname{sign} A(e^{i\theta}\xi_j) \right).\]
By the first point of Proposition~\ref{prop:JumpIsJump}, i.e by applying~\eqref{eq:jumpisjump} $n$ times, we obtain
\[\sum_{j=1}^n \left( \lim_{\theta\to 0^+}\operatorname{sign} A(e^{i\theta}\xi_j)-\lim_{\theta\to 0^-}\operatorname{sign} A(e^{i\theta}\xi_j) \right)=2\sum_{j=1}^n\delta\sigma_{(M,\lambda)}(\xi_j).\]
Hence
\[\operatorname{sign} A(\xi_0)-\lim_{\theta\to 0^+}\operatorname{sign} A(e^{i\theta})=\sum_{j=1}^n2\delta\sigma_{(M,\lambda)}(\xi_j).\]
This concludes the proof of the corollary in the case where $\det A(\xi_0) \neq 0$.
Next, we suppose that $\det A(\xi_0)=0$. Choose $\theta>0$ so that if $0<|\theta'|<\theta$, then we have $\det A(e^{i\theta'}\xi_0)\neq 0$. Using this condition, we apply the first part of the proof to $\xi'=e^{i\theta'}$ and obtain
$$\sigma_{(M,\lambda)}(\xi')=\operatorname{sign} A(\xi')-\lim_{\theta\to 0^+} \operatorname{sign} A(e^{i\theta})+\delta\sigma_{(M,\lambda)}(1).$$
In order to deduce the corresponding equality for $\xi_0$, note that since $\xi'$ is close to $\xi_0$, the definition of the signature function implies that $\sigma_{(M,\lambda)}(\xi_0)-\sigma_{(M,\lambda)}(\xi')=\delta\sigma_{(M,\lambda)}(\xi_0)+\sigma^{loc}_{(M,\lambda)}(\xi_0)$.
Combining these two observations, we see that
$$ \sigma_{(M,\lambda)}(\xi_0)=\delta\sigma_{(M,\lambda)}(\xi_0)+\sigma^{\text{loc}}_{(M,\lambda)}(\xi_0)+\operatorname{sign} A(\xi')-\lim_{\theta\to 0^+} \operatorname{sign} A(e^{i\theta})+\delta\sigma_{(M,\lambda)}(1). $$
Consequently, to conclude the proof, it suffices to show that $\delta\sigma_{(M,\lambda)}(\xi_0)+\sigma^{\text{loc}}_{(M,\lambda)}(\xi_0)+\operatorname{sign} A(\xi')$ is equal to $\operatorname{sign} A(\xi_0)$. Now this follows immediately from~\eqref{eq:jumpisjump_half} by noting that since $A(e^{i\theta})$ is non-singular for $\theta\in[\theta',\theta_0)$, we have $\operatorname{sign} A(\xi')=\lim_{\theta\to 0^{-}}\operatorname{sign} A(e^{i\theta}\xi_0)$.
\end{proof}
The second consequence of Proposition~\ref{prop:JumpIsJump} is the analogue of Corollary~\ref{cor:sigissig} for the averaged signature. Before stating this result, we introduce some notation. Namely, for $\xi\in S^1$, we write
\[\operatorname{sign}^{av}A(\xi)=\frac12\left(\lim_{\theta\to 0^+}\operatorname{sign} A(e^{i\theta}\xi)+\lim_{\theta\to 0^-}\operatorname{sign} A(e^{i\theta}\xi)\right)\]
and refer to it as the \emph{averaged signature} of $A(t)$ at $\xi$. The averaged signature of $(M,\lambda)$ can now be described in terms of the averaged signature of $A(t)$, without the extra $\delta \sigma_{(M,\lambda)}(1)$ term.
\begin{corollary}\label{cor:sigissig2}
If $(M,\lambda)$ is represented by a matrix $A(t)$, then for any $\xi \in S^1$ one has
\begin{equation}\label{eq:sigissig2}
\sigma^{av}_{(M,\lambda)}(\xi)
=\operatorname{sign}^{av} A(\xi)-\operatorname{sign}^{av}A(1).
\end{equation}
\end{corollary}
\begin{proof}
Combining Corollary~\ref{cor:sigissig} with the definition of the averaged signature, we obtain that
\begin{equation}\label{eq:sigthirdpart}
\sigma^{av}_{(M,\lambda)}(\xi)=\operatorname{sign}^{av}A(\xi)-\lim_{\theta\to 0^+}\operatorname{sign}^{av} A(e^{i\theta})+\delta\sigma_{(M,\lambda)}(1).
\end{equation}
Consequently, to conclude the proof of the corollary, it only
remains to show that
\begin{equation}\label{eq:withlimit}\lim_{\theta\to 0^+}\operatorname{sign}^{av} A(e^{i\theta})-\delta\sigma_{(M,\lambda)}(1)=\operatorname{sign}^{av}A(1).\end{equation}
Choose $\theta_0>0$ in such a way that for any $\theta\in(-\theta_0,\theta_0)$, $\theta\neq 0$, we have $\det A(e^{i\theta})\neq 0$. Assuming that~$\theta'$ and $\theta''$ lie in $(0,\theta_0)$, we obtain
\begin{align*}
\lim_{\theta\to 0^+}\operatorname{sign} A(e^{i\theta})&=\lim_{\theta\to 0^+}\operatorname{sign}^{av} A(e^{i\theta})=\operatorname{sign} A(e^{i\theta'})=\operatorname{sign}^{av}A(e^{i\theta'}),\\
\lim_{\theta\to 0^-}\operatorname{sign} A(e^{i\theta})&=\lim_{\theta\to 0^-}\operatorname{sign}^{av} A(e^{i\theta})=\operatorname{sign} A(e^{-i\theta''})=\operatorname{sign}^{av}A(e^{-i\theta''}).
\end{align*}
This allows us to deduce that
\begin{align*}
\operatorname{sign}^{av} A(1)&=\frac12\left(\operatorname{sign} A(e^{i\theta'})+\operatorname{sign} A(e^{-i\theta''})\right),\\
\delta\sigma_{(M,\lambda)}(1)&=\frac12\left(\operatorname{sign} A(e^{i\theta'})-\operatorname{sign} A(e^{-i\theta''})\right).
\end{align*}
Equation~\eqref{eq:withlimit} follows immediately, concluding the proof of the corollary.
\end{proof}
We can now prove the converse of Proposition~\ref{prop:inversejumps}
\begin{proposition}\label{prop:sumofjumps}
If $(M,\lambda)$ is a representable linking form over $\L{\F}$, then
\begin{equation}\label{eq:zerojump}\sum_{\xi\in S^1}\delta\sigma_{(M,\lambda)}(\xi)=0.\end{equation}
\end{proposition}
\begin{proof}
By Corollary~\ref{cor:sigissig2} applied to $\xi=1$, we infer that $\sigma^{av}_{(M,\lambda)}(1)=0$. Using Proposition~\ref{prop:avsig}, we also know that $\sigma^{av}_{(M,\lambda)}(1)=\sum_{\xi \in S^1} \delta\sigma_{(M,\lambda)}(\xi) $. Combining these two observations immediately leads to the desired result.
\end{proof}
Observe that Proposition~\ref{prop:sumofjumps} provides a second proof that $\mathfrak{e}(2n+1,0,\epsilon,\xi,\mathbb{C})$ is not representable, recall Proposition~\ref{prop:verystupidexample}. In fact, since Theorem~\ref{thm:metabolic_signature} ensures that the signature jumps are Witt invariants, Proposition~\ref{prop:sumofjumps} shows that this linking form is not even Witt equivalent to a representable one, recall Remark~\ref{rem:Litherland}.
Combining Proposition~\ref{prop:inversejumps} and Proposition~\ref{prop:sumofjumps}, we obtain the following result.
\begin{corollary}
\label{cor:WittRepresentability}
Over $\mathbb{C}[t^{\pm 1}]$, metabolic forms are representable, representability is invariant under Witt equivalence and the following statements are equivalent:
\begin{enumerate}
\item A linking form is representable.
\item A linking form is Witt equivalent to a representable one.
\item The \emph{total signature jump}~(\ref{eq:zerojump}) of a linking form vanishes.
\end{enumerate}
\end{corollary}
\begin{proof}
Using Theorem~\ref{thm:metabolic_signature}, metabolic forms have vanishing total signature jump and therefore Proposition~\ref{prop:inversejumps} guarantees representability. The second assertion follows immediately, while the three equivalences are consequences of Propositions~\ref{prop:sumofjumps} and Proposition~\ref{prop:inversejumps}.
\end{proof}
We conclude this subsection with two remarks. First observe that Proposition~\ref{cor:WittRepresentability} contrasts strongly with the real case in which all linking forms are representable. Secondly, using Remark~\ref{rem:Litherland}, Corollary~\ref{cor:WittRepresentability} can be understood as providing a computation of the image of the boundary map $W(\mathbb{C}(t)) \to W(\mathbb{C}(t),\mathbb{C}[t^{\pm 1}])$.
\subsection{Proof of Proposition~\ref{prop:JumpIsJump}}\label{sec:JumpIsJump}
Suppose that $A(t)$ is a matrix over $\L{\F}$ which represents the non-singular linking form $(M,\lambda)$.
Our strategy is to start with a particular case to which the general case will later be reduced.
\emph{Case 1.} We first suppose that $A(t)$ is a diagonal matrix with Laurent polynomials $a_1(t),\ldots,a_n(t)$ on its diagonal. This assumption implies both that the module $M$
splits as the direct sum of the cyclic submodules~$M_j=\L{\F}/a_j(t)$ and that the restriction of $\lambda$ to each of the $M_j$ is given by
\begin{equation}\label{eq:lambdaj}
\lambda_j \colon M_j\times M_j\to\O{\F}/\L{\F},\ (x,y)\mapsto\frac{xy^\#}{a_j(t)}.
\end{equation}
Therefore, when $A(t)$ is diagonal, it is sufficient to check~\eqref{eq:jumpisjump} and \eqref{eq:jumpisjump_half} for each $(M_j,\lambda_j)$. From now on, we fix a $j$ and write $n_j$ for the order of $a_j$ at $t=\xi$.
Depending on the sign of $a_j(e^{i\theta}\xi)$ near $\xi$, we consider four cases (here $\theta$ is a small real number):
\begin{itemize}
\item \emph{Case (I):} $a_j(e^{i\theta}\xi)$ changes sign from positive to negative as $\theta$ goes from negative to positive;
\item \emph{Case (II):} $a_j(e^{i\theta}\xi)$ changes sign from negative to positive as $\theta$ goes from negative to positive;
\item \emph{Case (III):} $a_j(e^{i\theta}\xi)$ is negative for $\theta\neq 0$ and $a_j(\xi)=0$;
\item \emph{Case (IV):} $a_j(e^{i\theta}\xi)$ is positive for $\theta\neq 0$ and $a_j(\xi)=0$.
\end{itemize}
We calculate the jumps of the signatures and put them in a table:
\smallskip
\begin{center}
\begin{tabular}
{|c|c|c|c|c|}\hline
Case & I & II & III & IV \\\hline
parity of $n_j$ & odd & odd & even & even \\\hline
value of $\epsilon_j$ & $+1$ & $-1$ & $-1$ & $+1$ \\ \hline
$\delta\sigma_{(M,\lambda)}(\xi)$ & $-1$ & $1$ & $0$ & $0$ \\ \hline
$\sigma_{(M,\lambda)}^{loc}$ & $0$ & $0$ & $1$ & $-1$ \\\hline
l.h.s. of \eqref{eq:jumpisjump} & $-2$ & $2$ & $0$ & $0$ \\ \hline
l.h.s. of \eqref{eq:jumpisjump_half} & $-1$ & $1$ & $1$ & $-1$ \\\hline
\end{tabular}
\end{center}
\smallskip
The sign of $\epsilon_j$ is calculated using Proposition~\ref{prop:cyclic_classif} above.
The left-hand sides of \eqref{eq:jumpisjump} and \eqref{eq:jumpisjump_half} are easily calculated using the $1\times 1$ matrix $A_j=(a_j)$.
From the table it immediately follows that \eqref{eq:jumpisjump} and \eqref{eq:jumpisjump_half} hold for $1\times 1$ matrices, and hence, also in the case when the matrix $A(t)$ is diagonal of arbitrary size.
\smallskip
\emph{Case 2.} We want to reduce the general case to the diagonal case.
In the real case, this is possible thanks to Proposition~\ref{prop:diagonalreal}.
Therefore, from now on, we will assume that $\mathbb{F}=\mathbb{C}$.
Since the proofs of \eqref{eq:jumpisjump} and of \eqref{eq:jumpisjump_half} are very similar, we focus on the proof
the former,
leaving the proof
the latter to the reader.
In order to avoid dealing with the delicate task of deciding whether $A(t)$ is diagonalizable over~$\L{\C}$, we instead pass to a local ring $\mathcal{O}_\xi$.
More precisely, recall from Subsection~\ref{sub:LocalizationDiagonalization} that~$\mathcal{O}_\xi$ denotes the local ring
of analytic functions near~$\xi$ and that by Lemma~\ref{lem:analytic_change}, we can find size~$n$ matrices~$P(t)$ and~$B(t)$ over $\mathcal{O}_\xi$ such that $B(t)=P(t)A(t){P(t)^\#}^T$ is diagonal for all~$t$ close to $\xi$ and $P(t)$ is invertible for all $t$ close to $\xi$.
In particular, the signature jump of $B(t)$ at~$\xi$ is equal to the jump of the signature of~$A(t)$:
\begin{equation}
\label{eq:dsignAdsignB}
\lim_{\theta\to 0^+} \sgn A(e^{i \theta}\xi)-\lim_{\theta\to 0^-}\sgn A(e^{i \theta}\xi)=
\lim_{\theta\to 0^+} \sgn B(e^{i \theta}\xi)-\lim_{\theta\to 0^-}\sgn B(e^{i \theta}\xi).
\end{equation}
We cannot immediately deduce \eqref{eq:jumpisjump} since $B(t)$ might not represent $(M,\lambda)$ over~$\L{\C}$: $B(t)$ is not even a matrix over $\L{\C}$.
On the other hand, $B(t)$ does represents the linking form $(\widehat{M},\widehat{\lambda})$ over~$\mathcal{O}_\xi$, where
\[\wh{M}=\mathcal{O}_\xi^n/B^T\mathcal{O}_\xi^n\,\ \ \wh{\lambda}(x,y)=xB^{-1}y^\#\in \Omega_\xi/\mathcal{O}_\xi.\]
Use $b_1(t),\ldots,b_n(t)$ to denote the diagonal elements of $B(t)$.
As $B(t)$ is diagonal, the linking form~$(\widehat{M},\widehat{\lambda})$ splits as a direct sum of linking forms $(\wh{M}_j,\widehat{\lambda}_j)$, where $\wh{M}_j:=\mathcal{O}_\xi/b_j(t)\mathcal{O}_\xi$ and
\[ \widehat{\lambda}_j \colon \wh{M}_j\times \wh{M}_j\to\Omega_\xi/\mathcal{O}_\xi,\ (x,y)\mapsto\frac{xy^\#}{b_j(t)}.\]
Since $\widehat{M}_j$ is a cyclic $\mathcal{O}_\xi$-module, Proposition~\ref{prop:classify_cyclic} guarantees that $(\wh{M}_j,\widehat{\lambda}_j)$ is isometric to the linking form $\wh{\mathfrak{e}}(n_j,0,\xi,\epsilon_j,\mathbb{C})$, where $n_j$ is the order of zero of $b_j(t)$ at $t=\xi$ (note that $n_j$ might be zero, if $b_j(\xi)\neq 0$) and $\epsilon_j$ is described as follows:
\begin{itemize}
\item if $n_j$ is even, then $\epsilon_j$ is the sign of $b_j(t)/((t-\xi)(t^{-1}-\ol{\xi}))^{n_j/2}$ at $t=\xi$;
\item if $n_j$ is odd then if $b_j(t)/((t-\xi)((t-\xi)(t^{-1}-\ol{\xi}))^{n_j/2})$ is $\xi$-positive, we set $\epsilon=1$, otherwise we set $\epsilon=-1$.
\end{itemize}
Note that when $n_j$ is odd, the first possibility corresponds precisely to the situation where $\theta\mapsto b_j(e^{i\theta})$
changes sign from positive to negative, when $\theta$ crosses $\theta_0$.
Next, we define
\[\wh{\delta\sigma}_B(\xi):=\sum_{\substack{j=1,\ldots,n\\ n_i\textrm{ odd}}} \epsilon_j.\]
The same arguments as in Case~1 (that is, essentially the table) imply that
\[2\wh{\delta\sigma}_B(\xi)= \lim_{\theta\to 0^+} \sgn B(e^{i \theta}\xi)-\lim_{\theta\to 0^-}\sgn B(e^{i \theta}\xi).\]
By~\eqref{eq:dsignAdsignB}, it only remains to show that $\wh{\delta\sigma}_B(\xi)=\delta\sigma_{(M,\lambda)}(\xi)$.
Note that $A(t)$ and $B(t)$ represent the same pairing over $\mathcal{O}_\xi$.
That is, $A(t)$ represents $\bigoplus_{j=1}^n \wh{\mathfrak{e}}(n_i,0,\xi,\epsilon_i)$ over $\mathcal{O}_\xi$.
On the other hand, $A(t)$ represents the pairing $(M,\lambda)$ over $\L{\C}$.
By Proposition~\ref{prop:tensor}, we infer that the $(t-\xi)$--primary part of $(M,\lambda)$ is precisely $\bigoplus_{i=1}^n \mathfrak{e}(n_i,0,\xi,\epsilon_i,\mathbb{C})$.
Therefore, by definition of the signature jump, we obtain
$$\delta\sigma_{(M,\lambda)}(\xi)=\sum_{\substack{j=1,\ldots,n\\ n_i\textrm{odd}}} \epsilon_j.$$
As this is the definition of $\delta \widehat{\sigma}_B(\xi)$, Case 2 is concluded and so is the proof of the proposition.
\section{Twisted homology and twisted Blanchfield pairings}
\label{sec:TwistedHomology}
This section is organized as follows: in Subection~\ref{sub:TwistedHomology}, we review twisted (co)homology, in Subsection~\ref{sub:Blanchfield}, we discuss twisted Blanchfield forms, and in Subsection~\ref{sub:Ore}, we relate twisted Blanchfield forms to their non-commutative counterparts.
\subsection{Twisted homology and cohomology}
\label{sub:TwistedHomology}
In this subsection, we briefly review twisted (co)ho\-mo\-lo\-gy and twisted intersection forms.
Standard references include~\cite{KirkLivingston, FriedlKim}.
\medbreak
Let $X$ be a CW complex and let $Y \subset X$ be a possibly empty subcomplex.
Use $p \colon \widetilde{X} \to X$ to denote the universal cover of $X$ and set $\widetilde{Y}:=p^{-1}(Y)$.
The left action of $\pi_1(X)$ on $\widetilde{X}$ endows the chain complex $C_*(\widetilde{X},\widetilde{Y})$ with the structure of a left $\mathbb{Z}[\pi_1(X)]$-module.
Moreover, let $R$ be a ring and let $M$ be a $(R,\mathbb{Z}[\pi_1(X)])$-module.
The chain complexes
\begin{align*}
& C_*(X,Y;M):=M \otimes_{\mathbb{Z}[\pi_1(X)]}C_*(\widetilde{X},\widetilde{Y}) \\
& C^*(X,Y;M):=\operatorname{Hom}_{\text{right-}\mathbb{Z}[\pi_1(X)]}(\makeithash{C_*(\widetilde{X}, \widetilde{Y} )},M)
\end{align*}
of~left $R$-modules will be called the \textit{twisted (co)chain complex} of $(X,Y)$ with coefficients in $M$.
The corresponding homology left $R$-modules~$H_*(X,Y;M)$ and $H^*(X,Y;M)$ will be called the \textit{twisted (co)homology} modules of $(X,Y)$ with coefficients in $M$.
Assume that $R$ is endowed with an involution $x \mapsto x^\#$. Let $M,M'$ be $(R,\mathbb{Z}[\pi_1(X)])$-bimodules and let $S$ be a $(R,R)$-bimodule.
Furthermore, let $\langle -,-\rangle \colon M \times M' \to S$ be a non-singular
$\pi_1(X)$-invariant sesquilinear pairing, in the sense that $\langle m\gamma , n\gamma \rangle=\langle m,n\rangle$ and $\langle rm,sn \rangle=r \langle m,n \rangle \makeithash{s}$ for all $\gamma \in \pi_1(X)$, all $r,s \in R$ and all $m\in M, n \in M'$.
Non-singularity means that the induced map $M\to \Hom_{\text{left-}R}(M',S)^\#$ is an isomorphism.
In this setting, as explained in~\cite[Section 4.2]{MillerPowell} (see also~\cite{FriedlKim, KirkLivingston}), there is an \textit{evaluation map}
$$ \ev \colon H^i(X,Y;M) \to \makeithash{\operatorname{Hom}_{\text{left-}R}(H_i(X,Y;M'),S)}. $$
This evaluation map need not be an isomorphism.
If $R$ is a principal ideal domain, then the universal coefficient theorem implies that $H^i(X,Y;M)$ decomposes as the direct sum of $\makeithash{\operatorname{Hom}_{\text{left-}R}(H_i(X,Y;M'),S)}$ and $\makeithash{\operatorname{Ext}^1_{\text{left-}R}(H_i(X,Y;M'),S)}$.
In general, the evaluation map can be studied using the universal coefficient spectral sequence~\cite[Theorem 2.3]{LevineKnotModules}.
Returning to $\pi_1(X)$-invariant pairings, we now discuss two examples which we shall use throughout this section.
\begin{example}
\label{ex:UsualCoeff}
Let $R$ be an integral domain with field of fractions $Q$.
Given a positive integer~$d$, let $\beta \colon \pi_1(X) \to GL_d(R)$ be a representation and use~$R^d_\beta$ to denote the $(R,\mathbb{Z}[\pi_1(X)])$-module whose right $\mathbb{Z}[\pi_1(X)]$-module structure is given by right multiplication by~$\beta(\gamma)$ on row vectors for~$\gamma \in \pi_1(X)$.
Use $\widecheck{\beta} \colon \pi_1(X) \to GL_d(R)$
to denote the representation defined by $\widecheck{\beta}(\gamma)=\makeithashT{\beta(\gamma^{-1})}$ and consider the pairings
\begin{align*}
\left(Q/R \otimes_R R_\beta^d\right) \times R_{\widecheck{\beta}}^d &\to Q/R \ \ \ \ \ \ \ \ \ \ \ \ \ \
R_\beta^d \times R_{\widecheck{\beta}}^d \to R \\
(q \otimes v, w) & \mapsto v \makeithashT{w} \cdot q \ \ \ \ \ \ \ \ \ \ \ \ \ \ (v, w) \mapsto v \makeithashT{w}.
\end{align*}
Note that the use of the representation $\widecheck{\beta}$ guarantees that these pairings are indeed $\pi_1(X)$-invariant.
In particular, if $\beta$ is a unitary representation (so that $\widecheck{\beta}=\beta$), then the two pairings displayed above give rise to evaluation maps $H^i(X,Y;Q/R \otimes_R R_\beta^d) \to \makeithash{\operatorname{Hom}_{\text{left-}R}(H_i(X,Y;R_{\beta}^d),Q/R)}$ and $H^i(X,Y; R_\beta^d) \to \makeithash{\operatorname{Hom}_{\text{left-}R}(H_i(X,Y; R_{\beta}^d),R)}$.
\end{example}
Next, we briefly review the definition of twisted intersection forms.
Let $W$ be a compact oriented $n$-manifold and let $M$ be a $(R,\mathbb{Z}[\pi_1(W)])$-bimodule.
It is known at the very least since Wall~\cite{Wall} that there are Poincar\'e duality isomorphisms
$$ H_k(W,\partial W;M) \cong H^{n-k}(W;M) \ \ \ \ \ H_k(W;M) \cong H^{n-k}(W,\partial W;M),$$
both of which we denote by $\text{PD}$.
In order to describe twisted intersection pairings, fix an $(R,R)$-bimodule $S$ and a non-singular $\pi_1(W)$-invariant sesquilinear pairing $M \times M \to S$.
Compose the homomorphism induced by the inclusion $(W,\emptyset) \to (W,\partial W)$ with Poincar\'e duality and the evaluation homomorphism described above.
The result is the following homomorphism of left $R$-modules.
$$ \Phi \colon H_k(W;M) \to H_k(W,\partial W;M) \xrightarrow{\text{PD}} H^{n-k}(W;M) \xrightarrow{\ev}\makeithash{\operatorname{Hom}_{\text{left-}R}(H_{n-k}(W;M),S)}.$$
Restricting to the case where $W$ is $2n$-dimensional, the \emph{twisted intersection pairing}
$$\lambda_{M,W} \colon H_n(W;M) \times H_n(W;M) \to S $$
is defined by $\lambda_{M,W}(x,y)=\Phi(y)(x)$.
Note that while $\lambda_{M,W}$ is $(-1)^n$-Hermitian, it may be singular: the submodule $\operatorname{im} (H_{n}(\partial W;M) \to H_{n}(W;M)) $ is annihilated by~$\lambda_{M,W}$ and the evaluation map need not be an isomorphism.
\subsection{The twisted Blanchfield pairing}
\label{sub:Blanchfield}
In this subsection, we review twisted Blanchfield pairings.
While the construction can be performed over non-commutative rings~\cite{Powell, CochranOrrTeichner,CochranNonCommutative}, we focus first on the case $R=\mathbb{F}[t^{\pm 1}]$.
References on twisted Blanchfield pairings include~\cite{MillerPowell, Powell}.
\medbreak
Let $N$ be a closed $3$-manifold.
Just as in Example~\ref{ex:UsualCoeff}, let $\beta \colon \pi_1(N) \to GL_d(\L{\F})$ be a unitary representation and use $\mathbb{F}[t^{\pm 1}]^d_\beta$ to denote the $(\L{\F},\mathbb{Z}[\pi_1(X)])$-bimodule whose right~$\mathbb{Z}[\pi_1(N)]$-module structure is given by right multiplication by $\beta(\gamma)$ on row vectors. In what follows, we shall think of $\mathbb{F}[t^{\pm 1}], \mathbb{F}(t)$ and $\mathbb{F}(t)/\mathbb{F}[t^{\pm 1}]$ as $(\mathbb{F}[t^{\pm 1}],\mathbb{F}[t^{\pm 1}])$-bimodules.
Since $\mathbb{F}[t^{\pm 1}] \otimes_{\mathbb{F}[t^{\pm 1}]}~\mathbb{F}[t^{\pm 1}]^d_\beta$ and $\mathbb{F}(t)/\mathbb{F}[t^{\pm 1}] \otimes_{\mathbb{F}[t^{\pm 1}]}~\mathbb{F}[t^{\pm 1}]^d_\beta$ are $(\mathbb{F}[t^{\pm 1}],\mathbb{Z}[\pi_1(N)])$-bimodules, there is a short exact sequence
$$ 0 \to C^*(N; \mathbb{F}[t^{\pm 1}]^d_\beta) \to C^*(N,\mathbb{F}(t) \otimes_{\mathbb{F}[t^{\pm 1}]} \mathbb{F}[t^{\pm 1}]^d_\beta) \to C^*(N;\mathbb{F}(t)/\mathbb{F}[t^{\pm 1}] \otimes_{\mathbb{F}[t^{\pm 1}]} \mathbb{F}[t^{\pm 1}]^d_\beta) \to 0$$
of cochain complexes of left $\mathbb{F}[t^{\pm 1}]$-modules, where we identified $\mathbb{F}[t^{\pm 1}]^d_\beta$ with $\mathbb{F}[t^{\pm 1}] \otimes_{\L{\F}} \mathbb{F}[t^{\pm 1}]^d_\beta $.
Passing to cohomology, we obtain a long exact sequence in which the connecting map
$$ \operatorname{BS} \colon H^{k}(N;\mathbb{F}(t)/\mathbb{F}[t^{\pm 1}] \otimes_{\mathbb{F}[t^{\pm 1}]} \mathbb{F}[t^{\pm 1}]^d_\beta) \to H^{k+1}(N;\mathbb{F}[t^{\pm 1}]^d_\beta) $$
is called the \emph{Bockstein homomorphism}.
We say that $\beta$ is \emph{$H_1$-null} if $H_1(N;\mathbb{F}(t) \otimes_{\L{\F}} \L{\F}_\beta^d$ vanishes, i.e.~if the $\L{\F}$-module $H_1(N;\L{\F}_\beta^d)$ is torsion.
Observe that if $\beta$ is $H_1$-null, then the corresponding Bockstein homomorphism is an isomorphim: indeed Poincar\'e duality and the universal coefficient theorem respectively imply that $H^2(N;\mathbb{F}(t) \otimes_{\mathbb{F}[t^{\pm 1}]} \mathbb{F}[t^{\pm 1}]^d_\beta)$ and $H^1(N;\mathbb{F}(t) \otimes_{\mathbb{F}[t^{\pm 1}]}~\mathbb{F}[t^{\pm 1}]^d_\beta)$ vanish.
Consider the composition
\begin{align*}
\Theta \colon H_1(N;\mathbb{F}[t^{\pm 1}]^d_\beta) \,
& \xrightarrow{PD}\, H^2(N;\mathbb{F}[t^{\pm 1}]^d_\beta) \\
& \xrightarrow{BS^{-1}} H^1(N;\mathbb{F}(t)/\mathbb{F}[t^{\pm 1}] \otimes_{\mathbb{F}[t^{\pm 1}]} \mathbb{F}[t^{\pm 1}]^d_\beta) \\
& \xrightarrow{\text{ev}} \makeithash{\Hom_{\mathbb{F}[t^{\pm 1}]}(H_1(N;\mathbb{F}[t^{\pm 1}]^d_\beta),\mathbb{F}(t)/\mathbb{F}[t^{\pm 1}])}
\end{align*}
of the three following $\mathbb{F}[t^{\pm 1}]$-homomorphisms: Poincar\'e duality, the inverse Bocktein and the evaluation map described in Example~\ref{ex:UsualCoeff}.
The main definition of this section is the following.
\begin{definition}
\label{def:Blanchfield}
Let $N$ be a closed oriented $3$-manifold and let $\beta \colon \pi_1(N) \to GL_d(\mathbb{F}[t^{\pm 1}])$ be a unitary $H_1$-null representation.
The \emph{twisted Blanchfield pairing associated to $\beta$}
$$ \Bl_\beta(N) \colon H_1(N;\mathbb{F}[t^{\pm 1}]^d_\beta) \times H_1(N;\mathbb{F}[t^{\pm 1}]^d_\beta) \to \mathbb{F}(t)/\mathbb{F}[t^{\pm 1}]$$
is defined as $\Bl_\beta(N)(x,y)=\Theta(y)(x)$.
If $N=M_K$ is the $0$-framed surgery along a knot $K$, then we write $\Bl_\beta(K)$ instead of $\Bl_\beta(M_K)$.
\end{definition}
We refer to~\cite{Powell} for a more general treatment of twisted Blanchfield pairings, but make two remarks nonetheless.
Firstly, Powell proved that $\Bl_\beta(N)$ is Hermitian~\cite{Powell}.
Secondly, since $\mathbb{F}[t^{\pm 1}]$ is a PID, it follows that $\Bl_\beta(N)$ is non-singular: indeed $\mathbb{F}(t)/\mathbb{F}[t^{\pm 1}]$ is an injective $\mathbb{F}[t^{\pm 1}]$-module and thus the evaluation map is an isomorphism.
Finally, although we mostly focus on \emph{closed} 3-manifolds, we conclude this subsection with a remark on manifolds with boundary. To describe this case, we say that a representation $\beta \colon \pi_1(N) \to~GL_d(\L{\F})$ is \emph{acyclic} if $H_i(N;\mathbb{F}[t^{\pm 1}]_\beta^d)$ is a torsion $\mathbb{F}[t^{\pm 1}]$-module for each $i$.
\begin{remark}
\label{rem:Boundary}
If $N$ is a 3-manifold with boundary and $\beta \colon \pi_1(N) \to GL_d(\L{\F})$ is a unitary acyclic representation, then one also obtains a twisted Blanchfield pairing $\operatorname{Bl}_\beta(N)$.
Indeed, one starts from the map $i \colon H_1(N;\L{\F}_\beta^d) \to H_1(N,\partial N;\L{\F}_\beta^d)$ induced by the inclusion $(N,\emptyset) \to~(N,\partial N)$ and then proceeds as in the closed case, using duality, the inverse of the Bockstein homorphism and evalutation.
The acyclicity assumption is used to guarantee that the Bockstein homomorphism is an isomorphism: $H_1$-nullity is not enough when the boundary is non-empty, although an alternative is to work on the torsion submodule of $H_1(N;\L{\F}_\beta^d)$ and to proceed as in~\cite{Powell, Hillman} or~\cite[Subsection 2.2]{ConwayBlanchfield}.
Finally, note that when $\partial N$ is non-empty, $\operatorname{Bl}_\beta(N)$ may be singular: the map~$i$ need not be an isomorphism.
\end{remark}
\subsection{Relation to Blanchfield pairings over noncommutative rings}
\label{sub:Ore}
Twisted Blanchfield pairing are also defined in the non-commutative setting, using Ore rings~\cite{CochranOrrTeichner, FriedlLeidyNagelPowell,Powell}.
In this subsection, after briefly reviewing some basics on these rings, we relate the resulting non-commutative Blanchfield pairings to the twisted Blanchfield pairings of Subsection~\ref{sub:Blanchfield}.
\medbreak
We start with some brief recollections on Ore rings, referring to~\cite[Chapter 4, Section 10]{Lam-modules} for further details.
Let $\mathcal{R}$ be a ring.
A multiplicative subset $S \subset \mathcal{R}$ is \emph{right permutable} if, for any~$r \in \mathcal{R}$ and $s \in S$, we have $r S \cap s\mathcal{R} \neq \emptyset$.
An element $r \in \mathcal{R}$ is \emph{regular} if it is neither a left zero divisor nor a right zero divisor.
We say that $\mathcal{R}$ is a \emph{right Ore ring} if the set $S$ of regular elements is right permutable.
In this case, $\mathcal{R}$ can be localized at $S$ and $\mathcal{R}S^{-1}$ is called the \emph{right ring of quotients of $\mathcal{R}$}.
The left analogues of these notions are defined similarly.
If $\mathcal{R}$ is both a left Ore ring and a right Ore ring, then $\mathcal{R}$ is called an \emph{Ore ring}.
In this case, the right and left rings of quotients agree, and we simply write $\mathcal{Q}$.
Commutative rings are Ore rings~\cite[Chapter 4, Section 10, 10.18]{Lam-modules}.
In what follows however, we shall mostly be concerned with rings of matrices.
\begin{example}
\label{ex:MatrixOre}
If $R$ is a commutative ring, then the matrix ring \(\mathcal{M}_{d}(R)\) is an Ore ring~\cite[Example 11.21(1)]{Lam-modules}.
For concreteness, we focus on the case where $R=\L{\F}$.
In this case, by~\cite[Example~11.21(1)]{Lam-modules} and~\cite[Proposition~10.21]{Lam-modules} the ring of quotients of \(\mathcal{M}_{d}(\mathbb{F}[t^{\pm1}])\) is isomorphic to \(\mathcal{M}_{d}(\mathbb{F}(t))\).
\end{example}
Next, suppose that \(R\) is a commutative ring with unit.
Right multiplication on row vectors endows $R^{d}$ with a right \(\mathcal{M}_{d}(R)\)-module structure and we write \(R^{d}_{\mathcal{M}}\) for emphasis.
An alternate description of \(R^{d}_{\mathcal{M}}\) is obtained by considering the elementary matrix \(E_{i,j}\) whose only nonzero coefficient is a~$1$ in its $(i,j)$-entry.
Indeed, one has the isomorphim \( E_{i,i}\mathcal{M}_d(R) \cong R^{d}_{\mathcal{M}}\) of right $\mathcal{M}_d(R)$-modules.
The next lemma shows that $\mathcal{M}_d(R)$ decomposes as a direct sum of $d$ copies of~$R^{d}_{\mathcal{M}}$.
\begin{lemma}\label{obs:proj-modules}
If \(R\) is a commutative ring with unit, then \(R^{d}_{\mathcal{M}}\) is a projective right \(\mathcal{M}_{d}(R)\)-module.
In fact, we have the following isomorphism of right \(\mathcal{M}_{d}(R)\)-modules:
\begin{equation}
\label{eq:DecompositionOre}
\mathcal{M}_{d}(R) \cong \bigoplus_{i=1}^d E_{i,i}\mathcal{M}_{d}(R) \cong \bigoplus_{i=1}^d R^{d}_{\mathcal{M}}.
\end{equation}
\end{lemma}
\begin{proof}
First, notice that the \(E_{i,i}\) are orthogonal idempotents: one has $E_{i,i} E_{j,j}=\delta_{ij}E_{i,j}$.
The conclusion of the lemma now follows since $\operatorname{id} = E_{1,1} + \cdots +E_{d,d}$.
\end{proof}
Next, assume that $X$ is a CW complex together with a homomorphism $\gamma \colon \mathbb{Z}[\pi_1(X)] \to \mathcal{R}$ to an Ore ring $\mathcal{R}$.
Just as in Subsection~\ref{sub:TwistedHomology}, this endows $\mathcal{R}$ with a right $\mathbb{Z}[\pi_1(X)]$-module structure which we denote by $\mathcal{R}_\gamma$ for emphasis.
Given a commutative ring $R$ with unit, note that any representation \(\gamma \colon \pi_{1}(X) \to GL_{d}(R)\) canonically extends to a ring homomorphism \(\gamma \colon \mathbb{Z}[\pi_{1}(X)] \to \mathcal{M}_{d}(R)\).
In particular $\gamma$ endows $R^d$ and $\mathcal{R}=\mathcal{M}_d(R)$ with right $\mathbb{Z}[\pi_{1}(X)]$-module structures.
The following proposition is a topological application of Lemma~\ref{obs:proj-modules}.
\begin{proposition}
\label{prop:OreTwisted}
Let \(R\) be a commutative ring with unit.
If $\gamma \colon \pi_1(X) \to GL_d(R)$ is a representation, then we have the following chain isomorphisms of chain complexes of left $R$-modules:
\begin{align*}
& C_*(X;\mathcal{M}_d(R)_\gamma) \cong \bigoplus_{i=1}^d C_*(X;R_\gamma^d), \\
&C^*(X;\mathcal{M}_d(R)_\gamma) \cong \bigoplus_{i=1}^d C^*(X;R_\gamma^d).
\end{align*}
Furthermore, one also has the following isomorphisms of left $R$-modules:
\begin{align*}
& H_*(X;R_\gamma^d) \cong R^d_{\mathcal{M}} \otimes_{\mathcal{M}_d(R)} H_*(X;\mathcal{M}_d(R)_\gamma) , \\
& H^*(X;R_\gamma^d) \cong R^d_{\mathcal{M}} \otimes_{\mathcal{M}_d(R)} H^*(X;\mathcal{M}_d(R)_\gamma).
\end{align*}
\end{proposition}
\begin{proof}
The chain isomorphisms follow from the decomposition of $\mathcal{M}_d(R)$ displayed in~\eqref{eq:DecompositionOre}.
To deal with the second assertion, we once again use Lemma~\ref{obs:proj-modules} to deduce that $\mathcal{M}_d(R)$ is a projective (and thus flat) $R_\mathcal{M}^d$-module.
We therefore obtain the following isomorphisms:
$$ H_*(X;R_\gamma^d) \cong H_*(X; R_{\mathcal{M}}^d \otimes_{\mathcal{M}_d(R)} \mathcal{M}_d(R)_\gamma) \cong R_{\mathcal{M}}^d \otimes_{\mathcal{M}_d(R)} H_*(X; \mathcal{M}_d(R)_\gamma).$$
The proof for cohomology is analogous.
This concludes the proof of the lemma.
\end{proof}
Next, we deal with non-commutative Blanchfield pairings.
Let $N$ be a closed oriented $3$-manifold.
Let $\mathcal{R}$ be an Ore ring with involution and use~$\mathcal{Q}$ to denote its ring of quotients.
Given a homomorphism $\gamma \colon \mathbb{Z}[\pi_1(N)] \to \mathcal{R}$, we can view $\mathcal{R},\mathcal{Q}$ and~$\mathcal{Q}/\mathcal{R}$ as right $\mathbb{Z}[\pi_1(N)]$-modules.
In particular, as in Subsections~\ref{sub:TwistedHomology} and~\ref{sub:Blanchfield}, there are Poincar\'e duality isomorphisms and a Bockstein homomorphism on cohomology. Thus, if $H_1(N;\mathcal{Q}_\gamma)=0$, then one can consider the composition
$$ \Phi \colon H_1(N;\mathcal{R}_\gamma) \stackrel{\operatorname{PD}}{\to} H^2(N;\mathcal{R}_\gamma) \stackrel{BS^{-1}}{\to} H^1(N;\mathcal{Q}/\mathcal{R}_\gamma) \stackrel{\operatorname{ev} \circ \kappa}{\to} \operatorname{Hom}_\mathcal{R}(H_1(N;\mathcal{R}_\gamma);\mathcal{Q}/\mathcal{R})^\#.$$
We refer to~\cite[Section 2]{FriedlLeidyNagelPowell} for further details, but the upshot is that there is a Blanchfield pairing in the non-commutative setting which is defined exactly as in Subsection~\ref{sub:Blanchfield}.
\begin{definition}
\label{def:NonCommutativeBlanchfield}
Let $N$ be a closed oriented $3$-manifold together with a map $\gamma \colon \mathbb{Z}[\pi_1(N)] \to \mathcal{R}$ to an Ore ring with involution.
Assuming that $H_1(N;\mathcal{Q}_\gamma)=0$, there is a \emph{non-commutative Blanchfield pairing}
$$ \Bl_{\mathcal{R},\gamma}(N) \colon H_1(N;\mathcal{R}_\gamma) \times H_1(N;\mathcal{R}_\gamma) \to \mathcal{Q}/\mathcal{R}$$
which is defined by setting $\Bl_{\mathcal{R},\gamma}(N)(x,y)=\Phi(y)(x)$.
\end{definition}
Note that in the litterature $\mathcal{R}$, is usually assumed to be an Ore \emph{domain}, however the definition extends \emph{verbatim} to the case of Ore rings.
The next proposition relates the non-commutative Blanchfield pairing to the twisted Blanchfield pairing.
\begin{proposition}
\label{prop:BlanchfieldOreTwisted}
Let \(N\) be a closed oriented \(3\)-manifold, and let \(\gamma \colon \pi_{1}(N) \to GL_{d}(\L{\F})\) be a representation.
The following assertions hold:
\begin{enumerate}
\item We have $H_*(N;\mathbb{F}(t) \otimes_{\L{\F}} \L{\F}^d_\gamma)=0$ if and only if \(H_*(N;\mathcal{M}_{d}(\mathbb{F}(t))_\gamma)=0\).
\item The following $\L{\F}$-linking forms are canonically isometric:
\[\mathbb{F}[t^{\pm1}]^{d}_{\mathcal{M}} \otimes_{\mathcal{M}_{d}(\mathbb{F}[t^{\pm1}])} \Bl_{\mathcal{M}_{d}(\mathbb{F}[t^{\pm1}]),\gamma}(N) \cong \Bl_{\gamma}(N).\]
\end{enumerate}
\end{proposition}
\begin{proof}
The first assertion follows immediately from the first assertion of Proposition~\ref{prop:OreTwisted}.
Next, we deal with the second assertion.
Using once again the first assertion of Proposition~\ref{prop:OreTwisted}, the linking form $\Bl_{\mathcal{M}_d(\L{\F}),\gamma}(N)$ can canonically be thought of as a linking form
$$\Bl_{\mathcal{M}_d(\L{\F}),\gamma}(N) \colon \bigoplus_{i=1}^d H_1(N;\L{\F}_\gamma^d) \times \bigoplus_{i=1}^d H_1(N;\L{\F}_\gamma^d) \to \mathcal{M}_d(\mathbb{F}(t))/\mathcal{M}_d(\L{\F}),$$
where the above decomposition is obtained by using the matrices \(E_{i,i}\), for \(1 \leq i \leq d\).
The result now follows by tensoring with $\L{\F}_\mathcal{M}^d$ and using the second assertion of Proposition~\ref{prop:OreTwisted}.
Indeed both Blanchfield pairings are defined by composing Poincar\'e duality isomorphisms, inverse Bockstein isomorphisms and evaluation maps.
The first two maps preserve the direct sum decomposition displayed above. After applying the adjoint of evaluation map, and using the identification from Proposition~\ref{prop:OreTwisted}, $\Bl_{\mathcal{M}_d(\L{\F})}$ is given by
\[(E_{i,i} \cdot x, E_{j,j} \cdot y) \mapsto\Bl_{\gamma}(x,y) E_{i,j} \in \mathcal{M}_{d}(\mathbb{F}(t) / \L{\F},\]
for \(x,y \in H_{1}(N;\L{\F}^{d}_{\gamma})\).
This concludes the proof of the proposition.
\end{proof}
Informally, Proposition~\ref{prop:BlanchfieldOreTwisted} implies that several statements that hold for non-commutative Blanchfield pairings also hold for twisted Blanchfield pairings.
This will be used in Theorem~\ref{thm:CablingTheorem}.
\section{Twisted signatures of 3-manifolds and knots}
\label{sec:TwistedSign3ManifKnot}
This section is organized as follows.
In Subsection~\ref{sub:TwistedSignatures}, we introduce our twisted signature function and describe its basic properties, in Subsection~\ref{sub:BordismSignature}, we provide conditions for its bordism invariance and in Subsection~\ref{sub:Satellite}, we investigate its behavior under satellite operations.
\subsection{Definition of the twisted signatures }
\label{sub:TwistedSignatures}
In this short subsection, we define our twisted signature invariants for knots and 3-manifolds.
In a nutshell, we apply Definitions~\ref{def:sig_func} and~\ref{def:average_signature} to the twisted Blanchfield pairings which were discussed in Section~\ref{sub:Blanchfield}.
\begin{definition}
\label{def:TwistedSignature}
Let $N$ be a closed connected oriented $3$-manifold and let $\beta \colon \pi_1(N) \to GL_d(\mathbb{F}[t^{\pm 1}])$ be a unitary $H_1$-null representation.
The \emph{twisted signature function} $\sigma_{N,\beta} \colon S^1 \to \mathbb{Z}$ of the pair~$(N,\beta)$ is the signature function of the non-singular linking form $\Bl_\beta(N)$.
The \emph{averaged signature function}~$\sigma_{N,\beta}^{\operatorname{av}}$ is the averaged signature function of $\Bl_\beta(N)$.
\end{definition}
The corresponding invariants for a knot $K$ are obtained by applying Definition~\ref{def:TwistedSignature} to the 0-framed surgery $M_K$.
\begin{definition}
\label{def:TwistedKnotSignature}
Let $K$ be an oriented knot and let $\beta \colon \pi_1(M_K) \to GL_d(\mathbb{F}[t^{\pm 1}])$ be a unitary~$H_1$-null representation.
The \emph{twisted signature function} $\sigma_{K,\beta}\colon S^1\to\mathbb{Z}$ of $K$ is the twisted signature function of~$M_K$ and similarly for the \emph{averaged signature function} $ \sigma_{K,\beta}^{\operatorname{av}}$.
\end{definition}
Before proving properties, we show how Definition~\ref{def:TwistedKnotSignature} recovers the Levine-Tristram signature.
\begin{remark}
\label{rem:LevineTristram}
If $\beta$ is the abelianization map $\phi_K$, then $\Bl_\beta(K)$ coincides with the untwisted Blanchfield pairing $\operatorname{Bl}(K)$ and the corresponding signature function recovers the Levine-Tristram signature $\sigma_K$ thanks to~\cite[Lemma 3.2]{BorodzikFriedl}. Note that $\operatorname{Bl}(K)$ is usually defined on the Alexander module~$H_1(X_K;\L{\F})$, where $X_K=S^3 \setminus \mathcal{N}(K)$ is the complement of a tubular neighborhood~$\mathcal{N}(K)$ of~$K$ in $S^3$.
It is however well known that $H_1(X_K;\L{\F})$ and $H_1(M_K;\L{\F})$ are isomorphic and that the corresponding Blanchfield pairings are isometric.
\end{remark}
Contrarily to $\sigma_K$, the twisted signature function~$\sigma_{K,\beta}$ need not be symmetric (i.e.~$\sigma_{K,\beta}(\overline{\omega})$ need not equal~$\sigma_{K,\beta}(\omega)$). Despite this difference, the next proposition shows that the twisted signature shares several properties with its classical counterpart.
\begin{proposition}
\label{prop:LocallyConstant}
Let $K$ be an oriented knot and let $\beta \colon \pi_1(M_K) \to GL_d(\mathbb{F}[t^{\pm 1}])$ be a unitary $H_1$-null representation.
\begin{enumerate}
\item The twisted signature $\sigma_{K,\beta}$ is constant on the complement in~$S^1$ of the zero set of the twisted Alexander polynomial $\Delta_{M_K}^\beta(t):=\operatorname{Ord}(H_1(M_K;\mathbb{F}[t^{\pm 1}]_\beta^d))$.
\item Let $\overline{K}$ denote the mirror image of $K$. The representation $\beta$ canonically defines a representation on $\pi_1(M_{\overline{K}})$ (which we also denote by $\beta$) and, for all $\omega \in S^1$, we have
\begin{align*}
&\Bl_\beta(\overline{K})=-\Bl_\beta(K), \\
&\sigma_{\overline{K},\beta}=-\sigma_{K,\beta}.
\end{align*}
\item Let $K^r$ denote $K$ with its orientation reversed. Assume that $\beta(g)=\phi_K(g)\rho(g)$, where $\rho \colon \pi_1(M_K) \to GL_n(\mathbb{F})$ is a unitary representation. The representation $\beta$ canonically defines a representation on $\pi_1(M_{K^r})$ (which we also denote by $\beta$) and, for all $\omega \in S^1$, we~have
\begin{align*}
& \Bl_{\beta}(K^r)(t) =\Bl_{\beta}(K^r)(t^{-1}) , \\
& \sigma_{K^r,\beta}(\omega) =\sigma_{K,\beta}(\overline{\omega}).
\end{align*}
\end{enumerate}
\end{proposition}
\begin{proof}
The proof of the first property follows by applying Proposition~\ref{prop:BasicProperties} to the twisted Blanchfield pairing. To prove the second assertion, let $h \colon S^3 \to S^3$ be an orientation reversing homeomorphism such that $h(K)=\overline{K}$. Note that $h$ induces an orientation reversion homeomorphism $h \colon M_K \to M_{\overline{K}}$. A diagram chase involving the definition of the Blanchfield pairing now shows that $h$ induces an isometry $\Bl_\beta(\overline{K})\cong -\Bl_\beta(K)$, the key point being that the square involving Poincar\'e duality anti-commutes thanks to the identity $h_*(h^*(\xi) \cap [M_K])=\xi \cap h_*([M_K])=-\xi \cap [M_{\overline{K}}]$ for any~$\xi \in H^2(M_K;\L{\F}^n_\beta)$. The result on the signature now follows from the definition. The third assertion is a consequence of the identity $\phi_K(\mu_{K^r})=\phi_K(\mu_K)^{-1}$.
\end{proof}
The order of $H_1(M_K;\mathbb{F}[t^{\pm 1}]_\beta^d)$ can be related to the twisted Alexander polynomial of $K$ which is usually defined on the exterior $X_K$ of $K$; see~\cite[Lemma 6.3]{KirkLivingston} and~\cite[Lemma 3]{FriedlVidussiSurvey}. In the abelian case, since $H_1(M_K;\L{\F}) \cong H_1(X_K;\L{\F})$, the first item of Proposition~\ref{prop:LocallyConstant} (and Remark~\ref{rem:LevineTristram}) recovers the well known fact that the Levine-Tristram signature function $\sigma_K$ is constant on the complement of the zero set of the Alexander polynomial $\Delta_K(t)$. The second item of Proposition~\ref{prop:LocallyConstant} is also a direct generalization of the corresponding property for the Levine-Tristram signature.
\begin{remark}
\label{rem:NotSymmetric}
Since the Levine-Tristram is symmetric, it satisfies $\sigma_{K^r}(\omega)=\sigma_{K}(\omega)$. This does not appear to be the case in general since the signature function of a general complex linking form need not be symmetric.
\end{remark}
Motivated by Remark~\ref{rem:LevineTristram} and Proposition~\ref{prop:LocallyConstant}, the next subsections investigate to what extent the properties of the Levine-Tristram signature generalize to our twisted signatures.
\subsection{Bordism and concordance invariance}
\label{sub:BordismSignature}
In this subsection, we give conditions under which the signatures of 3-manifolds (resp. knots) are bordism (resp. concordance) invariants.
\medbreak
The proof of the next lemma can be found in~\cite[Proposition 2.8]{Letsche} or~\cite[Theorem 4.4]{CochranOrrTeichner}.
\begin{lemma}\label{lem:metabolizer}
Let \(N_{1},\ldots,N_{k}\) be closed connected oriented \(3\)-manifolds.
Let \(\beta_{i} \colon \pi_{1}(N_{i}) \to GL_{d}(\L{\F})\) be a unitary acyclic representation for $i=1,\ldots, k$.
Assume that \(W\) is a \(4\)-manifold such that \(\partial W = N_{1} \sqcup \ldots \sqcup N_{k}\) and the two following conditions are satisfied:
\begin{itemize}
\item[(a)] for $i=1,\ldots, k, $ there exists a unitary representation \(\gamma \colon \pi_{1}(W) \to GL_{d}(\L{\F})\) extending the representations $\beta_{i}$;
\item[(b)] the following sequence is exact:
\begin{equation}
\label{eq:necessary-cond-metabolic-form}
TH_{2}(W, \partial W; \mathbb{F}[t^{\pm 1}]_{\gamma}^{d}) \xrightarrow{\partial} H_{1}(\partial W;\mathbb{F}[t^{\pm 1}]_{\gamma}^{d}) \xrightarrow{\iota_{\ast}} H_{1}(W; \mathbb{F}[t^{\pm 1}]_{\gamma}^{d}).
\end{equation}
\end{itemize}
Then the linking form \(\oplus_{i=1}^{k} \Bl_{\beta_{i}}(N_{i})\) is metabolic.
\end{lemma}
The following result uses Lemma~\ref{lem:metabolizer} to provide (slightly less restrictive) assumptions under which the twisted Blanchfield pairing is metabolic.
\begin{proposition}
\label{prop:MetabolicAcyclic}
Let \(N = N_{1} \sqcup \ldots \sqcup N_{k}\) be a disjoint union of closed \(3\)-manifolds and let \(\beta_{i} \colon \pi_{1}(N_{i}) \to~GL_{d}(\L{\F})\) be a unitary acyclic representation for $i=1,\ldots, k$.
If $N$ bounds a compact \(4\)-manifold~\(W\) and if the \(\beta_{i}\) extend to a unitary acyclic representation \(\gamma \colon \pi_{1}(W) \to~GL_{d}(\mathbb{F}[t^{\pm 1}])\), then \(\oplus_{i=1}^{k}\Bl_{\beta_{i}}(N_{i})\) is metabolic and, in particular, the sum $\sum_{i=1}^{k}\sigma^{\text{av}}_{N,\beta_{i}}$ of averaged signature functions is zero.
\end{proposition}
\begin{proof}
Since $\gamma$ is acyclic, $H_*(W;\mathbb{F}(t)_\gamma^d)$ vanishes.
Poincar\'e duality and the universal coefficient theorem now imply that $H_*(W,N;\mathbb{F}(t)_\gamma^d)$ also vanishes.
Since the $\beta_i$ are acyclic, the homology~$\mathbb{F}[t^{\pm 1}]$-modules of $N,W$ and $(W,N)$ are $\mathbb{F}[t^{\pm 1}]$-torsion and therefore the sequence~(\ref{eq:necessary-cond-metabolic-form}) is exact.
The proposition now follows by applying Lemma~\ref{lem:metabolizer} (to deduce that $\oplus_{i=1}^{k}\Bl_{\beta_{i}}(N_{i})$ is metabolic) and Corollary~\ref{cor:metabolic_signature} (to obtain that the sum of the signature functions is zero).
\end{proof}
Next, suppose that $K$ and $K'$ are two concordant knots in $S^3$ and let $M_K$ and $M_{K'}$ denote their respective $0$-framed surgeries.
Given a concordance \(\Sigma \subset S^{3} \times I\) from \(K\) to \(K'\), we can construct a cobordism \(W_{\Sigma}\) from \(M_{K}\) to \(M_{K'}\).
Indeed, using \(\mathcal{N}(\Sigma)\) to denote a tubular neighborhood of \(\Sigma\) (which can be identified with \(\mathcal{N}(\Sigma) \cong S^{1} \times D^{2} \times I\)), the desired cobordism is obtained by setting
\[W_{\Sigma} = ((S^{3} \times I) \setminus \mathcal{N}(\Sigma)) \cup_{h} S^{1} \times D^{2} \times I,\]
where \(h \colon S^{1} \times S^{1} \times I \to S^{1} \times S^{1} \times I\) is the homeomorphism defined by \(h(x,y,z) = (y,x,z)\).
The next result shows that the averaged twisted signature is a concordance invariant as long as the corresponding representations extend through the complement of the concordance to an acyclic representation.
\begin{proposition}\label{prop:witt-equiv-concordance}
Let \(K_{1}, K_{2}\) be two concordant knots in $S^3$, and let \(\beta_{i} \colon \pi_{1}(M_{K_{i}}) \to GL_{d}(\mathbb{F}[t^{\pm 1}])\) be a unitary acyclic representation for \(i=1,2\).
Assume there is a concordance \(\Sigma \subset S^{3} \times I\) from~\(K_{1}\) to~\(K_{2}\) and a unitary representation \(\gamma \colon \pi_{1}(W_{\Sigma}) \to GL_{d}(\mathbb{F}[t^{\pm 1}])\) such that
\begin{itemize}
\item[(a)] the representations \(\beta_{i}\) factor as \(\gamma \circ (\iota_{i})_{\ast}\), where \(\iota_{i} \colon M_{K_{i}} \hookrightarrow W_{\Sigma}\) is the inclusion for \(i=1,2\);
\item[(b)] the representation $\gamma$ is acyclic.
\end{itemize}
Then the twisted Blanchfield pairings \(\Bl_{\beta_{1}}(K_1)\) and \(\Bl_{\beta_{2}}(K_2)\) are Witt equivalent and, in particular, the averaged signature functions \(\sigma^{av}_{K_{1},\beta_{1}}\) and $\sigma^{av}_{K_{2},\beta_{2}}$ agree.
\end{proposition}
\begin{proof}
Set \(N = M_{K_{1}} \sqcup -M_{K_{2}}\).
By our assumptions, \(W_{\Sigma}\) is a null-bordism for \(N\) and satisfies the conditions of Proposition~\ref{prop:MetabolicAcyclic}.
Therefore, \(\Bl_{\beta_{1}}(M_{K_{1}}) \oplus \Bl_{\beta_{2}}(-M_{K_{2}})\) is metabolic.
Since \(\Bl_{\beta_{2}}(-M_{K_{2}}) = -\Bl_{\beta_{2}}(M_{K_{2}})\) (use the same reasoning as in the proof of Proposition~\ref{prop:LocallyConstant}), it follows that \(\Bl_{\beta_{1}}(K_1)\) and \(\Bl_{\beta_{2}}(K_2)\) are Witt equivalent and, consequently the averaged signature functions agree: \(\sigma^{\operatorname{av}}_{K_{1},\beta_{1}}=\sigma^{\operatorname{av}}_{K_{2},\beta_{2}}\).
\end{proof}
The same proof yields a result for slice knots with a representation which extends to an acylic representation over the slice disk exterior. We record this result for completeness.
\begin{proposition}\label{prop:SignatureSlice}
Let \(K\) be a slice knot and let \(\beta \colon \pi_{1}(M_K) \to GL_{d}(\mathbb{F}[t^{\pm 1}])\) be a unitary acyclic representation. If $D\subset D^4 $ is a slice disk for $K$ and there is a unitary acyclic representation \(\gamma \colon \pi_{1}(D^4 \setminus \mathcal{N}(D)) \to GL_{d}(\mathbb{F}[t^{\pm 1}])\) extending $\beta$, then the twisted Blanchfield pairings \(\Bl_{\beta}(K)\) is metabolic and, in particular, the averaged signature function \(\sigma^{av}_{K,\beta}\) is zero.
\end{proposition}
\subsection{Satellite formulas for twisted signatures}
\label{sub:Satellite}
In this subsection, we use a result of Friedl-Leidy-Nagel-Powell~\cite{FriedlLeidyNagelPowell} to understand the behavior of the twisted Blanchfield pairings and signatures under various satellite operations. In particular, these results provide twisted generalizations of well known theorems for the classical Blanchfield pairing and Levine-Tristram signature.
\medbreak
Let \(Y\) be a \(3\)-manifold with empty or toroidal boundary.
Suppose that we are given a simple closed curve \(\eta\) in \(Y\).
If \(K \subset S^{3}\) is a knot, then an \emph{infection} of~\(Y\) by \(K\) is a \(3\)-manifold \(Y_{K}\) obtained by gluing \(Y_{K} = (Y \setminus \mathcal{N}(\eta)) \cup_{\partial} S^{3} \setminus \mathcal{N}(K)\), where we glue a zero-framed longitude of \(K\) to the meridian of \(\eta\) and some longitude of \(\eta\) to the meridian of \(K\).
Notice that the ambiguity in the choice of the longitude of \(\eta\) may lead to different outcomes, however, as pointed out in~\cite{FriedlLeidyNagelPowell}, the resulting twisted Blanchfield pairing does not depend on this choice.
Note that if $Y$ has a boundary, then so does $Y_K$. In this case, we use Remark~\ref{rem:Boundary} which describes twisted Blanchfield pairings for $3$-manifolds with boundary.
Let \(P \subset S^{3}\) be a knot and let \(M_{P}\) be the zero-surgery on \(P\).
If \(K \subset S^{3}\) is another knot and \(\eta\) is a simple closed curve in the complement of \(P\), we can form the \emph{satellite knot} \(P(K,\eta)\) with \emph{pattern}~\(P\), \emph{companion} \(C\) and \emph{infection curve} \(\eta\) by looking at the image of \(P\) under the diffeomorphism \((S^{3}\setminus \mathcal{N}(\eta)) \cup_{\partial} (S^{3} \setminus \mathcal{N}(K)) \cong S^{3}\), where the gluing of the exteriors of \(\eta\) and \(K\) identifies the meridian of \(\eta\) with the zero-framed longitude of \(K\) and vice-versa.
From this description, it is easy to see that the zero-surgery on the satellite knot \(P(K,\eta)\) can be obtained by infection of~\(M_{P}\) by \(K\) along \(\eta \subset S^{3} \setminus P \subset M_{P}\):
\begin{equation}
\label{eq:DecompoSurgerySatellite}
M_{P(K,\eta)}=M_P \setminus \mathcal{N}(\eta) \cup_\partial S^3 \setminus \mathcal{N}(K).
\end{equation}
Let \(\mu_{\eta}\) denote a meridian of \(\eta\).
Our goal in this section is to express the Blanchfield form~\(\Bl_\gamma(M_{P(K,\eta)})\) in terms of the twisted Blanchfield forms on \(M_{P}\) and \(M_{K}\).
Using~\eqref{eq:DecompoSurgerySatellite}, notice that given a representation \(\gamma \colon \pi_{1}(M_{P(K,\eta)}) \to GL_{d}(\mathbb{F}[t^{\pm1}])\), there are induced representations
\[\gamma_{P} \colon \pi_{1}(M_{P} \setminus \mathcal{N}(\eta)) \to GL_{d}(\mathbb{F}[t^{\pm1}]), \quad \gamma_{K} \colon \pi_{1}(S^{3} \setminus \mathcal{N}(K)) \to GL_{d}(\mathbb{F}[t^{\pm1}]).\]
In general however, there is no guarantee that these representations can be extended to representations of the fundamental groups of \(M_{P}\) and \(M_{K}\).
\begin{remark}
\label{rem:EtaRegular}
In fact, \(\gamma_{P}\) and \(\gamma_{K}\) extend to representations of the respective zero-surgeries if and only if \(\gamma(\mu_{\eta}) = 1\).
This follows from the following two observations.
Firstly, \(\mu_{\eta}\) is identified with the zero-framed longitude of \(K\).
Secondly, by van Kampen's theorem, the inclusion \(M_{P} \setminus \mathcal{N}(\eta) \hookrightarrow M_{P}\) induces a surjection $\pi_{1}(M_{P} \setminus \mathcal{N}(\eta)) \to \pi_{1}(M_{P})$ with kernel normally generated by~\(\mu_{\eta}\).
\end{remark}
We say that a representation \(\gamma \colon \pi_{1}(M_{P(K,\eta)}) \to GL_{d}(\mathbb{F}[t^{\pm1}])\) is \(\eta\)-\emph{regular} if \(\gamma(\mu_{\eta}) = \operatorname{id}\) and \(\det(\operatorname{id} - \gamma(\eta)) \neq 0\).
As explained in Remark~\ref{rem:EtaRegular}, if $\gamma$ is $\eta$-regular, then it induces well defined representations $\gamma_P$ and $\gamma_K$ on $\pi_1(M_P)$ and $\pi_1(M_K)$.
The following theorem describes the twisted Blanchfield pairing of satellite knots.
\begin{theorem}
\label{thm:CablingTheorem}
Let $K,P$ be knots and let $\eta$ be a simple closed curve in $S^3 \setminus P \subset M_P$.
Let \(\gamma \colon \pi_{1}(M_{P(K,\eta)}) \to GL_{d}(\mathbb{F}[t^{\pm1}])\) be a unitary acyclic \(\eta\)-regular representation.
Let \(T \subset M_{P(K,\eta)}\) denote the common boundary of \(M_{P}\setminus \mathcal{N}(\eta)\) and \(S^{3} \setminus \mathcal{N}(K)\).
Define
\[L = \operatorname{im}(H_{1}(T;\L{\F}_{\gamma}^{d}) \to H_{1}(M_{P(K,\eta)};\L{\F}^{d}_{\gamma})).\]
The following assertions hold:
\begin{enumerate}
\item The representations $\gamma_P$ and $\gamma_K$ are unitary and acyclic.
\item The module \(L\) is isotropic and \(\ord L\) divides \(\det(1-\gamma(\eta))\).
\item The sublagrangian reduction of \(\Bl_\gamma(P(K,\eta))\) with respect to \(L\) is isometric to the direct sum \(\Bl_{\gamma_{P}}(P) \oplus \Bl_{\gamma_K}(K)\). In particular \(\Bl_\gamma(P(K,\eta))\) is Witt equivalent to $\Bl_{\gamma_{P}}(P) \oplus \Bl_{\gamma_K}(K)$.
\end{enumerate}
\end{theorem}
\begin{proof}
During this proof, in order to simplify the notation, we will omit coefficients.
As we mentioned in the statement, the torus \(T\) can be described either as \(\partial (M_{P} \setminus \mathcal{N}(\eta))$ or as $ \partial (S^{3} \setminus \mathcal{N}(K))\). As a consequence, we have~$M_{P(K,\eta)} = (M_{P} \setminus \mathcal{N}(\eta))\cup_{ T}(S^{3} \setminus \mathcal{N}(K))$.
Applying~\cite[Theorem~1.1]{FriedlLeidyNagelPowell}, we know that if $H_*(T)$ and~$H_*(M_{P(K,\eta)})$ are $\L{\F}$-torsion modules, then $H_*(M_{P} \setminus \mathcal{N}(\eta))$ and $H_*(S^{3} \setminus \mathcal{N}(K))$ are also $\L{\F}$-torsion modules and the direct sum of the canonical inclusions induces the following morphism linking forms:
\begin{equation}\label{eq:morphism-inking forms}
\psi \colon \Bl_{\gamma_P}(M_{P} \setminus \mathcal{N}(\eta)) \oplus \Bl_{\gamma_K}(S^{3} \setminus \mathcal{N}(K)) \to \Bl_\gamma(M_{P(K,\eta)}).
\end{equation}
Note that although~\cite[Theorem 1.1]{FriedlLeidyNagelPowell} is stated for Ore domain coefficients, Proposition~\ref{prop:BlanchfieldOreTwisted} implies that it also holds in the twisted setting.
The next paragraph is devoted to checking that~$H_*(M_{P(K,\eta)})$ and $H_*(T)$ are indeed $\L{\F}$-torsion modules.
Since the representation $\gamma$ was assumed to be acyclic, $H_*(M_{P(K,\eta)})$ is indeed torsion and we must only show that $H_*(T)$ is torsion.
We assert that
\begin{equation}
\label{eq:HomologyTorus}
H_{i}(T) =
\begin{cases}
\mathbb{F}[t^{\pm1}]^{d} / (1-\gamma(\eta)) & i = 0,1, \\
0 & \text{otherwise}.
\end{cases}
\end{equation}
Indeed, arguing as in~\cite[Proposition 3.11]{PowellThesis}, we see that the twisted chain complex \(C_{\ast}(T)\) is chain homotopy equivalent to
$ \mathbb{F}[t^{\pm1}]^{d} \xrightarrow{\left(1-\gamma(\eta) \quad\gamma(\mu_{\eta})-1\right)^{T}} \mathbb{F}[t^{\pm1}]^{2d} \xrightarrow{\left(1-\gamma(\mu_{\eta}) \quad 1-\gamma(\eta)\right)} \mathbb{F}[t^{\pm1}]^{d}$
with a basis of \(C_{1}(T) \cong \mathbb{F}[t^{\pm1}]^{2d} \) given by the \(\mu_{\eta} \otimes e_i \) and \(\eta \otimes e_i\) for $i=1,\ldots,d$.
Since $\gamma$ is $\eta$-regular, \(\gamma(\mu_{\eta})=1\) and the assertion follows.
For later use, note that this argument also proves that the~$\mathbb{F}[t^{\pm 1}]$-module \(H_{1}(T)\) is generated by the homology classes~\([e_i \otimes \mu_{\eta}]\) for $i=1,\ldots,d$.
Since~$\gamma$ is $\eta$-regular,~\eqref{eq:HomologyTorus} shows that $H_*(T)$ is torsion, concluding our verification of the assumptions of~\cite[Theorem 1.1]{FriedlLeidyNagelPowell} and thus establishing the existence of the morphism $\psi$ in~\eqref{eq:morphism-inking forms}.
Our goal is now to relate the Blanchfield pairings on $M_P,M_K$ and $M_{P(K,\eta)}$.
Observe that the $\L{\F}$-modules $H_1(M_{P(K,\eta)}), H_1(M_{P} \setminus \mathcal{N}(\eta),T)$ and $H_1(S^{3} \setminus \mathcal{N}(K),T)$ are torsion: for~$H_1(M_{P(K,\eta)})$, this follows from the $\eta$-regularity of $\gamma$.
For the two latter modules, this can be seen by looking at the long exact sequence of the pairs $(S^{3} \setminus \mathcal{N}(K),T)$ and $(M_{P} \setminus \mathcal{N}(\eta),T)$ together with the fact that~$H_1(M_{P} \setminus \mathcal{N}(\eta)),H_1(S^{3} \setminus \mathcal{N}(K))$ and $H_*(T)$ are all $\L{\F}$-torsion.
Combining this observation with $ H_{0}(M_{P} \setminus \mathcal{N}(\eta),T) = 0=H_{0}(M_{P} \setminus \mathcal{N}(\eta),T)$, Poincar\'e duality and the universal coefficient theorem, we obtain the following isomorphisms:
\begin{align}
H_{2}(M_{P(K,\eta)}) &\cong H^{1}(M_{P(K,\eta)}) \cong \operatorname{Ext}^{1}_{\text{left-}\L{\F}}(H_{0}(M_{P(K,\eta)}),\L{\F})^{\#} \cong H_{0}(M_{P(K,\eta)})^{\#}, \label{eq:H-2-1} \\
H_{2}(M_{P} \setminus \mathcal{N}(\eta)) &\cong H^{1}(M_{P}\setminus \mathcal{N}(\eta),T) \cong \operatorname{Ext}^{1}_{\L{\F}}(H_{0}(M_{P} \setminus \mathcal{N}(\eta),T),\L{\F}) = 0, \label{eq:H-2-2} \\
H_{2}(S^{3} \setminus \mathcal{N}(K)) &\cong H^{1}(S^{3}\setminus \mathcal{N}(K),T) \cong \operatorname{Ext}^{1}_{\L{\F}}(H_{0}(S^{3}\setminus \mathcal{N}(K),T),\L{\F}) = 0. \label{eq:H-2-3}
\end{align}
In more details, in each of the three lines, in our application of the universal coefficient theorem, we used our observation that the relevant first homology modules are $\L{\F}$-torsion.
In the third isomorphism of~\eqref{eq:H-2-1}, we also used that for any torsion \(\L{\F}\)-module \(T\), we have~$\operatorname{Ext}^{1}_{\L{\F}}(T,\L{\F}) \cong~T$.
Our strategy to relate the various Blanchfield pairings is to apply Theorem~\ref{thm:isoprojection}.
We now introduce the notation necessary to its application as well as verify one of its key assumptions.
\begin{lemma}
\label{claim:InclusionIsom}
The inclusions $\iota_{P} \colon M_{P} \setminus \mathcal{N}(\eta) \hookrightarrow M_{P}$ and $\iota_{K} \colon S^{3} \setminus \mathcal{N}(K) \hookrightarrow M_{K}$ induce isomorphisms on twisted homology in degree $j \neq 1,2$ and
epimorphisms $ (\iota_{P})_{\ast}, (\iota_{K})_{\ast}$ in degree one.
Moreover, \(\ker (\iota_{P})_{\ast}\) and \(\ker(\iota_{K})_{\ast}\) are generated by the~\([e_i \otimes \mu_{\eta}]\) for $i=1,\ldots, d$.
Additionally, the modules $H_1(M_K)$ and $H_1(M_P)$ are $\L{\F}$-torsion.
Finally, we have the following equalities
\begin{align*}
\ord H_{1}(M_{P} \setminus \mathcal{N}(\eta)) &\doteq \frac{\ord(H_{1}(M_{P})) \cdot \ord(H_{1}(T))}{\ord(H_{0}(M_{P}))^\#}, \\
\ord H_{1}(S^{3} \setminus \mathcal{N}(K)) &\doteq \frac{\ord(H_{1}(M_{K})) \cdot \ord(H_{1}(T))}{\ord(H_{0}(M_{K}))^\# }.
\end{align*}
\end{lemma}
\begin{proof}
We only prove the lemma for \(M_{P} \setminus \mathcal{N}(\eta)\): the proof for \(S^{3} \setminus \mathcal{N}(K)\) is completely analogous.
Using~\eqref{eq:HomologyTorus} and the Mayer-Vietoris sequence of \(M_{P} = (M_{P} \setminus \mathcal{N}(\eta)) \cup_T \mathcal{N}(\eta)\), we deduce that~$i_P$ does indeed induce an isomorphism on twisted homology in degree $j > 2$.
In degree zero, notice that the inclusion \(T \hookrightarrow \mathcal{N}(\eta)\) induces an isomorphism on \(H_{0}\) and therefore, with the aid of the Mayer-Vietoris sequence, we obtain an isomorphism \(H_{0}(M_{P} \setminus \mathcal{N}(\eta)) \cong H_{0}(M_{P})\).
To get the result in degrees $j=1,2$, consider the relevant part of the Mayer-Vietoris sequence:
\begin{equation}\label{eq:MV-sequence}
0 \xrightarrow{} H_{2}(M_{P}) \xrightarrow{} H_1(T) \xrightarrow{j_{P}} H_1(M_{P} \setminus \mathcal{N}(\eta)) \xrightarrow{(\iota_{P})_{\ast}} H_1(M_P) \to 0.
\end{equation}
Notice that the leftmost part of~\eqref{eq:MV-sequence} is exact by~\eqref{eq:H-2-2}.
Since $H_1(M_P \setminus \mathcal{N}(\eta))$ is $\L{\F}$-torsion,~\eqref{eq:MV-sequence} implies that this is also the case for $H_1(M_P)$.
Next,~\eqref{eq:MV-sequence} also implies that the map~\((\iota_{P})_{\ast}\) is an epimorphism with kernel generated by the~\([e_i \otimes \mu_{\eta}]\) (or technically by~\( j_P([e_i \otimes \mu_{\eta}])\), because~\(H_{1}(T)\) is generated by these elements.
The last assertion of the lemma also follows from~\eqref{eq:MV-sequence}: we use the multiplicativity of the orders and the same isomorphisms as in~\eqref{eq:H-2-1} to establish the isomorphism $H_{2}(M_{P}) \cong H_0(M_{P})^{\#}$ (this is possible since we now know that $H_1(M_P)$ is torsion).
This concludes the proof of the lemma.
\end{proof}
We now prove the first assertion of the theorem, namely that $\gamma_P$ and $\gamma_K$ are acyclic. We only check this for $\gamma_P$: the proof for $\gamma_K$ is identical. Since $H_*(M_P \setminus \mathcal{N}(\eta))$ is $\L{\F}$-torsion, Lemma~\ref{claim:InclusionIsom} implies that $H_i(M_P)$ is $\L{\F}$-torsion for $i \neq 2$. For $i=2$, we use the isomorphism $H_2(M_P)\cong H_0(M_P)^\#$ (that we established in the proof of Lemma~\ref{claim:InclusionIsom}) and the conclusion follows.
As a next step, consider the decomposition \(M_{P(K,\eta)} = \left(M_{P}\setminus \mathcal{N}(\eta)\right) \cup_T \left(S^{3} \setminus \mathcal{N}(K)\right)\).
The associated Mayer-Vietoris sequence can be decomposed into the following three exact sequences:
\begin{align}
&0 \xrightarrow{} H_{2}(M_{P(K,\eta)}) \xrightarrow{} H_{1}(T) \xrightarrow{} \mathcal{K} \xrightarrow{} 0, \label{eq:ex-seq-1}\\
&0 \xrightarrow{} \mathcal{K} \xrightarrow{} H_{1}(M_{P}\setminus \mathcal{N}(\eta)) \oplus H_{1}(S^{3}\setminus \mathcal{N}(K)) \xrightarrow{\psi} H_{1}(M_{P(K,\eta)}) \xrightarrow{} \mathcal{C} \xrightarrow{} 0, \label{eq:ex-seq-2}\\
&0 \xrightarrow{} \mathcal{C} \xrightarrow{} H_{0}(T) \xrightarrow{} H_{0}(M_{P}\setminus \mathcal{N}(\eta)) \oplus H_{0}(S^{3}\setminus \mathcal{N}(K)) \xrightarrow{} H_{0}(M_{P(K,\eta)}) \xrightarrow{} 0 \label{eq:ex-seq-3},
\end{align}
where $\mathcal{K}=\ker(\psi)$ and $\mathcal{C}=\coker(\psi)$.
Therefore, applying the multiplicativity of the orders, the use of~\eqref{eq:ex-seq-1} and~\eqref{eq:ex-seq-3} (as well as~\eqref{eq:H-2-1}) implies that
\begin{align*}
\ord (\mathcal{C}) &\doteq \frac{\ord(H_{0}(T)) \cdot \ord(H_{0}(M_{P(K,\eta)}))}{\ord(H_{0}(M_{P}\setminus \mathcal{N}(\eta))) \cdot \ord(H_{0}(S^{3} \setminus \mathcal{N}(K)))} \doteq \frac{\ord(H_{0}(T)) \cdot \ord(H_{0}(M_{P(K,\eta)}))}{\ord(H_{0}(M_{P})) \cdot \ord(H_{0}(M_{K}))}, \\
\ord (\mathcal{K} )&\doteq \frac{\ord(H_{1}(T))}{\ord(H_{2}(M_{P(K,\eta)}))} \doteq \frac{\ord(H_{1}(T))}{\ord(H_{0}(M_{P(K,\eta)}))^\#}.
\end{align*}
The key $\L{\F}$-module in our application of Theorem~\ref{thm:isoprojection} is defined as
\[A := \frac{H_{1}(M_{P}\setminus \mathcal{N}(\eta)) \oplus H_{1}(S^{3}\setminus \mathcal{N}(K))}{\mathcal{K}}.\]
Combining~\eqref{eq:ex-seq-2} with our computation of $\ord \mathcal{C}$, we obtain
\[\ord (A) \doteq \frac{\ord(H_{1}(M_{P(K,\eta)}))}{\ord \mathcal{C}} \doteq \frac{\ord(H_{1}(M_{P(K,\eta)})) \cdot \ord(H_{0}(M_{P})) \cdot \ord H_{0}(M_{K})}{\ord (H_{0}(T)) \cdot \ord(H_{0}(M_{P(K,\eta)}))}.\]
On the other hand, Lemma~\ref{claim:InclusionIsom} and our computation of $\ord (\mathcal{K})$ imply that
\begin{align*}
\ord (A) &\doteq \frac{\ord(H_{1}(M_{P} \setminus \mathcal{N}(\eta))) \cdot \ord(H_{1}(S^{3}\setminus \mathcal{N}(K)))}{\ord \mathcal{K}} \\
&\doteq \frac{\ord(H_{1}(M_{P})) \cdot \ord(H_{1}(M_{K})) \cdot \ord(H_{1}(T))^{2} \cdot \ord(H_{0}(M_{P(K,\eta)}))^\# }{\ord(H_{0}(M_{P}))^\# \cdot \ord(H_{0}(M_{K}))^\# \cdot \ord(H_{1}(T))} \\
&\doteq \frac{\ord(H_{1}(M_{P})) \cdot \ord(H_{1}(M_{K})) \cdot \ord(H_{1}(T)) \cdot \ord(H_{0}(M_{P(K,\eta)}))^\# }{\ord(H_{0}(M_{P}))^\# \cdot \ord(H_{0}(M_{K}))^\# }.
\end{align*}
Additionally, observe that $\ord(H_1(T))=\ord(H_0(T))^\#$ (this follows by duality and the universal coefficient theorem) and $\ord(H_1(M_{P(K,\eta)})=\ord(H_1(M_{P(K,\eta)})^\#$ (since $H_1(M_{P(K,\eta)})$ supports a non-singular linking form, recall Remark~\ref{rem:SplitInvolution}, or using duality and universal coefficients). Combining these observations with the above formulas for $\ord (A)$, we obtain
\begin{equation}
\label{eq:ord-A}
\ord (A) \ord (A)^\# \doteq \ord(H_{1}(M_{P})) \cdot \ord(H_{1}(M_{K})) \cdot \ord(H_{1}(M_{P(K,\eta)})),
\end{equation}
Observe that \((\iota_{P})_{\ast} \oplus (\iota_{K})_{\ast}\) and \(\psi\) vanish on $\mathcal{K}$.
As a consequence, we let \((\iota_{P})_{\ast} \oplus (\iota_{K})_{\ast} / \mathcal{K}\) and~\(\psi / \mathcal{K}\) denote the maps induced on \(A\).
Additionally, using \(\widetilde{L}\) to denote the kernel of \((\iota_{P})_{\ast} \oplus (\iota_{K})_{\ast} / \mathcal{K}\) and~\(\operatorname{Cok}\) to denote the cokernel of \(\psi / \mathcal{K}\), we obtain the following diagram of exact sequences:
\begin{equation}
\label{eq:ApplyAlgebra}
\xymatrix@R0.5cm{
\widetilde{L} \ar@{^{(}->}[d] \\
A \ar@{^{(}->}[r]^{\psi/\mathcal{K}} \ar@{->>}[d]_{(\iota_{P})_{\ast} \oplus (\iota_{K})_{\ast} / \mathcal{K}} & H_1(M_{P(K,\eta)}) \ar[r] & \operatorname{Cok}. \\
H_1(M_P) \oplus H_1(M_K)
}
\end{equation}
To check the second assertion of the theorem, we will show that $(\psi/\mathcal{K})(\widetilde{L}) = L$, where $L$ was defined as $\operatorname{im} (H_1(T) \to H_1(M_{P(K,\eta)}))$.
Use $j_P$ (resp. $j_K$) to denote the map induced by the inclusion $T \to M_P \setminus \mathcal{N}(\eta)$ (resp. $T \to S^3 \setminus \mathcal{N}(K)$).
The exact sequence in~\eqref{eq:ex-seq-2} implies that~\(\mathcal{K}\), treated as a submodule of \(H_{1}(M_{P}\setminus \mathcal{N}(\eta)) \oplus H_{1}(S^{3}\setminus \mathcal{N}(K))\), is generated by \((j_{P}([e_{i} \otimes \mu_{\eta}]), -j_{K}([e_{i}\otimes\mu_{\eta}]))\), for~$i=1,\ldots, d$.
Therefore, for any $i$, we have \((j_{P}([e_{i} \otimes \mu_{\eta}]),0) = (0,j_{K}([e_{i} \otimes \mu_{\eta}]))\) in \(A\).
Moreover,
\[\widetilde{L} =\ker \left((i_P)_{\ast} \oplus (i_K)_{\ast}\right) = \langle (j_{P}[e_{i}\otimes \mu_{\eta}], j_{K}[e_{i}\otimes \mu_{\eta}]) \mid 1 \leq i \leq d\rangle / \langle j_{P}[e_{i} \otimes \mu_{\eta}] = j_{K}[e_{i} \otimes \mu_{\eta}] \rangle,\]
and thus \((\psi/\mathcal{K})(\widetilde{L}) = L\). Since \(\psi/\mathcal{K}\) is injective, it follows that \(L\) and \(\widetilde{L}\) are isomorphic.
In order to prove that \(L\) is isotropic, observe that the homology classes \([e_{i}\otimes \mu_{\eta}]\) belong to the radical of~\(\Bl_{\gamma_{P}}(M_{P} \setminus \mathcal{N}(\eta))\) and \(\Bl_{\gamma_{K}}(S^{3} \setminus \mathcal{N}(K))\), because they come from the boundary of the respective manifolds.
Since \(\psi\) and \(\psi / \mathcal{K}\) are morphisms of linking forms, it follows that \(L\) is isotropic.
The constraint on the order of \(L\) comes from the fact that \(L\) is a quotient of \(H_{1}(T)\).
To prove the third statement, first observe that $\Bl_{\gamma_P}(M_P \setminus \mathcal{N}(\eta)) \oplus \Bl_{\gamma_K}(S^3 \setminus \mathcal{N}(K)) $ vanishes on $\mathcal{K}$: indeed $\mathcal{K}$ only contains homology classes which come from the boundary torus $T$.
As a consequence, $\Bl_{\gamma_P}(M_P \setminus \mathcal{N}(\eta)) \oplus \Bl_{\gamma_K}(S^3 \setminus \mathcal{N}(K))$ descends to $A$.
Next, we check the assumptions of Theorem~\ref{thm:isoprojection}.
First, it is clear that the maps \((\iota_{P})_{\ast} \oplus (\iota_{K})_{\ast} / \mathcal{K}\) and \(\psi / \mathcal{K}\) are morphisms of linking forms.
Next, by definition of $\mathcal{K}$, the map $\psi / \mathcal{K}$ is injective, while \((\iota_{P})_{\ast} \oplus (\iota_{K})_{\ast} / \mathcal{K}\) is surjective thanks to~\eqref{eq:MV-sequence}.
Since $M_P, M_K$ and $M_{P(K,\eta)}$ are closed manifolds, the corresponding Blanchfield forms are non-singular.
Finally, we already showed in~\eqref{eq:ord-A} that $\ord (A) \cdot (\ord (A))^\# \doteq \ord(H_{1}(M_{P})) \cdot \ord(H_{1}(M_{K})) \cdot \ord(H_{1}(M_{P(K,\eta)}))$.
We can therefore apply Theorem~\ref{thm:isoprojection} to the above diagram to obtain the third assertion of the theorem, concluding the proof.
\end{proof}
The next result shows that Blanchfield pairings need not be additive under satellite operations.
\begin{corollary}\label{cor:cabling-formula}
Using the same assumptions as in Theorem~\ref{thm:CablingTheorem}, if \(\ord (L) \) is relatively prime to \(\ord(H_{1}(M_{P};\L{\F}^{d}_{\gamma_{P}})) \cdot \ord(H_{1}(M_{K};\L{\F}^{d}_{\gamma_{K}}))\), then there is an isometry of linking forms
\[\Bl_\gamma(P(K,\eta)) \cong \Bl_{\gamma_P}(P) \oplus \Bl_{\gamma_K}(K) \oplus (X,\lambda_{X}),\]
where \((X,\lambda_{X})\) is a metabolic linking form and \(L \subset X\) is a metabolizer for \(\lambda_{X}\).
In particular, if the map \(H_{1}(T;\L{\F}^{d}_{\gamma}) \to H_{1}(M_{P(K,\eta)};\L{\F}^{d}_{\gamma})\) vanishes, then we obtain the following isometry of linking forms:
\[\Bl_{\gamma}(P(K,\eta)) \cong \Bl_{\gamma_P}(P) \oplus \Bl_{\gamma_K}(K).\]
\end{corollary}
\begin{proof}
We first recall some notations from the proof of Theorem~\ref{thm:CablingTheorem}. Recall that we set $\mathcal{K}=~\ker(\psi)$, where $\psi \colon \Bl_{\gamma_P}(M_{P} \setminus \mathcal{N}(\eta)) \oplus \Bl_{\gamma_K}(S^{3} \setminus \mathcal{N}(K)) \to \Bl_\gamma(M_{P(K,\eta)})$ is the morphism of linking forms described in~\eqref{eq:morphism-inking forms}. We once again omit coefficients and recall that $A:=(H_1(M_P \setminus \mathcal{N}(\eta) )\oplus H_1(S^3 \setminus \mathcal{N}(K)))/\mathcal{K}$. Since we saw that $\Bl_{\gamma_P}(P) \oplus \Bl_{\gamma_K}(K)$ descends to $A$; we let $\Bl_A$ denote the resulting pairing.
\begin{claim*}
There is a linking form $(X,\lambda_X)$ such that $ \Bl_\gamma(P(K,\eta)) \cong \Bl_{\gamma_{P}}(P) \oplus \Bl_{\gamma_{K}}(K) \oplus (X,\lambda_X)$.
\end{claim*}
\begin{proof}
Using the multiplicativity of orders on the vertical exact sequence in~\eqref{eq:ApplyAlgebra}, we see that $\ord (A) = \ord(\widetilde{L}) \cdot \ord(H_{1}(M_{P})) \cdot \ord(H_{1}(M_{K})).$
Using the primary decomposition theorem (since~$\L{\F}$ is a PID), the Chinese remainder theorem and \(\gcd(\ord (\widetilde{L}), \ord(H_{1}(M_{P})) \cdot \ord(H_{1}(M_{K}))) = 1\), we obtain an isomorphism
$A \cong \widetilde{L} \oplus B,$
where $B$ is isomorphic to $H_{1}(M_{P}) \oplus H_{1}(M_{K})$.
Observe that the surjectivity of \((\iota_{P})_{\ast} \oplus (\iota_{K})_{\ast}/\mathcal{K}\) implies that
\begin{equation}
\label{eq:BNonSingular}
\Bl_{A}|_{B} \cong \Bl_{\gamma_{P}}(P) \oplus \Bl_{\gamma_{K}}(K).
\end{equation}
Thus, $\psi/\mathcal{K} \colon A \to H_1(M_{P(K,\eta)})$ is an injective map of linking forms, $\Bl_\gamma(P(K,\eta))$ is non-singular (and thus split), and $(B,\Bl_{A}|_{B})$ is non-singular (by~\eqref{eq:BNonSingular}). We can therefore apply Proposition~\ref{prop:missing_in_action} to the injective morphism of linking forms $\psi/\mathcal{K}|_B \colon B \to H_1(M_{P(K,\eta)})$. This implies that for some linking form $(X,\lambda_X)$, we have the isometry \(\Bl_\gamma(P(K,\eta)) \cong (\psi/\mathcal{K})(\Bl_A|_{B}) \oplus~(X,\lambda_X)\).
This concludes the proof of the claim.
\end{proof}
Since Theorem~\ref{thm:CablingTheorem} shows that \(\Bl_\gamma(M_{P(K,\eta)})\) is Witt equivalent to $\Bl_{\gamma_P}(P) \oplus~\Bl_{\gamma_K}(K)$, the claim implies that
$(X,\lambda_X)$ is metabolic. It remains to show that a metabolizer for $(X,\lambda_X)$ is in fact given by $L:= (\psi/\mathcal{K})(\widetilde{L})$.
As $(\psi/\mathcal{K}) \colon A \to H_1(M_{P(K,\eta)}) $ is an injective morphism of linking forms, the isomorphisms $A \cong \widetilde{L} \oplus B$ and $H_1(M_{P(K,\eta)})~\cong \psi/\mathcal{K}(B) \oplus~X$ imply that \(L = (\psi/\mathcal{K})(\widetilde{L})\) is an isotropic submodule of $X$.
Since Theorem~\ref{thm:CablingTheorem} shows that $\Bl_{\gamma_P}(P) \oplus \Bl_{\gamma_K}(K)$ is isometric to the isotropic reduction of~\(\Bl_\gamma(M_{P(K,\eta)})\) with respect to $L$, the multiplicativity of orders implies that~$\ord (X)=\ord(L)\ord(L)^{\#}$. As $\L{\F}$ is a PID, we deduce that~\(L\) is a metabolizer in~\(X\), concluding the proof of the first assertion.
In order to prove the second assertion, notice that our hypothesis implies that the submodule~$L$ is trivial. Since $L$ is a metabolizer of $X$, we deduce that $X$ is itself trivial. The conclusion now follows from the first assertion.
\end{proof}
The situation simplifies considerably when \(\gamma_{K}\) is abelian, i.e.\ factors through the abelianization of \(\pi_{1}(M_{K})\).
Let \(\mathbb{F}[t_{K}^{\pm1}]\) be the group algebra of the abelianization of \(\pi_{1}(M_{K})\) with \(t_{K}\) denoting the generator corresponding to the meridian identified with \(\eta\).
In this case, \(\L{\F}^{d}_{\gamma_{K}}\) becomes a module over \(\mathbb{F}[t_{K}^{\pm1}]\) via the homomorphism \(\gamma_{K} \colon \mathbb{F}[t_K^{\pm 1}] \to \L{\F}^d_{\gamma_{K}} \), i.e.\ via the action $t_{K} \cdot v = v\gamma(\eta) $ with~$v \in \L{\F}^{d}_{\gamma_{K}}$.
In this setting, Theorem~\ref{thm:CablingTheorem} takes the following form (compare with~\cite[Theorem 1.3]{FriedlLeidyNagelPowell}).
\begin{corollary}\label{cor:cabling-formula-abelian}
Using the same assumptions as in Theorem~\ref{thm:CablingTheorem}, if \(\gamma_{K}\) is an abelian representation, then there exists an isometry of linking forms
\[\Bl_{\gamma}(P(K,\eta)) \cong \Bl_{\gamma_P}(P) \oplus \Bl_{\gamma_K}(K).\]
If we denote by \(\mathbb{F}[\eta] \subset \mathbb{F}[\pi(M_{P})]\) the subring generated by \(\eta\)
and we assume that \(\det(\gamma_{P}(x)) \neq 0\), for any \(x \in \mathbb{F}[\eta] \setminus \{0\}\), then $H_{1}(M_{K};\L{\F}^{d}_{\gamma})$ is isomorphic to $H_{1}(M_{K};\mathbb{F}[t_{K}^{\pm1}]) \otimes_{\mathbb{F}[t_{K}^{\pm1}]} \L{\F}^{d}_{\gamma_{K}},$ and there is an isometry
$$\Bl_{\gamma_K}(K) \cong \Bl(K) \otimes_{\mathbb{F}[t_{K}^{\pm1}]} \L{\F}^{d}_{\gamma_{K}}.$$
\end{corollary}
\begin{proof}
Let \(\lambda\) denote the zero-framed longitude of \(K\).
If \(F \subset S^{3} \setminus \mathcal{N}(K)\) is a Seifert surface for~\(K\), then it follows that the image of \(\pi_{1}(F)\) is contained in the commutator subgroup of \(\pi_{1}(S^{3} \setminus \mathcal{N}(K))\), therefore the inclusion of \(F\) lifts to a map to the infinite cyclic cover of \(S^{3} \setminus \mathcal{N}(K)\).
This means that for any \(\mathbb{F}[t_{K}^{\pm1}]\)-module \(M\), the homology class of \(\lambda\) is trivial in \(H_{1}(S^{3} \setminus \mathcal{N}(K);M)\) and~\(H_{1}(M_{K};M)\).
Thus, the hypothesis of Corollary~\ref{cor:cabling-formula} is satisfied
and we therefore obtain the desired isometry of linking forms.
If we have \(\det(\gamma(x))\neq 0\) for any \(x \in \mathbb{F}[\eta] \setminus \{0\}\), then \(\L{\F}^{d}_{\gamma_{K}}\) is a torsion-free \(\mathbb{F}[t_{K}^{\pm1}]\)-module and hence (since~$\L{\F}$ is a PID) flat~\(\mathbb{F}[t_{K}^{\pm1}]\)-module.
The last two assertions follow easily from the flatness of~\(\L{\F}^{d}_{\gamma_{K}}\).
\end{proof}
Next, we discuss the consequences of Theorem~\ref{thm:CablingTheorem} on signatures of satellite knots.
\begin{corollary}
\label{cor:SignatureSatellite}
Let $P(K,\eta)$ be a satellite knot with pattern $P$, companion $K$ and infection curve~$\eta$.
Let $\gamma \colon \pi_1(M_{P(K,\eta)}) \to GL_d(\L{\F})$ be a unitary acyclic $\eta$-regular representation.
If the assumptions of Theorem~\ref{thm:CablingTheorem} hold, then for any $\omega \in S^1$, the averaged twisted signatures of~$P(K,\eta),P$ and $K$ satisfy
$$ \sigma_{P(K,\eta),\gamma}^{\operatorname{av}}(\omega)=\sigma_{P,\gamma_P}^{\operatorname{av}}(\omega)+\sigma_{K,\gamma_K}^{\operatorname{av}}(\omega).$$
In particular, under these assumptions, the averaged twisted signature is additive on connected~sums:
$$ \sigma_{P \# K,\gamma}^{\operatorname{av}}(\omega)=\sigma_{P,\gamma_P}^{\operatorname{av}}(\omega)+\sigma_{K,\gamma_K}^{\operatorname{av}}(\omega).$$
\end{corollary}
\begin{proof}
Applying Theorem~\ref{thm:CablingTheorem},
\(\Bl_\gamma(P(K,\eta))\) is Witt equivalent to \(\Bl_{\gamma_{P}}(P) \oplus \Bl_{\gamma_K}(K)\).
The first assertion now follows from the Witt invariance of the averaged signature, recall Proposition~\ref{prop:avsig}.
The second assertion is immediate since the connected sum is obtained by taking the infection curve $\eta$ to be the meridian of $P$.
\end{proof}
Finally, in the abelian case, we recover results concerning the behavior of the Blanchfield pairing and Levine-Tristram signature under satellite operations (see~\cite{LivingstonMelvin} and~\cite{Shinohara, LitherlandIterated}).
\begin{corollary}
\label{cor:SignatureSatellite_abelian}
Let $P(K,\eta)$ be a satellite knot with pattern $P$, companion $K$, infection curve~$\eta$ and winding number \(w = \lk(\eta,P)\).
The Blanchfield pairing and Levine-Tristram signature sastify
\begin{align*}
\Bl(P(K,\eta))(t)&=\Bl(P)(t) \oplus \Bl(K)(t^w), \\
\sigma_{P(K,\eta)}(\omega)&=\sigma_{P}(\omega)+\sigma_{K}(\omega^w).
\end{align*}
\end{corollary}
\begin{proof}
The representation $\gamma$ is abelianization and the composition $ H_1(M_K;\mathbb{Z}) \to H_1(M_{P(K,\eta)};\mathbb{Z}) \cong~\mathbb{Z}$ is given by multiplication by $w$.
The result now follows from Remark~\ref{rem:LevineTristram} and Corollary~\ref{cor:cabling-formula-abelian}.
\end{proof}
\section{Metabelian Representations and Casson-Gordon invariants}
\label{sec:CassonGordon}
The aim of this section is to study signatures associated to metabelian representations.
In Subsection~\ref{sub:Metabelian}, we recall two equivalent constructions of a ``metabelian Blanchfield pairing", in Subsection~\ref{sub:MillerPowell}, we review a result due to Miller and Powell, in Subsection~\ref{sub:CassonGordon}, we recall the definition of the Casson-Gordon $\tau$-invariant and in Subsection~\ref{sec:Proof}, we relate it to our twisted signatures.
Throughout this section, we use $L_n(K)$ to denote the $n$-fold cover of $S^3$ branched along a knot $K$ and fix a character $\chi \colon H_1(L_n(K);\mathbb{Z}) \to \mathbb{Z}_m$. We also set \(\xi_{m} = e^{\frac{2 \pi i}{m}}\) and~$M_K$ will always denote the 0-framed surgery along $K$.
\subsection{The metabelian Blanchfield pairing}
\label{sub:Metabelian}
In this subsection, we review two equivalent definitions of a particular twisted Blanchfield pairing associated to $M_K$.
The first makes use of a metabelian representation $\alpha(n,\chi) \colon \pi_1(M_K) \to GL_n(\L{\C})$, while the second uses the $n$-fold cover $M_n \to M_K$.
References for this subsection include~\cite{HeraldKirkLivingston, FriedlEta, MillerPowell}.
\medbreak
Use $\phi_K \colon \pi_1(M_K) \to H_1(M_K;\mathbb{Z})\cong \mathbb{Z}=\langle t_K \rangle$ to denote the abelianization homomorphism.
Since~$\phi_K$ endows $\mathbb{Z}[t_K^{\pm 1}]$ with a right $\pi_1(M_K)$-module structure, it gives rise
to the twisted homology $\mathbb{Z}[t_K^{\pm1}]$-module $H_1(M_K;\mathbb{Z}[t_{K}^{\pm1}])$.
This module is known to coincide with the Alexander module of $K$. In what follows, we shall frequently identify $H_1(L_n(K);\mathbb{Z})$ with $H_1(M_K;\mathbb{Z}[t_K^{\pm 1}])/(t_K^n-1)$ as in~\cite[Corollary 2.4]{FriedlPowell}. We now consider the semidirect product $\mathbb{Z} \ltimes H_1(L_n(K);\mathbb{Z})$, where the group law is given by $ (t_K^i,v)\cdot (t_K^j,w)=(t_K^{i+j},t_K^{-j}v+w)$. Next, consider the representation
\begin{align}
\label{eq:Matrix}
\gamma_K(n,\chi) \colon \mathbb{Z} \ltimes H_1(L_n(K);\mathbb{Z}) &\to ~\operatorname{GL}_n(\L{\C}) \nonumber \\
(t_K^j,v) &\mapsto \begin{pmatrix}
0& 1 & \cdots &0 \\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 & \cdots & 1 \\
t & 0 & \cdots & 0
\end{pmatrix}^j
\begin{pmatrix}
\xi_{m}^{\chi(v)} & 0 & \cdots &0 \\
0 & \xi_{m}^{\chi(t_K \cdot v)} & \cdots &0 \\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 & \cdots & \xi_{m}^{\chi(t_K^{n-1} \cdot v)}
\end{pmatrix}.
\end{align}
We will now use $\gamma_K(n,\chi)$ to obtain a representation of $\pi_1(M_K)$. To achieve this, we identify the Alexander module $H_1(M_K;\mathbb{Z}[t_{K}^{\pm1}])$ with the derived quotient $\pi_1(M_K)^{(1)}/\pi_1(M_K)^{(2)}$ and consider the following composition of canonical projections
\[q_K \colon \pi_{1}(M_{K})^{(1)} \to H_1(M_K;\mathbb{Z}[t_K^{\pm 1}]) \to H_1(L_n(K);\mathbb{Z}).\]
Fix an element~$\mu_{K}$ in $\pi_1(M_K)$ such that $\phi_K(\mu_{K})=t_K$.
Note that for every $g \in \pi_1(M_K)$, we have~$\phi_K(\mu_K^{-\phi_K(g)}g)=1$.
Since $\phi_K$ is the abelianization map, we deduce that $\mu_K^{-\phi_K(g)}g$ belongs to~$\pi_1(M_K)^{(1)}$.
As a consequence, we obtain the following map:
\begin{align}
\label{eq:MapToSemiDirect}
\widetilde{\rho}_K \colon \pi_1(M_K) &\to \mathbb{Z} \ltimes H_1(L_n(K);\mathbb{Z}) \\
g &\mapsto (\phi_K(g),q_K(\mu_{K}^{-\phi_K(g)}g)). \nonumber
\end{align}
Precomposing the representation $\gamma_K(n,\chi)$ with $\widetilde{\rho}_K$ provides the unitary representation
\begin{equation}
\label{eq:ActionForMK}
\alpha(n,\chi)_K \colon \pi_1(M_K) \stackrel{\widetilde{\rho}}{\to} \mathbb{Z} \ltimes H_1(L_n(K);\mathbb{Z}) \stackrel{\gamma_K(n,\chi)}{\longrightarrow} \operatorname{GL}_n(\L{\C}).
\end{equation}
If $m$ is a prime power and $\chi \colon H_1(L_n(K);\mathbb{Z}) \to \mathbb{Z}_m$ is nontrivial, it is known that the unitary representation $\alpha(n,\chi)$ is acyclic~\cite{FriedlPowellInjectivity} (see also~\cite[Lemma 6.6]{MillerPowell}). When the knot $K$ is clear from the context, we drop it from the notation and simply write~$\alpha(n,\chi)$ instead of $\alpha_K(n,\chi)$.
The next lemma illustrates these notations and will be used to ensure that the hypotheses of Theorem~\ref{thm:CablingTheorem} hold.
\begin{lemma}
\label{lem:H0CG}
If $\chi \colon H_1(L_n(K);\mathbb{Z}) \to \mathbb{Z}_m$ is nontrivial, then $H_0(M_K;\L{\C}_{\alpha(n,\chi)}^n)=0$.
\end{lemma}
\begin{proof}
As $H_0(M_K;\L{\C}_{\alpha(n,\chi)}^n)=\L{\C}^n/V$, with
\[V=\langle (\alpha(n,\chi)(g)-\operatorname{id})v \ | \ v \in \L{\C}^n, g \in ~\pi_1(M_K) \rangle,\]
it is enough to find $g \in \pi_1(M_K)$ such that $\alpha(n,\chi)(g)-\operatorname{id}$ is invertible over $\L{\C}$.
Since~$\chi$ is non-trivial, we can pick $g \in \pi_1(M_K)^{(1)}=\ker(\phi_K)$ such that $\chi(q_K(g)) \neq~0$. Since~$\phi_K(g)=~0$, the definition of $\alpha(n,\chi)$ implies that $\alpha(n,\chi)$ is a diagonal matrix with the $\xi_m^{\chi(t_K^iq_K(g))}$ on its main diagonal. Since $\chi(q_K(g)) \neq~0$, we deduce that $\det (\alpha(n,\chi)(g)-\operatorname{id})$ is a non-zero complex number. It follows that $\alpha(n,\chi)(g)-\operatorname{id}$ is invertible over $\L{\C}$ and the lemma is proved.
\end{proof}
We now use $\alpha(n,\chi)$ to endow $\mathbb{C}[t^{\pm 1}]^n$ with a right $\mathbb{Z}[\pi_1(M_K)]$-module structure and apply Definition~\ref{def:Blanchfield} to define a twisted Blanchfield pairing on $M_K$.
\begin{definition}
\label{def:BlanchfieldMetabelian}
Let $K$ be an oriented knot, $n$ be an integer, $m$ be a prime power and choose a nontrivial character $\chi \colon H_1(L_n(K);\mathbb{Z}) \to~\mathbb{Z}_m$.
The \emph{metabelian Blanchfield pairing} is the Blanchfield pairing twisted by the unitary acyclic representation $\alpha(n,\chi)$:
$$\Bl_{\alpha(n,\chi)}(K) \colon H_1(M_K;\L{\C}^n_{\alpha(n,\chi)}) \times H_1(M_K;\L{\C}^n_{\alpha(n,\chi)}) \to \O{\C}/ \L{\C}.$$
\end{definition}
Next, we provide a description of $\Bl_{\alpha(n,\chi)}(K)$ in terms of the $n$-fold cyclic cover $p \colon M_n \to M_K$. The first step is to consider the abelian subgroup $n\mathbb{Z} \times H_1(L_n(K);\mathbb{Z})$ of $\mathbb{Z} \ltimes H_1(L_n(K);\mathbb{Z})$ and the following $1$-dimensional representation:
\begin{align}
\label{eq:Reprho(n,chi)}
\rho_K(n,\chi) \colon n \mathbb{Z} \times H_1(L_n(K);\mathbb{Z}) &\to \L{\C} \\
(t_K^{nk},v) &\mapsto t^k \xi_{m}^{\chi(v)}. \nonumber
\end{align}
We now use $\rho_K(n,\chi)$ to obtain a representation of $\pi_1(M_n)$. Consider the composition $\pi_1(M_n) \stackrel{p_*}{\rightarrow} \pi_1(M_K) \to H_1(M_K;\mathbb{Z}) \cong \mathbb{Z}$.
Since the image of this map is isomorphic to $n \mathbb{Z}$, mapping to it produces a surjective map $\alpha \colon \pi_1(M_n) \to n \mathbb{Z}$.
Using the decomposition $H_1(M_n;\mathbb{Z}) \cong H_1(L_n(K);\mathbb{Z}) \oplus~\mathbb{Z}$ additionally gives rise to a homomorphism $\rho \colon \pi_1(M_n) \to H_1(L_n(K);\mathbb{Z})$. Consider the $\mathbb{Z}[\pi_1(M_n)]$-module structure on~$\mathbb{C}[t^{\pm 1}]$ given by the following composition:
\begin{align}
\label{eq:ActionForMn}
\alpha \times \chi \colon \pi_1(M_n) \stackrel{\alpha \times \rho}{\to} n \mathbb{Z} \times H_1(L_n(K);\mathbb{Z}) \stackrel{\rho_K(n,\chi)}{\to} \L{\C}.
\end{align}
Under the assumption that $m$ is a prime power and that $\chi$ is a nontrivial character, Casson and Gordon show that $H_*(M_n;\mathbb{C}(t))$ vanishes~\cite[Corollary after Lemma 4]{CassonGordon2}.
In other words, under these conditions, the unitary representation $\alpha \times \chi$ is acyclic.
\begin{definition}
\label{def:BlanchfieldCover}
Let $n$ be an integer, let $m$ be a prime power and let $\chi \colon H_1(L_n(K);\mathbb{Z}) \to \mathbb{Z}_m$ be a nontrivial character.
The \emph{Blanchfield pairing of the cover} $\Bl_{\alpha \times \chi}(M_n)$ is the Blanchfield pairing corresponding to the unitary acyclic representation~$\alpha \times \chi:$
\begin{equation}
\label{eq:BlanchfieldCover}
\Bl_{\alpha \times \chi}(M_n) \colon H_1(M_n;\L{\C}) \times H_1(M_n;\L{\C}) \to \mathbb{C}(t) / \L{\C}.
\end{equation}
\end{definition}
Up to the end of this subsection, we shall use $\L{\C}_{\alpha \times \chi}$ to denote the $\mathbb{Z}[\pi_1(M_n)]$-module structure on $\L{\C}$ given by~(\ref{eq:ActionForMn}) and by $\L{\C}^n_{\alpha(n,\chi)}$ the $\mathbb{Z}[\pi_1(M_K)]$-module structure on $\L{\C}^n$ given by~(\ref{eq:ActionForMK}).
The following result is due to Herald, Kirk and Livingston~\cite[Theorem 7.1]{HeraldKirkLivingston}.
\begin{proposition}
\label{prop:HeraldKirkLivingston}
Let $n$ be a positive integer.
There is a canonical chain isomorphism of left~$\L{\C}$-modules between $C_*(M_n;\mathbb{C}[t^{\pm 1}]_{\alpha \times \chi})$ and $C_*(M_K;\mathbb{C}[t^{\pm 1}]^n_{\alpha(n,\chi)})$.
In particular, one obtains a canonical isomorphism between the following homology left $\L{\C}$-modules:
\begin{equation}
\label{eq:IsomorphismHKL}
H_1(M_n;\mathbb{C}[t^{\pm 1}]_{\alpha \times \chi}) \cong H_1(M_K;\mathbb{C}[t^{\pm 1}]^n_{\alpha(n,\chi)}).
\end{equation}
\end{proposition}
The proof of Proposition~\ref{prop:HeraldKirkLivingston} relies on the relation between the representation $\gamma_K(n,\chi)$ of~\eqref{eq:Matrix} and the representation $\rho_K(n,\chi)$ of~\eqref{eq:Reprho(n,chi)}. Since we shall use this relation in Subsection~\ref{sub:SatelliteMetabelian}, we give some further details.
\begin{remark}
\label{rem:InductionFunctor}
Let $H$ be a subgroup of a group $G$ and let $\rho$ be a representation of $H$.
It is known that there is an \emph{induced representation} $\operatorname{ind}_H^G\rho$ of $G$, we refer to \cite{HeraldKirkLivingston} for a construction and a discussion of its properties.
Taking $G=\mathbb{Z} \ltimes H_1(L_n(K);\mathbb{Z})$ and~$H=n\mathbb{Z} \times H_1(L_n(K);\mathbb{Z})$, the proof of Proposition~\ref{prop:HeraldKirkLivingston} relies heavily on the isomorphism
$$ \gamma_K(n,\chi)\cong\operatorname{ind}_H^G \rho_K(n,\chi). $$
\end{remark}
As the isomorphism described in~(\ref{eq:IsomorphismHKL}) arises from a canonical chain isomorphism (essentially Shapiro's lemma), we obtain the following result which appears to be implicit in~\cite{MillerPowell}:
\begin{corollary}
\label{cor:BlanchfieldDownstairs}
The isomorphism described in~(\ref{eq:IsomorphismHKL}) gives rise to an isometry between the twisted Blanchfield pairings $\Bl_{\alpha \times \chi}(M_n)$ and $ \Bl_{\alpha(n,\chi)}(K)$.
\end{corollary}
Summarizing the content of this section, we can choose to work alternatively with $\Bl_{\alpha(n,\chi)}(K)$ or with $\Bl_{\alpha \times \chi}(M_n)$.
The former is used in the paper of Miller-Powell~\cite{MillerPowell} and will be used in Subsections~\ref{sub:MillerPowell}~\ref{sub:SatelliteMetabelian} and Section~\ref{sec:TorusKnots}.
The latter is closer to the definition of the Casson-Gordon $\tau$ invariant with which we shall work in Subsections~\ref{sub:CassonGordon} and~\ref{sec:Proof}.
\subsection{A sliceness obstruction due to Miller and Powell}
\label{sub:MillerPowell}
The aim of this subsection is to state an obstruction to sliceness which is due to Miller and Powell~\cite{MillerPowell}.
\medbreak
Given a integer $n$, we use $\lambda_n$ to denote the non-singular $\mathbb{Q}/\mathbb{Z}$-valued linking form on the finite abelian group $H_1(L_n(K);\mathbb{Z})$.
Note that $H_1(L_n(K);\mathbb{Z})$ is also a $\mathbb{Z}[\mathbb{Z}_n]$-module.
A \emph{metabolizer} of~$\lambda_n$ is a $\mathbb{Z}[\mathbb{Z}_n]$-submodule $P$ of $H_1(L_n(K);\mathbb{Z})$ such that $P=P^\perp$, where the orthogonal complement is defined as~$P^\perp=\{y\in H_1(L_n(K);\mathbb{Z})\ | \ \forall x\in P,\ \lambda_n(x,y)=~0\}$.
Finally, given a prime $q$, integers $b \geq a$ and a character $\chi \colon H_1(L_n(K);\mathbb{Z}) \to~\mathbb{Z}_{q^a}$, we use $\chi_b$ to denote the composition of $\chi$ with the natural inclusion $\mathbb{Z}_{q^a} \to \mathbb{Z}_{q^b}$.
The following obstruction to sliceness is due to Miller and Powell~\cite[Theorem 6.9]{MillerPowell}.
\begin{theorem}
\label{thm:MillerPowell}
Let $K$ be a slice knot.
Then, for any prime power $n$, there exists a metabolizer~$P$ of~$\lambda_n$ such that for any prime power $q^a$, and any nontrivial character $\chi \colon H_1(L_n(K);\mathbb{Z}) \to \mathbb{Z}_{q^a}$ vanishing on $P$, we have some $b \geq a$ such that the metabelian Blanchfield pairing $Bl_{\alpha(n,\chi_b)}(K)$ is metabolic.
\end{theorem}
Miller and Powell actually prove Theorem~\ref{thm:MillerPowell} under the weaker assumption that $K$ is 2-solvable.
We refer to Cochran-Orr-Teichner's landmark paper~\cite{CochranOrrTeichner} for the definition of $n$-solvability but do not delve deeper into this issue.
Combining Theorem~\ref{thm:MillerPowell} with the earlier sections, we obtain the following result.
\begin{theorem}
\label{thm:MillerPowellSignature}
Let $K$ be a slice knot.
Then, for any prime power $n$, there exists a metabolizer~$P$ of~$\lambda_n$ such that for any prime power $q^a$, and any nontrivial character $\chi \colon H_1(L_n(K);\mathbb{Z}) \to \mathbb{Z}_{q^a}$ vanishing on $P$, we have some $b \geq a$ such that the signature function $\sigma^{av}_{K,\alpha(n,\chi_b)}$ is zero.
\end{theorem}
\begin{proof}
Theorem~\ref{thm:MillerPowell} implies that $\Bl_{\alpha(k,\chi_b)}(K)$ is metabolic.
Consequently, Corollary~\ref{cor:metabolic_signature} implies that the signature function of $\Bl_{\alpha(n,\chi_b)}(K)$ vanishes.
\end{proof}
The obstruction of Theorem~\ref{thm:MillerPowell} will be used in Subsection~\ref{sub:HeddenKirkLivingston} below to prove that a certain algebraic knot is not slice, thus recovering a result of Hedden-Kirk-Livingston~\cite{HeddenKirkLivingston}.
\subsection{The Casson-Gordon $\tau$-invariant}
\label{sub:CassonGordon}
Given an oriented knot $K$ and a prime power order character~$\chi \colon H_1(L_n(K);\mathbb{Z}) \to \mathbb{Z}_m$, we recall Casson-Gordon's construction of the Witt class~$\tau(K,\chi)$~\cite{CassonGordon2}.
\medbreak
Using the isomorphism $ H_1(M_n;\mathbb{Z}) \cong H_1(L_n(K);\mathbb{Z}) \oplus \mathbb{Z}$, the character $\chi$ induces a character on $H_1(M_n;\mathbb{Z})$ for which we use the same notation. Recalling that the projection $p \colon M_n \to M_K$ gives rise to a surjection $\alpha \colon \pi_1(M_n) \to n\mathbb{Z} \cong \mathbb{Z}$, we obtain a homomorphism
\begin{equation}
\label{eq:CGHomom}
\alpha \times \chi \colon \pi_1(M_n) \to \mathbb{Z} \times \mathbb{Z}_m.
\end{equation}
Since the bordism group $\Omega_3(\mathbb{Z} \times \mathbb{Z}_m)$ is finite, there is an integer $r$ such that $r$ copies of $(M_n,\alpha \times \chi)$ bound~$(V_n,\psi)$ for some $4$-manifold $V_n$ and for some homomorphism $\psi \colon \pi_1(V_n) \to \mathbb{Z} \times \mathbb{Z}_m$.
Consider the map $\mathbb{C}[\mathbb{Z} \times \mathbb{Z}_m] \to \mathbb{C}[t^{\pm 1}]$ which sends the generator of $\mathbb{Z}_m$ to $\xi_{m} = e^{2\pi i/m}$ and the generator of~$\mathbb{Z}$ to $t$.
Endow $\mathbb{C}(t)$ with the right $\mathbb{Z}[\pi_1(V_n)]$-module structure which arises from the composition
\begin{equation}
\label{eq:ModuleStructure}
\mathbb{Z}[\pi_1(V_n)] \xrightarrow{\psi} \mathbb{C}[\mathbb{Z} \times \mathbb{Z}_m] \to \L{\C} \to \mathbb{C}(t).
\end{equation}
We will denote by \(\mathbb{C}(t)_{\alpha \times \chi}\)
the module \(\mathbb{C}(t)\) equipped with the action of the group ring $\mathbb{Z}[\pi_1(V_n)]$
as described above.
Observe that~(\ref{eq:ModuleStructure}) also provides a right $\mathbb{Z}[\pi_1(V_n)]$-module structure on $\L{\C}$ (we write $\L{\C}_{\alpha \times \chi}$) which restricts to the one described in~(\ref{eq:ActionForMn}) on each boundary component of~$V_n$.
We will denote the related \(\mathbb{C}[\mathbb{Z}\times\mathbb{Z}_{m}]\)-module by~\(\L{\C}_{\mathbb{Z} \times \mathbb{Z}_m}\).
Next, consider the $\mathbb{C}(t)$-valued intersection form~$\lambda_{\mathbb{C}(t)_{\alpha \times \chi},V_n}$ on $H_2(V_n;\mathbb{C}(t)_{\alpha \times \chi})$.
Casson and Gordon show that if $\chi$ is a character of prime power order, then this Hermitian form is non-singular~\cite[Corollary following Lemma 4]{CassonGordon2}.
We record this result for later reference.
\begin{lemma}
\label{lem:CassonGordon}
Let $n$ be an integer and let $m$ be a prime power.
If $\chi$ is a nontrivial character, then~$H_*(M_n;\mathbb{C}(t)_{\alpha \times \chi})=0$ and the intersection pairing~$\lambda_{\mathbb{C}(t)_{\alpha \times \chi},V_n}$ is non-singular.
\end{lemma}
Since $\lambda_{\mathbb{C}(t)_{\alpha \times \chi},V_n}$ is non-singular, it gives rise to a Witt class $[\lambda_{\mathbb{C}(t)_{\alpha \times \chi},V_n}]$ in $W(\mathbb{C}(t))$.
On the other hand, the standard intersection pairing $\lambda_{\mathbb{Q},V_n}$ may very well be singular.
Consequently, we use~$\lambda_{\mathbb{Q},V_n}^{\text{nonsing}}$ to denote the form obtained by moding out the radical.
The Witt class of $\lambda_{\mathbb{Q},V_n}^{\text{nonsing}}$ provides an element in~$W(\mathbb{Q})$.
Since the inclusion map $\mathbb{Q} \to \mathbb{C}(t)$ induces a group homomorphism $i \colon W(\mathbb{Q}) \to W(\mathbb{C}(t))$, we therefore obtain an element $i([\lambda_{\mathbb{Q},V_n}^{\text{nonsing}}])$ in $W(\mathbb{C}(t))$.
\begin{definition}
\label{def:CassonGordonInvariant}
Let $K$ be an oriented knot, let $n$ be an integer, let $m$ be a prime power and let $\chi \colon H_1(L_n(K);\mathbb{Z}) \to \mathbb{Z}_m$ be a non-trivial character.
The \emph{Casson-Gordon $\tau$-invariant} is the Witt~class
$$\tau(K,\chi):=([\lambda_{\mathbb{C}(t)_{\alpha \times \chi},V_n}]-i([\lambda_{\mathbb{Q},V_n}^{\text{nonsing}}]) \otimes \frac{1}{r} \in W(\mathbb{C}(t))\otimes \mathbb{Q}.$$
\end{definition}
Casson and Gordon show that $\tau(K,\chi)$ is independent both of the choice of the 4-manifold $V_n$ and of the extension $\psi$ of $\alpha \times \chi$ to $\pi_1(V_n)$~\cite{CassonGordon2}.
Furthermore, as we shall recall later on, $\tau(K,\chi)$ provides an obstruction to sliceness.
\begin{remark}
\label{rem:Cyclotomic}
Casson and Gordon define their Witt class as an element of $W(\mathbb{Q}(\xi_m)(t))$~\cite{CassonGordon2}.
Since the natural map $\mathbb{Q}(\xi_m)(t) \to \mathbb{C}(t)$ induces a map on Witt groups, we also obtain a Witt class in~$W(\mathbb{C}(t))$.
Our reason for working with~$W(\mathbb{C}(t))$ instead of $W(\mathbb{Q}(\xi_m)(t))$ is the following: the former Witt group is much simpler than the latter and, in particular, it is more amenable to the machinery we developed in Sections~\ref{sec:Witt} and~\ref{sec:Signatures}.
\end{remark}
\subsection{Casson-Gordon signatures and averaged Blanchfield signatures}
\label{sec:Proof}
In this section, we investigate the relation between the Casson-Gordon invariant $\tau(K,\chi)$ and the twisted Blanchfield pairing $\Bl_{\alpha(n,\chi)}(K)$.
In order to make both these invariants more tractable, we use the averaged signatures from Section~\ref{sec:Signatures}.
\medbreak
Let $A(t)$ be a matrix over $\mathbb{C}(t)$.
The function $\operatorname{sign}_\omega(A(t)):=\operatorname{sign}(A(\omega))$ is a step-function with finitely many discontinuities, and at each discontinuity $\omega$, we can take the average of the one-sided limits in order to obtain a rational number $\operatorname{sign}_\omega^{\text{av}}(A(t))$.
As explained in~\cite[discussion preceding Theorem 3]{CassonGordon2}, (see also Proposition~\ref{prop:avsig}), one now obtains a well defined homomorphism $\operatorname{sign}^{\text{av}}_\omega \colon W(\mathbb{C}(t)) \to \mathbb{Q}$ by setting
$$\operatorname{sign}_\omega^{\text{av}}([A(t)]):=\text{sign}^{\text{av}}_\omega(A(t)).$$
This map also gives rise to a well defined map on $W(\mathbb{C}(t)) \otimes \mathbb{Q}$.
We now apply this discussion to extract signatures from the Casson-Gordon invariant $\tau(K,\chi)$ which is an element of $W(\mathbb{C}(t)) \otimes \mathbb{Q}$.
\begin{definition}
\label{def:CassonGordonSignature}
The \emph{Casson-Gordon signature} associated to $\omega \in S^1$ is $\operatorname{sign}_\omega^{\text{av}}(\tau(K,\chi))$.
\end{definition}
Before returning to Blanchfield pairings, let us make Definition~\ref{def:CassonGordonSignature} somewhat more concrete.
\begin{remark}
\label{rem:CassonGordonSignature}
Suppose $V_n$ is a $4$-manifold whose boundary consists of a disjoint union of $r$ copies of~$M_n$ and over which $\alpha \times \chi$ extends.
In other words, we have
\[\tau(K,\chi)= ([\lambda_{\mathbb{C}(t)_{\alpha\times \chi},V_n}]-i_*([\lambda_{\mathbb{Q},V_n}^{\text{nonsing}}])) \otimes \frac{1}{r}.\]
Since $\operatorname{sign}_\omega^{\text{av}}$ is a homomorphism, we obtain
\[\operatorname{sign}_\omega^{\text{av}}(\tau(K,\chi))=\frac{1}{r}(\operatorname{sign}_\omega^{\text{av}}([\lambda_{\mathbb{C}(t)_{\alpha\times \chi},V_n}])-\operatorname{sign}_\omega^{\text{av}}([\lambda_{\mathbb{Q},V_n}^{\text{nonsing}}])).\]
Clearly the latter term is simply equal to $\operatorname{sign}(V_n)$, the (untwisted) signature of $V_n$.
Consequently, if $A(t)$ represents $\lambda_{\mathbb{C}(t)_{\alpha\times \chi},V_n}$, then we deduce that
$$\operatorname{sign}_\omega^{\text{av}}(\tau(K,\chi))=\frac{1}{r}(\operatorname{sign}^{\text{av}}_\omega(A(t))-\operatorname{sign}(V_n)).$$
Note that $A(1)$ does \emph{not} represent the standard intersection form $\lambda_{\mathbb{Q},V_n}$.
It does however represent the twisted intersection form $\lambda_{\mathbb{C}_{\xi_{m}},V_n}$ which arises from the coefficient system $\pi_1(V_n) \to \mathbb{Z}_m \to \mathbb{C}$, where the latter map sends the generator of $\mathbb{Z}_m$ to $\xi_m=\text{exp}(\frac{2\pi i}{m})$.
\end{remark}
Returning to Blanchfield pairings, recall from Definition~\ref{def:BlanchfieldMetabelian} that $\Bl_{\alpha(n,\chi)}(K)$ denotes the twisted Blanchfield pairing associated to the representation $\alpha(n,\chi) \colon \pi_1(M_K) \to GL_n(\L{\C})$.
Recall furthermore that $\sigma_{K,\alpha(n,\chi)} \colon S^1 \to \mathbb{Z}$ denotes the associated twisted signature~function.
The following theorem relates our Blanchfield signatures to the Casson-Gordon signatures.
\begin{theorem}
\label{thm:BlanchfieldCG}
Let $K$ be an oriented knot and $n$ be an integer.
Let $L_n(K)$ be the $n$-fold cyclic cover of $S^3$ branched along $K$ and let $M_n$ be the $n$-fold cyclic cover of $M_K$.
If $\chi \colon H_1(L_n(K);\mathbb{Z}) \to~\mathbb{Z}_m$ is a nontrivial character of prime power order $m$, then the following statements hold:
\begin{itemize}
\item[(a)] The twisted Blanchfield pairing $\Bl_{\alpha(n,\chi)}(K)$ is representable.
\item[(b)] For each $\omega$ in $S^1$, we have the equality
$$ -\sigma^{\text{av}}_{K,\alpha(n,\chi)}(\omega)=\operatorname{sign}_\omega^{\text{av}}(\tau(K,\chi))-\operatorname{sign}_1^{\text{av}}(\tau(K,\chi)).$$
\end{itemize}
\end{theorem}
Before moving towards the proof, we start with a remark and a corollary to Theorem~\ref{thm:BlanchfieldCG}.
\begin{remark}
\label{rem:RemarkOnCGTheorem}
Although we stated Theorem~\ref{thm:BlanchfieldCG} using the metabelian Blanchfield pairing $\Bl_{\alpha(n,\chi)}(K)$, the proof will use the Blanchfield pairing of the cover $\operatorname{Bl}_{\alpha \times \chi}(M_n)$.
Indeed since Corollary~\ref{cor:BlanchfieldDownstairs} states that $\operatorname{Bl}_{\alpha(n,\chi)}(K)$ and $\operatorname{Bl}_{\alpha \times \chi}(M_n)$ are isometric, one linking form is representable if and only if the other one is and, for every $\omega \in S^1$, the second item of Theorem~\ref{thm:BlanchfieldCG} could be have been written using the signature function of $\operatorname{Bl}_{\alpha \times \chi}(M_n)$, since the following equality holds:
\begin{equation}
\label{eq:SignatureEquality}
\sigma^{\text{av}}_{K,\alpha(n,\chi)}(\omega)= \operatorname{sign}_\omega^{\text{av}}(\Bl_{\alpha \times \chi}(M_n)).
\end{equation}
\end{remark}
Since $\tau(K,\chi)$ is known to provide an obstruction to sliceness~\cite{CassonGordon2}, a similar conclusion holds for~$\sigma_{K,\alpha(n,\chi)}$, thus yielding a second proof of (a variation on) Theorem~\ref{thm:MillerPowellSignature}.
Indeed, recalling that~$\lambda_n$ denotes the $\mathbb{Q}/\mathbb{Z}$-valued linking form on $H_1(L_n(K);\mathbb{Z})$, the following result is an immediate corollary of Theorem~\ref{thm:BlanchfieldCG} and \cite[Theorem 2]{CassonGordon2}.
\begin{corollary}
\label{cor:CG}
If $K$ is slice, then there is a subgroup $G$ of $H_1(L_n(K);\mathbb{Z})$ such that $\lambda_n$ vanishes on $G$ and, for every character $\chi$ vanishing on $G$, the signature function $ \sigma^{\text{av}}_{K,\alpha(n,\chi)}(\omega)$ is zero.
\end{corollary}
We now move towards the proof of Theorem~\ref{thm:BlanchfieldCG}.
First of all, as we mentioned in Remark~\ref{rem:RemarkOnCGTheorem}, we can work with the pairing $\Bl_{\alpha \times \chi}(M_n)$ instead of $\Bl_{\alpha(n,\chi)(K)}$.
Next, as in Subsection~\ref{sub:CassonGordon}, we choose a $4$-manifold~$V_n$ such that the boundary of $V_n$ consists of $r$ copies of $M_n$ and such that the representation extends.
Without loss of generality, we can assume that $\pi_1(V_n)=\mathbb{Z}_m \times \mathbb{Z}$: indeed since~$\mathbb{Z}_m \times \mathbb{Z}$ is finitely normally presented, one can perform finitely many surgeries to obtain the desired fundamental group, while leaving the boundary fixed.
The following lemma collects some algebraic statements which we shall need later on.
\begin{lemma}
\label{lem:ev}
The following statements hold:
\begin{itemize}
\item[(a)] The $\L{\C}$-module $H_1(V_n;\L{\C}_{\alpha \times \chi})$ vanishes.
\item[(b)] The map $\ev \circ \kappa \colon H^2(V_n;\L{\C}_{\alpha \times \chi}) \rightarrow \makeithash{\Hom_{\L{\C}}(H_2(V_n;\L{\C}_{\alpha \times \chi}),\L{\C})}$ is an isomorphism.
\item[(c)] The $\L{\C}$-module $H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi})$ is free.
\item[(d)] The inclusion induced map $ H^2(V_n,\partial V_n;\mathbb{C}(t)_{\alpha\times \chi}) \to H^2(V_n;\mathbb{C}(t)_{\alpha\times \chi})$ is an isomorphism.
\end{itemize}
\end{lemma}
\begin{proof}
We start by proving (a).
Set $G:=\mathbb{Z}_m \times \mathbb{Z}$ for brevity. Recalling the notation introduced below~\eqref{eq:ModuleStructure}, we claim that \(\L{\C}_{G}\) is a projective \(\mathbb{C}[G]\)-module.
To see this, first note that \(\mathbb{C}[G] = \mathbb{C}[t^{\pm 1}] \otimes_{\mathbb{C}} \mathbb{C}[\mathbb{Z}_m]\) and \(\mathbb{C}[\mathbb{Z}_m] = \bigoplus_{j = 0}^{m-1}~\mathbb{C}_{\xi_m^j}\), where~\(\mathbb{C}_{\xi_m^j}\) denotes the irreducible complex representation of \(\mathbb{Z}_m\) with the action of the cyclic group given by multiplication by the root of unity \(\xi_m^j\), for \(\xi_m = \exp\left(\frac{2 \pi i}{m}\right)\).
Therefore, we have \(\mathbb{C}[G] = \bigoplus_{j=0}^{m-1} \mathbb{C}[t^{\pm 1}] \otimes_{\mathbb{C}}~\mathbb{C}_{\xi_m^j}\) and \(\L{\C}_{G} = \mathbb{C}[t^{\pm 1}] \otimes_{\mathbb{C}}~\mathbb{C}_{\xi_m}\), concluding the proof of the claim.
The claim now implies that $H_1(V_n; \L{\C}_{\alpha \times \chi})$ is a summand of~$H_1(V_n; \mathbb{C}[G])$.
Since $\pi_1(V_n)=G$, the corresponding $G$-cover of $V_n$ is simply-connected and thus~$H_1(V_n;\mathbb{C}[G])$ vanishes.
Consequently \(H_1(V_n; \L{\C}_{\alpha \times \chi})\) also vanishes since it is a direct summand of \(H_1(V_n; \mathbb{C}[G])\).
This concludes the proof of (a).
We now prove (b) according to which the evaluation map
\[\ev \circ \kappa \colon H^2(V_n;\L{\C}_{\alpha \times \chi}) \rightarrow \makeithash{\Hom_{\L{\C}}(H_2(V_n;\L{\C}_{\alpha \times \chi}),\L{\C})}\]
is an isomorphism.
Since $H_1(V_n;\L{\C}_{\alpha \times \chi})=0$ by the first statement and \(H_0(V_n;\L{\C}_{\alpha \times \chi})=~0\) (recall Lemma~\ref{lem:H0CG}), this follows immediately from the universal coefficient theorem.
We move on to (c) which asserts that $H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi})$ is a free $\L{\C}$-module.
Using Poincar\'e duality and the second statement, we deduce that $H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi})$ is isomorphic to $\makeithash{\Hom_{\L{\C}}(H_2(V_n;\L{\C}_{\alpha \times \chi}),\L{\C})}$ which is free, see \cite[Lemma 2.1]{BorodzikFriedl0}.
Finally, we deal with (d), that is we show that the inclusion induced map~$H^2(V_n,\partial V_n;\mathbb{C}(t)_{\alpha\times \chi}) \to H^2(V_n;\mathbb{C}(t)_{\alpha\times \chi})$ is an isomorphism.
Since $H_*(M_n;\mathbb{C}(t)_{\alpha\times \chi})$ vanishes by Lemma~\ref{lem:CassonGordon} and $\partial V_n$ consists of~$r$ disjoint copies of $M_n$, we deduce that~$H_*(\partial V_n;\mathbb{C}(t)_{\alpha\times \chi})$ also vanishes.
The long exact sequence of the pair~$(V_n,\partial V_n)$ implies that $H_2(V_n;\mathbb{C}(t)_{\alpha\times \chi}) \rightarrow H_2(V_n,\partial V_n;\mathbb{C}(t)_{\alpha\times \chi})$ is an isomorphism and the result follows by Poincar\'e duality.
This concludes the proof of (d) and thus the proof of the lemma.
\end{proof}
Since $\L{\C}$ is a PID, we can decompose any $\L{\C}$-module $H$ as the direct sum of its free part~$LH$ and its torsion part $TH$.
In particular, it follows that
\[\makeithash{\Hom_{\L{\C}}(H_2(V_n;\L{\C}_{\alpha \times \chi}),\L{\C})}=\makeithash{\Hom_{\L{\C}}(LH_2(V_n;\L{\C}_{\alpha \times \chi}),\L{\C})}.\]
Consequently, point (b) of Lemma~\ref{lem:ev} provides the following well defined isomorphism:
\begin{equation}
\label{eq:EvaluationFree}
ev \circ \kappa \colon H^2(V_n;\L{\C}_{\alpha \times \chi}) \rightarrow \makeithash{\Hom_{\L{\C}}(LH_2(V_n;\L{\C}_{\alpha \times \chi}),\L{\C})}.
\end{equation}
Composing the inclusion $LH_2(V_n;\L{\C}_{\alpha \times \chi}) \to H_2(V_n;\L{\C}_{\alpha \times \chi})$ with the inclusion induced map $H_2(V_n;\L{\C}_{\alpha \times \chi}) \to H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi})$ gives rise to a well defined map
\[LH_2(V_n;\L{\C}_{\alpha \times \chi}) \to H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi}).\]
Combining these remarks, there is a well defined intersection form
$$\lambda_{\L{\C}_{\alpha \times \chi},V_n} \colon LH_2(V_n;\L{\C}_{\alpha \times \chi}) \times LH_2(V_n;\L{\C}_{\alpha \times \chi}) \to \L{\C}.$$
Next, we describe the Blanchfield $\Bl_{\alpha \times \chi}(\partial V_n)$ on $H_1(\partial V_n;\L{\C}_{\alpha \times \chi})$.
As $\partial V_n$ consists of $r$ copies of~$M_n$, we deduce that $H_1(\partial V_n;\L{\C}_{\alpha \times \chi})$ is isomorphic to the direct sum of~$r$ copies of $H_1(M_n;\L{\C}_{\alpha \times \chi})$.
Since the latter $\L{\C}$-module is torsion, so is the former.
We deduce that $ \Bl_{\alpha\times \chi}(\partial V_n)$ decomposes as the direct sum $\Bl_{\alpha \times \chi}(M_n) \oplus \cdots \oplus \Bl_{\alpha \times \chi}(M_n)$, where there are $r$ direct summands.
The proof of the next result is postponed until the end of the proof of Theorem~\ref{thm:BlanchfieldCG}.
\begin{proposition}
\label{prop:Boundary}
Any matrix $A(t)$ representing $\lambda_{\L{\C}_{\alpha \times \chi},V_n}$ also represents~$\lambda_{\mathbb{C}(t)_{\alpha\times \chi},V_n}$ and~$-\Bl_{\alpha \times \chi}(\partial V_n)$.
\end{proposition}
\begin{proof}[Proof of Theorem~\ref{thm:BlanchfieldCG} assuming Proposition~\ref{prop:Boundary}]
We start by proving the second statement, namely the equality of signatures.
Let $A(t)$ be a matrix representing $\lambda_{\L{\C}_{\alpha \times \chi},V_n}$.
Since $A(t)$ represents the twisted intersection form $\lambda_{\mathbb{C}(t)_{\alpha\times \chi},V_n}$, the definition of the Casson-Gordon signature and Remark~\ref{rem:CassonGordonSignature} imply that
\begin{equation}
\label{eq:CassonGordonForProof}
\operatorname{sign}_\omega^{\text{av}}(\tau(K,\chi))=\frac{1}{r}(\operatorname{sign}^{\text{av}}_\omega(A(t))-\operatorname{sign}(V_n)).
\end{equation}
Since Proposition~\ref{prop:Boundary} tells us that $A(t)$ also represents $-\Bl_{\alpha \times \chi}(\partial V_n)$, Corollary~\ref{cor:sigissig2} implies that
$-\operatorname{sign}_\omega^{\text{av}}(\Bl_{\alpha \times \chi}(\partial V_n))=\operatorname{sign}^{\text{av}}_\omega(A(t))-\operatorname{sign}^{\text{av}}_1(A(t))$.
On the other hand, since $\partial V_n$ consists of $r$ disjoint copies of $M_n$, we observed previously that $ \Bl_{\alpha \times \chi}(\partial V_n)=\Bl_{\alpha \times \chi}(M_n) \oplus \ldots \oplus \Bl_{\alpha \times \chi}(M_n)$ and consequently $\operatorname{sign}_\omega^{\text{av}}(\Bl_{\alpha \times \chi}(\partial V_n))=r \cdot \operatorname{sign}_\omega^{\text{av}}(\Bl_{\alpha \times \chi}(M_n))$.
Combining these two observations, we obtain the following equality:
\begin{equation}
\label{eq:BlanchfieldSignatureProof2}
-\operatorname{sign}^{\text{av}}_\omega(\Bl_{\alpha \times \chi}( M_n))=\frac{1}{r}(\operatorname{sign}^{\text{av}}_\omega(A(t))-\operatorname{sign}^{\text{av}}_1(A(t))).
\end{equation}
Finally, adding and subtracting a $\text{sign}(V_n)$ term in~(\ref{eq:BlanchfieldSignatureProof2}) and combining the result with~(\ref{eq:CassonGordonForProof}), we can conclude the proof of the first assertion:
\begin{align*}
-\operatorname{sign}^{\text{av}}_\omega(\Bl_{\alpha \times \chi}( M_n))
&=\frac{1}{r}\left(\operatorname{sign}^{\text{av}}_\omega(A(t))-\operatorname{sign}(V_n)\right)-\frac{1}{r}\left(\operatorname{sign}^{\text{av}}_1(A(t))-\operatorname{sign}(V_n) \right) \\
&=\operatorname{sign}_\omega^{\text{av}}(\tau(K,\chi))-\operatorname{sign}_1^{\text{av}}(\tau(K,\chi)).
\end{align*}
It remains to prove the first statement, namely that the twisted Blanchfield pairing $\Bl_{\alpha \times \chi}(M_n)$ is representable.
In the previous paragraph, we showed that $ \Bl_{\alpha \times \chi}(\partial V_n)=\Bl_{\alpha \times \chi}(M_n) \oplus \ldots \oplus \Bl_{\alpha \times \chi}(M_n)$.
Since we know from Proposition~\ref{prop:Boundary} that $\Bl_{\alpha \times \chi}(\partial V_n)$ is representable, $\Bl_{\alpha \times \chi}(M_n) $ is itself representable ($\Bl_{\alpha\times\chi}(\partial V_n)$ satisfies the signature condition of Corollary~\ref{cor:WittRepresentability} and thus so does~$\Bl_{\alpha\times\chi}(M_n)$)
and thus the theorem is proved.
\end{proof}
We now prove Proposition~\ref{prop:Boundary}; the strategy follows very closely~\cite[Section 5.2]{ConwayFriedlToffoli} which itself is based on the proof of~\cite[Theorem 2.6]{BorodzikFriedl}.
\begin{proof}[Proof of Proposition~\ref{prop:Boundary}]
Consider the following diagram in which all homomorphisms are understood to be homomorphisms of $\L{\C}$-modules:
\begin{equation}
\label{eq:LargeDiagram}
\xymatrix@R0.6cm@C-1cm{
LH_2 \ar[rr] \ar[dd]^{\Theta} && H\ar[r]^-\partial \ar[d]^{PD} & H_1(\partial V_n;\Lambda_{\alpha \times \chi})\ar[d]^{PD}\ar[rr] \ar@/^/[dddrr]^{\Omega} &&0\\
&& H^2(V_n;\Lambda_{\alpha \times \chi}) \ar[r]\ar[d]\ar[lld]^{\ev \circ \kappa}& H^2(\partial V_n;\Lambda_{\alpha \times \chi})\ar[d]^{BS^{-1}} && \\
\makeithash{\Hom_\Lambda(LH_2,\Lambda)} \ar[d] && H^2(V_n;\Omega_{\alpha \times \chi})\ar[lld]^{\ev \circ \kappa}\ar[d]^{\cong} & H^1(\partial V_n;\Omega_{\alpha \times \chi}/\Lambda_{\alpha \times \chi})\ar[rrd]^{\ev \circ \kappa} \ar[d]&& \\
\makeithash{ \Hom_\Lambda(LH_2,\Omega) } \ar[d]^{\cong }&& H^2(V_n,\partial V_n;\Omega_{\alpha \times \chi})\ar@<-2pt>`d[r] `[r] [r]\ar[lld]^{\ev \circ \kappa}\ \ & H^2(V_n,\partial V_n;\Omega_{\alpha \times \chi}/\Lambda_{\alpha \times \chi}) \ar[rrd]^{\ev \circ \kappa}&& \makeithash{ \Hom_\Lambda(H_1(\partial V_n;\Lambda_{\alpha \times \chi}),\Omega/\Lambda) } \ar[d]_{\partial^{\#}} \\
\makeithash{\Hom_\Lambda(H,\Omega)} \ar[rrrr]^{} &&&&& \makeithash{ \Hom_\Lambda(H,\Omega/\Lambda) }.
}
\end{equation}
To make the diagram more concise we have used the following shorthands:
\begin{itemize}
\item $\Lambda=\L{\C}$;
\item $LH_2=LH_2(V_n;\L{\C}_{\alpha \times \chi})$;
\item $H=H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi})$;
\item $\Omega=\mathbb{C}(t)$.
\end{itemize}
The top horizontal line of \eqref{eq:LargeDiagram} is exact thanks to the long exact sequence of the pair $(V_n,\partial V_n)$ together with Lemma~\ref{lem:ev}: exactness at the rightmost end is guaranteed by the first point of Lemma~\ref{lem:ev}, while exactness at the middle follows from the third point.
In more details, since~$H:=H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi})$ is free, the image of $H_2(V_n;\L{\C}_{\alpha \times \chi})$ in $H$ is equal to the image of~$LH_2(V_n;\L{\C}_{\alpha \times \chi})$ in $H$.
The top middle square commutes by duality.
All the squares involving evaluation maps clearly commute, while the upper left (resp. right) square commutes by definition of the intersection (resp. Blanchfield) pairing.
Finally, the middle rectangle \textit{anti}-commutes thanks to~\cite[Lemma~5.4]{ConwayFriedlToffoli}, see~\cite[Appendix A]{ConwayBlanchfield} for a proof.
Inspired by~(\ref{eq:LargeDiagram}), we define a pairing $\theta$ on $H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi})$ by the composition
\begin{align*}
H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi}) &\xrightarrow{\operatorname{PD}} H^2(V_n;\L{\C}_{\alpha \times \chi}) \rightarrow H^2(V_n;\mathbb{C}(t)_{\alpha\times \chi}) \cong H^2(V_n,\partial V_n;\mathbb{C}(t)_{\alpha\times \chi}) \\
&\rightarrow \makeithash{\Hom_{\L{\C}}(H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi}),\mathbb{C}(t))},
\end{align*}
where the third map is an isomorphism thanks to the fourth point of Lemma~\ref{lem:ev}.
The commutativity of the diagram in equation~\eqref{eq:LargeDiagram} immediately implies the commutativity of
\begin{equation}
\label{eq:DiagramThreeLines}
\xymatrix@C1.4cm@R0.5cm{ LH_2(V_n;\L{\C}_{\alpha \times \chi}) \times LH_2(V_n;\L{\C}_{\alpha \times \chi}) \ar[r]^{\hskip 1cm \ \ \ \ \ \ \ \ \ \ \ \ -\lambda_{\L{\C}_{\alpha \times \chi},V_n}}\ar[d]^{} & \L{\C}\ar[d] \\
H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi}) \times H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi}) \ar[r]^{\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ -\theta}\ar[d]^{\partial \times \partial} & \mathbb{C}(t) \ar[d] \\
H_1(\partial V_n;\L{\C}_{\alpha \times \chi}) \times H_1(\partial V_n;\L{\C}_{\alpha \times \chi}) \ar[r]^{\hskip 1cm \ \ \ \ \ \ \ \ \ \ \ \Bl_{\alpha \times \chi}(\partial V_n)}& \mathbb{C}(t)/\L{\C}.
}
\end{equation}
We now pick bases in order to obtain matrices for the intersection form.
Namely, choose any basis~$\mathcal{C}$ of~$LH_2(V_n;\L{\C}_{\alpha \times \chi})$ and endow $\makeithash{\Hom_{\L{\C}}(LH_2(V_n;\L{\C}_{\alpha \times \chi}),\L{\C})}$ with the corresponding dual basis $\mathcal{C}^*$.
Let $A(t)$ denote the matrix of the $\L{\C}$-intersection form $\lambda_{\L{\C}_{\alpha \times \chi},V_n}$ with respect to these bases.
Next, we use $\mathcal{C}$ and $\mathcal{C}^*$ to base $H_2(V_n;\mathbb{C}(t)_{\alpha\times \chi})$ and $\makeithash{\Hom_{\L{\C}}(H_2(V_n;\mathbb{C}(t)_{\alpha\times \chi}),\mathbb{C}(t))}$.
In other words, as claimed in the first part of Proposition~\ref{prop:Boundary}, there are bases with respect to which both~$\lambda_{\mathbb{C}(t)_{\alpha\times \chi},V_n}$ and $\lambda_{\L{\C}_{\alpha \times \chi},V_n}$ are represented by $A(t)$.
Therefore to conclude the proof it only remains to show that $A(t)$ also represents the Blanchfield pairing on $\partial V_n$.
With this aim in mind, we start by providing a basis for $H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi})$ (which is free by Lemma~\ref{lem:ev}(b)).
This will allow us to represent the inclusion induced map $LH_2(V_n;\L{\C}_{\alpha \times \chi}) \to H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi})$ by a matrix.
First, consider the following commutative diagram of $\L{\C}$-homomorphisms:
\begin{equation}
\label{eq:DiagramForBasis}
\xymatrix@R0.5cm@C0.9cm{
LH_2(V_n;\L{\C}_{\alpha \times \chi})\ar[r]\ar[d]^-{PD}_{\cong} & H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi})\ar[d]^-{PD}_{\cong} \\
LH^2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi})\ar[r] & H^2(V_n;\L{\C}_{\alpha \times \chi})\ar[d]^-{\ev \circ \kappa}_{\cong}\\
&\makeithash{\Hom_{\L{\C}}(LH_2(V_n;\L{\C}_{\alpha \times \chi}),\L{\C})}.}
\end{equation}
Here the bottom-right map is an isomorphism thanks to the second point of Lemma~\ref{lem:ev}.
We now equip $H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi})$ with the basis induced from $\mathcal{C}^*$ and two isomorphisms in the right column of diagram in~(\ref{eq:DiagramForBasis}).
Arguing as in~\cite[Claim in Section 5.3]{ConwayFriedlToffoli}, elementary linear algebra shows that with respect to these bases, the map $LH_2(V_n;\L{\C}_{\alpha \times \chi}) \to H_2(V_n,\partial V_n;\L{\C}_{\alpha \times \chi})$ is represented by the matrix $A(t)^T$.
Rewriting the commutative diagram of equation (\ref{eq:DiagramThreeLines}) in terms of these bases, we obtain
\[ \xymatrix@C2.3cm@R0.5cm{\L{\C}^n \times \L{\C}^n \ar[r]^{(a,b) \mapsto -a^T A(t) \makeithash{b}}\ar[d]_{(a,b) \mapsto (A(t)^Ta,A(t)^Tb)} & \L{\C} \ar[d] \\
\L{\C}^n \times \L{\C}^n\ar[d]\ar[r]^{(a,b) \mapsto -a^T A(t)^{-1} \makeithash{b}} & \mathbb{C}(t) \ar[d] \\
H_1(\partial V_n;\L{\C}_{\alpha \times \chi}) \times H_1(\partial V_n;\L{\C}_{\alpha \times \chi})\ar[r]^{\hskip 1cm \ \ \ \ \ \ \ \ \ \ \ \Bl_{\alpha \times \chi}(\partial V_n)} & \mathbb{C}(t)/\L{\C}.
} \]
Here the middle horizontal map is determined by the top horizontal map, the vertical maps and the commutativity.
Since $H_1(V_n;\L{\C}_{\alpha \times \chi})$ vanishes by the first point of Lemma~\ref{lem:ev}, we deduce that the Blanchfield pairing on~$H_1(\partial V_n; \L{\C}_{\alpha \times \chi})$ is represented by $-A(t)$.
This concludes the proof of Proposition~\ref{prop:Boundary} and thus the proof of Theorem~\ref{thm:BlanchfieldCG}.
\end{proof}
\subsection{Satellite formulas for metabelian Blanchfield forms}
\label{sub:SatelliteMetabelian}
Given two knots $K,P$ and an unknotted curve $\eta$ in the complement of~$P$, we use $P(K,\eta)$ to denote the resulting satellite knot.
As described in Subsection~\ref{sub:Metabelian}, for a character
$H_1(L_n(P(K,\eta));\mathbb{Z}) \to \mathbb{Z}_{m}$, there is an associated metabelian representation $\alpha(n,\chi) \colon \pi_1(M_{P(K,\eta)}) \to GL_n(\mathbb{C}[t^{\pm 1}])$.
The goal of this subsection is to apply the satellite formula of Theorem~\ref{thm:CablingTheorem} to~$\alpha(n,\chi)$. On the level of signatures, the result is reminiscent of Litherland's description of the behavior of the Casson-Gordon invariants of satellite knots~\cite[Theorem 2]{Litherland}, but differs in the winding number zero case.
\medbreak
The composition $H_1(M_K;\mathbb{Z}) \cong H_1(X_K;\mathbb{Z}) \to H_1(M_{P(K,\eta)};\mathbb{Z}) \cong \mathbb{Z}$ is given by multiplication by~$w=\ell k(\eta,P)$.
Thus, the corresponding cover of $M_K$ is disconnected and has $|w|$ components (if~$w=0$, then the covering is trivial and so we temporarily disregard this case).
Using $t_Q$ (resp.~$t_K$) to denote the generator of the deck transformation group of the infinite cyclic cover of~$M_{P(K,\eta)}$ (resp. $M_K$), we note that~$t_Q^w=t_K$ and consider the following inclusion induced map
\begin{equation}
\label{eq:InclusionSatelliteAlexanderModule}
\iota_* \colon \bigoplus_{i=1}^{|w|} t_Q^{i-1} H_1(M_K;\mathbb{Z}[t_K^{\pm 1}]) \cong H_1(M_K;\mathbb{Z}[t_Q^{\pm 1}]) \to H_1(M_{P(K,\eta)};\mathbb{Z}[t_Q^{\pm 1}]).
\end{equation}
In order to obtain maps on the branched covers, we shall quotient both sides of~\eqref{eq:InclusionSatelliteAlexanderModule} by \(t_{Q}^{n}-1\). On the right hand side, the result is $H_1(L_n(P(K,\eta));\mathbb{Z})$ and so we focus on the left hand side.
Write $h:=\gcd(w,n)$
and observe that \(t_{K}^{n/h}=t_Q^{wn/h}=t_Q^{\operatorname{lcm}(n,w)}=1 \pmod{t_{Q}^{n}-1}\). Thus, the map~$i_*$ of~\eqref{eq:InclusionSatelliteAlexanderModule} descends to a map
$$ \bigoplus_{i=1}^{h} t_Q^{i-1} H_1(L_{n/h}(K);\mathbb{Z}) \to H_1(L_n(P(K,\eta));\mathbb{Z}),$$
where we are thinking of $H_{1}(L_{n/h}(K);\mathbb{Z})$ as $H_1(M_K;\mathbb{Z}[t_K^{\pm 1}])/(t_K^{n/h}-1)$. From now on, we fix a copy of $H_1(L_{n/h}(K);\mathbb{Z}) $ in this direct sum once and for all. Using this copy, we obtain a map
\begin{equation}
\label{eq:Defin}
\iota_n \colon H_1(L_{n/h}(K);\mathbb{Z}) \to H_1(L_n(P(K,\eta));\mathbb{Z}).
\end{equation}
Since $t_K^{n/h}-1=0$ mod $t_Q^n-1$, the character $\chi$ descends to a character on each $t_Q^{i-1} H_1(L_{n/h}(K);\mathbb{Z})$. Thus, for $i=1,\ldots,h$, the character \(\chi\) gives rise to the following characters:
\begin{align*}
\chi_{i} \colon H_{1}(L_{n/h}(K);\mathbb{Z}) &\to S^{1} \\
v &\mapsto \chi(t_{Q}^{i-1} \iota_n(v)).
\end{align*}
We use $\mu_Q$ (resp. $\mu_K$) to denote the meridian of~$M_{P(K,\eta)}$ (resp. $M_K$). Then, just as in Subsection~\ref{sub:Metabelian}, we consider the following composition of canonical projections:
$$q_{Q} \colon \pi_{1}(M_{P(K,\eta)})^{(1)} \to H_{1}(M_{P(K,\eta)};\mathbb{Z}[t_Q^{\pm1}]) \to H_{1}(L_{n}(P(K,\eta));\mathbb{Z}).$$
Observe that~$\mu_Q^{-w}\eta$ has trivial abelianization and therefore belongs to $\pi_{1}(M_{P(K,\eta)})^{(1)}$.
The main result of this section is the following.
\begin{theorem}\label{thm:metabelian-cabling-formula}
Let \(K,P\) be two knots in \(S^{3}\), let \(\eta\) be an unknotted curve in the complement of~\(P\) with meridian $\mu_\eta$, let \(w=\lk(\eta,P)\), let \(n>1\) and set \(h = \gcd(n,w)\).
\begin{itemize}
\item If \(w\neq 0\), then for any character \(\chi \colon H_1(L_n(P(K,\eta));\mathbb{Z}) \to \mathbb{Z}_{m}\) of prime power order, the metabelian representation~\(\alpha(n,\chi)\) is \(\eta\)-regular.
Moreover,
\begin{enumerate}
\item if \(w\) is divisible by \(n\), then there exists an isometry of linking forms
\[\Bl_{\alpha(n,\chi)}(P(K,\eta)) \cong \Bl_{\alpha(n,\chi_P)}(P) \oplus \bigoplus_{i=1}^{n} \Bl(K)(\xi_{m}^{\chi_{i}(q_Q(\mu_Q^{-w}\eta))}t^{w/n});\]
\item if \(w\) is not divisible by \(n\), then \(\Bl_{\alpha(n,\chi)}(P(K,\eta))\) is Witt equivalent to
\[\Bl_{\alpha(n,\chi_{P})} (P)\oplus \bigoplus_{i=1}^{h} \Bl_{\alpha(n/h,\chi_{i})}(K)(\xi_{m}^{\chi_{i}(q_Q(\mu_Q^{-w}\eta))}t^{w/h}).\]
\end{enumerate}
\item If $w=0$, then the representation \(\alpha(n,\chi)\) is
$\eta$-regular if and only if $\chi_i(q_Q(\mu_K^{-w}\eta)) \neq 0$ for each $i=1,\ldots,n$. In this case, there exists an isometry of linking forms
\[\Bl_{\alpha(n,\chi)}(P(K,\eta)) \cong \Bl_{\alpha(n,\chi_P)}(P).\]
\end{itemize}
\end{theorem}
On the level of averaged signatures, Theorem~\ref{thm:metabelian-cabling-formula} is reminiscent of Litherland's description of the behavior of the Casson-Gordon invariants of satellite knots~\cite[Theorem 2]{Litherland}.
\begin{remark}
\label{rem:ComparisonWithLitherland}
Taking averaged signatures in Theorem~\ref{thm:metabelian-cabling-formula} (and applying Proposition~\ref{prop:avsig}), we see that when the winding number $w$ is non-zero, our metabelian signatures behave in the same way as the Casson-Gordon signatures do~\cite[Theorem 2]{Litherland}. On the other hand, when $w=0$, the behaviors differ: namely we obtain $\sigma^{\text{av}}_{P(K,\eta),\alpha(n,\chi)}(\omega)=\sigma^{\text{av}}_{P,\alpha(n,\chi_P)}(\omega)$; the Levine-Tristram signatures of the companion knot do not contribute, contrarily do Litherland's formula. This can also be seen by combining Theorem~\ref{thm:BlanchfieldCG} with~\cite[Corollary~2]{Litherland}.
\end{remark}
Notice furthermore that Theorem~\ref{thm:metabelian-cabling-formula} takes a particularly simple form for connected sums.
Indeed, in this case, we have~$w=1$ (so~$h=1$) as well as $\eta=\mu_P$ so that we obtain the following corollary.
\begin{corollary}
\label{cor:MetabelianConnectedSum}
Let $K, P$ be two knots. If $\chi \colon H_1(L_n(K \# P);\mathbb{Z}) \to \mathbb{Z}_{m}$ is a character of prime power order, then \(\Bl_{\alpha(n,\chi)}(K \# P)\) is Witt equivalent to $\Bl_{\alpha(n,\chi_P)}(P) \oplus \Bl_{\alpha(n,\chi_K)}(K)$.
\end{corollary}
The remainder of this subsection is devoted to the proof of Theorem~\ref{thm:metabelian-cabling-formula}. We start by proving the $\eta$-regularity of $\alpha(n,\chi)$.
\begin{lemma}
\label{lemma:metabelian-eta-regularity}
Let $K,P$ be knots and let $\eta \subset S^3 \setminus P \subset M_P$ be a simple closed curve.
Choose a character \(\chi \colon H_1(L_n(P(K,\eta));\mathbb{Z}) \to \mathbb{Z}_{m} \) of finite order and let \(w = \lk(\eta,P)\).
\begin{enumerate}
\item If \(w \neq 0\), then \(\alpha(n,\chi)\) is \(\eta\)-regular.
\item For \(w=0\), \(\alpha(n,\chi)\) is \(\eta\)-regular if and only if $\chi_i(q_Q(\mu_K^{-w}\eta)) \neq 0$ for each $i=1,\ldots,n$.
\end{enumerate}
\end{lemma}
\begin{proof}
Notice that in $M_{P(K,\eta)}$, the meridian \(\mu_{\eta}\) of $\eta$ is identified with the zero-framed longitude of \(K\). In particular, since the longitude of~\(K\) is contained in \(\pi_{1}(S^{3} \setminus \mathcal{N}(K))^{(2)}\), it follows that~\(\mu_{\eta} \in \pi_{1}(M_{P(K,\eta)})^{(2)}\).
Therefore, since the representation $\alpha(n,\chi)$ is metabelian, we deduce that \(\alpha(n,\chi)(\mu_{\eta}) =~1\).
It remains to show that \(\det\left(\alpha(n,\chi)(\eta) - \operatorname{id}\right) \neq 0\). For \(a_{1},\ldots,a_{n}$ in~$\L{\C}\), define \(\operatorname{diag}(a_{1},\ldots,a_{n})\) to be the diagonal matrix with the $a_i$ on its main diagonal. The second assertion will quickly follow from the following observation:
\begin{claim*}
Set \(D = \operatorname{diag}(a_{1},a_{2},\ldots,a_{n})\) and~$A_n(t)= \left(\begin{smallmatrix}
0& 1 & \cdots &0 \\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 & \cdots & 1 \\
t & 0 & \cdots & 0
\end{smallmatrix}\right)$. For \(k\in \mathbb{Z}\) and \(h = \gcd(n,k)\), the following equality holds:
\[\det \left(A_{n}(t)^{k} D-\operatorname{id}\right) =
\begin{cases}
(a_1-1)\cdots(a_n-1)
\quad &\text{if } k = 0, \\
\pm \prod_{i=1}^{h}(1-t^{k/h} a_{i} a_{i+h} \cdots a_{i+k-h})
\quad &\text{if } k \neq 0.
\end{cases}
\]
\end{claim*}
\begin{proof}
If $k = 0$, then the claim is clear. We therefore assume that $k \neq 0$.
If \(h\neq 1\), then a computation using the definition of $A_n(t)$ implies that there is a permutation matrix \(B_{n,k}\)~such~that
\[B_{n,k} \left(A_{n}(t)^{k} D-\operatorname{id}\right) B_{n,k}^{-1} = \bigoplus_{i=1}^{h} \left(A_{n/h}(t)^{k/h} \operatorname{diag}(a_{i},a_{i+h},\ldots)-\operatorname{id}\right).\]
We can therefore restrict our attention to the case \(h=1\).
In this case, elementary row operations can be used to kill the entries below the diagonal.
Eventually, we end up with a matrix with zeros below the main diagonal and the following entries on the main diagonal:
\[\underbrace{-1,-1,\ldots,-1}_{n-1}, -1 + a_{1} a_{2} \cdots a_{n} t^{k}.\]
Thus, the determinant is equal to \((-1)^{n-1}(a_{1} a_{2} \cdots a_{n}t^{k}-1)\), concluding the proof of the claim.
\end{proof}
Use $\phi \colon \pi_1(M_{P(K,\eta)}) \to \mathbb{Z}$ to denote the abelianization map. We will apply the claim to~$k=\phi(\eta)=w$ and $a_i=\chi_i(q_Q(\mu_Q^{-w}\eta))$. If $w \neq 0$, then the claim implies that \(\det\left(\alpha(n,\chi)(\eta) - \operatorname{id}\right) \neq 0\), proving the first assertion. If $w=0$, then the claim implies that \(\det\left(\alpha(n,\chi)(\eta) - \operatorname{id}\right) \neq 0\) if and only if none of the $a_i$ is equal to $1$. This concludes the proof of the second assertion and therefore the proof of the lemma.
\end{proof}
We now suppose that $\alpha(n,\chi)$ is $\eta$-regular (this is automatic for $w \neq 0$). Thus $\alpha(n,\chi)$ restricts to representations $\alpha(n,\chi)_K$ and $\alpha(n,\chi)_P=\alpha(n,\chi_P)$ on the fundamental groups of $M_P$ and~$M_K$.
Before applying Theorem~\ref{thm:CablingTheorem} to obtain a decomposition of~$\operatorname{Bl}_{\alpha(n,\chi)}(M_{P(K,\eta)})$, we start by studying~$\alpha(n,\chi)_K$. The next remark leads to the main idea in this process.
\begin{remark}
\label{rem:MackeyMotivation}
Set $G:=\mathbb{Z} \ltimes H_1(L_n(P(K,\eta));\mathbb{Z})$. Recall from Subsection~\ref{sub:Metabelian} that the metabelian representation $\alpha(n,\chi) \colon \pi_1(M_{P(K,\eta)}) \to GL_n(\L{\C})$ is defined as the composition of the map $\widetilde{\rho}_{Q} \colon \pi_1(M_{P(K,\eta)}) \to~G$ of~\eqref{eq:MapToSemiDirect} with the representation $\gamma_Q(n,\chi) \colon G \to GL_n(\L{\C})$ described in~\eqref{eq:Matrix}. Recalling~the definition of $\iota_*$ from \eqref{eq:InclusionSatelliteAlexanderModule} as well as the definition of $\iota_n$ from~\eqref{eq:Defin}, the inclusion map $ S^3 \setminus \mathcal{N}(K) \to~M_{P(K,\eta)}$ gives rise to the following commutative diagram:
\begin{equation}
\label{eq:DiagramForCover}
\xymatrix@R0.5cm{
\pi_1(S^3\setminus \mathcal{N}(K)) \ar[r]^{\iota}\ar[d] \ar@/_7pc/[ddd]_{\widetilde{\rho}_{K,Q}}& \pi_1(M_{P(K,\eta)}) \ar[d] \ar@/^5pc/[dd]^{\widetilde{\rho}_Q} \\
\mathbb{Z} \ltimes H_1(M_K;\mathbb{Z}[t_K^{\pm 1}]) \ar[r]^{}\ar[d]& \mathbb{Z} \ltimes H_1(M_{P(K,\eta)};\mathbb{Z}[t_Q^{\pm 1}]) \ar[d] \\
\mathbb{Z} \ltimes \bigoplus_{i=1}^{h} t_Q^{i-1}H_1(L_{n/h}(K);\mathbb{Z})\ar[r]^{}
\ar@/^0.5pc/[d]^{\operatorname{proj}_0}& \mathbb{Z} \ltimes H_1(L_{n}(Q);\mathbb{Z}) \\
\mathbb{Z} \ltimes H_1(L_{n/h}(K);\mathbb{Z}),\ar[ru]_{\iota_n} \ar@/^0.5pc/[u]^{\operatorname{incl}_0}&
}
\end{equation}
where $\operatorname{proj}_{0}(t_{Q}^{\ell}, (x_{1}, t_{Q}x_{2},\ldots,t_{Q}^{h-1}x_{h})) = (t_{Q}^{\ell},x_{1})$ and $ \operatorname{incl}_{0}(t_{Q}^{\ell}, x) = (t_{Q}^{\ell}, (x, 0, \ldots, 0)).$
Since $\alpha(n,\chi)_K=\alpha(n,\chi) \circ \iota=\gamma_Q(n,\chi) \circ \widetilde{\rho}_Q \circ \iota =\gamma_Q(n,\chi) \circ \iota_n \circ \widetilde{\rho}_{K,Q}$, we must study the restriction of $\gamma_Q(n,\chi)$ to $H_1:=\operatorname{im}(\iota_n)$. We shall denote this restriction by $\res^{G}_{H_1} \gamma_Q(n,\chi)$. Next, set $H_2:=n\mathbb{Z} \times H_1(L_n(Q);\mathbb{Z})$ and recall from Remark~\ref{rem:InductionFunctor} that $\gamma_Q(n,\chi)$ is isomorphic to the induced representation $\ind^{G}_{H} \rho_Q(n,\chi).$ Thus, in order to understand $\alpha(n,\chi)_K$, we must study
$$\res^{G}_{H_1} \ind^{G}_{H_2} \rho_Q(n,\chi).$$
\end{remark}
Remark~\ref{rem:MackeyMotivation} calls for an application of Mackey's induction formula. In order to recall this result, we start with a group $G$ and two subgroups \(H_{1},H_{2} \subset G\).
For any \(a \in G\), there is a \emph{double \((H_{1},H_{2})\)-coset} \(H_{1} a H_{2} := \{h_{1} a h_{2} \mid h_{1} \in H_{1}, h_{2} \in H_{2}\} \subset G\).
It is easy to see that \(H_{1} a H_{2} = H_{1} b H_{2}\) if and only if there are \(h_{1} \in H_{1}\) and \(h_{2} \in H_{2}\) such that \(a = h_{1} b h_{2}\).
Denote the set of double \((H_{1},H_{2})\)-cosets by \(H_{1} \backslash G / H_{2}\).
If \(a \in G\) and \(\rho\) is a representation of \(H_{2}\), then we will denote by \({}^{a}\rho\) the representation of \(aH_{2}a^{-1}\) given by the formula ${}^{a}\rho(x)= \rho(a^{-1}xa).$
With these notations, Mackey's formula reads as follows; see \cite[Theorem 10.13]{CurtisReiner}.
\begin{theorem}\label{thm:mackey-thm}
Let $G$ be a group and let \(H_{1},H_{2} \subset G\) be finite index subgroups. If \(\rho\) is a representation of \(H_{2}\), then we have
\[\res^{G}_{H_{1}} \ind^{G}_{H_{2}}(\rho) = \bigoplus_{H_{1} a H_{2}} \ind^{H_{1}}_{aH_{2}a^{-1} \cap H_{1}} \res^{aH_{2}a^{-1}}_{aH_{2}a^{-1} \cap H_{1}} ({}^{a} \rho),\]
where the sum is taken over all \(H_{1} a H_{2} \in H_{1} \backslash G / H_{2}\).
\end{theorem}
We will now apply Mackey's formula to the group \(G = \mathbb{Z} \ltimes H_{1}(L_{n}(P(K,\eta));\mathbb{Z})\) and to the subgroups $H_2 = n\mathbb{Z} \times H_{1}(L_{n}(P(K,\eta));\mathbb{Z})$ and $H_1 = \operatorname{Im}(\iota_{n}).$ In order to carry this out, we need slight generalizations of our metabelian representations. Let $J$ be an arbitrary knot. Given a complex number \(\theta \in S^{1}\) and an integer~\(\ell~\in~\mathbb{Z}\), we consider the representation
\begin{align}
\label{eq:representation-beta}
\gamma_J(n,\chi,\theta,\ell) \colon &\mathbb{Z} \ltimes H_{1}(L_{n}(K);\mathbb{Z}) \to GL_{n}(\L{\C}) \nonumber \\
&(t_Q^{a},x) \mapsto
\begin{pmatrix}
0 & 1 & 0 & \cdots & 0 \\
0 & 0 & 1 & \cdots & 0 \\
\vdots & \vdots & \vdots & & \vdots \\
0 & 0 & 0 & \cdots & 1 \\
\theta t^{\ell} & 0 & 0 & \cdots & 0
\end{pmatrix}^{a}
\begin{pmatrix}
\xi_{m}^{\chi(x)} & 0 & 0 & \cdots & 0 \\
0 & \xi_{m}^{\chi(t_J \cdot x)} & 0 & \cdots & 0 \\
\vdots & \vdots & \vdots & & \vdots \\
0 & 0 & 0 & 0 & \xi_{m}^{\chi(t_J^{n-1}\cdot x)}
\end{pmatrix}.
\end{align}
By definition of this representation, we have the equality $\gamma_J(n,\chi,1,1) = \gamma_J(n,\chi)$. Next, we generalize the representation $\rho_J(n,\chi)$ by considering the representation
\begin{align}
\rho_J(n,\chi,\theta,\ell) \colon n\mathbb{Z} \times H_1(L_n(J);\mathbb{Z}) &\to \L{\C} \\
(t_J^{nk},x) &\mapsto (\theta t^{\ell})^{k} \cdot \xi_{m}^{\chi(x)}. \nonumber
\end{align}
Just as for the representation $\gamma_J(n,\chi,\theta,\ell)$, we observe that $\rho_J(n,\chi,1,1)=\rho_J(n,\chi)$. Furthermore, generalizing Remark~\ref{rem:InductionFunctor}, we also have the induction formula
$$\ind_H^G \rho_J(n,\chi,\theta,\ell)=\gamma_J(n,\chi,\theta,\ell).$$
The following lemma gathers some results for our application of Mackey's formula.
\begin{lemma}
\label{lem:DoubleCosets}
The following assertions hold:
\begin{enumerate}
\item Every double \((H_1,H_2)\)-coset is equal to~\(H_1(t_{Q}^{i},0)H_2\), for some \(1 \leq i \leq h\).
\item $H_2$ is a normal subgroup of $G$.
\item For any $(t_Q^k,x) \in G$ and any \(1 \leq i \leq h\), we have $${}^{t_{Q}^{i}} \rho_{Q}(n,\chi,\theta,\ell)(t_Q^k,x) = \rho_{Q}(n,\chi_{i},\theta,\ell)(t_Q^k,x).$$
\item For any $(t_K^{n/h},x) \in H_1 \cap H_2$, we have
$$\res_{H_2 \cap H_1}^{H_2} \rho_{Q}(n,\chi,\theta,\ell)(t_K^{n/h},x) = \rho_{K}(n/h,\chi,\theta^{w/h} \xi_{m}^{\chi(q_Q(\mu_K^{-w}\eta))},\ell w/h)(t_K^{n/h},x).$$
\end{enumerate}
\end{lemma}
\begin{proof}
Given \(g = (t_{Q}^d,x) \) in $G$, we study the $(H_1,H_2)$-coset $H_1gH_2$. We write $h=\operatorname{gcd}(n,w)$ as \(h=an + bw \) as well as \(d = hs + r\), for some \(0 \leq r \leq h-1\). Furthermore, for \(x \in H_{1}(L_{n}(P(K,\eta));\mathbb{Z})\), we set \(\nu_{p}(t_Q) = t_{Q}^{-p+1}+t_{Q}^{-p+2}+\cdots+t_{Q}^{-1}+1\). A computation now shows that
\[(t_{Q}^{d},x) = (t_{Q}^{w},q_Q(\mu_K^{-w}\eta))^{bs} (t_{Q}^{r},0) (t_{Q}^{asn}, x-t_Q^{-r-asn}\nu_{wbs}(t_Q)q_Q(\mu_K^{-w}\eta)).\]
To show that the right hand side belongs to~\(H_1(t_{Q}^r,0)H_2\), we only need to prove that $(t_{Q}^{w},q_Q(\mu_K^{-w}\eta))$ belongs to $H_1=\operatorname{im}(i_n)$. This follows from the diagram in~\eqref{eq:DiagramForCover} which implies that
$$ (t_{Q}^{w},q_Q(\mu_K^{-w}\eta))=\widetilde{\rho}_Q(\mu_K^{w} (\mu_K^{-w}\eta)=\widetilde{\rho}_Q(\iota (\mu_K))=\iota_n(\widetilde{\rho}_{K,Q}(\mu_K))=\iota_n(t_K,0). $$
The second assertion is a consequence of the definition of the group law on $G$. We now prove the third assertion. Given $(t_Q^k,x) \in G$ and $0 \leq i \leq h-1$, we apply successively the definition of $ {}^{a} \rho$, the group law in $G$ and the definition of $\chi_i$ to obtain the desired equality:
\begin{align*}
{}^{t_{Q}^{i}} \rho_{Q}(n,\chi,\theta,\ell)(t_{Q}^{k},x)
&= \rho_{Q}(n,\chi,\theta,\ell)((t_{Q}^{-i},0) (t_{Q}^{k},x)(t_{Q}^{i},0)) \\
&= \rho_{Q}(n,\chi,\theta,\ell)(t_{Q}^{k},t_{Q}^{-i}\cdot x) \\
&=\rho_{Q}(n,\chi_{i},\theta,\ell)(t_{Q}^{k},x) .
\end{align*}
To prove the fourth assertion, we first note that $H_2 \cap H_1=\operatorname{im} \left(\iota_{n}|_{(n/h)\mathbb{Z} \times H_{1}(L_{n/h}(K);\mathbb{Z})}\right)$. Next, using the relation \(\iota_{n}(t_{K},0) = (t_{Q}^{w},q_Q(\mu_K^{-w}\eta))\) and the definition of $\rho(n,\chi,\theta,\ell)$, we obtain
\begin{align*}
\left(\res_{H_2 \cap H_1}^{H_2} \rho_{Q}(n,\chi,\theta,w)\right)(t_{K}^{n/h},0)
&= \rho_{Q}(n,\chi,\theta,\ell)(t_{Q}^{wn/h},q_Q(\mu_K^{-w}\eta)) \\
&= (\theta t^{\ell})^{w/h} \xi_{m}^{\chi(q_Q(\mu_K^{-w}\eta))} \\
&=\rho_{K}(n/h,\chi,\theta^{w/h} \xi_{m}^{\chi(q_Q(\mu_K^{-w}\eta))},\ell w/h))(t_{K}^{n/h},0) .
\end{align*}
To get the equality for arbitrary elements $(t_{K}^{n/h},x) \in H_1 \cap H_2$, note that the value of both homomorphisms on $(1,x)$ is $\xi_m^{\chi(\iota_n(x))}.$ This concludes the proof of the lemma.
\end{proof}
Lemma~\ref{lem:DoubleCosets} is now used to apply Mackey's theorem to our setting.
\begin{proposition}
\label{prop:MackeyApplication}
Use $\gamma_Q(n,\chi)_{K}$ to denote the restriction of $\gamma_Q(n,\chi)$ to $\mathbb{Z} \ltimes H_1(L_{n/h}(K),\mathbb{Z})$. There exists an isomorphism of representations
\[\gamma_{Q}(n,\chi)_{K} \cong \bigoplus_{i=1}^{h} \gamma_{K}(n/h,\chi_{i},\xi_{m}^{\chi_{i}(q_Q(\mu_K^{-w}\eta))},w/h).\]
\end{proposition}
\begin{proof}
Since we saw in Remark~\ref{rem:MackeyMotivation} that $\gamma(n,\chi)_{K} = \res^{G}_{H_1} \ind^{G}_{H_2} \rho_{Q}(n,\chi),$ our goal is to apply Mackey's formula to $\rho_Q(n,\chi)=\rho_Q(n,\chi,1,1)$. For the remainder of the proof, we write $\rho$ as a shorthand for $\rho_Q(n,\chi,\theta,\chi)$.
Using consecutively Theorem~\ref{thm:mackey-thm} (as well as the first item of Lemma~\ref{lem:DoubleCosets}), and then the second, third and fourth items of Lemma~\ref{lem:DoubleCosets}, we obtain
\begin{align*}
\res^{G}_{H_{1}} \ind^{G}_{H_{2}}(\rho)
&= \bigoplus_{i=1}^{h} \ind^{H_1}_{H_2 \cap H_1} \res^{H_2}_{H_2 \cap H_{1}} ({}^{t_Q^i} \rho) \\
&= \bigoplus_{i=1}^{h} \ind^{H_1}_{H_2 \cap H_1} \res^{H_2}_{H_2 \cap H_{1}} (\rho_{Q}(n,\chi_{i},\theta,\ell)) \\
&= \bigoplus_{i=1}^{h} \ind^{H_1}_{H_2 \cap H_1}\rho_{K}(n/h,\chi_i,\theta^{w/h} \xi_{m}^{\chi_i(q_Q(\mu_K^{-w}\eta))},\ell w/h)).
\end{align*}
Since $H_2 \cap H_1=\operatorname{im} \left(\iota_{n}|_{(n/h)\mathbb{Z} \times H_{1}(L_{n/h}(K);\mathbb{Z})}\right)$, observe that $\ind_{H_1 \cap H_2}^G \rho_K(n,\chi,\theta,\ell)=\gamma_K(n,\chi,\theta,\ell)$. Additionally, recall that $\gamma_Q(n,\chi)_{K} = \res^{G}_{H_1} \ind^{G}_{H_2} \rho_{Q}(n,\chi,1,1)$ and $\gamma_Q(n,\chi)=\gamma_Q(n,\chi,1,1)$. Therefore taking $\theta=1$ and $\ell=1$ in the previous computation, we obtain
\begin{align*}
\gamma_Q(n,\chi)_{K}
=\res^{G}_{H_{1}} \ind^{G}_{H_{2}}(\rho_Q(n,\chi))
=\bigoplus_{i=1}^{h} \gamma_{K}(n/h,\chi_{i},\xi_{m}^{\chi_{i}(q_Q(\mu_K^{-w}\eta))},w/h).
\end{align*}
This concludes the proof of the proposition.
\end{proof}
Next, we return to Blanchfield pairings. Namely, we discuss the effect of the representation~$\gamma_Q(n,\chi,\theta,\ell)$ on the variable $t$ of the metabelian Blanchfield pairing.
\begin{remark}
\label{rem:BlanchfieldChangeVariable}
Recall from Subsection~\ref{sub:Metabelian} that for any knot $J$, the representation $\alpha_J(n,\chi)$ is obtained by precomposing the representation $\gamma_J(n,\chi) \colon \mathbb{Z} \ltimes H_1(L_n(J);\mathbb{Z}) \to GL_n(\L{\C})$ with the map $\widetilde{\rho}_J \colon \pi_1(M_J) \to \mathbb{Z} \ltimes H_1(L_n(J);\mathbb{Z})$. We shall adopt the same convention for $\gamma_J(n,\chi,\theta,\ell)$, thus obtaining a representation $\alpha_J(n,\chi,\theta,\ell)$ of $\pi_1(M_J)$. Since $\gamma_J(n,\chi,\theta,\ell)$ can be obtained from~\(\gamma_J(n,\chi)\) via the substitution~\(t \mapsto \theta t^{\ell}\), it follows that $\Bl_{\alpha(n,\chi,\theta,\ell)}(J)(t)=\Bl_{\alpha(n,\chi)}(J)(\theta t^\ell)$.
\end{remark}
Returning to satellite knots, recall from Remark~\ref{rem:MackeyMotivation} that the Blanchfield pairing~$\Bl_{\alpha(n,\chi)_K}(K)$ is obtained by precomposing the representation $\gamma_Q(n,\chi)_K \colon \mathbb{Z} \ltimes H_1(L_{n/h}(K);\mathbb{Z}) \to GL_n(\L{\C})$ with the map $\widetilde{\rho}_{K,Q} \colon \pi_1(M_K) \to \mathbb{Z} \ltimes H_1(L_{n/h}(K);\mathbb{Z})$. Thus, applying Proposition~\ref{prop:MackeyApplication} and Remark~\ref{rem:BlanchfieldChangeVariable}, we obtain the following isometry of linking forms:
\begin{equation}
\label{eq:MackeyBlanchfield}
\Bl_{\alpha(n,\chi)_K}(K)(t)
\cong \bigoplus_{i=1}^{h} \Bl_{\alpha_K(n/h,\chi)}(K)(\chi_{i}(q_Q(\mu_K^{-w}\eta))t^{w/h}).
\end{equation}
Using this isometry, we can now prove Theorem~\ref{thm:metabelian-cabling-formula}.
\begin{proof}[Proof of Theorem~\ref{thm:metabelian-cabling-formula}]
First of all, note that Lemma~\ref{lemma:metabelian-eta-regularity} ensures that $\alpha(n,\chi)$ is $\eta$-regular.
We~can therefore apply the satellite formula of Theorem~\ref{thm:CablingTheorem} to each of the cases which we shall now~distinguish.
We first assume that $w \neq~0$. As $\alpha(n,\chi)$ is $\eta$-regular, it restricts to representations $\alpha(n,\chi)_K$ and $\alpha(n,\chi)_P$ on $\pi_1(M_K)$ and~$\pi_1(M_P)$. We consider two cases.
\emph{Case 1.} If \(w\) is divisible by \(n\), then $\alpha(n,\chi)_K$ is abelian and is isomorphic to $\bigoplus_{i=1}^n \xi_m^{\chi_i} \otimes~\phi_K^{w/n}$, where $\xi_m^{\chi_i}$ is understood as the character mapping the meridian $\mu_K$ to $\xi_m^{\chi_i(q_Q(\mu_Q^{-w}\eta))}$. As a consequence,~\eqref{eq:MackeyBlanchfield} and Corollary~\ref{cor:cabling-formula-abelian} provide the desired isometry of linking forms:
\[\Bl_{\alpha(n,\chi)}(P(K,\eta)) \cong \Bl_{\alpha(n,\chi_P)}(P) \oplus \bigoplus_{i=1}^{n} \Bl(K)(\xi_m^{\chi_{i}(q_Q(\mu_K^{-w}\eta))}t^{w/n}).\]
\emph{Case 2.} If \(w\) is not divisible by \(n\), then~\eqref{eq:MackeyBlanchfield} and Theorem~\ref{thm:CablingTheorem} imply that
\(\Bl_{\alpha(n,\chi)}(P(K,\eta))\)
is Witt equivalent to
\[\Bl_{\alpha(n,\chi_{P})}(P) \oplus \bigoplus_{i=1}^{h} \Bl_{\alpha_K(n/h,\chi_{i})}(K)(\xi_m^{\chi_{i}(q_Q(\mu_K^{-w}\eta))}t^{w/h}).\]
Next, we assume that \(w=0\).
As the abelianization map $\phi \colon H_1(M_{P(K,\eta)};\mathbb{Z}) \to \mathbb{Z}$ restricts to the zero map on $H_1(M_K;\mathbb{Z})$, it follows that $\alpha(n,\chi)_K$ is abelian.
More precisely, \(\alpha(n,\chi)_{K} \cong \bigoplus_{i=1}^{n} \xi_m^{\chi_i}\), where $\xi_m^{\chi_i}$
is understood as the character mapping $\mu_K$ to $\xi_m^{\chi_i(q_Q(\eta))}$.
As a consequence, we obtain the isomorphism
$$H_{\ast}(M_{K};\L{\C}^{n}_{\alpha(n,\chi)}) \cong \bigoplus_{i=1}^n H_{1}(M_{K};\L{\C}_{\xi_m^{\chi_i}})\cong \bigoplus_{i=1}^n H_{1}(M_{K};\mathbb{C}_{\xi_m^{\chi_i}}) \otimes_\mathbb{C}\L{\C}. $$
The dimension of $H_{1}(M_{K};\mathbb{C}_{\xi_m^{\chi_i}}) $ is equal to $\eta_K(\xi_m^{\chi_i(q_Q(\eta))})$, the nullity of $K$ evaluated at the unit complex number $\xi_m^{\chi_i(q_Q(\eta))}=\xi_m^{\chi(t_Q^{i-1}q_Q(\eta))}$.
This value is non-zero if and only if the Alexander polynomial satisfies $\Delta_K(\xi_m^{\chi(t_Q^{i-1}q_Q(\eta))})=0$.
Since the character $\chi$ has prime power order; this can never happen~\cite[proof of Proposition 3.3]{FriedlEta} and we therefore deduce that $H_{\ast}(M_{K};\L{\C}^{n}_{\alpha(n,\chi)})=~0.$ Applying Corollary~\ref{cor:cabling-formula-abelian} now provides the desired isometry :
$$\Bl_{\alpha(n,\chi)}(P(K,\eta)) \cong \Bl_{\alpha(n,\chi_{P})}(P).$$
This concludes the proof of the theorem.
\end{proof}
\section{Metabelian Blanchfield forms of (iterated) torus knots}
\label{sec:TorusKnots}
In this section, we use the algorithm described in~\cite{PowellThesis, MillerPowell} in order to compute metabelian Blanchfield forms of the torus knots \(T_{2,2k+1}\), for \(k>0\).
In Subsection~\ref{sub:ChainComplexBlanchfield}, we recall the chain complex computation of the Blanchfield pairing which is due to Powell~\cite{PowellThesis}, in Subsections~\ref{sub:ChainComplex} and~\ref{sec:blanchf-forms-twist}, we use this algorithm on metabelian Blanchfield pairings of~\(T_{2,2k+1}\), in Subsection~\ref{sub:MetabelianSignaturesTorusKnot}, we determine the isometry type of these linking forms, and in Subsection~\ref{sub:HeddenKirkLivingston}, we recover results on the non sliceness of some linear combinations of algebraic knots.
Throughout this section, we will be working with the diagram of \(T_{2,2k+1}\) depicted in Figure~\ref{fig:diagram-T_2_2k+1}.
\begin{figure}[!htb]
\centering
\def0.7{0.7}
\import{./}{torus-knot.pdf_tex}
\caption{A diagram of \(T_{2,2k+1}\) together with generators of the knot group. Arrows indicate orientation of the respective meridian when going under the knot. The blue loop is \(a = x_{2k}x_{2k+1}=x_{1}x_{2}\).}
\label{fig:diagram-T_2_2k+1}
\end{figure}
\subsection{The chain complex computation of the Blanchfield pairing}
\label{sub:ChainComplexBlanchfield}
We briefly recall Powell's algorithm to compute the Blanchfield pairing~\cite{PowellThesis}.
We then discuss the example we wish to study and the steps that must be followed to perform Powell's algorithm.
\medbreak
In~\cite{Powell}, Powell defines twisted Blanchfield pairings for arbitrary $3$-dimensional symmetric chain complexes.
When $N$ is a closed oriented $3$-manifold and $\beta \colon \pi_1(N) \to GL_d(\L{\F})$ is a unitary acyclic representation, his definition yields a non-singular linking form
$$ \Bl^{\beta}(N) \colon H^2(N;\L{\F}^d_\beta) \times H^2(N;\L{\F}^d_\beta) \to \mathbb{F}(t)/\L{\F} $$
on the twisted \emph{co}homology of $N$.
Instead of recalling the definition of this pairing, we focus on the algorithm described in~\cite{MillerPowell} to compute it, and take the result as our definition of $ \Bl^{\beta}(N)$.
Use $\widetilde{N}$ to denote the universal cover of $N$ and let $(C^*(\widetilde{N}), \partial^*)$ be the resulting cochain complex of left $\mathbb{Z}[\pi_1(N)]$-modules.
It is well known that the Poincar\'e duality isomorphisms $H^{3-*}(\widetilde{N};\mathbb{Z}) \to H_*(\widetilde{N};\mathbb{Z})$ arise from chain homotopy equivalences $C^{3-*}(\widetilde{N}) \to C_*(\widetilde{N})$. In particular, we obtain a map $\Phi \colon C^2(\widetilde{N}) \to C_1(\widetilde{N})$.
Passing to twisted chain complexes, $\partial^*$ and $\Phi$ induce maps
\begin{align*}
&\beta(\partial^*) \colon C^*(N;\L{\F}^d_\beta) \to C^{*+1}(N;\L{\F}^d_\beta), \\
&\beta(\Phi) \colon C^2(N;\L{\F}^d_\beta) \to C_1(N;\L{\F}^d_\beta).
\end{align*}
For later use, we also recall that matrices are assumed to act on row vectors from the right and that the cohomological differentials are determined by the homological differentials via the formula~\(\beta(\partial^{i}) = (-1)^{i} \beta(\partial_{i})^{\#T}\).
The following definition is due to Powell~\cite{Powell} (see also~\cite{MillerPowell}).
\begin{definition}
\label{def:CohomBlanchfield}
Let $N$ be a closed oriented $3$-manifold and let $\beta \colon \pi_1(N) \to GL_d(\L{\F})$ be a unitary acyclic representation.
The \emph{cohomological twisted Blanchfield} is defined as
\begin{align}
\label{eq:ChainBlanchfieldFormula}
\Bl^{\beta}(N) \colon H^2(N;\L{\F}^d_\beta) \times H^2(N;\L{\F}^d_\beta) &\to \mathbb{F}(t)/\L{\F} \nonumber \\
([v],[w]) & \mapsto \frac{1}{s} \left(v \cdot \beta(\Phi) \cdot Z^{\# T}\right)^{\# T},
\end{align}
where $v,w \in C^2(N;\L{\F}^d_\beta)$ and $Z \in C^1(N;\L{\F}^d_\beta)$ satisfies $Z\beta(\partial^2)=sw$ for some~$s \in~\L{\F} \setminus \lbrace 0 \rbrace$.
\end{definition}
The cohomological twisted Blanchfield pairing is isometric to the twisted Blanchfield pairing from Definition~\ref{def:Blanchfield}.
Indeed, recalling that $\operatorname{PD}$ denotes Poincar\'e duality on the (co)homological level,~\cite[Proposition 5.3]{MillerPowell} implies that for $x,y \in H_1(N;\L{\F}^d_\beta)$, the pairings are related by
\[ \Bl^{\beta}(N)(x,y) = \Bl_{\beta}(N)(PD(x),PD(y)).\]
The goal of the next two subsections is to use Definition~\ref{def:CohomBlanchfield} to understand the (signatures of the) metabelian Blanchfield pairings $\Bl^{\alpha(2,\chi)}(T_{2,2k+1})$: in Subsection~\ref{sub:ChainComplex}, we will compute $\pi_1(M_{T_{2,2k+1}})$ and $C_*(\widetilde{M}_{T_{2,2k+1}})$, and in Subsection~\ref{sec:blanchf-forms-twist} we focus on $H^2(M_{T_{2,2k+1}};\mathbb{F}[t^{\pm 1}]^2_{\alpha(n,\chi)})$ and $ \beta(\Phi)$.
\subsection{The chain complex of the universal cover}
\label{sub:ChainComplex}
In this subsection, we explicitly describe the chain complex of the universal cover of the $0$-framed surgery $M_{T_{2,2k+1}}$.
\medbreak
We start by providing a presentation for the fundamental group of $M_{T_{2,2k+1}}$.
Such a presentation can be computed from the knot group once a word for the longitude of $T_{2,2k+1}$ is known.
Using Figure~\ref{fig:diagram-T_2_2k+1}, a Wirtinger presentation of the fundamental group of the complement of \(T_{2,2k+1}\) is
\[\pi := \pi_{1}(S^{3} \setminus T_{2,2k+1}) = \langle x_{1}, x_{2}, \ldots, x_{2k+1} \mid r_{1}, r_{2}, \ldots, r_{2k} \rangle,\]
where \(r_{i} = x_{i} x_{i+1} x_{i+2}^{-1} x_{i+1}^{-1}\) and the indices are taken mod \(2k+1\).
The next lemma describes an explicit isomorphism from $\pi$ to a simple presentation. While the statement is very well known, we
give a proof, because we will use it to describe the presentation of the fundamental group of the zero-framed surgery on a torus knot.
\begin{lemma}
\label{lem:PresentationTorusKnot}
There exists an isomorphism \(\phi \colon \pi \to G:=\langle a, b \mid a^{2k+1}b^{2} \rangle\) such that
\begin{align*}
\phi(x_{2(k-s)+1}) &= a^{s} (a^{k}b)^{-1} a^{-s}, \quad 0 \leq s \leq k, \\
\phi(x_{2(k-s)}) &= a^{s} (a^{k+1}b) a^{-s}, \quad 0 \leq s \leq k-1.
\end{align*}
\end{lemma}
\begin{proof}
We first prove that \(\pi\) is generated by \(x_{2k}\) and \(x_{2k+1}\).
Notice that the relation \(r_{i}\) implies that the equality \(x_{i}x_{i+1} = x_{i+1}x_{i+2}\) holds.
As a consequence, we obtain \(x_{i} x_{i+1} = x_{i+s}x_{i+s+1}\) for any \(i\) and \(s\) and, applying this formula recursively, we obtain
\begin{align*}
x_{2k-1} &\stackrel{r_{2k-1}}{=} x_{2k}x_{2k+1}x_{2k}^{-1}, \\
x_{2k-2} &\stackrel{r_{2k-2}}{=} x_{2k-1}x_{2k}x_{2k-1}^{-1} = (x_{2k} x_{2k+1}) x_{2k} (x_{2k}x_{2k+1})^{-1}, \\
x_{2k-3} &\stackrel{r_{2k-3}}{=} x_{2k-2}x_{2k-1}x_{2k-2}^{-1} \\
&= (x_{2k}x_{2k+1}x_{2k}x_{2k+1}^{-1}x_{2k}^{-1})(x_{2k}x_{2k+1}x_{2k}^{-1})(x_{2k}x_{2k+1}x_{2k}^{-1}x_{2k+1}^{-1}x_{2k}^{-1}) \\
&= (x_{2k}x_{2k+1})x_{2k}x_{2k+1}x_{2k}^{-1}(x_{2k}x_{2k+1})^{-1}.
\end{align*}
As a consequence, we eliminate the generators \(x_{1}, x_{2}, \ldots, x_{2k-1}\) and the relations \(r_{1}, r_{2}, \ldots, r_{2k-1}\) since we have the following equalities:
\begin{align*}
x_{2(k-s)+1} &= (x_{2k}x_{2k+1})^{s-1} x_{2k}x_{2k+1}x_{2k}^{-1} (x_{2k}x_{2k+1})^{-s+1}, \quad 1 \leq s \leq k, \\
x_{2(k-s)} &= (x_{2k}x_{2k+1})^{s} x_{2k+1} (x_{2k}x_{2k+1})^{s}, \quad 1 \leq s \leq k-1.
\end{align*}
Reformulating, we have proved that $\pi$ is indeed generated by $x_{2k}$ and $x_{2k+1}$ and we have obtained the presentation $\pi = \langle x_{2k}, x_{2k+1} \mid r_{2k}' \rangle$, where
\begin{align*}
r_{2k}' &= x_{2k} x_{2k+1} x_{1}^{-1} x_{2k+1}^{-1} = x_{2k} x_{2k+1} (x_{2k} x_{2k+1})^{k-1} x_{2k} x_{2k+1}^{-1} x_{2k}^{-1} (x_{2k}x_{2k+1})^{-k+1} x_{2k+1}^{-1} \\
&= (x_{2k} x_{2k+1})^{k} x_{2k} (x_{2k} x_{2k+1})^{-k} x_{2k+1}^{-1}.
\end{align*}
Using this presentation of $\pi$, it is now straightforward to verify that \(\phi \colon \pi \to G \) is a group isomorphism, where the inverse is defined by setting $\phi^{-1}(a) = x_{2k}x_{2k+1}$ and $\phi^{-1}(b) = (x_{2k+1}^{-1}x_{2k}^{-1})^{k} x_{2k+1}^{-1}$.
This concludes the proof of the lemma.
\end{proof}
As a consequence, we can describe the fundamental group of the 0-framed surgery on~$T_{2,2k+1}$.
\begin{lemma}
\label{lem:Presentation0Surgery}
The fundamental group of the \(0\)-surgery on \(T_{2,2k+1}\) admits the following presentation:
\begin{equation}\label{eq:presentation-fund-gp-zero-surgery}
G_{0} = \langle a, b \mid a^{2k+1}b^{2}, (a^{k}b)^{2k+1} a^{2k+1} (a^{k}b)^{2k+1}\rangle.
\end{equation}
\end{lemma}
\begin{proof}
Let \( (\mu,\lambda) \) be the meridian-longitude pair of \(T_{2,2k+1}\) expressed in words of $\pi$.
Using van Kampen's theorem, to prove the proposition, it suffices to show that $\phi(\lambda) = (a^{k}b)^{2k+1}a^{2k+1}(a^{k}b)^{2k+1}$, where $\phi$ is the isomorphism described in Lemma~\ref{lem:PresentationTorusKnot}.
Working with the Wirtinger presentation arising from Figure~\ref{fig:diagram-T_2_2k+1}, we choose $(\mu,\lambda)$ as follows:
\[\mu = x_{2k+1}, \quad \lambda = x_{2k+1}^{-2k-1} x_{1} x_{3} x_{5} \cdots x_{2k+1} x_{2} x_{4} \cdots x_{2k}.\]
Next, the definition of $\phi$ gives $\phi(x_{1} x_{3} \cdots x_{2k+1}) = a^{k} (b^{-1} a^{-k-1})^{k+1} a = a^{2k+1} (a^{-k-1}b^{-1})^{k+1} a^{-k}$ and $\phi(x_{2} x_{4} \ldots x_{2k}) = a^{k} (a^{k}b)^{k}$.
Since the relation \(a^{2k+1}b^{2}=1\) holds, we have \(a^{-k-1}b^{-1} = a^{k}b\), we deduce that $\phi(x_{1}x_{3} \cdots x_{2k+1} x_{2} \cdots x_{2k}) = a^{2k+1} (a^{k}b)^{2k+1}$, and the lemma is now concluded by recalling the definition of $\lambda$.
\end{proof}
In order to obtain a description of $C_*(\widetilde{M}_{T_{2,2k+1}})$, we need some additional terminology.
\begin{definition}[\expandafter{\cite[Definition 3.7]{Powell}}]
Let $\langle x_{1}, x_{2}, \ldots, x_{n} \mid r_{1}, r_{2} ,\ldots, r_{m}\rangle$ be a presentation for a group $G$, let \(F\) be the free group generated by \(x_{1}, \ldots, x_{n}\) and let \(P\) be the free group generated by~\(\rho_{1}, \ldots, \rho_{m}\).
Define a homomorphism \(\psi \colon F \ast P \to G\) by the formulas $\psi(x_{i}) = x_{i}, \psi(\rho_{j}) = r_{j}$. An \emph{identity of the presentation} is a word in \(\ker (\psi) \) that can be written as a product of words of the form \(w \rho_{j}^{\epsilon} w^{-1}\), where $w$ lies in $F$, \(1 \leq j \leq m\) and \(\epsilon = \pm 1\).
\end{definition}
The next lemma will be useful to compute $\Phi \colon C^2(\widetilde{M}_{T_{2,2k+1}}) \to C_1(\widetilde{M}_{T_{2,2k+1}})$.
\begin{lemma}\label{lemma:identity-of-presentation}
Set \(r = a^{2k+1}b^{2}\), \(\lambda = (a^{k}b)^{2k+1}a^{2k+1}(a^{k}b)^{2k+1}\) and \(\mu = (a^{k}b)^{-1}\).
The following equality is satisfied in the free group $F$ generated by $a$ and $b$:
\[\mu \lambda \mu^{-1} \lambda^{-1} = \left[\mu^{-2k}a^{k}\right] r \left[\mu^{-2k}a^{k}\right]^{-1} \mu^{-2k-1} r^{-1} \mu^{2k+1}.\]
Furthermore, if $P$ denotes the free group on $\rho_1,\rho_2$ and if we define \(\psi \colon F \ast P \to G_0\) by \(\psi(\rho_{1})=r\) and \(\psi(\rho_{2})=\lambda\), then
the following word is an identity for the presentation~\eqref{eq:presentation-fund-gp-zero-surgery} of $G_0$:
\begin{equation}\label{eq:identity-of-the-presentation}
R = \mu \rho_{2} \mu^{-1} \rho_{2}^{-1} \mu^{-2k-1} \rho_{1} \mu^{2k+1} \left(\mu^{-2k} a^{k}\right) \rho_{1}^{-1}\left(\mu^{-2k}a^{k}\right)^{-1}.
\end{equation}
\end{lemma}
\begin{proof}
The proof of the first assertion is a direct computation:
\begin{align*}
\mu \lambda \mu^{-1} \lambda^{-1} &= (a^{k}b)^{2k} a^{3k+1} b a^{-2k-1} (a^{k}b)^{-2k-1} \\
&= \left[(a^{k}b)^{2k+1} a^{k} \right] (a^{2k+1}b^{2}) b (b^{-2}a^{-2k-1}) (a^{k}b)^{-2k-1} \\
&= \left[\mu^{-2k-1} a^{k} \right] r \left[\mu^{-2k-1} a^{k} \right]^{-1} \left[\mu^{-2k-1} a^{k} \right] b \mu^{2k+1} \mu^{-2k-1} r^{-1} \mu^{2k+1} \\
&= \left[\mu^{-2k}a^{k}\right] r \left[\mu^{-2k}a^{k}\right]^{-1} \mu^{-2k-1} r^{-1} \mu^{2k+1}.
\end{align*}
The second assertion follows from the definition of an identity of a presentation.
\end{proof}
We now provide a description for the chain complex of the universal cover of $M_{T_{2,2k+1}}$.
\begin{proposition}\label{prop:cellular-chain-complex}
The cellular chain complex for the universal cover of the zero-surgery \(M_{T_{2,2k+1}}\) is given by the $\mathbb{Z}[G_0]$-module chain complex $\mathbb{Z}[G_{0}] \xrightarrow{\partial_{3}} \mathbb{Z}[G_{0}]^{2} \xrightarrow{\partial_{2}} \mathbb{Z}[G_{0}]^{2} \xrightarrow{\partial_{1}} \mathbb{Z}[G_{0}],$ where the differentials are described by the following formulas:
\begin{align*}
\partial_{3} &=
\begin{pmatrix}
\mu^{-2k-1}-\mu^{-2k} \cdot a^{k} & \mu-1 \\
\end{pmatrix}, \\
\partial_{2} &=
\begin{pmatrix}
\frac{\partial (a^{2k+1}b^{2})}{\partial a} & \frac{\partial (a^{2k+1}b^{2})}{\partial b} \\
\frac{\partial \lambda_{0}}{\partial a} & \frac{\partial \lambda_{0}}{\partial b} \\
\end{pmatrix}, \\
\partial_{1} &=
\begin{pmatrix}
a-1 & b-1
\end{pmatrix}^{T}.
\end{align*}
Moreover, if \(\Phi \colon C^{2}(\widetilde{M}_{T_{2,2k+1}}) \to C_{1}(\widetilde{M}_{T_{2,2k+1}})\) denotes the chain map representing Poincar\'e duality, then
\begin{equation}
\label{eq:Phi}
\Phi =
\begin{pmatrix}
\mu^{2k+1} \cdot \frac{\partial (\mu^{-2k-1})}{\partial a} - a^{-k} \cdot \mu^{2k} \cdot \frac{\partial (\mu^{-2k} \cdot a^{k})}{\partial a} & \mu^{2k+1} \cdot \frac{\partial (\mu^{-2k-1})}{\partial b} - a^{-k} \cdot \mu^{2k} \cdot \frac{\partial (\mu^{-2k} \cdot a^{k})}{\partial b} \\
\mu^{-1} \cdot \frac{\partial \mu}{\partial a} & \mu^{-1} \cdot \frac{\partial \mu}{\partial b} \\
\end{pmatrix}.
\end{equation}
\end{proposition}
\begin{proof}
The formulas for \(\partial_{1}\) and \(\partial_{2}\) are well known.
To compute the last differential, note that by~\cite{TrotterSystem}, \(\partial_{3}\) is expressible in terms of \(\frac{\partial R}{\partial \rho_{1}}\) and \(\frac{\partial R}{\partial \rho_{2}}\) via the fomula
\[\partial_{3} = \left(\begin{smallmatrix}\frac{\partial R}{\partial \rho_{1}} & \frac{\partial R}{\partial \rho_{1}} \end{smallmatrix}\right).\]
To obtain \(\Phi\), we use the formula discussed after the proof of Theorem~3.15 in~\cite{MillerPowell}, see also~\cite{TrotterSystem}.
\end{proof}
In fact, in later computations, we will need the explicit formulas for the $\partial_i$ and, for this reason, we record the following computations which only require Fox calculus:
\begin{align*}
\frac{\partial (a^{2k+1}b^{2})}{\partial a} &= 1+a+a^{2}+\ldots+a^{2k}, \\
\frac{\partial (a^{2k+1}b^{2})}{\partial b} &= a^{2k+1}(1+b), \\
\frac{\partial \lambda_{0}}{\partial a} &= (1 + (a^{k}b)^{2k+1} a^{2k+1}) \frac{\partial (a^{k}b)^{2k+1}}{\partial a} + (a^{k}b)^{2k+1} (1+a+a^{2}+\ldots+a^{2k}), \\
\frac{\partial \lambda_{0}}{\partial b} &= (1 + (a^{k}b)^{2k+1} a^{2k+1}) \frac{\partial (a^{k}b)^{2k+1}}{\partial b}, \\
\frac{\partial (a^{k}b)^{2k+1}}{\partial a} &= (1+a^{k}b+(a^{k}b)^{2}+\ldots+(a^{k}b)^{2k}) (1+a+a^{2}+\ldots+a^{k-1}), \\
\frac{\partial (a^{k}b)^{2k+1}}{\partial b} &= (1+a^{k}b+(a^{k}b)^{2}+\ldots+(a^{k}b)^{2k}) a^{k}.
\end{align*}
\subsection{Blanchfield forms twisted by dihedral representations}
\label{sec:blanchf-forms-twist}
In this subsection, we use Powell's algorithm to compute metabelian Blanchfield pairings of $T_{2,2k+1}$.
\medbreak
Use \(L_{2}(T_{2,2k+1})\) to denote the double cover of \(S^{3}\) branched along \(T_{2,2k+1}\) and let \(\xi = \xi_{2k+1}\) be a primitive root of unity of order \(2k+1\).
Recall from Subsection~\ref{sub:Metabelian} that for every character $\chi \colon H_1(L_{2}(T_{2,2k+1});\mathbb{Z}) \to \mathbb{Z}_{2k+1}$, there is a metabelian representation
\[ \alpha(2,\chi) \colon \pi_1(M_{T_{2,2k+1}}) \to GL_2(\mathbb{C}[t^{\pm 1}]).\]
Since $\pi_1(M_{T_{2,2k+1}}) \cong G_0$ has a simple presentation (with the meridians \(x_{2k},x_{2k+1}\) as generators) and since $L_2(T_{2,2k+1})$ is a lens space, the representation~$\alpha(2,\chi)$ can be described quite explicitly.
This is the goal of the next few paragraphs.
As $H_1(L_{2}(T_{2,2k+1});\mathbb{Z})\cong \mathbb{Z}_{2k+1}$, characters $\chi \colon H_1(L_{2}(T_{2,2k+1});\mathbb{Z}) \to \mathbb{Z}_{2k+1}$ are determined by the image of a generator.
We shall use $x_{2k}$ and $x_{2k+1}$ to describe such a generator explicitly.
We start by producing generators of the Alexander module of $T_{2,2k+1}$.
Since the abelianization map
$$\operatorname{Ab} \colon G_{0} \cong \pi_1(M_{T_{2,2k+1}})\to \mathbb{Z}$$
sends $x_{2k}$ and $x_{2k+1}$ to $1$, the commutator subgroup \(G_{0}^{(1)} = [G_{0},G_{0}]\) consists of words in \(x_{2k}, x_{2k+1}\) such that the sum of the exponents is zero.
In particular, \(x_{2k}x_{2k+1}^{-1} \in G_{0}^{(1)}\) and it is easy to check that~\(G_{0}^{(1)}\) is normally generated by \(x_{2k} x_{2k+1}^{-1}\).
Therefore, the image of \(x_{2k} x_{2k+1}^{-1}\) generates the Alexander module \(H_{1}(M_{K};\mathbb{Z}[t^{\pm1}]) = G_{0}^{(1)} / G_{0}^{(2)}\) as a \(\mathbb{Z}[t^{\pm1}]\)-module.
In order to describe a generator of $H_{1}(L_{2}(T_{2,2k+1});\mathbb{Z}) \cong \mathbb{Z}_{2k+1}$, consider the projection
\[q \colon G_{0}^{(1)} \to H_{1}(M_{K};\mathbb{Z}[t^{\pm1}]) \stackrel{}{\to} H_{1}(L_{2}(T_{2,2k+1});\mathbb{Z}) \]
Since the image of $x_{2k}x_{2k+1}^{-1}$ generates the Alexander module, it follows that $q(x_{2k} x_{2k+1}^{-1})$ generates $H_{1}(L_{2}(T_{2,2k+1});\mathbb{Z})$.
As a consequence, the $2k+1$ characters $H_{1}(L_{2}(T_{2,2k+1});\mathbb{Z}) \to \mathbb{Z}_{2k+1}$ can be described by imposing that
\[\chi_{\theta} \colon H_{1}(L_{2}(T_{2,2k+1});\mathbb{Z}) \to \mathbb{Z}_{2k+1}\]
satisfies \(\chi_{\theta}(q(x_{2k} x_{2k+1}^{-1})) = \theta\) for $\theta=0,\ldots,2k$.
We shall now use these observations to compute the value of the metabelian representation~$\rho_\theta:=\alpha(2,\chi_{\theta})$ on the generators $a$ and $b$ of $G_0$.
Using~\eqref{eq:Matrix}, we have~
$\rho_{\theta}(x_{2k} x_{2k+1}^{-1}) =
\left(\begin{smallmatrix} \xi^{\theta} & 0 \\
0 & \xi^{-\theta} \\
\end{smallmatrix}\right)$ and~$ \rho_{\theta}(x_{2k+1}) =\left(\begin{smallmatrix} 0 & 1 \\ t & 0 \\ \end{smallmatrix}\right) $, and we therefore obtain
\[\rho_{\theta}(x_{2k}) =
\begin{pmatrix}
0 & 1 \\
t & 0 \\
\end{pmatrix}
\begin{pmatrix}
\xi^{\theta} & 0 \\
0 & \xi^{-\theta} \\
\end{pmatrix}
=
\begin{pmatrix}
0 & \xi^{-\theta} \\
t \xi^{\theta} & 0 \\
\end{pmatrix}.
\]
Next, recall from Lemma~\ref{lem:PresentationTorusKnot} that the generators $a$ and $b$ of $G_0$ are related to the generators~$x_{2k}$ and~$x_{2k+1}$ by the formulas $a=\phi(x_{2k}x_{2k+1})$ and $b=\phi(x_{2k+1}^{-1}x_{2k}^{-1})^kx_{2k+1}^{-1}$.
As a consequence, the metabelian representation $\rho_\theta$ is entirely described by
\begin{align*}
\rho_\theta(a) &=
\begin{pmatrix}
t\xi^{-\theta} & 0 \\
0 & t \xi^{\theta} \\
\end{pmatrix}, \\
\rho_\theta(b) &=
\begin{pmatrix}
0 & t^{-k-1} \xi^{k \cdot \theta} \\
t^{-k} \xi^{-k \cdot \theta} & 0 \\
\end{pmatrix}.
\end{align*}
We can now work towards the description of the module $H^{2}(M_{T_{2,2k+1}};\mathbb{C}[t^{\pm 1}]_{\rho_{\theta}}^{2})$ which supports the cohomological Blanchfield pairing.
Consider the polynomials $P_{k}(t) = 1+t+t^{2}+\ldots+t^{k}$ and~$R_\eta(t)=t+t^{-1}-2 \operatorname{Re}(\eta)$, for $\eta \in S^1$.
The latter polynomial is the basic polynomial from Example~\ref{BasicPairingXiS1Real}.
Furthermore, we shall also need the following symmetric polynomial:
\[\Delta_{\theta}(t) = t^{-k} \frac{P_{2k}(t)}{{}^\R\! B_{\xi^{\theta}}(t)} = \prod_{\stackrel{i=1}{i \neq \theta}}^{k} {}^\R\! B_{\xi^{i}}(t).\]
The next lemma describes the cohomology $\mathbb{C}[t^{\pm 1}]$-module $H^{2}(M_{T_{2,2k+1}};\mathbb{C}[t^{\pm 1}]_{\rho_{\theta}}^{2})$.
\begin{proposition}
\label{prop:TwistedModuleComputation}
The module $H^{2}(M_{T_{2,2k+1}};\mathbb{C}[t^{\pm 1}]_{\rho_{\theta}}^{2})$ is isomorphic to $ \mathbb{C}[t^{\pm 1}] / (\Delta_{\theta}(t))$ and admits a generator \([v_2] \in H^{2}(M_{T_{2,2k+1}};\mathbb{C}[t^{\pm 1}]_{\rho_{\theta}}^{2})\) such that
\[\Bl^{\rho_{\theta}}(T_{2,2k+1})([v_2],[v_2]) = \frac{\frac{1}{2}(t^{2k+1} \xi^{k\theta} - t^{k+1} -t^{k}+\xi^{(k+1)\theta})(t^{-2k-1}+1)}{t^{-k}P_{2k}(t)}.\]
\end{proposition}
\begin{proof}
Since $H^{2}(M_{T_{2,2k+1}};\mathbb{C}[t^{\pm 1}]_{\rho_{\theta}}^{2})$ is equal to the quotient $\ker({{\partial_3}^{\#}}^T)/\operatorname{im}({{\partial_2}^{\#}}^T)$, we first compute $\ker({{\partial_3}^{\#}}^T)$, before studying $\operatorname{im}({{\partial_2}^{\#}}^T)$.
First of all, note that $\rho_{\theta}(\mu) = \rho_{\theta}(x_{2k+1}) = \left(\begin{smallmatrix} 0 & 1 \\ t & 0 \end{smallmatrix}\right) $.
Using this computation and looking back to the definition of $\partial_3$ (Proposition~\ref{prop:cellular-chain-complex}), we deduce that
\[\rho_{\theta}(\partial_{3}) =
\begin{pmatrix}
-\xi^{-k\theta} & t^{-k-1} & -1 & 1 \\
t^{-k} & -\xi^{k\theta} & t & -1 \\
\end{pmatrix}.
\]
The differential \(\rho_{\theta}(\partial_{3})^{\#T}\) has rank two, hence \(\ker (\rho_{\theta}(\partial_{3})^{\#T})\) is $2$-dimensional.
In fact, we claim that the kernel of \(\rho_{\theta}(\partial_{3})^{\# T}\) is freely generated by
\begin{align*}
v_{1} &= \left(t^{-k}\xi^{-k\theta}, t^{-2k-1}, \xi^{-k\theta} P_{2k}(t^{-1}), \xi^{-k\theta}P_{2k}(t^{-1})\right), \\
v_{2} &= \left(0, t^{-1}-1, \xi^{(k+1)\theta}-t^{k+1}, \xi^{-k\theta}-t^{k}\right).
\end{align*}
Indeed, it is easy to check that \(v_{1} \cdot \rho_{\theta}(\partial_{3})^{\# T} = v_{2} \cdot \rho_{\theta}(\partial_{3})^{\# T} = 0\) and that \(v_{1}\) and \(v_{2}\) are linearly independent.
We now turn to \( \operatorname{im}(\rho_{\theta}(\partial_{2})^{\# T}) \).
We first compute \(\rho_{\theta}(\partial_{2})^{\# T}\), then find a basis for~\(C^{1}(M_{T_{2,2k+1}};\mathbb{C}[t^{\pm 1}]_{\rho_{\theta}}^{2})\) and finally compute the image.
To obtain the first line of $\rho_\theta(\partial_2)$, we compute the Fox derivatives of \(r=a^{2k+1}b^2\):
\begin{align*}
\rho_{\theta}\left(\frac{\partial r}{\partial a}\right) &=
\begin{pmatrix}
P_{2k}(t\xi^{-\theta}) & 0 \\
0 & P_{2k}(t\xi^{\theta}) \\
\end{pmatrix}, \\
\rho_{\theta}\left(\frac{\partial r}{\partial b}\right) &=
\begin{pmatrix}
t^{2k+1} & t^{k} \xi^{k\theta} \\
t^{k+1} \xi^{-k\theta} & t^{2k+1} \\
\end{pmatrix}.
\end{align*}
In order to obtain the second line of $\rho_\theta(\partial_{2})$, we first compute
\begin{align*}
\rho_{\theta}\left(\frac{\partial (a^{k}b)}{\partial a}\right) &=
\begin{pmatrix}
P_{k-1}(t\xi^{-\theta}) & 0 \\
0& P_{k-1}(t\xi^{\theta}) \\
\end{pmatrix}, \\
\rho_{\theta}\left(\frac{\partial (a^{k}b)}{\partial b}\right) &=
\begin{pmatrix}
t^{k}\xi^{-k\theta} & 0 \\
0 & t^{k}\xi^{k\theta} \\
\end{pmatrix}, \\
\rho_{\theta}\left(\frac{\partial (a^{k}b)^{2k+1}}{\partial a}\right) &=
\begin{pmatrix}
P_{k}(t^{-1}) & t^{-1}P_{k-1}(t^{-1}) \\
P_{k-1}(t^{-1}) & P_{k}(t^{-1})
\end{pmatrix} \cdot
\begin{pmatrix}
P_{k-1}(t\xi^{-\theta}) & 0 \\
0 & P_{k-1}(t\xi^{\theta}) \\
\end{pmatrix} \\
&=
\begin{pmatrix}
P_{k}(t^{-1})P_{k-1}(t\xi^{-\theta}) & t^{-1}P_{k-1}(t^{-1})P_{k-1}(t\xi^{\theta}) \\
P_{k-1}(t^{-1})P_{k-1}(t\xi^{-\theta}) & P_{k}(t^{-1})P_{k-1}(t\xi^{\theta}) \\
\end{pmatrix}, \\
\rho\left(\frac{\partial (a^{k}b)^{2k+1}}{\partial b}\right) &=
\begin{pmatrix}
P_{k}(t^{-1}) & t^{-1}P_{k-1}(t^{-1}) \\
P_{k-1}(t^{-1}) & P_{k}(t^{-1})
\end{pmatrix} \cdot
\begin{pmatrix}
t^{k}\xi^{-k\theta} & 0 \\
0 & t^{k} \xi^{k\theta} \\
\end{pmatrix} \\
&=
\begin{pmatrix}
P_{k}(t)\xi^{-k\theta }& P_{k-1}(t)\xi^{k\theta} \\
t P_{k-1}(t) \xi^{-k\theta} & P_{k}(t) \xi^{k\theta} \\
\end{pmatrix}.
\end{align*}
As a consequence, the second line of $\partial_2$ is given by
\begin{align*}
\rho_{\theta}\left(\frac{\partial \lambda_{0}}{\partial a}\right) &=
\begin{pmatrix}
t^{-k}P_{k-1}(t\xi^{-\theta})P_{2k}(t) & t^{-k}P_{k-1}(t\xi^{\theta})P_{2k}(t) + t^{-k-1}P_{2k}(t\xi^{\theta}) \\
t^{-k+1}P_{k-1}(t\xi^{-\theta})P_{2k}(t) + t^{-k}P_{2k}(t\xi^{-\theta}) & t^{-k}P_{k-1}(t\xi^{\theta})P_{2k}(t) \\
\end{pmatrix} \\
&=
\begin{pmatrix}
t^{-k}P_{k-1}(t\xi^{-\theta})P_{2k}(t) & \frac{P_{2k}(t)}{t\xi^{\theta}-1} (\xi^{k\theta}-t^{-k-1})\\
\frac{P_{2k}(t)}{t\xi^{-\theta}-1}(t\xi^{-k\theta}-t^{-k}) & t^{-k}P_{k-1}(t\xi^{\theta})P_{2k}(t) \\
\end{pmatrix}, \\
\rho_{\theta}\left(\frac{\partial \lambda_{0}}{\partial b}\right) &=
\begin{pmatrix}
\xi^{-k\theta} P_{2k}(t) & \xi^{k\theta} P_{2k}(t) \\
t \xi^{-k\theta} P_{2k}(t) & \xi^{k\theta} P_{2k}(t) \\
\end{pmatrix}.
\end{align*}
This completes the first step of our computation of \( \operatorname{im}(\rho_{\theta}(\partial_{2})^{\# T}) \).
To carry out the second step, consider the following set of vectors of \(C^{1}(M_{T_{2,2k+1}};\mathbb{C}[t^{\pm 1}]_{\rho_{\theta}}^{2})\):
\begin{align}
\label{eq:Z1}
Z_{1} &= \left(\frac{\xi^{-(k-1)\theta}t^{k+1}}{\xi^{\theta}-\xi^{-\theta}}, \frac{\xi^{-\theta}}{\xi^{\theta}-\xi^{-\theta}}, 0, -\frac{t^{2k+1}\xi^{\theta}P_{2k}(t^{-1}\xi^{\theta})}{\xi^{\theta}-\xi^{-\theta}}\right), \\
Z_{2} &= \left((t^{-1}\xi^{\theta}-1)\xi^{-k\theta}t^{k+1}, t^{-1}\xi^{-\theta}-1, 0, t^{2k+1}-1\right), \nonumber \\
Z_{3} &= \left(0, t^{-k-2}\xi^{(k-1)\theta}-\xi^{k\theta}t^{-k-1}, 1, -\xi^{k\theta}t^{-k-1}\right), \nonumber\\
Z_{4} &= \left(0,0,0,1\right) \nonumber.
\end{align}
We claim that this collection of vectors yields a basis of \(C^{1}(M_{T_{2,2k+1}};\mathbb{C}[t^{\pm 1}]_{\rho_{\theta}}^{2})\).
Indeed, if~\(Z\) denotes the matrix whose rows are \(Z_{1}, Z_{2}, Z_{3}\) and \(Z_{4}\), then \(\det (Z) = -t^{k+1}\xi^{-k\theta}\): this determinant can be computed by first expanding along the fourth row and then expanding along the third column.
Finally, we carry out the third and last step: using this basis of vectors, we compute the image of $\rho_\theta(\partial_2)^{\# T} $ by observing that
\begin{align}
\label{eq:Z1v2}
Z_{1} \cdot \rho_{\theta}(\partial_{2})^{\# T} &= t^{-k+1} \cdot\Delta_{\theta}(t^{-1}) \cdot v_{2}, \\
Z_{2} \cdot \rho_{\theta}(\partial_{2})^{\# T} &= 0, \nonumber \\
Z_{3} \cdot \rho_{\theta}(\partial_{2})^{\# T} &= 0, \nonumber \\
Z_{4} \cdot \rho_{\theta}(\partial_{2})^{\# T} &= v_{1}. \nonumber
\end{align}
We therefore deduce that the twisted homology $\mathbb{C}[t^{\pm 1}]$-module $H^{2}(M_{T_{2,2k+1}};\mathbb{C}[t^{\pm 1}]_{\rho_{\theta}}^{2})$ is cyclic with order \(\Delta_{\theta}(t)\).
This concludes the proof of the first assertion of the proposition.
In order to compute the twisted Blanchfield pairing, we first compute \(\rho_\theta(\Phi)\) using \eqref{eq:Phi} and then compute the Blanchfield pairing.
We start with the calculation of the coefficients of the first line of $\rho_\theta(\Phi)$.
Looking back to~\eqref{eq:Phi}, we see that each of the two blocs consists of a difference of two expressions.
We compute each of these terms separately. First, note that we have
\begin{align*}
\rho_{\theta}\left(\mu^{2k+1} \cdot \frac{\partial \mu^{-2k-1}}{\partial a}\right) &=
\begin{pmatrix}
0 & t^{k} \\
t^{k+1} & 0 \\
\end{pmatrix}
\begin{pmatrix}
P_{k}(t^{-1})P_{k-1}(t\xi^{-\theta}) & t^{-1}P_{k-1}(t^{-1})P_{k-1}(t\xi^{\theta}) \\
P_{k-1}(t^{-1})P_{k-1}(t\xi^{-\theta}) & P_{k}(t^{-1})P_{k-1}(t\xi^{\theta}) \\
\end{pmatrix} \\
&=
\begin{pmatrix}
t^{k}P_{k-1}(t^{-1})P_{k-1}(t\xi^{-\theta}) & t^{k}P_{k}(t^{-1})P_{k-1}(t\xi^{\theta}) \\
t^{k+1}P_{k}(t^{-1})P_{k-1}(t\xi^{-\theta}) & t^{k}P_{k-1}(t^{-1})P_{k-1}(t\xi^{\theta}) \\
\end{pmatrix}, \\
\rho_{\theta}\left(\mu^{2k+1} \cdot \frac{\partial \mu^{-2k-1}}{\partial b}\right) &=
\begin{pmatrix}
0 & t^{k} \\
t^{k+1} & 0 \\
\end{pmatrix}
\begin{pmatrix}
P_{k}(t)\xi^{-k\theta }& P_{k-1}(t)\xi^{k\theta} \\
t P_{k-1}(t) \xi^{-k\theta} & P_{k}(t) \xi^{k\theta} \\
\end{pmatrix} \\
&=
\begin{pmatrix}
t^{k+1}\xi^{-k\theta}P_{k-1}(t) & t^{k} \xi^{k\theta}P_{k}(t) \\
t^{k+1} \xi^{-k\theta}P_{k}(t) & t^{k+1} \xi^{k\theta}P_{k-1}(t) \\
\end{pmatrix}.
\end{align*}
The computation of the upper left block of $\rho_\theta(\Phi)$ also requires us to compute
\begin{align*}
\frac{\partial (\mu^{-2k} \cdot a^{k})}{\partial a} &= \frac{\partial{\mu^{-2k}}}{\partial a} + \mu^{-2k} \cdot \frac{\partial{a^{k}}}{\partial a} \\
&= (1 + \mu^{-1} + \mu^{-2} + \cdots + \mu^{-2k+1}) \cdot \frac{\partial (a^{k} b)}{\partial a} + \mu^{-2k} \cdot (1+a+a^{2}+\cdots+a^{k-1}) \\
&= (1 + \mu^{-1}) \cdot (1 + \mu^{-2} + \cdots (\mu^{-2})^{k-1}) \cdot \frac{\partial a^{k}b}{\partial a} + \mu^{-2k} \cdot (1+a+a^{2}+\cdots+a^{k-1}) \\
\rho_{\theta}\left(\frac{\partial(\mu^{-2k} \cdot a^{k})}{\partial a}\right) &=
\begin{pmatrix}
1 & t^{-1} \\
1 & 1 \\
\end{pmatrix} \cdot
\begin{pmatrix}
P_{k-1}(t^{-1}) & 0 \\
0 & P_{k-1}(t^{-1}) \\
\end{pmatrix} \cdot
\begin{pmatrix}
P_{k-1}(t\xi^{-\theta}) & 0 \\
0 & P_{k-1}(t\xi^{\theta}) \\
\end{pmatrix} \\
&+
\begin{pmatrix}
t^{-k} & 0 \\
0 & t^{-k} \\
\end{pmatrix} \cdot
\begin{pmatrix}
P_{k-1}(t\xi^{-\theta}) & 0 \\
0 & p_{k-1}(t\xi^{\theta}) \\
\end{pmatrix} \\
&=
\begin{pmatrix}
P_{k}(t^{-1}) P_{k-1}(t\xi^{-\theta}) & t^{-1}P_{k-1}(t^{-1})P_{k-1}(t\xi^{\theta}) \\
P_{k-1}(t^{-1}) P_{k-1}(t\xi^{-\theta}) & P_{k}(t^{-1})P_{k-1}(t\xi^{\theta})\\
\end{pmatrix},
\end{align*}
and, similarly, the upper right block of $\rho_\theta(\Phi)$ requires that we compute
\begin{align*}
\frac{\partial (\mu^{-2k} \cdot a^{k})}{\partial b} &= \frac{\partial \mu^{-2k}}{\partial b} = (1+\mu^{-1}) \cdot (1+\mu^{-2}+\cdots+(\mu^{-2})^{k-1}) \cdot \frac{\partial a^{k}b}{\partial b}, \\
\rho_{\theta}\left(\frac{\partial (\mu^{-2k} \cdot a^{k})}{\partial b}\right) &=
\begin{pmatrix}
1 & t^{-1} \\
1 & 1 \\
\end{pmatrix} \cdot
\begin{pmatrix}
P_{k-1}(t^{-1}) & 0 \\
0 & P_{k-1}(t^{-1}) \\
\end{pmatrix} \cdot
\begin{pmatrix}
t^{k} \xi^{-k\theta} & 0 \\
0 & t^{k} \xi^{k\theta} \\
\end{pmatrix} \\
&=
\begin{pmatrix}
t^{k}\xi^{-k\theta}P_{k-1}(t^{-1}) & t^{k-1} \xi^{k\theta} P_{k-1}(t^{-1}) \\
t^{k}\xi^{-k\theta}P_{k-1}(t^{-1}) & t^{k}\xi^{k\theta}P_{k-1}(t^{-1})
\end{pmatrix},
\end{align*}
Consequently, using the two above sequences of computations, we have
\begin{align*}
\rho_{\theta}\left(a^{-k} \cdot \mu^{2k} \cdot \frac{\partial(\mu^{-2k} \cdot a^{k})}{\partial a}\right) &=
\begin{pmatrix}
\xi^{k\theta} & 0 \\
0 & \xi^{-k\theta} \\
\end{pmatrix}
\begin{pmatrix}
P_{k}(t^{-1}) P_{k-1}(t\xi^{-\theta}) & t^{-1}P_{k-1}(t^{-1})P_{k-1}(t\xi^{\theta}) \\
P_{k-1}(t^{-1}) P_{k-1}(t\xi^{-\theta}) & P_{k}(t^{-1})P_{k-1}(t\xi^{\theta})\\
\end{pmatrix} \\
&=
\begin{pmatrix}
\xi^{k\theta} P_{k}(t^{-1})P_{k-1}(t\xi^{-\theta}) & \xi^{k\theta}t^{-1}P_{k-1}(t^{-1}) P_{k-1}(t\xi^{\theta})\\
\xi^{-k\theta} P_{k-1}(t^{-1})P_{k-1}(t\xi^{-\theta}) & \xi^{-k\theta}P_{k}(t^{-1})P_{k-1}(t\xi^{\theta}) \\
\end{pmatrix}, \\
\rho_{\theta}\left(a^{-k} \cdot \mu^{2k} \cdot \frac{\partial(\mu^{-2k} \cdot a^{k})}{\partial b}\right) &=
\begin{pmatrix}
\xi^{k\theta} & 0 \\
0 & \xi^{-k\theta} \\
\end{pmatrix}
\begin{pmatrix}
t^{k}\xi^{-k\theta}P_{k-1}(t^{-1}) & t^{k-1} \xi^{k\theta} P_{k-1}(t^{-1}) \\
t^{k}\xi^{-k\theta}P_{k-1}(t^{-1}) & t^{k}\xi^{k\theta}P_{k-1}(t^{-1}) \\
\end{pmatrix} \\
&=
\begin{pmatrix}
t^{k}P_{k-1}(t^{-1}) & t^{k-1}\xi^{2k\theta}P_{k-1}(t^{-1}) \\
t^{k}\xi^{-2k\theta}P_{k-1}(t^{-1}) & t^{k}P_{k-1}(t^{-1}) \\
\end{pmatrix}.
\end{align*}
Looking at~\eqref{eq:Phi}, assembling these computations and taking the appropriate differences provides an explicit understanding of the first row of $\rho_\theta(\Phi)$.
Next, we compute entries in the second row of $\rho_\theta(\Phi)$:
\begin{align*}
\rho_{\theta}\left(-\mu^{-1}\frac{\partial \mu}{\partial a}\right) &=
\begin{pmatrix}
-P_{k-1}(t\xi^{-\theta}) & 0 \\
0 & -P_{k-1}(t\xi^{\theta}) \\ \end{pmatrix}, \\
\rho_{\theta}\left(-\mu^{-1}\frac{\partial \mu}{\partial b}\right) &=
\begin{pmatrix}
-t^{k} \xi^{-k\theta} & 0 \\
0 & -t^{k}\xi^{k\theta}
\end{pmatrix}.
\end{align*}
Finally, using $\rho_\theta(\Phi)$ and~\eqref{eq:ChainBlanchfieldFormula}, we can compute the cohomological twisted Blanchfield pairing of~$M_{T_{2,2k+1}}$.
In more details, we know from~\eqref{eq:ChainBlanchfieldFormula} that if $v \in Z^2(N;\L{\F}^d_{\rho_\theta})$, then
\[\Bl^{\rho_\theta}(T_{2,2k+1})([v],[v])=\frac{1}{s} \left(v \cdot \rho_\theta(\Phi) \cdot Z^{\# T}\right)^{\# T},\]
where $Z \in C^1(N;\L{\F}^d_{\rho_\theta})$ satisfies $Z\beta(\partial^2)=sv$ for some~$s \in~\mathbb{C}[t^{\pm 1}] \setminus \lbrace 0 \rbrace$.
In our case, we take $v=v_2$ and observe that $ Z=Z_1$ is described in~\eqref{eq:Z1} and $s=t^{-k+1}\Delta_\theta(t^{-1})=t^{-k+1}\Delta_\theta(t)$ (recall~\eqref{eq:Z1v2}). Therefore, we start by computing
\begin{align*}
Z_{1} \cdot \rho_{\theta}(\Phi)^{\# T} \cdot v_{2}^{\# T}
&= \frac{1}{\xi^{\theta}-\xi^{-\theta}} \cdot \left[\xi^{-(k-1)\theta} t^{k+1} P_{k-1}(t^{-1}\xi^{\theta})(1-\xi^{k\theta}t^{k}) + \xi^{-\theta} P_{k-1}(t^{-1}\xi^{-\theta}) (1-\xi^{k\theta}t^{k+1})\right] -\\
&- \frac{1}{\xi^{\theta}-\xi^{-\theta}}t^{k+1}\xi^{-(k-1)\theta}(1-\xi^{k\theta}t^{k})P_{2k}(t^{-1}\xi^{\theta}) = \\
&= \frac{1}{\xi^{\theta}-\xi^{-\theta}} \cdot \left[\xi^{-(k-1)\theta} t^{k+1} P_{k-1}(t^{-1}\xi^{\theta})(1-\xi^{k\theta}t^{k}) + \xi^{-\theta} P_{k-1}(t^{-1}\xi^{-\theta}) (1-\xi^{k\theta}t^{k+1})\right] -\\
&- \frac{1}{\xi^{\theta}-\xi^{-\theta}} \cdot t^{k+1}\xi^{-(k-1)\theta}(1-\xi^{k\theta}t^{k})\left[P_{k-1}(t^{-1}\xi^{\theta}) + t^{-k}\xi^{k\theta}P_{k}(t^{-1}\xi^{\theta})\right] = \\
&= \frac{1}{\xi^{\theta}-\xi^{-\theta}}\left[\xi^{-\theta}P_{k-1}(t^{-1}\xi^{-\theta})(1-\xi^{-(k+1)\theta}t^{k+1}) - t \cdot \xi^{\theta} (1-t\xi^{\theta}) P_{k-1}(t\xi^{\theta})P_{k}(t^{-1}\xi^{\theta})\right] = \\
&= \frac{P_{k-1}(t\xi^{\theta}) P_{k}(t\xi^{-\theta})}{\xi^{\theta} - \xi^{-\theta}} \left[t^{-k+1} \xi^{-k\theta} (1-t\xi^{-\theta}) - t^{-k+1}\xi^{(k+1)\theta}(1-t\xi^{\theta})\right] = \\
&= \xi^{(k+1)\theta}t^{-k+2}P_{k-1}(t\xi^{\theta})P_{k}(t\xi^{-\theta}).
\end{align*}
To facilitate computations, we will now arrange that both the numerator and denominator of $\Bl^{\rho_\theta}(T_{2,2k+1})$ are symmetric. To that effect, we observe that if $b$ is a symmetric polynomial and $(a/b)^\#=a/b$ in $\mathbb{F}(t)/\L{\F}$, then $\frac{a}{b}=\frac{\frac{1}{2}(a+a^\#)}{b}$ in $\mathbb{F}(t)/\L{\F}$. Applying this remark to the symmetric polynomial $b=\Delta_\theta(t)$, using $\Bl^{\rho_\theta}$ as a shorthand for $\Bl^{\rho_\theta}(T_{2,2k+1})$ and recalling from above that $\Bl^{\rho_\theta}([v_2],[v_2])=\frac{1}{s}Z_1 \cdot \rho_\theta(\Phi)^{\#T} \cdot v_2^{\#T}$ with $s=t^{-k+1}\Delta_\theta(t)$, we obtain
\begin{align*}
\Bl^{\rho_\theta}([v_{2}],[v_{2}]) &= \frac{\frac{1}{2}\xi^{(k+1)\theta}t P_{k-1}(t\xi^{\theta})P_{k}(t\xi^{-\theta})}{\Delta_{\theta}(t)} + \frac{\frac{1}{2}\xi^{-(k+1)\theta}t^{-1}P_{k-1}(t^{-1}\xi^{-\theta})P_{k}(t^{-1}\xi^{\theta})}{\Delta_{\theta}(t)} \\
&= \frac{\frac{1}{2}\xi^{(k+1)\theta}t P_{k-1}(t\xi^{\theta})P_{k}(t\xi^{-\theta})(t^{-2k-1}+1)}{\Delta_{\theta}(t)} \\
&= \frac{\frac{1}{2}(t^{2k+1} \xi^{k\theta} - t^{k+1} -t^{k}+\xi^{(k+1)\theta})(t^{-2k-1}+1)}{t^{-k}P_{2k}(t)}.
\end{align*}
This concludes the computation of the twisted Blanchfield pairing on the generator of the twisted cohomology $\mathbb{C}[t^{\pm 1}]$-module $H^{2}(M_{T_{2,2k+1}};\mathbb{C}[t^{\pm 1}]_{\rho_{\theta}}^{2})$ and thus the proof of the proposition.
\end{proof}
\subsection{The isometry type of $\Bl^{\rho_{\theta}}(T_{2,2k+1})$.}
\label{sub:MetabelianSignaturesTorusKnot}
The aim of this subsection is to determine the isometry type of the twisted Blanchfield forms~$\Bl^{\rho_\theta}(T_{2,2k+1})$, where $\rho_\theta=\alpha(2,\chi_\theta)$. The following proposition implicitly contains the values of the twisted signature function $\sigma_{T_{2,2k+1},\rho_\theta} \colon S^1 \to \mathbb{Z}$.
\begin{proposition}
\label{prop:DecompoTwistedBlanchfield}
Set \(\xi = \exp\left(\frac{2 \pi i}{2k+1}\right)\).
For any \(k>0\) and \(1 \leq \theta \leq k\), there exists an isometry
\begin{align*}
\Bl^{\rho_{\theta}}(T_{2,2k+1}) \cong \lambda_{\theta}^{even} \oplus \lambda_{\theta}^{odd},
\end{align*}
where the linking forms $\lambda_{\theta}^{even}$ and $\lambda_{\theta}^{odd}$ are as follows:
\begin{align*}
\lambda_{\theta}^{odd} &= \bigoplus_{\stackrel{1 \leq e \leq k}{2 \nmid \theta + e}} \left(\mathfrak{e}(1,0,1, \xi^{e},\mathbb{C}) \oplus \mathfrak{e}(1,0,-1,\xi^{-e},\mathbb{C})\right),
\end{align*}
and
\begin{align*}
\lambda_{\theta}^{even} &= \bigoplus_{\stackrel{1 \leq e < \theta}{2 \mid \theta+e}}\left(\mathfrak{e}(1,0,1,\xi^{e}, \mathbb{C}) \oplus \mathfrak{e}(1,0,-1,\xi^{-e},\mathbb{C})\right) \oplus \\
&\bigoplus_{\stackrel{\theta < e \leq k}{2 \mid \theta+e}}\left(\mathfrak{e}(1,0,-1,\xi^{e}, \mathbb{C}) \oplus \mathfrak{e}(1,0,1,\xi^{-e},\mathbb{C})\right).
\end{align*}
\end{proposition}
\begin{proof}
Using Proposition~\ref{prop:TwistedModuleComputation}, we can choose a generator $x$ of the cyclic module $H^{2}(M_{T_{2,2k+1}};\mathbb{C}[t^{\pm 1}]_{\rho_{\theta}}^{2})$ so that $\Bl^{\rho_{\theta}}(T_{2,2k+1})(x,x) =F(t)$, where
\[F(t) = \frac{\frac{1}{2}(t^{2k+1} \xi^{k\theta} - t^{k+1} -t^{k}+\xi^{(k+1)\theta})(t^{-2k-1} + 1)}{t^{-k}P_{2k}(t)}.\]
Since $\Delta_\theta(t)=\prod_{\stackrel{e=1}{e \neq \theta}}^{k} {}^\R\! B_{\xi^{e}}(t)$, we know that the isometry type of $\Bl^{\rho_{\theta}}(T_{2,2k+1})$ will involve a direct sum of the basic pairings $ \mathfrak{e}(1,0,\pm 1,\xi^{e},\mathbb{C})$. In order to determine the correct signs, and prove the proposition, we will apply the third item of Proposition~\ref{prop:cyclic_classif}. This proposition indicates that the aforementioned signs depend on whether $F(t)(t-\xi^e)$ is \(\xi^{e}\)-positive or \(\xi^{e}\)-negative, for \(1 \leq e \leq 2k\) and \(e \neq \theta, 2k+1-\theta\). Notice that $F(t) = \frac{G(t)}{H_{e}(t)} \cdot \frac{1}{{}^\R\! B_{\xi^{e}}(t)}$, where, for \(1 \leq e \leq 2k\), we set
\begin{align*}
G(t) &= \frac{1}{2}(t^{2k+1} \xi^{k\theta} - t^{k+1} -t^{k}+\xi^{(k+1)\theta})(t^{-2k-1} + 1), \\
H_{e}(t) &= \frac{t^{-k}P_{2k}(t)}{{}^\R\! B_{\xi^{e}}(t)} = \prod_{\stackrel{i=1}{i \neq e}}^{k} {}^\R\! B_{\xi^{i}}(t).
\end{align*}
Reformulating, the proposition will immediately follow once we determine the $e$ for which the following function is $\xi^e$-positive or $\xi^e$-negative:
\begin{equation}\label{eq:twisted-Bl-torus-knot}
F(t) \cdot (t-\xi^{e}) = (-1)^{\theta+e}\left[(-1)^{\theta+1}G(t)\right] \cdot \left[\frac{(-1)^{e-1}}{H_{e}(t)(1-t^{-1}\xi^{-e})}\right].
\end{equation}
In Lemma~\ref{lem:HxiPositive} below, we show that $\frac{(-1)^{e-1}}{H_{e}(t)(1-t^{-1}\xi^{-e})}$ is $\xi^e$-positive, while in Lemma~\ref{lem:GxiPositive}, we study the sign of $(-1)^{\theta+1}G(\xi^e)$.
\begin{lemma}
\label{lem:HxiPositive}
For any \(1 \leq e \leq 2k\), the following function is \(\xi^{e}\)-positive:
\[\frac{(-1)^{e-1}}{H_{e}(t)(1-t^{-1}\xi^{-e})}.\]
\end{lemma}
\begin{proof}
Notice that \(\frac{1}{1-t^{-1}\xi^{-e}}\) is \(\xi^{e}\)-positive if \(1 \leq e \leq k\) and \(\xi^{e}\)-negative if \(k+1 \leq e \leq 2k\).
As a consequence, the lemma boils down to determining the sign of $(-1)^{e-1}H_e(\xi^e)$. We first check this when \(1 \leq e \leq k\). For any \(1 \leq i \leq k\) such that \(e \neq i\), we have
\[{}^\R\! B_{\xi^{i}}(\xi^{e}) = 2\re(\xi^{e}) - 2 \re(\xi^{i})
\begin{cases}
>0 & \text{if } e<i, \\
<0 & \text{if } e>i. \\
\end{cases}
\]
Combining these inequalities with the definition of $H_e(t)$, we obtain
\[(-1)^{e-1}H_{e}(\xi^{e}) = \underbrace{(-1)^{e-1}\prod_{i=1}^{e-1}{}^\R\! B_{\xi^{i}}(\xi^{e})}_{>0} \cdot \underbrace{\prod_{i=e+1}^{k}{}^\R\! B_{\xi^{i}}(\xi^{e})}_{>0}>0.\]
The case \(k+1 \leq e \leq 2k\) can be reduced the previous one.
Indeed, notice that \(H_{e}(t) = H_{2k+1-e}(t)\) and \({}^\R\! B_{\xi^{i}}(\xi^{e}) = {}^\R\! B_{\xi^{i}}(\xi^{2k+1-e})\), because \({}^\R\! B_{\xi^{i}}(t)\) is a real polynomial. This implies that
\[(-1)^{e-1}H_{e}(\xi^{e}) = (-1)^{2k+1-e}H_{2k+1-e}(\xi^{2k+1-e})<0.\]
This concludes the proof of the lemma.
\end{proof}
Let us now study the sign of the numerator of~\eqref{eq:twisted-Bl-torus-knot}.
Notice first that for \(1 \leq e \leq k\), we have
\[G(\xi^{e}) = G(\xi^{-e}) = G(\xi^{2k+1-e}),\]
and it is therefore sufficient to determine the sign of \(G(\xi^{e})\) for \(1 \leq e \leq k\).
\begin{lemma}\label{lem:GxiPositive}
For \(1 \leq e \leq k\) such that \(e \neq \theta\) we have
\[(-1)^{\theta+1}G(\xi^{e})
\begin{cases}
>0, & \text{ if } 1 \leq e < \theta \leq k \text{ and } 2 \mid \theta + e, \\
<0, & \text{ if } 1 \leq \theta < e \leq k \text{ and } 2 \mid \theta+e, \\
<0, & \text{ if } 2 \nmid \theta+e.
\end{cases}
\]
\end{lemma}
\begin{proof}
Using successively the definition of $G(t)$, the fact that $\xi=\operatorname{exp}(\frac{2\pi i}{2k+1})$ and the identities
\[\cos (x) - \cos (y) = -2 \sin\left(\frac{x+y}{2}\right) \sin\left(\frac{x-y}{2}\right),\]
we obtain the following equality:
\begin{align*}
G(\xi^{e}) &= 2 (\re(\xi^{k\theta})-\re(\xi^{ke})) \\
&= 2 \left(\cos\left(\frac{2\pi k\theta}{2k+1}\right) - \cos\left(\frac{2\pi ke}{2k+1}\right)\right) \\
&= -4 \sin\left(\frac{\pi k (\theta+e)}{2k+1}\right) \sin\left(\frac{\pi k(\theta-e)}{2k+1}\right) \\
&= -4 \sin\left(\frac{(\theta+e)\pi}{2} - \frac{(\theta+e)\pi}{4k+2}\right) \sin\left(\frac{(\theta-e)\pi}{2}-\frac{(\theta-e)\pi}{4k+2}\right) = (\ast).
\end{align*}
As consequence, we consider two cases depending on the parity of \(\theta + e\).
First, suppose that \(\theta+e\) is even.
Using the identity $\sin(y+n\pi)=(-1)^{n}\sin(y)$, we deduce that previous expression becomes
\begin{align*}
(\ast) &= 4(-1)^{\theta+1}\sin\left(\frac{(\theta+e)\pi}{4k+2}\right) \sin\left(\frac{(\theta-e)\pi}{4k+2}\right).
\end{align*}
Since $\sin\left(\frac{(\theta+e)\pi}{4k+2}\right)>0$ and \(\sin\left(\frac{(\theta-e)\pi}{4k+2}\right)>0\) if \(\theta>e\), we deduce that for \(\theta+e\) even
\[(-1)^{\theta+1}(\ast)
\begin{cases}
>0, & 1 \leq e < \theta \leq k, \\
<0, & 1 \leq \theta < e \leq k. \\
\end{cases}
\]
Next, we suppose that \(\theta+e\) is odd.
Using the identities \(\sin(y+\pi n) = (-1)^{n} \sin(y)\) and \(\sin(y + \pi/2) = \cos(y)\), we obtain
\begin{align*}
(\ast) &= 4(-1)^{\theta}\cos\left(\frac{(\theta+e)\pi}{4k+2}\right) \cos\left(\frac{(\theta-e)\pi}{4k+2}\right).
\end{align*}
Since \(1 \leq e \leq k\), we have $\cos\left(\frac{(\theta+e)\pi}{4k+2}\right)>0$ and $\cos\left(\frac{(\theta-e)\pi}{4k+2}\right)>0$.
Consequently, if \(\theta+e\) is odd, \((-1)^{\theta+1}(\ast)<0\).
This concludes the proof of the lemma.
\end{proof}
The proof of Proposition~\ref{prop:DecompoTwistedBlanchfield} is now concluded by using Lemmas~\ref{lem:HxiPositive} and~\ref{lem:GxiPositive} as well as~\eqref{eq:twisted-Bl-torus-knot} and the remarks which were made at the beginning of the proof.
\end{proof}
\subsection{Metabelian signatures and cabling}
\label{sub:HeddenKirkLivingston}
In this subsection, we use Proposition~\ref{prop:DecompoTwistedBlanchfield} and Subsection~\ref{sub:SatelliteMetabelian} in order to compute metabelian signatures of a concrete linear combination of algebraic knots.
Throughout this subsection, for an integer \(\ell>0\) we set~\(\xi_{\ell} = \exp\left(2\pi i/\ell\right)\).
\medbreak
Given a knot $K$, use $K_{2,d}$ to denote its $(2,d)$-cable.
In other words, $K_{2,d}$ is the satellite knot with pattern the $(2,d)$ torus knot $T_{2,d}$, companion $K$ and infection curve $\eta = a$ (using notation from Section~\ref{sub:ChainComplex})~depicted in Figure~\ref{fig:diagram-T_2_2k+1}.
The preimage of $\eta$ in the 2-fold branched cover $L_2(T_{2,d})$ consists of two curves $\widetilde{\eta}_1,\widetilde{\eta}_2$.
Denote by~$\mu_\eta$ and~$\lambda_\eta$ the meridian and longitude of $\eta$ and write $\widetilde{\mu}_i$ and $\widetilde{\lambda}_i$ for some meridian-longitude pair of the boundary of the tubular neighboorhood of $\widetilde{U}_i \subset L_2(T_{2,d})$ for $i=1,2$.
Note that $\widetilde{\mu}_1$ and~$\widetilde{\mu}_2$ vanish in $H_1(L_2(T_{2,d});\mathbb{Z})=\mathbb{Z}_d$, while the lift $\widetilde{\lambda}_1$ generates $H_1(L_2(T_{2,d});\mathbb{Z})$ and $\widetilde{\lambda}_2=-\widetilde{\lambda}_1$ in $H_1(L_2(T_{2,d});\mathbb{Z})$.
We refer to~\cite[Section 2]{HeddenKirkLivingston} for further details as well as to~\cite[Lemma 2.2, item 3)]{HeddenKirkLivingston} for the proof of the following proposition.
\begin{proposition}
\label{prop:HKLCable}
Let $\ell$ be an odd prime.
To any character $\chi \colon H_1(L_2(K_{2,d});\mathbb{Z}) \to \mathbb{Z}_{\ell}$, one can associate an integer $\theta$ modulo $\ell$ by the condition $\chi(\widetilde{\lambda}_1)=\xi_{\ell}^{\theta}$.
This character is denoted $\chi_{\theta}$.
In particular, this sets up a bijective correspondence between $\mathbb{Z}_{\ell}$-valued characters on $H_1(L_2(K_{2,d});\mathbb{Z})$ and on $ H_1(L_2(T_{2,d});\mathbb{Z})$.
\end{proposition}
Given an oriented knot $K$, as is customary in knot concordance, we use $-K$ to denote the mirror image of $K$ with the reversed orientation, i.e. $-K=\overline{K}^r$. The next remark, which follows~\cite[Lemma 3.2]{HeddenKirkLivingston}, describes the characters on $H_1(L_2(-K);\mathbb{Z})$.
\begin{remark}
\label{rem:ReverseMirror}
By definition of the reverse mirror image, there is an orientation reversing homeomorphism $h \colon L_2(K) \to L_2(-K)$ and, from now on, it will be understood that we identify the characters on $H_1(L_2(K);\mathbb{Z})$ and with those on $H_1(L_2(-K);\mathbb{Z})$ via this homeomorphism. With this convention, the same proof as in Proposition~\ref{prop:LocallyConstant} shows that $\Bl_{\alpha(2,\chi)}(-K)=-\Bl_{\alpha(2,\chi)}(K)$. Note however that the character we fixed on $-K$ is \emph{not} the one obtained by combining the second and third items of Proposition~\ref{prop:LocallyConstant}.
\end{remark}
Denote by \(T_{\ell,d;r,s}\) the \((r,s)\)-cable of the \((\ell,d)\)-torus knot.
From now on, we consider the following algebraic knot which was thoroughly studied by Hedden, Kirk and Livingston~\cite{HeddenKirkLivingston}:
\begin{align}
\label{eq:HKLKnot}
K &= T_{2,3;2,13} \# T_{2,15} \# -T_{2,3;2,15} \# -T_{2,13} \\
&= K_{1} \# K_{2} \# K_{3} \# K_{4}. \nonumber
\end{align}
Our goal is to study metabelian Blanchfield pairings of the form $\Bl_{\alpha(2,\chi)}(K)$.
We start by discussing characters on $H_1(L_2(K);\mathbb{Z})$.
Using the decomposition of $K$ as $K_{1} \# K_{2} \# K_{3} \# K_{4}$, we obtain the direct sum decomposition $H_1(L_2(K);\mathbb{Z})=H_1(L_2(K_1);\mathbb{Z}) \oplus \ldots \oplus H_1(L_2(K_4);\mathbb{Z})$.
Furthermore, by Proposition~\ref{prop:HKLCable}, we have the isomorphisms $H_1(L_2(K_i);\mathbb{Z})\cong H_1(L_2(T_{2,13});\mathbb{Z})$ for $i=1,4$ and $H_1(L_2(K_i);\mathbb{Z}) \cong H_1(L_2(T_{2,15});\mathbb{Z})$ for $i=2,3$.
Since these isomorphisms identify the corresponding characters, we have obtained the following lemma.
\begin{lemma}
\label{lem:CharactersOfK}
Let $\ell$ be an odd prime. For the knot $K$ described in~\eqref{eq:HKLKnot}, every character $\chi \colon H_1(L_2(K);\mathbb{Z}) \to \mathbb{Z}_{\ell}$ can be written as $\chi= \chi_{1}+\chi_{2}+\chi_{3}+\chi_{4}$ with $\chi_i:=\chi_{\theta_i} \colon H_1(L_2(K_i);\mathbb{Z}) \to~\mathbb{Z}_{\ell}$, where \(0 \leq \theta_{1},\theta_{4}\leq 12\) and \(0 \leq \theta_{2},\theta_{3} \leq 14\).
\end{lemma}
\begin{remark}
\label{rem:HalfCharacters}
To study these metabelian Blanchfield signatures, it is enough to consider the cases \(0 \leq \theta_{1},\theta_{4}\leq 6\) and \(0 \leq \theta_{2},\theta_{3} \leq 7\), indeed this follows from the fact that the representations~\(\alpha(2,\chi_{\alpha})\) and \(\alpha(2,\chi_{-\alpha})\) are equivalent.
To be more precise, if we set ~\(A = \left(\begin{smallmatrix} 0&1\\ t&0 \end{smallmatrix}\right) \), then we~get~\(A \alpha(2,\chi_{\alpha}) A^{-1} =~\alpha(2,\chi_{-\alpha})\).
\end{remark}
The next proposition describes the Witt class of the metabelian Blanchfield pairing $ \Bl_{\alpha(2,\chi)}(K) $.
\begin{proposition}
\label{prop:MetabelianBlanchfieldHKLKnot}
Let $K$ be the knot described in~\eqref{eq:HKLKnot} and let $\chi \colon H_1(L_2(K);\mathbb{Z}) \to \mathbb{Z}_{\ell} $ be a character.
Write $\chi=\chi_1+\chi_2+\chi_3+\chi_4$ as in Lemma~\ref{lem:CharactersOfK}, where $\chi_i=\chi_{\theta_i}$ with \(0 \leq \theta_{1},\theta_{4}\leq 6\) and \(0 \leq \theta_{2},\theta_{3} \leq 7\).
Then the metabelian Blanchfield form $ \Bl_{\alpha(2,\chi)}(K) $ is Witt equivalent to
\begin{align}
\label{eq:HKLMetabelianDecompo}
& \Bl_{\alpha(2,\chi_{1})}(T_{2,13}) \oplus -\Bl_{\alpha(2,\chi_{4})}(T_{2,13}) \\
&\oplus \Bl_{\alpha(2,\chi_{2})}(T_{2,15}) \oplus -\Bl_{\alpha(2,\chi_{3})}(T_{2,15}) \nonumber \\
&\oplus \Bl(T_{2,3})(\xi_{13}^{\theta_{1}}t) \oplus \Bl(T_{2,3})(\xi_{13}^{-\theta_{1}}t) \oplus -\Bl(T_{2,3})(\xi_{15}^{\theta_{3}}t) \oplus -\Bl(T_{2,3})(\xi_{15}^{-\theta_{3}}t). \nonumber
\end{align}
\end{proposition}
\begin{proof}
Since we know from Corollary~\ref{cor:MetabelianConnectedSum} that metabelian Blanchfield pairings are additive, up to Witt equivalence, we need only study the metabelian Blanchfield pairing of $(2,d)$-cables of $(2,2k+1)$-torus knots (here, we also used Remark~\ref{rem:ReverseMirror}).
The proposition will follow from the claim that given a $(2,2k+1)$-torus knot \(K'\) and a character $\chi=\chi_\theta$ on $H_1(L_2(K'_{2,d});\mathbb{Z})$, there is an isometry
$$\Bl_{\alpha(2,\chi_{\theta})}(K'_{2,d}) \cong \Bl_{\alpha(2,\chi_{\theta})}(T_{2,d}) \oplus \Bl(K')(\xi_{\ell}^{-\theta}t) \oplus \Bl(K')(\xi_{\ell}^{\theta}t).$$
Using the notation from Section~\ref{sub:Satellite}, $K'_{2,d}$ is a satellite knot with pattern \(P = T_{2,d}\), companion~$K'$, and the infection curve \(\eta\) is in fact the curve which was denoted by \(a\) in Subsection~\ref{sec:blanchf-forms-twist}.
Since the winding number is $w=2$, the first assertion of Theorem~\ref{thm:metabelian-cabling-formula} implies that $\alpha(2,\chi)$ is $\eta$-regular and therefore restricts to a representation $\alpha(2,\chi)_{K'}$ on $\pi_1(M_{K'})$.
Since $n=2$ divides $w=2$, the representation~$\alpha(2,\chi)_{K'}$ is abelian.
As the curve $a$ is a generator of $H_1(M_{K'};\mathbb{Z})$, we see that~$\alpha(2,\chi)_{K'}$ is determined by $\alpha(2,\chi_{\theta})(a) =
\left(\begin{smallmatrix}
t \xi_{\ell}^{-\theta} & 0 \\
0 & t \xi_{\ell}^{\theta} \\
\end{smallmatrix}\right)$.
The claim (and thus the proposition) now follow by applying Theorem~\ref{thm:metabelian-cabling-formula}.
\end{proof}
Next, we determine the conditions under which $\Bl_{\alpha(2,\chi)}(K)$ is metabolic.
\begin{proposition}
\label{prop:MetabolicCriterion}
Let $K$ be the knot described in~\eqref{eq:HKLKnot} and let $\chi \colon H_1(L_2(K);\mathbb{Z}) \to \mathbb{Z}_{\ell} $ be a character.
Write $\chi=\chi_1+\chi_2+\chi_3+\chi_4$ as in Lemma~\ref{lem:CharactersOfK} where $\chi_i=\chi_{\theta_i}$ with \(0 \leq \theta_{1},\theta_{4}\leq 6\) and \(0 \leq \theta_{2},\theta_{3} \leq 7\).
Then the metabelian Blanchfield pairing \(\Bl_{\alpha(2,\chi)}(K)\) is metabolic if and only if \(\theta_{1}=\theta_{2}=\theta_{3}=\theta_{4}=0\).
\end{proposition}
\begin{proof}
Substituting $\theta_{1}=\theta_{2}=\theta_{3}=\theta_{4}=0$ into~\eqref{eq:HKLMetabelianDecompo} shows that $\Bl_{\alpha(2,\chi)}(K)$ is metabolic.
We now prove the converse in two steps.
Firstly, we show that if $\theta_1 \neq \theta_4$ (or $\theta_2 \neq \theta_3$), then $\Bl_{\alpha(2,\chi)}(K)$ is not metabolic.
Secondly, we show that when $\theta_1=\theta_4$ and $\theta_3=\theta_2$, the metabelian Blanchfield form $\Bl_{\alpha(2,\chi)}(K)$ is metabolic if and only if $\theta_{1} = \theta_{3} = 0$.
Assume that \(\theta_{1} \neq \theta_{4}\).
We assert that the signature jump of \(\Bl_{\alpha(2,\chi)}(K)\) at \(\xi_{13}^{\theta_{1}}\) is \(\pm1\).
To see this, first note that Proposition~\ref{prop:DecompoTwistedBlanchfield} (or a glance at the twisted Alexander polynomial) implies that in~\eqref{eq:HKLMetabelianDecompo}, only the summand \(\Bl_{\alpha(2,\chi_{1})}(T_{2,13}) \oplus -\Bl_{\alpha(2,\chi_{4})}(T_{2,13})\) can contribute a non-trivial signature jump at \(\xi_{13}^{\theta_{1}}\): indeed Proposition~\ref{prop:DecompoTwistedBlanchfield} shows that $\Bl_{\alpha(2,\chi_{2})}(T_{2,15}) \oplus -\Bl_{\alpha(2,\chi_{3})}(T_{2,15})$ can only jump at powers of $\xi_{15}$.
Since the untwisted terms in~\eqref{eq:HKLMetabelianDecompo} do not contribute to the signature jump either, we focus on $\Bl_{\alpha(2,\chi_{1})}(T_{2,13}) \oplus -\Bl_{\alpha(2,\chi_{4})}(T_{2,13}) $.
Proposition~\ref{prop:DecompoTwistedBlanchfield} implies that the signature function of~ \(\Bl_{\alpha(2,\chi_{1}) }(T_{2,13})\) jumps at $\xi_{13}^e$ when $e \neq \theta_1$.
Therefore, the signature jump of \(\Bl_{\alpha(2,\chi_{1})}(T_{2,13})\) at~\(\xi_{13}^{\theta_{1}}\) is trivial and, since $\theta_4 \neq \theta_1$, the signature jump of \(-\Bl_{\alpha(2,\chi_{4})}(T_{2,13})\) at~\(\xi_{13}^{\theta_{1}}\) is~\(\pm1\).
This concludes the proof of the assertion.
Using this assertion and Theorem~\ref{thm:WittClassificationComplex}, we deduce that~\(\Bl_{\alpha(2,\chi)}(K)\) is not metabolic.
The case where \(\theta_{2} \neq \theta_{3}\) is treated analogously.
Next, we assume that \(\theta_{1} = \theta_{4}\) and \(\theta_{2}=\theta_{3}\).
Using~\eqref{eq:HKLMetabelianDecompo}, this assumption implies that the metabelian Blanchfield form $\Bl_{\alpha(2,\chi)}(K) $ is Witt equivalent to
\begin{equation}
\label{eq:UntwistedBlanchfieldTorusKnots}
\Bl(T_{2,3})(\xi_{13}^{\theta_{1}}t) \oplus \Bl(T_{2,3})(\xi_{13}^{-\theta_{1}}t) \oplus -\Bl(T_{2,3})(\xi_{15}^{\theta_{3}}t) \oplus -\Bl(T_{2,3})(\xi_{15}^{-\theta_{3}}t).
\end{equation}
To determine whether $\Bl_{\alpha(2,\chi)}(K) $ is metabolic, Theorem~\ref{thm:WittClassificationComplex} implies that we must study the jumps of the signature function of the linking form in~\eqref{eq:UntwistedBlanchfieldTorusKnots}.
Since we are dealing with untwisted Blanchfield forms, Remark~\ref{rem:LevineTristram} implies that these jumps are the signature jumps of the corresponding Levine-Tristram signature function.
The proof of \cite[Theorem 7.1]{hedden_non-slice_2012} shows that for distinct $a_i$, the jumps (as $\omega$ varies along $S^1$) of the Levine-Tristram signatures $\sigma_{T_{m,n}}(\xi_\ell^{a_i} \omega)$ occur at distinct points (here $m,n$ are prime).
Consequently, we deduce that \(\Bl_{\alpha(2,\chi)}(K)\) is metabolic if and only if \(\theta_{1} = \theta_{3} = 0\).
This concludes proof of the theorem.
\end{proof}
We recover a result of Hedden, Kirk and Livingston~\cite{HeddenKirkLivingston}.
\begin{theorem}
\label{thm:HKLNotSlice}
The knot \(K\) from~\eqref{eq:HKLKnot} is algebraically slice but not slice.
\end{theorem}
\begin{proof}
Hedden-Kirk-Livingston show that \(K\) is algebraically slice~\cite[Lemma 2.1]{HeddenKirkLivingston}.
By means of contradiction, assume that $K$ is slice.
Theorem~\ref{thm:MillerPowell} implies that for any prime power $\ell$, there exists a metabolizer~$P$ of~$\lambda_{\ell}$ such that for any prime power $q^a$, and any non-trivial character $\chi \colon H_1(L_{\ell}(K);\mathbb{Z}) \to \mathbb{Z}_{q^a}$ vanishing on $P$, we have some $b \geq a$ such that the metabelian Blanchfield pairing $\Bl_{\alpha(n,\chi_b)}(K)$ is metabolic.
As Proposition~\ref{prop:MetabolicCriterion} shows that the Blanchfield pairing~$\Bl_{\alpha(n,\chi_b)}(K)$ is metabolic if and only if the character is trivial, we obtain the desired~contradiction.
This concludes the proof of the theorem.
\end{proof}
|
\section{Introduction}
$sk$-Splines were introduced and their basic theory was developed by Kushpel [2-4,7,8]. These results were
developed in [1,5,6,9]. In this cycle of works, $sk
$-splines were proved to be useful in Kolmogorov's problem on $n$-widths and
approximation of smooth, infinitely smooth, analytic and entire functions on
the circle, $d$-torus and $d$-dimensional sphere. The theory of $sk$-splines
and methods developed were applied by different authors for calculation of
sharp values of $n$-widths (see e.g. [11,13] and
references therein). Here we introduce $sk$-splines on ${\bf R}^{n}$. Let
$\Lambda :=\left\{ {\bf x}_{{\bf k}}\right\} $ be an additive group of
lattice points in ${\bf R}^{n}$ \ and $K\left( \cdot \right) :{\bf R
^{n}\rightarrow {\bf R}$ be a fixed kernel function. An $sk$-Spline is a
function of the for
\[
sk\left( {\bf x}\right) =\sum_{{\bf x}_{{\bf k}}\in {\bf \Lambda
}}c_{{\bf k}}K\left( {\bf x-x}_{{\bf k}}\right) \in L_{2}({\bf R}^{n}).
\
Let us consider in more details the problem of interpolation in ${\bf R
^{n}$. For a fixed $\Lambda $ and any continuous function $f$ $:{\bf R
^{n}\rightarrow {\bf R}$ we need to find such $c_{{\bf j}}$, ${\bf
}\in {\bf Z}^{n}$ that
\[
\sum_{{\bf x}_{{\bf j}}\in \Lambda }c_{{\bf j}}K\left( {\bf x}_
{\bf k}}-{\bf x}_{{\bf j}}\right) =f\left( {\bf x}_{{\bf k
}\right) ,
\
for any ${\bf x}_{{\bf k}}\in \Lambda $. Assuming some general
conditions on the kernel function $K$ we \ give an explicit solution of this
problem in the case $\Lambda ={\bf A}{\bf Z}^{n}\subset {\bf R}^{n}
, where ${\bf A}$ is an arbitrary nonsingular matrix. Such grids are
analogs for ${\bf R}^{n}$ of number theoretic Korobov's grids and their
various modifications, such as sparse grids. These grids proved to be useful
in high dimensional integration and interpolation (see e.g. [1]).
Let $L_{p}({\bf R}^{n})$ be the usual space of $p$-integrable functions
equipped with the norm
\[
\Vert f\Vert _{p}=\Vert f\Vert _{L_{p}({\bf R}^{n})}:=\left\{
\begin{array}{cc}
\left( \int_{{\bf R}^{n}}\left\vert f({\bf x})\right\vert ^{p}d{\bf
x}\right) ^{1/p}, & 1\leq p<\infty , \\
\mathrm{ess}\,\,\sup_{{\bf x}\in {\bf R}^{n}}|f({\bf x})|, &
p=\infty
\end{array
\right.
\]
Let ${\bf x}$ and ${\bf y}$ be two vectors in ${\bf R}^{n}$ and
\left\langle {\bf x,y}\right\rangle =\sum_{k=1}^{n}{\bf x}_{k}{\bf y}_{k}$
be the usual scalar product, $|{\bf x}|=\left\langle {\bf x,x}\right\rangle^{1/2}$. For an integrable on ${\bf R}^{n}$ function,
i.e., $f({\bf x})\in L_{1}\left( {\bf R}^{n}\right) $ define its
Fourier transform
\[
{\bf F}f({\bf y})=\int_{{\bf R}^{n}}\exp \left( -i\left\langle
{\bf x,y}\right\rangle \right) f({\bf x})d{\bf x}.
\
and its formal inverse as
\[
\left( {\bf F}^{-1}f\right) ({\bf x})=\frac{1}{\left( 2\pi \right) ^{n
}\int_{{\bf R}^{n}}\exp \left( i\left\langle {\bf x,y}\right\rangle
\right) f({\bf y})d{\bf y}.
\]
We will need the following well-known results (see [12] for more
information).
{\bf Theorem 1} \begin{em}
(Plancherel's theorem) The Fourier transform is a
linear continuous operator from $L_{2}\left( {\bf R}^{n}\right) $ onto
L_{2}\left( {\bf R}^{n}\right) .$ The inverse Fourier transform, $F^{-1},$
can be obtained by letting
\[
\left( {\bf F}^{-1}g\right) \left( {\bf x}\right) =\frac{1}{(2\pi )^{n
}\left( {\bf F}g\right) \left( -{\bf x}\right)
\
for any $g\in L_{2}\left( {\bf R}^{n}\right) .$
\end{em}
{\bf Theorem 2}
\begin{em}
(Poisson summation formula) Suppose that for some $C>0$ and $\delta >0$ we have $\max \left\{ f({\bf x}),{\bf F}f
{\bf x})\right\} \leq C\left( 1+\left\vert {\bf x}\right\vert \right)
^{-n-\delta }$. Then
\[
\sum_{{\bf m}\in {\bf Z}^{n}}f\left( {\bf x}+{\bf m}\right)
=\sum_{{\bf m}\in {\bf Z}^{n}}{\bf F}f\left( 2\pi {\bf m}\right)
\exp \left( 2\pi i\left\langle {\bf m},{\bf x}\right\rangle \right)
\
and the series converges absolutely.
\end{em}
\section{Interpolation by $sk$-splines on ${\bf R}^{n}$}
Let ${\bf A}$ be a nonsingular $n\times n$ matrix. Consider the set
\Omega _{{\bf A}}$ of grid points ${\bf x}_{{\bf m}}:={\bf Am}$,
where ${\bf m }\in{\bf Z}^{n}$. For a fixed continuous kernel function
$K$, the space $SK\left( \Omega _{{\bf A}}\right) $ of $sk$-splines on
\Omega _{{\bf A}}$ is the space of functions representable in the form
\[
sk\left( {\bf x}\right) =\sum_{{\bf m }\in{\bf Z}^{n}}c_{{\bf m
}K\left( {\bf x-x}_{{\bf m}}\right) ,
\
where $c_{{\bf m}}\in {\bf R}$, ${\bf m\in }{\bf Z}^{n}$. Let $f\left( {\bf x
\right) $ be a continuous function, $f:{\bf R}^{n}\rightarrow {\bf R}
. Consider the problem of interpolation by $sk$-splines, $sk\left( {\bf x
_{\bf s}\right) =f\left( {\bf x}_{\bf s}\right) $, where
{\bf s }\in{\bf Z}^{n}$. Even in the one-dimensional case the problem
of interpolation does not always have a solution. If the solution exists then
the $sk$-spline interpolant can be written in the form
\[
sk\left( {\bf x}\right) =\sum_{{\bf s }\in{\bf Z}^{n}}f\left(
{\bf x}_{{\bf s}}\right) \widetilde{sk}\left( {\bf x}-{\bf x}_
{\bf s}}\right) ,
\
where $\widetilde{sk}\left( {\bf x}\right) $ is a fundamental $sk
-spline, i.e
\[
\widetilde{sk}\left( {\bf x}_{{\bf s}}\right) =\left\{
\begin{array}{cc}
1, & {\bf s=0,} \\
0, & {\bf s\neq 0.
\end{array
\right.
\]
{\bf Theorem 3}
\begin{em}
Let ${\bf A}$ be a
nonsingular matrix, $K:{\bf R}^{n}\rightarrow {\bf R}$,$K\in
L_{2}\left( {\bf R}^{n}\right) \cap L_{1}\left( {\bf R}^{n}\right)
\cap C\left( {\bf R}^{n}\right) .$ Assume that
\[
\sum_{{\bf m}\in {\bf Z}^{n}}{\bf F}\left( K\right) \left( {\b
z+}2\pi \left( {\bf A}^{-1}\right) ^{T}{\bf m}\right) \neq 0,\forall
{\bf z }\in2\pi \left( {\bf A}^{-1}\right) ^{T}{\bf Q},
\
where ${\bf Q}:=\left\{ {\bf x|x=}\left( x_{1},\cdot \cdot \cdot
,x_{n}\right) \in {\bf R}^{n},0\leq x_{k}\leq 1,1\leq k\leq
n\right\} ,$ and the function
\[
\Upsilon \left( {\bf z}\right) :=\frac{1}{\sum_{{\bf m}\in {\bf Z
^{n}}{\bf F}\left( K\right) \left( {\bf z+}2\pi \left( {\bf A
^{-1}\right) ^{T}{\bf m}\right) }
\
can be represented by its Fourier series, i.e. for any ${\bf z\in
{\bf R}^{n}$, \
\begin{equation}
\Upsilon \left( {\bf z}\right) =\sum_{{\bf s}\in {\bf Z}^{n}}\alpha
_{{\bf s}}\exp \left( -i\left\langle {\bf As},{\bf z}\right\rangle
\right) . \label{repr}
\end{equation
The
\[
\widetilde{sk}\left( {\bf x}\right) =\frac{\det \left( {\bf A}\right)
}{\left( 2\pi \right) ^{n}}\int_{{\bf R}^{n}}\Upsilon \left( {\bf z
\right) {\bf F}\left( K\right) \left( {\bf z}\right) \exp \left(
i\left\langle {\bf z},{\bf x}\right\rangle \right) d{\bf z}
\
and this representation is unique.
\end{em}
{\bf Proof} Since $K\in L_{2}\left( {\bf R}^{n}\right) \cap
L_{1}\left( {\bf R}^{n}\right) $ and \ using (\ref{repr}) we get
\[
\widetilde{sk}\left( {\bf x}\right) =\frac{\det \left( {\bf A}\right)
}{\left( 2\pi \right) ^{n}}\int_{{\bf R}^{n}}\Upsilon \left( {\bf z
\right) {\bf F}\left( K\right) \left( {\bf z}\right) \exp \left(
i\left\langle {\bf x},{\bf z}\right\rangle \right) d{\bf z}
\
\[
=\frac{\det \left( {\bf A}\right) }{\left( 2\pi \right) ^{n}}\int_
{\bf R}^{n}}{\bf F}\left( K\right) \left( {\bf z}\right) \left(
\sum_{{\bf s}\in {\bf Z}^{n}}\alpha _{{\bf s}}\exp \left(
-i\left\langle {\bf As},{\bf z}\right\rangle \right) \right) \exp
\left( i\left\langle {\bf x,z}\right\rangle \right) d{\bf z}
\
\[
=\frac{\det \left( {\bf A}\right) }{\left( 2\pi \right) ^{n}}\sum_
{\bf s}\in {\bf Z}^{n}}\alpha _{\bf s}\int_{{\bf R}^{n}
{\bf F}\left( K\right) \left( {\bf z}\right) \exp \left( i\left\langle
{\bf x}-{\bf As},{\bf z}\right\rangle \right) d{\bf z.}
\
Since $K\in L_{2}\left( {\bf R}^{n}\right) $ then by Theorem
\[
\widetilde{sk}\left( {\bf x}\right) =\det \left( {\bf A}\right) \sum_
{\bf s}\in {\bf Z}^{n}}\alpha _{{\bf s}}K\left( {\bf x}-{\bf
As}\right) ,
\
so that $\widetilde{sk}\left( {\bf x}\right) \in SK\left( \Omega _
{\bf A}}\right) $. Next, we calculate $\widetilde{sk}\left( {\bf Am
\right) $ for ${\bf m}\in {\bf Z}^{n}$,
\[
\widetilde{sk}\left( {\bf Am}\right) =\frac{\det \left( {\bf A}\right)
}{\left( 2\pi \right) ^{n}}\int_{{\bf R}^{n}}\Upsilon \left( {\bf z
\right) {\bf F}\left( K\right) \left( {\bf z}\right) \exp \left(
i\left\langle {\bf z},{\bf Am}\right\rangle \right) d{\bf z}
\
\[
=\frac{\det \left( {\bf A}\right) }{\left( 2\pi \right) ^{n}}\sum_
{\bf l}\in {\bf Z}^{n}}\int_{2\pi \left( {\bf A}^{-1}\right)^{T
{\bf l}+2\pi \left( {\bf A}^{-1}\right) ^{T}{\bf Q}}\Upsilon \left(
{\bf z}\right) {\bf F}\left( K\right) \left( {\bf z}\right) \exp
\left( i\left\langle {\bf z},{\bf Am}\right\rangle \right) d{\bf z}
\
\[
=\frac{\det \left( {\bf A}\right) }{\left( 2\pi \right) ^{n}}
\
\[
\times \sum_{{\bf l}\in {\bf Z}^{n}}\int_{2\pi \left( {\bf A
^{-1}\right) ^{T}{\bf Q}}\frac{{\bf F}\left( K\right) \left( {\bf z
}2\pi \left( {\bf A}^{-1}\right) ^{T}{\bf l}\right)
}{\sum_{{\bf m}\in {\bf Z}^{n}}{\bf F}\left( K\right) \left(
{\bf z-}2\pi \left( {\bf A}^{-1}\right) ^{T}{\bf l
+2\pi \left( {\bf A}^{-1}\right) ^{T}{\bf m}\right) }
\
\[
\times \exp \left( i\left\langle {\bf z-}2\pi \left( {\bf A
^{-1}\right) ^{T}{\bf l},{\bf Am}\right\rangle \right)
{\bf z.}
\
Changing ${\bf l}$ by $-{\bf l}$ in $\sum_{{\bf l}\in {\bf Z
^{n}}$ we ge
\[
\widetilde{sk}\left( {\bf Am}\right) =\frac{\det \left( {\bf A}\right)
}{\left( 2\pi \right) ^{n}}
\
\[
\times \sum_{{\bf l}\in {\bf Z}^{n}}\int_{2\pi \left( {\bf A
^{-1}\right) ^{T}{\bf Q}}\frac{{\bf F}\left( K\right) \left( {\bf z
}2\pi \left( {\bf A}^{-1}\right) ^{T}{\bf l}\right)
}{\sum_{{\bf m}\in {\bf Z}^{n}}{\bf F}\left( K\right) \left(
{\bf z+}2\pi \left( {\bf A}^{-1}\right) ^{T}{\bf l
+2\pi \left( {\bf A}^{-1}\right) ^{T}{\bf m}\right) }
\
\[
\times \exp \left( i\left\langle {\bf z+}2\pi \left( {\bf A
^{-1}\right) ^{T}{\bf l},{\bf Am}\right\rangle \right)
{\bf z.}
\
Since
\[
\sum_{{\bf m}\in {\bf Z}^{n}}{\bf F}\left( K\right) \left( {\b
z+}2\pi \left( {\bf A}^{-1}\right) ^{T}{\bf l}+2\pi
\left( {\bf A}^{-1}\right) ^{T}{\bf m}\right)
\
\[
=\sum_{{\bf m}\in {\bf Z}^{n}}{\bf F}\left( K\right) \left( {\b
z+}2\pi \left( {\bf A}^{-1}\right) ^{T}{\bf m}\right)
=\Upsilon ^{-1}\left( {\bf z}\right)
\
and
\[
\exp \left( i\left\langle {\bf z+}2\pi \left( {\bf A
^{-1}\right) ^{T}{\bf l},{\bf Am}\right\rangle \right) =\exp
\left( i\left\langle {\bf z},{\bf Am}\right\rangle \right)
\
for any ${\bf l}\in {\bf Z}^{n}$ then
\[
\widetilde{sk}\left( {\bf Am}\right) =\frac{\det \left( {\bf A}\right)
}{\left( 2\pi \right) ^{n}}\sum_{{\bf l}\in {\bf Z}^{n}}\int_{2\pi
\left( {\bf A}^{-1}\right) ^{T}{\bf Q}}\Upsilon \left( {\bf z
\right) {\bf F}\left( K\right) \left( {\bf z+}2\pi \left( {\bf A
^{-1}\right) ^{T}{\bf l}\right)
\
\[
\times \exp \left( i\left\langle {\bf z+}2\pi \left( {\bf A
^{-1}\right) ^{T}{\bf l},{\bf Am}\right\rangle \right)
{\bf z}
\
\[
=\frac{\det \left( {\bf A}\right) }{\left( 2\pi \right) ^{n}}\int_{2\pi
\left( {\bf A}^{-1}\right) ^{T}{\bf Q}}\exp \left( i\left\langle
{\bf z},{\bf Am}\right\rangle \right) \Upsilon \left( {\bf z
\right) \sum_{{\bf l}\in {\bf Z}^{n}}{\bf F}\left( K\right) \left(
{\bf z}+2\pi \left( {\bf A}^{-1}\right) ^{T}{\bf l}\right)
d{\bf z}
\
\[
=\frac{\det \left( {\bf A}\right) }{\left( 2\pi \right) ^{n}}\int_{2\pi
\left( {\bf A}^{-1}\right) ^{T}{\bf Q}}\exp \left( i\left\langle
{\bf z},{\bf Am}\right\rangle \right) d{\bf z}
\
\[
=\int_{{\bf Q}}\exp \left( i\left\langle {\bf z},{\bf m
\right\rangle \right) d{\bf z}=\left\{
\begin{array}{cc}
1, & m_{k}=0,1\leq k\leq n, \\
0, & otherwise
\end{array
\right. .
\]
The representation of the fundamental $sk-$spline is unique since the
functions $\widetilde{sk}\left( {\bf x}-{\bf x}_{{\bf m}}\right)
{\bf x}_{{\bf m}}\in \Omega _{{\bf A}}$ are linearly independent.
Observe that Theorem 3 is a multidimensional analog of Schoenberg's result
on interpolation by polynomial splines of odd degree on ${\bf R}$ [10].
Let ${\bf Q}$ be a nonsingular matrix and $f\in L_{2}\left(
{\bf R}^{n}\right) $. The
\[
{\bf F}\left( f\left( {\bf Q}\cdot \right) \right) \left( {\bf x
\right) =\frac{1}{\det \left( {\bf Q}\right) }{\bf F}\left( f\left(
\cdot \right) \right) \left( \left( {\bf Q}^{-1}\right) ^{T}{\bf x
\right) .
\
Hence for any $K\in L_{2}\left( {\bf R}^{n}\right) $, by Theorem 1,
\[
{\bf F}\left( {\bf F}\left( K\right) \left( 2\pi \left( {\bf A
^{-1}\right) ^{T}\left(\cdot \right) \right) \right) \left(
{\bf y}\right)
\
\[
=\frac{1}{\det \left( 2\pi \left( {\bf A}^{-1}\right) ^{T}\right)
{\bf F\circ F}\left( K\left( \cdot \right) \right) \left( \left( 2\pi
\left( {\bf A}^{-1}\right) ^{T}\right) ^{-1}{\bf y}\right)
\
\[
=\frac{\det \left( {\bf A}\right) }{\left( 2\pi \right) ^{n}}\left( 2\pi
\right) ^{n}{\bf F}^{-1}\circ { \bf F}\left( K\left( \cdot \right)
\right) \left( -\left( 2\pi \left( {\bf A}^{-1}\right) ^{T}\right) ^{-1
{\bf y}\right)
\
\begin{equation}
=\det \left( {\bf A}\right) K\left( -\frac{1}{2\pi }{\bf A}^{T}{\bf
y}\right) . \label{sk111}
\end{equation
Assume that $K$ satisfies the conditions of Theorem 2. Let ${\bf z}=2\pi
\left( {\bf A}^{-1}\right) ^{T}{\bf x}$ then using (\ref{sk111}) we ge
\[
\Upsilon ^{-1}\left( {\bf z}\right) =\sum_{{\bf m}\in {\bf Z}^{n}
{\bf F}\left( K\right) \left( {\bf z+}2\pi \left( {\bf A
^{-1}\right) ^{T}{\bf m}\right)
\
\[
=\sum_{{\bf m}\in {\bf Z}^{n}}{\bf F}\left( K\right) \left( 2\pi
\left( {\bf A}^{-1}\right) ^{T}\left( {\bf x}+{\bf m}\right)
\right)
\
\[
=\det \left( {\bf A}\right) \sum_{{\bf m}\in {\bf Z}^{n}}K\left(
{\bf A}^{T}{\bf m}\right) \exp \left( 2\pi i\left\langle {\bf m,x
\right\rangle \right)
\
\[
=\det \left( {\bf A}\right) \sum_{{\bf m}\in {\bf Z}^{n}}K\left(
{\bf A}^{T}{\bf m}\right) \exp \left( 2\pi i\left\langle {\bf m,
\left( 2\pi \left( {\bf A}^{-1}\right) ^{T}\right) ^{-1}{\bf z
\right\rangle \right)
\
\[
=\det \left( {\bf A}\right) \sum_{{\bf m}\in {\bf Z}^{n}}K\left(
{\bf A}^{T}{\bf m}\right) \exp \left( i\left\langle {\bf Am,z
\right\rangle \right) .
\
Hence we proved
{\bf Corollary 1}
\begin{em}
Let $K$ satisfies the conditions of
Theorem 1-Theorem 3 then
\[
\widetilde{sk}\left( {\bf x}\right) ={\bf F}^{-1}\left( \frac{{\bf
}\left( K\right) \left( \cdot \right) }{\sum_{{\bf m}\in {\bf Z
^{n}}K\left( -{\bf A}^{T}{\bf m}\right) \exp \left( i\left\langle
{\bf Am,\cdot }\right\rangle \right) }\right) \left( {\bf x}\right) .
\]
\end{em}
{\bf Example 1} Let ${\bf B}$ be a nonsingular matrix and $K\left(
{\bf x}\right) $ be a Gaussian of the form $K\left( {\bf x}\right)
=\exp \left( -\left\vert {\bf Bx}\right\vert ^{2}\right) ,$ then
\[
{\bf F}\left( K\right) \left( {\bf z}\right) =\frac{\pi ^{n/2}}{\det
\left( {\bf B}\right) }\exp \left( -\left\vert \left( {\bf B
^{-1}\right) ^{T}{\bf z}\right\vert ^{2}\right) .
\
It is easy to check that in this case $K$ satisfies the conditions of
Theorem 1-Theorem 3. Hence for the interpolation by ${\bf A}{\bf Z}^{n}
$-shifts of Gaussians we have
\[
\widetilde{sk}\left( {\bf x}\right) =\frac{\pi ^{n/2}}{\det \left(
{\bf B}\right) }{\bf F}^{-1}\left( \frac{\exp \left( -\left\vert
\left( {\bf B}^{-1}\right) ^{T}\cdot \right\vert ^{2}\right) }
\sum_{{\bf m}\in {\bf Z}^{n}}\exp \left( -\left\vert {\bf BA}^{T
{\bf m}\right\vert ^{2}+i\left\langle {\bf Am,\cdot }\right\rangle
\right) }\right) \left( {\bf x}\right) .
\]
|
\section*{Table of Contents}
\begin{center}\small
\begin{minipage}[c]{0.98\textwidth}\vspace*{-8ex}%
\tableofcontents
\end{minipage}
\end{center}
\section{Motivation, Background, and Summary of Contribution}
Arguably most computational problems in Science and Engineering
are concerned with continuous rather than with discrete data \cite{BC06,Bra13}.
Here the Theory of Computability exhibits new topological ---
and continuous complexity theory furthermore metric --- aspects
that trivialize, and are thus invisible, in the discrete realm.
In particular input and output require rather careful a choice
of the underlying encoding as sequences of bits
to be read, processed, and written by a Turing machine. For example,
\begin{itemize}
\item
encoding real numbers via their binary expansion
$x=\sum_{n=0}^\infty b_n2^{-n-1}$,
and thus operating on the Cantor space of infinite binary sequences $\bar b=(b_0,b_1,\ldots b_n,\ldots)$,
renders arithmetic averaging $[0;1]^2\ni(x,y)\mapsto (x+y)/2\in[0;1]$
\emph{dis}continuous and \emph{un}computable
\cite{Turing37},\cite[Exercise~7.2.7]{Wei00}.
\item
Encoding real numbers via a sequence of (numerators and denominators, in binary, of)
rational approximations up to absolute error $\leq2^{-n}$ does render averaging computable
\cite[Theorem~4.3.2]{Wei00},
but admits no worst-case bound on computational cost \cite[Examples~7.2.1+7.2.3]{Wei00}.
\item
The \emph{dyadic} representation encodes
$x\in[0;1]$ as any integer sequence $a_n\in\{0,\ldots 2^n\}$ (in binary without leading 0)
s.t. $|x-a_n/2^n|\leq2^{-n}$; and similarly encode $y\in[0;1]$ as $(b_n)$.
Then the/an integer $c_n$ closest to $(a_{n+1}+b_{n+1})/4$
satisfies $\big|(x+y)/2-c_n/2^n\big|\leq2^{-n}$,
and is easily computed --- but requires first reading/writing
$a_m,b_m,c_m$ for all $m<n$: a total of $\Theta(n^2)$ bits.
\item
Encoding $x\in[0;1]$ as \emph{signed} binary expansion
$x=\sum_{n\geq0} (2b_n+b_{n+1}-1)\cdot2^{-n-1}$ with $b_n\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}$ s.t. $2b_n+b_{n+1}\in\{-1,0,1\}$,
and similarly $y$, renders averaging computable in linear time $\mathcal{O}(n)$ \cite[Theorem~7.3.1]{Wei00}.
\end{itemize}
The signed binary expansion is thus asymptotically `optimal' up to a constant factor,
the dyadic representation is still optimal up to a quadratic polynomial,
the rational representation is `unbounded',
and the binary expansion is unsuitable.
But how to choose and quantitatively assess complexity-theoretically appropriate
encodings of spaces $X$ other than $[0;1]$,
such as those common in the analysis and solution theory of PDEs \cite{Triebel}?
The present work refines the existing classification of encodings from the
computability theory of general continuous data while guided by and generalizing
the well-established theory of computational complexity over real numbers.
There, the binary expansion is known to violate the technical condition
of \emph{admissibility}; and we introduce and investigate quantitative strengthenings
\emph{linear} admissibility (satisfied by the signed binary,
but neither by the dyadic nor by the rational representation) and
\emph{polynomial} admissibility (satisfied by the signed binary
and by the dyadic, but not by the rational representation).
\subsection{Computability over Continuous Data, Complexity in Real Computation}
Here we review established notions and properties
of computability and complexity theory over the reals,
as well as notions and properties of
computability theory over more general abstract spaces:
as guideline to the sensible complexity theory
of more general abstract spaces developed in the sequel.
\begin{definition}
\label{d:Type2}
A \emph{Type-2 Machine} $\mathcal{M}$ is a Turing machine with
dedicated one-way output tape and infinite read-only input tape
{\rm\cite[Definitions~2.1.1+2.1.2]{Wei00}}.
Naturally operating on infinite sequences of bits,
$\mathcal{M}$ \emph{computes}
a partial function $F:\subseteq\mathcal{C}\to\mathcal{C}$
on the Cantor space $\mathcal{C}=\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^\omega$ of infinite binary sequences if,
when run with any input $\bar b\in\operatorname{dom}(F)$ on its tape,
$\mathcal{M}$ keeps printing the symbols of $F(\bar b)$ one by one;
while its behaviour on other inputs may be arbitrary.
$\mathcal{M}$ computes $F$ in \emph{time} $t:\mathbb{N}\to\mathbb{N}$ if
it prints
the $n$-th symbol of $F(\bar b)$ after at most $t(n)$ steps
regardless of $\bar b\in\operatorname{dom}(F)$.
For a fixed predicate $\varphi:\mathcal{C}\to\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}$,
a Type-2 Machine with \emph{oracle} $\varphi$
can repeatedly query $\varphi(\vec z)\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}$
for finite strings $\vec z$ during its computation.
\end{definition}
Concerning topological spaces $X$ of continuum cardinality beyond real numbers,
the \emph{Type-2 Computability Theory} systematically studies and compares
encodings, formalized as follows \cite[\S3]{Wei00}:
\begin{definition}
\label{d:TTE}
\begin{enumerate}
\item[a)]
A \emph{representation} of a set $X$ is a partial surjective
mapping $\xi:\subseteq\mathcal{C}:=\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{\mathbb{N}}\twoheadrightarrow X$
on the Cantor space of infinite streams of bits.
\item[b)]
The \emph{product} of representations $\xi:\subseteq\mathcal{C}\twoheadrightarrow X$
and $\upsilon:\subseteq\mathcal{C}\twoheadrightarrow Y$ is
$\xi\times\upsilon:\subseteq\mathcal{C}\ni (b_0,b_1,\ldots b_n,\ldots)\mapsto
\big(\xi(b_0,b_2,b_4,\ldots),\upsilon(b_1,b_3,\ldots)\big)\in X\times Y$.
\item[c)]
For representations $\xi:\subseteq\mathcal{C}\twoheadrightarrow X$
and $\upsilon:\subseteq\mathcal{C}\twoheadrightarrow Y$,
a $(\xi,\upsilon)$-\emph{realizer} of a function $f:X\to Y$
is a partial function $F:\operatorname{dom}(\xi)\to\operatorname{dom}(\upsilon)\subseteq\mathcal{C}$ on Cantor space
such that $f\circ\xi=\upsilon\circ F$ holds; see Figure~\ref{f:commdiag}.
\item[d)]
$(\xi,\upsilon)$-\emph{computing} $f$ means to compute some $(\xi,\upsilon)$-realizer $F$ of $f$
in the sense of Definition~\ref{d:Type2}.
\item[e)]
A \emph{reduction} from representation $\xi\twoheadrightarrow X$ to $\xi'\twoheadrightarrow X$
is a $(\xi,\xi')$-realizer of the identity $\operatorname{id}:X\to X$; that is,
a partial function $F:\operatorname{dom}(\xi)\to\operatorname{dom}(\xi')$ on Cantor space such that $\xi=\xi'\circ F$.
We write $\xi\preccurlyeq_{\rm T}\xi'$ to express that a \emph{continuous} reduction exists,
where $\mathcal{C}$ is equipped with the Cantor space metric $d_{\mathcal{C}}(\bar b,\bar a)=2^{-\min\{n:b_n\neq a_n\}}$.
\end{enumerate}
\end{definition}
Examples~\ref{x:Binary}, \ref{x:Rational}, \ref{x:Dyadic}, and \ref{x:SignedDigit} below
formalize the above binary, rational, dyadic, and signed encodings of the reals
as representations $\beta$, $\rho$, $\delta$, and $\sigma$, respectively.
It is well-known that the latter three, but not $\beta$,
are pairwise continuously reducible \cite[Theorem~7.2.5]{Wei00}
and thus equivalent with respect to the notions of computability they induce on reals;
but only $\delta$ and $\sigma$
admit mutual reductions with \emph{polynomial} modulus of continuity.
\begin{myremark}
Recall {\rm\cite[\S6]{Wei03}} that a \emph{modulus\footnote{%
From the general logical perspective this constitutes a
\emph{Skolemization} as canonical quantitative refinement of qualitative continuity \cite{Koh08a}.
We consider a modulus as an integer function due to its close connection
with asymptotic computational cost: Fact~\ref{f:Proper}a+b).
For the continuous conception of a modulus common in Analysis see Lemma~\ref{l:Modulus} below.} continuity}
of a function $f:X\to Y$ between compact metric spaces $(X,d)$ and $(Y,e)$
is a non-decreasing mapping $\mu:\mathbb{N}\to\mathbb{N}$ such that
$d(x,x')\leq2^{-\mu(n)}$ implies $e\big(f(x),f(x')\big)\leq2^{-n}$.
\\
Every uniformly continuous function has a (pointwise minimal) modulus of continuity;
Lipschitz-continuity corresponds to moduli $\mu(n)=n+\mathcal{O}(1)$,
and H\"{o}lder-continuity to linear moduli $\mu(n)=\mathcal{O}(n)$:
see Fact~\ref{f:Topology}c).
\end{myremark}
According to the sometimes so-called \emph{Main Theorem} of Computable Analysis,
a real function $f$ is continuous iff $f$ is computable by some oracle Type-2 Machine
w.r.t. $\rho$ and/or $\delta$ and/or $\sigma$ \cite[Definitions~2.1.1+2.1.2]{Wei00}.
For spaces beyond the reals, Kreitz and Weihrauch \cite{KW85} have identified
\emph{admissibility} as central condition on `sensible' representations
in that these
make the Main Theorem generalize \cite[Theorem~3.2.11]{Wei00}:
\begin{fact} \label{f:Main}
Let $X$ and $Y$ denote second-countable T$_0$ spaces
equipped with \emph{admissible} representations $\xi$ and $\upsilon$, respectively.
A function $f:X\to Y$ is continuous ~iff~
it admits a continuous $(\xi,\upsilon)$-\emph{realizer}.
Recall {\rm\cite[Theorem~3.2.9.2]{Wei00}} that
a representation $\xi:\subseteq\mathcal{C}\twoheadrightarrow X$ is \emph{admissible} iff
(i) it is continuous and (ii) every continuous representation $\zeta:\subseteq\mathcal{C}\twoheadrightarrow X$
satisfies $\zeta\preccurlyeq_{\rm T}\xi$.
\end{fact}
Computability-theoretically `sensible' representations $\xi$ are thus those
maximal, among the continuous ones, with respect to continuous reducibility.
The present work refines these considerations and notions from qualitative computability to
complexity. For representations satisfying our proposed strengthening of \emph{admissibility},
Theorem~\ref{t:Main2} below asymptotically optimally translates
quantitative continuity between functions $f:X\to Y$ and their realizers $F$.
Such translations heavily depend on the co/domains $X,Y$ under consideration:
\begin{myexample}
\label{x:Max}
\begin{enumerate}
\item[a)]
A function $f:[0;1]\to[0;1]$ has polynomial
modulus of continuity ~iff~ it has a $(\delta,\delta)$-realizer
w.r.t. the dyadic and/or signed binary expansion
(itself having a polynomial modulus of continuity)
~iff~ $f$ has a $(\sigma,\sigma)$-realizer
of polynomial modulus of continuity;
cmp. {\rm\cite[Theorem~2.19]{Ko91},\cite[Exercise~7.1.7]{Wei00},\cite[Theorem~14]{DBLP:conf/lics/KawamuraS016}}.
\item[b)]
For the compact space $[0;1]'_1:=\operatorname{Lip}_1([0;1],[0;1])$ of non-expansive (=Lipschitz-continuous with constant 1)
$f:[0;1]\to[0;1]$ equipped with the supremum norm,
the application function\emph{al} $[0;1]'_1\times[0;1]\ni (f,r)\mapsto f(r)\in[0;1]$
is non-expansive,
yet admits \emph{no} realizer with \emph{sub-}exponential modulus of continuity
for \emph{any} product representation (Definition~\ref{d:TTE}b)
of $[0;1]'_1\times[0;1]$ {\rm\cite[Example~6g]{KMRZ15}}.
\end{enumerate}
\end{myexample}
\begin{figure}[htb]
\begin{center}\includegraphics[width=0.5\textwidth]{commdiag}\end{center}
\caption{\label{f:commdiag}Realizers of $f:X\to Y$ with respect to different representations.}
\end{figure}
It is naturally desirable that, like in the discrete setting,
every computation of a total function $f:X\to Y$ admit some (possibly fast growing, but pointwise finite)
worst-case complexity bound $t=t(n)$; however already for the real unit interval $X=[0;1]=Y$
this requires the representation $\xi$ of $X$ to be chosen with care,
avoiding both binary $\beta$ and rational $\rho$ \cite[Example~7.2.3]{Wei00}.
Specifically, one wants its domain $\operatorname{dom}(\xi)\subseteq\mathcal{C}$ to be compact \cite{Sch95,Wei03,Sch04}:
cmp. Example~\ref{x:Rational}.
\begin{fact} \label{f:Proper}
\begin{enumerate}
\item[a)]
Suppose Type-2 Machine $\mathcal{M}$ (with/out some fixed oracle) computes a function
$F:\subseteq\mathcal{C}\to\mathcal{C}$ with compact $\operatorname{dom}(F)$.
Then $\mathcal{M}$ admits a bound $t(n)\in\mathbb{N}=\{0,1,2,\ldots\}$
on the number of steps it takes to
print the first $n$ output symbols of the value $F(\bar b)$
\emph{regardless} of the argument $\bar b\in\operatorname{dom}(F)$;
see {\rm\cite[Exercise~7.1.2]{Wei00}}.
\item[b)]
If $F:\subseteq\mathcal{C}\to\mathcal{C}$ is computable (with/out some fixed oracle) in time $t$,
then $n\mapsto t(n)$ constitutes a modulus of continuity of $F$.
\item[c)]
Conversely to every continuous $F:\subseteq\mathcal{C}\to\mathcal{C}$
with modulus of continuity $\mu$ there exists an oracle\footnote{%
In our model, query tape and head are not altered by an oracle query.}
$\varphi$ and Type-2 Machine $\mathcal{M}^\varphi$
computing $F$ in time $\mathcal{O}\big(n+\mu(n)\big)$;
cmp. {\rm\cite[Theorem~2.3.7.2]{Wei00}}.
\end{enumerate}
\end{fact}
Item~b) expresses that, on Cantor space, quantitative continuity
basically coincides with time-bounded relativized computability.
Item~a) requires a continuous representation $\xi$ to map closed subsets
of Cantor space to closed subsets of $X$ --- hence one cannot expect
$\xi$ to be an \emph{open} mapping, as popularly posited
in computability \cite[Lemma~3.2.5.(2+3+5)]{Wei00}
and ingredient to (the proof of) the aforementioned Main Theorem;
cmp. \cite[Theorem~3.2.9]{Wei00}.
\subsection{Summary of Contribution}
We establish a quantitative refinement of the Main Theorem
for arbitrary compact metric spaces,
tightly relating moduli of continuity of functions $f:X\to Y$ to those of their
realizers $F$ \emph{relative} to the entropies of co/domains $X$ and $Y$:
Recall {\cite{Kolmogorov},\rm\cite[\S6]{Wei03}} that the \emph{entropy}\footnote{%
From the general logical perspective, entropy constitutes a
\emph{Skolemization} as canonical quantitative refinement of qualitative precompactness \cite{Koh08a}.}
of a compact metric space $(X,d)$ is the mapping $\eta:\mathbb{N}\to\mathbb{N}$ such that
$X$ can by covered by $2^{\eta(n)}$, but not by $2^{\eta(n)-1}$,
closed balls of radius $2^{-n}$.
The real unit interval $[0;1]$ has entropy $\eta(n)=n-1$;
whereas $[0;1]'_1=\operatorname{Lip}_1([0;1],[0;1])$ has entropy $\eta'_1(n)=\Theta(2^n)$;
see Example~\ref{x:Entropy} for further spaces.
\begin{myremark}
\label{r:Admissible}
By Example~\ref{x:Entropy}f), for any modulus of continuity $\kappa$ of a representation
$\xi:\subseteq\mathcal{C}\twoheadrightarrow X$, the space $X$ has entropy $\eta\leq\kappa$;
and we require a \emph{linearly admissible} $\xi$ to (i) have modulus of continuity
$\kappa(n)\leq\mathcal{O}\Big(\eta\big(n+\mathcal{O}(1)\big)\Big)$ almost optimal:
permitting asymptotic `slack' a constant factor in value and constant shift in argument.
Moreover a linearly admissible $\xi$ must satisfy that,
(ii) to every representation $\zeta:\subseteq\mathcal{C}\twoheadrightarrow X$
with modulus of continuity $\nu$ there exists a mapping $F:\operatorname{dom}(\zeta)\to\operatorname{dom}(\xi)$
with modulus of continuity $\mu$ such that $\zeta=\xi\circ F$
\emph{and} $\big(\mu\circ\kappa\big)(n)\leq \nu\big(\mathcal{O}(n)\big)$.
\end{myremark}
This new Condition~(ii) strengthens previous qualitative continuous reducibility ``$\zeta\preccurlyeq_{\rm T}\xi$''
to what we call \emph{linear} metric reducibility ``$\zeta\preccurlyeq_{\rm O}\xi$'',
requiring a $(\zeta,\xi)$-realizer $F$ with almost optimal modulus of continuity:
For functions $\varphi:X\to Y$ and $\psi:Y\to Z$ with respective moduli of continuity $\mu$ and $\kappa$,
their composition $\psi\circ\varphi:X\to Z$ is easily seen to have modulus of continuity $\mu\circ\kappa$.
Abbreviating $\operatorname{lin}(\nu):=\mathcal{O}\Big(\nu\big(\mathcal{O}(n)\big)\Big)$
and with the semi-inverse $\loinv{\nu}(n):=\min\{ m\::\: \nu(m)\geq n\}$,
our results are summarized as follows:
\smallskip
\begin{enumerate}
\item[a)]
Let $(X,d)$ and $(Y,e)$ denote infinite compact metric spaces
with entropies $\eta$ and $\theta$
and equipped with \emph{linearly} admissible representations $\xi$ and $\upsilon$.
If $f:X\to Y$ has modulus of continuity $\mu$,
it admits a realizer $F$ with modulus of continuity
$\operatorname{lin}(\eta)\circ \mu\circ\operatorname{lin}\big(\loinv{\theta}\big)$.
Conversely if $F$ is a realizer of $f$ with modulus $\nu$,
then $f$ has modulus
$\operatorname{lin}\big(\loinv{\eta}\big)\circ\nu\circ\operatorname{lin}(\theta)$.
\item[b)]
Every compact metric space $(X,d)$ admits a linearly admissible representation $\xi$.
For `popular' spaces $X,Y$ having linear/polynomial entropy $\eta,\theta$,
the moduli of continuity of functions and their realizers are thus
linearly/polynomially related; yet according to Examples~\ref{x:Entropy}d+e)
there exist both spaces of entropy growing arbitrarily slow and arbitrarily fast.
Still, estimates (a) are asymptotically tight in a sense explained in Remark~\ref{r:Tight}.
\item[c)]
The category of quantitatively admissible representations is Cartesian closed:
Given linearly admissible representations for spaces $X_j$ ($j\in\mathbb{N}$),
we construct one for the product space $\prod_j X_j$
w.r.t. the `Hilbert Cube' metric $d_{\mathcal{H}}=\sup_j d_j/2^j$
(cmp. Example~\ref{x:Entropy}a)
such that the canonical projections $(x_0,\ldots x_j,\ldots)\mapsto x_j$
and embeddings $x_j\mapsto (x_0,\ldots x_j,\ldots)$
admit realizers with optimal modulus of continuity in the sense of (a).
For the compact space $\mathcal{K}(X)$ of non-empty compact subsets w.r.t. Hausdorff Distance
we construct a canonical \emph{polynomially} admissible representation;
and for the compact space $X'_\mu:=\mathcal{C}_\mu(X,[0;1])$ of functions
$f:X\to[0;1]$ having modulus of continuity $\mu$, equipped with the supremum norm,
one such that the application functional $X'_\mu\times X\ni (f,x)\mapsto f(x)$
has a realizer with optimal modulus of continuity in the sense of (a).
\end{enumerate}
\smallskip\noindent
See Theorems~\ref{t:Main2} and \ref{t:Admissible} and \ref{t:Cartesian} and \ref{t:Functions}
for the precise statements.
Example~\ref{x:SignedDigit} verifies that the signed digit expansion
of the interval $[0;1]$ is linearly admissible;
hence our quantitative ``Main Theorem'' (a) indeed generalizes the real case
as well as quantitatively refining the qualitative Fact~\ref{f:Main}.
\medskip
Note in (a) the typical form of transition maps,
similar for instance to change-of-basis in linear algebra or change-of-chart in differential geometry.
It thus captures the information-theoretic `external' influence of the co/domain
according to Fact~\ref{f:Proper}b), and allows to separate that from the
`inherent' recursion-theoretic complexity of a computational problem:
Informally speaking, an algorithm operating on continuous data is not
to be `blamed' for incurring large cost if the underlying domain has
large entropy, as in Example~\ref{x:Max}b):
see Remark~\ref{r:SecondOrder} below.
\COMMENTED{
\begin{definition}
\label{d:Adapted}
Fix compact metric spaces $(X,d)$ and $(Y,e)$
with entropies $\eta$ and $\theta$ and
with linearly/polynomially admissible representations
$\xi$ and $\upsilon$, respectively.
\emph{Computing} a function $f:X\to Y$
with modulus of continuity $\mu$
in linear/polynomial time
means to compute a $(\xi,\upsilon)$-realizer $F$ of $f$
within time linear/polynomial in
$\operatorname{lin}(\operatorname{id}+\log\eta)\circ \mu\circ\operatorname{lin}\big(\loinv{\operatorname{id}+\log\theta}\big)$.
\end{definition}
This straightforwardly generalizes to \emph{multi}functions \cite{PZ13}.
It consistently `overloads' Definition~\ref{d:Type2} regarding Cantor space,
and recovers the real case $X,Y\in\big\{\mathcal{C},[0;1]\big\}$ having
linear entropies $\eta,\theta$.
}
\subsection{Previous and Related Work, Current Ideas and Methods}
\label{ss:Related}
Computability theory of real numbers was initiated by Alan Turing (1936),
then generalized to real functions by Grzegorzcyk (1957),
to Euclidean subsets by Kreisel and Lacombe (1957),
to Banach Spaces by Pour-El and Richards (1989),
to topological T$_0$ spaces by Weihrauch \cite[\S3.2]{Wei00},
and furthermore to so-called \emph{QCB} spaces by Schr\"{o}der \cite{Sch02}:
These works had introduced the fundamental notions
which in turn have enabled the abundance of investigations
that constitute the contemporary computability theory over continuous data.
\medskip
Computational (bit-)complexity theory of real numbers and functions
was introduced by Harvey Friedman and Ker-I Ko \cite{KF82,Bra05e}.
It differs from the discrete setting for instance by measuring
computational cost in dependence on the output approximation error $2^{-n}$.
Some effort, a careful choice of representation, and the hypothesis
of a compact domain is needed to prove that any total computable real function
actually admits a finite runtime bound depending only on $n$ \cite{Sch04}.
It took even more effort, as well as guidance from discrete
Implicit Complexity Theory \cite{DBLP:journals/siamcomp/KapronC96,Lam06},
to proceed from this Complexity Theory of real functions \cite{Ko91}
to a suitable definition of computational complexity for real operators \cite{KC12}.
The latter involves a modified model of computation discussed in Subsection~\ref{ss:Hyper}.
Again, only the notions introduced in the above works
have enabled the present plethora of investigations
and rigorous classifications of common numerical problems,
such as \cite{Ko91,RW02,BBY06,BBY07a,BGP11,KMRZ15,Poisson}.
And their sensible further generalization to abstract function
(e.g. Sobolev) spaces common in analysis is still in under
development and debate \cite{KSZ16,DBLP:journals/lmcs/Steinberg17}.
Indeed the real co/domain is special in that it has linear entropy;
hence the impact of co/domain on the computational complexity of
problems had been hidden before our quantitative ``Main Theorem'' (a).
In \cite{DBLP:conf/lics/KawamuraS016} we had picked up from \cite{Wei03}
towards a general theory of computational complexity for compact
metric spaces $(X,d)$: exhibiting its entropy $\eta$ as a lower bound
on the bit-cost of real 1-Lipschitz functions $f:X\to[0;1]$,
and constructing a generic representation
with modulus of continuity $\kappa(n)\leq\mathcal{O}\big(n\cdot\eta(n)\big)$
that allows an appropriate (oracle) Type-2 Machine
to compute any fixed such $f$ in time polynomial in $\eta$.
The present work generalizes and extends this as follows:
\begin{itemize}
\item
Theorem~\ref{t:Linear} constructs a generic representation $\xi$
with (i) modulus of continuity $\kappa\leq\operatorname{lin}(\eta)$ \emph{linear} in the entropy $\eta$
\item
and (ii) establishes said $\xi$ \emph{maximal}/complete w.r.t. linear metric reduction
``$\zeta\preccurlyeq_{\rm O}\xi$'' among all uniformly continuous representations $\zeta$ of $X$.
\item
We propose (i) and (ii) as axioms
and quantitative strengthening of classical, qualitative admissibility
for \emph{complexity-}theoretically sensible representations.
\item
Theorem~\ref{t:Main2} strengthens the classical, qualitative Main Theorem
by \emph{quantitatively} characterizing (moduli of) continuous functions $f:X\to Y$
in terms of (moduli of) their $(\xi,\upsilon)$-realizers and the entropies of co/domain $X,Y$.
\item
Theorem~\ref{t:Admissible}d) confirms the classical categorical binary Cartesian product
of representations \cite[Definition~3.3.3.1]{Wei00} to maintain \emph{linear admissibility} (i) and (ii).
\item
The classical categorical countable Cartesian product of representations \cite[Definition~3.3.3.2]{Wei00}
does \emph{not} maintain linear (nor polynomial) admissibility; but our modified construction
exhibited in Theorem~\ref{t:Cartesian} does.
In particular (ii) it is \emph{maximal}/complete w.r.t. linear metric reductions.
\item
Moreover, as opposed to some linearly admissible representation constructed from `scratch'
by invoking Theorem~\ref{t:Linear}, the representation from Theorem~\ref{t:Cartesian} additionally
guarantees the canonical componentwise projections and embeddings
to admit continuous realizers with optimal moduli of continuity.
\item
Theorem~\ref{t:Functions} constructs from any linearly admissible representation $\xi$
of compact metric $(X,d)$ a \emph{polynomially} admissible representation $\xi'_\mu$
of $X'_\mu=\mathcal{C}_\mu(X,[0;1])$.
In particular (ii) it is \emph{maximal}/complete w.r.t. polynomial metric reductions.
\item
Moreover
said representation $\xi'_\mu$
guarantees the application functional $X'_\mu\times X\ni (f,x)\mapsto f(x)\in[0;1]$
to have a continuous realizer with optimal modulus of continuity.
\end{itemize}
Revolving around notions like entropy and modulus of continuity,
our considerations and methods are mostly information-theoretic:
carefully constructing representations and realizers,
analyzing the dependence of their value on the argument,
and comparing thus obtained bounds on their modulus of continuity
to bounds on the entropy of the space under consideration,
estimated from above by constructing coverings with `few' balls of given radius $2^{-n}$
as well as bounded from below by constructing subsets of points
of pairwise distance $>2^{-n}$.
\subsection{Overview}
Section~\ref{s:Admissible} formally introduces our
conception of quantitatively (polynomially/linearly) admissible representations.
Subsection~\ref{ss:Real} re-analyzes the above three representations of $[0;1]$
from this new perspective.
And Subsection~\ref{ss:Topology} collects metric properties of other popular spaces;
including new ones constructed
via binary and countable Cartesian products (w.r.t. Hilbert Cube metric),
the hyperspace of non-empty compact subsets w.r.t. Hausdorff metric,
and function spaces.
Section~\ref{s:Standard} recalls and rephrases
our previous results \cite[\S3]{DBLP:conf/lics/KawamuraS016}
from this new perspective:
constructing a generic polynomially-admissible representation
for any compact metric space $(X,d)$.
And Subsection~\ref{ss:Linear} improves that to linear admissibility.
Section~\ref{s:Category} finally formally states our
complexity-theoretic Main Theorem~\ref{t:Main2} in quantitative detail;
and presents categorical constructions for new
quantitatively admissible representations from given ones,
paralleling the considerations from Subsection~\ref{ss:Topology}:
binary and countable Cartesian products,
Hausdorff hyperspace of non-empty compact subsets, and function spaces.
Section~\ref{s:Conclusion} collects some questions
about possible refinements/strengthenings,
such as improving/dropping constant factors in our results.
Subsection~\ref{ss:Hyper} finally extends our considerations
to generalized representations for higher types in the sense of \cite{KC12};
and Subsection~\ref{ss:Future} puts them into the larger context
of quantitatively-universal compact metric spaces.
\section{Intuition and Definition of Quantitative Admissibility}
\label{s:Admissible}
In order to refine computability to a sensible theory of computational complexity
we propose in this section
two quantitative refinements of qualitative \emph{admissibility}
formalized in Definition~\ref{d:Admissible} below.
But first let us briefly illustrate how a reasonable representation
can be turned into an unreasonable one,
and how that affects the computational complexity of a function:
to get an impression of what quantitative admissibility should prohibit.
\medskip
Consider `padding' a given representation $\xi$ with some
fixed strictly increasing $\varphi:\mathbb{N}\to\mathbb{N}$ in order
to obtain a new representation $\xi_\varphi$ defined by
$\xi_\varphi(\bar b):=\xi\big((b_{\varphi(n)})_{_n}\big)$
for $\bar b=(b_n)_{_n}\in\mathcal{C}$.
Then $\operatorname{dom}(\xi_\varphi)$ is compact whenever $\operatorname{dom}(\xi)$ is;
but computing some $(\xi_\varphi,\upsilon)$-realizer
now may require `skipping' over $\varphi(n)$ bits of any
given $\xi_\varphi$-name before reaching/collecting
the same information as contained in only the first $n$ bits
of a given $\xi$-name when computing a $(\xi,\upsilon)$-realizer:
possibly increasing the time complexity from $t$ to $t\circ\varphi$,
definitely increasing its optimal modulus of continuity.
On the other end computing a $(\xi,\upsilon_\varphi)$-realizer
might become easier, as now as many as $\varphi(n)$ bits of the
padded output can be produced from only $n$ bits of the unpadded one:
possibly decreasing the time complexity from $t$ to $t\circ\upinv{\varphi}$,
see Definition~\ref{d:Admissible}c) below.
\begin{definition}
\label{d:Discrete}
\begin{enumerate}
\item[a)]
We abbreviate $\bar x|_{<n}:=(x_0,\ldots x_{n-1})$ and
\[ (x_0,\ldots x_{n-1})\circ Z^\mathbb{N} \;=\; \big\{(x_0,\ldots x_{n_1},z_n,z_{n+1} \ldots):z_n,z_{n+1} \ldots\in Z\big\} \enspace . \]
\[ \{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^n \ni \vec x = (x_0,\ldots x_{n-1}) \mapsto
\langle \vec x\rangle := (\textup{\texttt{0}}\xspace\,x_0\,\textup{\texttt{0}}\xspace\,x_1\,\textup{\texttt{0}}\xspace x_2\ldots\,\textup{\texttt{0}}\xspace\,x_{n-1}\,\textup{\texttt{1}}\xspace)\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{2n+1} \]
denotes a self-delimiting encoding of finite binary strings.
Let $\lfloor r\rceil\in\mathbb{Z}$ mean the integer closest to $r\in\mathbb{R}$
with ties broken towards 0:
$\big\lfloor \pm n+\tfrac{1}{2}\big\rceil=\pm n$ for $n\in\mathbb{N}$.
\item[b)]
Let $\operatorname{Reg}$ denote the set of all non-decreasing
unbounded mappings $\nu:\mathbb{N}\to\mathbb{N}$.
The \emph{lower} and \emph{upper semi-inverse} of $\nu\in\operatorname{Reg}$ are
\[
\loinv{\nu}(n) \;:=\; \min\{ m\::\: \nu(m)\geq n\}
, \quad \upinv{\nu}(n) \;:=\;
\min\{m \::\: \nu(m+1)>n \}
\enspace . \]
\item[c)]
Extend Landau's class of asymptotic growth
\begin{eqnarray*}
\nu \leq \mathcal{O}(\mu) &\Leftrightarrow&
\exists C\in\mathbb{N} \; \forall n\in\mathbb{N} : \; \nu(n)\leq C\cdot\mu(n)+C \\
\nu \leq \mathcal{P}(\mu) &:=& \mathcal{O}\big(\mu(n)\big)^{\mathcal{O}(1)}
\Leftrightarrow\; \exists C\;\forall n:\; \nu(n)\leq \big(C+C\cdot \mu(n)\big)^C \\
\nu \leq \mathcal{S}(\mu) &:=& \mu+\mathcal{O}(1)
\Leftrightarrow\; \exists C\;\forall n:\; \nu(n)\leq \mu(n)+C \\
\nu\leq \mathcal{\scriptscriptstyle O}(\mu) &:=& \mu\circ\mathcal{O}(\operatorname{id}) \quad\Leftrightarrow\quad \exists C\;\forall n:\; \nu(n)\leq \mu(C\cdot n+C)\\
\nu \leq \mathcal{\scriptscriptstyle P}(\mu) &:=& \mu\circ\mathcal{P}(\operatorname{id}) \quad\Leftrightarrow\quad \exists C\;\forall n:\; \nu(n)\leq \mu(C\cdot n^C+C)\\
\nu \leq \mathcal{\scriptscriptstyle S}(\mu) &:=& \mu\circ\mathcal{S}(\operatorname{id}) \quad\Leftrightarrow\quad \exists C\;\forall n:\; \nu(n)\leq \mu(n+C)\\
\nu \leq \operatorname{lin}(\mu) &:=& \mathcal{O}\big(\mathcal{\scriptscriptstyle O}(\mu)\big)
= \mathcal{\scriptscriptstyle O}\big(\mathcal{O}(\mu)\big)
\;\Leftrightarrow\;\; \exists C\;\forall n:\; \nu(n)\leq C+C\cdot\mu(C+C\cdot n) \\
\nu \leq \operatorname{poly}(\mu) &:=& \mathcal{P}\big(\mathcal{\scriptscriptstyle P}(\mu)\big) \;\Leftrightarrow\; \exists C\;\forall n:\; \nu(n)\leq \big(n+C+C\cdot \mu(C\cdot n^C+C)\big)^C
\end{eqnarray*}
to denote sequences bounded linearly/polynomially/additively by,
and/or after linearly/polynomially/additively growing the argument to, $\mu$.
Here $\operatorname{id}:\mathbb{N}\to\mathbb{N}$ is the identity mapping.
\end{enumerate}
\end{definition}
In Item~c), classes $\operatorname{lin}(\mu)\leq\operatorname{poly}(\mu)$ capture `relative' asymptotics
in increasing granularity. They are transitive and compositional in the following sense:
\begin{observation}
\label{o:Growth}
\begin{enumerate}
\item[a)]
$\mu\leq\mathcal{O}(\nu)$ and $\nu\leq\mathcal{O}(\kappa)$ implies $\mu\leq\mathcal{O}(\kappa)$. \\
$\mu\leq\mathcal{P}(\nu)$ and $\nu\leq\mathcal{P}(\kappa)$ implies $\mu\leq\mathcal{P}(\kappa)$. \\
$\mu\leq\mathcal{S}(\nu)$ and $\nu\leq\mathcal{S}(\kappa)$ implies $\mu\leq\mathcal{S}(\kappa)$.
\item[b)]
$\mu\leq\mathcal{\scriptscriptstyle O}(\nu)$ and $\nu\leq\mathcal{\scriptscriptstyle O}(\kappa)$ implies $\mu\leq\mathcal{\scriptscriptstyle O}(\kappa)$. \\
$\mu\leq\mathcal{\scriptscriptstyle P}(\nu)$ and $\nu\leq\mathcal{\scriptscriptstyle P}(\kappa)$ implies $\mu\leq\mathcal{\scriptscriptstyle P}(\kappa)$. \\
$\mu\leq\mathcal{\scriptscriptstyle S}(\nu)$ and $\nu\leq\mathcal{\scriptscriptstyle S}(\kappa)$ implies $\mu\leq\mathcal{\scriptscriptstyle S}(\kappa)$.
\item[c)]
$\mu\leq\mathcal{O}(\nu)$ implies $\mu\circ\kappa \leq \mathcal{O}(\nu\circ\kappa)$. \\
$\mu\leq\mathcal{P}(\nu)$ implies $\mu\circ\kappa \leq \mathcal{P}(\nu\circ\kappa)$. \\
$\mu\leq\mathcal{S}(\nu)$ implies $\mu\circ\kappa \leq \mathcal{S}(\nu\circ\kappa)$.
\item[d)]
$\mu\leq\mathcal{\scriptscriptstyle O}(\nu)$ implies $\kappa\circ\mu \leq \mathcal{\scriptscriptstyle O}(\kappa\circ\nu)$. \\
$\mu\leq\mathcal{\scriptscriptstyle P}(\nu)$ implies $\kappa\circ\mu \leq \mathcal{\scriptscriptstyle P}(\kappa\circ\nu)$. \\
$\mu\leq\mathcal{\scriptscriptstyle S}(\nu)$ implies $\kappa\circ\mu \leq \mathcal{\scriptscriptstyle S}(\kappa\circ\nu)$.
\item[e)]
$\mathcal{\scriptscriptstyle O}(\mu)\circ\mathcal{O}(\nu)
\;=\; \mathcal{\scriptscriptstyle O}(\mu)\circ\nu
\;=\; \mu\circ\mathcal{O}(\nu)$. \\
$\mathcal{\scriptscriptstyle P}(\mu)\circ\mathcal{P}(\nu)
\;=\; \mathcal{\scriptscriptstyle P}(\mu)\circ\nu
\;=\; \mu\circ\mathcal{P}(\nu)$. \\
$\mathcal{\scriptscriptstyle S}(\mu)\circ\mathcal{S}(\nu)
\;=\; \mathcal{\scriptscriptstyle S}(\mu)\circ\nu
\;=\; \mu\circ\mathcal{S}(\nu)$.
\end{enumerate}
\end{observation}
In particular polynomial `absolute' growth means $\operatorname{poly}(\operatorname{id})=\mathcal{P}(\operatorname{id})=\mathcal{\scriptscriptstyle P}(\operatorname{id})$;
and linear means $\operatorname{lin}(\operatorname{id})=\mathcal{O}(\operatorname{id})=\mathcal{\scriptscriptstyle O}(\operatorname{id})$.
We also collect some properties of the semi-inverses:
\begin{lemma}
\label{l:seminv}
\begin{enumerate}
\item[a)]
For $\mu\in\operatorname{Reg}$, $\loinv{\mu}$ and $\upinv{\mu}$ are again in $\operatorname{Reg}$
and $\loinv{\mu}\leq\upinv{\mu}$.
\item[b)]
Every $\mu\in\operatorname{Reg}$ satisfies
$\loinv{\mu}\circ\mu\;\leq\;\operatorname{id}\;\leq\;\upinv{\mu}\circ\mu$, \\
with equality in case $\mu$ is injective
(necessarily growing at least linearly); \\[0.3ex]
and every $\mu\in\operatorname{Reg}$ satisfies
$\mu\circ\upinv{\mu}\;\leq\;\operatorname{id}\;\leq\mu\circ\loinv{\mu}$, \\
with equality in case $\mu$ is surjective
(necessarily growing at most linearly).
\item[c)]
For $\nu,\kappa\in\operatorname{Reg}$ it holds
\[
\mu\circ\nu\;\leq\;\kappa \;\;\Leftrightarrow\;\;
\nu\;\leq\;\upinv{\mu}\circ\kappa \;\;\Leftrightarrow\;\;
\mu\;\leq\;\kappa\circ\loinv{\nu} \]
\item[d)]
Suppose $a,b,c,d\in\mathbb{N}$ with $b,d\geq1$ and $\mu,\nu\in\operatorname{Reg}$ satisfy
$\forall n: \;\nu(n)\leq a+b\cdot \mu(c+d\cdot n)$.
Then it holds $\forall m: \;\loinv{\mu}(m)\leq c+d\cdot\loinv{\nu}(a+b\cdot m)$
and $\upinv{\mu}(m)\leq d\cdot\upinv{\nu}(a+b\cdot m)+c+d-1$.
\qquad\qquad\qquad
In particular, \\
$\mathcal{O}\big(\upinv{\mu}\big)=\upinv{\mathcal{\scriptscriptstyle O}(\mu)}$
and $\mathcal{\scriptscriptstyle O}\big(\upinv{\mu}\big)=\upinv{\mathcal{O}(\mu)}$
and $\operatorname{lin}\big(\upinv{\mu}\big)=\upinv{\operatorname{lin}(\mu)}$;
similarly
$\mathcal{O}\big(\loinv{\mu}\big)=\loinv{\mathcal{\scriptscriptstyle O}(\mu)}$
and $\mathcal{\scriptscriptstyle O}\big(\loinv{\mu}\big)=\loinv{\mathcal{O}(\mu)}$
and $\operatorname{lin}\big(\loinv{\mu}\big)=\loinv{\operatorname{lin}(\mu)}$.
\item[e)]
There are $\mu,\nu\in\operatorname{Reg}$ s.t. $\mu\leq\operatorname{poly}(\nu)$
but $\loinv{\nu}\not\leq\operatorname{poly}\big(\loinv{\mu}\big)$
and $\upinv{\nu}\not\leq\operatorname{poly}\big(\upinv{\mu}\big)$.
There are $\mu,\nu\in\operatorname{Reg}$ s.t.
$\mu\circ\loinv{\mu}\not\in\operatorname{poly}(\operatorname{id})$
and $\upinv{\nu}\circ\nu\not\in\operatorname{poly}(\operatorname{id})$
and $\loinv{\nu}\circ\mathcal{O}(\nu)\not\in\operatorname{poly}(\operatorname{id})$
and $\mu\circ\mathcal{O}\big(\upinv{\mu}\big)\not\in\operatorname{poly}(\operatorname{id})$.
\item[f)]
$\max\big\{\upinv{\mu},\upinv{\nu}\big\} =\upinv{\min\{\mu,\nu\}}$ and
$\min\big\{\upinv{\mu},\upinv{\nu}\big\} =\upinv{\max\{\mu,\nu\}}$ and
$\max\big\{\loinv{\mu},\loinv{\nu}\big\} =\loinv{\min\{\mu,\nu\}}$ and
$\min\big\{\loinv{\mu},\loinv{\nu}\big\} =\loinv{\max\{\mu,\nu\}}$.
\end{enumerate}
\end{lemma}
Here finally comes our formal definition of quantitative admissibility:
\begin{definition}
\label{d:Admissible}
\begin{enumerate}
\item[a)]
Consider a compact subset $K$ of a metric space $(X,d)$.
Its \emph{relative entropy}
is the non-decreasing integer mapping
$\eta=\eta_{X,K}:\mathbb{N}\to\mathbb{N}$ such that some $2^{\eta(n)}$,
but no $2^{\eta(n)-1}$, closed balls $\overline{\operatorname{B}}(x,r)$
of radius $r:=2^{-n}$ with centers $x\in X$ can cover $K$.
If $X$ itself is compact, we write $\eta_X:=\eta_{X,X}$
for its (intrinsic) entropy.
\item[b)]
Consider a uniformly continuous representation $\xi$ of the compact metric space $(X,d)$
and uniformly continuous mapping $\zeta:\subseteq\mathcal{C}\to X$.
Refining Definition~\ref{d:TTE}e), call a reduction $F:\operatorname{dom}(\zeta)\to\operatorname{dom}(\xi)$
\emph{polynomial} (``$\zeta\preccurlyeq_{\rm P}\xi$'') if it has a modulus of continuity $\mu$
and $\xi$ has a modulus $\kappa$ satisfying
$\mu\circ\kappa\leq\mathcal{\scriptscriptstyle P}(\nu)$ for every modulus $\nu$ of $\zeta$.
\\
$F$ is \emph{linear} (``$\zeta\preccurlyeq_{\rm O}\xi$'') if it holds $\mu\circ\kappa\leq\mathcal{\scriptscriptstyle O}(\nu)$.
\item[c)]
A representation $\xi$ of the compact metric space $(X,d)$ is
\emph{polynomially admissible} iff (i) it has a modulus of continuity
$\kappa\leq\mathcal{P}\mathcal{\scriptscriptstyle O}(\eta)$, i.e.,
bounded polynomially in the entropy with linearly transformed argument,
and (ii) every uniformly continuous representation $\zeta:\subseteq\mathcal{C}\twoheadrightarrow X$
satisfies $\zeta\preccurlyeq_{\rm P}\xi$ in the sense of (b).
\item[d)]
Representation $\xi$ of $(X,d)$ is \emph{linearly admissible} iff (i) it has a modulus of continuity
$\kappa\leq\mathcal{O}\mathcal{\scriptscriptstyle S}(\eta)$, i.e., not exceeding
the entropy by more than a constant factor in value and constant shift in argument,
and (ii) every uniformly continuous representation $\zeta:\subseteq\mathcal{C}\twoheadrightarrow X$
satisfies $\zeta\preccurlyeq_{\rm O}\xi$.
\end{enumerate}
\end{definition}
According to Example~\ref{x:Entropy}f) below,
any representation's modulus of continuity satisfies $\kappa\geq\eta$,
i.e., is bounded from below by the entropy;
and Condition~(i) in Definition~\ref{d:Admissible}c+d)
requires a complexity-theoretically appropriate representation
to be close to that optimum ---
which itself can be arbitrarily small/large according to Example~\ref{x:Entropy}d+e).
The converse Condition~(ii) in Definition~\ref{d:Admissible}c+d)
similarly requires that $\mu\circ\kappa$, a modulus of continuity of $\xi\circ F=\zeta$
be `close' to that of $\zeta$.
Note that linear admissible representations may (i) exceed the entropy by
a constant factor in value and by an additive constant in the argument
while (ii) linear reduction only allows for the latter:
because (i) is what we can achieve in Theorem~\ref{t:Linear}
while (ii) guarantees transitivity; similarly for the polynomial case.
\begin{myremark}
\label{r:Transitive}
We record that relations
``$\preccurlyeq_{\rm P}$'' and ``$\preccurlyeq_{\rm O}$'' are transitive:
Fix $\alpha$ with modulus of continuity $\lambda$,
$\beta$ with modulus $\mu$, and
$\gamma$ with $\nu$;
and linear reduction $F:\operatorname{dom}(\alpha)\to\operatorname{dom}(\beta)$ with modulus of continuity $\iota$
such that $\alpha=\beta\circ F$ and $\iota\circ\mu\leq\mathcal{\scriptscriptstyle O}(\lambda)$;
as well as linear reduction $G:\operatorname{dom}(\beta)\to\operatorname{dom}(\gamma)$ with modulus of continuity $\kappa$
such that $\beta=\gamma\circ G$ and $\kappa\circ\nu\leq\mathcal{\scriptscriptstyle O}(\mu)$.
Then $\alpha=\gamma\circ G\circ F$, where reduction $G\circ F:\operatorname{dom}(\alpha)\to\operatorname{dom}(\gamma)$
has modulus $\iota\circ\kappa$ satisfying
$(\iota\circ\kappa)\circ \nu\leq \iota\circ \mathcal{\scriptscriptstyle O}(\mu)\leq\mathcal{\scriptscriptstyle O}(\lambda)$
by Observation~\ref{o:Growth}e).
\\
Also note that, according to Lemma~\ref{l:seminv}d),
condition $\mu\circ\kappa\leq\mathcal{\scriptscriptstyle O}(\nu)$ is equivalent to
$\mu\leq\nu\circ\mathcal{O}\big(\loinv{\kappa}\big)=\mathcal{\scriptscriptstyle O}(\nu)\circ\loinv{\kappa}$;
similarly for
$\mu\circ\kappa\leq\mathcal{\scriptscriptstyle P}(\nu)$.
\end{myremark}
Definition~\ref{d:Admissible}c) coincides with \cite[Definition~18]{DBLP:conf/lics/KawamuraS016}
and strengthens the computability-theoretically common qualitative notion of \emph{admissibility}
from \cite[Definitions~2.1.1+2.1.2]{Wei00};
while Definition~\ref{d:Admissible}d) in turn strengthens (c) from polynomial to linear.
\begin{proof}[Lemma~\ref{l:seminv}]
Most claims are immediate, we thus only expand on Item~d):
\begin{enumerate}
\item[d)]
First suppose $c=0$ and $d=1$. Then
\begin{eqnarray*}
\loinv{\mu}(m) &=& \min\underbrace{\{n:a+b\cdot \mu(n)\geq a+b\cdot m\}}_{
\rotatebox[origin=c]{270}{$\supseteq$}}
\\ &\leq& \min\overbrace{\{n:\nu(n)\geq a+b\cdot m\}} \;=\; \loinv{\nu}(a+b\cdot m) \quad\text{ and} \\[1ex]
\upinv{\mu}(m) &=& \min\underbrace{\{n:a+b\cdot \mu(n+1)> a+b\cdot m\}}_{
\rotatebox[origin=c]{270}{$\subseteq$}}
\\ &\leq& \min\{n:\nu(n+1)> a+b\cdot m\} \;=\; \upinv{\nu}(a+b\cdot m) \enspace . \\[1ex]
\loinv{\mu}(m) &=& \min\underbrace{\{n:\mu(n)\geq m\}}_{
\rotatebox[origin=c]{270}{$\supseteq$}}
\\ &\leq& \min\overbrace{\{c+d\cdot n':\mu(c+d\cdot n')\geq m\}}
\;\leq\;
\min\{c+d\cdot n':\nu(n')\geq m\}
\end{eqnarray*}
$=\; c+d\cdot\loinv{\nu}(m)$ in case $a=0$ and $b=1$. If additionally $d=1$, then
\begin{eqnarray*}
\upinv{\mu}(m)&=&
\min\big\{n:\mu(n+1)> m\big\}
\;\leq\; \min\big\{c+n:\mu(c+n+1)> m\} \\
&\leq& \min\big\{c+n:\nu(n+1)> m\}
\;=\; c+\upinv{\nu}(m)
\end{eqnarray*}
Finally in case $a=0=c$ and $b=1\neq d$,
\begin{eqnarray*}
\upinv{\mu}(m)&=&
\min\big\{n:\mu(n+1)> m\big\}
\;\leq\; \min\big\{d\cdot \lceil n/d\rceil :\mu(n+1)> m\big\} \\
&\leq& \min\big\{d\cdot n' + (d-1) :\mu(c\cdot n'+1)> m\big\} \\
&\leq& \min\big\{d\cdot n' + (d-1) :\nu(n'+1)> m\big\}
\end{eqnarray*}
$=\; d\cdot\upinv{\nu}(m)+(d-1)$.
\qed\end{enumerate}\end{proof}
\subsection{Real Examples}
\label{ss:Real}
Here we formally recall, and analyze from the perspective of admissibility,
the three representations of the real unit interval mentioned in the introduction:
binary, dyadic and signed binary.
Let us record that the real unit interval $[0;1]$
has entropy $\eta_{[0;1]}(n)=n-1$ for all integers $n\geq1$.
\begin{myexample}[Binary Representation] \label{x:Binary}
The binary representation of the real unit interval
\[
\beta: \; \mathcal{C} \;\ni\; \bar b \;\mapsto\; \sum\nolimits_n b_n2^{-n+1} \;\in\; [0;1]
\]
is surjective and 1-Lipschitz, i.e., has the identity $\operatorname{id}:\mathbb{N}\ni n\mapsto n\in\mathbb{N}$ as modulus of continuity:
coinciding with the entropy up to shift 1, i.e., optimal!
However it is not (even qualitatively) admissible {\rm\cite[Theorem~4.1.13.6]{Wei00}},
does not admit a continuous realizer of, e.g.,
the continuous mapping $[0;1/3]\ni x\mapsto 3x\in[0;1]$;
cmp. {\rm\cite[Example~2.1.4.7]{Wei00}}.
\end{myexample}
\begin{myexample}[Rational Representation] \label{x:Rational}
Consider the binary encoding of non-negative integers without leading $\textup{\texttt{0}}\xspace$:
\begin{equation}
\label{e:Binary}
\mathrm{bin} \;:\; \mathbb{N} \;\ni\; 2^n-1+\sum\nolimits_{0\leq j<n} b_j2^j \;\mapsto\; (b_0,\ldots b_{n-1})\; \in\; \{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^n \enspace .
\end{equation}
The \emph{rational} representation of $[0;1]$ is the mapping
$\rho:\subseteq\mathcal{C}\twoheadrightarrow[0;1]$ with
\begin{multline*}
\big(
\langle\mathrm{bin}(a_0)\rangle\: \langle\mathrm{bin}(c_0)\rangle\;
\langle\mathrm{bin}(a_1)\rangle\: \langle\mathrm{bin}(c_1)\rangle\;
\ldots
\langle\mathrm{bin}(a_n)\rangle\: \langle\mathrm{bin}(c_n)\rangle\;
\ldots \big)
\;\mapsto\; \lim\nolimits_j a_j/c_j , \\[0.5ex]
\operatorname{dom}(\rho) \:=\: \big\{
\big(
\ldots\:
\langle\mathrm{bin}(a_n)\:\langle\mathrm{bin}(c_n)\;\rangle\:\ldots\big) \::\:
\exists x\in[0;1] \: |a_n/c_n-x|\leq 2^{-n}\big\} \enspace .
\end{multline*}
Representation $\rho$ is continuous, but not uniformly continuous
(its domain is not compact) and thus has no modulus of continuity.
\end{myexample}
\begin{proof}
Consider a $\rho$-name of $r=1/2$
starting with any $a_0\in\mathbb{N}$ and $b_0:=2a_0$.
Increasingly long $a_0$ thus give rise to a sequence of $\rho$-names of $r$
with no converging subsequence.
Moreover $a_0/b_0=1/2$ fixes $r$ up to error $2^{-n}$ only for $n:=0$;
but requires `knowing' the first $\mu(0)\geq\log_2(a_0)\to\infty$ bits of its $\rho$-name.
\qed\end{proof}
\begin{myexample}[Dyadic Representation] \label{x:Dyadic}
The \emph{dyadic representation} of the real unit interval $[0;1]$
\begin{multline*}
\delta:\subseteq\mathcal{C}\;\ni\;\big(
\langle\mathrm{bin}(a_0)\rangle\:
\langle\mathrm{bin}(a_1)\rangle\:
\langle\mathrm{bin}(a_2)\rangle\:
\ldots
\langle\mathrm{bin}(a_n)\rangle\:
\ldots \big)
\;\mapsto\; \lim\nolimits_j a_j/2^j , \\[0.5ex]
\operatorname{dom}(\delta) \:=\: \big\{
\big(
\ldots\:
\langle\mathrm{bin}(a_n)\rangle\:\ldots\big) \::\:
2^n \geq a_n\in\mathbb{N}, \;
|a_n/2^n-a_m/2^m|\leq 2^{-n}+2^{-m}\big\}
\end{multline*}
\begin{enumerate}
\item[i)] has a quadratic modulus of continuity $\kappa(n):=2\cdot(n+1)\cdot(n+2)$
but no sub-quadratic one and in particular is not H\"{o}lder-continuous.
\item[ii)] To every partial function $\zeta:\subseteq\mathcal{C}\to [0;1]$
with modulus of continuity $\nu$
there exists a mapping $F:\operatorname{dom}(\zeta)\to\operatorname{dom}(\delta)$
with modulus of continuity $\nu$
such that $\zeta=\delta\circ F$ holds. \\
In particular $\delta$ is polynomially admissible.
\item[iii)]
To every $m\in\mathbb{N}$ and every $r,r'\in[0;1]$ with $|r-r'|\leq2^{-m-1}$,
there exist $\delta$-names $\bar y_r$ and $\bar y'_{r'}$
of $r=\delta(\bar y_r)$ and $r'=\delta(\bar y'_{r'})$
with $d_{\mathcal{C}}(\bar y_r,\bar y'_{r'})\leq2^{-m-1}$.
\item[iv)]
If $(Y,e)$ is a compact metric space and
$f:[0;1]\to Y$ such that $f\circ\delta:\operatorname{dom}(\delta)\subseteq\mathcal{C}\to Y$
has modulus of continuity $\nu$,
then $f$ has modulus of continuity $\nu$.
\end{enumerate}
\end{myexample}
Here and as opposed to Definition~\ref{d:Admissible},
(ii) applies also to non-surjective $\zeta$.
\begin{proof}
\begin{enumerate}
\item[i)]
Record
that $0\leq a_n \leq 2^n$ implies $\langle\mathrm{bin}(a_n)\rangle\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^*$
to have length between 1 and $2n+1$; and
$\langle\mathrm{bin}(a_0)\rangle\: \ldots \langle\mathrm{bin}(a_{n})\rangle$
has binary length between $n+1$ and $\kappa(n)$.
Therefore perturbing a $\delta$-name $\bar y$ of some $r\in[0;1]$
to $\bar y'$ with $d_{\mathcal{C}}(\bar y,\bar y')\leq2^{-\kappa(n)}$
will keep (the binary expansions of) $a_0,\ldots a_{n}$ unmodified;
and thus satisfies $|\delta(\bar y)-\delta(\bar y')|\leq 2^{-n}$.
Hence $\kappa(n)$ is a modulus of continuity.
On the other hand consider the $\delta$-name $\bar y$
of $r:=3/4$ with $a_n:=3\cdot 2^{n-2}$ for $n\neq2$
has $\langle\mathrm{bin}(a_n)\rangle$ of length $2n-1$
and starts at bit position $\sum_{m<n} (2m-1)\geq\Omega(n^2)$;
yet changing $a'_{m}:\equiv3\cdot 2^{m-2}+2^{m-n}$ for all $m>n$
turns it into a $\delta$-name $\bar y'$ of $r':=r+2^{-n}$
with $d_{\mathcal{C}}(\bar y,\bar y')\leq2^{\mathcal{O}(n^2)}$.
So $\delta$ has no sub-quadratic modulus of continuity.
\item[ii)]
We construct $F:\operatorname{dom}(\zeta)\to\operatorname{dom}(\delta)$
as limit $F(\bar x)=\lim_n F_n\big(\bar x|_{<\nu(n+1)}\big)$
of a sequence of partial functions
$F_n:\subseteq\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{\nu(n+1)}\to\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{<\kappa(n)}$
which is monotone in that $F_{n+1}\big(\bar x|_{<\nu(n+2)}\big)$
contains $F_{n}\big(\bar x|_{<\nu(n+1)}\big)$ as initial segment.
To every $n$ and each (of the finitely many) $\bar x|_{<\nu(n+1)}$ with $\bar x\in\operatorname{dom}(\zeta)$,
fix some $r_n=r_n\big(\bar x|_{<\nu(n+1)}\big)\in\zeta\big[\bar x|_{<\nu(n+1)}\circ\mathcal{C}]\subseteq[0;1]$.
Then given $\bar x\in\operatorname{dom}(\delta)$ and iteratively for $n=0,1,\ldots$ let
\[ F_{n}\big(\bar x|_{<\nu(n+1)}\big) \;:=\; F_{n-1}\big(\bar x|_{<\nu(n)}\big) \:\circ\: \langle\mathrm{bin}(a_n)\rangle,
\quad a_n:=\lfloor r_n\cdot2^n\rceil\in\{0,\ldots 2^n\} \enspace . \]
Since $\nu$ is a modulus of continuity of $\zeta$, it follows
\[ \zeta\big[\bar x|_{<\nu(n+1)}\circ\mathcal{C}\big] \;\subseteq\; \big[r_n-2^{-n-1};r_n+2^{-n-1}\big]
\;\subseteq\; \big[\tfrac{a_n}{2^n}-2^{-n};\tfrac{a_n}{2^n}+2^{-n}\big] \]
as $|r_n-a_n/2^n|\leq2^{-n-1}$. Thus it holds
$|a_n/2^n-r|\leq2^{-n}$ for $r:=\zeta(\bar x)$ since $|r-r_n|\leq2^{-n-1}$;
$F(\bar x) = \lim\nolimits_n F_n(\bar x_n) =
\big(\langle\mathrm{bin}(a_0)\rangle\:\ldots\: \langle\mathrm{bin}(a_n)\rangle\: \ldots \big)$
is a $\delta$-name of $r$;
and fixing the first $\nu(n+1)$ symbols of $\bar x\in\operatorname{dom}(\zeta)$
fixes $\vec y:=F_n\big(\bar x|_{<\nu(n+1)}\big)$ as well as $(a_0,\ldots a_n)$
and therefore also (at least) the first $n+1$ symbols of $\bar y:=F(\bar x)$:
hence $F$ has modulus of continuity $\nu(n)$.
\\
As recorded above, $[0;1]$ has entropy $\eta(n)=n-1$;
hence (i) and (ii) imply polynomial admissibility
according to Definition~\ref{d:Admissible}c).
\item[iii)]
xTo $r\in [0;1]$ consider the $\delta$-name $\bar y_r:=\big(\ldots \langle\mathrm{bin}(a_n)\rangle\:\ldots \big)\in\mathcal{C}$ of $r$
with $a_n:=\lfloor r\cdot2^n\rceil$.
For every $m\in\mathbb{N}$, its initial segment
$\big( \langle\mathrm{bin}(a_0)\rangle\: \ldots \langle\mathrm{bin}(a_m)\rangle\big)$
has binary length between $m+1$ and $\kappa(m)$;
and, for every $r'\in[0;1]$ with $|r-r'|\leq2^{-m-1}$,
can be extended to a $\delta$-name $\bar y'_{r'}$
via $a'_{m'}:=\lfloor r'\cdot2^{m'}\rceil$ for all $m'>m$.
\item[iv)]
Applying (iii) to $m:=\nu(n)$, the hypothesis implies
$e\big(f(r),f(r')\big)=e\big(f\circ\delta(\bar y_r),f\circ\delta(\bar y'_{r'})\big)\leq 2^{-n}$.
\qed\end{enumerate}\end{proof}
The dyadic representation in Example~\ref{x:Dyadic} has modulus of continuity
quadratic (i.e. polynomial), but not linear, in the entropy. This overhead comes
from the `redundancy' of the precision-$n$ approximation $a_n/2^n$ of binary length
$\mathcal{O}(n)$ superseding all previous $a_m/2^m$, $m<n$. The signed binary representation
on the other hand achieves precision $2^{-n}$ by appending one `signed' digit
$\tilde b_{n-2}\in\{-1,0,1\}$, encoded as two binary digits $(b_{2n-4},b_{2n-3})\in\{\textup{\texttt{0}}\xspace\sdzero,\textup{\texttt{0}}\xspace\textup{\texttt{1}}\xspace,\textup{\texttt{1}}\xspace\textup{\texttt{0}}\xspace\}$
via $\tilde b_{n-2}=2b_{2n-4}+b_{2n-3}-1$, to the previous
approximation up to error $2^{-n+1}$, yielding a modulus of continuity linear in the entropy:
\begin{myexample}[Signed Binary Represent.] \label{x:SignedDigit}
The signed binary representation, considered as total mapping
\begin{equation}
\label{e:SignedDigit}
\sigma:\subseteq\{\textup{\texttt{0}}\xspace\sdzero,\textup{\texttt{0}}\xspace\textup{\texttt{1}}\xspace,\textup{\texttt{1}}\xspace\textup{\texttt{0}}\xspace\}^\mathbb{N}\subseteq\mathcal{C}
\;\ni \bar b \mapsto\;
\tfrac{1}{2}+\sum\limits_{m=0}^{\infty} (2b_{2m}+b_{2m+1}-1) \cdot 2^{-m-2}
\;{\in}\; [0;1]
\end{equation}
\begin{enumerate}
\item[i)]
is surjective and has modulus of continuity $\kappa(n)=2n$,
i.e., is H\"{o}lder-continuous.
\item[ii)]
To every partial function $\zeta:\subseteq\mathcal{C}\to [0;1]$
with modulus of continuity $\nu$
there exists a mapping $F:\operatorname{dom}(\zeta)\to\operatorname{dom}(\sigma)$
with modulus of continuity $\kappa:2m\mapsto\nu(m+1)$
such that $\zeta=\sigma\circ F$ holds. \\
In particular $\sigma$ is linearly admissible.
\item[iii)]
To every $n\in\mathbb{N}$ and every $r,r'\in[0;1]$ with $|r-r'|\leq2^{-n}$,
there exist $\sigma$-names $\bar y_r$ and $\bar y'_{r'}$
of $r=\sigma(\bar y_r)$ and $r'=\sigma(\bar y'_{r'})$
with $d_{\mathcal{C}}(\bar y_r,\bar y'_{r'})\leq2^{-2n}$.
\item[iv)]
If $(Y,e)$ is a compact metric space and
$f:[0;1]\to Y$ such that $f\circ\sigma:\operatorname{dom}(\sigma)\subseteq\mathcal{C}\to Y$
has modulus of continuity $2\nu$, then $f$ has modulus of continuity $\nu$.
\end{enumerate}
\end{myexample}
\begin{proof}[Example~\ref{x:SignedDigit}]
\begin{enumerate}
\item[i)]
Note that appropriate choice of $\tilde b_{0},\tilde b_{1},\ldots\in\{-1,0,1\}$
yields precisely any possible value $[-1/2;+1/2]\ni\sum_{m=0}^\infty \tilde b_{m-2} \cdot 2^{-m}$;
hence $\sigma$ is total and surjective. In one worst case,
changing $(b_{2m},b_{2m+1})=\textup{\texttt{0}}\xspace\sdzero$ (encoding the signed digit $-1$)
to $(b'_{2m},b'_{2m+1})=\textup{\texttt{1}}\xspace\textup{\texttt{0}}\xspace$ (encoding $+1$) for all $m\geq n$
changes the real number $r$ from Equation~(\ref{e:SignedDigit})
to $r'=r+\sum\nolimits_{m\geq n} (2) \cdot 2^{-m-2}=r+2^{-n-1}$;
while $\bar b'$ agrees with $\bar b$ up to position $2n-1$:
This asserts $\kappa(n)=2n$ to be a modulus of continuity and,
in view of $[0;1]$ having entropy $\eta(n)=n-1$,
establishes Condition~(i) of Definition~\ref{d:Admissible}b).
\item[ii)]
Similarly to the proof of Example~\ref{x:Dyadic},
for every $n$ and every $\bar x|_{<\nu(n+1)}$ with $\bar x\in\operatorname{dom}(\zeta)$,
consider the compact set $\zeta\big[\bar x|_{<\nu(n+1)}\circ\mathcal{C}\big]\subseteq[0;1]$:
Having diameter $\leq2^{-(n+1)}$ by the definition of $\nu$,
it is contained in $[r_n-2^{-n-2};r_n+2^{-n-2}]$ for
$r_n:=\big(\min\zeta\big[\bar x|_{<\nu(n+1)}\circ\mathcal{C}\big]+\max\zeta\big[\bar x|_{<\nu(n+1)}\circ\mathcal{C}\big]\big)/2\in[0;1]$;
hence $|r_{n+1}-r_n|\leq2^{-n-2}$.
Now let $r'_1:=\tfrac{1}{2}$ capture the constant term $\tilde b_{-1}:=-1$ in Equation~(\ref{e:SignedDigit})
such that $|r_1-r'_1|\leq\tfrac{1}{4}$ with $r_1\in\big(\tfrac{1}{4};\tfrac{3}{4}\big)$;
and for $n=2,3,\ldots$ inductively append one additional signed digit
\[
2b_{2n-4}+b_{2n-3}-1\;=\;\tilde b_{n-1}\;:=\;\lfloor 2^{n}\cdot(\underbrace{r_n-r'_{n-1}}_{\leq\pm3\cdot2^{-n-1}}
)\rceil\in\{-1,0,+1\} \]
such that $r'_n\;:=\;\tfrac{1}{2}+\sum\nolimits_{m=1}^{n} \tilde b_{m-2} \cdot 2^{-m}$
again satisfies $|r_n-r'_n|\leq2^{-n-1}$ and $|r_{n+1}-r'_n|\leq|r_{n+1}-r_n|+|r_n-r'_n|\leq3\cdot2^{-n-2}$
and $|r-r'_n|\leq|r-r_n|+|r_n-r'_n|\leq2^{-n}$ for $r=\zeta(\bar x)\in\zeta\big[\bar x|_{<\nu(n+1)}\circ\mathcal{C}\big]$:
Hence $F(\bar x):=(b_0,\ldots b_{2n-4},b_{2n-3},\ldots)$ is a $\sigma$-name of $r$,
and the thus defined function $F$ has modulus of continuity $2n\mapsto\nu(n+1)$.
\item[iii)]
To $r\in [0;1]$ and $n\in\mathbb{N}$ consider signed digits
$\tilde b_0,\ldots \tilde b_{n-2}\in\{-1,0,1\}$
and $r'_n\;:=\;\tfrac{1}{2}+\sum\nolimits_{m=1}^{n} \tilde b_{m-2} \cdot 2^{-m}$
with $|r-r'_n|\leq2^{-n-1}$ as in (ii).
As in (i), appropriate choice of $\tilde b_{n-1},\tilde b_{n},\ldots\in\{-1,0,1\}$
yields any possible value $[-2^{-n};+2^{-n}]\ni\sum_{m=n-1}^\infty \tilde b_{m-2} \cdot 2^{-m}$;
hence every $r'\in[0;1]$ with $|r-r'|\leq2^{-n}$ admits a signed binary expansion
$r'=\tfrac{1}{2}+\sum_{m=1}^{\infty} \tilde b_{m-2}\cdot2^{-m}$
extending $(\tilde b_0,\ldots \tilde b_{n-2})$,
and $\sigma$-name $\bar y_{r'}$ coinciding on the first $2n$ binary symbols.
\item[iv)]
Applying (iii) to $n:=\nu(m)$, the hypothesis implies
$e\big(f(r),f(r')\big)=e\big(f\circ\sigma(\bar y_r),f\circ\sigma(\bar y'_r)\big)\leq 2^{-m}$.
\qed\end{enumerate}\end{proof}
The signed binary representation renders real addition
computable by a finite-state transducer:
\begin{center}
\includegraphics[width=0.9\textwidth]{transducer}
\end{center}
Starting off in state \texttt{C},
in each round $\#n=0,1,\ldots$ it reads the next signed digits $a_{n},b_{n}\in\{-1,0,1\}$
in the respective expansions of real arguments $x=\sum_n a_n2^{-n}$ and $y=\sum_n b_n2^{-n}$,
and follows that edge whose \emph{first} label agrees with $a_n+b_n$
while outputting the \emph{second} label $c_{n-2}\in\{-1,0,1\}$
of said edge such that $x+y=\sum_n c_n2^{-n}$.
The transducer works by storing for each state the accumulated value from previous input except those already output.
\medskip
The signed-digit expansion's modulus of continuity
leaves a constant-factor gap to the entropy, attained by the binary expansion.
One can trade between both, namely permit $\textup{\texttt{\={1}}}\xspace$ only at asymptotically fewer positions (i)
while incurring possible `carry ripples' between them over asymptotically longer ranges (ii):
\begin{myexample} \label{x:Gleb}
Fix a strictly increasing function $\varphi:\mathbb{N}\to\mathbb{N}$ with $\varphi(0)=0$.
Representation $\sigma_\varphi$
`interpolates' between Examples~\ref{x:Binary} and \ref{x:SignedDigit}
by considering signed binary expansions
$\sum\limits_{m=1}^{\infty} \tilde c_{m-1} \cdot 2^{-m}$
with $\tilde c_{m}\in\{\textup{\texttt{\={1}}}\xspace,\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}$ for every $m\in\operatorname{range}\varphi=\varphi[\mathbb{N}]$
but $\tilde c_m\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}$ for all $m\in\mathbb{N}\setminus\varphi[\mathbb{N}]$.
Each $\tilde c_m$ with $m\in\mathbb{N}\setminus\varphi[\mathbb{N}]$ is encoded as one bit,
each $\tilde c_m$ with $m\in\varphi[\mathbb{N}]$ as two bits.
Thus $\sigma_{\operatorname{id}}=\sigma$ recovers Example~\ref{x:SignedDigit}.
\begin{enumerate}
\item[i)]
$\sigma_\varphi$ is surjective and has modulus of continuity $n\mapsto n+\upinv{\varphi}(n)$.
\item[ii)]
There exists a mapping $F_\varphi:\operatorname{dom}(\sigma)\to\operatorname{dom}(\sigma_\varphi)$
with modulus of continuity $2\varphi\circ\upinv{(\operatorname{id}+\varphi)}$
such that $\sigma=\sigma_\varphi\circ F_\varphi$ holds.
\end{enumerate}
In particular $\sigma_{n^2}:=\sigma_{n\mapsto n^2}$ has modulus of continuity $n+\mathcal{O}(\sqrt{n})$;
and to every partial function $\zeta:\subseteq\mathcal{C}\to [0;1]$
with modulus of continuity $\nu$
there exists a mapping $F_{n^2}:\operatorname{dom}(\zeta)\to\operatorname{dom}(\sigma_{n^2})$
with modulus of continuity $n\mapsto\nu(n+1)+\mathcal{O}\big(\sqrt{\nu(n+1)}\big)$
such that $\zeta=\sigma_{n^2}\circ F_{n^2}$.
\end{myexample}
Note that $\varphi(n):=2^n$ has
$\upinv{(\operatorname{id}+\varphi)}(n)\geq \log_2(n)+1$ infinitely often
and therefore $\varphi\circ\upinv{(\operatorname{id}+\varphi)}(n)\geq 2\cdot n$:
yielding a linear reduction $\zeta\preccurlyeq_{\rm O}\sigma_\varphi$,
but no better.
\begin{proof}[Example~\ref{x:Gleb}]
\begin{enumerate}
\item[i)]
Similarly to the proof of Example~\ref{x:SignedDigit}i),
the first $n$ digits $\tilde c_0,\ldots \tilde c_{n-1}$ of an expansion
fix the value up to absolute error $<2^{-n}$.
Differing from Example~\ref{x:SignedDigit},
this initial segment of the expansion occupies not $2n$
but $n+\upinv{\varphi}(n)$ bits since `signed' digits (permitted)
only at the $\upinv{\varphi}(n)$ positions $\varphi[\mathbb{N}]\cap\{0,\ldots n-1\}$.
\item[ii)]
We describe a transformation $F_\varphi$ converting a given signed-digit expansion
$r=\tfrac{1}{2}+\sum\limits_{m=1}^{\infty} \tilde b_{m-2} \cdot 2^{-m}$
with $\tilde b_m\in\{\textup{\texttt{\={1}}}\xspace,\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}$ to the required form
$r=\sum\limits_{m=1}^{\infty} \tilde c_{m-1} \cdot 2^{-m}$
with $\tilde c_m\neq\textup{\texttt{\={1}}}\xspace$ except for positions $m\in\varphi[\mathbb{N}]$.
Reflecting the constant term in Equation~(\ref{e:SignedDigit}),
initially let $c_0:=\textup{\texttt{1}}\xspace$, tentatively.
Now iteratively for $k=1,2,\ldots$ re-code the signed integer
\begin{eqnarray*}
-2^{\varphi(k)}+1 &=&
0 \:- 2^{\varphi(k)-1} \:- 2^{\varphi(k)-2} \:-\ldots\: -2 \:-1 \\
&\leq&
2^{\varphi(k)}\cdot c_{\varphi(k-1)}\:+\:
2^{\varphi(k)-1}\cdot \tilde b_{\varphi(k-1)} \:+\:
2^{\varphi(k)-2}\cdot \tilde b_{\varphi(k-1)+1} \:+\ldots \\
&& \ldots+\: 2 \cdot \tilde b_{\varphi(k)-2} \:+\: \tilde b_{\varphi(k)-1} \\
&=:&
2^{\varphi(k)}\cdot \tilde c_{\varphi(k-1)} \:+\:
2^{\varphi(k)-1}\cdot \tilde c_{\varphi(k-1)+1} \:+\: 2^{\varphi(k)-2}\cdot \tilde c_{\varphi(k-1)+2} \:+\ldots \\
&& \ldots+\: 2 \cdot \tilde c_{\varphi(k)-1} \:+\: c_{\varphi(k)}
\end{eqnarray*}
uniquely with $\tilde c_{\varphi(k-1)}\in\{\textup{\texttt{\={1}}}\xspace,\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}$ and
$\tilde c_{\varphi(k-1)+1},\ldots \tilde c_{\varphi(k)-1},c_{\varphi(k)}\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}$:
the latter again only tentatively.
Thus the $\varphi(k)+k$ bits of $(\tilde c_0,\ldots \tilde c_{\varphi(k)-1})$
depend precisely on the $2\varphi(k)$ bits of $(\tilde b_0,\ldots \tilde b_{\varphi(k)-1})$:
the transformation $F_\varphi$ on Cantor space thus
has modulus of continuity $\varphi(k)+k\mapsto2\varphi(k)$ for all $k\in\mathbb{N}$,
and $2\varphi\circ\upinv{(\operatorname{id}+\varphi)}$ in general.
\qed\end{enumerate}\end{proof}
\subsection{Abstract Examples}
\label{ss:Topology}
This subsection collects some properties, relations, and examples of
moduli of continuity and entropies of spaces.
\begin{fact}
\label{f:Topology}
\begin{enumerate}
\item[a)]
Every compact metric space $(X,d)$ can be covered by
finitely many open balls $\operatorname{B}(x,r)$; therefore its entropy $\eta$ is well-defined.
If $X$ is infinite
then $\eta\in\operatorname{Reg}$.
\item[b)]
Every continuous function $f:X\to Y$ between compact metric spaces $(X,d)$ and $(Y,e)$
is uniformly continuous and therefore has a modulus $\mu$ of continuity.
\item[c)]
Lipschitz-continuous functions have moduli of continuity
$\mu(n)=n+\mathcal{O}(1)$, and vice versa.
H\"{o}lder-continuous functions have linear moduli of continuity
$\mu(n)=\mathcal{O}(n)$, and vice versa.
\item[d)]
Proceeding from metric $d$ on $X$ to a metric $d'$ with
$d'\leq2^{-n}$ whenever $d\leq2^{-\nu(n)}$
changes the entropy $\eta$ to $\eta'\leq\eta\circ\nu$.
Additionally proceeding from $e$ on $Y$ to $e'$ satisfying $e'\leq2^{-\kappa(n)}\Rightarrow e\leq2^{-n}$
turns a modulus of continuity $\mu$ of $f:X\to Y$ into $\mu'$
with $\mu\leq\nu\circ\mu'\circ\kappa$.
\end{enumerate}
\end{fact}
\COMMENTED{%
\begin{proof}[Fact~\ref{f:Topology}]
\begin{enumerate}
\item[c)]
Suppose $e\big(f(x),f(x')\big)\leq 2^\ell\cdot d(x,x')^{1/k}$ for some positive $k\in\mathbb{N}$.
Then $d(x,x')\leq2^{-\mu(n):=-k\cdot(n+\ell)}$ implies $e\big(f(x),f(x')\big)\leq 2^{-n}$.
Conversely suppose $\mu(n)=k\cdot (n+\ell)$ is a modulus of continuity of $f$
and $L\in\mathbb{N}$ is such that $2^{L+1}\geq\max_{x,x'} e\big(f(x),f(x')\big)$.
For $x,x'\in X$, in case $d(x,x')\leq2^{-k\ell}$ consider the unique $n\in\mathbb{N}$
such that $2^{-k\cdot(n+\ell+1)}<d(x,x')\leq2^{-k\cdot(n+\ell)}$:
then $e\big(f(x),f(x')\big)\leq2^{-n}\leq 2^{\ell+1}\cdot d(x,x')^{1/k}\leq 2^{L+k\ell+1}\cdot d(x,x')^{1/k}$;
while in case $d(x,x')\geq2^{-k\ell}$ it also follows
$e\big(f(x),f(x')\big)\leq2^{L+1}\leq 2^{L+1+k\ell}\cdot d(x,x')^{1/k}$.
\item[e)]
In the $N:=t(n)$ steps the (oracle) Type-2 Machine $\mathcal{M}$ makes
before printing the $n$-th symbol $y_n$ of output
$\bar y=F(\bar x)$, it cannot have read more than
symbols $x_0,\ldots x_{N-1}$ of the input $\bar x\in\operatorname{dom}(F)$.
Therefore executing $\mathcal{M}$, instead of of input $\bar x$,
on input $\bar x'$ with $d_{\mathcal{C}}(\bar x,\bar x')\leq2^{-N+1}$
will result in the same first $N$ steps and
in particular output of the same $n$-th symbol $y'_n=y_n$:
requiring $d_{\mathcal{C}}(\bar y,\bar y')\leq 2^{-n-1}$
whenever $d_{\mathcal{C}}(\bar x,\bar x')\leq 2^{-t(n)+1}$.
\\
For the converse consider the oracle
\begin{multline*}
\big\{
big( \langle x_0,\ldots x_{\mu(0)-1}\rangle \;
\langle x_{\mu(0)},\ldots x_{\mu(1)-1}\rangle \;
\ldots
\langle x_{\mu(n-1)},\ldots x_{\mu(n)-1}\rangle \; b \big) \;:\;
n\in\mathbb{N}, \\ b,x_0,\ldots x_{\mu(n)-1}\in \{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}, \;
\exists \bar y\in\mathcal{C} : \;
F\big(x_0,\ldots x_{\mu(n)-1}\,\bar y\big)_n=b \big\}
\;\subseteq\;\mathcal{C}
\end{multline*}
encoding the $n$-th symbol $b$ shared by all $F(\bar x')$
with $d_{\mathcal{C}}\big(\bar x',x_0,\ldots x_{\mu(n)-1}\,\bar y\big)\leq2^{-\mu(n)}$.
A Type-2 Machine equipped with this oracle can,
on input $\bar x\in\operatorname{dom}(F)$ and iteratively for each $n\in\mathbb{N}$,
for each $m=0,1,\ldots$ and each $b\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}$
repeatedly query the oracle for
$\langle x_0,\ldots x_{\mu(0)-1}\rangle \: \ldots
\langle x_{\mu(n-1)},\ldots x_{\mu(n-1)+m}\rangle \: b$
in time $\mathcal{O}(m)$ each:
The first positive answer then yields
$\mu(n)=\mu(n-1)+m+1$ and permits to output
$F(\bar x)_n=b$.
\qed\end{enumerate}\end{proof}
}
\begin{myexample}[Entropy]
\label{x:Entropy}
\begin{enumerate}
\item[a)]
The real unit interval $[0;1]$
has entropy $\eta_{[0;1]}(n)=n-1$ for all integers $n\geq1$.
Cantor space has entropy $\eta_{\mathcal{C}}=\operatorname{id}$.
The Hilbert Cube $\mathcal{H}=\prod_{j\geq0}[0;1]$ with metric $d_{\mathcal{H}}(\bar x,\bar y)=\sup_j |x_j-y_j|/2^j$
has entropy $\eta_{\mathcal{H}}(n)=\Theta(n^2)$.
\item[b)]
Let compact $(X,d)$ and $(Y,e)$ have entropies $\eta_X$ and $\eta_Y$, respectively.
Then the entropy $\eta_{X\times Y}$ of compact $\big(X\times Y,\max\{d,e\}\big)$ satisfies
\[ \forall n: \quad \eta_X(n)+\eta_Y(n) \;\leq\; \eta_{X\times Y}(n+1)+1 \;\leq\; \eta_X(n+1)+\eta_Y(n+1)+1 \enspace . \]
\item[c)]
Let compact $(X_j,d_j)$ all have diameter $\leq1$ and entropy $\eta_j$, $j\in\mathbb{N}$.
Then $\big(\prod_jX_j,\sup_j d_j/2^j\big)$ is compact and has entropy $\eta$ satisfying
\[ \forall n: \quad \sum\nolimits_{j\leq n} \eta_j(n-j) \;\leq\;
\eta(n+1)+\lceil n/2\rceil \;\leq\; \sum\nolimits_{j\leq n} \eta_j(n+1-j) +\lceil n/2\rceil \enspace . \]
For $X_j\equiv[0;1]$ this recovers the Hilbert Cube $\prod_j[0;2^{-j}]$.
\item[d)]
Let $X$ be a compact space with metric $d\leq1$ and entropy $\eta$.
Then $D(x, y) := 1/\big(\log_2 2/d(x,y)\big)\leq1$
constitutes a topologically equivalent metric yet
inducing entropy $H(n) = \eta(2^n - 1)$.
\item[e)]
Fix an arbitrary non-decreasing unbounded $\varphi:\mathbb{N}\to\mathbb{N}$
and re-consider Cantor space, now equipped with
$d_\varphi:(\bar a,\bar b)\mapsto 2^{-\varphi(\min\{m:a_m\neq b_m\})}\in[0;1]$.
This constitutes a metric, topologically equivalent to $d_{\mathcal{C}}=d_{\operatorname{id}}$
but with entropy $\eta_\varphi=\loinv{\varphi}$.
\item[f)]
For $K$ a closed subset of compact $(X,d)$,
it holds $\eta_{X,K}\leq\eta_{X,X}$
but not necessarily $\eta_K\leq\eta_X$.
The image $Z:=f[X]\subseteq Y$ has entropy
$\eta_{Z}\leq \eta_X\circ\mu$,
where $\mu$ denotes a modulus of continuity of $f$.
Every connected compact metric space $X$ has entropy at least linear $\eta(n)\geq n+\Omega(1)$.
\item[g)]
Fix a compact metric space $(X,d)$ with entropy $\eta$.
Let $\mathcal{K}(X)$ denote the set of non-empty closed subsets of $X$
and equip it with the Hausdorff metric
$D(V,W)=\max\big\{\sup\{ d_V(w) :w\in W\},\sup\{d_W(v):v\in V\}\big\}$,
where $d_V:X\ni x\mapsto\inf\{ d(x,v) : v\in V\}$ denotes the distance function.
Then $\big(\mathcal{K}(X),D\big)$ constitutes a compact metric space \cite[Exercise~8.1.10]{Wei00}.
It has entropy $H\leq2^{\eta}$ with $2^{\eta(n)-1}<H(n+1)$.
\item[h)]
Fix a connected compact metric space $(X,d)$
with entropy $\eta$,
and consider the convex metric space $X':=\mathcal{C}(X,[0;1])$ of continuous real functions
equipped with the supremum norm $|f|=\sup_{x\in X} |f(x)|$.
Its subset $X'_1:=\operatorname{Lip}_1(X,[0;1])$ of non-expansive functions $f:X\to[0;1]$
is compact by Arzel\'a-Ascoli; it has relative entropy
$\eta'_1(n):=\eta_{X',X'_1}(n)=\Theta\big(2^{\eta(n\pm\mathcal{O}(1))}\big)$;
more precisely: $2^{\eta(n-1)-1} \;<\;
\eta'_1(n) \;\leq\;\mathcal{O}\big(2^{\eta(n+2)}\big)$.
\end{enumerate}
\end{myexample}
Item~(d) yields spaces with asymptotically large entropy;
and Item~(e) does similarly for small entropy ---
of a totally disconnected space in view of Example~\ref{x:Entropy}f).
The connection between modulus of continuity and entropy
(Item~f) had been observed in \cite[Lemma~3.1.13]{SteinbergDisse}.
We emphasize that Item~h) refers to $X'_1$ as
subset of $X'$, not as metric space of its own:
see also Question~\ref{q:Future}e) below.
According to Item~(d) of the following Lemma,
analyses of function spaces $\mathcal{C}_\mu(X,[0;1])$
may indeed w.l.o.g. suppose $\mu=\operatorname{id}$, i.e.,
the consider the non-expansive case.
\begin{lemma}
\label{l:Modulus}
For non-decreasing unbounded $\mu:\mathbb{N}\to\mathbb{N}$ let
\[
\omega_\mu \;:\; [0;\infty) \;\ni\;
t \;\mapsto\; \inf\Big\{ \sum\limits_{j=1}^J 2^{-n_j} \::\: J,n_1,\ldots n_j\in\mathbb{N}, \: t\leq \sum\limits_{j=1}^J 2^{-\mu(n_j)} \Big\}
\;\in\; [0;\infty)
\]
\begin{enumerate}
\item[a)]
It holds $\omega_\mu(0)=0$ and $\omega_\mu(t)>0$ for $t>0$.
$\omega_\mu$ is subadditive: $\omega(s+t)\leq\omega(s)+\omega(t)$.
$\omega_\mu$ has modulus of continuity $\mu$.
\item[b)]
If $\mu$ is strictly increasing,
then $\mu(n)=\min\big\{ m\in\mathbb{N} \::\: \omega_\mu(2^{-m})\leq2^{-n}\big\}$.
\item[c)]
For a compact \emph{convex} metric space $(X,d)$ and
any $x,y\in X$, there exists an isometry $\imath:[0;d(x,y)]\to X$
with $\imath(0)=x$ and $\imath\big(d(x,y)\big)=y$.
\item[d)]
If $(X,d)$ is compact convex and $\mu$ a modulus of continuity of $f:X\to\mathbb{R}$,
then $|f(x)-f(x')|\leq \omega_\mu\big(d(x,x')\big)$ for all $x,x'\in X$. \\
In particular $\mathcal{C}_\mu\big((X,d),\mathbb{R}\big)=\operatorname{Lip}_1\big((X,\omega_\mu\circ d),\mathbb{R}\big)$
holds for every strictly increasing $\mu$.
\end{enumerate}
\end{lemma}
Recall that a (not necessarily linear) metric space $X$ is called
\emph{convex} if, to any distinct $x,y\in X$,
there exists a $z\in X\setminus\{x,y\}$ with $d(x,y)=d(x,z)+d(z,y)$.
Examples include compact convex subsets of Euclidean space with its inherited metric,
but also connected compact subsets when equipped with the intrinsic (=shortest-path) distance,
while Cantor space is not convex.
\begin{myexample}[Modulus of Continuity]
\label{x:Topology}
\begin{enumerate}
\item[a)]
The function $(0;1]\ni t\mapsto 1/\ln(e/t)\in(0;1]$
extends uniquely continuously to $0$
and has an exponential, but no polynomial,
modulus of continuity.
\item[b)]
Picking up on Example~\ref{x:Entropy}c),
let $\xi_j:\subseteq\mathcal{C}\twoheadrightarrow X_j$ have modulus of continuity $\kappa_j$
and fix some injective `pairing' function
\[ \mathbb{N}\times\mathbb{N} \;\ni\; (n,m) \;\mapsto\;\langle n,m\rangle
\;\in\; \mathbb{N} \]
such as Cantor's $\langle n,m\rangle\;=\;(n+m)\cdot(n+m+1)/2+n$. Then
\[ \prod\nolimits_j x_j:\subseteq\mathcal{C}\;\ni\;\bar b\;\mapsto\;
\Big(\xi_j\big(b_{\langle j,0\rangle},
\ldots b_{\langle j,n\rangle},\ldots\big)_j\Big)
\;\in\;\prod\nolimits_j X_j \]
has modulus of continuity $n\mapsto\sup_{j<n} \langle j,\kappa_j(n-j)\rangle$.
\item[c)]
If $\xi$ is a representation of $X$ with modulus of continuity $\mu$,
then the following $2^\xi$ is a representation of $\mathcal{K}(X)$ with modulus
of continuity $m\mapsto 2^{\mu(m)+1}-1$:
$(b_0,b_1,\ldots b_m,\ldots)\in\mathcal{C}$ is a $2^\xi$-name of $A\in\mathcal{K}(X)$ iff,
for every $n\in\mathbb{N}$ and every $\vec v\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{<\mu(n)}$ it holds:
\begin{eqnarray} \label{e:Hausdorff}
(\vec v\circ\mathcal{C})\cap\operatorname{dom}(\xi)\neq\emptyset
\;\wedge\; b_{\mathrm{bin}(\vec v)}=\textup{\texttt{1}}\xspace &\Rightarrow&
\xi[\vec v\circ\mathcal{C}]\cap\overline{\operatorname{B}}(A,2^{-n})\neq\emptyset \\ \nonumber
(\vec v\circ\mathcal{C})\cap\operatorname{dom}(\xi)\neq\emptyset
\;\wedge\; b_{\mathrm{bin}(\vec v)}=\textup{\texttt{0}}\xspace &\Rightarrow&
\xi[\vec v\circ\mathcal{C}]\cap\overline{\operatorname{B}}(A,2^{-n-1})=\emptyset
\end{eqnarray}
where $\vec v\circ\mathcal{C} \;:=\; \{\vec v\bar w: \bar w\in\mathcal{C}\} \;\subseteq\; \mathcal{C}$
and $\mathrm{bin}(v_0,\ldots v_{n-1})=v_0+2v_1+4v_2+\cdots+2^{n-1}v_n+2^n-1$
and $\overline{\operatorname{B}}(A,r):=\bigcup_{a\in A}\overline{\operatorname{B}}(a,r)$.
\end{enumerate}
\end{myexample}
Note that $\mu\leq\mathcal{O}(\eta)$ implies $2^{\mu}\leq\mathcal{P}(2^\eta)$:
reflected in Theorem~\ref{t:Admissible}d) and Theorem~\ref{t:Functions} below.
According to Example~\ref{x:Entropy}c),
Example~\ref{x:Topology}b) does \emph{not} preserve linear admissibility
already in case of spaces $X_j$ with quadratic entropy:
a more sophisticated construction is needed in Theorem~\ref{t:Cartesian}.
\subsection{Proofs}
\begin{proof}[Example~\ref{x:Topology}]
\begin{enumerate}
\item[c)]
$2^\xi$ is a representation, as $A\in\mathcal{K}(X)$ can be recovered from any name $\bar b$:
On the one hand, for every $\xi$-name $\bar v$ of $a\in A$
and every $n\in\mathbb{N}$, $\displaystyle b_{\mathrm{bin}(v_0,\ldots v_{\mu(n)-1})}=\textup{\texttt{1}}\xspace$;
on the other hand, for every $\xi$-name $\bar v$ of $a\not\in A$,
$\operatorname{B}(a,2^{-n})\cap A=\emptyset$ implies
$\displaystyle b_{\mathrm{bin}(v_0,\ldots v_{\mu(n)-1})}=\textup{\texttt{0}}\xspace$.
Since $\mathrm{bin}(v_0,\ldots v_{\mu(n)-1})<2^{\mu(n)+1}-1$,
this also establishes $2^{\mu+1}-1$ as modulus of continuity of $2^\xi$.
\qed\end{enumerate}\end{proof}
\begin{proof}[Lemma~\ref{l:Modulus}]
\begin{enumerate}
\item[a)] Regarding subadditivity, $\omega_\mu(s+t)=$
\begin{eqnarray*}
&=&
\inf \Big\{ \sum\limits_{j=1}^J 2^{-n_j} + \sum\limits_{k=1}^K 2^{-m_j}
\::\: J,K,n_1,\ldots n_j,m_1,\ldots m_K\in\mathbb{N}, \\
&& \phantom{\inf } \underbrace{\phantom{\Big\{}
\qquad\qquad\qquad s+t\leq \sum\limits_{j=1}^J 2^{-\mu(n_j)} +
\sum\limits_{k=1}^K 2^{-\mu(m_k)} \Big\}}_{\rotatebox[origin=c]{270}{$\supseteq$}} \\
&\leq&
\inf\overbrace{\Big\{ \sum\limits_{j=1}^J 2^{-n_j} +\sum\limits_{k=1}^K 2^{-m_j}
\::\: J,K,n_1,\ldots n_j,m_1,\ldots m_K\in\mathbb{N}, } \\
&& \phantom{\inf \Big\{ } \qquad s\leq \sum\nolimits_{j=1}^J 2^{-\mu(n_j)} \;\wedge\;
t\leq \sum\nolimits_{k=1}^K 2^{-\mu(m_k)} \Big\} \\
&=& \omega_\mu(s)\;+\; \omega_\mu(t)
\end{eqnarray*}
By definition ($J:=1$) it holds $0\leq\omega_\mu(t)\leq 2^{-n}$ for $t\leq2^{-n}$,
and in particular $\omega_\mu(0)=0$. By subadditivity
and whenever $0\leq\delta\leq2^{-\mu(n)}$, we have both
$\omega_\mu(t)\leq\omega_\mu(t+\delta)\leq\omega_\mu(t)+\omega_\mu(\delta)\leq\omega_\mu(t)+2^{-n}$ and
$\omega_\mu(t)-2^{-n}\leq\omega_\mu(t)-\omega_\mu(\delta)=\omega_\mu(t-\delta+\delta)-\omega_\mu(\delta)
\leq\omega_\mu(t-\delta)\leq\omega_\mu(t)$.
\item[b)]
For every $t\leq2^{-\mu(n)}$ it holds $\omega_\mu(t)\leq2^{-n}$ by definition, and hence
$\tilde\mu(n):=\min\big\{ m\in\mathbb{N} \::\: \omega_\mu(2^{-m})\leq2^{-n}\big\}\leq \mu(n)$.
Conversely for $m\leq n_1,\ldots,n_j\in\mathbb{N}$,
\[ \sum\nolimits_j 2^{-\mu(n_j)} \;\leq\;
\sum\nolimits_j 2^{-\mu(m)-n_j+m}
\;=\; 2^{-\mu(m)}\cdot 2^{m}\cdot\sum\nolimits_j 2^{-n_j} \]
from strict monotonicity $\mu(n_j)=\mu(m+n_j-m)\geq\mu(m)+(n_j-m)$ by induction.
So $\sum_j 2^{-\mu(n_j)}\geq 2^{-\mu(m)}$ implies $\sum_j 2^{-n_j}\geq2^{-m}$
and $\omega_\mu\big(2^{-\mu(m)}\big)\geq2^{-m}$
and $\tilde\mu(n)\geq\mu(n)$.
\item[c)]
Using transfinite induction and completeness,
\cite[Exercise~5.1.17]{Kaplansky}
constructs a $z\in X$ with $d(x,z)=d(z,y)=d(x,y)/2$.
Now iterating with both $(x,z)$ and $(z,y)$ in place of $(x,y)$
yields a sequence of refinements $z_n\in X$, $n=0,\ldots,N=2^k$
$z_0=x$ and $z_N=y$ and $d(z_n,z_{n+1})=d(x,y)/N$.
Again by completeness,
$\imath_k(t):=z_{\min\{n: n/2^k\geq t\}}$ thus converges
uniformly to the claimed isometry.
\item[d)]
For $t:=d(x,x')$, and to any $J\in\mathbb{N}$ and $n_1,\ldots,n_j\in\mathbb{N}$
with $t\leq \sum\nolimits_{j=1}^J 2^{-\mu(n_j)}$,
c) yields $x=:x_0,x_1,\ldots,x_J=x'\in X$ with $d\big(x_{j-1},x_j\big)\leq2^{-\mu(n_j)}$.
Hence
\[
\big|f(x)-f(x')\big|
\;\leq\; \sum\nolimits_{j=1}^J \big|f(x_{j-1})-f(x_j)\big|
\;\leq\; \sum\nolimits_{j=1}^J 2^{-n_j} \;\leq\; \omega_\mu(t)
\]
\qed\end{enumerate}\end{proof}
\begin{fact}
\label{f:Extension}
Fix a compact metric space $(X,d)$,
non-empty $Z\subseteq X$ and $L>0$.
For $L$-Lipschitz $f:Z\to\mathbb{R}$, the functions
\begin{equation}
\label{e:Extension}
\extl{f}:x\;\mapsto\; \sup\nolimits_{z\in Z} f(z)-L\cdot d(z,x),
\quad \exth{f}:x\;\mapsto\;\inf\nolimits_{z\in Z} f(z)+L\cdot d(z,x)
\end{equation}
extend $f$ to $X$ while preserving $L$-Lipschitz continuity.
Moreover every $L$-Lipschitz extension $\tilde f:X\to\mathbb{R}$ of $f$ to $X$
satisfies $\extl{f}\leq\tilde f\leq\exth{f}$,
where $\exth{f}-\extl{f}\leq 2L |d_Z|=2L\sup_x d_Z(x)$.
The extension operator
$\operatorname{Lip}_L(Z,\mathbb{R})\ni f\mapsto \extm{f}:=(\extl{f}+\exth{f})/2\in\operatorname{Lip}_L(X,\mathbb{R})$
is a well-defined isometry of compact metric spaces w.r.t. the supremum norm.
\end{fact}
$\big(\extl{f},\exth{f}\big)$
is known as \emph{McShane-Whitney pair} \cite{Petrakis}.
For the purpose of self-containment, we include a proof:
\begin{proof}[Fact~\ref{f:Extension}]
W.l.o.g. $L=1$.
For $x\in Z$, choosing $z:=x$ shows $\extl{f}(x)\geq f(x)$
while $f(z)-d(z,x)\geq f(z)-|f(z)-f(x)|\geq f(x)$
implies $\extl{f}(x)\leq f(x)$.
Moreover, for every $z\in Z$, we have
\[ -\sup\big\{f(z')-d(z',x'):z'\in Z\big\}
\;\leq\; -f(z)+d(z,x') \;\leq\; -f(z)+d(z,x)+d(x,x') \]
and hence
\[ \extl{f}(x)-\extl{f}(x') \;\leq\;
\sup\big\{ f(z)+d(z,x) - f(z)+d(z,x)+d(x,x'): z\in Z\big\}
\;=\; d(x,x') . \]
The estimates for $\exth{f}$ proceed similarly.
For $x\in X$ and with $z,z',w,w'$ ranging over $Z$,
$\displaystyle
\big(\extl{f}(x)+\exth{f}(x)\big)-\big(\extl{g}(x)+\exth{g}(x)\big) \;=$
\begin{eqnarray*}
&=&\sup\limits_z f(z)\!-\!d(z,x)
\:-\: \sup\limits_w g(w)\!-\!d(w,x)
\:-\: \inf\limits_{w'} g(w')\!+\!d(w',x)
\:+\: \inf\limits_{z'} f(z')\!+\!d(z',x)
\\
&=& \sup_z f(z)\!-\!d(z,x)
\:+ \inf_w -g(w)+d(w,x)
\:+ \sup_{w'} -g(w')-d(w',x)
\:+ \inf_{z'} f(z')+d(z',x) \\
&\leq& \sup\limits_z f(z)-d(z,x)
\:-\: g(z)+d(z,x)
\:+\: \sup\limits_{w'} -g(w')-d(w',x)
\:+\: f(w')+d(w',x) \\
&=& 2\cdot \sup_z f(z)-g(z)\;\leq\; 2\cdot\sup\nolimits_z |g(z)-f(z)| \enspace .
\qed \end{eqnarray*}
\end{proof}
\begin{proof}[Example~\ref{x:Entropy}]
Let $\mathcal{H}_X(n)$ denote the least number of closed balls of radius $2^{-n}$ covering $X$,
so that $\eta_X(n)=\lceil\log_2\mathcal{H}(n)\rceil$.
Let $\mathcal{C}_X(n)$ denote the largest number of points in $X$
of pairwise distance $>2^{-n}$, also known as \emph{capacity}.
(Since $X$ is not an integer function, there is not danger of
confusion this notation with that of a space of continuous functions\ldots)
Then $\mathcal{C}_X(n)\leq\mathcal{H}_X(n+1)$:
To cover $X$ requires covering the $\mathcal{C}_X(n)$ points as above;
but any closed ball of radius $2^{-(n+1)}$ can contain at most one
of those points having distance $>2^{-n}$.
On the other hand $\mathcal{H}_X(n)\leq\mathcal{C}_X(n)$,
since balls of radius $2^{-n}$ whose centers form a maximal set $X_n$ of pairwise distance $>2^{-n}$
cover $X$: if they missed a point, that had distance $>2^{-n}$ to all centers in $X_n$
and thus could be added to $X_n$: contradicting its maximality.
\begin{enumerate}
\item[a)]
Cover $[0;1]$ by $2^{n-1}$ closed intervals $I_{n,j}:=\big[j\cdot2^{-(n-1)};(j+1)\cdot2^{-(n-1)}\big]$,
$j=0,\ldots 2^{n-1}-1$, of radius $2^{-n}$ around centers $(2j+1)2^{-n}$: optimally. \\
Cover $\mathcal{C}$ by $2^n$ closed balls $\vec x\circ\mathcal{C}$
of radius $2^{-n}$ around centers $\vec x\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^n$: optimally.
Cover $\mathcal{H}$ by $2^{n-1}\cdot2^{n-2}\cdots 2\cdot 1=2^{n(n+1)/2}$ closed balls
\[ I_{n,j_0}\times I_{n-1,j_1} \times \cdots \times I_{1,j_{n-1}} \times \prod\nolimits_{j\geq n} [0;1] \]
of radius $2^{-n}$ with indices ranging as follows:
$0\leq j_0<2^{n-1}, \quad 0\leq j_1<2^{n-2}, \quad \cdots \quad j_{n-1}=0$.
\item[b)]
Obviously $\mathcal{H}_{X\times Y}\leq \mathcal{H}_X\cdot \mathcal{H}_Y$
and $\mathcal{C}_{X\times Y}\geq\mathcal{C}_X\cdot\mathcal{C}_Y$:
$\mathcal{H}_X(n)\cdot\mathcal{H}_Y(n) \leq$
\[ \leq\; \mathcal{C}_X(n)\cdot\mathcal{C}_Y(n) \;\leq\; \mathcal{C}_{X\times Y}(n)
\;\leq\; \mathcal{H}_{X\times Y}(n+1) \;\leq\; \mathcal{H}_X(n+1)\cdot\mathcal{H}_Y(n+1) \]
Also record $\lceil s\rceil+\lceil t\rceil \geq \lceil s+t\rceil\geq \lceil s\rceil+\lceil t\rceil-1$
for all $s,t>0$.
\item[c)]
Abbreviating $\mathcal{H}_j:=\mathcal{H}_{X_j}$ and $\mathcal{C}_j:=\mathcal{C}_{X_j}$,
we have $\mathcal{H}(n)\leq \prod_{j<n} \mathcal{H}_j(n-j)$
and $\mathcal{C}(n)\geq \prod_{j<n} \mathcal{C}_j(n-j)$:
note that $\mathcal{H}_j(0)=1=\mathcal{C}_j(0)$ as $X_j$ has diameter $\leq1$.
Finally $\sum_{j=0}^{n-1} \lceil t_j\rceil \geq \lceil\sum_{j=0}^{n-1} t_j\rceil \geq \sum_{j=0}^{n-1} \lceil t_j\rceil -\lfloor n/2\rfloor$.
\item[f)]
For a counterexample to $\eta_K\leq\eta_X$
consider a circle/hyper-/sphere with and without center.
\\
Regarding the lower bound for connected compact metric spaces,
consider $N:=2^{\eta(n)}$
and $x_1,\ldots x_N\in X$ such that balls with centers $x_j$ and radius $2^{-n}$ cover $X$:
$X\subseteq\bigcup_{n=1}^N \overline{\operatorname{B}}(x_n,2^{-n})$.
Consider the finite undirected graph $G_n=(V_n,E_n)$
with vertices $V_n=\{1,\ldots N\}$ and edges
$\{i,j\}\in E_n\Leftrightarrow\operatorname{B}(x_i,2^{-n+1})\cap\operatorname{B}(x_j,2^{-n+1})\neq\emptyset$
whenever the two open balls with centers $x_i,x_j$ and radius twice $2^{-n}$ intersect.
This graph is connected: If $I,J\subseteq V_n$ were distinct connected components,
then $\bigcup_{n\in I}\operatorname{B}(x_n,2^{-n+1})$ and $\bigcup_{n\not\in I}\operatorname{B}(x_n,2^{-n+1})$
were two disjoint open subsets covering $X$.
Therefore any two vertices $i,j\in V_n$ are connected via $\leq N-1$ edges;
and for every edge $\{a,b\}$, it holds $d(x_a,x_b)<2^{-n+2}$ by definition of $E_n$:
Hence $x_i$ and $x_j$ have metric distance $d(x_i,x_j)$ at most $(N-1)\cdot 2^{-n+2}$;
and any two $x,y\in X$ have $d(x,y)\leq N\cdot2^{-n+2}$:
requiring $2^{\eta(n)}=N\geq d(x,y)\cdot 2^{n-2}$.
\item[g)]
Obviously $\mathcal{H}_{\mathcal{K}(X)}\leq2^{\mathcal{H}_X}$
and $\mathcal{C}_{\mathcal{K}(X)}\geq2^{\mathcal{C}_X}$.
\item[h)]
Fix $n\in\mathbb{N}$ and consider a maximal set
$X_n\subseteq X$ of $N:=\mathcal{C}_X(n)$ points of pairwise distance $>2^{-n}$.
There are $2^{\mathcal{C}_X(n)}$ different $f:X_n\to\{0,2^{-n}\}$;
each is 1-Lipschitz, and extends to $\extm{f}:X\to[0;1]$;
and, according to Fact~\ref{f:Extension},
different such $f$ give rise to $\extm{f}$ of
mutual supremum distance $\geq2^{-n}$:
Hence $\mathcal{C}_{X'_1}(n)\geq 2^{\mathcal{C}_X(n)}$,
and $X'_1\subseteq X'$
has intrinsic entropy $\eta'_1(n)\geq
\log_2 \mathcal{C}_{X'_1}(n-1)
\geq \mathcal{C}_X(n-1)\geq \mathcal{H}_X(n-1)>2^{\eta(n-1)-1}$.
\\
Conversely, for any 1-Lipschitz $f:X\to[0;1]$,
consider $f'_n:=\lfloor 2^n\cdot f\big|_{X_n}\rceil/2^n$:
still $(1+1/2)$-Lipschitz since rounding affects the value by at most $2^{-n-1}$
on arguments of distance $>2^{-n}$.
As argued before, maximality of $X_n$ implies
that the closed balls around centers $x\in X_n$
of radius $2^{-n}$ cover $X$ (hence $d_{X_n}\leq 2^{-n}$);
consequently so do the open balls with radius $2^{-n+1}$.
Similarly to the proof of (f),
consider the finite undirected and connected graph $G_n=(X_n,E_n)$ with edge
$\{x,y\}\in E_n\;:\Leftrightarrow\;\operatorname{B}(x,2^{-n+1})\cap\operatorname{B}(y,2^{-n+1})\neq\emptyset$.
Any vertex $y$ of $G_n$ adjacent to some $x$ has distance $d(x,y)<2^{-n+2}$;
and since $f'_n:X_n\to\mathbb{D}_n$ is $\tfrac{3}{2}$-Lipschitz,
this implies $\mathbb{D}_n\ni|f'_n(x)-f'_n(y)|\leq \tfrac{3}{2}\cdot2^{-n+2}$
leaving no more than 13 possible values for $f'_n(x)-f'_n(y)\in\big\{-6\cdot2^{-n},\ldots 0,\ldots +6\cdot2^{-n}\big\}$.
Connectedness of $G_n$ with $N$ vertices thus limits
the number of different $\tfrac{3}{2}$-Lipschitz $f'_n:X_n\to\mathbb{D}_n$
to $\leq(1+2^n)\cdot 13^{N-1}\leq 2^{\mathcal{O}(\mathcal{C}_X(n))}$
in view of (f). And by Fact~\ref{f:Extension}
each such $f'_n$ extends to some $\tfrac{3}{2}$-Lipschitz $\extm{f'_n}:X\to[0;1]$.
Moreover $|d_{X_n}|\leq 2^{-n}$ implies
$\big|f-\extm{f_n}\big|\leq \big|\exth{f_n}-\extl{f_n}\big|/2 \leq \tfrac{3}{2}\cdot 2^{-n}$
for the $\tfrac{3}{2}$-Lipschitz (!) extension of the restriction $f_n:=f\big|_{X_n}$.
Since $g\mapsto\extm{g}$ is an isometry,
this implies $\big|f-\extm{f'_n}\big|
\leq \big|f-\extm{f_n}\big|+\big|f_n-f'_n\big|
\leq \tfrac{3}{2}\cdot2^{-n}+2^{-n-1}=2^{-n+1}$.
The $2^{\mathcal{O}(\mathcal{C}_X(n))}$ closed balls of radius $2^{-n+1}$
around centers $\extm{f'_n}\in\operatorname{Lip}_{3/2}(X,[0;1])\subseteq X'$ thus cover $\operatorname{Lip}_1(X,[0;1])=X'_1$:
$\eta'_1(n-1)=\eta_{X',X'_1}(n-1)\leq \mathcal{O}\big(\mathcal{C}_X(n)\big) \leq \mathcal{O}\big(\mathcal{H}_X(n+1)\big) \leq \mathcal{O}\big(2^{\eta(n+1)}\big)$.
\qed\end{enumerate}\end{proof}
\section{Concise Standard Representations}
\label{s:Standard}
\cite[Definitions~3.2.2+3.2.7]{Wei00} introduce qualitative
admissibility in terms of a \emph{standard} representation
which \cite[Lemma~3.2.5]{Wei00} then shows to satisfy properties (i) and (ii) from Fact~\ref{f:Main}.
Here we first recall from \cite[Definition~15]{DBLP:conf/lics/KawamuraS016}
the construction of a \emph{concise} standard representation
of any fixed compact metric space $(X,d)$
that generalizes Example~\ref{x:Dyadic}:
For each $n$, fix a covering of $X$ by $\leq 2^{\eta(n)}$ balls of radius $2^{-n}$ according to the entropy;
assign to each ball a binary string $\vec a_n$ of length $\eta(n)$;
then every $x\in X$ can be approximated by the center of some of these balls;
finally define a name of $x$ to be such a sequence $(\vec a_n)_{_n}$ of binary strings.
Theorem~\ref{t:Polynomial} establishes that this representation is polynomially admissible,
provided the balls' radius is reduced to $2^{-n-1}$.
Subsection~\ref{ss:Linear} improves the construction to yield a linearly admissible standard representation.
\begin{definition}
\label{d:Standard}
Let $(X,d)$ denote a compact metric space
with entropy $\eta$.
For each $n\in\mathbb{N}$ fix some partial mapping
$\xi_n:\subseteq\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{\eta(n+1)}\to X$
such that $X=\bigcup_{\vec a\in\operatorname{dom}(\xi_n)} \overline{\operatorname{B}}\big(\xi_n(\vec a),2^{-n\pmb{-1}}\big)$,
where $\overline{\operatorname{B}}(x,r)=\{x'\in X:d(x,x')\leq r\}$
denotes the closed ball around $x$ of radius $r$.
The \emph{standard} representation of $X$
(with respect to the family $\xi_n$ of partial dense enumerations)
is the mapping
\begin{gather}
\label{e:Dyadic}
\xi \; :\subseteq \mathcal{C} \;\ni\;
\big( (\vec a_0) \: (\vec a_1) \:
\ldots
(\vec a_n) \:
\ldots \big)
\;\mapsto\;
\lim\nolimits_n \xi_n(\vec a_n) \;\in\; X ,
\\[0.5ex] \nonumber
\operatorname{dom}(\xi) :=\: \big\{ \big( \:\ldots\:
(\vec a_n) \: \ldots\big) \::\: \vec a_n\in\operatorname{dom}(\xi_n), \;
d\big(\xi_n(\vec a_n),\xi_m(\vec a_m)\big)\:\leq\: 2^{-n}\!+\!2^{-m}\big\}
\end{gather}
\end{definition}
Fact~\ref{f:Topology}a) asserts such $\xi_n$ to exist.
The real Example~\ref{x:Dyadic} is\footnote{%
Well, almost: $\mathrm{bin}(a_n)$ has length between $1$ and $2n+1$ while
here we make all strings in $\operatorname{dom}(\xi_n)$ have the same length $=\eta(n+1)$:
Using strings of varying length $<\eta(n+1)$ would additionally require encoding delimiters.}
a special case of this definition
with $\eta_{[0;1]}(n+1)=n$ according to Example~\ref{x:Entropy}a) and
\[ \delta'_n \;:\; \{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^n \;\ni\; \vec a\;\mapsto\;
\big(\tfrac{1}{2}+a_0+2a_1+4a_2+\cdots+2^{n-1}a_{n-1}\big)/2^n
\enspace . \]
The covering balls' radius being $2^{-n\pmb{-1}}$ instead of $2^{-n}$ is exploited in the following theorem:
\begin{theorem}
\label{t:Polynomial}
\begin{enumerate}
\item[i)]
The standard representation $\xi$ of $(X,d)$ w.r.t. $(\xi_n)$ according to Definition~\ref{d:Standard}
has modulus of continuity $\kappa(n):=\sum_{m=0}^{n} \eta(m+1)$.
\item[ii)]
To every partial function $\zeta:\subseteq\mathcal{C}\to X$
with modulus of continuity $\nu$
there exists a mapping $F:\operatorname{dom}(\zeta)\to\operatorname{dom}(\xi)$
with modulus of continuity $\mu=\nu\circ\big(1+\loinv{\kappa}\big):\kappa(n)\mapsto\nu(n+1)$
such that $\zeta=\xi\circ F$ holds. \\
In particular $\xi$ is polynomially admissible,
provided that the entropy grows at least with some positive power
$\eta(n)\geq\Omega(n^{\epsilon})$, $\epsilon>0$.
\item[iii)]
To every $m\in\mathbb{N}$ and every $x,x'\in X$ with $d(x,x')\leq2^{-m-1}$,
there exist $\xi$-names $\bar y_x$ and $\bar y'_{x'}$
of $x=\xi(\bar y_x)$ and $x'=\xi(\bar y'_{x'})$
with $d_{\mathcal{C}}(\bar y_x,\bar y'_{x'})\leq2^{-\kappa(m)}$.
\item[iv)]
If $(Y,e)$ is a compact metric space and
$f:X\to Y$ such that $f\circ\xi:\operatorname{dom}(\xi)\subseteq\mathcal{C}\to Y$
has modulus of continuity $\kappa\circ\nu$,
then $f$ has modulus of continuity $\nu+1$.
\end{enumerate}
\end{theorem}
Again (ii) strengthens Definition~\ref{d:Admissible}c)
in applying to not necessarily surjective $\zeta$.
In view of Lemma~\ref{l:seminv}c), (iv) can be rephrased as follows:
$f\circ\xi$ with modulus of continuity $\nu$
implies $f$ to have modulus of continuity $1+\loinv{\kappa}\circ\nu$.
However (ii) is \emph{not} saying that
$\zeta$ with modulus of continuity $\nu\circ\mu$
yields $F$ with modulus of continuity $n\mapsto\nu(n+1)$.
\begin{proof}[Theorem~\ref{t:Polynomial}]
\begin{enumerate}
\item[i)]
First observe that $\xi$ is well-defined:
as compact metric space, $X$ is complete
and the dyadic sequence $\xi_n(\vec a_n)\in X$ therefore converges.
Moreover $\xi$ is surjective:
To every $x\in X$ and $n\in\mathbb{N}$ there exists by hypothesis some
$\vec a_n\in\operatorname{dom}(\xi_n)$ with $x\in\overline{\operatorname{B}}\big(\xi_n(\vec a_n),2^{-n-1}\big)$;
hence $d\big(\xi_n(\vec a_n),\xi_m(\vec a_m)\big)\leq 2^{-n}+2^{-m}$
and $\lim_n\xi_n(\vec a_n)=x$.
Furthermore, $\vec a_n$ has binary length $\eta(n+1)$; hence
$\big(\vec a_0 \: \ldots \vec a_{n}\big)$
has length $\kappa(n)$ as above; and fixing this initial segment of a $\xi$-name $\bar x$
implies $d\big(\xi(\bar x),\xi_n(\vec a_n)\big)\leq2^{-n}$ by Equation~(\ref{e:Dyadic}).
\item[ii)]
To every $n$ and each (of the finitely many) $\bar x|_{<\nu(n+1)}$ with $\bar x\in\operatorname{dom}(\zeta)$,
fix some $\vec a_n=\vec a_n\big(\bar x|_{<\nu(n+1)}\big)\in\operatorname{dom}(\xi_n)$ such that
\[ \xi_n(\vec a_n)\;\in\;\zeta\big[\bar x|_{<\nu(n+1)}\circ\mathcal{C}\big]\;\subseteq\;\bigcup\nolimits_{\vec a\in\operatorname{dom}(\xi_n)}
\overline{\operatorname{B}}\big(\xi_n(\vec a),2^{-n-1}\big) \enspace . \]
Then iteratively for $n=0,1,\ldots$ let
similarly to the proof of Example~\ref{x:Dyadic},
\[ F_{n}\big(\bar x|_{<\nu(n+1)}\big) \;:=\; F_{n-1}\big(\bar x|_{<\nu(n)}\big) \:\circ\: (\vec a_n)
\;\in\; \{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{\kappa(n-1)+\eta(n+1)=\kappa(n)} \enspace . \]
This makes $F(\bar x):=\lim_n F_n\big(\bar x|_{<\nu(n+1)}\big)\in\mathcal{C}$
well-defined with modulus of continuity $\kappa:\kappa(n)\mapsto\nu(n+1)$.
Moreover it holds $F(\bar x)\in\operatorname{dom}(\xi)$ and $\xi\big(F(\bar x)\big)=\zeta(\bar x)$
since $\xi_n(\vec a_n),\xi(\bar x)\in\zeta\big[\bar x|_{<\nu(n+1)}\circ\mathcal{C}\big]\subseteq\overline{\operatorname{B}}\big(\xi_n(\vec a_n),2^{-n}\big)$
because $\nu$ is a modulus of continuity of $\zeta$.
\\
Finally observe $\eta(n+1)\leq\kappa(n)\leq(n+1)\cdot\eta(n+1)\in\operatorname{poly}(\eta)$;
hence (i) and (ii) of Definition~\ref{d:Admissible}c) hold.
\item[iii)]
To $x\in X$ consider the $\xi$-name $\bar y_x:=\big( \ldots \: (\vec a_m) \: \ldots \big)$ of $x$
with $d\big(\xi_m(\vec a_m),x\big)\leq2^{-m-1}$.
For $m\in\mathbb{N}$ its initial segment $\big( (\vec a_0)\: \ldots \: (\vec a_m)\big)$
has binary length $\kappa(m)$;
and, for every $x'\in X$ with $d(x,x')\leq2^{-m-1}$,
can be extended to a $\xi$-name $\bar y'_{x'}$.
\item[iv)]
Applying (iii) to $m:=\nu(n)$, the hypothesis implies
$e\big(f(x),f(x')\big)=e\big(f\circ\xi(\bar y_x),f\circ\xi(\bar y'_{x'})\big)\leq 2^{-n}$.
\qed\end{enumerate}\end{proof}
\subsection{Improvement to Linear Admissibility}
\label{ss:Linear}
The generic representation $\xi$
of a compact metric space $(X,d)$ according to Definition~\ref{d:Standard}
being `only' polynomially admissible,
this subsection improves the construction to achieve linear admissibility.
Note that $\kappa(n)=\sum_{m=0}^{n} \eta(m+1)$ according to Theorem~\ref{t:Polynomial}a)
already is in $\mathcal{O}\big(\eta(m+1)\big)$ whenever $\eta(m)\geq 2^{\Omega(m)}$ grows at least exponentially;
hence we focus on spaces with sub-exponential entropy.
To this end fix some unbounded non-decreasing $\varphi:\mathbb{N}\to\mathbb{N}$
and define a representation $\xi^\varphi$ of $X$
(with respect to the family $\xi_n$ of partial dense enumerations)
based on the \emph{sub}sequence $\xi_{\varphi(n)}$ of $\xi_n$:
\begin{gather}
\label{e:Dyadic2}
\xi^\varphi \; :\subseteq \mathcal{C} \;\ni\;
\big( \vec a_0 \:
\ldots
\vec a_n \:
\ldots \big)
\;\mapsto\;
\lim\nolimits_n \xi_{\varphi(n)}\big(\vec a_n\big) \;\in\; X ,
\qquad \operatorname{dom}(\xi^\varphi):= \\[0.5ex]
\big\{ \big( \vec a_0\: \ldots\:
\vec a_n\:\ldots\big) \::\: \vec a_n\in\operatorname{dom}\big(\xi_{\varphi(n)}\big), \;
d\big(\xi_{\varphi(n)}(\vec a_n),\xi_{\varphi(m)}(\vec a_m)\big)\leq 2^{-n}\!+\!2^{-m}\big\}
\nonumber \end{gather}
Intuitively, proceeding to a subsequence $\xi_{\varphi(n)}$
amounts to `skipping' intermediate precisions/error bounds
and `jumping' directly from $2^{-\varphi(n-1)}$ to $2^{-\varphi(n)}$.
It formalizes a strategy implemented for instance by the \texttt{iRRAM C++}
library for Exact Real Computation \cite{Mue01} which
starts with $\varphi(0)=50$ bits \texttt{double} precision and
in phases $n=\#1,\#2,\ldots$
increases to $\varphi(n)=\lfloor \tfrac{6}{5}\cdot \varphi(n-1)\rceil+20$.
The proof of Theorem~\ref{t:Polynomial} carries over literally to see:
\begin{enumerate}
\item[i)]
Representation $\xi^\varphi$
has modulus of continuity
\[ \kappa^\varphi(n)\;:=\;\sum_{m=0}^{\loinv{\varphi}(n)} \eta\big(\varphi(m)+1\big) \]
\item[ii)]
To every partial function $\zeta:\subseteq\mathcal{C}\to X$
with modulus of continuity $\nu$
there exists a mapping $F:\operatorname{dom}(\zeta)\to\operatorname{dom}(\xi^\varphi)$
with modulus of continuity $\nu\circ\big(1+\loinv{\kappa^\varphi}\big)$
such that $\zeta=\xi^\varphi\circ F$ holds.
\item[iii)]
To every $m\in\mathbb{N}$ and every $x,x'\in X$ with $d(x,x')\leq2^{-m-1}$,
there exist $\xi^\varphi$-names $\bar y_x$ and $\bar y'_{x'}$
of $x=\xi^\varphi(\bar y_x)$ and $x'=\xi^\varphi(\bar y'_{x'})$
with $d_{\mathcal{C}}(\bar y_x,\bar y'_{x'})\leq2^{-\kappa^\varphi(m)}$.
\item[iv)]
If $(Y,e)$ is a compact metric space and
$f:X\to Y$ such that $f\circ\xi^\varphi:\operatorname{dom}(\xi)\subseteq\mathcal{C}\to Y$
has modulus of continuity $\kappa^\varphi\circ\nu$,
then $f$ has modulus of continuity $\nu+1$.
\end{enumerate}
\begin{theorem}
\label{t:Linear}
Let $(X,d)$ denote a compact metric space of entropy $\eta$,
equipped with partial mappings $\xi_n:\subseteq\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{\eta(n+1)}\to X$
such that $X=\!\!\!\!\!\bigcup\limits_{\vec a\in\operatorname{dom}(\xi_n)} \!\!\!\!\!\overline{\operatorname{B}}\big(\xi_n(\vec a),2^{-n-1}\big)$.
There exists an unbounded non-decreasing $\varphi:\mathbb{N}\to\mathbb{N}$
such that the representation $\xi^\varphi$ from Equation~(\ref{e:Dyadic2})
has modulus of continuity $\kappa^\varphi(n)\leq\tfrac{27}{4}\cdot\eta(n+1)$
and $\kappa^\varphi(n)\geq \eta(n+1)$.
In particular $\xi^\varphi$ is linearly admissible.
\end{theorem}
The proof of Theorem~\ref{t:Linear} follows immediately from Item~d) of the following lemma,
applied to $c:=3/2$ with $\eta(n+1)$ in place of $\eta(n)$.
\begin{lemma}
\label{l:Donghyun}
Let $\eta:\mathbb{N}\to\mathbb{N}$ be unbounded and non-decreasing and fix $c>1$.
Then there exists a strictly increasing mapping $\varphi:\mathbb{N}\to\mathbb{N}$ such that it holds
\begin{enumerate}
\item[a)]
$\displaystyle\forall m\in\mathbb{N}: \quad \eta\big(\varphi(m+1)\big)\;\leq\; c^2\cdot \eta\big(\varphi(m)+1\big)$.
\item[b)]
$\displaystyle\forall m\in\mathbb{N}: \quad c\cdot\eta\big(\varphi(m)\big)\;\leq\; \eta\big(\varphi(m+1)\big)$.
\item[c)]
$\displaystyle\forall m\in\mathbb{N}: \quad c\cdot\eta\big(\varphi(m)+1\big)\;\leq\; \eta\big(\varphi(m+1)+1\big)$.
\item[d)]
$\displaystyle \sum\nolimits_{m=0}^{\loinv{\varphi}(n)} \eta\big(\varphi(m)\big) \;\leq\; \tfrac{c^3}{c-1}\cdot\eta(n)$.
\end{enumerate}
\end{lemma}
Think of an infinite roll of toilet papers with numbers $\eta(0)$, $\eta(1)$, \ldots printed on them.
We shall cut this roll into appropriate \emph{runs} from sheet $\#\varphi(m)$ to $\#\varphi(m+1)-1$.
Item~a) asserts that integers on sheets within the same run differ by no more than factor $c^2$.
Items~b) and c) formalizes that labels on consecutive runs grow at least exponentially.
\begin{proof}[Lemma~\ref{l:Donghyun}]
We will construct an infinite subset of $\mathbb{N}$ by picking elements one by one. Its elements in increasing order will then constitute the sequence $\varphi$.
First, in case there exists $x \in \mathbb{N}$ such that $\varphi(x) = 0$, pick the largest such $x$. And pick all those $x \in \mathbb{N}$ satisfying
$$0 < \varphi(x) \cdot c \le \varphi(x+1).$$
Possibly we have picked only finitely many elements. Let $m$ be the largest number picked so far. Pick $x > m+1$ such that
$$c \le \frac{\varphi(x)}{\varphi(m+1)} < c^2.$$
Such $x$ is guaranteed to exist so that we can choose. Now take $m = x$ and repeat this process infinitely. We can mechanically check that conditions (b) and (c) are met now. What remains is to pick more numbers so that (a) be satisfied while maintaining (b) and (c).
We will pick some more numbers for each $i \in \mathbb{N}$ that fails condition (a). Suppose that $i \in \mathbb{N}$ fails (a). Denote for convenience $a := f(i)+1$ and $b := f(i+1)$. There are two cases.
\textbf{Case i)} Suppose that $\frac{\varphi(b)}{\varphi(a)} \in [c^{2k}, c^{2k+1})$ for some $k$. Pick $x_1, x_2, \cdots x_k$ such that the followings hold for $j=1, 2, \cdots, k$:
$$\frac{\varphi(x_j)}{\varphi(a)} < c^{2j-1} \le \frac{\varphi(x_j + 1)}{\varphi(a)}.$$
\textbf{Case ii)} Suppose that $\frac{\varphi(b)}{\varphi(a)} \in [c^{2k+1}, c^{2k+2})$ for some $k$. Pick $x_1, x_2, \cdots x_k$ such that the followings hold for $j=1, 2, \cdots, k$:
$$\frac{\varphi(x_j)}{\varphi(a)} < c^{2j} \le \frac{\varphi(x_j + 1)}{\varphi(a)}.$$
It is now mechanical to check that all conditions (a), (b), and (c) are fulfilled.
\qed\end{proof}
\section{Quantitative Main Theorem and Categorical Constructions}
\label{s:Category}
We can now establish the quantitative Main Theorem
strengthening the classical qualitative one \cite[Theorem~3.2.11]{Wei00}.
\begin{theorem}[Main Theorem of Type-2 \emph{Complexity} Theory]
\label{t:Main2}
Let $(X,d)$ be compact with entropy $\eta$
and linearly admissible representation $\xi$
of modulus of continuity $\kappa$.
Let $(Y,e)$ be compact with entropy $\theta$
and linearly admissible representation $\upsilon$
of modulus of continuity $\lambda$.
\begin{enumerate}
\item[a)]
If $f:X\to Y$ has modulus of continuity $\mu$,
then it admits a $(\xi,\upsilon)$-realizer $F$
with modulus of continuity
\[ \nu \;=\; \kappa\circ(1+\mu)\circ\big(\loinv{\lambda}+\mathcal{O}(1)\big)
\;\in\; \operatorname{lin}(\eta)\circ\mu\circ\operatorname{lin}\big(\loinv{\theta}\big) \]
\item[b)]
If $f:X\to Y$ has $(\xi,\upsilon)$-realizer $F$
with modulus of continuity $\nu$,
then $f$ has modulus
\[ \mu \;=\; \loinv{\kappa}\circ\nu\circ\lambda(1+\operatorname{id})+\mathcal{O}(1)
\;\in\; \operatorname{lin}\big(\loinv{\eta}\big)\circ\nu\circ\operatorname{lin}(\theta) \]
\end{enumerate}
\end{theorem}
The estimated moduli of continuity are (almost) tight:
\begin{myremark}\label{r:Tight}
Applying first (a) and then (b) always recovers
$f$ to have modulus of continuity $\mu':n\mapsto\mu\big(n+\mathcal{O}(1)\big)+\mathcal{O}(1)$
in place of $\mu$, that is, optimal up to a constant shift; recall Lemma~\ref{l:seminv}c).
\\
On the other hand applying first (b) and then (a) in general recovers $F$ only to have modulus of continuity
$\nu'=\kappa\circ\loinv{\kappa}\circ\nu\circ\lambda\circ\big(\loinv{\lambda}+\mathcal{O}(1)\big)+\mathcal{O}(1)$:
which simplifies to $m\mapsto \nu\big(m+\mathcal{O}(1)\big)+\mathcal{O}(1)$ under additional hypotheses such as
\begin{itemize}
\itemsep0pt
\item Both $\kappa$ and $\lambda$ being surjective (and hence growing at most linearly), or
\item $\nu$ being of the form $\kappa\circ\nu'\circ\loinv{\lambda}$.
\end{itemize}
Since the real unit cube $[0;1]^d$
has linear modulus of continuity (Example~\ref{x:Entropy}a+b)
and the signed binary representation is linearly admissible (Example~\ref{x:SignedDigit}),
Theorem~\ref{t:Main2} yields the following strengthening of Example~\ref{x:Max}a): \\
For any fixed $d,e\in\mathbb{N}$ and non-decreasing $\mu:\mathbb{N}\to\mathbb{N}$,
a function $f:[0;1]^d\to[0;1]^e$ has modulus of continuity $\operatorname{lin}(\mu)$
~iff~ it admits a $(\sigma,\sigma)$-realizer with modulus of continuity $\operatorname{lin}(\mu)$.
\end{myremark}
Recall (Remark~\ref{r:Admissible}) that
\emph{linear metric} reducibility ``$\zeta\preccurlyeq_{\rm O}\xi$''
refines continuous reducibility ``$\zeta\preccurlyeq_{\rm T}\xi$''
by requiring a reduction $F:\operatorname{dom}(\zeta)\to\operatorname{dom}(\xi)$ with $\zeta=\xi\circ F$
to have modulus of continuity $\mu\circ\mathcal{O}\big(\loinv{\kappa}\big)=\mathcal{\scriptscriptstyle O}(\nu)\circ\loinv{\kappa}$
for every modulus of continuity $\nu$ of $\zeta$ and some $\kappa$ of $\xi$.
\begin{theorem}
\label{t:Admissible}
\begin{enumerate}
\item[a)]
Every infinite compact metric space $(X,d)$ of diameter $\leq1$
admits a linearly admissible representation $\xi$ of $X$.
Linear metric reducibility is transitive.
A representation $\zeta$ is linearly admissible iff
(i) it has a modulus of continuity in $\operatorname{lin}(\eta)$, and
(ii) admits a linear metric reduction $\xi\preccurlyeq_{\rm O}\zeta$.
\item[b)]
If $\xi:\subseteq\mathcal{C}\twoheadrightarrow X$ is a linearly admissible
representation of the same compact space $X$ with respect to two metrics $d$ and $d'$,
then it holds $d\leq_\nu\leq d'\leq_\nu d$
for some $\nu\in\operatorname{lin}\big(\loinv{\eta}\big)\circ\operatorname{lin}(\eta)\circ\operatorname{lin}(\eta)$.
\item[c)]
Let $\xi$ and $\upsilon$ be linearly admissible representations for compact $(X,d)$ and $(Y,e)$, respectively.
Then $\xi\times\upsilon:\subseteq\mathcal{C}\ni\bar b\mapsto\big(\xi(b_0,b_2,b_4,\ldots),\upsilon(b_1,b_3,b_5,\ldots)\big)$
is linearly admissible for $\big(X\times Y,\max\{d,e\}\big)$.
\\
Moreover it satisfies the following universal properties:
The projections $\pi_1:X\times Y\ni(x,y)\mapsto x\in X$
has a $\big(\xi\times\upsilon,\xi\big)$-realizer
with linear modulus of continuity $n\mapsto2n$;
$\pi_2:X\times Y\ni(x,y)\mapsto y\in Y$
has a $\big(\xi\times\upsilon,\upsilon\big)$-realizer
with modulus of continuity $n\mapsto2n+1$.
Conversely for every fixed $y\in Y$
the embedding $\imath_{2,y}:X\ni x\mapsto (x,y)\in X\times Y$
has a $\big(\xi,\xi\times\upsilon\big)$-realizer
with modulus of continuity $2n\mapsto n$;
and for every fixed $x\in X$
the embedding $\imath_{1,x}:Y\ni y\mapsto (x,y)\in X\times Y$
has a $\big(\upsilon,\xi\times\upsilon\big)$-realizer
with modulus of continuity $2n+1\mapsto n$.
\item[d)]
Let $\xi$ be a linearly admissible representation of connected compact $(X,d)$.
Then the representation $2^\xi$ of $\mathcal{K}(X)$ from Example~\ref{x:Topology}c)
is \emph{polynomially} admissible.
\end{enumerate}
\end{theorem}
Note that linear `slack' in a modulus of continuity of $\xi$
translates to polynomial one in $2^\xi$.
Item~(c) justifies \cite[Definition~3.3.3.1]{Wei00}
constructing a representation of $X\times Y$
\emph{from} such of $X$ and $Y$ that it
(as opposed one created one `from scratch' by invoking a)
is compatible with the canonical morphisms $X\times Y$.
However \cite[Definition~3.3.3.2]{Wei00} for countable products
does \emph{not} preserve linear admissibility; and neither does
Example~\ref{x:Topology}b), already in case of
spaces $X_j$ with quadratic entropy
according to Example~\ref{x:Entropy}c).
For that purposes a more careful construction is needed:
\begin{theorem}
\label{t:Cartesian}
Fix compact metric spaces $(X_j,d_j)$ of entropies $\eta_j$ and diameters between $1/2$ and $1$, $j\in\mathbb{N}$.
Let $\xi_j:\subseteq\mathcal{C}\twoheadrightarrow X_j$ be \emph{uniformly} linearly admissible
in that (i) it has modulus of continuity $\kappa_j(n)\leq c+c\cdot\eta_j(c+c\cdot n)$
and (ii) to every representation $\zeta_j:\subseteq\mathcal{C}\twoheadrightarrow X_j$
with modulus of continuity $\nu_j$
there exists a mapping $F_j:\operatorname{dom}(\zeta_j)\to\operatorname{dom}(\xi_j)$
with modulus of continuity $\mu_j$
with $\mu_j\big(\kappa_j(n)\big)\leq\nu_j(c+c\cdot n)$
for some $c\in\mathbb{N}$ \emph{in}dependent of $j$.
\\
Let a name of $(x_0,x_1,\ldots x_j,\ldots)\in\prod_j X_j$
be any infinite binary sequence
\begin{multline}
\label{e:Cartesian}
\bar b^{(0)}|_{\kappa_0(0):\kappa_0(1)}, \qquad
\bar b^{(0)}|_{\kappa_0(1):\kappa_0(2)}, \quad
\bar b^{(1)}|_{\kappa_1(0):\kappa_1(1)}, \\
\bar b^{(0)}|_{\kappa_0(2):\kappa_0(3)}, \quad
\bar b^{(1)}|_{\kappa_1(1):\kappa_1(2)}, \quad
\bar b^{(2)}|_{\kappa_2(0):\kappa_2(1)},
\qquad\ldots\ldots \\[1ex] \ldots\ldots\qquad
\bar b^{(0)}|_{\kappa_0(n-1):\kappa_0(n)}, \;
\bar b^{(1)}|_{\kappa_1(n-2):\kappa_1(n-1)}, \;
\ldots \\ \ldots \;
\bar b^{(j)}|_{\kappa_j(n-j-1):\kappa_j(n-j)}, \;
\ldots \;
\bar b^{(n-1)}|_{\kappa_{n-1}(0):\kappa_{n-1}(1)}, \qquad \ldots\ldots
\end{multline}
such that $\bar b^{(j)}$ is a $\xi_j$-name of $x_j$.
Here $\bar b|_{k:\ell}$ abbreviates the finite segment $\bar b_{k},\ldots b_{\ell-1}$ of $\bar b$.
The thus defined representation $\xi:=\prod_j\xi_j:\subseteq\mathcal{C}\twoheadrightarrow\prod_j X_j=:X$
has modulus of continuity $\kappa:n\mapsto\sum\nolimits_{j<n} \kappa_j(n-j)$
and is linearly admissible for $\big(\prod_j X_j,\sup_j d_j/2^j\big)$.
\\
Moreover the projection $\pi_j:\prod_j X_j\ni (x_0,\ldots x_j,\ldots)\mapsto x_j\in X_j$
has a $\big(\prod_i\xi_i,\xi_j\big)$-realizer with modulus of continuity
$m\mapsto \kappa\big(\loinv{\kappa_j}(m)+j\big)$;
and for every fixed $\bar x\in X$ and $j\in\mathbb{N}$, the embedding
$\imath_{j,\bar x}:X_j\ni x_j\mapsto (x_0,\ldots x_j,\ldots)\in X$
has a $\big(\xi_j,\prod_i\xi_i\big)$-realizer
with modulus of continuity
$m\mapsto \kappa_j\big(\max\big\{0,\loinv{\kappa}(m)-j\big\}\big)$.
\end{theorem}
Note that the derived moduli of continuity of the canonical morphisms
$\pi_j$ and $\imath_{j,\bar x}$ agree linearly
with those predicted by Theorem~\ref{t:Main2},
are thus optimal in the sense of Remark~\ref{r:Tight}.
For Cartesian closure we finally treat function spaces,
in view of Lemma~\ref{l:Modulus}c) w.l.o.g. the 1-Lipschitz case:
\begin{theorem}
\label{t:Functions}
Fix convex compact metric space $(X,d)$.
To any linearly admissible representation $\xi$ of $X$
there exists a polynomially admissible representation $\xi'_1$
of the convex compact space $X'_1=\operatorname{Lip}_1(X,[0;1])$ of
non-expansive functions $f:X\to[0;1]$.
It is canonical in that it asserts the application functional
$X'_1\times X\ni (f,x)\mapsto f(x)\in[0;1]$
admit a $(\xi'_1\mu\times\xi,\sigma)$-realizer $F$
with asymptotically optimal modulus of continuity $\leq\operatorname{lin}(\eta'_1)$.
\end{theorem}
Recall that Condition~(ii) of polynomial admissibility means
polynomial metric reducibility ``$\zeta\preccurlyeq_{\rm P}\xi'_1$''
to $\xi'_1$ of any other continuous representation $\zeta$ of $X'_1$.
\subsection{Proofs} \label{s:Proofs}
\begin{proof}[Theorem~\ref{t:Main2}]
\begin{enumerate}
\item[a)]
First suppose $\xi=\xi^\varphi$ the representation from Theorem~\ref{t:Linear}
with modulus of continuity $\kappa(n)\leq\mathcal{O}\big(\eta(n+1)\big)$
and similarly $\upsilon$ with $\lambda(n)\leq\mathcal{O}\big(\theta(n+1)\big)$.
Applying (ii) to $\zeta:=f\circ\xi:\operatorname{dom}(\xi)\subseteq\mathcal{C}\to Y$
with modulus of continuity $\kappa\circ\mu$ yields
a $(\xi,\upsilon)$-realizer $F$ with
modulus $\nu=\kappa\circ\mu\circ\big(1+\loinv{\lambda}\big)$.
\\
Next consider arbitrary linearly admissible $\xi'$ with modulus $\kappa'$
and $\upsilon'$ with $\lambda'$.
Applying (ii) to $\xi'$ yields $G:\operatorname{dom}(\xi')\to\operatorname{dom}(\xi)$
with modulus $\kappa'\circ\big(1+\loinv{\lambda'}\big)$
such that $\xi'=\tilde\xi\circ G$;
and applying (ii) of Definition~\ref{d:Admissible}b) to $\upsilon$ yields $H:\operatorname{dom}(\upsilon)\to\operatorname{dom}(\upsilon')$
with modulus $\lambda\circ\big(\loinv{\lambda'}+\mathcal{O}(1)\big)$
such that $\upsilon=\upsilon'\circ H$.
Together, $F':=H\circ F\circ G$ constitutes a $(\xi',\upsilon')$-realizer of $f$
with modulus of continuity
\begin{eqnarray*}
\nu' &=& \kappa'\circ\big(1+\loinv{\lambda'}\big)\;\circ\;
\kappa\circ\mu\circ\big(1+\loinv{\lambda}\big) \;\circ\;
\lambda\circ\big(\loinv{\lambda'}+\mathcal{O}(1)\big) \\
&\leq& \kappa' \circ (1+\mu) \circ \big(\loinv{\lambda'}+\mathcal{O}(1)\big)
\;\in\; \operatorname{lin}(\eta)\circ\mu\circ\operatorname{lin}\big(\loinv{\theta}\big)
\end{eqnarray*}
by Lemma~\ref{l:seminv}c+e)
since $\eta\leq\kappa'\in\operatorname{lin}(\eta)$ and $\theta\leq\lambda'\in\operatorname{lin}(\theta)$
according to Definition~\ref{d:Admissible}b\,i) and Example~\ref{x:Entropy}f).
\item[b)]
As in (a) first suppose $F$ is a $(\xi,\upsilon)$-realizer of $f$
with modulus $\nu$, for $\xi=\xi^\varphi$ from Theorem~\ref{t:Linear}
with modulus of continuity $\kappa(n)\leq\mathcal{O}\big(\eta(n+1)\big)$
and similarly $\upsilon$ with $\lambda(n)\leq\mathcal{O}\big(\theta(n+1)\big)$.
Then $f\circ\xi=\upsilon\circ F:\operatorname{dom}(\xi)\subseteq\mathcal{C}\to Y$
has modulus of continuity $\nu\circ\lambda\leq \kappa\circ\loinv{\kappa}\circ\nu\circ\lambda$
by Lemma~\ref{l:seminv}c); and (iv) implies $f$ to have modulus
$\mu=1+\loinv{\kappa}\circ\nu\circ\lambda$.
\\
Next consider arbitrary linearly admissible $\xi'$ with modulus $\kappa'$
and $\upsilon'$ with $\lambda'$;
and let $F'$ be a $(\xi',\upsilon')$-realizer of $f$ with modulus $\nu'$.
(ii) yields $H'$ with $\upsilon'=\upsilon\circ H'$ of modulus $\lambda'\circ\big(1+\loinv{\lambda}\big)$;
and $G'$ with $\xi=\xi'\circ G'$ of modulus $\kappa\circ\big(\loinv{\kappa'}+\mathcal{O}(1)\big)$
according to Definition~\ref{d:Admissible}b).
Together, $F:=H'\circ F'\circ G'$ constitutes a $(\xi,\upsilon)$-realizer of $f$
with modulus $\nu=\kappa\circ\big(\loinv{\kappa'}+\mathcal{O}(1)\big)\circ \nu'\circ \lambda'\circ\big(1+\loinv{\lambda}\big)$.
So our initial consideration implies $f$ to have modulus
\begin{eqnarray*}
\mu' &=& 1+\loinv{\kappa}\;\circ\; \kappa\circ\big(\loinv{\kappa'}+\mathcal{O}(1)\big)\circ \nu'\circ \lambda'\circ\big(1+\loinv{\lambda}\big)
\;\circ\; \lambda \\
&\leq& \mathcal{O}(1)+\loinv{\kappa'}\circ\nu'\circ \lambda'(1+\operatorname{id})
\; \in \; \operatorname{lin}\big(\loinv{\eta}\big)\circ\nu'\circ\operatorname{lin}(\theta)
\end{eqnarray*}
by Lemma~\ref{l:seminv}c+e)
since $\eta\leq\kappa'\in\operatorname{lin}(\eta)$ and $\theta\leq\lambda'\in\operatorname{lin}(\theta)$
according to Definition~\ref{d:Admissible}b\,i) and Example~\ref{x:Entropy}f).
\qed\end{enumerate}\end{proof}
\begin{proof}[Theorem~\ref{t:Admissible}]
\begin{enumerate}
\item[a)]
Theorem~\ref{t:Linear} asserts the first claim.
For the second
let $\zeta$ have modulus $\nu$ and $\zeta'$ have modulus $\nu'$ and $\zeta''$ have modulus $\nu''$;
let $F:\operatorname{dom}(\zeta)\to\operatorname{dom}(\zeta')$ with $\zeta=\zeta'\circ F$
have modulus of continuity $\nu\circ\mathcal{O}\big(\loinv{\nu'}\big)$
and $F':\operatorname{dom}(\zeta')\to\operatorname{dom}(\zeta'')$ with $\zeta'=\zeta''\circ F'$
have modulus of continuity $\nu'\circ\mathcal{O}\big(\loinv{\nu''}\big)$.
Then $\zeta=\zeta''\circ F'\circ F$, where $F'\circ F$ has modulus
\[ \nu\circ\mathcal{O}\big(\loinv{\nu'}\big)
\;\circ\; \nu'\circ\mathcal{O}\big(\loinv{\nu''}\big)
\;=\; \nu\circ\mathcal{O}\big(\loinv{\nu''}\big) \enspace . \]
Finally, $\tilde\xi\preccurlyeq_{\rm O}\zeta$ is necessary according to
Definition~\ref{d:Admissible}b\,ii);
while any other representation $\zeta'\preccurlyeq_{\rm O}\tilde\xi\preccurlyeq_{\rm O}\zeta$
since $\tilde\xi$ is admissible.
\item[b)]
First consider $\tilde\xi=\xi^\varphi$ the representation from Theorem~\ref{t:Linear}
with modulus $\tilde\kappa$.
By (iii), to every $x,x'\in X$ with $d(x,x')\leq2^{-m-1}$,
there exist $\tilde\xi$-names $\bar y_x$ and $\bar y'_{x'}$
of $x=\tilde\xi(\bar y_x)$ and $x'=\tilde\xi(\bar y'_{x'})$
with $d_{\mathcal{C}}(\bar y_x,\bar y'_{x'})\leq2^{-\tilde\kappa(m)}$.
Hence $d'(x,x')=d'\big(\tilde\xi(\bar y_x),\tilde\xi(\bar y'_{x'})\big)\leq2^{-n}$
by (i) for $\tilde\kappa'\in\operatorname{lin}(\eta)$ modulus of continuity of $\tilde\xi$ w.r.t. $d'$
and $\tilde\kappa(m)\geq\tilde\kappa'(n)$, that is
for $m=\loinv{\tilde\kappa}\circ\tilde\kappa'(n)$
by Lemma~\ref{l:seminv}c):
$d'\leq_\nu d$ for
$\nu:=1+\loinv{\tilde\kappa}\circ\tilde\kappa'\in\operatorname{lin}\big(\loinv{\eta}\big)\circ\operatorname{lin}(\eta)$.
\\
Now let $\xi$ be linearly admissible with modulus $\kappa$.
Then $\xi=F\circ\tilde\xi$ for some $F:\operatorname{dom}(\xi)\to\operatorname{dom}(\tilde\xi)$
with modulus $\mu\leq\kappa\circ\mathcal{O}\big(\loinv{\tilde\kappa}\big)$
Hence $\bar z_x:=F(\bar y_x)$ and $\bar z'_{x'}=F(\bar y'_{x'})$
have $d_{\mathcal{C}}(\bar z_x,\bar z'_{x'})\leq 2^{-m'}$
for $\tilde\kappa(m)\geq\mu(m')$;
and again $d'(x,x')\leq2^{-n}$
for $m'\geq\tilde\kappa'(n)$ modulus of continuity of $\xi$ w.r.t. $d'$:
$d'\leq_\nu d$ for
$\nu:=1+\loinv{\tilde\kappa}\circ\mu\circ\tilde\kappa'\in\operatorname{lin}\big(\loinv{\eta}\big)\circ\operatorname{lin}(\eta)\circ\operatorname{lin}(\eta)$.
\item[c)]
Let $X$ and $Y$ have entropies $\eta$ and $\theta$, respectively;
$\xi$ and $\upsilon$ moduli of continuity $\mu\leq\operatorname{lin}(\eta)$ and $\nu\leq\operatorname{lin}(\theta)$.
(i) $X\times Y$ has entropy at least $\eta(n-1)+\theta(n-1)-1$ by Example~\ref{x:Entropy}b);
and $\xi\times\upsilon$ has modulus of continuity
$2\cdot\max\{\mu,\nu\}\leq\operatorname{lin}\big(n\mapsto \eta(n-1)+\theta(n-1)-1\big)$.
(ii) Let $\zeta:\subseteq\mathcal{C}\twoheadrightarrow X\times Y$
be any representation with modulus $\kappa$.
Then its projection onto the first component $\zeta_1$
constitutes a continuous representation of $X$ with some modulus $\kappa$.
Hence there exist by hypothesis $F:\operatorname{dom}(\zeta)\to\operatorname{dom}(\xi)$
of modulus $\kappa\circ\mathcal{O}\big(\loinv{\mu}\big)$
such that $\zeta_1=\xi\circ F$;
similarly $\zeta_2=\upsilon\circ G$ with $G:\operatorname{dom}(\zeta)\to\operatorname{dom}(\upsilon)$
of modulus $\kappa\circ\mathcal{O}\big(\loinv{\nu}\big)$.
Then $F\times G:=(F_0,G_0,F_1,G_1,F_2,\ldots):\operatorname{dom}(\zeta)\to\operatorname{dom}(\xi\times\upsilon)$
satisfies $\zeta=(\xi\times\upsilon)\circ(F\times G)$ and has modulus
\[ 2\cdot\max\Big\{\kappa\circ\mathcal{O}\big(\loinv{\mu}\big),\kappa\circ\mathcal{O}\big(\loinv{\nu}\big)\Big\}
\;=\; 2\cdot\kappa\circ \mathcal{O}\big(\loinv{\min\{\mu,\nu\}}\big) \]
$\leq\; \kappa\circ\mathcal{O}\big(\loinv{2\cdot\max\{\mu,\nu\}}\big)$
by Lemma~\ref{l:seminv}h).
\\
Finally, $(b_0,b_1,b_2,b_3,b_4,\ldots)\mapsto (b_0,b_2,b_4,\ldots)$
is a $\big(\xi\times\upsilon,\xi\big)$-realizer of $\pi_1$
with modulus of continuity $n\mapsto2n$;
and $(b_0,b_1,b_2,b_3,b_4,\ldots)\mapsto (b_1,b_3,\ldots)$
a $\big(\xi\times\upsilon,\upsilon\big)$-realizer of $\pi_2$
with modulus of continuity $n\mapsto2n+1$.
For any fixed $\upsilon$-name $(b_1,b_3,\ldots)$ of $y$,
$(b_0,b_2,\ldots)\mapsto (b_0,b_1,b_2,b_3,b_4,\ldots)$
is a realizer of $\imath_{2,y}$ with modulus of continuity $2n\mapsto n$;
and any fixed $\xi$-name $(b_0,b_2,\ldots)$ of $x$,
$(b_1,b_3,\ldots)\mapsto (b_0,b_1,b_2,b_3,b_4,\ldots)$
is a realizer of $\imath_{1,x}$ with modulus of continuity $2n+1\mapsto n$.
\item[d)]
Combine Example~\ref{x:Topology}c) with Example~\ref{x:Entropy}g).
\qed\end{enumerate}\end{proof}
\begin{proof}[Theorem~\ref{t:Cartesian}]
Since $X_j$ has diameter between $1/2$ and $1$, w.l.o.g. $\eta_j(0)=0$ and $\eta_j(n)\geq1$ for $n\geq2$ and w.l.o.g. $\kappa_j(0)=0$.
$X:=\prod_j X_j$ has entropy $\eta(n)\geq\sum_{j<n}\eta_j(n-1-j)-\lfloor n/2\rfloor$ by Example~\ref{x:Entropy}c).
On the other hand the initial segment $\bar b^{(j)}|_{0:\kappa_j(n-j)}$
of a $\xi_j$-name $\bar b^{(j)}$ determines $x_j=\xi_j\big(\bar b^{(j)}\big)$
up to error $2^{-(n-j)}$ w.r.t. $d:=d_j/2^j$;
and is located among the first $\kappa_0(n)+\kappa_1(n-1)+\cdots+\kappa_n(0)=\kappa(n)$
symbols of a $\xi$-name of $(x_j)_{_j}$ according to Equation~(\ref{e:Cartesian});
recall $\kappa_n(0)=0$.
Therefore (i) $\xi$ has modulus of continuity $\kappa(n)$,
which is $\leq\operatorname{lin}\big(\eta(n)\big)$ since
$\kappa_j(n)\leq c+c\cdot\eta_j(c+c\cdot n)$
and $\eta_j(n)\geq1$ for $n\geq2$ `covers' the $\lfloor n/2\rfloor$.
\\
Regarding (ii), let $\zeta:\subseteq\mathcal{C}\twoheadrightarrow X$
have modulus of continuity $\nu$.
The projection $\pi_j:X\ni (x_0,\ldots x_j,\ldots)\mapsto x_j\in X_j$
has modulus of continuity $n\mapsto n+j$ since $X$ is equipped with metric $d=\sup_j d_j/2^j$.
The representation $\zeta_j:=\pi_j\circ\zeta:\operatorname{dom}(\zeta)\subseteq\mathcal{C}\twoheadrightarrow X_j$
thus has modulus $\nu_j:n\mapsto\nu(n+j)$.
By hypothesis (ii) on $\xi_j$, there exists a mapping $F_j:\operatorname{dom}(\zeta)\to\operatorname{dom}(\xi_j)$
whose modulus of continuity $\mu_j$ satifies $\mu_j\big(\kappa_j(n)\big)\leq\nu_j(c+c\cdot n)$
such that it holds $\pi_j\circ\zeta=\xi_j\circ F_j$.
Now let $F:=$
\begin{multline*}
F_0|_{\kappa_0(0):\kappa_0(1)}, \qquad
F_0|_{\kappa_0(1):\kappa_0(2)}, \quad
F_1|_{\kappa_1(0):\kappa_1(1)}, \\
F_0|_{\kappa_0(2):\kappa_0(3)}, \quad
F_1|_{\kappa_1(1):\kappa_1(2)}, \quad
F_2|_{\kappa_2(0):\kappa_2(1)},
\qquad\ldots\ldots \\[1ex] \ldots\ldots\qquad
F_0|_{\kappa_0(n-1):\kappa_0(n)}, \;
F_1|_{\kappa_1(n-2):\kappa_1(n-1)}, \;
\ldots \\ \ldots \;
F_j|_{\kappa_j(n-j-1):\kappa_j(n-j)}, \;
\ldots \;
F_{n-1}|_{\kappa_{n-1}(0):\kappa_{n-1}(1)}, \qquad \ldots\ldots
\end{multline*}
so that $\zeta=\xi\circ F$.
Moreover $F_j|_{0:\kappa_j(n-j)}$ depends on
the first $\mu_j\big(\kappa_j(n-j)\big)$ symbols of its argument;
hence $F$ has modulus of continuity $\mu$ with
\[ \mu\big(\kappa(n)\big)
\;=\; \sup\nolimits_{j<n} \mu_j\big(\kappa_j(n-j)\big)
\;\leq\; \sup\nolimits_{j<n} \nu_j\big(c+c\cdot(n-j)\big)
\;=\; \nu(c+c\cdot n). \]
Mapping an infinite binary sequence according to Equation~(\ref{e:Cartesian})
to $\bar b^{(j)}$ constitutes a $\big(\prod_i\xi_i,\xi_j\big)$-realizer
of $\pi_j$; and for $n\geq j$ the first $\kappa_j(n-j)$ bits of $\bar b^{(i)}$
are determined by the first $\kappa(n)$ bits of the given $\xi$-name:
hence $\kappa_j(n-j)\mapsto\kappa(n)$ a modulus of continuity.
Conversely mapping $\xi_j$-name $\bar b^{(j)}$ of $x_j$
to the infinite binary sequence from Equation~(\ref{e:Cartesian})
constitutes a $\big(\xi_j,\prod_i\xi_i\big)$-realizer of $\imath_{j,\bar x}$
with modulus of continuity $\kappa(n)\mapsto \kappa_j\big(\max\{0,n-j\}\big)$.
\qed\end{proof}
\begin{proof}[Theorem~\ref{t:Functions}]
Let $\eta$ denote the entropy of $X$
and $\kappa\leq\operatorname{lin}(\eta)$ a modulus of continuity of $\xi$.
Only for notational simplicity, consider the case $\mu=\operatorname{id}$
of 1-Lipschitz functions $X'_1:=\operatorname{Lip}_1(X,[0;1])$.
We pick up on, and refine, the entropy analysis from the proof of Example~\ref{x:Entropy}h).
Fix $n\in\mathbb{N}$ and, for every $\vec w\in \{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{\kappa(n)}$
with $\xi\big[\vec w\,\mathcal{C}\big]\neq\emptyset$,
choose\footnote{Different choices lead to different representations $\xi'_1$ of $X'_1$.\label{f:contingent}}
some $x_{\vec w}\in \xi\big[\vec w\,\mathcal{C}\big]$.
Record that, as centers of closed balls of radius $2^{-n}$, these cover $X$.
\\
Next choose\footref{f:contingent}
some subset $W_n\subseteq \{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{\kappa(n)}$
such that any two distinct $\vec w\in W_n$
satisfy $d\big(x_{\vec w},x_{\vec v}\big)>2^{-n}$
while the closed balls $\overline{\operatorname{B}}\big(x_{\vec w},2^{-n+1}\big)$
of double radius still cover $X$.
Such $W_n$ can be created greedily
by repeatedly and in arbitrary order weeding out
one of $(\vec w,\vec v)$ whenever $d\big(x_{\vec w},x_{\vec v}\big)\leq2^{-n}$:
observe $\overline{\operatorname{B}}\big(x_{\vec w},2^{-n+1}\big)\supseteq \overline{\operatorname{B}}\big(x_{\vec v},2^{-n}\big)$
and abbreviate $X_n:=\{ x_{\vec w}:\vec w\in W_n\}$.
\\
We now formalize the idea that a $\xi'_1$-name of $f\in X'_1$
encodes a sequence $f_n:X_n\to\mathbb{D}_n$ of $\tfrac{3}{2}$-Lipschitz functions
whose $\tfrac{3}{2}$-Lipschitz extensions $\extm{f_n}$ approximate
$f$ up to error $2^{-n+1}$:
As in the proof of Example~\ref{x:Entropy}h),
$f'_n:=\lfloor 2^n\cdot f\big|_{X_n}\rceil/2^n$
satisfies this condition, hence asserting every $f\in X'_1$ to have a $\xi'_1$-name,
i.e. $\xi'_1$ be surjective.
\\
In order to encode the $f_n$ succinctly, and make $\xi'_1$ have modulus
of continuity $\kappa'_1\leq\operatorname{lin}(\eta'_1)$ for the entropy $\eta'_1$ of $X'_1$,
recall the connected undirected graph $G_n=(X_n,E_n)$ from the proof of Example~\ref{x:Entropy}f+h)
with edge $(x,x')$ present iff the open balls
of radius $2^{-n+\pmb{2}}$ around centers $x,x'$ intersect.
Choose\footref{f:contingent} some directed spanning tree $F_n\subseteq E_n$ of $G_n$
with root $x_{n,0}$ and remaining nodes $x_{n,1},\ldots x_{n,N_n-1}$
in some \footref{f:contingent} topological order, where $N_n:=\operatorname{Card}(W_n)\leq 2^{\kappa(n)}$.
As in the proof of Example~\ref{x:Entropy}h),
every $\tfrac{3}{2}$-Lipschitz $f_n:X_n\to\mathbb{D}_n$
is uniquely described by $f_n\big(x_{n,0}\big)\in\mathbb{D}_n\cap[0;1]$
together with the sequence
\[ f_n\big(x_{n,m}\big)\;-\;f_n\big(x_{n,m-1}\big) \;\in\; \big\{-6\cdot2^{-n},\ldots 0,\ldots +6\cdot2^{-n}\big\} \quad , 1\leq m<N_n \enspace . \]
The first takes $n+1$ bits to describe, the latter $(N_n-1)\times \log_2(13)$ bits:
in view of Example~\ref{x:Entropy}f) a total of $\mathcal{O}\big(2^{\kappa(n)}\big)$ to encode $f_n$ in some $\vec u_n$.
So the initial segment $(\vec u_0,\ldots \vec u_{n+1})$ of a thus defined $\xi'_1$-name $\bar u=(\vec u_0,\vec u_1,\ldots)$ of $f$,
encoding $f_0,\ldots f_{n+1}$ and thus determining $f$ up to error $2^{-n}$,
has length $\kappa'_1(n)=\mathcal{O}\big(2^{\kappa(0)}\big)+\cdots+\mathcal{O}\big(2^{\kappa(n+1)}\big)\leq\mathcal{O}\big(2^{\kappa(n+2)}\big)\leq\operatorname{lin}(\eta'_1)$
again by Example~\ref{x:Entropy}f+h): thus establishing Condition~(i).
\\
Regarding the application functional $(f,x)\mapsto f(x)$,
consider the following `algorithm' and recall Theorem~\ref{t:Admissible}c):
Given a $(\xi'_1\times\xi)$-name $(u_0,v_0,u_1,v_1,\ldots u_m,v_m,\ldots)$
of $(f,x)$ and $n\in\mathbb{N}$,
let $\vec v:=\bar v_{<\kappa(n)}$ and `find' some $\vec w\in W_n$
with $d\big(x_{\vec w},x_{\vec v}\big)\leq2^{-n+1}$;
then `trace' the path in spanning tree $(W_n,F_n)$ from
its root $x_{n,0}$ to $x_{\vec w}=x_{n,M}$: all information contained
within the first $\kappa'_1(n)$ bits of $\bar u$
encoding $\tfrac{3}{2}$-Lipschitz $f_n:X_n\to\mathbb{D}_n$
whose extension approximates $f$ up to error $2^{-n+1}$,
sufficient to recover the value
\[ y_n \;:=\; f_n\big(x_{\vec w}\big) \;=\; f_n\big(x_{n,0}\big) \;+\;
\sum\nolimits_{m=1}^M \Big(f_n\big(x_{n,m}\big)-f_n\big(x_{n,m-1}\big)\Big)
\;\in\;\mathbb{D}_n \]
satisfying $|y_n-f(x)|\leq |y_n-f_n(x)|+|f_n(x)-f(x)|\leq \tfrac{3}{2}\cdot 2^{-n+1}+2^{-n+1}\leq 2^{-n+3}$.
The (initial segment of length $n+3$ of) sequence $y_0,\ldots y_n,\ldots$ in turn is easily converted
to (an initial segment of length $n$ of) a signed binary expansion of $f(x)$
\cite[Lemma~7.3.5]{Wei00}: yielding a $(\xi'_1\times\xi,\sigma)$-realizer of $(f,x)\mapsto f(x)$ with
asymptotically optimal modulus of continuity $\mu(n)=\max\{2\kappa(n+3),2\kappa'_1(n+3)\}\leq\operatorname{lin}\big(\eta'_1(n+3)\big)$
by Example~\ref{x:Entropy}f+h).
\qed\end{proof}
\section{Conclusion and Perspective}
\label{s:Conclusion}
For an arbitrary compact metric space $(X,d)$
we have constructed a generic representation $\xi$
with optimal modulus of continuity, namely agreeing with the space's entropy
up to a constant factor. And we have shown this representation to exhibit
properties similar to the classical \emph{standard} representation of a topological T$_0$ space
underlying the definition of qualitative \emph{admissibility},
but now under the quantitative perspective crucial for a generic
resource-bounded complexity theory for computing with continuous data:
$\xi$ is maximal with respect to optimal metric reduction
among all continuous representations.
The class of such metrically optimal representations
is closed binary and countable Cartesian products,
and gives rise to metrically optimal representations
of the Hausdorff space of compact subsets
and of the space of non-expansive real functions.
Moreover, with respect to such \emph{linearly} admissible representations
$\xi$ and $\upsilon$ of compact metric spaces $X$ and $Y$,
optimal moduli of continuity of functions $f:X\to Y$
and their $(\xi,\upsilon)$-realizers $F:\operatorname{dom}(\xi)\to\operatorname{dom}(\upsilon)$
are linearly related up to
composition with (the lower semi-inverse) of the entropies
of $X$ and $Y$, respectively.
All our notions (entropy, modulus of continuity) and arguments are information-theoretic:
according to Fact~\ref{f:Proper}b) these precede, and under suitable oracles coincide with,
complexity questions.
They thus serve as general guide to investigations over concrete advanced
spaces of continuous data, such as of integrable or weakly differentiable
functions employed in the theory of Partial Differential Equations
\cite{DBLP:journals/lmcs/Steinberg17}.
\COMMENTED{
\begin{definition}
\label{d:Adapted}
Fix compact metric spaces $(X,d)$ and $(Y,e)$
with entropies $\eta$ and $\theta$ and
with linearly/polynomially admissible representations
$\xi$ and $\upsilon$, respectively.
\emph{Computing} a function $f:X\to Y$
with modulus of continuity $\mu$
in linear/polynomial time
means to compute a $(\xi,\upsilon)$-realizer $F$ of $f$
within time linear/polynomial in
$\operatorname{lin}(\operatorname{id}+\log\eta)\circ \mu\circ\operatorname{lin}\big(\loinv{\operatorname{id}+\log\theta}\big)$.
\end{definition}
This straightforwardly generalizes to \emph{multi}functions \cite{PZ13}.
It consistently `overloads' Definition~\ref{d:Type2} regarding Cantor space,
and recovers the real case $X,Y\in\big\{\mathcal{C},[0;1]\big\}$ having
linear entropies $\eta,\theta$.
}
In order to strengthen our Main Theorem~\ref{t:Main2},
namely to further decrease the gap between (a) and (b)
according to Remark~\ref{r:Tight}, we wonder:
\begin{myquestion}
\label{q:Future}
Which infinite compact metric spaces $(X,d)$ with entropy $\eta$ admit%
\begin{enumerate}\itemsep0pt%
\item[a)]
a
representation
with $\eta\big(n+\mathcal{O}(1)\big)+\mathcal{O}(1)$ as modulus of continuity?
\item[b)]
an \emph{admissible} representation
with modulus of continuity $\eta\big(n+\mathcal{O}(1)\big)+\mathcal{O}(1)$?
\item[c)]
a \emph{linearly} admissible representation
with modulus $\eta\big(n+\mathcal{O}(1)\big)+\mathcal{O}(1)$ ?
\smallskip
\item[d)]
If $\xi$ is a linearly admissible representation of $(X,d)$
and $Z\subseteq X$ closed, \\ is the restriction
$\xi|^{Z}$ then again linearly admissible?
{\rm\cite[Lemma~3.3.2]{Wei00}}
\item[e)]
In view of Example~\ref{x:Entropy}f),
how large can be the asymptotic gap between
intrinsic $\eta_{K,K}$ and relative entropy $\eta_{X,K}$ ?
\item[f)]
(How) does Theorem~\ref{t:Functions} generalize
from real $[0;1]$ to other compact metric codomains $Y$?
\item[g)]
How do the above considerations carry over from the Type-2
setting of computation on streams to that of oracle arguments
{\rm\cite{Ko91,KC12}}?
\end{enumerate}
\end{myquestion}
\subsection{Algorithmic Cost and Representations for Higher Types}
\label{ss:Hyper}
It seems counter-intuitive that the application functional
should be as hard to compute as Example~\ref{x:Max} predicts.
To `mend' this artefact, the Type-2 Machine model
from Definition~\ref{d:Type2} --- computing on `streams'
of infinite sequences of bits
$\bar b=(b_0,b_1,\ldots b_n,\ldots)\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^\mathbb{N}$
with sequential/linear-time access --- is commonly modified
for problems involving continuous subsets or functions as arguments
to allow for `random'/logarithmic-time access {\cite{KC12}}:
\begin{definition}
\label{d:Type3}
\begin{enumerate}
\item[a)]
Abbreviate with $\mathcal{C}':=\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^*}$
the set of total finite string predicates $\varphi:\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^*\to\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}$,
equipped with the metric $d_{\mathcal{C}'}(\varphi,\psi)=2^{-\min\{|\vec u|:\varphi(\vec u)\neq\psi(\vec u)\}}$,
where $|\vec u|\in\mathbb{N}$ denotes the length $n$ of $(u_0,\ldots u_{n-1})\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^*$.
\item[b)]
A \emph{Type-3 Machine} $\mathcal{M}^?$ is an ordinary oracle Turing machine
with variable oracle. It \emph{computes} the partial function $\mathcal{F}:\subseteq\mathcal{C}'\to\mathcal{C}'$
if, for every $\varphi\in\operatorname{dom}(\mathcal{F})\subseteq\mathcal{C}'$,
$\mathcal{M}^\varphi$ computes $\mathcal{F}(\varphi)\in\mathcal{C}'$ in that it
accepts all inputs $\vec v\in\big(\mathcal{F}(\varphi)\big)^{-1}[\textup{\texttt{1}}\xspace]$
and rejects all inputs $\vec v\in\big(\mathcal{F}(\varphi)\big)^{-1}[\textup{\texttt{0}}\xspace]$;
The behaviour of $\mathcal{M}^\varphi$ for $\varphi\not\in\operatorname{dom}(\mathcal{F})$ may be arbitrary.
\item[c)]
For $\psi\in\mathcal{C}'$,
$\mathcal{M}^{\psi,?}$ denotes a Type-3 Machine with fixed oracle $\psi$ and additional variable oracle
in that it operates, for every $\varphi\in\operatorname{dom}(\mathcal{F})\subseteq\mathcal{C}'$,
like $\mathcal{M}^{\psi\otimes\varphi}$, where
\[ \psi\otimes\varphi: (\textup{\texttt{0}}\xspace\,\vec u)\mapsto\psi(\vec u),
\quad (\textup{\texttt{1}}\xspace\,\vec u)\mapsto\varphi(\vec u) \enspace . \]
\item[d)]
$\mathcal{M}^?$ computes $\mathcal{F}$ in \emph{time} $t:\mathbb{N}\to\mathbb{N}$ if
$\mathcal{M}^\varphi$ on input $\vec v\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^n$
stops after at most $t(n)$ steps
regardless of $\varphi\in\operatorname{dom}(\mathcal{F})$;
similarly for $\mathcal{M}^{\psi,?}$.
\end{enumerate}
\end{definition}
Encoding a continuous space using oracles $\psi\in\mathcal{C}'$
rather than sequences $\bar b\in\mathcal{C}$ gave rise to a
different, new notion of representation \cite[\S3.4]{KC12}.
Here we shall call it \emph{hyper-}representation,
and avoid confusion with the previous conception
now referred to as \emph{stream} representation.
\begin{definition}
\label{d:Hyper}
\begin{enumerate}
\item[a)]
A \emph{hyper-}representation of a space $X$ is
a partial surjective mapping $\Xi:\subseteq\mathcal{C}'\twoheadrightarrow X$.
\item[b)]
For hyper-representations $\Xi:\subseteq\mathcal{C}'\twoheadrightarrow X$
and $\Upsilon:\subseteq\mathcal{C}'\twoheadrightarrow Y$,
a $(\Xi,\Upsilon)$-\emph{realizer} of a function $f:X\to Y$
is a partial function $\mathcal{F}:\operatorname{dom}(\Xi)\to\operatorname{dom}(\Upsilon)$
such that $f\circ\Xi=\Upsilon\circ\mathcal{F}$ holds.
\item[c)]
A \emph{reduction} from $\Xi:\subseteq\mathcal{C}'\twoheadrightarrow X$
to $\Xi':\subseteq\mathcal{C}'\twoheadrightarrow X$ is a
$(\Xi,\Xi')$-realizer of the identity $\operatorname{id}:X\to X$.
\item[d)]
$(\Xi,\Upsilon)$-\emph{computing} $f$ means to compute some $(\Xi,\Upsilon)$-realizer $\mathcal{F}$ of $f$
in the sense of Definition~\ref{d:Type3}b).
\item[e)]
The \emph{product} hyper-representation
of $\Xi:\subseteq\mathcal{C}'\twoheadrightarrow X$
and $\Upsilon:\subseteq\mathcal{C}'\twoheadrightarrow Y$ is
\[ \Xi\times\Upsilon:\subseteq\mathcal{C}'
\;\ni\;
\varphi\;\mapsto\; \Big(\Xi\big(\vec v\mapsto \varphi(\textup{\texttt{0}}\xspace\,\vec v)\big),\Upsilon\big(\vec v\mapsto\varphi(\textup{\texttt{1}}\xspace\,\vec v)\big)\Big)
\;\in\; X\times Y \]
\item[f)]
Consider the hyper-representation (sic!)
$\imath_\mathcal{C}: \mathcal{C}' \;\ni\; \varphi \;\mapsto\; \big( \varphi(\textup{\texttt{1}}\xspace^n)_{_n}\big) \;\in\; \mathcal{C}$.
For stream representation $\xi:\subseteq\mathcal{C}\twoheadrightarrow X$,
let $\xi\circ\imath_\mathcal{C}:\subseteq\mathcal{C}'\twoheadrightarrow X$
denote its induced \emph{unary} hyper-representation.
\item[g)]
Abusing notation, consider the hyper-representation
\[ \mathrm{bin}: \mathcal{C}' \;\ni\; \varphi \;\mapsto\; \Big(\varphi\big(\mathrm{bin}(n)\big)_{_n}\Big) \;\in\; \mathcal{C} \enspace . \]
For stream representation $\xi:\subseteq\mathcal{C}\twoheadrightarrow X$,
let $\xi\circ\mathrm{bin}\subseteq\mathcal{C}'\twoheadrightarrow X$
denote its induced \emph{binary} hyper-representation.
\end{enumerate}
\end{definition}
Indeed, according to \cite[\S4.3]{KC12},
appropriate hyper-representations now
allow to compute the application functional in
time more reasonable than in Example~\ref{x:Max}b);
cmp. Item~d) of the following Proposition.
Items~a) to c) correspond to Fact~\ref{f:Proper}.
\begin{proposition}
\label{p:Type3}
\begin{enumerate}
\item[a)]
$\mathcal{C}'$ is compact of entropy $\eta_{\mathcal{C}'}=2^{\operatorname{id}}-1$.
If $\mathcal{F}:\subseteq\mathcal{C}'\to\mathcal{C}'$ is computed by
$\mathcal{M}^{?}$ and if $\operatorname{dom}(\mathcal{F})$ is compact,
then this computation admits a time bound $t:\mathbb{N}\to\mathbb{N}$
in the sense of Definition~\ref{d:Type3}d);
similarly for $\mathcal{M}^{\psi,?}$\ldots
\item[b)]
If $\mathcal{M}^{\psi,?}$ computes $\mathcal{F}:\subseteq\mathcal{C}'\to\mathcal{C}'$ in time $t(n)$,
then $\mathcal{F}$ has modulus of continuity $t(n)$.
\item[c)]
If $\mathcal{F}:\subseteq\mathcal{C}'\to\mathcal{C}'$ has modulus
of continuity $\mu:\mathbb{N}\to\mathbb{N}$, then there exists an oracle $\psi\in\mathcal{C}'$
and a Type-3 Machine $\mathcal{M}^{\psi,?}$ computing $\mathcal{F}$
in time $\mathcal{O}\big(n+2^{t(n)}\big)$.
\item[d)]
The compact space $[0;1]'_1$ from Example~\ref{x:Max}b)
admits a hyper-representation $\Delta_1$ such that application
$[0;1]'_1\times[0;1]\ni(f,r)\mapsto f(r)\in[0;1]$
is $(\Delta_1\times\tilde\delta,\delta\circ\imath_\mathcal{C})$-computable in polynomial time
for the unary hyper-representation $\delta\circ\imath_\mathcal{C}$
induced by the aforementioned dyadic stream representation $\delta$ of $[0;1]$.
\item[e)]
Hyper-representation (sic!) $\imath_\mathcal{C}:\subseteq\mathcal{C}'\twoheadrightarrow\mathcal{C}$ is an isometry.
Stream representation $\xi:\subseteq\mathcal{C}\twoheadrightarrow X$ has modulus of continuity $\kappa$
iff induced unary hyper-representation $\xi\circ\imath_\mathcal{C}$ does.
\item[f)]
For $(\xi,\upsilon)$-realizer $F:\subseteq\mathcal{C}\to\mathcal{C}$ of $f:X\to Y$,
$\mathcal{F}:=\imath_\mathcal{C}^{-1}\circ F\circ\imath_\mathcal{C}:\subseteq\mathcal{C}'\to\mathcal{C}'$
is a $\big(\xi\circ\imath_\mathcal{C},\upsilon\circ\imath_\mathcal{C}\big)$-realizer
of $f:X\to Y$. $F$ has modulus of continuity $\mu$ iff $\mathcal{F}$ does.
\item[g)]
Hyper-representation $\mathrm{bin}:\subseteq\mathcal{C}'\twoheadrightarrow\mathcal{C}$ has logarithmic modulus of continuity.
Its inverse, the stream representation $\mathrm{bin}:\mathcal{C}\twoheadrightarrow\mathcal{C}'$
has exponential modulus of continuity. Both are optimal.
Stream representation $\xi:\subseteq\mathcal{C}\twoheadrightarrow X$ has modulus of continuity $\mathcal{O}\big(2^\kappa\big)$
iff induced binary hyper-representation $\xi\circ\mathrm{bin}:\subseteq\mathcal{C}'\twoheadrightarrow X$
has modulus of continuity $\kappa$.
\end{enumerate}
\end{proposition}
Note the gap between Type-3 Proposition~\ref{p:Type3}b+c),
absent in the Type-2 Fact~\ref{f:Proper}b+c).
From a high-level perspective,
the exponential lower complexity bound to the application functional
in Example~\ref{x:Max}b) is due to $[0;1]'$ having
exponential entropy while stream representations'
domain $\mathcal{C}$ has only linear entropy:
see Example~\ref{x:Entropy}a+h).
Proposition~\ref{p:Type3}d)
avoids that information-theoretic bottleneck
by proceeding to hyper-representations
with domain $\mathcal{C}'$ also having exponential entropy.
In fact {\cite[\S4.3]{KC12}} extends the representation
$\Delta_1$ of $[0;1]'_1=\operatorname{Lip}_1([0;1],[0;1])$ from Proposition~\ref{p:Type3}d)
to entire $\mathcal{C}\big([0;1],[0;1]\big)=\bigcup_{\mu} \mathcal{C}_\mu\big([0;1],[0;1]\big)$,
where the union ranges over all strictly increasing $\mu:\mathbb{N}\to\mathbb{N}$.
Lacking compactness, in view of Proposition~\ref{p:Type3}a)
one cannot expect a time bound
on entire $\mathcal{C}\big([0;1],[0;1]\big)$
depending only on the output precision $n$.
Instead {\cite[\S3.2]{KC12}} considers runtime \emph{polynomial}
if bounded by some term $P=P(n,\mu)$
in both the integer output precision parameter $n$ and a
modulus of continuity $\mu:\mathbb{N}\to\mathbb{N}$ of the given function argument $f$:
a higher-type parameter \cite{EikeFlorian}.
The considerations in the present work suggest
a natural
\begin{myremark}
\label{r:SecondOrder}
According to Example~\ref{x:Entropy}h)
$\mathcal{C}_\mu([0;1],[0;1])$ has entropy $\eta=\Theta\big(2^{\mu}\big)$
such that $\log\eta=\Theta(\mu)$ `recovers' the modulus of continuity.
Moreover our complexity-theoretic ``Main'' Theorem~\ref{t:Main2}
confirms that even metrically well-behaved (e.g. 1-Lipschitz) functionals $\Lambda:X\to[0;1]$
can only have realizers with modulus of continuity/time complexity
growing in $X$'s entropy $\eta$. This suggests generalizing
second-order polynomial runtime bounds $P=P(n,\mu)$
from spaces of continuous real functions {\cite[\S3.2]{KC12}}
to $P(n,\log\eta)$ for compact metric spaces $X$ beyond $\mathcal{C}_\mu([0;1],[0;1])$.
The logarithm is consistent with the quantitative properties of hyper-representations
expressed in Proposition~\ref{p:Type3}.
\end{myremark}
\begin{proof}[Proposition~\ref{p:Type3}]
\begin{enumerate}
\item[a)]
Compactness of $\mathcal{C}'$ follows from K\"{o}nig's Lemma:
it is an infinite finitely (only, as opposed to $\mathcal{C}$,
increasingly) branching tree.
Cover $\mathcal{C}'$ by $2^{2^n-1}$ closed balls
$\displaystyle\big\{\varphi:\varphi\big|_{\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{<n}}=\psi\big\}$,
$\psi:\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^n\to\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}$ of radius $2^{-n}$: optimally.
\\
Consider the number $N:=t_\mathcal{M}(\varphi,\vec u)\in\mathbb{N}$
$\mathcal{M}^{\varphi}$ makes on input $\vec u\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^n$
for $\varphi\in\operatorname{dom}(\mathcal{F})\subseteq\mathcal{C}'$.
During this execution, $\mathcal{M}^{\varphi}$
can construct and query oracle $\varphi$
only on strings $\vec v$ of length $|\vec v|<N$:
Replacing $\varphi$ with some $\psi\in\mathcal{C}'$
of distance $d_{\mathcal{C}'}(\varphi,\psi)\leq2^{-N}$
will remain undetected, that is,
$\mathcal{M}^{\psi}$ on input $\vec u$
will behave the same way, and in particular
still terminate after $N$ steps.
This establishes continuity of $t_\mathcal{M}(\cdot,\vec u)$.
By compactness of $\operatorname{dom}(\mathcal{F})$,
the following maxima thus exist:
\[ t_\mathcal{M}(\vec u) \;:=\; \max \big\{t_\mathcal{M}(\varphi,\vec u)\::\:\varphi\in\mathcal{F}\big\}
\qquad
t_\mathcal{M}(n)\;:=\;\max\big\{ t_\mathcal{M}(\vec u)\::\:
\vec u\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{n}\big\} \]
\item[b)]
As mentioned in the proof of (a),
$\mathcal{M}^{\varphi}$ and $\mathcal{M}^{\psi}$
will behave identically on all inputs $\vec u\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^n$
for $\varphi,\psi\in\operatorname{dom}(\mathcal{F})$
with $d_{\mathcal{C}'}(\varphi,\psi)\leq2^{-t(n)}$:
Meaning $\mathcal{F}(\varphi)$ and $\mathcal{F}(\psi)$
have distance $\leq 2^{-n}$.
\item[c)]
By hypothesis, $\mathcal{F}\big(\varphi\big)(\vec u)\in\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}$
depends only on the restriction $\displaystyle\varphi\big|_{\{\textup{\texttt{0}}\xspace,\textup{\texttt{1}}\xspace\}^{<m}}$
for $m:=\mu(n)$ and $n:=|\vec u|$. Thus
\[ \psi\big(\vec u,\varphi(),\varphi(\textup{\texttt{0}}\xspace),\varphi(\textup{\texttt{1}}\xspace),\varphi(\textup{\texttt{0}}\xspace\sdzero),\ldots\varphi(\textup{\texttt{1}}\xspace^{m-1})\big)
\;:=\;\mathcal{F}\big(\varphi\big)(\vec u) \]
is well-defined an oracle.
And, for given $\vec u$ and $\varphi$,
making this query of length $\mathcal{O}(n+2^m)$
recovers the value $\mathcal{F}\big(\varphi\big)(\vec u)$.
\qed\end{enumerate}\end{proof}
\subsection{Representation Theory of Compact Metric Spaces}
\label{ss:Future}
Both stream and hyper representations
translate (notions of computability and bit-complexity) to their co-domains
from `universal' \cite{Benyamini}
compact structures (which are already naturally equipped with formal conceptions of computing,
namely) $\mathcal{C}$ and $\mathcal{C}'$, respectively.
Matthias Schr\"{o}der [personal communication 2017] has suggested a third candidate domain
of generalized representations: the Hilbert Cube $\mathcal{H}=\prod_{j\geq0} [0;1]$
equipped with metric $d_{\mathcal{H}}(\bar x,\bar y)=\sup_j |x_j-y_j|/2^j$; recall Example~\ref{x:Entropy}a).
This parallels earlier developments in continuous computability theory
considering \emph{equilogical} \cite{DBLP:journals/tcs/BauerBS04}
and \emph{quotients of countably-based topological} (=QCB)
spaces \cite{DBLP:conf/cie/Schroder06}.
And its suggests the following generalization:
\begin{definition}
\label{d:Representation}
Fix a compact metric space $(\frakX,D)$ with entropy $\Theta$.
\begin{enumerate}
\item[a)]
A \emph{$\frakX$-representation} of another compact metric space $(X,d)$
is a surjective partial mapping
$\xi:\subseteq\frakX\twoheadrightarrow X$.
\item[b)]
Let $\eta$ denote the entropy of $(X,d)$. \\
Call $\xi$ \emph{linearly admissible} if
it has a modulus of continuity $\kappa$ such that \\
(i) $\Theta\circ\kappa\leq\mathcal{O}\mathcal{\scriptscriptstyle S}(\eta)$, i.e.,
$\exists C\in\mathbb{N} \; \forall n\in\mathbb{N}: \; \big(\Theta\circ\kappa\big)(n)\leq C+C\cdot \eta(n+C)$ \\
and (ii) for every uniformly continuous partial surjection
$\zeta:\subseteq\frakX\twoheadrightarrow X$
it holds $\zeta\preccurlyeq_{\rm O}\xi$, meaning:
There exists a map $F:\subseteq\operatorname{dom}(\zeta)\to\operatorname{dom}(\xi)$
such that $\zeta=\xi\circ F$ and,
for every modulus of continuity $\nu$ of $\zeta$,
$F$ has a modulus of continuity $\mu$
satisfying $\mu\circ\kappa\leq\mathcal{\scriptscriptstyle O}(\nu)$.
\item[c)]
Call $\xi$ \emph{polynomially admissible} if
it has a modulus of continuity $\kappa$ such that \\
(i) $\Theta\circ\kappa\leq\mathcal{P}\mathcal{\scriptscriptstyle O}(\eta)$, and \\
(ii) for every uniformly continuous partial surjection
$\zeta:\subseteq\frakX\twoheadrightarrow X$
it holds $\zeta\preccurlyeq_{\rm P}\xi$, meaning:
There exists a map $F:\subseteq\operatorname{dom}(\zeta)\to\operatorname{dom}(\xi)$
such that $\zeta=\xi\circ F$ and,
for every modulus of continuity $\nu$ of $\zeta$,
$F$ has a modulus of continuity $\mu$
satisfying $\mu\circ\kappa\leq\mathcal{\scriptscriptstyle P}(\nu)$.
\item[d)]
$(\frakX,D)$ is linearly/polynomially \emph{universal}
if every compact metric space $(X,d)$
admits a linearly/polynomially admissible $\frakX$-representation.
\end{enumerate}
\end{definition}
Note how (b) and (c) boil down to Definition~\ref{d:Admissible}d+e)
in case $(\frakX,D)=(\mathcal{C},d_{\mathcal{C}})$ of linear entropy $\Theta=\operatorname{id}$.
And Theorem~\ref{t:Linear} now means that
Cantor space is linearly universal.
It seems worthwhile to identify and classify linearly/polynomially
universal compact metric spaces.
\bibliographystyle{alpha}
|
\section{INTRODUCTION}
\subsection{Motivation}
Complex systems are encountered in many applications, including cooperative autonomous agents, sensor fusion, and biological systems.
Referring to something as complex implies that it consists of interconnected agents which adapt and respond to their local and global environment.
As we move towards increasingly complex systems \cite{Malikopoulos2015},
new control approaches are needed to optimize the impact on system
behavior of the interaction between its entities \cite{Malikopoulos2015b,Malikopoulos}.
Robotic swarm systems can exhibit complex behavior and have attracted considerable attention in many applications, e.g., transportation\cite{Malikopoulos2018, Malikopoulos2020a, Ren2007}, construction \cite{Lindsey2012, Joshi2014}, and surveillance\cite{Cort??s2009}.
A common requirement for swarms is to move into a desired formation. However, due to cost constraints imposed on individual agents in a swarm, e.g., limited computation capabilities, battery capacity, and sensing abilities, any efficient control approach must take into account energy consumption. The task of moving in a specified formation has been explored in the literature \cite{Oh2017,Oh2015,Brambilla2013}. However, achieving formations with minimum energy consumption during operation has not yet been thoroughly investigated.
Several approaches to building cohesive formations in robotic systems have been proposed, such as formations built from triangular sub-structures \cite{Guo2010, Hanada2007}, where a scalable formation is achieved through the construction of a series of isosceles triangles. Methods inspired by crystal growth \cite{Song} and lattices structures \cite{Lee2008} have also shown promise.
Other control methods using only scalar, bearing, or distance measurements were presented by Swartling et al. \cite{Swartling2014}. This approach was generalized to include the case where only a single leader agent was able to make distance or bearing measurements.
The problem of generating a desired formation was solved via scheduling by Turpin et al. \cite{Turpin}, where an initial assignment is achieved using a scheduling-based heuristic run on a central computer with global information. A significant amount of work, e.g., Wang and Xin \cite{Wang2013}, Sun and Cassandras \cite{Sun}, Xu and Carrillo \cite{Xu2015}, and Rajasree and Jisha \cite{Rajasree}, used optimization techniques in their solutions. However, these methods optimized the
position of each agent in a virtual potential field and did not consider energy consumption by individual agents.
The contribution of this paper is an assignment and trajectory generation algorithm which uses only local information for each agent. Other approaches, such as those by Turpin et al. \cite{Turpin2014}, Morgan et al. \cite{Morgan2016}, or Rubenstein et al. \cite{Rubenstein2012}, required global information in terms of a priori assignment, characteristics about the communication network size, or specifically oriented seed agents, respectively.
Our proposed formulation is valid for any feasible initial and final conditions, requiring only that the initial and final positions be non-overlapping.
In addition, the formulation does not rely on potential fields \cite{Wang2013,Sun,Xu2015}, and instead produces energy-optimal trajectories which use proactive steering to avoid collisions.
The remainder of this paper proceeds as follows. In Section \ref{sec:problem}, we formulate the decentralized optimal control problem for each agent. In Section \ref{sec:solution}, we provide the problem formulation and solution approach of the assignment and trajectory generation and discuss implications on robustness. We present a numerical case study in Section \ref{sec:simulation}, which shows the behavior of the proposed method. Finally, we draw concluding remarks and discuss some ideas about future work in Section \ref{sec:conclusion}.
\section{PROBLEM FORMULATION} \label{sec:problem}
We consider the set $\mathcal{A} = \{1, \dots, N\},\, N \in \mathbb{N}_{>0},$ to index a system of autonomous agents in $\mathbb{R}^2$. The agents are moving into a desired formation indexed by a set of $\mathcal{F} = \{1, \dots, M\},\, M \in \mathbb{N}_{>0},$ goals. We consider the case where $N \leq M$, i.e., no redundant agents are brought to fill the formation, as shown in Fig. \ref{fig:introFigure}. This requirement can be relaxed by defining a behavior for excess agents, such as idling \cite{Turpin2014}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.75\linewidth]{IntroFigure.png}
\caption{A group of $N=5$ agents entering a formation consisting of $M=7$ goals in $\mathbb{R}^2$.}
\label{fig:introFigure}
\end{figure}
Each agent, $i\in\mathcal{A}$, is modeled as a double integrator
\begin{align}
\mathbf{\dot{p}}_i(t) &= \mathbf{v}_i(t), \label{eqn:pDynamics} \\
\mathbf{\dot{v}}_i(t) &= \mathbf{u}_i(t),\label{eqn:vDynamics}
\end{align}
where $\mathbf{p}_i(t)\in\mathbb{R}^2$ and $\mathbf{v}_i(t)\in\mathbb{R}^2$ are the time-varying position and velocity vectors respectively, and $\mathbf{u}_i(t)\in\mathbb{R}^2$ is the control input over time $t\in[0, T_i]$, where $T_i\in \mathbb{R}_{>0},$ is the arrival time for agent $i$ to its assigned goal. Each agent's velocity and control input are bounded, namely,
\begin{align}
v_{\min} \leq ||\mathbf{v}_i(t)|| \leq v_{\max},
\label{eqn:vBounds}\\
u_{\min} \leq ||\mathbf{u}_i(t)|| \leq u_{\max},
\label{eqn:uBounds}
\end{align}
where $||\cdot||$ denotes the Euclidean norm, and $v_{\min}$, $v_{\max}$ and $u_{\min}$, $u_{\max}$ are the minimum and maximum allowable speed and control input respectively for each agent $i\in\mathcal{A}$. The state of each agent is the time-varying vector
\begin{equation}
\mathbf{x}_i(t) =
\begin{bmatrix}
\mathbf{p}_i(t) \\ \mathbf{v}_i(t)
\end{bmatrix}.
\end{equation}
Our objective is to develop a framework for the $N$ agents to optimally, in terms of energy, create any feasible formation of $M$ points while avoiding collisions between agents.
The energy consumption of each agent $i\in\mathcal{A}$ is given by
\begin{equation} \label{eqn:energy}
\dot{E}_i(t) = \frac{1}{2} ||\mathbf{u}_i(t))||^2.
\end{equation}
By minimizing the $L^2$ norm of the control input (acceleration/deceleration) we will have direct benefits in energy consumption.
\begin{definition} \label{def:goals}
The \textit{desired formation} is the set of time-varying goals $\mathcal{G}(t) = \{\mathbf{p}_k(t) : \mathbb{R} \to \mathbb{R}^2 ~ | ~ k \in \mathcal{F}\}$. The set $\mathcal{G}$ can be prescribed offline, i.e., by a human designer, or online by a high-level planner.
\end{definition}
Next, we present our modeling framework, which outlines the approach and assumptions used to solve the minimum energy desired formation problem.
\subsection{Modeling Framework}
In this framework, the agents can communicate with each other. The maximum sensing and communication range, $h\in\mathbb{R}_{>0}$, is used to define a neighborhood for each agent.
\begin{definition} \label{def:neighborhood}
The \textit{neighborhood} of agent $i\in\mathcal{A}$ is defined as the time-varying set
\begin{equation*}
\mathcal{N}_i(t) = \{j\in\mathcal{A} ~ | ~ \big|\big|\textbf{p}_i(t) - \textbf{p}_j(t) \big|\big| \leq h\}.
\end{equation*}
\end{definition}
An agent $i\in\mathcal{A}$ is able to measure the relative position of any neighboring agent $j\in\mathcal{N}_i$. This leads to a natural definition of the scalar separating distance.
\begin{definition} \label{def:separating}
The \textit{scalar separating distance} is defined as
\begin{equation*}
r_{ij}(t) = \big|\big|\mathbf{p}_i(t) - \mathbf{p}_j(t)\big|\big|.
\end{equation*}
\end{definition}
Each agent $i\in\mathcal{A}$ occupies a closed disk of radius $R$. To guarantee no collisions between any agents $i,j\in\mathcal{A}$, we impose the following conditions on the system
\begin{align} \label{eqn:collisionCondition}
r_{ij}(t) &> 2R,\,\, t\in\mathbb{R}_{>0}, \\
h &>> 2R. \label{eqn:sensingCondition}
\end{align}
To ensure each goal in the formation is feasible, the following condition should hold
\begin{equation} \label{eqn:formationSpacing}
\min_{\mathbf{p}(t),\mathbf{q}(t) \in\mathcal{G}}\{||\mathbf{p}(t) - \mathbf{q}(t) ||\} <2R, ~ t\in\mathbb{R}_{>0}.
\end{equation}
In our modeling framework we impose the following assumptions:
\begin{assumption} \label{smp:perfect}
The state $\mathbf{x}_i(t)$ for each agent $i\in\mathcal{A}$ is perfectly observed and there is negligible communication delay between the agents.
\end{assumption}
Assumption \ref{smp:perfect} is required to evaluate the idealized deterministic performance of the generated optimal solution.
\begin{assumption} \label{smp:identical}
All agents are homogeneous, and any agent may fill any goal in the formation.
\end{assumption}
This assumption simplifies the trajectory generation and assignment problems, and it can generally be relaxed by adding goal types as a constraint on the goal assignment.
\begin{assumption} \label{smp:energy}
The energy cost of communication is negligible; the only energy consumption is in the form of (\ref{eqn:energy}).
\end{assumption}
The strength of this assumption is application dependent. For cases with long-distance communications or high data rates, the trade-off for communication cost can be controlled by the selection of $h$.
Under this framework, the energy-optimal desired formation problem can be solved. This problem can be decomposed into two coupled subproblems: (1) goal assignment and (2) trajectory generation. Both of these problems are described in the following section, with emphasis on the goal assignment.
\section{Solution Approach} \label{sec:solution}
The decentralized desired formation problem is solved by decomposing it into the coupled goal assignment and trajectory generation subproblems. To decouple these problems the minimum energy objective in the assignment problem is approximated by the minimum Euclidean distance. Prior work, \cite{Turpin2014, Morgan2016}, has shown that this approximation is generally sufficient. This enables the assignment problem to be solved independently, which results in the endpoint constraints for the minimum-energy trajectory generation.
\subsection{Assignment Problem}
The objective of the assignment problem is to assign each agent to a goal such that the total distance traveled by all agents is minimized. In the decentralized case, each agent $i\in\mathcal{A}$ only has information about the positions of its neighbors, $j\in\mathcal{N}_i$, and the available goals, $\mathcal{G}$. A local assignment can be realized with the use of a local assignment matrix, $\mathbf{A}_i$,
\begin{equation}
\begin{pmatrix}
\mathbf{p}_1(T_1) \\ \mathbf{p}_2(T_2) \\ \vdots \\ \mathbf{p}_n(T_n)
\end{pmatrix}
=
\mathbf{A}_i
\begin{pmatrix}
g_1 \\ g_2 \\ \vdots \\ g_M
\end{pmatrix},
\label{eqn:assignment}
\end{equation}
where $\mathbf{p}_j(T_j),~j\in\mathcal{N}_i$ are the final positions of each agent being assigned, $g_k, ~ k\in\mathcal{F},$ are the indices for each goal, and the elements $a_{jk}\in \mathbf{A}_i$ are binary assignment variables.
Each agent $i\in\mathcal{A}$ can solve (\ref{eqn:assignment}) independently as a linear program, and use the solution to select the prescribed goal.
\begin{definition} \label{def:prescribedGoal}
For each agent $i\in\mathcal{A}$ the \textit{prescribed goal}, $\mathbf{p}^a_i(t)$, is defined as the goal assigned to agent $i$ for which
\begin{equation}
\mathbf{p}_i^a(t) \in \{\mathbf{p}_k \in \mathcal{G} ~|~ k\in\mathcal{F}, ~ a_{ik}=1,~ a_{ik} \in\mathbf{A}_i\},
\end{equation}
where the right hand side is a singleton set
\end{definition}
It is possible for multiple agents to have the same prescribed goal. This occurs when two agents $i\in\mathcal{A},~ j\in\mathcal{N}_i,$ have different neighborhoods and use conflicting information to solve their assignment problem. This conflict is resolved by introducing the \textit{banned goal} set, defined next.
\begin{definition}\label{def:bannedGoal}
For any agent $i\in\mathcal{A}$, the \textit{banned goal} set is defined as the set $\mathcal{B}_i \subset \mathcal{G}$ which consists of all goals that agent $i$ is \emph{permanently banned from} when solving the goal assignment (\ref{eqn:assignment})
\end{definition}
The following definitions and algorithm are presented for some agent $i\in\mathcal{A}$. However, all steps are performed \emph{simultaneously by all agents}.
For this agent $i$, the banned goal set is initially empty. Goals may be added to this set whenever
the following condition is not satisfied
\begin{equation}
\begin{aligned}
\mathbf{p}^a_i(t) \neq \mathbf{p}^a_j(t),\,\, \forall j\in \mathcal{N}_i(t).
\end{aligned}
\label{eqn:NoConflicts}
\end{equation}
In the case that (\ref{eqn:NoConflicts}) is not satisfied, some agent(s) must be \emph{permanently banned} from the conflict goal, defined for agent $i$ as
\begin{equation} \label{eqn:conflictGoal}
\mathbf{p}_c(t) \coloneqq \mathbf{p}_i^a(t).
\end{equation}
Banning is achieved by sequential application of ``tiebreaker" heuristics which compare:
\begin{enumerate}
\item the size of each agent's neighborhood,
\item the distance between each agent and the goal, and
\item the index of each agent.
\end{enumerate}
Since the metrics of criteria 1, 2, and 3 are perfectly measurable (Assumption \ref{smp:perfect}), it follows that all agents must agree on the tiebreaker resolutions.
The tiebreaker hierarchy allows the banned goal set to be broken into three partitions,
\begin{equation}\label{eqn:banPartitions}
\mathcal{B}_i(t) = \mathcal{B}^1_i(t) \cup \mathcal{B}^2_i(t) \cup \mathcal{B}^3_i(t),
\end{equation}
where superscripts $1$, $2$, and $3$ refer to the three tiebreakers, respectively. The tiebreakers are performed by all agents in the set of competing agents, defined next.
\begin{definition}\label{def:competingAgents}
The set of \textit{competing agents} for agent $i\in\mathcal{A}$ is defined as
\begin{equation*}
\mathcal{C}_i(t) = \Big\{k\in\mathcal{N}_i(t) ~|~ \mathbf{p}_a^k(t) = \mathbf{p}_c(t) \Big\}.
\end{equation*}
\end{definition}
When $|\mathcal{C}_i| > 1$ there are at least two agents, $i,j\in\mathcal{N}_i$ assigned to $\mathbf{p}_c$. Similarly to (\ref{eqn:banPartitions}), the set of competing agents can be split into three decreasing subsets,
\begin{align}
\mathcal{C}_i^3 \subseteq \mathcal{C}_i^2 \subseteq \mathcal{C}_i^1 = \mathcal{C}_i
\end{align}
where the superscripts 1, 2, and 3 correspond to the agents which are comparing the three tiebreaker heuristics.
For each agent $i\in\mathcal{A}$, the banned goal sets partitions in \eqref{eqn:banPartitions} are defined as
\begin{align} \label{eqn:bannedUpdate}
\mathcal{B}^m_i(t) = &\Big\{\bigcup_{\tau=0}^t \Big(\{\mathbf{p}_i^a(\tau)\} \cap \Phi_i^m(\tau)\Big)\Big\},
\end{align}
where $\Phi^m_i(t)$ is given by
\begin{equation} \label{eq:indicators}
\Phi_i^m(t) =
\begin{cases}
\mathcal{G}, & \text{if } m=1, ~ i\neq \underset{j\in\mathcal{C}_i^1(t)}{\text{argmax}}\{|\mathcal{N}_j|(t)\}, \\
\mathcal{G}, & \text{if } m=2, ~ i\neq \underset{j\in\mathcal{C}_i^2(t)}{\text{argmax}}\{||\mathbf{p}_c(t)-\mathbf{p}_j(t)||\}, \\
\mathcal{G}, & \text{if } m=3, ~ i\neq \underset{j\in\mathcal{C}_i^3(t)}{\text{argmin}}\{j\}, \\
\emptyset, & \text{otherwise}.
\end{cases}
\end{equation}
where $m\in\{1, 2, 3\}$ again corresponds to the three tiebreaker heuristics.
To begin the tiebreaker process for agent $i\in\mathcal{A}$, consider the first conflict set $\mathcal{C}^1_i$ with the neighborhood heuristic. Every agent $j\in\mathcal{C}_i^1$ which satisfies
\begin{equation}\label{eqn:neighborGoalCondition}
j = \arg\max_{k\in\mathcal{C}_i^1}\{|\mathcal{N}_k(t)|\},
\end{equation}
is eligible to be assigned to goal $\mathbf{p}_c$. If agent $i\in\mathcal{A}$ uniquely satisfies \eqref{eqn:neighborGoalCondition}, then the conflict test is complete and $i$ is assigned to $\mathbf{p}_c$. If $i$ does not satisfy \eqref{eqn:neighborGoalCondition}, then the goal $\mathbf{p}_c$ is added to $\mathcal{B}^1_i(t)$ as designated by \eqref{eq:indicators}. Finally, if agent $i$ does not uniquely satisfy \eqref{eqn:neighborGoalCondition} then the second criteria, distance to goal, must be compared. This comparison is done over a reduced conflict set,
\begin{equation}
\mathcal{C}_i^2(t) = \Big\{j\in\mathcal{C}_i^1(t) ~|~ |\mathcal{N}_j(t)| = |\mathcal{N}_i(t)| \Big\}.
\end{equation}
The second tiebreaker, maximum distance, is a minimax strategy which seeks to minimize the maximum distance traveled by any agent to the conflict goal. Again, every agent $j\in\mathcal{C}_i^2$ which satisfies
\begin{equation}\label{eqn:claimDistCondition}
j = \underset{k\in\mathcal{C}_i^2}{\text{argmax}}\{||\mathbf{p}_k - \mathbf{p}_c||\},
\end{equation}
is eligible to be assigned to goal $\mathbf{p}_c$. If agent $i$ uniquely satisfies \eqref{eqn:claimDistCondition}, then the conflict test is complete and $i$ is assigned to $\mathbf{p}_c$. If $i$ does not satisfy \eqref{eqn:claimDistCondition}, then the goal $\mathbf{p}_c$ is added to $\mathcal{B}^2_i$ per \eqref{eq:indicators}. Finally, if $i$ satisfies \eqref{eqn:claimDistCondition}, but not uniquely, the final test must be taken over a further reduced conflict set, given by
\begin{equation}
\mathcal{C}_i^3(t) = \Big\{j\in\mathcal{C}_i^2(t) ~|~ ||\mathbf{p}_c - \mathbf{p}_j|| = ||\mathbf{p}_c - \mathbf{p}_i|| \Big\},
\end{equation}
where the agent $k$ satisfying
\begin{equation} \label{eqn:idAgent}
k = \min\Big\{ j\in\mathcal{C}_i^3 \Big\},
\end{equation}
is assigned to the goal, and all other agents add $\mathbf{p}_c$ to $\mathcal{B}_i^3$ as designated by \eqref{eq:indicators}.
After the conflicts are resolved, if the size of $B_i$ has increased then the value of $T_i$ must also increase to
\begin{equation}
T_i = t + T,
\end{equation}
where $t$ is the current time, and $T$ is a system parameter. This allows agent $i$ a sufficient amount of time to reach its new goal.
Finally, for each subsequent assignment involving agent $i\in\mathcal{A}$, when $\mathcal{B}_i(t) \neq \emptyset$ agent $i$ must broadcast its banned goal set to all $j\in\mathcal{N}_i$.
The assignment and banning process is iterated by all $j\in\mathcal{N}_i$ until (\ref{eqn:NoConflicts}) is satisfied in the entire neighborhood. The banned and restricted goal information is enforced through a constraint on the assignment problem, which follows.
\begin{problem}[Goal Assignment] \label{prb:assignment}
Each agent assigns itself a goal independently by solving the linear minimum-distance assignment (\ref{eqn:assignment}).
For each agent $i\in\mathcal{A}$, we have
\begin{align}
\underset{a_{jk}\in\mathbf{A}_i}{\text{min}} \Bigg\{
\sum_{k\in\mathcal{N}_i} \sum_{j\in\mathcal{G}} a_{jk} \big|\big|\mathbf{p}_k(t) - \mathbf{p}_j^*(T_k)\big|\big| \Bigg\},\\
\mathbf{p}_k^0 \in \mathcal{N}_i,\, \mathbf{p}_j^*(t) \in \mathcal{G},\nonumber
\end{align}
subject to
\begin{align}
\sum_{j\in\mathcal{G}} a_{jk} &= 1, ~~~ k\in\mathcal{N}_i, \label{eqn:p11} \\
\sum_{k\in\mathcal{N}_i} a_{jk} &\leq 1, ~~~ j\in\mathcal{G},\label{eqn:p12}\\
a_{jk} &= 0, ~~~ k\in\mathcal{N}_i, ~ \mathbf{p}_j\in\mathcal{B}_k,\label{eqn:p13} \\
a_{jk} &\in \{0, 1\}.\nonumber
\end{align}
Each agent independently solves Problem \ref{prb:assignment} as a linear program and selects its assigned goal. This process is repeated by each agent, $i\in\mathcal{A}$, until $|\mathcal{C}_i|=1$.
\end{problem}
As the safety constraints of Problem \ref{prb:assignment} explicitly depend on the neighborhood of agent $i\in\mathcal{A}$, the optimization must be recalculated each time the cardinality of the neighborhood of agent $i$ changes. Under weak assumptions about the trajectories of each agent, the assignments generated by Problem \ref{prb:assignment} is guaranteed to bring each agent to a unique goal as it is shown next.
\begin{lemma} \label{lma:solutionExistance}
For every agent $i\in\mathcal{A}$, if $\big|\big(\bigcup_{j\in\mathcal{N}_i} \mathcal{B}_j\big) \setminus \mathcal{G}\big| \geq |\mathcal{N}_i|$, then the feasible region of Problem \ref{prb:assignment} is always nonempty.
\end{lemma}
\begin{proof}
Let the set of goals available to all agents in the neighborhood of agent $i\in\mathcal{A}$ be denoted by the set
\begin{equation} \label{eqn:feasibleGoals}
\mathcal{V}_i(t) = \{\mathbf{p}\in\mathcal{G} ~|~ \mathbf{p}\not\in\mathcal{B}_j(t), ~ \forall j\in\mathcal{N}_i(t) \}.
\end{equation}
Let the injective function $m_i : \mathcal{N}_i(t) \to \mathcal{V}_i(t)$ map each agent to a goal. As $|\mathcal{N}_i| \leq |\mathcal{V}_i(t)|$, the function $m_i$ must always exist and imposes a mapping from each agent to a unique goal.
Since $m_i$ is injective, it satisfies (\ref{eqn:p11}) and (\ref{eqn:p12}). Likewise, $\mathcal{V}_i\subset\mathcal{B}_j^c$ for all $j\in\mathcal{N}_i$, and therefore the imposes mapping satisfies (\ref{eqn:p13}). Therefore, the mapping imposed by the function $m_i$ is a feasible solution to Problem \ref{prb:assignment}.
%
%
%
\end{proof}
For a sufficiently large value of $T$, the convergence of all agents to goals is guaranteed by Theorem \ref{thm:assignmentConvergence}.
\begin{theorem}[Assignment Convergence] \label{thm:assignmentConvergence}
Under the assumptions of Lemma \ref{lma:solutionExistance}, for a sufficiently large value of $T$, and if the energy-optimal trajectories for each agent always move toward their assigned goal, then
all $i\in\mathcal{A}$ must reach an assigned goal in finite time.
\end{theorem}
\begin{proof}
Let $\{g_n\}_{n\in\mathbb{N}}$ be the sequence of goals assigned to agent $i\in\mathcal{A}$ as designated by the solution of Problem \ref{prb:assignment}. From Lemma \ref{lma:solutionExistance}, $\{g_n\}_{n\in\mathbb{N}}$ is not empty, and the elements of this sequence are integers bounded by $1\leq g_n \leq |\max{\mathcal{F}}|$. Thus, the range of this sequence is compact and must be (1) finite, ($2$) convergent, or ($3$) periodic.
($1$) For a finite sequence, $T_i$ is bounded by $T \cdot |\mathcal{G}|$.
($2$) Under the discrete metric, an infinite convergent sequence requires that there exists $N\in\mathbb{N}_{>0}$ such that $g_n = p$ for all $n>N$ for some formation index $p\in\mathcal{F}$. This reduces to case $1$, as $T_i$ does not increase for repeated assignments to the same goal.
($3$) By the Bolzano-Weierstrass Theorem, an infinite non-convergent sequence $\{g_n\}_{n\in\mathbb{N}}$ must have a convergent subsequence, i.e., agent $i$ is assigned to some subset of goals $\mathcal{I}\subseteq\mathcal{G}$ infinitely many times with some number of intermediate assignments for each goal $\mathbf{g}\in\mathcal{I}$. From the construction of the banned goal set, we must have $\mathcal{I}\bigcap\mathcal{B}_i(t) = \emptyset$ for all $t\in[0,T_i]$. This implies that, by the update method of $T_i$, the position of all goals, $g(t)\in\mathcal{I}$ must only be considered at time $T_i$, which we denote as $\mathbf{g}(T_i)\in\mathcal{I} = \mathbf{g}\in\mathcal{I}$.
This implies that the goals available to agent $i$, i.e., $\mathcal{I} = \mathcal{G}\setminus\mathcal{B}_i$, must be shared between $n>0$ other periodic agents. This implies at some time $t_1$ that a goal, $\mathbf{g}\in\mathcal{I}$, must be an optimal assignment for agent $i$, a non optimal assignment at time $t_2>t_1$ and an optimal assignment at time $t_3>t_2$. This implies the distance between agent $i$ and goal $\mathbf{g}$ satisfy
\begin{align}
|\mathbf{p}_i(t_1) - \mathbf{g}| &< |\mathbf{p}_i(t_1) - \mathbf{g'}|, \\
|\mathbf{p}_i(t_2) - \mathbf{g'}| &< |\mathbf{p}_i(t_2) - \mathbf{g}|, \\
|\mathbf{p}_i(t_3) - \mathbf{g}| &< |\mathbf{p}_i(t_3) - \mathbf{g'}|,
\end{align}
for some goal $g'\in\mathcal{I}, ~ g'\neq g$. Agent $i$ must not increase his distance from his assigned goal, which implies
\begin{align}
|\mathbf{p}_i(t_1) - \mathbf{g}| &> |\mathbf{p}_i(t_2) - \mathbf{g}|, \\
|\mathbf{p}_i(t_2) - \mathbf{g'}| &> |\mathbf{p}_i(t_3) - \mathbf{g'}|,\\
\end{align}
and hence
\begin{align}
|\mathbf{p}_i(t_1) - \mathbf{g'}| & > |\mathbf{p}_i(t_3) - \mathbf{g'}|,
\end{align}
which is satisfied for all goals $g'\in\mathcal{I}$. This is only possible if agent $i$ simultaneously approaches all goals in $\mathcal{I}$, which implies they are arbitrarily close. This contradicts \eqref{eqn:formationSpacing}, and thus no such periodic behavior may exist.
\end{proof}
\subsection{Optimal Trajectory Generation}
After the goal assignment is determined, each agent must generate a collision-free and energy-optimal trajectory to their assigned goal.
The initial and final condition constraints for any agent $i\in\mathcal{A}$ are given by
\begin{align}
\Big(\mathbf{p}_i(t_0) - \mathbf{p}_{i,0},\, \mathbf{v}_i(t_0) - \mathbf{v}_{i,0}\Big) &= \Big(\mathbf{0},\, \mathbf{0}\Big),
\label{eqn:ICs}\\
\Big(\mathbf{p}_i(T_i) - \mathbf{p}_{i,f},\, \mathbf{v}_i(T_i) - \mathbf{v}_{i,f}\Big) &= \Big(\mathbf{0},\, \mathbf{0}\Big),
\label{eqn:BCs}
\end{align}
where
$\mathbf{p}_{i,f}, \mathbf{v}_{i,f}$ are the result of solving Problem \ref{prb:assignment}.
To resolve the coupling introduced by collision avoidance, each agent $i\in\mathcal{A}$ predicts the optimal trajectories of its neighbors, $j\in\mathcal{N}_i$ to select its prescribed trajectory.
\begin{definition}
The \textit{prescribed trajectory}, $u_i^a(t)$, is the trajectory assigned to agent $i$ after solving for the optimal trajectories of every agent in its neighborhood, $j\in\mathcal{N}_i$.
\end{definition}
For agent $i$ to calculate its prescribed trajectory, $u_i^a(t)$, the trajectory optimization problem must be solved over the set
\begin{equation*}
\mathbf{U}_i(t) = \big\{u_j(t) : j\in\mathcal{N}_i\big\},
\end{equation*}
such that
\begin{equation*}
u_i^a(t) = u_i(t) \in \mathbf{U}_i(t).
\end{equation*}
This can be achieved by the quadratic optimization problem given by:
\begin{problem}[Trajectory Generation] \label{prb:trajectory}
For each agent $i\in\mathcal{A}$, we have
\begin{align}
\underset{\mathbf{u}_i\in\mathbf{U}}{\text{min}} \Bigg\{
\sum_{j\in\mathcal{N}_i} \int_{\tau=t}^{T_i} ||\mathbf{u}_j(\tau)|| \text{d}t \Bigg\},
\end{align}
subject to
\begin{align*}
\text{Dynamic constraints }&(\ref{eqn:pDynamics}), (\ref{eqn:vDynamics}), \\
\text{State and control bounds }&(\ref{eqn:vBounds}), (\ref{eqn:uBounds}), \\
\text{Collision avoidance }&(\ref{eqn:collisionCondition}),\\
\text{Boundary constraints }&(\ref{eqn:ICs}), (\ref{eqn:BCs}).\\
\end{align*}
\end{problem}
Problem \ref{prb:trajectory} can be solved as an iterated quadratic program with a similar conflict framework as Problem \ref{prb:assignment}, where one agent fixes its trajectory and others steer to avoid it.
Problems \ref{prb:assignment} and \ref{prb:trajectory} are solved sequentially at time $t=0$ to achieve an initial set of assignments and corresponding optimal trajectories. As both optimizations only use local information, each agent must resolve each problem whenever their neighborhood changes. This ensures that every agent is using all available information to optimize their trajectories while guaranteeing collision avoidance.
\begin{remark}\label{rmk:centralized}
The solutions of Problems \ref{prb:assignment} and \ref{prb:trajectory} reduce to the centralized case as $h \to \infty.$
\end{remark}
Remark \ref{rmk:centralized} relies on the fact that, as $h \to\infty$, it must be true that $\mathcal{N}_i(t_0) = \mathcal{A}$ and $\mathcal{B}_i(t_0) = \emptyset$ $\forall i\in\mathcal{A}$.
Hence, problems \ref{prb:assignment} and \ref{prb:trajectory} simply reduce to each agent solving the centralized problem individually.
\section{Simulation Case Study}\label{sec:simulation}
To give insight into the behavior of the agents a series of simulations were performed in Matlab. Each simulation lasted for $20$ s or until all agents reach their assigned goal, whichever was longer. The centroid of the formation moved with a fixed velocity, while the leftmost and rightmost three goals included additional periodic motion
The minimum separating distance between agents, total energy consumed, and maximum velocity for the unconstrained solutions to Problem \ref{prb:trajectory} are given as a function of the horizon in Table \ref{tab:resultsEnergy}. A graph of each agent's position over time for two cases is given in Fig. \ref{fig:trajectory1} and \ref{fig:trajectory3}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\columnwidth]{rinf-fullTraj.png}
\caption{Agent trajectories for the centralized case, here the assignment globally minimizes distance travelled and trajectories are evenly spaced.}
\label{fig:trajectory1}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.9\columnwidth]{r05-fullTraj.png}
\caption{Agent trajectories for $R=0.5$m, one grid cell in diameter. Agents use very little information and tend to group at the nearest goal position.}
\label{fig:trajectory3}
\end{figure}
\begin{table}[h]
\centering
\begin{tabular}{cccc}
$h$ [m] & min. separation [cm]& $E$ [kJ/kg] & $t_f$ [s] \\\toprule
$ \inf $ & $16.5$ & $63.77$ & $20$\\
$ 1.60 $ & $0.82$ & $83.76$ & $27$\\
$ 1.50 $ & $1.21$ & $56.43$ & $20$\\
$ 1.40 $ & $0.38$ & $140.7$ & $41$\\
$ 1.30 $ & $5.25$ & $52.26$ & $20$\\
$ 1.10 $ & $0.32$ & $96.13$ & $34$\\
$ 0.95 $ & $0.54$ & $41.61$ & $20$\\
$ 0.75 $ & $0.60$ & $227.7$ & $42$ \\
$ 0.50 $ & $2.41$ & $140.1$ & $39$
\end{tabular}
\caption{Numerical results for N=M=10 agents and goals, with a time parameter of $T=10$ s and various sensing distances.}
\label{tab:resultsEnergy}
\end{table}
The results in Table \ref{tab:resultsEnergy} generally show no correlation between energy consumption and sensing horizon. In fact, the minimum energy consumption occurs near $R=1.3$ m rather than the centralized case. This is likely a result of the minimum distance approximation, which does not account for the required change in velocity for a dynamic formation with moving goals.
\section{Conclusion}\label{sec:conclusion}
In this paper, we proposed an approach for solving the desired formation problem of a group of autonomous agents. We presented a formulation of the formation reconfiguration problem and introduced a concept of prescribed goals and trajectories. The robustness and convergence properties of the system were discussed, and the performance was characterized relative to the centralized approach. A numerical solution was presented for $N=M=10$ agents and goals, and the system performance metrics were compared relative to the sensing radius.
Future areas of research include: relaxing the assumptions on Lemma \ref{lma:solutionExistance} to characterize when solutions exist, incorporating information from outside the neighborhood into goal assignment, analyzing the effect on sensing radius on communication cost versus convergence and propulsion energy, reducing the computational load in calculating Problem \ref{prb:trajectory}, and characterizing the optimality of the tiebreaker heuristics.
\bibliographystyle{IEEEtran}
|
\section{INTRODUCTION}
Following people is a highly desirable skills for mobile robots to support daily chores. To achieve robust and efficient person-following capabilities, perception, robot gaze control, and navigation need to be effectively integrated.
Vision-based human recognition has dramatically improved with new softwares that rely on deep learning based technologies such as YoLo \cite{redmon2016you} and
OpenPose\cite{cao2017realtime}. On the flip side, they have a limited range of sight \cite{morales2012people,basso2013fast}. To resolve this problem, laser-based methods \cite{chung2012detection,kawarazaki2015development},
and various sensor fusion techniques combining face recognition and leg detection have been introduced \cite{scheutz2004fast,fritsch2003multi,bellotto2009multisensor}. However, major difficulties include handling occlusions, identifying target people among crowds, and difficulty on effectively detecting human faces \cite{yuan2015multisensor}. To surpass these limitations this, new techniques have been devised relying on extra features such as the detection of clothes, bags, and shoes \cite{ahmed2015improved}.
\begin{figure}[t]
\centering
\includegraphics[width=0.98 \linewidth]{snapshot_final.png}
\protect\caption{Person-following with a mobile robot}
\label{Snapshots}
\end{figure}
Another problem is due to using passive perception techniques where the robot stays stationary thus loosing the target. It is therefore best for robots to achieve active perception such that people can be followed despite their movement \cite{fitzparick2003perception}. We are interested in these questions:
i) where should the robot navigate to? and ii) what should the robot look at? Many researchers have studied this topic withing the topic of active perception or visual sensor planning \cite{chen2011active}. This kind of problem is intractable because there are too many variables. However, using prior knowledge, context, and logical assumptions about the environment it is possible to find solution approximations. If a robot is aware of the connectivity between spaces, when the target suddenly disappears from the view of the robot, one strategy could be to navigate to the last observed location to look for the target. This space connectivity can be simplified by the use of a topology map or graph \cite{portugal2012extracting,thrun2003robotic}.
Robot skills should be integrated in harmony with the perceptual processes to improve a robot\textquoteright s ability to adapt to the various
dynamic circumstances. For example, actions such as searching for a target, tracking, and navigating should be properly coordinated. To achieve this coordination, a Behavior-Tree framework is applied to sequence the skills \cite{colledanchise2017behavior}.
In that light, the main contribution of this paper is on integrating sensor fusion, context-base motion planning, person movement prediction, and behavior decision making. The rest of paper is organized as follows: Section II gives an overview of the methods to track and follow people. Experimental setup and results are describes in section III. In section IV, we draw conclusions and, finally, we present ideas for future improvements.
\section{FRAMEWORK}
\subsection{Behavior-Tree}
A behavior tree (BT) framework \cite{colledanchise2017behavior} allows a robot to achieve autonomous planning in response to various situations. Similarly to (hierarchical) finite state machines, BT's are being utilized because of their modularity, efficiency, and intuitive usage. BT's are considered as generalizations of three classical concepts including Subsumption architecture \cite{brooks1986robust}, Sequential behavior composition \cite{burridge1999sequential}, and Decision trees \cite{nehaniv2002imitation}. BT's contain four types of control flow nodes (fall back, sequence, parallel, and decorator) and two execution nodes (action and condition) \cite{colledanchise2017behavior}. BT's start from a root node toward child nodes with a predefined frequency, and child nodes return one of the values \textit{running}, \textit{success}, or \textit{failure}. Autonomous behaviors of robots can be designed with two fallback nodes and two sequence nodes as shown in Fig.\ref{Figure:BT_framework}.
\begin{figure}[t]
\centering
{\includegraphics[width=1.0\linewidth]{RobotBT.png}}
\caption{Behavior-Tree architecture for person-following. }
\label{Figure:BT_framework}
\end{figure}
\subsection{Map Representation}
To efficiently describe spatial information of the environment, a polygon-based decomposition method can be used. A map consists of sub regions, while each region can be bounded by a polygon. The connectivity information can be obtained from the map. The main functionality of a high-level map representation is to infer the robot's current location and reason about where it is heading to. This information can be used as an important clue for robots to follow or search for target objects. To compute where a robot is headed, a person-following algorithm is described in Algorithm \ref{waypoint}. The main idea is to compare the relative distance between a robot and each neighborhood and find the location for which relative distance has decreased for a certain duration of time.
\begin{algorithm}[t]
\caption{Way-point\_search()} \label{waypoint}
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\begin{algorithmic}
\Require {$x$, $g$, $L$, $h_{\tau-T:\tau}$ (robot, graph, map, history)}
\Ensure {$l^*$ (next way-point)}
\State $l_r\gets Get\_current\_location(x)$
\ForAll{${neighbor}^i\in\mathit{Neighbors(l_r)}$}
\For{$t=\tau-T$ to $\tau$}
\State $d^i \gets dist(x,{neighbor}^i)$
\State $\Delta d^i = d^i_{t+1} -d^i_{t}$
\State $\sum d^i =\sum d^i+ \Delta d^i$
\EndFor
\State $\phi^i = \frac{1}{T} \sum_{\tau-T}^{\tau} \Delta d^i \gets Get\_average\_delta()$
\EndFor
\State $ \Phi = \cup_i\{ \phi^i \} $
\State $l^{*} = l(i) \gets \argmin_{i}\Phi \gets Get\_heading\_way\-point()$
\While {$m$ is not covered}
\State $F_c \gets Extract\_Frontier\_Clusters()$
\ForAll{$f^i\in\mathit{F_c}$}
\State $c^i\gets$ the closest point in $f^i$ to $l^*$
\EndFor
\State $c^{*}\gets \argmax{\bold{U}(c,x,f,l^*)}$
\EndWhile
\Return $c^{*}$
\end{algorithmic}
\end{algorithm}
As for low-level map representations, an occupancy grid mapping scheme, which aims to geometrically represent surroundings with occupancy probabilities, is adopted. This filter uses three types of occupancies including occupied, free, and unknown: $m_k \in \{O(1),F(0),U(0.5)\}$ to characterize a map $m$. From the initial distribution of the occupancy grid which is set to 0.5 for all cells, sequential sensor measurements can update the occupancy grid using inverse sensor models $p_z(m|z)$ for every time step. For the grid within the field of view (FOV), the posterior occupancy probability at time $k+1$, $p_{k+1}(m_{k+1})$ can be obtained via the equation \cite{nuss2016random}:
\begin{eqnarray}
\frac{p_{z_{k+1}}(m_{k+1}|z_{k+1})\cdot p_{k}(m_k)}{p_{z_{k+1}}(m_{k+1}|z_{k+1})\cdot p_{k}(m_k)+p_{z_{k+1}}(\bar{m}_{k+1}|z_{k+1})\cdot p_{k}(\bar{m}_k)}
\end{eqnarray}
where $p(\bar{m}) = 1-p(m)$.
\subsection{Frontier-Based Exploration}
Generally, the purpose of exploration for mobile robots is to cover the environment. This exploration strategy can be linked to an uncertainty of the map, described with Shannon's entropy \cite{crupi2016generalized}. Assuming there exist map boundaries, the entropy can be defined as $H(M_t) = -\sum_{i=1}^{N}m^i_t\log(m^i_t)$, where, $m^i_t$ is the $i$-th entry of the map state from the 2D occupancy grid and $N$ denotes the total number of grids. The entropy of map has higher value when there remains many unknown grids. Using this concept, one possible approach is for a robot to choose the best sensing spot that maximizes information gain of the current map. Regarded as the best solution, frontier-based exploration \cite{yamauchi1997frontier} has been widely used, where a frontier reveals the boundaries between known (occupied or free) and unknown areas. This boundaries are potentially informative because they are close to unknown areas. Geometrically, frontiers can be characterized with following equation \cite{jadidi2014exploration},
\begin{equation}
F = ||\nabla \Lambda||_1 - \beta (||\nabla \Lambda_o||_1+ \Lambda_o-0.5)
\end{equation}
where $\nabla$ is the gradient operator, and $\beta$ is a weight factor for the effect of obstacle boundaries. $||\nabla \Lambda||_1$ defines all boundaries, while the second and third terms meas occupied regions including obstacles and their boundaries. The last constant term, is subtracted to remove the biased probability for unknown region in the occupancy map. the The resulting frontier map $F$ contains only known-free and unknown boundaries and by clustering points in this map we can select goals for further exploration.
Although this approach was originally developed for SLAM, the core concept of frontiers is quite useful for target search. Frontiers can provide a a practical representation for the selection of the observation location for target search. Therefore, this type of information can be used in combination with a priory knowledge or observation models for the target, to achieve synergistic effects.
\subsection{Utility Function}
A utility function can be defined using frontier clusters. Counting the number of unknown cells in a frontier cluster can be regarded as the amount of information gain as the robot observes that area. In other words, the entropy of clusters is regarded as the number of unknown cells. Therefore, the utility function can be modeled with this information gain. Using this idea, the utility function can be written as:
\begin{equation}
U(c,x,l^*)= \alpha N-{d(x,c)+d(c,l^*)},
\end{equation}
where $d ( \cdot )$ is a distance function and $\alpha$ is the weight factor that affects the information gain and the travel cost. As described in Algorithm \ref{waypoint}, the proposed algorithm will find the best sensing spot to look for the target.
\subsection{People Detection and Tracking }
In order to be practical, the real-time detection and tracking of people are essential. In our study, RGB-D camera and laser scanners are combined to detect, identify, and track humans. The use of a muti-modal sensor fusion technique can improve the accuracy of recognition and tracking people.
\begin{figure}[t]
\centering
{\includegraphics[width=0.98\linewidth]{person_detection.png}}
\caption{Person detection and tracking concept}
\label{Figure:tracking_framework}
\end{figure}
\paragraph{Leg Detection and Tracking}
Using a laser range finder, human leg patterns can be recognized. we adopt a random forest classifier as described in \cite{arras2007using}. Each leg position is tracked by an Extended Kalman Filter with a constant velocity assumption. The filter consists of prediction and correction steps. Each filter estimates the state of each leg candidate position and velocity, $x$, $\dot{x}$. The extended Kalman Filter uses a set of linear dynamical systems and a measurement model, $\dot{x}=Ax+Bu+w$, and $z=Hx+v$, respectively. $A$ is a state transition matrix, $B$ is the input matrix, $u$ is input variable, and $w$ is a white Gaussian noise with co-variance $Q$. The measurement variable, $z$, can be modeled with the observation matrix, $H$, and $v$ is the observation noise variable, with co-variance $R$.
In the case of multi-object tracking, data associations are required. In other words, during the update step, the filter needs to select the best observations to update the current objects being tracked. Here, the Nearest Neighbor based data association method \cite{muja2009fast} is adopted to link new candidates ($i$) and objects ($j$). The main idea is finding pairs $(i,j)$ for all observations and the existing target to minimize the total sum of the distances among all the individual assignments.
\paragraph{Human Pose Detection}
Recently, a real-time convolution neural network based algorithm named OpenPose \cite{wei2016cpm} \cite{cao2017realtime} to estimate 2D human poses with a skeleton tracking has been developed and widely used. A major advantage of this algorithm is that it can robustly detect and track multiple people while providing not only a bounding box of the people but also recognizing the human body parts. It is possible to compute the distance to the person from a robot by using the average coordinates of the recognized body parts, and clustering the point clouds of that average. Therefore, 3D bounding boxes of the human can be obtained.
\begin{algorithm}[b]
\caption{Target Identification Strategy}\label{Target Identification Strategy}
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\begin{algorithmic}
\Require {$\mathnormal{B}=\{h_1, h_2, \cdots,h_k\}$ (Current human belief), \\\qquad $k$ : Number of human candidates}
\Ensure{$True$ \ ($H^*$: Human target) or $False$}
\State {\bf Step 1} \textit{Filterwithlegs()}
\State {\bf Step 2} \textit{Filterwithface()}
\State {\bf Step 3} \textit{Filterwithclothes()}
\end{algorithmic}
\end{algorithm}
\paragraph{Person Identification}
The face recognition package \cite{ageitgey2013} is also applied for our proposed system. It basically uses Histograms of Oriented Gradients (HOG) \cite{dalal2005histograms} to detect faces and face landmark estimation \cite{kazemi2014one} to extract face features. Then, the extracted features are used to train a Deep Convolutional Neural Network to recognize faces. Therefore, human faces can be recognized from the image stream, which leads to identifying people. Because the face is the most definite feature that distinguishes a person from another, the highest level of trust is given to the face recognition process. However, using face information has limitations when the robot is following people. For this, our framework is required to have other clues to identify people such as identifying clothes \cite{lee2008inter}. Thus, firstly, the OpenPose recognition tool is used to extract the region of interest for clothe detection. Then, the color or pattern of that bounding box can be characterized using the histogram intersection algorithm \cite{swain1991color}, which is known for its invariance to translations, scaling, and robustness to occlusions with other objects. Given a learned template ($T$), a histogram intersection is defined as $\sum_{j=1}^n min(I_j,T_j)$, where $n$ is the number of bins. To obtain similarity values between input images $I$ and $T$ it uses the normalized ratio given the following formula, $S=\frac{\sum_{j=1}^n min(I_j,T_j)}{\sum_{j=1}^nT_j}$. Using this histogram-based metric, each image extracted from OpenPose, and the corresponding similarity is computed. Then, a person having the similarity exceeding the critical similarity can be regarded as a target.
\paragraph{Human Belief}
We start by combining laser-based detection and vision-based detection using Algorithm \ref{Target Identification Strategy}. It is assumed that vision information is more reliable because lasers provide candidates of human legs, rather than precise information. An important point when fusing is that the FOV of the laser sensor and that of the RGB-D camera are different, so that the robot has to be aware of the position of the people being recognized. Similarly to the low-level map representation discussed earlier, the human belief (the probability that human exists in a grid cell) can be expressed using occupancy grid mapping with a Bayesian inference method. Keeping track of the human belief, the robot considers a human to be present in that region until it observes that region. The grid cell inside a current FOV can be iteratively updated over time with following equation,
\begin{align}
p(h|o) &= \frac{p(o|h)P(o)}{p(o|h)P(h) + P(o|h^c)(1-P(h))}
\end{align}
where $o$ stands for the OpenPose detection results.
\begin{figure*}[t]º
\centering
\includegraphics[width=0.85\linewidth]{svr_result.png}
\protect\caption{Trajectory prediction experiments}
\label{Figure:Prediction_trajectory}
\end{figure*}
\paragraph{Trajectory prediction}
An efficient person-following robot should be able to anticipate where the target might be when it suddenly disappears. The basic idea is to build a regression model from the past history and to extrapolate the person's possible trajectory. A Support Vector Machine Regression (SVR) is adopted to predict the trajectory of people. This can approximate nonlinear relationships, and provide a parsimonious fit, since it relies on kernel functions. Assuming that we have a set of training data where $\boldsymbol{\mathbf{x}}=\left(x_{1},\,x_{2},\,...,\,x_{k}\right)$
is a vector that comprises the input variables ($k$ being the number
of such variables), and that we have $n$ observations: $\left(\mathbf{x}_{1},\,y_{1}\right),\:(\mathbf{x}_{2},\,y_{2}),\,...,\:(\mathbf{x}_{n},\,y_{n})\,,$ where
$y$ is the variable to predict. The problem becomes, according to
the theory of Support Vector Machines (SVM) \cite{Vapnik1998} as that of
finding a function $f(\mathbf{x})=\mathbf{\mathbf{\boldsymbol{\mathbf{\omega}}}}^{T}\phi(\mathbf{x})+b\,,$
and using that function to fit the training data, where: $\boldsymbol{\mathbf{\mathbf{\omega}}}$
is the vector that contains the weights that affect each
predictor; $b$ is a real number; and $\phi$ is a non-linear mapping. SVM theory \cite{Vapnik1999} states that the solution
to this problem is the same as the solution of the equation:
\begin{equation}
\underset{_{\mathbf{\boldsymbol{\omega}},b,\xi}}{\min}\left(-\mathbf{\boldsymbol{\omega}}^{T}\mathbf{\boldsymbol{\omega}}+C\sum_{i=1}^{n}\xi_{i}\right),\label{eq:6}
\end{equation}
subject to $y_{i}\left[\mathbf{\boldsymbol{\omega}}^{T}\phi\left(\mathbf{x}_{i}\right)+b\right]\geq1-\xi_{i}\,\,;$$\:\xi_{i}\geq0\,;$$\,\:\left(i=1,\,2,\,...,\,n\right),$
where: $\xi_{i}$ is the error between observed and predicted values,
that is, $\max\left\{ 0,\,\mid y_{i}-f(\mathbf{x}_{i})\mid-\varepsilon\right\} ;$
the parameter $\varepsilon>0$ determines an insensitivity zone around
the fitted model where the error is not taken into account and, $C$
is the penalty parameter that weighs the error in the function that
is minimized. Thus, the term $C\sum_{i=1}^{n}\xi_{i}$ in equation
\ref{eq:6} represents the losses on the training set.
Once $C$, $\varepsilon$ and the parameters of the kernel function
have been selected, this problem has a unique solution. Different
parameters will give different solutions or models. Therefore, the
parameters must be tuned to optimize the model \cite{Garcia-Gonzalo2016}.
The simplest way of performing the parameter tuning is grid search.
\paragraph{Robot control}
To track the human target, robot gaze control is essential. The gazing behavior is designed to seek for human candidates. For example, when the human target is not visible, a gaze planner forces a robot to look where humans might exist. This information can be obtained from human belief. If the target doesn't exist, the robot will seek for the target using leg candidates. From the geometric relationship between the robot and the target positions, the desired joint values are easily obtained and a PD joint position control is used.
Our navigation strategy is to use the $TURN$ and $GO\_STRAIGHT$ commands and the $MOVE\_BASE$ command from the ROS navigation stack \cite{rosnavigation2012}. While the first command is activated when no obstacles exist between the robot and the target, the second command is activated when there exists obstacles in the desired path. The first command is beneficial: since the laser sensor is the most important observation source, it is effective to track the location of a person in the center of the FOV. Consequently, in order to place a person at the center of the FOV, the first strategy is turning the robot base toward the person and going straight toward the target.
\begin{figure*}[t]
\centering
{\includegraphics[width=1.0\linewidth]{timeline.png}}
\protect\caption{Information timeline during person-following.}
\label{Figure:timeline}
\end{figure*}
\section{RESULTS}
\subsection{System Description}
The Toyota Human Support Robot (HSR) is a mobile manipulator \cite{hashimoto2013field} which has been used as a hardware platform for this study. The mobile base of the HSR consists of two omni-wheels and three caster wheels which are located at the front and rear of the robot. The maximum speed of the HSR is approximately $0.22\nicefrac{m}{s}$, maximum step size of the mobile base is $5mm$ and the maximum incline that it can climb is $5^{\circ}$. As for the vision system information, two stereo cameras are mounted around the eyes of the robot, a wide angle camera is on the forehead, a depth camera (Xtion, Asus) is placed on the top of the head to get RGB-D video stream. Furthermore, a laser range scanner, Hokuyo, is mounted at the front bottom the of mobile base platform. The HSR uses two different computers, the main pc is for most sensing and navigation tasks and an Alienware laptop (Intel Core i7-7820HK, GTX 1080) is used for running OpenPose for human detection. All sub-programs for the robot are able to communicate useful information to each other via ROS interfaces. The tested environment is at UT Austin's Anna Hiss Gymnasium (AHG). A prototype of a home-like arena was built to perform tasks including perception, navigation, manipulation, and more.
\subsection{Results}
\subsubsection{Trajectory prediction}
Human walking is quite unpredictable, so that future positions can only be estimated in a limited time and from the previous positions in a short time margin. For that reason, the last observations are the most significant for trajectory prediction. To address this relative importance in the trajectory estimation, increased weight is given to the last samples. In the case of SVR, the sample weighting re-scales the $C$ parameter, which means that the model puts more emphasis on getting these points right \cite{Han2014}. To compare the experimental results, a three-degree polynomial regression is also implemented.
The parameters of SVR with radial basis function (RBF) kernel were selected while the grid search method and the optimal values that we have used are $C=1000.0$, $\varepsilon=0.01$, and $\gamma=1.0$. The prediction algorithms were tested with three difficult, but common situations regarding person-following: when the target goes through a door, she hides completely from the robot and can turn left, go straight or turn right. The results obtained are shown in figure \ref{Figure:Prediction_trajectory}. In the case of turning right or left, the polynomial regression prediction diverges from the real trajectory, while the SVR prediction gets a good approximation. However, we must take into account the limitations of the prediction algorithm since the extrapolation process is full of uncertainties and can produce meaningless predictions. The divergence with the actual trajectory is a characteristic of extrapolation methods. Thus, estimations might only be considered valid within a limited time range.
\begin{figure}[t]
\centering
\includegraphics[width=0.99\linewidth]{map_trajectory.png}
\protect\caption{Trajectories projected on the house map with map annotations.}
\label{Figure:trajectory}
\end{figure}
\subsubsection{Person-Following Performance}
We address the case of person-following in indoor environments to validate performance of the proposed architecture. The result is shown in Fig. \ref{Figure:timeline} and Fig.\ref{Figure:trajectory}. After learning the target's face and clothing information, robot following is initiated from a starting point. At a certain moment, the target was lost during its way from the kitchen to the office. The first strategy of the robot is to try to predict where person has gone via SVR-based prediction using the input data. Then, the robot decided to go to that location to look for the existence of the person. Since it failed to seek the target using the robot's gaze for that position, the way-point search is activated for further search. Using the Algorithm \ref{waypoint}, the robot navigated to the office location. There, the robot re-identified the target and started to follow her again. The robot status of awareness of target and corresponding actions during the experiment are well described in the second and third rows in Fig.\ref{Figure:timeline}.
Three main achievements should be highlighted. Firstly, the versatility of implementing the person-following architecture with the support of a behavior tree. As it can be seen in the action sequence it confers the robot the ability to perform complex tasks based on reasoning about simple tasks. Secondly, regarding person tracking, Fig. \ref{Figure:timeline} shows that the robot is able to track, follow and successfully re-identify the target in a dynamic environment. Finally, the robustness of the architecture has been achieved by using active search techniques such as the SVR based trajectory prediction and the way-point search that allowed the robot to re-identify the target and complete its task successfully.
\section{CONCLUSIONS}
This paper proposes a novel architecture for a human-following robot that allows robust and efficient tracking in highly dynamic environments. Autonomous behavior planning can successfully coordinate perception and navigation for the robot. Information regarding regions on the map and frontiers from the current observations are used to obtain best observational location. A trajectory prediction algorithm based on Support Vector Machine (SVM) has also been implemented. Finally, we achieved robust autonomous person-following using the HSR.
Although the proposed framework is effective, there are still a lot of situations that can cause the robot to freeze or fail. Similarly to humans, to cope with uncertain situations, reinforcement learning can improve the performance. A probabilistic model can be developed that contributes to solve the uncertainties that arise. On the other hand, regarding following behaviors, the utility function or the cost must be formulated mathematically. What should be the cost function for person-following? These kind of criteria could be learned through exploration of human-robot interactions.
\bibliographystyle{IEEEtran}
|
\section{Introduction}%
\label{sec:intro}
\hl{The wide variety of cells in a multicellular organism show that cells with identical copies of DNA may differentiate in different cell types.
In the late 40's, Max Delbruck at Caltech suggested that each type of cell could correspond to a distinct steady state
in the dynamics of their shared gene expression network.
In order to analyze such large networks, Ren\'e Thomas conjectured in 1980 that the existence of a positive (resp.~negative) feedback loop
was a necessary condition for multistationarity (resp.~sustained oscillations)}~\cite{Thomas81sss}.
\hl{Those conjectures were later proved in various formalisms (Boolean or discrete transition systems, differential equations)
with various degrees of generality.
In 2003, Christophe Soul\'e finally proved Thomas's necessary condition for multistationarity with full generality for dynamical systems defined by differential equations}~\cite{Soule03complexus}.
\hl{In his mathematical formalization of the conjecture, Soul\'e considers a differentiable mapping $F$ from a finite dimensional real vector space to itself,
and for each point $a$, the directed graph $G(a)$ where the arcs are the non-zero entries of the Jacobian matrix of $F$, labeled by their sign.
He shows that if $F$ has at least two non-degenerate zeroes, there exists $a$ such that $G(a)$ has a positive circuit.
}
\hl{When applied to (protein) reaction networks however, Thomas's necessary condition for multistationarity fails short since it is trivially satisfied
as soon as there exists either a bimolecular or a reversible reaction.
}
Indeed,
a bimolecular reaction such as a complexation reaction immediately creates a mutual inhibition between the two reactants, i.e.~a
positive circuit, and a reversible reaction produces a mutual activation, i.e.~again
a positive circuit, making Thomas's necessary condition always true in those networks.
Nevertheless, reaction models are widespread in computational systems biology and it would be very desirable to be able to predict the absence of multistationarity
by systematically checking such conditions with efficient algorithms.
For instance, the BioModels database\footnote{\url{http://biomodels.net/}}~\cite{CLN13issb} is a
repository of more than 600 hand-curated models written in the Systems Biology Markup
Language (SBML)~\cite{Hucka08sbml2} mostly with reaction rules, over several tenths or hundred of molecular species.
There are hundreds more models in the non-curated branch, and thousands of models imported from metabolic networks databases with even larger numbers of reactions and species.
\hl{%
Soul\'e's proof, as most preceding and following proofs, uses the fact that the existence of multiple steady states
implies a non-injectivity property which is shown to be equivalent
to a determinant being zero for some values of reaction rate constants.
One approach, called the Jacobian approach, is thus to use symbolic computation methods to directly compute the roots of that determinant.
If it is non-zero, one can conclude to the absence of multistationarity.
This is the approach taken by Feliu and Wiuf in}~\cite{FW13bi}.
\hl{%
Interestingly, they evaluated their algorithm, implemented in Maple 16, on the curated branch of BioModels (323 networks in their case),
showing that 31,6\% were injective and that only 8,3\% of the networks of this benchmark caused memory overflow by that method.
On the sequences of $r$ phosphorylation cycles of}~\cite{WS08jmb},
\hl{%
they could check non-injectivity up to $r=17$ cycles in 1200 seconds.
In this paper, we follow the alternative graphical approach to multistationarity analyses.
We describe a graph rewriting algorithm which deals with sequences of $r=1000$ phosphorylation cycles in a second,
and analyzes the curated branch of BioModels (506 networks in our case)
with a maximum computation time of 50 milliseconds per network (including large networks of size up to 430 species),
while concluding to the non existence of multiple steady states in 160 networks of size up to 54 species in that benchmark, i.e.~with a similar ratio of 31.6\% of results concluding to non-multistationarity.
This algorithm is based on a refinement of the graphical requirements of Soul\'e}~\cite{Soule03complexus}\hl{%
given by the third author in}~\cite{Soliman13bmb} \hl{as a necessary condition for the existence of multiple steady states in (biochemical) reaction networks.
Similar graphical requirements have also been given in}~\cite{BC10aam}\hl{ without restriction to mass-action law kinetics,
but to our knowledge, it is the first time that they are implemented and evaluated systematically in model repositories.
}
For instance, we are not aware of similar evaluations obtained
with the Chemical Reaction Network Toolbox\footnote{\url{https://crnt.osu.edu/toolbox-history-and-explanation}} for \hl{systematically} checking the
graphical conditions for multistationarity of Feinberg's Chemical
Reaction Network Theory (CRNT)~\cite{Feinberg77crt,CF06siamjam}.
\hl{More specifically, we present a series of graph rewriting algorithms for checking the different graphical requirements} of~\cite{Soliman13bmb},
and \hl{analyze their practical performance} in the curated models of BioModels, in order to:
\begin{itemize}
\item evaluate when the original condition of Thomas allows \hl{us} to rule out
multistationarity;
\item evaluate when the following three extra conditions given in~\cite{Soliman13bmb} become conclusive, namely:
\begin{enumerate}
\item the positive circuit must not come from twice the same reaction;
\item the positive circuit must not come from a reaction and its
reverse reaction;
\item the positive circuit must not involve all species of a
conservation law;
\end{enumerate}
\item evaluate when even stronger conditions based on the
rewirings detailed in~\cite{Soule03complexus,Soliman13bmb} are necessary to conclude, namely
\begin{enumerate}
\item by sign change of incoming arcs on a set of species,
\item or by permuting the arcs to a set of target species.
\end{enumerate}
\end{itemize}
For this study, we used our software modelling environment BIOCHAM\footnote{\url{http://lifeware.inria.fr/biocham4}}~\cite{biocham4,cfs06bi}
to load all models from the curated branch of BioModels,
improve their writing in SBML with well-formed reactions using the algorithm described in~\cite{FGS15tcs},
compute the conservation laws~\cite{Soliman12amb},
compute their influence multigraph labelled by the reactions~\cite{FMRS18tcbb,FS08fmsb}
and export the labelled multigraph in the Lemon library format\footnote{\url{http://lemon.cs.elte.hu/}}.
Then we used an implementation in C++
of the algorithm presented in this paper to search for positive circuits
with the different refined conditions on the labelled influence multigraph,
and evaluate their respective contributions for the analysis of multistationarity in BioModels.
\hl{All the computation times obtained with this algorithm given in this paper were obtained on a standalone desktop Linux machine with an Intel Xeon 3.6 GHz processor\footnote{For the sake of reproducibility, our programs and data are available at \url{https://lifeware.inria.fr/wiki/Main/Software\#JTB18}}.}
The rest of this article is organized as follows. The next section presents the refined necessary conditions for multistationarity in reaction networks
\hl{described} in~\cite{Soliman13bmb} and detailed here with five levels of conditions.
The following section presents a graph rewriting algorithm for checking those conditions, and evaluates its computational complexity.
Section~\ref{sec:biomodels} \hl{shows the remarkable performance of this algorithm by applying it} systematically to the curated part of the model repository BioModels,
including models out of reach of Jacobian-based symbolic computation methods,
\hl{and details the effect of the five levels of refined conditions in this benchmark}.
Section~\ref{futile} considers the models of double phosphorylation cycles of Wang and Sontag \cite{WS08jmb} and
shows a very low quadratic empirical complexity of the graphical algorithm, again in sharp contrast to
symbolic computation methods.
Section~\ref{sec:erk} focusses on model 270 of ERK signalling that contains 33 species and 42 reactions resulting in an influence multigraph of 126 arcs
with many positive and negative feedback loops, yet for which our graphical algorithm demonstrates the absence of multistationarity.
\hl{These examples illustrate} the importance of \hl{modelling} the intermediate complexes in enzymatic reactions
\hl{to obtain multiple steady states}, and show the sensitivity of both the dynamical properties of the models and of our graphical conditions to
the writing of enzymatic reactions with or without intermediate complexes.
We conclude on the \hl{remarkable} performance of the graphical approach to analyze multistationarity in reaction models of large size,
and on some perspectives to further improve our algorithm and generalize this approach.
\section{Necessary Condition for Multistationarity in Reaction Networks}
Let us consider a biochemical reaction system with $n$ species
$S_1,\dots,S_n$ and $m$ reactions $R_1,\dots,R_m$. Using notations from~\cite{Kaltenbach12arxiv} we write:
\[R_j = \sum_{i=1}^n y_{ij}S_i \longrightarrow \sum_{i=1}^n y'_{ij}S_i\]
The $y$ and $y'$ represent the stoichiometric coefficients of the reactants
and products of the reaction.
The rate law associated with reaction $R_j$ will be written $v_j$. This
defines a dynamical system, in the form of an Ordinary Differential Equation
(ODE): $\dot x = F(x)$
where $x_i$ is the concentration of species $S_i$ and
\[f_i(x) = \sum_j v_j(x)\cdot (y'_{ij} - y_{ij})\]
This kind of reaction-based system encompasses most of the systems biology
models developed nowadays and made available in model repositories like BioModels.
In particular, SBML reaction models can be translated with our notations,
basically by splitting reversible reactions into forward and backward
reactions, and by including modifiers on both sides of the reaction.
Reaction systems are often graphically represented as a Petri-net, i.e., a bipartite
graph for species and reactions~\cite{Ivanova79kk,IT79kk}. Using the same
bipartite vertices but different arcs and labels, it is possible to represent
the Directed Species-Reaction (DSR) graph of Kaltenchbach~\cite{Kaltenbach12arxiv}.
\hl{This graph is a variant of the DSR graph} of~\cite{BC09cms,BC10aam} \hl{with different labels and no sign}.
\hl{Here the arcs of the DSR graph are defined and identified by their label $\lambda$} as follows: \[ \lambda(S_i, R_j) =
\frac{\partial v_j}{\partial x_i}\qquad \lambda(R_j, S_i) = y'_{ij} - y_{ij}
\]
If $\lambda$ is zero, then there is no arc. $\lambda$ is extended to
paths (resp.\ subgraphs) as the product of the labels of all arcs in the path
(resp.\ subgraph). For a path $P$, we shall write $\lambda_{SR}(P)$ (resp.\
$\lambda_{RS}(P)$) for the product of labels considering only species to
reaction (resp.\ reaction to species) arcs.
Intuitively, $\lambda_{SR}$ represent\hl{s} the contribution of species to each
reaction rate, whereas $\lambda_{RS}$ describe\hl{s} the stoichiometric effect
of reactions on each species.
Fig.~\ref{fig:dsr} shows the DSR graph for the chemical reaction network
corresponding to the enzymatic reaction $S + E \rightleftarrows ES
\longrightarrow E + P$.
\begin{figure}[htb]
\begin{center}
\begin{tikzpicture}
\matrix[row sep=9mm,column sep=14mm] {%
& & \node[place] (E) {E}; & & &\\
& & \node[transition] (r1) {$R_1$}; & & &\\
& \node[place] (S) {S}; & & \node[place] (ES) {ES};
& \node[transition] (r3) {$R_2$}; & \node[place] (P) {P};\\
& & \node[transition] (r2) {$R_{-1}$}; & & &\\
};
\draw[pre] (r1) to[bend right] node[right] {$\frac{\partial v_1}{\partial E}$} (E);
\draw[post] (r1) to[bend left] node[left] {-1} (E);
\draw[pre] (r1) to[bend right] node[left] {$\frac{\partial v_1}{\partial S}$} (S);
\draw[post] (r1) to[bend left] node[right] {-1} (S);
\draw[post] (r1) -- node[above] {1} (ES);
\draw[pre] (r2) to[bend right] node[right] {$\frac{\partial
v_{-1}}{\partial ES}$} (ES);
\draw[post] (r2) to[bend left] node[left] {-1} (ES);
\draw[post] (r2) -- node[above] {1} (S);
\draw[pre] (r3) to[bend right] node[above] {$\frac{\partial
v_2}{\partial ES}$} (ES);
\draw[post] (r3) to[bend left] node[below] {-1} (ES);
\draw[post] (r3) -- node[above] {1} (P);
\draw[round, post] (r2) -- node[above] {1} ++(-35mm,0) |- (E);
\draw[round, post] (r3) -- node[right] {1} (r3 |- E) -- (E);
\end{tikzpicture}
\end{center}
\caption{DSR graph of the enzymatic reaction:
$S + E \rightleftarrows ES \longrightarrow E + P$.}\label{fig:dsr}
\end{figure}
\begin{definition}\cite{Kaltenbach12arxiv}
A \emph{species hamiltonian hooping} of the DSR graph is a collection of
cycles covering each of the species nodes exactly once.
\end{definition}
The set of all species hamiltonian hoopings will be denoted by $\H$.
Thanks to the fact that $\lambda(H) = \lambda_{SR}(H) \lambda_{RS}(H)$,
Kaltenbach~\cite{Kaltenbach12arxiv} proposed to group all species hamiltonian
hoopings having the same species-to-reaction arcs using an equivalence
relation noted $\sim$, writing $[H] = \{H'\in\H\mid H'\sim H\}$ for the equivalence class of a hooping
$H$, and
$\H/\sim$ for the quotient set.
\begin{theorem}[\cite{Kaltenbach12arxiv}]\label{thm:kaltenbach}
\[\det(J) = \sum_{[H] \in \H/\sim}\Lambda([H]) \lambda_{SR}(H)
\quad\text{ with }\quad
\Lambda([H]) = \sum_{H' \in [H]}\sigma(H') \lambda_{RS}(H')\]
\end{theorem}
Considering Soulé's proof of Thomas's conjecture for dynamical
systems~\cite{Soule03complexus} and applying Thm.~\ref{thm:kaltenbach} to each
sub-DSR-graph corresponding to a principal minor of $-J$,~\cite{Soliman13bmb}
notes that a necessary condition for multistationarity is that some term of
the sum is negative. This again states the usual condition about the existence
of a positive cycle in the influence graph of $J$.
\hl{Now, the usual labelling of the arcs of the influence graph between molecular species by the sign of the Jacobian matrix coefficient
can be augmented to
contain not only the sign but also the reaction used for each arc. There is
thus an arc in this reaction-labelled influence multigraph for each
species-to-species path of length two in the DSR graph. This leads to a
one-to-one correspondence between hamiltonian hoopings of the
reaction-labelled influence multigraph and species hamiltonian hoopings of the DSR
graph. Fig.~}\ref{fig:labelled} \hl{illustrates this on the example of
Fig.~}\ref{fig:dsr}.
\begin{figure}[htb]
\begin{center}
\begin{tikzpicture}
\matrix[row sep=3cm,column sep=2cm] {%
& \node[place] (E) {E}; & & &\\
\node[place] (S) {S}; & & \node[place] (ES) {ES};
& & \node[place] (P) {P};\\
};
\draw[posreg,bend left] (E) to node[auto,black] {$R_1$} (ES);
\draw[posreg] (S) to node[auto,black] {$R_1$} (ES);
\draw[posreg,bend left] (ES) to node[auto,black] {$R_{-1}$} (E);
\draw[posreg,bend left] (ES) to node[auto,black] {$R_{-1}$} (S);
\draw[posreg] (ES) to node[auto,black] {$R_2$} (E);
\draw[posreg] (ES) to node[auto,black] {$R_2$} (P);
\draw[negreg,bend left] (E) to node[auto,black] {$R_1$} (S);
\draw[negreg,bend left] (S) to node[auto,black] {$R_1$} (E);
\end{tikzpicture}
\end{center}
\caption{Influence \hl{multigraph} associated to the Michaelis-Menten reaction
system of the three reactions $S + E \rightleftarrows ES \longrightarrow E + P$.
The influence arcs are
labelled both by their sign, as usual and by the unique reaction from which they originate.
Note for instance that there are two positive
arcs from \lstinline|ES| to \lstinline|E|. Negative self-loops are omitted for
clarity}\label{fig:labelled}
\end{figure}
\hl{Let us denote by $|_H$ the} restriction of the reaction system to a species hooping $H$,
i.e.~the
system where reactions $\{R_i \mid i \in I\}$ not appearing in $H$ are
omitted.
\begin{theorem}[\cite{Soliman13bmb}]\label{thm:fullrank}
Let $F$ be any differentiable map from $\Omega$ to $\mathbb{R}^n$
corresponding to a biochemical reaction system. If $\Omega$ is open and $F$
has two nondegenerate zeroes in $\Omega$ then there exists some $a$ in
$\Omega$ such that:
\begin{enumerate}
\item The reaction-labelled influence graph $G$ of $F$ at point $a$
contains a positive circuit $C$;
\item There exists a hooping $H$ in $G$, such that $C$ is subcycle of
$H$ with $(Y' - Y)|_H$ of full rank.
\end{enumerate}
\end{theorem}
\hl{This theorem gives several graphical requirements for multistationarity.}
\begin{corollary}\label{cor:once}
A necessary condition for the multistationarity of a biochemical reaction
system is that there exists a positive cycle in its influence \hl{multigraph}, using
\emph{at most once each reaction}.
\end{corollary}
This condition actually only requires the reaction-labelled influence \hl{multigraph}.
It is immediate to check that the mutual inhibition resulting from bimolecular
reactions---like that between \lstinline|E| and \lstinline|S| in our running
example---cannot fulfill these necessary conditions, since the
same reaction---$R_1$ in Fig.~\ref{fig:labelled}---will be repeated twice.
\begin{corollary}\label{cor:reversible}
A necessary condition for the multistationarity of a biochemical reaction
system is that there exists a positive cycle in its influence \hl{multigraph}, not
using \emph{both forward and backward directions} of any reversible
reaction.
\end{corollary}
The mutual activation resulting from reversible
reactions---like that between \lstinline|ES| and \lstinline|S| through $R_1$
and $R_{-1}$ in our running example---cannot thus fulfill these necessary conditions.
That condition is also a corollary of the conditions given in \cite{BC10aam}
since, in their setting, reversible reactions give rise to a unique undirected edge.
Another information that can be extracted from the stoichiometry is the
(structural) conservation laws, i.e., P-invariants of the underlying Petri
net, or more simply the left kernel of the stoichiometry matrix. Finding all
the conservation laws of a biochemical model might be computationally
expensive, though in practice that does not seem to be the
case~\cite{Soliman12amb}.
\begin{corollary}\label{cor:pinv}
A necessary condition for the multistationarity of a biochemical reaction
system is that there exists a positive cycle in its influence \hl{multigraph}, not
using \emph{all species involved in a conservation law}.
\end{corollary}
In our running example,
\hl{the species }\lstinline|E| and \lstinline|ES|\hl{, mutually activated through $R_1$
and $R_{2}$, form a conservation law,}
which violates the necessary conditions for multistationarity.
The three Corollaries~\ref{cor:once},~\ref{cor:reversible} and~\ref{cor:pinv}
thus rule out all the cases for which Thomas's condition was
satisfied in this example.
Furthermore, in~\cite{Soule03complexus} and later in~\cite{Soliman13bmb}, the two following
corollaries relying on some particular graph rewirings were also mentionned without much clue to check them:
\begin{corollary}\label{cor:inversed}
A necessary condition for the multistationarity of a biochemical reaction
system is that there exist positive cycles fulfilling condition 2 of
Theorem~\ref{thm:fullrank} in the influence \hl{multigraph} corresponding to its
Jacobian, \textbf{and} in any graph obtained from it choosing a set of
species and by reversing the sign of all arcs that have as target some
species belonging to that set.
\end{corollary}
\begin{corollary}\label{cor:permuted}
A necessary condition for the multistationarity of a biochemical reaction
system is that there exist positive cycles fulfilling condition 2 of
Theorem~\ref{thm:fullrank} in the influence \hl{multigraph} corresponding to its
Jacobian, \textbf{and} in any graph obtained from it by choosing a permutation
of the species and by rewiring the arcs' target according to the
permutation.
\end{corollary}
\section{Graph-Theoretic Algorithm for Proving Non-Multistationarity}\label{sec:alg}
\subsection{{Computing the Labelled Influence Multigraph of a Reaction Model written in SBML}}
The signs of the arcs in the reaction-labelled influence
multigraph of a reaction system, are given by the sign of $\partial v_i/\partial x_j$ instead of that
of $\partial f_i/\partial x_j$. Even without precise kinetic values,
this can be easily computed under the general condition of well-formedness of the reactions~\cite{FMRS18tcbb,FS08tcs}.
This condition is satisfied by the commonly used kinetics such as
mass action law, Michaelis-Menten and
Hill kinetics, \hl{and provides a sanity check for the writing in SBML of ODE models~}\cite{FGS15tcs}.
In the following, and to ensure in a simple and systematic way
that the structure of the reactions, and of the computed influence \hl{multigraph}, do correspond to the continuous dynamics of the model,
all SBML models considered here are first automatically \emph{sanitized} as explained
in~\cite{FGS15tcs}, by exporting the system of ordinary
differential equations, and reimporting it as a well-formed reaction system.
Algorithm~\ref{alg:sanitize} summarizes the main steps of this procedure.
\begin{algorithm}[htb]
\begin{algorithmic}
\Function{extract\_labelled\_influence\_graph}{$sbmlModel$}
\State $Model\gets$\Call{load\_sbml\_model}{$sbmlModel$}
\State $System\gets$\Call{compute\_odes}{$Model$}
\State $Model\gets$\Call{infer\_reaction\_model\_from\_odes}{$System$}\\
\Comment{as explained in~\cite{FGS15tcs}}
\State $Graph\gets$\Call{infer\_influence\_graph}{$Model$}\\
\Comment{as explained in~\cite{FS08fmsb} but adding reactions as labels on the edges}
\State \Return $Graph$
\EndFunction
\end{algorithmic}
\caption{Algorithm for computing the labelled influence multigraph of a reaction model \hl{written} in SBML~\cite{FGS15tcs}.}\label{alg:sanitize}
\end{algorithm}
\hl{This algorithm needs to determine the sign of a partial derivative.
In our implementation this is done by a simple symbolic derivation algorithm and a heuristic to determine the sign of the expressions.
In case of indeterminacy, both signs are assumed.
In general, the result that is computed is thus an over-approximation of the real influence multigraph.}
\subsection{Absence of Positive Circuit with the Conditions of Cor.~\ref{cor:once}~\ref{cor:reversible}~\ref{cor:pinv}}
Tarjan's \hl{depth-first tree traversal of a graph} provides a classical algorithm for testing the existence of a circuit,
\hl{by just checking the existence of a back edge during this traversal~}\cite{Tarjan72siam}.
\hl{Generalizing this algorithm to check the absence of circuits
satisfying the previous conditions on the signs and on the reactions at the origin of the arcs
is however non obvious. This may explain why the previous refined graphical requirements had not been implemented before.
In this section, we present an algorithm which proceeds by graph rewriting.
This algorithm will generalize the following graph simplification rules which show that a graph is acyclic
if and only if it reduces to the empty graph by using them}~\cite{ll88joa,pqr99jco,FL06cor}:
\begin{itemize}
\item \textbf{IN0($v$)}: Remove vertex $v$ and all associated edges if $v$
has no incoming edge.
\item \textbf{OUT0($v$)}: Remove vertex $v$ and all associated edges if $v$
has no outgoing edge.
\item \textbf{IN1($v$)}: Remove vertex $v$ if $v$ has exactly one incoming
edge and connect this edge to all the outgoing edges of $v$.
\item \textbf{OUT1($v$)}: Remove vertex $v$ if $v$ has exactly one outgoing
edge and connect all incoming edges to it.
\end{itemize}
In order to check the conditions of Cor.~\ref{cor:once}~\ref{cor:reversible}~\ref{cor:pinv},
we consider here labelled \hl{multigraph}s, where each arc is labelled by a couple: its sign
and the reaction from which it originates.
Instead of stopping when the previous rules do not apply, and conclude to the cyclicity of the graph if it is not empty,
we extend this set of rules to any number of incoming or outgoing edges and add a restriction on the created edges that must satisfy the conditions of the corollaries.
\hl{This is described by the following single graph rewriting rule which subsumes} the four previous ones:
\begin{itemize}
\item \textbf{INOUTi($v$)}: Remove vertex $v$ if $v$ has exactly $i$ incoming or outgoing
edges, and create the incoming-outcoming edges labeled by the product of the signs and the union of the reactions
if, and only if, those labels satisfy the conditions of the corollaries.
\end{itemize}
This \hl{generic rewriting rule removes} one vertex and all the attached edges
and creates a new edge for every pair of \textit{incoming-outgoing} edge of
the vertex satisfying the conditions. When creating such arcs, the reactions
and species involved in the process \hl{are memorized in order} to check the conditions given by
the previous corollaries and to eliminate the edges, now representing paths,
that do not respect them.
\hl{In this way, this rewriting rule} preserves
all circuits satisfying the conditions of the corollaries.
This rule is applied successively to the nodes of the graph \hl{by choosing a vertex of minimum degree $i$ at each step.
This is done with a simple data structure that maintains the degree of each vertex.}
This algorithm terminates when the first positive self-loop is found, denoting
that a positive circuit satisfying the conditions of the three
Corollaries~\ref{cor:once},~\ref{cor:reversible} and~\ref{cor:pinv}
has been found in the original graph, or when the graph is empty, proving that
no such circuit exists.
The main steps of the decision procedure are summarized in Alg.~\ref{alg:multistat} and~\ref{alg:red}.
\begin{algorithm}[htb]
\begin{algorithmic}[5]
\Function{CheckAcyclicity}{$G$}
\While{\Call{CountVertices}{$G$} $> 0$}
\State $v\gets$ vertex with the least number of incoming or outgoing arcs
\State \Call{RemoveVertex}{$v$}
\For{$L \in \mathit{SelfLoops}$}\label{algline:forloop}
\If{$L$ is positive}
\State \Return False
\Else
\State Delete $L$
\EndIf
\EndFor
\EndWhile
\State \Return True
\EndFunction
\end{algorithmic}
\caption{Acyclicity check.}\label{alg:multistat}
\end{algorithm}
\begin{algorithm}[htb]
\begin{algorithmic}
\Procedure{RemoveVertex}{$v$}
\For{$w\longrightarrow_l v \in IncomingEdges(v)$}
\For{$v\longrightarrow_m x \in OutgoingEdges(v)$}
\State Create a new label $n\gets l\cdot v\cdot m$
\If{$n$ does not contain twice the same reaction, a reaction and
its inverse or all species of a conservation}
\State Create $w\longrightarrow_n x$
\Else
\State Discard $n$
\EndIf
\State Delete $v\longrightarrow_m x$
\EndFor
\State Delete $w\longrightarrow_l v$
\EndFor
\State Delete $v$
\EndProcedure
\end{algorithmic}
\caption{Graph reduction preserving acyclicity.}\label{alg:red}
\end{algorithm}
\begin{proposition}\label{prop:complexity}
The time complexity of Alg.~\ref{alg:multistat} is $\mathcal{O}\left(k^{2^n}\right)$ where $n$ the number of nodes
and $k$ is the maximum degree of the graph.
\end{proposition}
\begin{proof}
Let us write $k_i$ the maximum indegree or outdegree of the graph after the $i^{\text{th}}$ loop of Alg.~\ref{alg:multistat} ($k_0 = k$). The call to remove a vertex (Alg~\ref{alg:red}) is done in at most $k_i^2$ steps and creates at most the same number of edges, we then have the relation: $k_{i+1} = k_i^2$ which gives: $k_i = k^{2^i}$.
Alg.~\ref{alg:multistat} goes through at most $n$ loops, therefore, the number of steps to complete the algorithm is at most given by:
\[C(k,n) = \sum_{i=1}^{n}k^{2^i} \leqslant \sum_{j=1}^{2^{n-1}} k^{2j} = \mathcal{O} \left(k^{2^n}\right)\]
\end{proof}
\hl{We do not know whether this doubly exponential complexity can be reached in some networks.
It is worth noting that this bound does not take into account the fact that the number of edges strictly decreases
when the rule INOUTi is applied to vertices of degree $i\le 1$},
nor that the edges that do not satisfy the conditions of the corollaries are not created.
Furthermore the degree of the nodes in the initial graph is also a limiting factor as it is generally low \hl{in the context of biochemical networks~}\cite{NMFS16constraints}.
\hl{These considerations explain the much better practical complexity reported in Sections~}\ref{sec:biomodels}, and~\ref{futile}\hl{,
where we will show for instance that the time taken to analyze one model of BioModels is
empirically $\mathcal{O}(e\log(n))$ where $e$ is the number of edges.}
\subsection{Sign Changes}
\hl{We show here that} the condition given by Corollary~\ref{cor:inversed} can be done by solving a linear system in
\hl{Galois field} $GF(2)$, \hl{i.e.~$\mathbb Z/2\mathbb Z$}, in which each species is a variable (valued to 1 if the sign of the incoming arcs needs to be reversed). Each simple loop satisfying Corollaries~\ref{cor:once},~\ref{cor:reversible} and~\ref{cor:pinv} in the graph is then modelled by an equation on the sum of all the species involved in the loop, equal to 0 if the loop is negative and 1 otherwise.
As an example, let us consider the influence graph shown in Fig.~\ref{fig:mapk}. This graph contains two positive circuits which satisfy the three Corolaries~\ref{cor:once},~\ref{cor:reversible} and~\ref{cor:pinv} ($ K \xrightarrow[]{R_1} MK \xrightarrow[]{R_2} K $ and $ K \xrightarrow[]{R_3} MpK \xrightarrow[]{R_4} K $) and only one negative circuit satisfying the same Corollaries ($ K \xrightarrow[]{R_1} MK \xrightarrow[]{R_2} Mp \xrightarrow[]{R_3} K $). The system associated to these three loops is therefore:
\[
\left\lbrace\begin{array}{cccccccc}
x_K &+& x_{MK} & & & & & = 1 \\
x_K & & &+& x_{MpK} & & & = 1 \\
x_K &+& x_{MK} & & &+& x_{Mp} & = 0
\end{array}\right.
\]
If this system has a solution, then the reaction graph for which we reverse the sign of every arc that has as target any species which variable evaluates to 1 in the solution, does not contain any positive loop satisfying the three Corollaries~\ref{cor:once},~\ref{cor:reversible} and~\ref{cor:pinv}.
\begin{figure}[htb]
\begin{center}
\begin{tikzpicture}
\matrix[row sep=1.5cm,column sep=1.5cm] {%
& & \node[place] (K) {K}; & & & \\
\node[place] (M) {M}; & & & & \node[place] (MpK) {MpK}; & \node[place] (Mpp) {Mpp}; \\
\\
& \node[place] (MK) {MK}; & & \node[place] (Mp) {Mp}; & & \\
};
\draw[bend right=20,negreg] (K) to node[auto,black,above] {$R_1$} (M);
\draw[bend right=20,negreg] (M) to node[auto,black] {$R_1$} (K);
\draw[bend left=10,negreg] (Mp) to node[auto,black] {$R_3$} (K);
\draw[bend left=10,negreg] (K) to node[auto,black] {$R_3$} (Mp);
\draw[bend left=10,posreg] (M) to node[auto,black] {$R_1$} (MK);
\draw[bend left=10,posreg] (MK) to node[auto,black] {$R_{-1}$} (M);
\draw[posreg] (MK) to node[auto,black] {$R_{2}$} (K);
\draw[bend left=20,posreg] (MK) to node[auto,black] {$R_{-1}$} (K);
\draw[bend left=20,posreg] (K) to node[auto,black] {$R_{1}$} (MK);
\draw[posreg] (MK) to node[auto,black] {$R_{2}$} (Mp);
\draw[bend left=10,posreg] (Mp) to node[auto,black] {$R_{3}$} (MpK);
\draw[bend left=10,posreg] (MpK) to node[auto,black] {$R_{-3}$} (Mp);
\draw[posreg] (MpK) to node[auto,black,above] {$R_{-3}$} (K);
\draw[bend right=20,posreg] (MpK) to node[auto,black,above] {$R_{4}$} (K);
\draw[bend right=20,posreg] (K) to node[auto,black,below] {$R_{3}$} (MpK);
\draw[posreg] (MpK) to node[auto,black] {$R_{4}$} (Mpp);
\end{tikzpicture}
\end{center}
\caption{influence multigraph of $M + K \rightleftarrows MK \longrightarrow K + Mp
\rightleftarrows MpK \longrightarrow K + Mpp$. Negative self-loops are omitted for clarity}\label{fig:mapk}
\end{figure}
Solving such system can be done by using a simple Gaussian elimination. This process is applied every time a new loop is found by adding the corresponding equation in the system and checking the new equation does not yield a contradiction (which can only be the equation $0 = 1$). This process allows us not to compute every possible loop in the graph if a contradiction emerges.
The previous system obtained from Fig.~\ref{fig:mapk} has two solutions, one of which is $x_K = x_{Mp} = 1$ and the over variables are put to zero. Therefore, the influence graph for which the sign of the incoming arcs for nodes $K$ and $Mp$ are reversed (Fig.~\ref{fig:mapk_sign}) does not contain any positive circuit satisfying Corollaries~\ref{cor:once},~\ref{cor:reversible} and~\ref{cor:pinv}. The biochemical system cannot display any multistationarity.
\begin{figure}[htb]
\begin{center}
\begin{tikzpicture}
\matrix[row sep=1.5cm,column sep=1.5cm] {%
& & \node[place] (K) {K}; & & & \\
\node[place] (M) {M}; & & & & \node[place] (MpK) {MpK}; & \node[place] (Mpp) {Mpp}; \\
\\
& \node[place] (MK) {MK}; & & \node[place] (Mp) {Mp}; & & \\
};
\draw[bend right=20,negreg] (K) to node[auto,black,above] {$R_1$} (M);
\draw[bend right=20,posreg] (M) to node[auto,black] {$R_1$} (K);
\draw[bend left=10,posreg] (Mp) to node[auto,black] {$R_3$} (K);
\draw[bend left=10,posreg] (K) to node[auto,black] {$R_3$} (Mp);
\draw[bend left=10,posreg] (M) to node[auto,black] {$R_1$} (MK);
\draw[bend left=10,posreg] (MK) to node[auto,black] {$R_{-1}$} (M);
\draw[negreg] (MK) to node[auto,black] {$R_{2}$} (K);
\draw[bend left=20,negreg] (MK) to node[auto,black] {$R_{-1}$} (K);
\draw[bend left=20,posreg] (K) to node[auto,black] {$R_{1}$} (MK);
\draw[negreg] (MK) to node[auto,black] {$R_{2}$} (Mp);
\draw[bend left=10,posreg] (Mp) to node[auto,black] {$R_{3}$} (MpK);
\draw[bend left=10,negreg] (MpK) to node[auto,black] {$R_{-3}$} (Mp);
\draw[negreg] (MpK) to node[auto,black,above] {$R_{-3}$} (K);
\draw[bend right=20,negreg] (MpK) to node[auto,black,above] {$R_{4}$} (K);
\draw[bend right=20,posreg] (K) to node[auto,black,below] {$R_{3}$} (MpK);
\draw[posreg] (MpK) to node[auto,black] {$R_{4}$} (Mpp);
\end{tikzpicture}
\end{center}
\caption{influence multigraph of $M + K \rightleftarrows MK \longrightarrow K + Mp
\rightleftarrows MpK \longrightarrow K + Mpp$ for which arcs ending in $K$ and $Mp$ have changed sign. Self-loops are omitted for clarity}\label{fig:mapk_sign}
\end{figure}
Gaussian elimination can be done directly on the species involved in the loop by noticing that adding two equations of the system in $GF(2)$ corresponds to taking the symmetric difference between the list of species of the two loops involved and changing the sign accordingly. This last process is described in Alg.~\ref{alg:gauss}. This function is to be
incorporated in Alg.~\ref{alg:multistat} by replacing the body of the
\textbf{for} loop starting on line~\ref{algline:forloop}, stopping on False
and continuing on True.
\begin{algorithm}[htb]
\begin{algorithmic}
\Function{AddToLoopSystem}{$l$}
\For{$m \in LoopSystem$}
\State Let $P \gets Pivot(m)$
\If{$P \in l$}
\State $species(l) \gets species(l)\ \Delta\ species(m)$
\State $Sign(l) \gets Sign(l) + Sign(m)$
\EndIf
\EndFor
\If{$l$ is positive and $Species(l) = \emptyset$}
\State \Return False
\Else
\State Add $l$ to $LoopSystem$
\State $Pivot(l) \gets \mathit{FirstSpecies}(l)$
\State \Return True
\EndIf
\EndFunction
\end{algorithmic}
\caption{Gaussian Elimination}\label{alg:gauss}
\end{algorithm}
\subsection{Permutations}
Checking the condition given by Corollary~\ref{cor:permuted} requires more computation as we do not know beforehand the effects of applying a permutation on the circuits of the graph.
As an example, if we apply a swapping between $K$ and $MpK$ in the previous case of Fig.~\ref{fig:mapk}, i.e.~changing the target of every edge that points to $K$ to $MpK$ (including self-loops) and vice-versa, we obtain the graph shown in Fig~\ref{fig:mapk_swap} for which there are 4 positive loops ($K \xrightarrow[]{R_1} MpK \xrightarrow[]{R_4} K$, $K \xrightarrow[]{R_1} MpK \xrightarrow[]{R_{-3}} K$, $K \xrightarrow[]{R_3} MpK \xrightarrow[]{R_4} K$ and $K \xrightarrow[]{R_1} MK \xrightarrow[]{R_2} Mp \xrightarrow[]{R_3} K$).
Changing the sign of $K$ would then transform those positive circuits into negative ones and once again rule out the possibility for multistationarity.
Note that because of the conditions of the original theorems
of~\cite{Soule03complexus} on the diagonal of the Jacobian, only the vertices that
have at least one incoming and one outgoing arc are considered for rewiring.
In other words \textbf{IN0} and \textbf{OUT0} are performed before
permutations.
\hl{Because we did not find any efficient algorithm to propagate target permutation constraints,
we restricted ourselves to simple permutations made of \emph{one single swapping} between tow molecular species not eliminated by \textbf{IN0} and \textbf{OUT0},
and systematically tried beforehand in a generate-and-test manner.}
\begin{figure}[htb]
\begin{center}
\begin{tikzpicture}
\matrix[row sep=2.5cm,column sep=2.5cm] {%
\node[place] (M) {M};&&\node[place] (K) {K};&\\
&\node[place] (Mp) {Mp};&&\\
\node[place] (MK) {MK};&&\node[place] (MpK) {MpK};&\node[place] (Mpp) {Mpp};\\
};
\coordinate[shift={(-10mm,-2mm)}] (n) at (MK.south east);
\draw[bend right=20,negreg] (K) to node[auto,black, below] {$R_1$} (MpK);
\draw[negreg] (K) to node[auto,black, below] {$R_3$} (MpK);
\draw[bend right = 20, negreg] (MpK) to node[auto,black, above] {$R_4$} (K);
\draw[bend right = 40, negreg] (MpK) to node[auto,black, above] {$R_{-3}$} (K);
\draw[bend right=20,negreg] (K) to node[auto,black,above] {$R_1$} (M);
\draw[bend right=45, negreg] (M) to node[auto,black] {} (n) to node[auto,black] {$R_1$} (MpK);
\draw[bend left=10,negreg] (Mp) to node[auto,black] {$R_3$} (MpK);
\draw[bend left=10,negreg] (K) to node[auto,black] {$R_3$} (Mp);
\draw[bend left=10,posreg] (M) to node[auto,black] {$R_1$} (MK);
\draw[bend left=10,posreg] (MK) to node[auto,black] {$R_{-1}$} (M);
\draw[posreg] (MK) to node[auto,black, below] {$R_{2}$} (MpK);
\draw[bend left=20,posreg] (MK) to node[auto,black] {$R_{-1}$} (MpK);
\draw[bend right=30,posreg] (K) to node[auto,black, above] {$R_{1}$} (MK);
\draw[posreg] (MK) to node[auto,black] {$R_{2}$} (Mp);
\draw[bend left=10,posreg] (Mp) to node[auto,black] {$R_{3}$} (K);
\draw[bend left=10,posreg] (MpK) to node[auto,black] {$R_{-3}$} (Mp);
\draw[posreg] (MpK) to node[auto,black] {$R_{4}$} (Mpp);
\end{tikzpicture}
\end{center}
\caption{influence multigraph of $M + K \rightleftarrows MK \longrightarrow K + Mp
\rightleftarrows MpK \longrightarrow K + Mpp$ for which arcs ending in $K$ and $MpK$
have been swapped. Self-loops are omitted for clarity}\label{fig:mapk_swap}
\end{figure}
\section{Evaluation on the BioModels Repository}\label{sec:biomodels}
\subsection{Reaction Networks from BioModels}
To evaluate the information brought by \hl{the previous graphical requirements for multistationarity and the performance of our graph rewriting algorithms}, we downloaded the
latest release of the BioModels database\footnote{\url{http://biomodels.net/},
31st release, dated 26th of June, 2017}~\cite{CLN13issb} and applied our
method in a systematic way. First, a labelled influence \hl{multigraph} is extracted as
per Alg.~\ref{alg:sanitize}. Out of the 640 curated models, the extraction led
to 506 models with a non-trivial influence \hl{multigraph}. The other models rely on
events, assignment-rules, etc.\ to enforce their dynamics, or simply do not
contain reaction but flux-balance constraints, gene-regulations, etc.\ even
after re-import of their ODE dynamics~\cite{FGS15tcs}.
\subsection{Results of the Graphical Algorithm}
\begin{table}[htb]
\begin{center}
\begin{tabular}{lrrrrr}
\toprule
Conditions verified& Number & \multicolumn{2}{r}{Nb of species} & \multicolumn{2}{r}{Computation time}\\
& of graphs & avg. & max. & avg. (s) & max (s)\\ \midrule
All graphs & 506 & 21.24 & 430 & & \\ \midrule
\hl{No negative circuit} & \hl{70} & \hl{6.87} & \hl{57} & \hl{$< 0.01$} & \hl{$0.05$}\\ \midrule
No positive circuit & 48 & 3.42 & 18 & $< 0.01$ & $<0.01$\\
Cor.~\ref{cor:once}~\ref{cor:reversible}~\ref{cor:pinv} & 105 &
6.22 & 46 & $< 0.01$ & $<0.01$\\
Cor.~\ref{cor:once}~\ref{cor:reversible}~\ref{cor:pinv}~\ref{cor:inversed} & 160 &
8.23 & 54 & $< 0.01$ & $0.05$\\
Cor.~\ref{cor:once}~\ref{cor:reversible}~\ref{cor:pinv}~\ref{cor:inversed}~\ref{cor:permuted} & 180 & 8.38 & 54 & 5.90 & 980.1 \\
\bottomrule
\end{tabular}
\end{center}
\caption{Analysis of \hl{the 506 sanitized reaction models from the curated branch of
the BioModels repository. The table reports the proportion of models, and their size, for which
the non-existence of multiple steady states is proved using Thomas's positive circuit condition and using
the refined conditions expressed in the corollaries described above.
The computation times are given for the whole set of models. The maximum computation time is obtained for checking the last condition on model number 574.}\label{tab:biomodels}}
\end{table}
Table~\ref{tab:biomodels} summarizes the results of our experiments.
\hl{It is worth noting that the maximum running time of 50ms for checking our main graphical requirements is remarkably low.
It concerns all models of the benchmark, including the largest model number 235 that contains 430 species
and an influence multigraph of 1875 arcs.}
\hl{Another} observation is that \hl{not only the
number of models for which the absence of multistationarity is proved}
more than doubles when using
Corollaries~\ref{cor:once}~\ref{cor:reversible}~\ref{cor:pinv} on top of
Thomas's simple condition, the models that are added are of \hl{much larger size} than the one dealt with the original conditions.
Indeed with the simple condition, only very small models
with less than 18 species and linear reactions were shown to have no multistationarity, whereas the stronger conditions
allow us to prove the absence of multistationarity in models of size up to 46 species and including non-linear reactions. This is far below the size of the biggest
models of the BioModels repository (for which the existence of multiple steady states is generally unknown) but shows that the supplementary conditions
do change the scope of use of the method.
\begin{figure}[htb]
\begin{center}
\begin{tikzpicture}[scale=1]
\begin{scope}[transparency group]
\begin{scope}[blend mode=luminosity]
\fill[opacity=0.5, line width=1pt, rotate=-40, red] (-0.5,-2) ellipse (3.8 and 2);
\fill[opacity=0.5, line width=1pt, rotate=-40, blue] (0,0) ellipse (3.8 and 2);
\fill[opacity=0.5, line width=1pt, rotate=40, green] (0,0) ellipse (3.8 and 2);
\fill[opacity=0.5, line width=1pt, rotate=40, yellow] (0.5,-2) ellipse (3.8 and 2);
\fill[opacity=0.5, line width=1pt, white] (0,-1.8) ellipse (0.65 and 0.3);
\end{scope}
\end{scope}
\node (a) at (0,-0.8) {${160}$};
\node (a) at (0,-1.8) {$\mathbf{180}$};
\node (a) at (0.9,-2.3) {$154$};
\node (a) at (-0.9,-2.3) {$126$};
\node (a) at (0,-3.3) {$118$};
\node (a) at (2.2,-1.8) {$112$};
\node (a) at (-2.2,-1.8) {$60$};
\node (a) at (1.3,-0.2) {$123$};
\node (a) at (-1.3,-0.2) {$105$};
\node (a) at (3.9,0.2) {$81$};
\node (a) at (-3.9,0.2) {$55$};
\node (a) at (0,1.3) {$73$};
\node (a) at (2.5,1) {$95$};
\node (a) at (-2.5,1) {$97$};
\node (a) at (1.8,2.1) {$52$};
\node (a) at (-1.8,2.1) {$72$};
\node (a) at (0,3.2) {$48$};
\node (a) at (0,-1.4) {permutation};
\node[fill=red!50, rotate=-50] (a) at (-4.1,-1.6) {same reactions};
\node[fill=blue!50] (a) at (-2.5,2.9) {reversed reactions};
\node[fill=green!50] (a) at (1.8,2.9) {invariant};
\node[fill=yellow!50, rotate=50] (a) at (4.1,-1.6) {sign change};
\end{tikzpicture}
\end{center}
\caption{\hl{Number of models among the 506 curated reaction models of BioModels for which multistationarity can be ruled out by using respectively original Thomas's positive circuit condition, Cor.~}\ref{cor:once} (no same reactions),~\ref{cor:reversible} (no reversed reactions),~\ref{cor:pinv} (no invariant) and~\ref{cor:inversed} (sign change), plus~\ref{cor:permuted} (permutation).}\label{fig:venn}
\end{figure}
\hl{Fig.}~\ref{fig:venn} shows a Venn diagram which details the contribution of the different graphical conditions.
One can note that Cor.~\ref{cor:pinv} was in fact useful in only eight of the new models found with the other two corollaries
combined.
The condition of Cor.~\ref{cor:inversed} (sign change of all incoming edges to a set of vertices) is responsible for concluding
to the absence of multistationarity in 55 more models, of size up to 54 species.
Corollary~\ref{cor:permuted} \hl{allows us to rule out multistationarity} in 20 new models, but with the same maximum size.
Even with the restriction to \hl{single transpositions} as explained above, the \hl{maximum running time on the whole benchmark} becomes much higher than
for the simpler conditions, by five orders of magnitude.
However, the increase in the number of models for which
multistationarity is proved not to be possible
with this restricted strategy is relatively high (20) and thus encouraging \hl{for further improvements.
Indeed, better heuristics or more efficient
propagation of the permutation constraint, might lead to even more conclusive results on even larger size problems}.
\subsection{Practical Complexity}
\begin{figure}[htb]
\begin{center}
\includegraphics[width=\textwidth]{graph_exec_time.png}
\caption{Execution time of the 506 models of BioModels tested with Alg.~\ref{alg:multistat} and~\ref{alg:gauss}
with the conditions of Cor.\ref{cor:once}~\ref{cor:reversible}~\ref{cor:pinv}~\ref{cor:inversed}, relatively to the value of $e\log(n)$.
The points in blue represent the models for which multistationarity is proved impossible, and the ones in red, those for which the algorithm exhibits a circuit that satisfies the conditions of the Corollaries.}\label{fig:exec_time}
\end{center}
\end{figure}
The computation times presented in Table~\ref{tab:biomodels} with the use of the first four corollaries are far better than the theoretical complexity bound given in Prop~\ref{prop:complexity}.
It is known that reducing the graph with only the 4 original rules IN0, IN1, OUT0 and OUT1 can be done with a time complexity in $\mathcal{O}\left(e\log(n)\right)$~\cite{ll88joa} where $e$ is the number of edges in the graph.
In Alg.~\ref{alg:multistat}, these simplification rules are \hl{in fact} used with a higher priority than the rule INOUTi which are taken with the increasing order on $i$.
Fig.~\ref{fig:exec_time} plots the computation time for each model relatively to the value of $e\log(n)$.
The linear shape of the curve suggests that the empirical time complexity on BioModels is close to $\mathcal{O}\left(e\log(n)\right)$,
i.e.~the complexity of the 4 original rules, and that the extra rules INOUTi with $i \ge 2$
(although used to conclude on non multistationarity in 65 \hl{over the 160} models)
do not significantly increase the computation time apart from very few cases (models 014, 365 and 574) up to 50ms.
\hl{More precisely}, the rule INOUTi is used with an average maximum value of $i = 3.5$ and 139 models do not use more than $i = 1$ which explains why it does not add much to the computation time. Models 014 and 365 use the rule INOUTi with $i=36$ and $i=98$ respectively which explains their higher computation time. Model 574 uses the rule INOUTi with $i = 8$ which is not uncommon but this graph is dense and each node has at least a degree of 3. In this case, the simple rules with $i = 0$ or $i = 1$ that give a complexity in $\mathcal{O}\left(e\log(n)\right)$ are never used which \hl{basically} explains why this model is the longest to check.
\subsection{Comparison to the Jacobian-based Symbolic Computation Method}
In~\cite{FW13bi}, \hl{Feliu and Wiuf have presented a symbolic computation algorithm implemented in Maple 16
to directly check the existence of roots of some matrix determinant which is equivalent to a non-injectivity property implied by the existence of multiple steady states.
That condition is in principle stronger than the graphical requirements we consider.
Interestingly, they evaluated their algorithm on BioModels, with a version at that time of 365 curated models.
Their method showed that 31.6\% of the networks do not have multiple steady states,
i.e.~the same proportion as us (160 out of 506 networks) when checking the first four corollaries,
but their Maple program failed by memory overflow on 8\% of the networks
whereas our maximum computation time is 0.05s.}
\hl{
Furthermore, the proportion of conclusive analyses raises in our case to 35,5\% (180 out of 506) by using the last corollary
but currently with a high computational cost and a restricted implementation of that condition.
}
\section{Analysis of Multiple Phosphorylation Cycles and MAPK Signalling Models}
\subsection{Wang and Sontag's Futile Phosphorylation Cycles}\label{futile}
\begin{figure}
\begin{center}
\includegraphics[width=.5\textwidth]{mhk04fig1}
\caption{Fig.~1 of~\cite{MHK04jcb} displaying the general
double-phosphorylation cycle involved in all the models studied by
Markevich et al.~and by Wang and Sontag~\cite{WS08jmb}.}\label{fig:markevich}
\end{center}
\end{figure}
In~\cite{FW13bi}, \hl{Feliu and Wiuf also evaluated their method on the
the $r$-site phosphorylation cycles of Wang and Sontag
who showed the existence of multiple steady states in those networks for $r\ge 2$}~\cite{WS08jmb}\hl{.
The case $r=2$, schematized in Fig.}~\ref{fig:markevich}\hl{, was extensively studied by Markevich et al.~in~}\cite{MHK04jcb}\hl{
in a series of models, numbered from 26 to 31 in the
BioModels repository (see~}\cite{GSF10bi} \hl{for the model reduction relationships between these models found by subgraph epimorphisms),
showing in all cases the existence of multiple steady states.}
\hl{The symbolic computation method used by Feliu and Wiuf
grew rapidly in time as a function of the number $r$ of phosphorylations, and became impractical after $r=17$ for which it needed 1200 seconds.
Our graphical method has a very low computational complexity on these networks, taking only 1.2s for $r=1000$.
It is worth noting that our method checks necessary conditions for the non-injectivity of the system whereas the symbolic method of Feliu and Wiuf directly determines that property.
In both cases, though, one cannot conclude that the system does have multiple stationary states since the non-injectivity property is itself a necessary not sufficient condition.
}
\hl{Table}~\ref{tab:r_site_phosphorylation} \hl{summarizes our results.
The first two computation time columns refer to the original model of Wang and Sontag
in which each phosphorylation and dephosphorylation transformation is modelled
by three reactions with mass action law kinetics
with an explicit representation
of the intermediary complexes,
by repeating the following pattern:}
\begin{equation}
\left\{\begin{array}{l}
E+S_i \rightleftarrows ES_i \rightarrow E+S_{i+1} \\
F+S_{i+1} \rightleftarrows FS_{i+1} \rightarrow F+S_i
\end{array}\right.
\label{eq:WangSontag}
\end{equation}
\hl{This gives the following differential equations:}
$$
\begin{array}{rclr}
\dfrac{dS_0}{dt} &=& -k_{\on_0}S_0E + k_{\off_0}ES_0 + l_{cat_0}FS_1 \\
\dfrac{dS_i}{dt} &=& -k_{\on_i}S_iE + k_{\off_i}ES_i + k_{cat_{i-1}}ES_{i-1} \\ && - l_{\on_{i-1}}S_if + l_{\off_{i-1}}FS_i + l_{cat_i}FS_{i+1} &, i = 1, ..., n-1\\
\dfrac{dES_j}{dt} &=& k_{\on_j}S_jE - \left(k_{\off_j} + k_{cat_j}\right)ES_j &, j = 0, ..., n-1\\\\
\dfrac{dFS_k}{dt} &=& l_{\on_{k-1}}S_kf - \left(l_{\off_{k-1}} + l_{cat_{k-1}}\right)FS_k &, k = 1, ..., n\\
\end{array}
$$
\hl{The computation times given in Table}~\ref{tab:r_site_phosphorylation} \hl{indicate that, on these networks, our method
has an empirical complexity of the order of $10^{-3}r^2$.}
\begin{table}[!htb]
\begin{center}
\begin{tabular}{ccccc}
\toprule
\hl{$r$} & \hl{Jacobian Method}~\cite{FW13bi} & \multicolumn{3}{r}{\hl{Graphical Method (Alg}~\ref{alg:multistat} \hl{\&}~\ref{alg:gauss}\hl{)}}\\
& \hl{model }\eqref{eq:WangSontag}& \eqref{eq:WangSontag} &
\eqref{eq:Markevich}& \eqref{eq:CatalyticReactions} \\ \midrule
\hl{ 1 }&\hl{ 4 }&\hl{ 0.3 }&\hl{ 1.6 }&\hl{ 0.6 }\\
\hl{ 2 }&\hl{ 75 }&\hl{ 0.5 }&\hl{ 2 }&\hl{ 1 }\\
\hl{ 3 }&\hl{ 44 }&\hl{ 1 }&\hl{ 2 }&\hl{ 1 }\\
\hl{ 4 }&\hl{ 81 }&\hl{ 1 }&\hl{ 3 }&\hl{ 1 }\\
\hl{ 5 }&\hl{ 191 }&\hl{ 1 }&\hl{ 4 }&\hl{ 1 }\\
\hl{ 6 }&\hl{ 256 }&\hl{ 1 }&\hl{ 4 }&\hl{ 1 }\\
\hl{ 7 }&\hl{ 444 }&\hl{ 1 }&\hl{ 5 }&\hl{ 1 }\\
\hl{ 8 }&\hl{ 795 }&\hl{ 2 }&\hl{ 5 }&\hl{ 1 }\\
\hl{ 9 }&\hl{ 1169 }&\hl{ 2 }&\hl{ 6 }&\hl{ 2 }\\
\hl{ 10 }&\hl{ 2195 }&\hl{ 2 }&\hl{ 6 }&\hl{ 2 }\\
\hl{ 11 }&\hl{ 3998 }&\hl{ 2 }&\hl{ 6 }&\hl{ 2 }\\
\hl{ 12 }&\hl{ 7696 }&\hl{ 2 }&\hl{ 7 }&\hl{ 2 }\\
\hl{ 13 }&\hl{ 15180 }&\hl{ 2 }&\hl{ 7 }&\hl{ 2 }\\
\hl{ 14 }&\hl{ 32180 }&\hl{ 3 }&\hl{ 7 }&\hl{ 2 }\\
\hl{ 15 }&\hl{ 67740 }&\hl{ 3 }&\hl{ 7 }&\hl{ 2 }\\
\hl{ 16 }&\hl{ 171700 }&\hl{ 3 }&\hl{ 8 }&\hl{ 2 }\\
\hl{ 17 }&\hl{ 1199000 }&\hl{ 4 }&\hl{ 8 }&\hl{ 2 }\\
\hl{ 50 }&\hl{ $\times$ }&\hl{ 12 }&\hl{ 17 }&\hl{ 4 }\\
\hl{ 100 }&\hl{ $\times$ }&\hl{ 26 }&\hl{ 40 }&\hl{ 6 }\\
\hl{ 500 }&\hl{ $\times$ }&\hl{ 343 }&\hl{549 }&\hl{ 34}\\
\hl{ 1000 }&\hl{ $\times$ }&\hl{ 1200 }&\hl{ 1874 }&\hl{ 98 }\\
\bottomrule
\end{tabular}
\end{center}
\caption{Execution times given in milliseconds for the analysis of the $r$-site phosphorylation system of~\cite{WS08jmb},
first as reported in~\cite{FW13bi} for the Jacobian method using symbolic computation, then obtained with our graphical algorithm on the same model and on two variants concerning the writing of the dephosphorylation and phosphorylation reactions.}\label{tab:r_site_phosphorylation}
\end{table}
\hl{The third computation time column refers to the writing of the
dephosphorylations with two intermediate complexes, as follows:}
\begin{equation}
\left\{\begin{array}{l}
E+S_i \rightleftarrows ES_i \rightarrow E+S_{i+1} \\
F+S_{i+1} \rightleftarrows FS_{i+1}^\star \rightarrow FS_i \rightleftarrows F+S_i
\end{array}\right.
\label{eq:Markevich}
\end{equation}
\hl{
This writing of the dephosphorylations corresponds to the first model of Markevich et al.~}\cite{MHK04jcb}.
\hl{On this reaction pattern }\eqref{eq:Markevich}\hl{, our graph algorithm
has execution times similar to those obtained on reaction pattern
}\eqref{eq:WangSontag}.
This is due to the resemblance of their influence multigraphs.
Nevertheless, the rule INOUTi is used \hl{here} with $i \le 12$,
while on model pattern~\eqref{eq:WangSontag}\hl{ it is used with value at most $9$.
This is responsible for a slight difference in response time.}
\hl{The second model of Markevitch et al.~}\cite{MHK04jcb} \hl{
is a reduction of the previous model using Michaelian kinetics.
The intermediary complexes are eliminated
but}
the writing of the kinetics for the
dephosphorylation of Mp by phosphatase MKP3, named $v_4$ in the original
article~\cite{MHK04jcb}, is \hl{not a naive Michaelis-Menten kinetics but the following one}:
\[v_4 = \frac{k_4^{cat}\cdot{[\mathit{MKP3}]}_{tot}\cdot [Mp]/K_{m3}}{(1 +
[Mpp]/K_{m3} + [Mp]/K_{m4} + [M]/K_{m5})}\]
\hl{Mpp appears as inhibitor in this
kinetic expression to represent} the sequestration of the phosphatase in the
reversible last step of dephosphorylation,
Such a sequestration results in a negative influence of Mpp on M and
a positive influence of Mpp on Mp (i.e., a positive term in
$\frac{\partial\dot{Mp}}{\partial Mpp}$) as this reaction consumes Mp to
produce M.
Intuitively the fact that Mpp can actively sequestrate the phosphatase MKP3
makes it inhibit the dephosphorylation of Mp and therefore stabilizes Mp.
\hl{Therefore, while the positive circuit between Mp and Mpp by $v_2$ and $v_3$ that can be easily seen from Fig.~}\ref{fig:markevich}
\hl{is immediately rule out since $v_2$ and $v_3$ are opposite reactions},
the complex Michaelian kinetics of \cite{MHK04jcb}
gives rise to a completely different positive circuit between those two
species (with kinetics $v_4$ and $v_2$). This circuit cannot be removed by sign-changes or
any single swap and is indeed responsible for the appearance of bistability in those models.
\hl{The fourth computation time column refers to that model structure,
with catalytic reactions instead of intermediary complexes, but using mass action law (or simple Michaelis-Menten) kinetics,
with the following pattern:}
\begin{equation}
\left\{\begin{array}{l}
E+S_i \rightarrow E+S_{i+1} \\
F+S_{i+1} \rightarrow F+S_i
\end{array}\right.
\label{eq:CatalyticReactions}
\end{equation}
\hl{The differential equations for this pattern with simple Michaelis-Menten kinetics are as follows :}
$$
\begin{array}{rclr}
\dfrac{dS_0}{dt} &=& -\dfrac{V_0ES_0}{K_0+S_0} + \dfrac{V_0^*fS_{1}}{K_0^*+S_1}\\
\dfrac{dS_i}{dt} &=& -\dfrac{V_iES_i}{K_i+S_i} + \dfrac{k_{i-1}ES_{i-1}}{K_{i-1}+S_{i-1}} \\
&& + \dfrac{V_i^*fS_{i+1}}{K_i^*+S_{i+1}} - \dfrac{V_{i-1}^*fS_i}{K_{i-1}^* + S_i} &~~~~, i=1,...,n-1 \\
\dfrac{dS_n}{dt} &=& \dfrac{V_{n-1}ES_{n-1}}{K_{n-1}+S_{n-1}} - \dfrac{V_{n-1}^*fS_n}{K_{n-1}^*+S_n}
\end{array}
$$
\hl{In this modelling of the system, the possibilities of multistationarity disappear,
and this is shown by the result of our graphical algorithm.
On large instances, the computation time is also significantly lower.
This is because the graph does not contain any positive cycle satisfying our conditions,
and the graph algorithm has only to remove nodes using the rule INOUTi with $i\le 1$.
}
\subsection{MAPK Signalling Models}\label{sec:erk}
\begin{figure}[htb]
\begin{center}
\includegraphics[width=.8\textwidth]{F1A}
\caption{Figure 1A of~\cite{SMHWKKLTK09msb}, on the right is the
distributive model encoded as model 270 in the BioModels repository.
Several other variables appear in the model because of the encoding with ODEs of
Delay Differential Equations involved close to the receptor.}\label{fig:erk}
\end{center}
\end{figure}
\hl{
Double phosphorylation cycles are integral parts of MAPK signalling cascades
and one might expect to observe multistationarity in MAPK models.
However, as shown in the previous section, this depends on the way the phosphorylation and dephosphorylation reactions are modeled.
}
\hl{Model 270 of the BioModels database} describes a complete Epo-induced ERK signalling cascade, from
receptor binding to cell fate decision, corresponding to the \emph{distributive} model of~\cite{SMHWKKLTK09msb}, schematized in
Fig.~\ref{fig:erk}.
It includes four reversible
double-phosphorylation stages (MEK2, MEK1, ERK1 and ERK2) and many dummy
variables introduced at the beginning of the cascade to encode delay ordinary
equations into simple ODEs. The resulting reaction model has 33 species and 42 reactions,
and leads to a labelled influence \hl{multigraph} containing 126 arcs with many positive and
negative feedback loops.
During the run of Alg.~\ref{alg:multistat}, only 2
paths are removed thanks to one of the 9 conservation laws (Cor.~\ref{cor:pinv}), but 58 are removed
thanks to Cor.~\ref{cor:once} and~\ref{cor:reversible}, resulting in no single
positive feedback loop satisfying the conditions of our theorem.
The fact that multistationarity is not possible for such a model is consistent
with the data shown by the authors in the article, i.e.~functional dose-response diagrams without hystereses.
This is however not evident, nor perhaps expected, from the
model itself since memory effects resulting in hysteresis might have been
possible at many different places in the network due to the dephosphorylation loops.
\hl{However, as explained in the previous section with model~}\ref{eq:CatalyticReactions}
\hl{this is not the case when simple Michaelis-Menten kinetics are used}.
These examples show that the existence of multiple steady states in reaction networks
is sensitive to the explicit representation of the intermediate complexes in enzymatic reactions,
or at least to the explicit inclusion of their inhibitors (by sequestration) in the kinetics.
Interestingly, our refined conditions are similarly sensitive to these subtle modelling choices
and allow us to conclude differently according to the impact of the writing of the reactions
on the multistationarity properties of the system.
The role of intermediate complexes in multistationarity was analysed in detail in~\cite{FC13rsi}.
In particular, it was shown that if the network does not have conservation laws,
then multistationarity cannot arise after the introduction of intermediate complexes.
These remarks also go in the same direction to what has been observed for oscillations in the
MAPK cascade again, where the absence of complexation removes the negative feedbacks
going upwards in the cascade and therefore the negative feedback loops and
the corresponding possibility of oscillations~\cite{VSM08plos}. If the intermediary complexes are
explicitly represented, then oscillations can be found~\cite{QNKS07plos},
without any external negative feedback reaction such as receptor desensitization~\cite{Kholodenko00ejb}.
\hl{In many other networks in} BioModels, Alg.~\ref{alg:multistat}
actually shows as side-effect the existence of numerous negative feedback loops.
\section{Conclusion}\label{sec:conclu}
This experiment is, to our knowledge, the first systematic evaluation of
\hl{graphical requirements for multistationarity in the reaction networks of model repositories in large scale.
We have shown that Thomas's necessary condition for multistationarity, and its refinement for reaction models given in~}\cite{Soliman13bmb},
can be implemented with a graph rewriting algorithm that brings useful information for
many models in BioModels,
by proving the non-existence of multiple steady states independently of the parameter values and of the precise form of the rate functions.
\hl{Though the
original Thomas's conditions show the absence of multistationarity in some small models,
the refined conditions are conclusive in many more cases: 180 vs~48, including
much bigger models: up to 54 vertices vs~18.
Furthermore this is achieved at a remarkably low computational cost, below 0.05 second per network for the main conditions,
even on models with several hundreds of molecular species and thousands of influence arcs, currently out of reach of symbolic computation methods}.
It is worth noting that our graph-theoretic algorithm is not limited to reaction systems with mass action law kinetics,
but relies on a simple symbolic derivation algorithm for computing an over-approximation of the signs of the partial derivatives in the Jacobian matrix.
\hl{In case of indeterminacy of the sign, both signs are assumed which may lead to the existence of circuits that would have been ruled out by a more accurate
determination of the sign.}
Our graphical algorithm could also be improved by a more efficient \hl{and complete} use of the \hl{condition dealing with target species permutations},
for instance by recourse to constraint propagation algorithms~\cite{FL06cor} instead of the current generate-and-test procedure for single swappings.
\hl{This might further increase the number of conclusive cases}.
\hl{Another way would be to use the condition noted (*)
in}~\cite{BC09cms} \hl{to rule out
the positive circuits that do not intersect another positive circuit on a
species-to-reaction path, which is necessary for multistationarity.}
Since our procedure, and more precisely Alg.~\ref{alg:sanitize}, goes
through an import of the ODE system \hl{and infers a reaction network, it can be readily used} on dynamical systems that do not stem from reaction
networks but may exhibit similar symmetries. Such use of the refined conditions in general ODE systems would probably benefit much less from the structural
conditions added on top of Thomas's rules, but by identifying similar terms in the
ODEs, our algorithm should be able to automatically prove the absence of
multistationarity in interesting cases, \hl{as also suggested in}~\cite{BC09cms}.
A comparison to Feinberg's CRNT-based approaches would also be interesting,
\hl{%
by considering the different approaches summarized for instance in Table 3
of
}~\cite{FW13bi}.
\hl{In particular, our circuit conditions on the influence multigraph depend on the signs of the entries of the Jacobian matrix but are independent
of not only the values of the kinetic parameters, but also of the form of the reaction rate functions which can be any partially differentiable function,
i.e.~without any monotonicity, non-autocatalytic, or such restriction.}
\hl{Finally, this study focussed on multistationarity, but we saw that most
models of the benchmark also have negative circuits.
A systematic study of the oscillation conditions in reaction model repositories,
possibly using a similar theoretical refinement of Thomas-Snoussi's necessary
conditions for sustained oscillations}~\cite{Snoussi98jbs},
\hl{would be worth investigating
as natural systems indeed provide many oscillators and}
even models not conceived to oscillate have been shown capable of exhibiting unexpected sustained oscillations in non-standard conditions~\cite{QNKS07plos}.
\subsection*{Acknowledgments.}
This work has been supported by the bilateral project ANR-17-CE40-0036
SYMBIONT ({https://www.symbiont-project.org}).
We are also grateful to the reviewers for their useful comments for improving the presentation of our results.
\section*{References}
\bibliographystyle{elsarticle-harv}
|
\section{Introduction}
In this paper, we consider convergence of the weighted nonlocal Laplacian (WNLL) on high dimensional randomly distributed data.
WNLL is proposed in \cite{WGL} for high dimensional point cloud interpolation.
High dimensional point cloud interpolation is a fundamental problem in machine learning which can be formulated as: Let
$P = \{\bm{p}_1, \cdots, \bm{p}_n\}$ and $S =\{\bm{s}_1, \cdots, \bm{s}_m\}$ be two sets of points in $\mathbb{R}^d$
Suppose $u$ is a function defined on the point cloud $\bar{P}=P\cup S$ which is known only over $S$, denoted as $b(\bm{s})$ for any $\bm{s} \in S$. The interpolation methods are used to compute $u$ over the whole point cloud $\bar{P}$ from the given values over $S$.
In nonlocal Laplacian, which is widely used in nonlocal methods for image processing \cite{BCM05,BCM06,GO07,GO08}, the interpolation function is obtained by minimizing the energy functional
\begin{equation}
\label{eq:ob-gl}
\mathcal{J}(u) = \frac{1}{2}\sum_{\bm{x},\bm{y}\in \bar{P}} w(\bm{x},\bm{y})(u(\bm{x})-u(\bm{y}))^2,
\end{equation}
with the constraint
\begin{equation}
\label{eq:constraint}
u(\bm{x})=b(\bm{x}),\quad \bm{x}\in S.
\end{equation}
Here $w(\bm{x},\bm{y})$ is a given weight function, typically chosen to be Gaussian, i.e.,
$w(\bm{x},\bm{y})=\exp(-\frac{\|\bm{x}-\bm{y}\|^2}{\sigma^2})$, $\sigma$ is a parameter, $\|\cdot\|$ is the Euclidean norm in $\mathbb{R}^d$.
In graph theory and machine learning literatures, nonlocal Laplacian is also called graph Laplacian \cite{Chung:1997,ZhuGL03}.
Graph Laplacian works very well with high labeling rate, i.e., there is a large portion of data been labeled. However, when the labeling rate is low, i.e., $|S|/|\bar{P}|\ll 1$, the solution of the graph Laplacian is found to be discontinuous at the labeled points \cite{Shi:harmonic,WGL}.
WNLL is proposed to fix this problem. In WNLL, energy functional in \eqref{eq:ob-gl} is modified by adding a weight, $\frac{|\bar{P}|}{|S|}$, to balance the labeled and unlabeled terms, which leads to
\begin{align}
\label{opt:wgl}
\hspace{-3mm} \min_{u} \sum_{\bm{x}\in P}\left(\sum_{\bm{y}\in \bar{P}} w(\bm{x},\bm{y})(u(\bm{x})-u(\bm{y}))^2\right)
+\frac{|\bar{P}|}{|S|}\sum_{\bm{x}\in S}\left(\sum_{\bm{y}\in \bar{P}} w(\bm{x},\bm{y})(u(\bm{x})-u(\bm{y}))^2\right),
\end{align}
with the constraint
\begin{align*}
u(\bm{x})=b(\bm{x}),\quad \bm{x}\in S.
\end{align*}
When the labeling rate is high, WNLL is close to graph Laplacian.
When the labeling rate is low, the weight forces the solution to be close to the given values near the labeled points, such that the discontinuities are removed.
With a symmetric weight function, i.e. $w(\bm{x},\bm{y})=w(\bm{y},\bm{x})$, the corresponding Euler-Lagrange equation of \eqref{opt:wgl} is a simple linear system
\begin{align}
2\sum_{\bm{y}\in P} w(\bm{x},\bm{y})\left(u(\bm{x})-u(\bm{y})\right)
+ \left(\frac{|P|}{|S|}+2\right)\sum_{\bm{y}\in S} w(\bm{y},\bm{x})(u(\bm{x})-b(\bm{y}))&=0,\quad \bm{x}\in P,\nonumber\\
u(\bm{x})&=b(\bm{x}),\quad \bm{x}\in S.\nonumber
\end{align}
This linear system can be solved efficiently by conjugate gradient iteration. The superiority of the WNLL compared to the graph Laplacian has been shown evidently in image inpainting \cite{Shi:harmonic,WGL}, scientific data interpolation \cite{Zhu:2018JCP}, and more recently deep learning \cite{Wang:2018}.
\subsection{Main Result}
We consider error of the WNLL in a model problem. The whole computational domain is set to be a $k$-dimensional closed manifold $\mathcal{M}$ embedded in $\mathbb{R}^d$. The point cloud $P$ gives a discrete representation
of $\mathcal{M}$ which is assumed to be uniformly distributed on $\mathcal{M}$.
$\mathcal{D}\subset \mathcal{M}$ be a subset of $\mathcal{M}$ which has been labeled, and $S$ is a uniform sample of $\mathcal{D}$. In $S$, we have $u(\bm{x})=b(\bm{x})$.
An illustration of the computational domain and the point cloud is shown in Fig. \ref{fig:domain_large}.
In WNLL, in order to extend the label function $u$ to the entire dataset $P$,
we solve the linear system
\begin{align}
\label{eq:wgl}
\sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y})\left(u_\delta(\bm{x})-u_\delta(\bm{y})\right)+ \mu \sum_{\bm{y}\in S} K_\delta(\bm{x},\bm{y})(u_\delta(\bm{x})-b(\bm{y}))&=0,\quad \bm{x}\in P,\\
u_\delta(\bm{x})&=b(\bm{x}),\quad \bm{x}\in S.\nonumber
\end{align}
where $R_\delta(\bm{x},\bm{y})$, $K_\delta(\bm{x},\bm{y})$ are kernel functions given as
\begin{equation}
\label{eq:kernel}
R_\delta(\bm{x}, \bm{y}) = C_\delta R\left(\frac{|\bm{x} -\bm{y}|^2}{4\delta^2}\right),\quad K_\delta(\bm{x}, \bm{y}) = C_\delta K\left(\frac{|\bm{x} -\bm{y}|^2}{4\delta^2}\right),
\end{equation}
where $C_\delta = \frac{1}{(4\pi \delta^2)^{k/2}}$ is the normalization factor.
$R, K\in C^2(\mathbb{R}^+) $ are two kernel functions satisfying the conditions listed in Assumption \ref{assumptions}.
\begin{figure}[H]
\centering
\begin{tabular}{c}
\includegraphics[width=0.6\textwidth]{domain_large_3.eps}
\end{tabular}
\caption{Illustration of the computational domain. Gray points: sample of $\mathcal{M}$; Black points: sample of $\mathcal{D}\subset \mathcal{M}$.}\label{fig:domain_large}
\end{figure}
As the continuous counterpart, we consider the Laplace-Beltrami equation on a closed smooth manifold $\mathcal{M}$
\begin{align}
\label{eq:laplace-large}
\left\{
\begin{array}{rcll}
\Delta_\mathcal{M} u(\bm{x})&=&0, & \bm{x}\in \mathcal{M},\\
u(\bm{x})&=&b(\bm{x}),& \bm{x}\in \mathcal{D},
\end{array}\right.
\end{align}
where $\Delta_\mathcal{M}=\text{div}(\nabla)$ is the Laplace-Beltrami operator on $\mathcal{M}$. Let $\Phi: \Omega\subset \mathbb{R}^k\rightarrow \mathcal{M}\subset\mathbb{R}^d$ be a local parametrization of $\mathcal{M}$ and $\theta\in \Omega$.
For any differentiable function $f:\mathcal{M}\rightarrow \mathbb{R}$,
we define the gradient on the manifold
\begin{align}
\label{eq:diff-M}
\nabla f(\Phi(\theta))&=\sum_{i,j=1}^m g^{ij}(\theta)\frac{\partial \Phi}{\partial\theta_i}(\theta)\frac{\partial f(\Phi(\theta))}{\partial\theta_j}(\theta).
\end{align}
And for vector field $F:\mathcal{M}\rightarrow T_{\bm{x}}\mathcal{M}$ on $\mathcal{M}$, where $T_{\bm{x}}\mathcal{M}$ is the tangent space of $\mathcal{M}$ at $\bm{x}\in \mathcal{M}$, the divergence is defined as
\begin{align}
\label{eq:diver}
\text{div} (F)&= \frac{1}{\sqrt{\det G}}\sum_{k=1}^d\sum_{i,j=1}^m\frac{\partial}{\partial \theta_i}\left(\sqrt{\det G}\,g^{ij}F^k(\Phi(\theta))\frac{\partial \Phi^k}{\partial\theta_j}\right)
\end{align}
where $(g^{ij})_{i,j=1,\cdots,k}=G^{-1}$, $\det G$ is the determinant of matrix $G$ and $G(\theta)=(g_{ij})_{i,j=1,\cdots,k}$ is the first fundamental form with
\begin{eqnarray}
\label{eq:remainn}
g_{ij}(\theta)=\sum_{k=1}^d\frac{\partial \Phi_k}{\partial\theta_i}(\theta)\frac{\partial \Phi_k}{\partial\theta_j}(\theta),\quad i,j=1,\cdots,m.
\end{eqnarray}
and $(F^1(\bm{x}),\cdots,F^d(\bm{x}))^T$ is the representation of $F$ in the embedding coordinates.
To prove the convergence, we need the following assumptions.
\begin{assumption}
\label{assumptions}
\begin{itemize}
\item[]
\item \rm Assumptions on the manifold: $\mathcal{M}$ be a $k$-dimensional closed $C^\infty$ manifold isometrically embedded in a Euclidean space $\mathbb{R}^d$. $\mathcal{D}$ and $\partial\mathcal{D}$ are smooth submanifolds of $\mathbb{R}^d$. Moreover, $b(\bm{x})\in C^1(\mathcal{D})$.
\item \rm Assumptions on the kernel functions:
\begin{itemize}
\item[\rm (a)] \rm Smoothness: $K(r), R(r)\in C^2(\mathbb{R}^+)$;
\item[(b)] Nonnegativity: $R(r), K(r)\ge 0$ for any $r\ge 0$.
\item[(c)] Compact support:
$R(r) = 0$ for $\forall r >1$; $K(r) = 0$ for $\forall r > r_0\ge 2$.
\item[(d)] Nondegeneracy:
$\exists \delta_0>0$ such that $R(r)\ge\delta_0$ for $0\le r\le 1/2$ and $K(r)\ge\delta_0$ for $0\le r\le 2$.
\end{itemize}
\item Assumptions on the point cloud: $P$ and $S$ are uniformly distributed on $\mathcal{M}$ and $\mathcal{D}$, respectively.
\end{itemize}
\end{assumption}
In this paper, we use the notation $C$ to denote any constant which may be different in different places.
The main contribution of this paper is to analyze relation between
the solutions of the Laplace-Beltrami equation \eqref{eq:laplace-large} and the WNLL \eqref{eq:wgl}. More precisely, we prove the following theorem:
\begin{theorem}
\label{thm:main}
Let $u_\delta$ solves \eqref{eq:wgl} and $u$ solves \eqref{eq:laplace-large}.
Under the assumptions in Assumption \ref{assumptions}, with probability at least $1-1/(2n),\; where \ n=|P|$, we have
\begin{equation}
|u_\delta-u|\le C\delta,\nonumber
\end{equation}
as long as
\begin{equation}
\label{eq:condition-mu-intro}
\mu\sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y})\ge C \sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y}),\quad \bm{x}\in P\cap\mathcal{D}_\delta.
\end{equation}
$\mathcal{D}_\delta=\{\bm{x}\in \mathcal{M}: \text{dist}(\bm{x},\mathcal{D})\le 2\delta\}$, $C=C(\mathcal{M},\mathcal{D},R,K)>0$ is a constant independent of $\delta$, $P$ and $S$.
\end{theorem}
In the above theorem, \eqref{eq:condition-mu-intro} actually gives a condition for the weight $\mu$.
Notice that $$ \frac{1}{n} \sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y})\approx \frac{1}{|\mathcal{M}| }\int_\mathcal{M} R_\delta(\bm{x},\bm{y})\mathrm{d} \bm{y}= O(1),\quad \bm{x}\in P\cap\mathcal{D}_\delta.$$
$S$ samples $\mathcal{D}$, if $S$ is dense enough, we have that
\begin{equation}
\frac{1}{|S|}\sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y})\approx\frac{1}{|\mathcal{D}| }\int_\mathcal{D} K_\delta(\bm{x},\bm{y})\mathrm{d} \bm{y}, \quad \bm{x}\in P\cap\mathcal{D}_\delta.\nonumber
\end{equation}
Here, we need the assumption on $K$ such that $K(r)\ge\delta_0>0, \forall 0\le r\le 2$. This implies that
\begin{equation*}
\int_\mathcal{D} K_\delta(\bm{x},\bm{y})\mathrm{d} \bm{y}=O(1), \quad \bm{x}\in P\cap\mathcal{D}_\delta.
\end{equation*}
Hence, from \eqref{eq:condition-mu-intro}, we have
\begin{equation}
\mu\sim \frac{|P|}{|S|}.\nonumber
\end{equation}
This explains the scaling of $\mu$ in WNLL.
\begin{figure}[H]
\centering
\begin{tabular}{c}
\includegraphics[width=0.6\textwidth]{domain_sparse_2.eps}
\end{tabular}
\caption{Illustration of the computational domain with extremely low labeling rate.}\label{fig:domain_sparse}
\end{figure}
On the other hand, if sample $S$ is extremely sparse such that $\bigcup_{\bm{x}\in S}B(\bm{x};4\delta)$ does not cover $\mathcal{D}_\delta$
as shown in Fig. \ref{fig:domain_sparse}, $\sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y})$ may be zero for some $\bm{x}\in P\cap\mathcal{D}_\delta$.
In this case, condition \eqref{eq:condition-mu-intro} does not hold. Then we can not guarantee the convergence even in WNLL.
With extremely low labeling rate, actually, the whole framework of harmonic extension fails \cite{NS09,ZB11}. We should use other approach to get a smooth interpolation.
Theorem \ref{thm:main} is a direct consequence of the maximum principle (Theorem \ref{thm:maximum-principle}) and the error estimation (Theorem \ref{thm:error}).
\begin{theorem}
\label{thm:maximum-principle}
Under the assumptions in Assumption \ref{assumptions}, with probability at least $1-1/(2n)$, $n=|P|$, $L_{\delta,n}$ has the comparison principle, i.e.
\begin{align}
|L_{\delta,n} u (\bm{x})|\le L_{\delta,n} v (\bm{x})\quad \rightarrow\quad |u|\le v,\nonumber
\end{align}
where
\begin{align}
\label{eq:Ld}
L_{\delta,n} u(\bm{x})=\sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y})\left(u(\bm{x})-u(\bm{y})\right)+ \mu \sum_{\bm{y}\in S} K_\delta(\bm{x},\bm{y})u(\bm{x}),\quad \bm{x}\in P.
\end{align}
\end{theorem}
\begin{theorem}
\label{thm:error}
Let $u_\delta$ and $u$ solve \eqref{eq:wgl} and \eqref{eq:laplace-large} respectively. $v$ is the solution of \eqref{eq:laplace-large-ref},
\begin{align}
\label{eq:laplace-large-ref}
\left\{
\begin{array}{rcll}
-\Delta_\mathcal{M} v(\bm{x})&=&1, & \bm{x}\in \mathcal{M}\backslash \mathcal{D},\\
v(\bm{x})&=&1,& \bm{x}\in \mathcal{D}.
\end{array}\right.
\end{align}
Under the assumptions in Assumption \ref{assumptions}, with probability at least $1-1/(2n)$, $n=|P|$,
\begin{align}
|L_{\delta,n} (u_\delta-u)|\le C \delta L_{\delta,n} v,\nonumber
\end{align}
as long as
\begin{equation}
\mu\sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y})\ge C \sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y}),\quad \bm{x}\in P\cap\mathcal{D}_\delta.\nonumber
\end{equation}
$\mathcal{D}_\delta=\{\bm{x}\in \mathcal{M}: \text{dist}(\bm{x},\mathcal{D})\le 2\delta\}$, $C=C(\mathcal{M},\mathcal{D},R,K)>0$ is a constant independent on $\delta$, $P$ and $S$.
\end{theorem}
The above two theorems will be proved in Section \ref{sec:max} and Section \ref{sec:convergence}, respectively.
In Section \ref{sec:entropy}, we prove a technical theorem used in the analysis. Some discussions are made in Section \ref{sec:concl}.
\section{Maximum Principle (Theorem \ref{thm:maximum-principle})}
\label{sec:max}
First, we introduce some notations.
For any two points $\bm{x},\bm{y}\in P$, we say that they are neighbors if and only if $R_\delta(\bm{x},\bm{y})>0$, denoted as $\bm{x}\sim \bm{y}$.
For $\bm{x}\in S,\bm{y}\in P\cup S$, they are neighbors if and only if $K_\delta(\bm{x},\bm{y})>0$, denoted also by $\bm{x}\sim \bm{y}$ or $\bm{y}\sim \bm{x}$. $\bm{x}$ and $\bm{y}$ are connected if there exist
$\bm{z}_1,\cdots,\bm{z}_m\in P\cup S$ such that $$\bm{x}\sim \bm{z}_1\sim\cdots\sim\bm{z}_m\sim\bm{y}.$$
We say point cloud $P$ is $S$-connected if for any point $\bm{x}\in P$, there exists $\bm{y}\in S$, such that $\bm{x}$ and $\bm{y}$ are connected.
If $P$ is $S$-connected, it is easy to check that $L_{\delta,n}$ has the maximum principle, i.e.
\begin{align}
\label{eq:max}
L_{\delta,n} u (\bm{x})\ge 0,\; \bm{x} \in P \quad \rightarrow \quad u(\bm{x})\ge 0,\; \bm{x} \in P,\\
L_{\delta,n} u (\bm{x})\le 0,\; \bm{x} \in P \quad \rightarrow \quad u(\bm{x})\le 0,\; \bm{x} \in P.
\end{align}
and consequently
\begin{align}
\label{eq:compare}
|L_{\delta,n} u (\bm{x})|\le L_{\delta,n} v (\bm{x})\quad \rightarrow\quad |u|\le v.
\end{align}
In the rest of this section, we will prove that with high probability, $P$ is $S$-connected. To prove this, we need a theorem from the empirical process theory \cite{entropy}.
\begin{theorem}
\label{thm:entropy-bound}
With probability at least $1-1/(2n)$, $n=|P|$,
\begin{align}
\sup_{f\in \mathcal{R}_\delta}|I(f)-I_n(f)|\le \frac{C}{\delta^k\sqrt{n}}\left(\ln n-2\ln \delta +1\right)^{1/2},
\end{align}
where $k$ is the dimension of $\mathcal{M}$,
$$I(f)=\frac{1}{|\mathcal{M}|}\int_\mathcal{M} f(\bm{x})\mathrm{d} \bm{x},\quad I_n(f)=\frac{1}{n}\sum_{\bm{x}\in P}f(\bm{x}),$$
$|\mathcal{M}|$ is the volume of $\mathcal{M}$ and $\mathcal{R}_\delta$ is a function class defined as
$$\mathcal{R}_\delta=\{R_\delta(\bm{x},\cdot): \bm{x}\in \mathcal{M}.\}$$
\end{theorem}
This theorem will be proved in Section \ref{sec:entropy}.
Suppose $P$ is not $S$-connected. Let $$\bar{S}=\{\bm{x}\in P\cup S: \bm{x} \;\text{is connected to $S$}\},\quad \bar{S}^c=(P\cup S)\backslash \bar{S}.$$
Then $\bar{S}^c\ne \emptyset$.
Denote
\begin{align*}
\bar{S}_\delta=\left(\bigcup_{\bm{x}\in \bar{S}}B(\bm{x};\delta/2)\right)\cap \mathcal{M},\quad \bar{S}_\delta^c=\left(\bigcup_{\bm{x}\in \bar{S}^c}B(\bm{x};\delta/2)\right)\cap \mathcal{M}
\end{align*}
where $B(\bm{x};\delta)=\{\bm{y}\in \mathbb{R}^d: |\bm{x}-\bm{y}|\le \delta\}$.
Using the definition of $\bar{S}$ and $\bar{S}^c$, we know that
$\bar{S}_\delta\cap \bar{S}_\delta^c=\emptyset$, hence
$$\partial \bar{S}_\delta\cap \bar{S}_\delta^c=\emptyset,$$
where $\partial \bar{S}_\delta$ is the boundary of $\bar{S}_\delta$ in $\mathbb{R}^d$.
Furthermore, since $\mathcal{M}$ is connected, we have
$$\partial \bar{S}_\delta \cap \mathcal{M} \ne \emptyset.$$
Choose any $\bm{x}_0\in \partial \bar{S}_\delta \cap \mathcal{M}$, we also have that $\bm{x}_0\notin \bar{S}_\delta^c$, which implies that
$$R_{\delta/4}(\bm{x}_0,\bm{y})=0,\quad \forall \bm{y}\in P.$$
It follows that
$$I_n(R_{\delta/4}(\bm{x}_0,\cdot))=0.$$
On the other hand, $I(R_{\delta/4}(\bm{x}_0,\cdot))=O(1)$. Using Theorem \ref{thm:entropy-bound}, we know that the probability is less than $1/(2n)$, which proves that
$P$ is $S$-connected with probability at least $1-1/(2n)$.
So far, we have proved Theorem \ref{thm:maximum-principle}.
\section{Error Estimate (Theorem \ref{thm:error})}
\label{sec:convergence}
Let $e_\delta(\bm{x})=u_\delta(\bm{x})-u(\bm{x})$. $u_\delta$ and $u$ solve \eqref{eq:wgl} and \eqref{eq:laplace-large} respectively.
Direct calculation shows that
\begin{align}
\label{eq:error-large}
L_{\delta,n} e_\delta(\bm{x})&=\sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y})(u(\bm{x})-u(\bm{y}))+\mu\sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y})(u(\bm{x})-b(\bm{y}))\mathrm{d} \bm{y},
\quad \bm{x}\in P,\\
\label{eq:error-large-2}
e_\delta(\bm{x})&=0,\quad \bm{x}\in S.
\end{align}
Next, we will find an upper bound of the right hand side in \eqref{eq:error-large}.
An upper bound of the second term of \eqref{eq:error-large} is relatively easy to find by using the smoothness of $u$ and $b$:
\begin{align}
\label{eq:error-2}
\left|\sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y})(u(\bm{x})-b(\bm{y}))\right|\le C\delta \sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y})
\end{align}
To find an upper bound of the first term, we need the following theorem which can be found in \cite{Shi-iso}.
\begin{theorem}
Let $u(\bm{x})\in C^3(\mathcal{M})$ and
\begin{align}
\label{eq:error_boundary}
I_{bd} =\sum_{j=1}^d \int_{\partial\mathcal{M}}n^j(\bm{y})(\bm{x}-\bm{y})\cdot\nabla(\nabla^ju(\bm{y}))\bar{R}_t(\bx, \by)\mathrm{d} \tau_{\bm{y}},
\end{align}
and
\begin{align}
I_{in}=\frac{1}{\delta^2}\int_{\mathcal{M}} R_\delta(\bm{x},\bm{y})(u(\bm{x})-u(\bm{y}))\mathrm{d} \bm{y}+\int_\mathcal{M} \bar{R}_\delta(\bm{x},\bm{y}) \Delta_\mathcal{M} u(\bm{y})\mathrm{d} \bm{y}-\int_{\partial\mathcal{M}} \bar{R}_\delta(\bm{x},\bm{y})\frac{\partial u}{\partial \mathbf{n}}(\bm{y})
\mathrm{d} \tau_{\bm{y}}-I_{bd}\nonumber.
\end{align}
where $\mathbf{n}(\bm{y})=(n^1(\bm{y}),\cdots,n^d(\bm{y}))$ is the out normal vector of $\partial\mathcal{M}$ at $\bm{y}$, $\nabla^j$ is the $j$th component of gradient $\nabla$,
$\bar{R}_\delta(\bm{x},\bm{y})=C_\delta\bar{R}\left(\frac{|\bm{x}-\bm{y}|^2}{4\delta^2}\right)$ and $\bar{R}(r)=\int_{r}^{\infty}R(s)\mathrm{d} s$.
Then there exist constants $C, T_0$ depending only on $\mathcal{M}$ and $p(\bm{x})$, so that,
\begin{eqnarray}
\label{eqn:integral_error_int}
\left|I_{in}\right|\le C\delta\|u\|_{C^3(\mathcal{M})},\quad
\end{eqnarray}
as long as $\delta\le T_0$.
\label{thm:integral_error}
\end{theorem}
According to the above theorem, we have
\begin{align}
\frac{1}{\delta^2}\left|\int_{\mathcal{M}} R_\delta(\bm{x},\bm{y})(u(\bm{x})-u(\bm{y}))\mathrm{d} \bm{y}\right|\le C\delta,\quad \bm{x}\in \mathcal{M}\backslash \mathcal{D}_\delta,
\end{align}
where $\mathcal{D}_\delta=\{\bm{x}\in \mathcal{M}: \text{dist}(\bm{x},\mathcal{D})\le 2\delta\}$. Notice that for $\bm{x}\in \mathcal{M}\backslash \mathcal{D}$, $R_\delta(\bm{x},\bm{y})$ has no intersection with
$\mathcal{D}$, so all boundary terms vanish.
To get an upper bound of the first term in \eqref{eq:error-large}, we need to estimate the difference between $\int_{\mathcal{M}} R_\delta(\bm{x},\bm{y})(u(\bm{x})-u(\bm{y}))\mathrm{d} \bm{y}$
and $\sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y})(u(\bm{x})-u(\bm{y}))$. This is given by the following theorem.
\begin{theorem}
\label{thm:entropy-bound-2}
With probability at least $1-1/(2n)$, $n=|P|$,
\begin{align}
\sup_{f\in \bar{\mathcal{R}}_\delta}|I(f)-I_n(f)|\le \frac{C}{\delta^k\sqrt{n}}\left(\ln n-2\ln \delta +1\right)^{1/2},
\end{align}
where $k$ is the dimension of $\mathcal{M}$,
$$I(f)=\frac{1}{|\mathcal{M}|}\int_\mathcal{M} f(\bm{x})\mathrm{d} \bm{x},\quad I_n(f)=\frac{1}{n}\sum_{\bm{x}\in P}f(\bm{x}),$$
$|\mathcal{M}|$ is the volume of $\mathcal{M}$ and $\mathcal{R}_\delta$ is a function class defined as
$$\bar{\mathcal{R}}_\delta=\{R_\delta(\bm{x},\cdot), R_\delta(\bm{x},\cdot)u(\cdot), R_\delta(\bm{x},\cdot)v(\cdot): \bm{x}\in \mathcal{M},\; u \;and\; v\; \text{solves \eqref{eq:laplace-large} and
\eqref{eq:laplace-large-ref} respectively.}\}$$
\end{theorem}
This theorem will be proved in Section \ref{sec:entropy} using the empirical process theory \cite{entropy}.
Using Theorem \ref{thm:entropy-bound-2}, we have
\begin{align}
\label{eq:error-1-1}
\left|\frac{1}{n}\sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y})(u(\bm{x})-u(\bm{y}))\right|\le \frac{C}{\delta^k\sqrt{n}}(\ln n-2\ln \delta +1)^{1/2}+C\delta^3,\quad \bm{x}\in P\cap (\mathcal{M}\backslash \mathcal{D}_\delta).
\end{align}
For $\bm{x}\in P\cap \mathcal{D}_\delta$, the bound is straightforward, just using the smoothness of $u$,
\begin{align}
\label{eq:error-1-2}
\left|\frac{1}{n}\sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y})(u(\bm{x})-u(\bm{y}))\right|\le \frac{C\delta}{n} \sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y}),\quad \bm{x}\in P\cap \mathcal{D}_\delta.
\end{align}
Substituting \eqref{eq:error-2}, \eqref{eq:error-1-1} and \eqref{eq:error-1-2} in \eqref{eq:error-large}, we have
\begin{align}
\label{eq:error-D}
| L_{\delta.n} e_\delta(\bm{x})|\le C\delta \left(\frac{1}{n}\sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y})\right)+C\delta\left(\frac{\mu}{n} \sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y})\right)
, \quad \bm{x}\in P\cap \mathcal{D}_\delta.
\end{align}
and
\begin{align}
| L_{\delta.n} e_\delta(\bm{x})|\le \frac{C}{\delta^k\sqrt{n}}(\ln n-2\ln \delta +1)^{1/2}+C\delta^3+C\delta\left(\frac{\mu}{n} \sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y})\right),
\quad \bm{x}\in P\cap (\mathcal{M}\backslash \mathcal{D}_\delta).\nonumber
\end{align}
Suppose the number of sample points $n$ is large enough such that
\begin{equation}
\label{eq:p-large}
\frac{C}{\delta^k\sqrt{n}}(\ln n-2\ln \delta +1)^{1/2}\le \delta^3,
\end{equation}
then we have
\begin{align}
\label{eq:error-P}
| L_{\delta.n} e_\delta(\bm{x})|\le C\delta^3+C\delta \left(\frac{\mu}{n}\sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y})\right), \quad \bm{x}\in P\cap (\mathcal{M}\backslash \mathcal{D}_\delta).
\end{align}
\eqref{eq:error-D} and \eqref{eq:error-P} give an upper bound for $| L_{\delta.n} e_\delta(\bm{x})|$.
Next, we want to get a lower bound of $ L_{\delta.n} v(\bm{x})$ with $v$ given in \eqref{eq:laplace-large-ref}.
By Theorem \ref{thm:integral_error} and \eqref{eq:laplace-large-ref}, we have
\begin{align}
\frac{1}{\delta^2}\int_{\mathcal{M}} R_\delta(\bm{x},\bm{y})(v(\bm{x})-v(\bm{y}))\mathrm{d} \bm{y}\ge \int_{\mathcal{M}} \bar{R}_\delta(\bm{x},\bm{y})\mathrm{d} \bm{y} -C\delta,
\quad \bm{x}\in \mathcal{M}\backslash \mathcal{D}_\delta.
\end{align}
Also using Theorem \ref{thm:entropy-bound-2}
\begin{align}
&\frac{1}{n}\sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y})(v(\bm{x})-v(\bm{y}))\nonumber\\
\ge& \; \bar{w}_\delta\delta^2 -C\delta^3-\frac{C}{\delta^k\sqrt{n}}(\ln n-2\ln \delta +1)^{1/2}
\ge \bar{w}_\delta\delta^2/2
,\quad \bm{x}\in P\cap (\mathcal{M}\backslash \mathcal{D}_\delta).
\label{eq:error-v1-1}
\end{align}
with $\displaystyle\bar{w}_\delta=\min_{\bm{x}\in \mathcal{M}\backslash \mathcal{D}_\delta}\frac{1}{|\mathcal{M}|}\int_{\mathcal{M}} \bar{R}_\delta(\bm{x},\bm{y})\mathrm{d} \bm{y}$. Here, we also use the assumption that $n$ is
large enough, \eqref{eq:p-large}.
In $P\cap \mathcal{D}_\delta$, we have
\begin{align}
\label{eq:error-v1-2}
\frac{1}{n}\sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y})(v(\bm{x})-v(\bm{y}))\ge &-\frac{C\delta}{n} \sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y}),\quad \bm{x}\in P\cap \mathcal{D}_\delta,
\end{align}
this is due to the smoothness of $v$.
Also notice that
\begin{align}
\label{eq:error-v2}
\frac{1}{n}\sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y})v(\bm{x})= \frac{v(\bm{x})}{n} \sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y}) \ge \frac{1}{n}\sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y}).
\end{align}
Combining \eqref{eq:error-v1-1}, \eqref{eq:error-v1-2} and \eqref{eq:error-v2}, we obtain
\begin{align}
\label{eq:error-v-D}
L_{\delta.n} v(\bm{x})\ge \frac{\mu}{n}\sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y})-\frac{C\delta}{n} \sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y}), \quad \bm{x}\in P\cap \mathcal{D}_\delta.
\end{align}
and
\begin{align}
\label{eq:error-v-P}
L_{\delta,n} v(\bm{x})\ge&\; \frac{\bar{w}_\delta}{2}\delta^2
+\frac{\mu}{n}\sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y}), \quad \bm{x}\in P\cap (\mathcal{M}\backslash \mathcal{D}_\delta).
\end{align}
Comparing \eqref{eq:error-P} and \eqref{eq:error-v-P}, we have
\begin{equation}
|L_{\delta,n} e_\delta(\bm{x})|\le C \delta L_{\delta.n} v(\bm{x}),\quad \bm{x}\in P\cap (\mathcal{M}\backslash \mathcal{D}_\delta).
\end{equation}
Meanwhile, \eqref{eq:error-D} and \eqref{eq:error-v-D} show that
\begin{equation}
|L_{\delta,n} e_\delta(\bm{x})|\le C \delta L_{\delta.n} v(\bm{x}),\quad \bm{x}\in P\cap\mathcal{D}_\delta,
\end{equation}
as long as
\begin{equation}
\label{eq:condition-mu}
\frac{\mu}{n}\sum_{\bm{y}\in S} K_\delta(\bm{x}, \bm{y})\ge \frac{C}{n} \sum_{\bm{y}\in P} R_\delta(\bm{x},\bm{y}),\quad \bm{x}\in P\cap\mathcal{D}_\delta.
\end{equation}
The proof of Theorem \ref{thm:error} is completed
\section{Entropy bound}
\label{sec:entropy}
In this section, we will prove Theorem \ref{thm:entropy-bound} and \ref{thm:entropy-bound-2}. The method we use is to estimate the covering number of the function classes.
First we introduce the definition of the covering number.
Let $(Y, d)$ be a metric space and set $F\subset Y$ . For every $\epsilon>0$, denote by $N (\epsilon, F, d)$ the
minimal number of open balls (with respect to the metric $d$) that are needed to cover $F$. That is,
the minimal cardinality of the set $\{y_1 , \cdots, y_m\}\subset Y$ with the property that every $f \in F$
has some $y_i$ such that $d(f, y_i ) < \epsilon$. The set $\{y_1 , \cdots, y_m \}$ is called an $\epsilon$-cover of $F$ . The
logarithm of the covering numbers is called the entropy of the set.
For every sample $\{x_1 , \cdots, x_n \}$,
let $\mu_n$ be the empirical measure supported on that sample. For $1 \le p <\infty$ and a
function $f$ , put $\|f\|_{L_p (\mu_n )} =\left(\frac{1}{n}\sum_{i=1}^n |f (x_i )|^p\right)^{1/p}$
and set $\|f\|_\infty = \max_{1\le i\le n} |f (x_i )|$. Let
$N(\epsilon , F, L_p (\mu_n )$ be the covering numbers of $F$ at scale $\epsilon$ with respect to the $L_p (\mu_n )$ norm.
We will use the following theorem which is well known in empirical process theory.
\begin{theorem}(Theorem 2.3 in \cite{entropy})
\label{thm:entropy}
Let $F$ be a class of functions from $\mathcal{M}$ to $[-1,1]$ and set $\mu$ to be a probability measure on $\mathcal{M}$.
Let $(\bm{x}_i)_{i=1}^\infty$ be independent
random variables distributed according to $\mu$. For any $\epsilon>0$ and every $n\ge 8/\epsilon^2$,
\begin{eqnarray}
\mathbb{P}\left(\sup_{f\in F}|\frac{1}{n}\sum_{i=1}^nf(\bm{x}_i)-\int_\mathcal{M} f(\bm{x})\mu(\bm{x})\mathrm{d} \bm{x}|>\epsilon\right)
\le 8\mathbb{E}_\mu[ N(\epsilon/8,F,L_1(\mu_n))]\exp(-n\epsilon^2/128)
\end{eqnarray}
\end{theorem}
Note that
\begin{align*}
L_1(\mu_n)\le L_\infty(\mu_n)\le L_\infty
\end{align*}
where $\|f\|_{L_\infty}=\max_{\bm{x}\in\mathcal{M}}|f(\bm{x})|$. Then we get following corollary.
\begin{corollary}
\label{cor:entropy-0}
Let $F$ be a class of functions from $\mathcal{M}$ to $[-1,1]$ and set $\mu$ to be a probability measure on $\mathcal{M}$.
Let $(\bm{x}_i)_{i=1}^\infty$ be independent
random variables distributed according to $\mu$. For any $\epsilon>0$ and every $n\ge 8/\epsilon^2$,
\begin{eqnarray}
\mathbb{P}\left(\sup_{f\in F}|\frac{1}{n}\sum_{i=1}^n f(\bm{x}_i)-\int_\mathcal{M} f(\bm{x})\mu(\bm{x})\mathrm{d} \bm{x}|>\epsilon\right)
\le 8 N(\epsilon/8,F,L_\infty)\exp(-n\epsilon^2/128)
\end{eqnarray}
where $N(\epsilon , F, L_\infty)$ is the covering numbers of $F$ at scale $\epsilon$ with respect to the $L_\infty$ norm
\end{corollary}
\begin{corollary}
\label{cor:entropy}
Let $F$ be a class of functions from $\mathcal{M}$ to $[-1,1]$. Let $(\bm{x}_i)_{i=1}^\infty$ be independent
random variables distributed according to $p$, where $p$ is the probability distribution. Then
with probability at least $1-\delta$, we have
\begin{eqnarray}
\sup_{f\in F}|p(f)-p_n(f)|\le \sqrt{\frac{128}{n}\left(\ln N(\sqrt{\frac{2}{n}},F,L_\infty)
+\ln \frac{8}{\delta}\right)},\nonumber
\end{eqnarray}
where
\begin{align}
\label{eq:note-mc}
p(f)=\int_\mathcal{M} f(\bm{x})p(\bm{x})\mathrm{d} \bm{x},\quad p_n(f)=\frac{1}{n}\sum_{i=1}^nf(\bm{x}_i).
\end{align}
\end{corollary}
\begin{proof}
Using Corollary \ref{cor:entropy-0}, with probability at least $1-\delta$,
\begin{eqnarray}
\sup_{f\in F}|p(f)-p_n(f)|\le \epsilon_\delta,\nonumber
\end{eqnarray}
where $\epsilon_\delta$ is determined by
\begin{eqnarray}
\epsilon_\delta=\sqrt{\frac{128}{n}\left(\ln N(\epsilon_\delta/8,F,L_\infty)+\ln \frac{8}{\delta}\right)}.\nonumber
\end{eqnarray}
Obviously,
\begin{align*}
\epsilon_\delta\ge \sqrt{\frac{128}{n}}=8\sqrt{\frac{2}{n}}
\end{align*}
which gives that
\begin{align*}
N(\epsilon_\delta/8,F,L_\infty)\le N(\sqrt{\frac{2}{n}},F,L_\infty)
\end{align*}
Then, we have
\begin{eqnarray}
\epsilon_\delta \le \sqrt{\frac{128}{n}\left(\ln N(\sqrt{\frac{2}{n}},F,L_\infty)
+\ln \frac{8}{\delta}\right)}\nonumber
\end{eqnarray}
which proves the corollary.
\end{proof}
The above corollaries provide a tool to estimate the integral error on random samples. To apply the above corollaries in our problem, the key point is to obtain the estimates of the covering number of function class $\mathcal{R}_\delta$.
Since the kernel $R\in C^1(\mathcal{M})$ and $\mathcal{M}\in C^\infty$, we have
for any $\bm{x},\bm{y}\in \mathcal{M}$
\begin{eqnarray}
|R\left(\frac{\|\bm{x}-\bm{y}\|^2}{4\delta^2}\right)-R\left(\frac{\|\bm{z}-\bm{y}\|^2}{4\delta^2}\right)|\le \frac{C}{\delta}\|\bm{x}-\bm{z}\|.\nonumber
\end{eqnarray}
This gives an easy bound of $N(\epsilon, \mathcal{R}_\delta,L_\infty)$,
\begin{eqnarray}
\label{eq:bound-R}
N(\epsilon, \mathcal{R}_\delta,L_\infty)\le \left(\frac{C}{\epsilon\delta}\right)^k
\end{eqnarray}
Using Corollary \ref{cor:entropy}, with probability at least $1-1/(2n)$,
\begin{align}
\label{eq:bound-w-prob}
\sup_{f\in \mathcal{R}_\delta}|p(f)-p_n(f)|\le \frac{C}{\delta^k \sqrt{n}}\left(\ln n-2\ln \delta +1\right)^{1/2}
\end{align}
This proves Theorem \ref{thm:entropy-bound}. Theorem \ref{thm:entropy-bound-2} can be proved similarly using the fact that $u$ (solution of \eqref{eq:laplace-large})
and $v$ (solution of \eqref{eq:laplace-large-ref}) are both smooth.
\section{Discussion and Future Works}
\label{sec:concl}
In this paper, we analyzed convergence of the weighted nonlocal Laplacian (WNLL) on random point cloud. The analysis reveals that the weight is very important in the convergence and it should have the same order as $|P|/|S|$, i.e. $\mu\sim |P|/|S|$. The result in this paper provides the WNLL a solid theoretical foundation.
Furthermore, our analysis also shows that the convergence may fail with extremely low labeling rate. As discussed in Section \ref{sec:convergence}, in this case, we should consider other approaches. One interesting option is to minimize $L_\infty$ norm of the gradient instead of the $L_2$ norm, i.e. to solve the following optimization problem
\begin{align*}
\min_u \left(\max_{\bm{x}\in P\cup S} \left(\sum_{\bm{y}\in P\cup S}w(\bm{x},\bm{y})(u(\bm{x})-u(\bm{y}))^2\right)^{1/2}\right),
\end{align*}
with the constraint
\begin{align*}
u(\bm{x})=b(\bm{x}),\quad \bm{x}\in S.
\end{align*}
This approach is closely related to the infinity Laplacian \cite{GEL11,EDLL14}. The above optimization problem can be solved by the split Bregman iteration. An interesting observation
is that the WNLL can accelerate convergence of the split Bregman iteration and improve efficiency. This will be further explore in our future work.
\vspace{0.2in}
\noindent
\textbf{Acknowledgment.}
This material is based, in part, upon work supported by the U.S. Department of Energy, Office of Science and by National Science Foundation, and National Science Foundation of China, under Grant Numbers DOE-SC0013838 and DMS-1554564, (STROBE), NSFC 11671005.
\newpage
|
\section{Introduction}
Several problems in graph theory and combinatorial optimization involve determining if a given graph $G$ has a subgraph with certain properties.
Examples include seeking paths, cycles, trees, dominating sets, cliques, vertex covers, matching, independent sets, bipartite subgraphs, etc.
Related optimization problems include finding a maximum clique, a maximum (connected) vertex cover, a maximum independent set, a minimum (connected) dominating set, etc.
These well-studied problems have significant theoretical interest and many practical applications.
In this paper, we consider the problem in which we are given a simple connected graph $G=(V,E)$ whose vertex set $V$ has each node being ``red'' or ``blue''
(note, the color assignment might not be a proper $2$-coloring of the vertices, i.e., we allow nodes of the same color to be adjacent in $G$).
We seek a maximum-cardinality subset $V'\subseteq V$ of the nodes such that $V'$ is {\em color-balanced}, i.e. having same number of red and blue nodes in $V'$, and such that the induced subgraph $H$ by $V'$ in $G$ is connected. We refer to this problem as the \colb{Balanced Connected Subgraph ({\textit{BCS}})} problem:
\begin{framed}
\noindent {\bf \textit{{\emph{Balanced Connected Subgraph (\textit{BCS}) Problem}}}}\\
{\bf Input:} A graph $G=(V,E)$, with node set $V=V_R\cup V_B$ partitioned into red nodes ($V_R$) and blue nodes ($V_B$).\\
{\bf Goal:} Find a maximum-cardinality color-balanced subset $V'\subseteq V$ that induces a connected subgraph $H$.
\end{framed}
Notice that, the {\textit{BCS}}~problem is a special case of the \colb{Maximum Node Weight Connected Subgraph ({\textit{MNWCS}})} problem \cite{Johnson1985}.
In the {\textit{MNWCS}}~problem, we are given a connected graph $G(V,E)$, with a (possibly negative) integer weight $w(v)$ associated with each node $v\in V$, and an integer bound $B$; the objective is to decide whether there exists a subset $V'\subseteq V$ such that the subgraph induced by $V'$ is connected and the total weight of the vertices in $V'$ is at least $B$. In the {\textit{MNWCS}}~problem, if we assign the weight of each vertex is either $+1$ (red) or $-1$ (blue), then deciding whether there exist a $V'\subseteq V$ such that $|V'| \geq k$, subgraph induced by $V'$ in $G$ is connected and total of vertices in $V'$ is exactly zero
is equivalent as the {\textit{BCS}}~problem. The {\textit{MNWCS}}~problem along with its variations have numerous practical application in various fields. This includes designing fiber-optic networks \cite{Felix1998}, oil-drilling \cite{Hochbaum1994}, systems biology \cite{Backes2012,Dittrich2008,Yamamoto2009}, wildlife corridor design \cite{Dilkina2010}, computer vision \cite{Chen2017,Chen2012}, forest planning \cite{Carvajal2013}, and many more (see \cite{El-Kebir2014} and the references therein).
Some of these applications are best suited to the {\textit{BCS}}~problem.
\subsection{Related work}
The bichromatic inputs, often referred in the literature as red-blue input, has appeared extensively in numerous
problems. For bipartite trees, see \cite{abellanas1999bipartite}. In \cite{biniaz2014bottleneck,dumitrescu2001matching,dumitrescu2002partitioning}
colored points have been considered in the context of matching and partitioning problem. For a detailed survey on geometric problems with red-blue points; see \cite{kaneko2003discrete}.
In \cite{aichholzer2015balanced}, Aichholzer et al. considered the balanced island problem and devised polynomial algorithms for points considered on plane. From combinatorial side, Balanchandran et al. \cite{balachandran2017system} studied the problem of unbiased representatives in a set of bicolorings. In this paper, they have mentioned the usefulness of the unbiased representatives in drug testing. While the drugs are tested over a large population, the effectiveness of a new drug is measured under various attributes e.g., weight, height, age etc. One would require to sample representative in certain \emph{balanced} manner. Kaneko et al. \cite{kaneko2017balancing} considered the problem of balancing the colored points on the line. Subsequently, Bereg et al. \cite{bereg2015balanced} studied the balanced partitions of $3$-colored geometric sets on the plane.
On the other hand, finding a certain type of subgraph in a graph is considered to be a fundamental algorithmic question. In \cite{feige2001dense}, Feige et al. studied the dense $k$-subgraph problem where given a graph $G$ and a parameter $k$, the goal is to find a set of $k$ vertices with maximum average degree in the subgraph induced by this set. From parameterized algorithms side, Crowston et al. \cite{crowston2013maximum} considered the balanced subgraph problem. Kierstead et al. \cite{kierstead1992colorful} studied the problem of finding colorful induced subgraph in a properly colored graph. This led us to study the balanced connected subgraph problem on graphs.
In \cite{derhy2009finding}, Derhy and Picouleau considered the problem of finding induced trees on both weighted and unweighted graphs and obtained hardness and algorithmic results.
They have studied some particular classes of graphs like the bipartite graphs or the triangle-free graphs. Moreover, they have considered the case where the number of prescribed vertices is bounded.
\subsection{Our contributions}
In this paper, we consider the balanced connected subgraph problem on various graph families and present several hardness and algorithmic results.
On the hardness side, in Section \ref{np-hardness}, we prove that the {\textit{BCS}}~problem is {NP}-hard on general graphs, even for planar graphs, bipartite graphs (with a general red/blue color assignment, not necessarily a proper 2-coloring), and chordal graphs.
Furthermore, we show that the existence of a balanced connected subgraph containing a specific vertex is {NP}-complete. In addition to that, we prove that finding the maximum balanced path in a graph is {NP}-hard.
On the algorithmic side, in Section \ref{algorithmicresults}, we devise polynomial-time algorithms
for trees (in $O(n^4)$ time), split graphs (in $O(n^2)$ time), bipartite graphs with a proper 2-coloring (in $O(n^2)$ time), and graphs with diameter~$2$ (in $O(n^2)$ time). Here $n$ is the number of vertices in the input graphs.
\section{Hardness results}\label{np-hardness}
\subsection{{\textit{BCS}}~problem}\label{NPBCP}
In this section we prove that the {\textit{BCS}}~problem is {NP}-hard for bipartite graph with a general red/blue color assignment, not necessarily a proper 2-coloring. We give a reduction from the \colb{Exact-Cover-by-3-Sets ({\it EC3Set})} problem \cite{Garey1979}.
In this {\it EC3Set}~problem, we are given a set $U$ with $3k$ elements and a collection $S$ of $m$ subsets of $U$ such that each $s_i\in S$ contains exactly $3$ elements.
The objective is to find an exact cover for $U$ (if exists), i.e., a sub-collection $S'\subseteq S$ such that every element of $U$ occurs in exactly one member of $S'$. During the reduction, we generate an instance $G=(R\cup B,E)$ of {\textit{BCS}}~problem from an instance $X(S,U)$ of the {\it EC3Set}~problem as follows:
\noindent {\bf Reduction:} For each set $s_i\in S$, we take a blue vertex $s_i \in B$. For each element $u_j\in U$, we take a red vertex $u_j \in R$. Now consider a set $s_i \in S$ which contains three elements $u_\alpha, u_\beta$, and $u_\gamma$, then we add 3 edges $(s_i,u_\alpha)$, $(s_i,u_\beta)$, and $(s_i,u_\gamma)$ in $E$. Additionally, we consider a path of $5k$ blue vertices starting and ending with vertices $b_1$ and $b_{5k}$ respectively. Similarly, we consider a path of $3k$ red vertices starting and ending with vertices $r_1$ and $r_{3k}$ respectively. We connect these two paths by joining the vertices $r_{3k}$ and $b_1$ by an edge. Finally, we connect each vertices $s_i$ with $b_{5k}$ by edges. This completes the construction. See Figure \ref{fig:my_label} for the complete construction.
\begin{figure}[ht]
\centering
\includegraphics[scale=.6]{balance1.eps}
\caption{Construction of the instance $G$ of the {\textit{BCS}}~problem.}\label{fig:my_label}
\end{figure}
Clearly, the number of vertices and edges in $G$ are polynomial in terms of number of elements and sets in $X$. Hence the construction can be done in polynomial time. We now prove the following theorem.
\begin{lemma}\label{lem-bcs-np-hard}
The instance $X$ of the {\it EC3Set}~problem has a solution if and only if the instance $G$ of the {\textit{BCS}}~problem has a connected balanced subgraph $T$ with $12k$ vertices ($6k$ red and $6k$ blue).
\end{lemma}
\begin{proof}
Assume that {\it EC3Set}~problem has a solution. Let $S^*$ be an optimal solution in it. We choose the corresponding vertices of $S^*$ in $T$. Since this solution covers all $u_j$'s. So we select all $u_j$'s in $T$. Finally we select all the $5k$ blue and $3k$ red vertices in $T$, resulting in a total of $6k$ red and $6k$ blue vertices.
On the other hand, assume that there is a balanced tree $T$ in $G$ with $6k$ vertices of each color. The solution must pick the $5k$ blue vertices $b_1, \ldots, b_{5k}$. Otherwise, it exclude the $3k$ red vertices $r_1, \ldots,r_{3k}$, and reducing the size of the solution. Since the graph $G$ has at most $6k$ red vertices, at most $k$ vertices can be picked from the set $s_1, \ldots,s_m$ and need to cover all the $3k$ red vertices corresponding to $u_j$ for $1\leq j\leq 3k$. Hence, this $k$ sets give an exact cover. \qed
\end{proof}
It is easy to see that the graph we constructed from the \colb{Exact-Cover-by-3-Sets ({\it EC3Set})} problem in Figure \ref{fig:my_label} is indeed a bipartite graph. Hence we have the following theorem.
\begin{theorem}\label{thm-bcs-np-hard-bipartite}
{\textit{BCS}}~problem is {NP}-hard for bipartite graphs.
\end{theorem}
\subsection{{NP}-hardness: {\textit{BCS}}~problem over special classes of graphs}
In this section, we show that the {\textit{BCS}}~problem is {NP}-hard even if we restrict the graph classes to chordal, or planar graphs.
\subsubsection{Chordal graphs:}
We prove that the {\textit{BCS}}~problem is {NP}-hard where the input graph is a chordal graph. The hardness construction is similar to the construction in Section~\ref{NPBCP}; we modify the construction so that the graph is chordal. In particular, we add edges between $s_{i}$ and $s_{j}$ for each $i \neq j, 1 \leq i,j \leq m$. For this modified graph, it is easy to see that a lemma identical to Lemma~\ref{lem-bcs-np-hard} holds. Hence, we conclude that the {\textit{BCS}}~ problem is {NP}-hard for chordal graphs.
\subsubsection{Planar graphs:}
In this section we prove that {\textit{BCS}}~problem is {NP}-hard for planar graphs. We give a reduction from the Steiner tree problem in planar graphs ({\it STPG}) \cite{Garey1979}. In this problem, we are given a planar graph $G = (V,E)$, a subset $X \subseteq V$, and a positive integer $k \in \mathbb{N}$. The objective is to find a tree $T =(V',E')$ with at most $k$ edges such that $ X \subseteq V'$.
\vspace{.2cm}
\noindent {\it Reduction:} We generate an instance $H=(R \cup B,E(H))$ for the {\textit{BCS}}~problem from an instance $G = (V,E)$ of the {\it STPG}~problem. We color all the vertices in $G$ as blue.
We now create red color vertices and connect to these vertices. For each vertex $u_i \in X$, we add a vertex $u'_i$ in $H$ whose color is red add connect $u'_i$ to $u_i$ via an edge. Additionally, we take a set $Z$ of $(k+1-|X|)$ red vertices in $H$ and the edges $(z_{j},u'_1)$ into $E(H)$, for each $z_{j} \in Z$. Hence we have, $B =V$, and $R= Z \cup \{u'_i\} | \ 1 \leq i \leq |X|\} $. Note that $|R|<|B|$ and $ |R|=(k+1) $. This completes the construction. For an illustration see Figure \ref{fig:planar}.
\begin{figure}[ht]
\centering
\includegraphics[scale=.6]{balance4.eps}
\caption{Schematic construction for planar graphs.}
\label{fig:planar}
\end{figure}
Clearly the number of vertices and edges in $H$ are polynomial in terms of vertices in $G$. Hence the construction can be done in polynomial time. We now prove the following theorem.
\begin{theorem}\label{theoBCS}
{\it STPG}~has a solution if and only if $H$ of the {\textit{BCS}}~problem has a balanced connected subgraph with $(k+1)$ vertices of each color.
\end{theorem}
\begin{proof}
Assume that {\it STPG}~has a solution. Let $T=(V',E')$ be the resulting Steiner tree which contains at most $k$ edges and $X\subseteq V'$. If $ |V'|=(k+1) $ then the subgraph of $H$ induced by $(V' \cup R)$ is connected and balanced with $(k+1)$ vertices of each color. If $ |V'| < (k+1) $ then we take a set $ Y $ of $((k+1)-|V'|)$ many vertices from $ V $ such that the subgraph of $G$ induced by $(V' \cup Y)$ is connected. Clearly $ |V'|=(k+1) $. Now the subgraph of $H$ induced by $(V' \cup Y \cup R)$ is connected and balanced with $(k+1)$ vertices of each red and blue color.
On the other hand, assume that there is a balanced connected subgraph $H'$ of $H$ with $(k+1)$ vertices of each color. Note that, except vertex $u'_1$, in $H$ all the red vertices are of degree $1$ and connected to blue vertices. Let $G'$ be the subgraph of $G$ induced by all blue vertices in $H'$. Since $H$ is connected and there is no edge between any two red vertices, $G'$ is connected. Since $G'$ contains $(k+1)$ vertices, any spanning tree $T$ of $H'$ contains $k$ edges. So $T$ is a solution of {\it STPG}~problem. \qed
\end{proof}
Hence we have the following theorem.
\begin{theorem}\label{thm-bcs-np-hard}
{\textit{BCS}}~problem is {NP}-hard for planar graphs.
\end{theorem}
%
%
%
%
%
%
\subsection{NP-completeness for {\textit{BCS}}~problem for a specific vertex.}
In this section we prove that the existence of a balanced subgraph containing a specific vertex is {NP}-complete. We call this problem the {\textit{BCS}}-existence problem. The reduction is similar to the reduction used in showing the {NP}-hardness of the {\textit{BCS}}~problem; we also use here a reduction from the {\it EC3Set}~problem (see Section \ref{NPBCP} for the definition).
\noindent {\bf Reduction:} Assume that we are given a {\it EC3Set}~problem instance $X=(U,S)$, where set $U$ contains $3k$ elements and a collection $S$ of $m$ subsets of $U$ such that each $s_i\in S$ contains exactly $3$ elements. We generate an instance $G(R,B,E)$ of the {\textit{BCS}}-existence problem from $X$ as follows. The red vertices $R$ are the elements $u_j\in U$; i.e., $R=U$. The blue vertices $B$ are the $3$-element sets $s_i\in S$; i.e., $B=S$.
For each blue vertex $s_i=\{u_\alpha,u_\beta,u_\gamma\} \in S=B$, we add the 3 edges $(s_i,u_\alpha)$, $(s_i,u_\beta)$, and $(s_i,u_\gamma)$ to the set $E$ of edges of $G$.
%
%
We instantiate an additional set of $2k$ blue vertices, $\{b_1,\ldots,b_{2k}\}$, and add edges to $E$ to link them into a path $(b_1,b_2,\ldots,b_{2k})$. Finally, we add an edge from $b_{2k}$ to each of the blue vertices $s_i$. Refer to Figure \ref{fig:vertex}.
\begin{figure}[ht]
\centering
\includegraphics[scale=.6]{balance3.eps}
\caption{Construction of the instance $G$ of the {\textit{BCS}}~problem containing $ b_{1}$.}\label{fig:vertex}
\end{figure}
Clearly, the number of vertices and edges in $G$ are polynomial in terms of number of elements and sets in the size of the {\it EC3Set}~problem instance $X$, and the construction can be done in polynomial time. We now prove the following lemma.
\begin{lemma}\label{lemma-bcs-np-hard}
The instance $X$ of the {\it EC3Set}~problem has a solution iff the instance $G$ of the corresponding {\textit{BCS}}~existence problem has a balanced subgraph $T$ containing the vertex $b_1$.
\end{lemma}
\begin{proof}
Assume that the {\it EC3Set}~problem has a solution, and let $S^*$ be the collection of $k=|S^*|$ sets of $S$ in the solution.
%
%
Then, we obtain a balanced subgraph $T$ that contains $b_1$ as follows: $T$ is the induced subgraph of the $3k$ red vertices $U$, together with the $k$ blue vertices $S^*$ and the $2k$ blue vertices $b_1,\ldots,b_{2k}$. Note that $T$ is balanced and connected and contains $b_1$.
Conversely, assume there is a balanced connected subgraph $T$ containing $b_{1}$. Let $t$ be the number of (blue) vertices of $S$ within $T$. First, note that $t\leq k$. (Since $T$ is balanced and contains at most $3k$ red vertices, it must contain at most $3k$ blue vertices, $2k$ of which must be $\{b_1,\ldots,b_{2k}\}$, in order that $T$ is connected.)
%
%
Next, we claim that, in fact, $t\geq k$. To see this, note that each of the $t$ blue vertices of $T$ that corresponds to a set in $S$ is connected by edges to $3$ red vertices; thus, $T$ has at most $3t$ red vertices. Now, $T$ has $2k+t$ blue vertices (since it has $t$ vertices other than the path $(b_1,\ldots,b_{2k})$), and $T$ is balanced; thus, $T$ has exactly $2k+t$ red vertices, and we conclude that $2k+t\leq 3t$, implying $k\leq t$, as claimed.
%
%
Therefore, we need to select exactly $k$ blue vertices corresponding to the sets $S$, and these vertices connect to all $3k$ of the red vertices. The $k$ sets corresponding to these $k$ blue vertices is a solution for the {\it EC3Set}~problem. \qed
\end{proof}
It is easy to prove that the {\textit{BCS}}~existence problem is in {NP}. Hence, we have the following theorem.
\begin{theorem}
It is {NP}-complete to decide if there exists a connected balanced subgraph that contains a specific vertex.
\end{theorem}
\subsection{{NP}-hardness: balanced connected path problem}
In this section we consider the balanced connected path ({\textit{BCP}}) Problem and prove that it is {NP}-hard. In this problem instead of finding a balanced connected subgraph, our goal is to find a balanced path with a maximum cardinality of vertices. To prove the {\textit{BCP}}~problem is {NP}-hard we give a polynomial time reduction from the \colb{Hamiltonian Path (Ham-Path)} problem which is known to be {NP}-complete \cite{Garey1979}. In this problem, we are given an undirected graph $Q$, and the goal is to find a Hamiltonian path in $Q$ i.e., a path which visits every vertex in $Q$ exactly once. In the reduction we generate an instance $G$ of the {\textit{BCP}}~problem from an instance $Q$ of the {\it Ham-Path}~problem as follows:
\noindent {\bf Reduction:} We make a new graph $Q'$ from $Q$. Let us assume that the graph $Q$ contains $m$ vertices. If $m$ is even then $Q'=Q$. If $m$ is odd, then we add a dummy vertex $u$ in $Q$ and connect to every other vertices in $Q$ by edges with $u$. The resulting graph is our desired $Q'$. It is easy to observe that, $Q$ has a Hamiltonian path if and only if $Q'$ has a Hamiltonian path.
Now we have a {\it Ham-Path}~instance $Q'$ with even number of vertices, say $n$. We arbitrary choose any $n/2$ vertices in $Q'$ and color them red and color the remaining $n/2$ vertices blue. Let $G$ be the colored graph.
This completes the construction. Clearly, this can be done in polynomial time. We now have the following lemma.
\begin{lemma} $Q'$ has a Hamiltonian path $T$ if and only if $G$ has a balanced path $P$ with exactly $n$ vertices.
\end{lemma}
\begin{proof} Assume that $Q'$ has a Hamiltonian path $T$. This implies that, $T$ visits every vertex in $Q'$. Since by the construction there are exactly half of the vertices in $G$ is red and remaining are blue, the same path $T$ is balanced with $n/2$ vertices of each color.
On the other hand, assume that there is a balanced path $P$ in $G$ with exactly $n/2$ vertices of each color. Since, $G$ has a total of $n$ vertices, the path $P$ visits every vertex in $G$. Hence, $P$ is a Hamiltonian path. \qed
\end{proof}
Therefore, we have the following theorem.
\begin{theorem}
{\textit{BCP}}~problem is {NP}-hard for general graph.
\end{theorem}
\section{Algorithmic results}\label{algorithmicresults}
In this section, we consider several graph families and devise polynomial time algorithms for the
{\textit{BCS}}~problem.
Notice that, if the graph is a path or cycle, the optimal solution is just a path. Hence, one can do brute-force search to obtain the maximum balanced path.
In case of a complete graph $K_n$, we output a sub-graph $H$ of $K_{n}$ induced by $V$,
where $|V|= 2|B|$, $B \subset V$, and $B$ is the set of all blue vertices in $K_{n}$ (assuming that, the number of blue vertices is at most the number of red vertices in $K_{n}$).
Clearly, $H$ is the maximum-cardinality balanced sub-graph in $K_{n}$.
We consider trees, split graphs, bipartie graphs (properly colored),
graphs of diameter~$2$, and present polynomial algorithms for each of them.
\subsection{Trees}
In this section we give a polynomial time algorithm for the {\textit{BCS}}~problem where the input graph is a tree. We first consider the following problem.
\noindent {\bf Problem 1:} Given a tree $T=(V,E)$, and a root $t \in V$ where $V=V_R\cup V_B$. The vertices in $V_R$ and $V_B$ are colored red and blue,respectively.
The objective is to find maximum balanced tree with root $t$.
We now design an algorithm to solve this problem. Let $v$ be a vertex in $G$. We associate a set $P_v$ of \colb{pairs} of the form $(r,b)$ to $v$, where $r$ is the count of red vertices and $b$ is the count of blue vertices. A single pair $(r,b)$ associated with vertex $v$ indicates that there is a subtree rooted at $v$ having $r$ red and $b$ blue vertices. Note that $r$ may not be equal to $b$. Now for any $k$ pairs, the sum is also a pair which is defined as the element-wise sum of these $k$ pairs. Let $A_1,A_2, \ldots,A_k$ be $k$ sets. The Minkowski sum $^M\sum_{i=1}^k A_i$ denotes the set of sums of $k$ elements one from each set $A_i$ i.e., $^M\sum_{i=1}^k A_i = A_1 \oplus A_2 \oplus \ldots \oplus A_k$. We use $\oplus$ to denote Minkowski sum between sets. For example, for the Minkowski sum of the sets $A$ and $B$, we write $A \oplus B$ and it means $A \oplus B=\{a+b \colon a\in A, b\in B\}$.
Now we are ready to describe the algorithm to solve Problem 1. In Algorithm 1, we describe how to get maximum balanced subtree with root $t$ for a tree $T$ rooted at $t$.
\IncMargin{1em}
\begin{algorithm}[H]
\SetKwData{Left}{left}\SetKwData{This}{this}\SetKwData{Up}{up}
\SetKwFunction{Union}{Union}\SetKwFunction{FindCompress}{FindCompress}
\SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output}
\Input{$(i)$ A rooted tree $T=(B \cup R,E)$ with root $t$.\\ $(ii)$ $B$ and $R$ are colored blue and red respectively.}
\Output{A set of pairs at each node in $T$.}
\BlankLine
\If{$v$ is a leaf with red color}{
{\ $P_{v}=\{(0,0),(1,0)\}$;}
}
\If{$v$ is a leaf with blue color}{
{\ $P_{v}=\{(0,0),(0,1)\}$;}
}
\If{$v$ be a vertex with red color and $v$ has $k$ children $u_{1}, u_{2},...,u_{k}$ in $T$ with root at $r$,}{
{\ $P_{v} = \{(0,0)\} \cup \{ ^M\sum_{i=1}^k P_{u_{i}} \oplus \{(1,0)\} \}$;}
}
\If{$v$ be a vertex with blue color and $v$ has $k$ children $u_{1}, u_{2},...,u_{k}$ in $T$ with root at $r$,}{
{\ $P_{v} = \{(0,0)\} \cup \{ ^M\sum_{i=1}^k P_{u_{i}} \oplus \{(0,1)\} \}$}\tcp*[r]{\ \colb{ \color{blue} Here $\oplus$ denotes Minkowski Set sum.}}
}
\Return {$P_{t}$}
\caption{ Construct red-blue pair-sets in a rooted tree.}\label{algo_1}
\end{algorithm}\DecMargin{1em}
In Algorithm \ref{algo_1} we compute a finite set $P_{t}$ of pairs $\{(r,b)\}$ at the root $t$ in $T$. To do so, we recursively calculate the set of pairs from leaf to the root. For an internal vertex $v$, the set $P_v$ is calculated as follows: let the color of $v$ is red and it has $k$ children $u_1,u_2,\ldots,u_k$. Then, $P_{v} = \{(0,0)\} \cup \{^M\sum_{i=1}^{k} P_{u_i} \oplus \{(1,0)\}\}$.
We now prove the following lemma.
\begin{lemma}
Let $T$ be rooted tree with $t$ as a root. Then Algorithm \ref{algo_1} produces all possible balanced subtree rooted at $t$ in $O(n^6)$ time.
\end{lemma}
\begin{proof}
Notice that in Algorithm \ref{algo_1}, at each node $v\in T$, we store a set $P_{v}$ of pairs $\{(r_{i},b_{i})\}$, where each $(r_{i},b_{i})$ indicates that there exists a subtree $T'$ with root $v$ such that number of red and blue vertices in $T'$ are $r_{i}$ and $b_{i}$, respectively. Note that $r_{i}$ may not be same as $b_{i}$. When we construct the set $P_v$, all the sets corresponding to its children are already calculated. Finally, in steps $6$ and $8$ of Algorithm \ref{algo_1} we calculate the set $P_v$ based on the color of $v$. Hence, when Algorithm \ref{algo_1} terminates, we get the set $P_t$ where $t$ is the root of $T$.
Now we calculate the time taken by Algorithm \ref{algo_1}. Clearly, steps $2$ and $4$ take $O(1)$ time to construct the $p_v$ when $v$ is a leaf. Note that, the size of $P_v$, for an internal node $v$ is $O(n^2)$. Since there are at most $n$ blue and red vertices in the subtree rooted at $v$. If $ v $ has $k$ children then we have to take Minkowski sum of the sets corresponds to the children of $v$. To get the sum of two sets it takes $ O(n^4) $ time. As there are at most $ n $ children of node $ v $, so the time taken by steps 6 and 8 are $O(n^5)$. Finally, we traverse the tree from bottom to the root. Hence, the total time taken by the algorithm is $O(n^6)$. \qed
\end{proof}
We can now improve the time complexity by slightly modifying the Algorithm \ref{algo_1}. For an internal vertex $v$, we actually don't need all the pairs to get the maximum balanced subtree. Suppose there are two pairs $(a,b)$ and $(c,d)$ in $P_{v}$, where $(b-a)=(d-c)$ and $a<c$. Then, instead of using the subtree with pair $(a,b)$, it is better to use the subtree with pair $(c,d)$, since it may help to construct a larger balance subtree. Therefore, in a set $P_{v}$ if there are $k$ pairs $\{(a_{i},b_{i}); 1 \leq i \leq k \}$ such that $(b_{i}-a_{i})=(b_{j}-a_{j})$ whenever $i \neq j$, $1 \leq i,j \leq k$. Then we remove the $(k-1)$ pairs and store only the pair which is largest among all these $k$ pairs. We say $(a_{m},b_{m})$ is largest when $a_{m} > a_{i}$ and $b_{m}> b_{i}$ for $1 \leq i \leq k, i \neq m$. So we reduce the size of $P_{v}$ for each vertex $v \in T$ from $ O(n^2) $ to $O(n)$. Let $T(n)$ be the time to compute red-blue pairset for the root vertex $t$ in the tree $T$ with size $n$. If $r$ has $k$ children $u_{1}, u_{2},...,u_{k}$ with size $n_{1},n_{2},...,n_{k}$. Then the recurrence is $T(n)= T(n_{1})+T({n_{2})+...+T(n_{k})+O(\sum_{i=1}^{k-1} (n_{1}+ n_{2}+ \dots + n_{i})n_{i+1})}$. Now $ \sum_{i=1}^{k-1} (n_{1}+ n_{2}+ \dots n_{i})n_{i+1} \leq \sum_{i=1}^{k-1} nn_{i+1} = n \sum_{i=1}^{k-1} n_{i+1} \leq n^2$.
which gives the solution that $T(n)= O(n^3)$.
Hence, we conclude the following lemma.
\begin{lemma}
Let $T$ be rooted tree with $t$ as a root. We can produces all possible balanced subtree rooted at $t$ in $O(n^3)$ time and
$ O(n^2) $ space complexity.
\end{lemma}
\subsection*{Optimal solution for BCS problem in tree}
If there are $n$ nodes in the tree $T$, then, for each node $v_{i}, 1 \leq i \leq n$, we consider $T$ to be a tree rooted at $v_{i}$. We then apply Algorithm $\ref{algo_1}$ to find maximum-cardinality balanced subtree rooted at $v_{i}$; let $T_i$ be the resulting balanced subtree, having $m_{i}$ vertices of each color. Then, to obtain an optimal solution for the {\textit{BCST}}~problem in $T$ we choose a balanced subtree that has $max \{m_i; 1 \leq i \leq n\}$ vertices of each color. Now we can state the following theorem.
\begin{theorem}
Let $T$ be a tree whose $n$ vertices are colored either red or blue. Then, in $O(n^4)$ time and $O(n^2)$ space, one can compute a maximum-cardinality balanced subtree of~$T$.
\end{theorem}
\subsection{Split graphs}
A graph $G = (V, E)$ is defined to be a split graph if there is a
partition of $V$ into two sets $S$ and $K$ such that $S$ is an independent set and $K$ is a complete graph. There is no restriction on edges between vertices of $S$ and $K$.
Here we give a polynomial time algorithm for the {\textit{BCS}}~problem where the input graph $G=(V,E)$ is a split graph. Let $S$ and $K$ be two disjoint partition of $V$ where $S$ is an independent set and $K$ is a complete graph. Also, let $S_B$ and $S_R$ be the sets of blue and red vertices in $S$, respectively. Similarly, let $K_B$ and $K_R$ be the sets of blue and red vertices in $K$, respectively. We argue that there exists a balanced connected subgraph in $G$,
having $\min\{|S_B\cup K_B|,|S_R\cup K_R|\}$ vertices of each color.
Note that if $|S_{B}\cup K_{B}|=|S_{R}\cup K_{R}|$ then $G$ itself is balanced.
Now, w.l.o.g., we can assume that $|S_{B}\cup K_{B}|<|S_{R}\cup K_{R}|$.
We will find a connected balanced subgraph $H$ of $G$, where the number of vertices in $H$ is exactly $2|S_{B}\cup K_{B}|$. To do so, we first modify the graph $G=(V,E)$ to a graph $G'=(V,E')$. Then, from $G'$, we will find the desired balanced subgraph with $|S_B\cup K_B|$ many vertices of each color. Moreover, this process is done in two steps.
\begin{description}
\item[Step~1:] Construct $G'=(V,E')$ from $G=(V,E)$.\\
For each $u \in S_{B}$, if $u$ is adjacent to at least a vertex $u'$ in $K_{R}$, then remove all adjacent edges with $u$ except the edge $(u,u')$. Similarly, for each $v \in S_{R}$, if $v$ is adjacent to at least a vertex $v'$ in $K_{B}$, then remove all adjacent edges with $v$ except the edge $(v,v')$.
\vspace{.2cm}
\item[Step~2:]
Delete $|S_{R}\cup K_{R}|-|S_{B}\cup K_{B}|$ vertices from $G'$.\\
Let $k= |S_{R}\cup K_{R}|-|S_{B}\cup K_{B}|$. Now we we have following cases.
\begin{description}
\item[Case 1:] $|S_{R}| \geq k$. We remove $k$ vertices from $S_{R}$ in $G'$. Clearly, after this modification, $G'$ is connected, and we get a balanced subgraph having $|S_B\cup K_B|$ vertices of each color.
\item[Case 2:] $|S_{R}|< k$.
Then we know, $|K_{R}| > |K_{B}\cup S_{B}|$.
Let $S'_{B} \subseteq S_{B}$ be the set of vertices in $G'$ such that each vertex of $S'_{B}$ has exactly one neighbor in $K_R$.
Then, we take a set $ X\subset K_{R} $ with cardinality $|K_{B} \cup S_{B}|$ such that $X$ contains all adjacent vertices of $S'_{B}$.
Now we take the subgraph $H$ of $G'$ induced by
$(S_{B} \cup K_{B} \cup X )$. $H$ is optimal and balanced.
\end{description}
\end{description}
\vspace{.2cm}
\noindent {\bf Running time:} Step 1 takes $O(|E|)$ time to construct $G'$ from $G$. Now in step~2, both Case~1 and Case~2 take $O(|V|)$ time to delete $|S_{R}\cup K_{R}|-|S_{B}\cup K_{B}|$ vertices from $G'$. Hence, the total time taken is $O(n^2)$, where $n$ is the number of vertices in $G$. We conclude in the following theorem.
\begin{theorem}
Given a split graph $G$ of $n$ vertices, with $r$ red and $b$ blue ($n=r+b$) vertices, then, in $O(n^2)$ time we can find a balanced connected subgraph of $G$ having $\min \{b,r\}$ vertices of each color.
\end{theorem}
\subsection{Bipartite graphs, properly colored}
In this section, we describe a polynomial-time algorithm for the {\textit{BCS}}~problem where the input graph is a bipartite graph whose nodes are colored red/blue according to proper 2-coloring of vertices in a graph. We show that there is a balanced connected subgraph of $G$ having $\min \{b,r\}$ vertices of each color where $G$ contains $r$ red vertices and $b$ blue vertices. Note that we earlier showed that the {\textit{BCS}}~problem is NP-hard in bipartite graphs whose vertices are colored red/blue arbitrarily; here, we insist on the coloring being a proper coloring (the construction in the hardness proof had adjacent pairs of vertices of the same color).
We begin with the following lemma.
\begin{lemma}\label{lemma-2colorable}
Consider a tree $T$ (which is necessarily bipartite) and a proper $2$-coloring of its nodes, with $r$ red nodes and $b$ blue nodes. If $r<b$, then $T$ has at least one blue leaf.
\end{lemma}
\begin{proof}
We prove it by contradiction. Let there is no blue leaf. Now assign any blue node say $b_{r}$ as a root. Note that it always exists. Now $b_{r}$ is at level $0$ and $b_{r}$ has degree at least $2$. Otherwise, $b_{r}$ is a leaf with blue color. We put all the adjacent vertices of $b_{r}$ in level~$1$. This level consists of only red vertices. In level~$2$ we put all the adjacent vertices of level~$1$. So level~$2$ consists of only blue vertices. This way we traverse all the vertices in $T$ and let that we stop at $k^{th}$-level. $k$ cannot be even as all the vertices in even level are blue. So $k$ must be odd. Now for each $0 \leqslant i \leqslant \frac{k-1}{2}$, in the vertices of (level $2i$ $\cup$ level $(2i+1)$), number of blue vertices is at most the number of red vertices. Which leads to the contradiction that $r<b$. Hence there exists at least one leaf with blue color. \qed
\end{proof}
We are now ready to describe the algorithm. We first find a spanning tree $T$ in $G$. If $r=b$ then $T$ itself is a maximum balanced subtree (subgraph also) of $G$. Without loss of generality assume that $r<b$. So by Lemma~\ref{lemma-2colorable}, $T$ has at least $1$ blue vertex. Now we remove that blue vertex from $T$. Using similar reason, we repetitively remove $(b-r)$ blue vertices from $T$. Finally, $T$ becomes balanced subgraph of $G$, with $r$ many vertices of each color.
\vspace{.2cm}
\noindent {\bf Running time:}
Finding a spanning tree in $G$ requires $O(n^2)$ time. To find all the leaves in the tree $T$ requires $O(n^2)$ time (breadth first search). Hence the total time is needed is $O(n^2)$.
Now, we state the following theorem.
\begin{theorem}
Given a bipartite graph $G$ with a proper 2 coloring ($r$ red or $b$ blue vertices), then in $O(n^2)$ time we can find a balanced connected subgraph in $G$ having $\min \{b,r\}$ vertices of each color.
\end{theorem}
\subsection{Graphs of diameter~2}
In this section, we give a polynomial time algorithm which solves the {\textit{BCS}}-problem where the input graph has diameter 2. Let $G(V,E)$ be such a graph which contains $b$ blue vertex set $B$ and $r$ red vertex set $R$. We find a balanced connected subgraph $H$ of $G$ having $\min\{b,r\}$ vertices of each color. Assume that $b<r$. This can be done in two phases. In phase 1, we generate an induced connected subgraph $G'$ of $G$ such that (i) $G'$ contains all the vertices in $B$, and (ii) the number of vertices in $G'$ is at most $(2b-1)$. In phase 2, we find $H$ from $G'$.
\begin{description}
\item[Phase 1] To generate $G'$, we use the following result.
\begin{lemma}\label{dia_2}
Let $G=(V,E)$ be a graph of diameter 2. Then for any pair of non adjacent vertices $u$ and $v$ from $G$, there always exists a vertex $w$ such that both $(u,w)\in E$ and $(v,w) \in E$.
\end{lemma}
We first include $B$ in $G'$. Now we have the following two cases.
\begin{description}
\item \noindent{\textbf{Case 1:}} The induced subgraph $G[B]$ of $B$ is connected. In this case, $G'$ is $G[B]$.
\item \noindent{\textbf{Case 2:}} The induced subgraph $G[B]$ of $B$ is not connected.
Assume that $G[B]$ has $k(>1)$ components. Let $B_{1},B_{2},..., B_{k}$ be $k$ disjoints sets of vertices such that each induced subgraph $G[B_i]$ of $B_i$ in $G$ is connected. Now using Lemma \ref{dia_2}, any two vertices $v_i \in B_i$ and $v_j \in B_j$ are adjacent to a vertex say $u_\ell \in R$. We repetitively apply Lemma \ref{dia_2} to merge all the $k$ subgraphs into a larger graph. We need at most $(k-1)$ red vertices to merge $k$ subgraph. We take this larger graph as the graph $G'$.
\end{description}
\vspace{.2cm}
\item[Phase 2] In this phase, we find the balanced connected subgraph $H$ with $b$ vertices of each color. Note that the graph $G'$ generated in phase 1 contains $b$ blue and at most $(b-1)$ red vertices. Assume that $G'$ contains $b'$ red vertices. We add $(b-b')$ red vertices from $G \setminus G'$ to $G'$. This is possible since $G$ in connected.
\end{description}
\vspace{.2cm}
\noindent {\bf Running time:} In phase 1, first finding all the blue vertices and it's induced subgraph takes $O(n^2)$ time. Now to merge all the $k$ components into a single component which is $G'$ needs $O(n^2)$ time. In phase 2, adding $(b-b')$ red vertices to $G'$ takes $O(n^2)$ time as well. Hence, total time requirement is $O(n^2)$.
\begin{theorem}
Given a graph $G = (V, E)$ of diameter~$2$, where the vertices in $G$ are colored either red or blue. If $G$ has $b$ blue and $r$ red vertices then, in $O(n^2)$ time we can find a balanced connected subgraph in $G$ having $\min \{b,r\}$ vertices of each color.
\end{theorem}
\section{Conclusions and open questions}
We have introduced the problem of finding largest size (cardinality of the vertex set) balanced connected subgraph in a simple connected graph. We have seen that this problem is {NP}-complete for bipartite graphs, chordal graphs, or planar graph. We have given polynomial time algorithms for solving this problem for trees, graphs with proper $2$ coloring, split graphs and graphs with diameter~$2$. So the obvious question is can other special classes of graphs be found to yield polynomial time algorithms? For example, outer planar graphs, interval graphs, regular graphs, permutation graphs etc. Here we give another open question. Let $G$ be a given graph and $ OPT$ be the number of vertices in an optimal solution of {\textit{BCS}}~problem. Is there any polynomial time $(\alpha,\beta)$ approximation algorithm which yields a solution $ H $ such that minimum number of blue and red vertices in $ H $ is at most $ \alpha \times OPT $ and difference between the number of blue and red vertices in $ H $ is at most $\beta$?
|
\section{Introduction}
\label{sec:intro}
This file is documentation for the SIAM \LaTeX\ style, including how
to typeset the main document, the \BibTeX\ file, and any supplementary
material. More information
about SIAM's editorial style can be found in the style manual, available
at \url{https://www.siam.org/journals/pdf/stylemanual.pdf}.
The major changes in the SIAM standard class are summarized in \cref{sec:changes}.
The SIAM \LaTeX\@ files can be found at
\url{https://www.siam.org/journals/auth-info.php}. The files that
are distributed for the standard macros are given below.
\begin{itemize}
\item \texttt{siamart171218.cls} (required): Main SIAM standard \LaTeX\ class file.
\item \texttt{siamplain.bst} (required): Bibliographic style file for
\BibTeX.
\item \texttt{docsiamart.tex}: Produces this documentation.
\item \texttt{references.bib}: \BibTeX\ database for this
documentation and examples.
\item \texttt{ex\_article.tex}: Template for article.
\item \texttt{ex\_supplement.tex}: Template for supplement.
\item \texttt{ex\_shared.tex}: Template for shared information for
article and supplement.
\end{itemize}
To use these files, put \texttt{siamart171218.cls} and
\texttt{siamplain.bst} in the directory with your
paper or, alternatively, into your \LaTeX\@ and \BibTeX\@ paths,
respectively.
The outline of a SIAM \LaTeX\ article is shown in
\cref{ex:outline}. Templates are provided and discussed in more detail
in \cref{sec:template}.
\begin{example}[label={ex:outline},listing only
listing options={style=siamlatex,{morekeywords=[1]{maketitle},
morekeywords=[2]{siamart171218}},}
{Document outline}
\documentclass{siamart171218}
\begin{document}
\maketitle
\end{document}
\end{example}
\section{Class options}
\label{sec:class-options}
Class options can be included in the bracketed argument of the
command, separated by commas. The possible class options are:
\begin{itemize}
\item \code{review} --- Recommended for submitting your manuscript to
a SIAM journal. Adds line numbers as well as the statement ``This
manuscript is for review purposes only'' to the bottom of each page.
\item \code{final} --- Turns off the black boxes that help authors
identify lines that are too long. The final published version will
have this option on.
\item \code{supplement} --- Specifies that the file is a supplement
and not the main document, causing changes in the appearance of the
title and numbering; see \cref{sec:supplement} for details.
\item \code{hidelinks} --- Turns off colors on hyperlinks; see
\cref{sec:cr+hyp}. The hyperlinks still exist, but there is no color
to differentiate them.
The final published version will have this option on.
\end{itemize}
\section{Front matter}
\label{sec:front}
The title and author parts are formatted using the standard
\code{\title}, \code{\author}, and \code{\maketitle} commands as
described in Lamport \cite{La86}. The title and author should be
declared in the preamble. The title and author names are automatically
converted to uppercase in the document.
If there is more than one author, each additional author should be preceded by
the \code{\and} command.
The addresses and support acknowledgments are added via
\code{\thanks}. Each author's thanks should specify their address.
The support acknowledgment should be put in the title thanks,
unless specific support needs to be specified for individual authors,
in which case it should follow the author address.
The header for this file was produced by the code in \cref{ex:header}, including an
example of a shared footnote. Each thanks produces a footnote, so the
footnote of the second author is \#3.
The command
\code{\headers{title}{authors}} command, with
the title (possibly shortened to fit) and the authors' names, creates
the page headers, automatically converted to uppercase.
\examplefile[label={ex:header},listing only
listing options={style=siamlatex
deletetexcs={and,thanks,title,author}
{moretexcs=[2]{and,thanks,title,author,maketitle,headers,email}}}
]{Title and authors in preamble}{tmp_\jobname_header.tex}
\newpage
Following the author and title is the abstract, key words listing, and AMS subject
classifications, designated using the \code{abstract}, \code{keywords}, and \code{AMS}
environments.
Authors are responsible for providing AMS numbers which can be found
on the AMS web site \cite{AMSMSC2010}. The abstract, keywords, and AMS subject classifications
for this document are specified in \cref{ex:abstract}.
\examplefile[label={ex:abstract}
before upper={\preamble{\bs newcommand\{\bs BibTeX\}\{\{\bs scshape Bib\}\bs TeX\bs xspace\}}},
listing only
listing options={style=siamlatex
{morekeywords=[2]{abstract,keywords,AMS}}}
]{Abstract, keywords, and AMS classifications}{tmp_\jobname_abstract.tex}
A more complete example, including a PDF supplement, that uses the
included files \texttt{ex\_article.tex}, \texttt{ex\_supplement.tex},
and \texttt{ex\_shared.tex} is discussed in
\cref{sec:template}. The example files can be used as a starting point
for producing a document.
\section{Cross references and hyperlinks}
\label{sec:cr+hyp}
SIAM now supports cross references and hyperlinks via the
\texttt{cleveref} and \texttt{hyperef} packages, which are loaded by
the class file.
\subsection{Cleveref}
\label{sec:cleveref}
SIAM strongly recommends using the commands provided by
the \texttt{cleveref} package for cross referencing.
The package is automatically loaded and already customized
to adhere to SIAM's style guidelines.
To create a cross reference, use the command \code{\cref} (inside
sentence) or \code{\Cref} (beginning of a sentence) in place of the
object name and \code{\ref}.
The \texttt{cleveref}
package enhances \LaTeX's cross-referencing features, allowing
the format of cross references to be determined automatically
according to the ``type" of cross reference (equation, section, etc.)
and the context in which the cross reference is used.
So, the package
\emph{automatically} inserts the object name as well as the
appropriate hyperlink; see \cref{ex:cref}.
It may require two \LaTeX\@ compilations for the references to show up
correctly.
Additional examples are shown in the sections below
for equations, tables, figures, sections, etc.
\begin{example}[label=ex:cref,bicolor,listing options={style=siamlatex
{morekeywords=[2]{cref,ref}}}]{Advantage of using cleveref}
The normal way to get a cross reference with a hyperlink requires a
lot of typing: \hyperref[thm:mvt]{Theorem~\ref*{thm:mvt}}.
The \texttt{cleveref} package gets both the name and hyperlink
automatically using a single macro: \cref{thm:mvt}.
It also handles multiple references with the same macro, such as
\cref{thm:mvt,fig:pgfplots,fig:testfig}.
\end{example}
\subsection{Hyperef}
\label{sec:hyperef}
Hyperlinks are created with the \code{\href} and \code{\url} commands,
as shown in \cref{ex:href}.
SIAM has also defined the \code{\email} command, as shown in
\cref{ex:header}.
You can hide links (i.e., turn off link colors) with the \code{hidelinks} option.
\begin{example}[label={ex:href},bicolor
listing options={style=siamlatex
{morekeywords=[2]{href,url}}}]{Creating hyperlinks}
The \href{https://www.siam.org}{SIAM homepage} has general information.
Note that the colored text will \emph{not} appear in the print version
nor will the hyperlink be active, so the writer may want to specify
the location explicitly instead by using \url{https://www.siam.org}.
\end{example}
Note that homepage links via \code{\url} in the \code{\thanks} environment require
special formatting for the tilde (\string~) character. The formatting
is used in the template and shown in \cref{ex:shared}.
\section{Math and equations}
\label{sec:math}
Here we show some example equations, with numbering, and examples of
referencing the equations. SIAM now includes the package
\texttt{amsmath} by default, and we include some of its features as
well, although the reader should consult the package user manual for
further guidance \cite{amsmath,shortmath}.
Several of the example are adapted
from Mittlebach and Goossen's guide to \LaTeX~\cite{MiGo04}.
\Cref{ex:textmath} is a straightforward example of inline mathematics
equations that does not use any special packages or features.
\begin{example}[label={ex:textmath},bicolor]{Inline math}
The following shows an example of math in text:
Let $S=[s_{ij}]$ ($1\leq i,j\leq n$) be a $(0,1,-1)$-matrix of order $n$.
\end{example}
In \cref{ex:bbm}, we show the recommended method for getting
blackboard fonts using the \texttt{amsfonts} package. This is not
loaded by default and must be included in the preamble.
\begin{example}[label={ex:bbm},bicolor,before upper={\preamble{\bs
usepackage\{amsfonts\}}}
listing options={style=siamlatex
{morekeywords=[2]{mathbb}}}]{Blackboard math}
Blackboard bold characters, such as $\mathbb{C}$ and $\mathbb{R}$,
should be created with the \texttt{amsfonts} package, although this
is not included by default.
\end{example}
\Cref{ex:smallmatrix} shows the \code{smallmatrix} environment for an
inline matrix from the \texttt{amsmath} package, which is included by
default.
\begin{example}[label={ex:smallmatrix},bicolor
listing options={style=siamlatex
{morekeywords=[2]{smallmatrix}}}]{Inline matrix}
Matrices of no more than two rows appearing in text can be created
as shown in the next example:
$B = \bigl[ \begin{smallmatrix} B_{11} & B_{12} \\
B_{21} & B_{22} \end{smallmatrix} \bigr]$.
\end{example}
Bigger matrices can be rendered with environments from
the \texttt{amsmath} package, such as \code{bmatrix} and
\code{pmatrix} used in
\cref{ex:matrices}.
\begin{example}[label={ex:matrices},bicolor
listing options={style=siamlatex
{morekeywords=[2]{bmatrix,pmatrix}}}]{Creating matrices}
Display matrices can be rendered using environments from \texttt{amsmath}:
\begin{equation}\label{eq:matrices}
S=\begin{bmatrix}1&0\\0&0\end{bmatrix}
\quad\text{and}\quad
C=\begin{pmatrix}1&1&0\\1&1&0\\0&0&0\end{pmatrix}.
\end{equation}
\Cref{eq:matrices} shows some example matrices.
\end{example}
\newpage
\Cref{ex:dmo} shows how to use the \code{\DeclareMathOperator} command
from the \texttt{amsopn} package to declare the \code{\Range} macro.
(This example also uses the \texttt{braket} package for the
\code{\set} macro, but this is not necessarily recommended by SIAM.)
\begin{example}[label={ex:dmo}
before upper={\preamble{\bs usepackage\{braket,amsfonts,amsopn\}}\\
\noindent\preamble{\bs DeclareMathOperator\{\bs Range\}\{Range\}}}
bicolor
listing options={style=siamlatex
{moretexcs=[2]{Range}}}
]{Declaring math operators}
An example of a math operator:
\begin{equation}\label{eq:range}
\Range(A) = \set{ y \in \mathbb{R}^n | y = Ax }.
\end{equation}
\end{example}
\Cref{ex:foo} shows how to use the \code{align} environment from
\texttt{amsmath} to easily align multiple equations.
\begin{example}[label={ex:foo},bicolor
listing options={style=siamlatex
{morekeywords=[2]{align}}}]{Aligned equations}
\Cref{eq:a,eq:b,eq:c} show three aligned equations.
\begin{align}
f &= g, \label{eq:a} \\
f' &= g', \quad\text{and} \label{eq:b} \\
\mathcal{L}f &= \mathcal{L}g \label{eq:c}.
\end{align}
\end{example}
Another way to number a set of equations is the
\code{subequations} environment from \texttt{amsmath}, as shown in \cref{ex:aligned}.
\begin{example}[label={ex:aligned},bicolor
listing options={style=siamlatex
{morekeywords=[2]{subequations}}}]{Subequations}
We calculate the Fr\'{e}chet derivative of $F$ as follows:
\begin{subequations}
\begin{align}
F'(U,V)(H,K)
&= \langle R(U,V),H\Sigma V^{T} + U\Sigma K^{T} -
P(H\Sigma V^{T} + U\Sigma K^{T})\rangle \label{eq:aa} \\
&= \langle R(U,V),H\Sigma V^{T} + U\Sigma K^{T}\rangle
\nonumber \\
&= \langle R(U,V)V\Sigma^{T},H\rangle +
\langle \Sigma^{T}U^{T}R(U,V),K^{T}\rangle. \label{eq:bb}
\end{align}
\end{subequations}
\Cref{eq:aa} is the first line, and \cref{eq:bb} is the last line.
\end{example}
~
For an equation split over multiple lines, \cref{ex:ml} shows the
usage of the \code{multline} environment provided by \texttt{amsmath}.
~
\begin{example}[label={ex:ml},bicolor
listing options={style=siamlatex
{morekeywords=[2]{multline}}}]{Equation split across lines}
We claim that the projection $g(U,V)$ is given by the pair of matrices:
\begin{multline} \label{eq:ml}
g(U,V) = \biggl( \frac{R(U,V)V\Sigma^{T}U^{T}
- U\Sigma V^{T}R(U,V)^{T}}{2}U,\\
\frac{R(U,V)^{T}U\Sigma V^{T}-V \Sigma^{T}U^{T}R(U,V)}{2}V \biggr).
\end{multline}
\end{example}
\section{Theorem-like environments}
\label{sec:thm}
SIAM loads \texttt{ntheorem} package and uses it to define the following
theorem-like environments:
\code{theorem},
\code{lemma},
\code{corollary},
\code{definition}, and
\code{proposition}.
SIAM also defines a \code{proof} environment that automatically
inserts the symbol ``$\,\proofbox\,$'' at the end of any proof, even if it ends in an
equation environment. \emph{Note that the document may need to be
compiled twice for the mark to appear.}
Some of the calculus examples were adapted from \cite{CalcI}.
\Cref{ex:theorem} shows usage of the \code{theorem} environment.
An optional argument can be used to name the theorem.
\Cref{ex:cor} illustrates a corollary, without a name, and the
proof environment.
~
\begin{example}[label=ex:theorem,bicolor,parbox=false
listing options={style=siamlatex
{morekeywords=[2]{theorem}}}]{Theorem}
\begin{theorem}[Mean Value Theorem]\label{thm:mvt}
Suppose $f$ is a function that is continuous on the closed interval
$[a,b]$. and differentiable on the open interval $(a,b)$.
Then there exists a number $c$ such that $a < c < b$ and
\begin{displaymath}
f'(c) = \frac{f(b)-f(a)}{b-a}.
\end{displaymath}
In other words, $f(b)-f(a) = f'(c)(b-a)$.
\end{theorem}
\end{example}
\begin{example}[label=ex:cor,bicolor,parbox=false
listing options={style=siamlatex
{morekeywords=[2]{corollary,proof}}}
{Corollary and proof}
\begin{corollary}
Let $f(x)$ be continuous and differentiable everywhere. If $f(x)$
has at least two roots, then $f'(x)$ must have at least one root.
\end{corollary}
\begin{proof}
Let $a$ and $b$ be two distinct roots of $f$.
By \cref{thm:mvt}, there exists a number $c$ such that
\begin{displaymath}
f'(c) = \frac{f(b)-f(a)}{b-a} = \frac{0-0}{b-a} = 0.
\end{displaymath}
\end{proof}
\end{example}
SIAM also defines commands to create your own theorem- and
remark-like environments:
\begin{itemize}
\item \code{newsiamthm} --- Small caps header, italized body.
\item \code{newsiamremark} --- Italics header, roman body.
\end{itemize}
Each command takes two arguments. The first is the environment name,
and the second is the name to show in the document. These commands should
be used instead of \code{\newtheorem}.
\Cref{ex:claim,ex:ref} shows how to use the commands above, including how to
specify the plural version for \texttt{cleveref} if it is unusual.
\begin{example}[label=ex:claim,bicolor
before upper={\preamble{\bs newsiamthm\{claim\}\{Claim\}}\\
\noindent\preamble{\bs newsiamremark\{hypothesis\}\{Hypothesis\}}\\
\noindent\preamble{\bs crefname\{hypothesis\}\{Hypothesis\}\{Hypotheses\}}}
parbox=false
listing options={style=siamlatex
{morekeywords=[2]{claim,proof,hypothesis}}}]{New theorem-like environment}
\begin{claim}\label{cl:constant}
If $f'(x) = 0$ for all $x \in (a,b)$ then $f(x)$ is constant on $(a,b)$.
\end{claim}
\begin{hypothesis}\label{hyp1}
The function $f$ is continuously differentiable.
\end{hypothesis}
\begin{hypothesis}\label{hyp2}
The random variable is normally distributed.
\end{hypothesis}
\end{example}
\begin{example}[label=ex:ref,bicolor,listing options={style=siamlatex
{morekeywords=[2]{cref}}}]{References}
We can reference multiple types of objects with a single reference:
\cref{cl:constant,thm:mvt,hyp1,hyp2}.
\end{example}
\section{Tables}
\label{sec:tab}
Table captions should go above the tables.
\Cref{ex:simpletable} shows the code to generate a
\cref{tab:simpletable}.
A more complicated example is shown in
\cref{ex:table}, which generates \cref{tab:KoMa14}. This
example uses subfloats via the \texttt{subfig} package, as well as
special column options from the \texttt{array} package.
\begin{tcbverbatimwrite}{tmp_\jobname_simpletable.tex}
\begin{table}[tbhp]
{\footnotesize
\caption{Example table}\label{tab:simpletable}
\begin{center}
\begin{tabular}{|c|c|c|} \hline
Species & \bf Mean & \bf Std.~Dev. \\ \hline
1 & 3.4 & 1.2 \\
2 & 5.4 & 0.6 \\ \hline
\end{tabular}
\end{center}
}
\end{table}
\end{tcbverbatimwrite}
\examplefile[label={ex:simpletable}
listing only, listing options={style=siamlatex}
{Example table.}{tmp_\jobname_simpletable.tex}
\input{tmp_\jobname_simpletable.tex}
\begin{tcbverbatimwrite}{tmp_\jobname_table.tex}
\newcolumntype{R}{>{$}r<{$}}
\newcolumntype{V}[1]{>{[\;}*{#1}{R@{\;\;}}R<{\;]}}
\begin{table}[tbhp]
{\footnotesize
\captionsetup{position=top}
\caption{Example table adapted from Kolda and Mayo \rm{\cite{KoMa14}}.}\label{tab:KoMa14}
\begin{center}
\subfloat[$\beta=1$]{
\begin{tabular}{|r|R|V{3}|c|r@{\,$\pm$\,}l|} \hline
occ. & \multicolumn{1}{c|}{$\lambda$} & \multicolumn{4}{c|}{$\mathbf{x}$} &
fevals & \multicolumn{2}{c|}{time (sec.)}\\ \hline
718 & 11.3476 & 0.5544 & 0.3155 & 1.2018 & 0.0977 & 45 & 0.17 & 0.06 \\ \hline
134 & 3.7394 & 0.2642 & -1.1056 & 0.2657 & -0.3160 & 31 & 0.12 & 0.05 \\ \hline
4 & \multicolumn{6}{c|}{\emph{--- Failed to converge ---}} & 0.21 & 0.10 \\ \hline
\end{tabular}}
\subfloat[$\beta=-1$]{
\begin{tabular}{|r|R|V{3}|c|r@{\,$\pm$\,}l|} \hline
occ. & \multicolumn{1}{c|}{$\lambda$} & \multicolumn{4}{c|}{$\mathbf{x}$} &
fevals & \multicolumn{2}{c|}{time (sec.)}\\ \hline
72 & -1.1507 & 0.2291 & 0.6444 & 0.3540 & -0.8990 & 34 & 0.14 & 0.06 \\ \hline
624 & -6.3985 & 0.1003 & 0.1840 & 0.5305 & 1.2438 & 48 & 0.19 & 0.08 \\ \hline
2 & \multicolumn{6}{c|}{\emph{--- Failed to converge ---}} & 0.23 & 0.02 \\ \hline
\end{tabular}}
\end{center}
}
\end{table}
\end{tcbverbatimwrite}
\examplefile[label={ex:table}
before upper={\preamble[\scriptsize]{\bs usepackage\{array\}}\\[-0.4em]
\noindent\preamble[\scriptsize]{\bs usepackage[caption=false]\{subfig\}}}
listing only, listing options=
style=siamlatex,basicstyle=\ttfamily\scriptsize}
{Example table with subtables.}{tmp_\jobname_table.tex}
\input{tmp_\jobname_table.tex}
\section{Figures}
\label{sec:fig}
It is recommended that all figures be generated in high resolution.
In the past, SIAM has required encapsulated
postscript (EPS) format for final production. This is still an
acceptable format, but SIAM also now allows high-resolution
PDF, JPEG, and PNG figures.
If working with EPS images and using \texttt{pdflatex}, we
recommend the package \texttt{epstopdf} to automatically convert EPS
images to PDF for inclusion in PDF documents created by
\texttt{pdflatex}.
\Cref{ex:fig} shows the code to generate \cref{fig:testfig}. This
example uses the \texttt{graphicx} package for the
\code{\includegraphics} command.
\begin{tcbverbatimwrite}{tmp_\jobname_fig.tex}
\begin{figure}[tbhp]
\centering
\subfloat[$\epsilon_{\max}=5$]{\label{fig:a}\includegraphics{lexample_fig1}}
\subfloat[$\epsilon_{\max}=0.5$]{\label{fig:b}\includegraphics{lexample_fig2}}
\caption{Example figure using external image files.}
\label{fig:testfig}
\end{figure}
\end{tcbverbatimwrite}
\examplefile[label={ex:fig}
before upper={\preamble[\scriptsize]{\bs usepackage\{graphicx,epstopdf\}}\\[-0.4em]
\noindent\preamble[\scriptsize]{\bs usepackage[caption=false]\{subfig\}}}
listing only, listing options=
style=siamlatex,basicstyle=\ttfamily\scriptsize}
{Example figure with subfigures and external files}{tmp_\jobname_fig.tex}
\input{tmp_\jobname_fig.tex}
Another option for figures is a graphics-generator that is platform- and
format-independent.
PGF is a TeX macro package for generating such graphics and works together with the most important TeX backend drivers, including pdftex and dvips.
The user-friedly syntax layer called TikZ.
Here we show an example using \texttt{PGFPLOTS}, useful for drawing high-quality
plots directly in \LaTeX.
\Cref{ex:data} and \cref{ex:pgfplots} shows the data and code, respectively, to generate \cref{fig:pgfplots}, adapted from \cite{pgfplots}.
\examplefile[label={ex:data},listing only,
listing options={style=siamlatex,basicstyle=\ttfamily\scriptsize}
{Example data file (data.dat)}{data.dat}
\begin{tcbverbatimwrite}{tmp_\jobname_tikz.tex}
\begin{figure}[tbhp]
\centering
\begin{tikzpicture}
\begin{loglogaxis}[height=2.75in, grid=major,
xlabel={Degrees of Freedom}, ylabel={$L_2$ Error},
legend entries={$d=2$,$d=3$}]
\addplot table [x=d2_dof,y=d2_l2_err] {data.dat};
\addplot table [x=d3_dof,y=d3_l2_err] {data.dat};
\end{loglogaxis}
\end{tikzpicture}
\caption{Example \texttt{PGFPLOTS} figure.}
\label{fig:pgfplots}
\end{figure}
\end{tcbverbatimwrite}
\examplefile[label={ex:pgfplots}
before upper={\preamble[\scriptsize]{\bs usepackage\{pgfplots\}}}
listing only, listing options=
style=siamlatex}
{Example TikZ/PGF for platform-independent graphics.}{tmp_\jobname_tikz.tex}
\input{tmp_\jobname_tikz.tex}
\section{Algorithms}
\label{sec:algs}
SIAM automatically includes the \texttt{algorithm} package in the
class definition. This provides the float environment.
Users have the choice of \texttt{algpseudocode},
\texttt{algorithmic}, and other packages for actually formatting the algorithm.
For example, \cref{alg:buildtree} is produced by the code in \cref{ex:alg}.
In order to reference lines within the algorithm, we need to tell the
\texttt{cleveref} package how to do the referencing, which is the
second line of \cref{ex:alg}. Then we can use the code
\code{\cref{line3}} to produce \cref{line3}.
\begin{tcbverbatimwrite}{tmp_\jobname_alg.tex}
\begin{algorithm}
\caption{Build tree}
\label{alg:buildtree}
\begin{algorithmic}[1]
\STATE{Define $P:=T:=\{ \{1\},\ldots,\{d\}$\}}
\WHILE{$\#P > 1$}
\STATE\label{line3}{Choose $C^\prime\in\mathcal{C}_p(P)$ with $C^\prime := \operatorname{argmin}_{C\in\mathcal{C}_p(P)} \varrho(C)$}
\STATE{Find an optimal partition tree $T_{C^\prime}$ }
\STATE{Update $P := (P{\setminus} C^\prime) \cup \{ \bigcup_{t\in C^\prime} t \}$}
\STATE{Update $T := T \cup \{ \bigcup_{t\in\tau} t : \tau\in T_{C^\prime}{\setminus} \mathcal{L}(T_{C^\prime})\}$}
\ENDWHILE
\RETURN $T$
\end{algorithmic}
\end{algorithm}
\end{tcbverbatimwrite}
\examplefile[float=htpb,label={ex:alg}
before upper={\preamble[\scriptsize]{\bs usepackage\{algorithmic\}}\\[-0.4em]
\preamble[\scriptsize]{\bs Crefname\{ALC@unique\}\{Line\}\{Lines\}}}
listing only, listing options=
style=siamlatex,basicstyle=\ttfamily\scriptsize}
{Example algorithm}{tmp_\jobname_alg.tex}
\input{tmp_\jobname_alg.tex}
\section{Sections}
\label{sec:sec}
Sections are denoted using standard \LaTeX\ section commands, i.e.,
\code{\section}, \code{\subsection}, etc.
If you wish to end the section title with something other that a period (the default), you have to add the command \code{\nopunct} at the end of the title.
Appendices are created with the normal sectioning commands, following
the command \code{
\section{Introduction}\label{Sec_intro}
We consider smooth unconstrained nonconvex optimization over networks, in the following form:\vspace{-0.2cm}
\begin{equation}
\min_{ {\bt}\in \mathbb{R}^m}\, F(\bt)\triangleq \sum_{i=1}^n f_i(\bt),
\tag{P}\label{eq:P}\vspace{-0.2cm}
\end{equation}
where $n$ is the number of agents in the network; and $f_i:\mathbb{R}^m\rightarrow\mathbb{R}$ is the cost function of agent $i$, assumed to be smooth and known only to agent $i$.
Agents are connected through a communication network, modeled as a (possibly directed, strongly) connected graph. No specific topology is assumed for the graph (such as star or hierarchical structure). In this setting, agents seek to cooperatively solve
Problem \eqref{eq:P} by exchanging information with their immediate neighbors in the network.
Distributed \emph{nonconvex} optimization in the form \eqref{eq:P} has found a wide range of applications in several areas, including network information processing, machine learning, communications, and multi-agent control; see, e.g., \cite{Scutari_Ying_LectureNote}. For instance, this is the typical scenario of in-network data-intensive (e.g., sensor-network) applications wherein data are scattered across the agents (e.g., sensors, clouds, robots), and the sheer volume and spatial/temporal disparity of data render centralized processing and storage infeasible or inefficient. Communication networks modeled as \emph{directed} graphs capture simplex communications between adjacent nodes. {This is the case, e.g., in several wireless (sensor) networks wherein nodes transmit at different power and/or communication channels are not symmetric.}
\textbf{Main objective:} We call $\bt$ a critical point of $F$ if $\nabla F(\bt)=\mathbf{0}$; a critical point $\bt$ is a \emph{strict saddle} of $F$ if $\nabla^2 F(\bt)$ has at least one negative eigenvalue; and it is a \emph{Second-order Stationary} (SoS) solution if $\nabla^2 F(\bt)$ is positive semidefinite. Critical points that are not minimizers are of little interest in the nonconvex setting. It is thus desirable to consider methods for \eqref{eq:P} that are not attracted to such points.
When $F$ has a favorable structure, stronger guarantees can be claimed. For instance, a wide range of salient objective functions arising from applications in machine learning and signal processing have been shown to enjoy the so-called \emph{strict saddle} property: all the critical points of $F$ are either strict saddles or local minimizers. Examples include principal component analysis and fourth order tensor factorization \cite{pmlr-v40-Ge15}, low-rank matrix completion \cite{Ge_localMinsSpurious_NIPS16}, and some instances of neural networks \cite{NIPS2016_6112}, just to name a few. In all these cases, converging to SoS solutions--and thus circumventing strict saddles--guarantees finding a local minimizer.
This paper studies for the first time second-order guarantees of two renowned distributed gradient-based algorithms for Problem \eqref{eq:P}, namely: the Distributed Gradient Descent (DGD) \cite{Nedic2009,Nedic2010} and the family of distributed algorithms based on gradient-tracking \cite{dilorenzo2015distributed,NEXT,xu2015augmented}. The former is implementable on undirected graphs while the latter is suitable also for directed graphs. Convergence of these schemes applied to {\it convex} instances of \eqref{eq:P} is well understood; however, less is known in the {\it nonconvex} case, let alone second-order guarantees; the relevant works
are discussed next.\vspace{-0.1cm
\subsection{Literature review} { Recent years have witnessed many studies proving asymptotic solution- and {convergence} rate-guarantees of a variety of algorithms for specific classes of nonconvex optimization problems (e.g., satisfying suitable regularity conditions); a good overview can be found in \cite{chi2019nonconvex}. Since these analyses are heavily tailored to specific applications and it is unclear how to generalize them to a wider class of nonconvex functions, we omit further details and discuss next only results of centralized and distributed algorithms for {\it general} nonconvex instances of \eqref{eq:P}. }\vspace{-0.2cm}
\subsubsection{Second-order guarantees of centralized optimization algorithms}
Second-order guarantees of centralized solution methods for general nonconvex optimization \eqref{eq:P} have been extensively studied in the literature.
\textbf{Hessian-based methods:} Algorithms based on \emph{second-order} information have long been known to converge to SoS solutions of \eqref{eq:P}; they rely on computing the Hessian to distinguish between first- and second-order stationary points. The classical cubic-regularization \cite{Griewank_technical_report, Nesterov2006, Cartis2011_part1, Cartis2011_part2, Agarwal:2017:FAL:3055399.3055464} and trust region (e.g. \cite{doi:10.1137/0904038,Powell1984, Curtis2017,Dauphin:2014:IAS:2969033.2969154}) methods can provably find {\it approximate} SoS solutions in polynomial time (by approximate SoS we mean $\boldsymbol{\theta}$ such that $||\nabla F(\boldsymbol{\theta})||\leq \epsilon_g$ and $\lambda_{\min}(\nabla^2 F(\boldsymbol{\theta}))\geq -\epsilon_h$, for small $\epsilon_g,\epsilon_h>0$); they however require access to the full Hessian matrix. A recent line of works \cite{Carmon_2018, Agarwal_et_al_2017, Carmon_Duchi_2016} show that the requirement of full Hessian access can be relaxed to Hessian-vector products in each iteration, hence solving simpler sub-problems per iteration, but at the cost of requiring more iterations to reach approximate SoS solutions.
\textbf{First-order methods:} For general nonconvex problems, Gradient Descent (GD) is known to find a stationary point in polynomial time \cite{Nesterov04}.
In \cite{pmlr-v49-lee16}, it was proved that randomly initialized GD with a fixed step-size converges to SoS solutions almost surely.
The elegant analysis of \cite{pmlr-v49-lee16}, leveraging tools from the theory of dynamical systems (e.g., the Stable Manifold Theorem), has been later extended in a number of follow-up works establishing same kind of second-order guarantees of a variety of first-order methods, including the proximal point algorithm, block coordinate descent, mirror descent \cite{Jordan_FOM_AvoidsSaddles}; the heavy-ball method and the Nesterov's accelerated method \cite{SWright_AGDavoidssaddles}; block coordinate descent and alternating minimization \cite{Li2019}; and a primal-dual optimization procedure for solving linear equality constrained nonconvex optimization problems \cite{Mingyi_SOAlgs_AvoidsSaddles}.
These results are all asymptotic in nature and it is unclear whether polynomial convergence rates can be obtained for these methods. In \cite{du2017gradient} it was actually proven that, even with fairly natural random initialization schemes and for non-pathological functions, GD can be significantly slowed down by saddle points, taking exponential time to escape. Recent work has analyzed variations of GD that include stochastic perturbations. It has been shown that when perturbations are incorporated into GD at each step the resulting algorithm can escape strict saddle points in polynomial time \cite{pmlr-v40-Ge15}; the same conclusion was earlier established in \cite{Pemantle90} for stochastic gradient methods, although without escape time guarantees. It has also been shown that episodic perturbations suffice; in particular, \cite{pmlr-v70-jin17a} introduced an algorithm that occasionally adds a perturbation to GD, and proved that
the number of iterations to escape saddle points depends only poly-logarithmically on dimension (i.e., it is nearly dimension-independent). Fruitful follow-up results show that other first-order perturbed algorithms escape from strict saddle points efficiently \cite{Jin2018,Songtao2019}.
\subsubsection{Distributed algorithms for \eqref{eq:P} and guarantees}\label{sec:intro_distributed} Distributed algorithms for {\it convex} instances of \eqref{eq:P} have a long history; less results are available for nonconvex objectives. Since the focus on this paper is on nonconvex problems, next, we mainly comment on distributed algorithms for minimizing nonconvex objectives.
\noindent \textbf{$\bullet$ DGD and its variants:} DGD (and its variants) is unquestionably among the first and most studied decentralizations of the gradient descent algorithm for \eqref{eq:P} \cite{Nedic2009,Nedic2010}. The instance of DGD considered in this paper reads: given $\mathbf{x}_i^0\in \mathbb{R}^m$, $i\in[n]$, \vspace{-0.4cm}\begin{equation}
\label{DGD_intro}
\mathbf{x}_i^{\nu+1}=\sum_{j=1}^n D_{ij}\,\mathbf{x}_j^\nu-\alpha \nabla f_i(\mathbf{x}_i^\nu),\quad i\in[n],\vspace{-0.3cm}
\end{equation}
where $\mathbf{x}_i^\nu$ is the agent $i$'s estimate at iteration $\nu$ of the vector variable $\bt$; $\{D_{ij}\}_{i,j}$ are suitably chosen set of nonnegative weights (cf. Assumption \ref{DoublyStochastic_D}), matching the graph topology (i.e., $D_{ij}>0$ if there is a link between node $i$ and $j$, and $D_{ij}=0$ otherwise); and $\alpha>0$ is the step-size. Roughly speaking, the update of each agent $i$ in \eqref{DGD_intro} is the linear combination of two components: i) the gradient $\nabla f_i$ evaluated at the agent's latest iterate (recall that agents do not have access to the entire gradient $\nabla F$); and ii) a convex combination of the current iterates of the neighbors of agent $i$ (including agent $i$ itself). The latter term (a.k.a. consensus step) is instrumental to asymptotically enforcing agreement among the agents' local variables.
When each $f_i$ in \eqref{eq:P} is (strongly) {\it convex}, convergence of DGD is well understood. With a diminishing step-size, agents' iterates converge to a consensual {\it exact} solution; if a constant step-size is used, convergence is generally faster but only to a neighborhood of the solution, and exact consensus is not achieved.
When \eqref{eq:P} is {\it nonconvex}, the available convergence guarantees are weaker.
In \cite{WYin_ncvxDGD_SIAM2018} it was shown that if a constant step-size is employed, every limit point $(\bx_1^{\infty},\ldots, \bx_n^{\infty})$ of the sequence generated by \eqref{DGD_intro} satisfies $\sum_{i=1}^n \nabla_{x_i} f_i(\bx_i^\infty)=\mathbf{0}$; the limit points of agents' iterates are not consensual; asymptotic consensus is achieved only using a diminishing step-size.
Since in general $f_i$ are all different, such limit points are {\it not} critical points of $F$. Nothing is known about the connection of the critical points of $\sum_{i=1}^n f_i(\bx_i)$ and those of $F$, {\it let alone its second-order guarantees}. { A first contribution of this paper is to establish second-order guarantees of DGD \eqref{DGD_intro} applied to \eqref{eq:P} over undirected graphs}.
Several extensions/variants of the vanilla DGD followed the seminal works \cite{Nedic2009,Nedic2010}. The projected (stochastic) DGD for nonconvex constrained instances of \eqref{eq:P} was proposed in \cite{bianchi2013convergence}; with a diminishing step-size, the algorithm converges to a stationary solution of the problem (almost surely, if noisy instances of the local gradients are used). The extension of DGD to
{\it digraphs} was studied in \cite{Nedic2015} for convex unconstrained optimization, and later extended in \cite{tatarenko2017nonconvex} to nonconvex objectives. The algorithm, termed push-sum DGD, combines a local gradient step with the push-sum algorithm \cite{benezit2010weighted}. When a diminishing step-size is employed, push-sum DGD
converges to an exact stationary solution of \eqref{eq:P}; and its noisy perturbed version almost surely converges to local minimizers, provided that $F$ does not have any saddle point \cite{tatarenko2017nonconvex}. { To our knowledge, no other guarantees are known for DGD-like algorithms in the nonconvex setting. In particular, it is unclear whether DGD \eqref{DGD_intro} escapes strict saddles of $F$. }
\smallskip
\noindent\textbf{$\bullet$ Gradient tracking-based methods:}\label{DOGT_method_intro}
To cope with the speed-accuracy dilemma of DGD, \cite{dilorenzo2015distributed,NEXT} proposed a new class of distributed gradient-based methods that converge to an \emph{exact} consensual
solution of nonconvex (constrained) problems while using a \emph{fixed step-size}. The algorithmic framework, termed NEXT,
introduces the idea of \emph{gradient tracking} to correct the
DGD direction and cancel the steady state error in it while using a fixed step-size: each agent updates its own local variables along a surrogate direction that
tracks the gradient $\nabla F$ of the entire objective (the same idea was proposed independently in \cite{xu2015augmented} for convex unconstrained smooth problems).
The generalization of NEXT to digraphs--the SONATA algorithm-
was proposed in \cite{sun2016distributed,Scutari_Ying_LectureNote,SONATA_technical_report, YingDanScuConvergence18}, with \cite{SONATA_technical_report,YingDanScuConvergence18} proving convergence of the agents' iterates to consensual stationary solutions of nonconvex problems at a sublinear rate. {\it No second-order guarantees have been established for these methods}. Extensions of the SONATA family based on different choices of the weight matrices were later introduced in \cite{xin2018linear,ANedich_GeoConvAlg_arxive_2018} for {\it convex} smooth unconstrained problems. In this paper we consider the following family of distributed algorithms based on gradient tracking, which encompasses the majority of the above schemes (see, e.g., \cite[Sec. 5]{SONATA_technical_report}), and refer to it as
Distributed Optimization with Gradient Tracking (\algname/):
\vspace{-0.25cm}
\begin{align}
\mathbf{x}_i^{\nu+1} &=\sum_{j=1}^n{R}_{ij}\bx_j^\nu\ -\alpha\,\mathbf{y}^\nu_i, \label{eq:DOGT_x_update_intro}
\\
\mathbf{y}^{\nu+1}_i &=\sum_{j=1}^n{C}_{ij}\by_j^\nu +\nabla f_i\big(\mathbf{x}^{\nu+1}_i\big)-\nabla f_i\big( \mathbf{x}^\nu_i\big),\quad\text{(Gradient Tracking)}\label{eq:DOGT_track_update_intro}
\end{align}
where $(R_{ij})_{i,j}$ and $(C_{ij})_{i,j}$ are suitably chosen nonnegative weights compliant to the graph structure (cf. Assumption \ref{matrix_C_R}); and $\mathbf{y}_i\in \mathbb{R}^m$ is an auxiliary variable, controlled by agent $i$ via the update (\ref{eq:DOGT_track_update_intro}), which aims at tracking locally the gradient sum $\sum_i\nabla f_i(\mathbf{x}^\nu_i)$. Overall, the update (\ref{eq:DOGT_track_update_intro}) in conjunction with the consensus step in (\ref{eq:DOGT_x_update_intro})
is meant to ``correct'' the local gradient direction $-\nabla f_i(\mathbf{x}^\nu_i)$ (as instead used in the DGD algorithm) and thus nulls asymptotically the steady error $\nabla f_i(\mathbf{x}^\nu_i)-\nabla F(\mathbf{x}^\nu_i)$. This permits the use of a constant step-size $\alpha$ while still achieving {\it exact} consensus without penalizing the convergence rate. Another important difference between DOGT and DGD in (\ref{DGD_intro}) is that the former serves as a unified platform for distributed algorithms applicable over both undirected and directed graphs. Convergence of DOGT in the form \eqref{eq:DOGT_x_update_intro}-(\ref{eq:DOGT_track_update_intro}) when $F$ is nonconvex remains an open problem, let alone second-order guarantees. A second contribution of this paper is to fill this gap and provide a first- and second-order convergence analysis of DOGT.
\noindent\textbf{$\bullet$ Primal-dual distributed algorithms:} \label{PrimalDual_Algs_Subsec}
We conclude this literature review by commenting on distributed algorithms for nonconvex \eqref{eq:P} using a primal-dual form \cite{Zhu:gm,hong2016decomposing,P-ProxPDA}.
Because of their primal-dual nature, all these schemes are implementable only over {\it undirected} graphs.
In \cite{Zhu:gm}
a distributed approximate dual (sub)gradient algorithm, coupled with a consensus step is introduced.
Assuming zero-duality gap, the algorithm is proved to asymptotically find a pair of primal-dual solutions of an auxiliary problem, which however might not be critical points of $F$; also, consensus is not guaranteed. No rate analysis is provided. In \cite{hong2016decomposing}, a proximal primal-dual algorithm is proposed; the algorithm, termed Prox-PDA, employs either a constant or increasing penalty parameter (which plays the role of the step-size); a sublinear convergence rate of a suitably defined primal-dual gap is proved.
A perturbed version of Prox-PDA, P-Prox-PDA, was introduced in \cite{P-ProxPDA}, which can also deal with nonsmooth convex, additive functions in the objective of \eqref{eq:P}. P-Prox-PDA converges to an $\epsilon$-critical point (and thus also to {\it inexact} consensus), under a proper choice of the penalty parameters that depends on $\epsilon$. A sublinear convergence rate is also proved. No second-order guarantees have been established for the above schemes. The only primal-dual algorithms we are aware of with provable convergence to SoS solutions is the one in \cite{Mingyi_SOAlgs_AvoidsSaddles}, proposed for a linearly constrained nonconvex optimization problem. When linear constraints are used to enforce consensus, the primal-dual method \cite{Mingyi_SOAlgs_AvoidsSaddles} becomes distributed and applicable to Problem \eqref{eq:P}, but only for {\it undirected} graphs (DOGT is instead implementable also over digraphs). Second-order guarantees of such a scheme are established under slightly stronger assumptions than those required for DOGT (cf. Remark \ref{DOGT_vs_ProxPDA_remark}, Sec. \ref{sec:SOS_main}). Finally, notice that, since \cite{Mingyi_SOAlgs_AvoidsSaddles} substantially differs from DGD and DOGT--the former is a primal-dual scheme while the latter are primal methods--the convergence analysis put forth in \cite{Mingyi_SOAlgs_AvoidsSaddles} is not applicable to DGD and DOGT. Since DGD and DOGT in their general form encompass two classic algorithms for distributed optimization,
the open problem of their second-order properties leaves a significant gap in the literature.
\vspace{-0.1cm}
\subsection{Major results} We establish for the first time second-order guarantees of DGD (\ref{DGD_intro}) and DOGT \eqref{eq:DOGT_x_update_intro}-\eqref{eq:DOGT_track_update_intro}. The main results are summarized next.\vspace{-0.2cm}
\subsubsection{DGD (\ref{DGD_intro})}\label{sec:DGD_intro_contribution} We prove that: \begin{description}
\item[(i)] { For a sufficiently small step-size $\alpha$, agents' iterates $\{\bx^\nu\}$ generated by (\ref{DGD_intro}) converge to an $O(\alpha)$-critical point of $F$ for all initializations--see {Lemma \ref{L_criticalPoints_are_alpha_statPoints}}; neighborhood convergence to critical points is also established (cf.Theorem \ref{prop_DGD_local_convergence}). This complements the convergence results in \cite{WYin_ncvxDGD_SIAM2018}; }
\item[(ii)] The average sequence $\{\overline{\bx}^\nu\triangleq (1/n) \sum_{i=1}^n \bx_i^\nu\}$ converges almost surely to a neighborhood of a SoS solution of \eqref{eq:P}, where the probability is taken over the initializations--see {Theorem \ref{SOG_of_DGD_thm}}.
\end{description}
To prove \textbf{(ii)}, we employ a novel analysis, which represents a major technical contribution of this work. In fact, existing techniques developed to established second-order guarantees of the centralized GD are not readily applicable to DGD--roughly speaking, this is due to the fact that DGD (\ref{DGD_intro}) converges only to a neighborhood of critical points of $F$ [fixed points of (\ref{DGD_intro}) are not critical points of $F$]. We elaborate next on this challenge and outline our analysis.
The elegant roadmap developed in \cite{pmlr-v49-lee16,Jordan_FOM_AvoidsSaddles} to establish second-order guarantees of the centralized GD builds on the Stable Manifold theorem: roughly speaking, fixed-points of the gradient map corresponding to strict saddles of the objective function are ``unstable'' (more formally, the stable set\footnote{ Given $\mathcal{X}\subseteq \mathbb{R}^m$, $g:\mathbb{R}^m\to \mathbb{R}^m$, and the fixed-point iterate $\bx^{\nu+1}=g(\bx^\nu)$, the stable set of $\mathcal{X}$ is $\{\bx:\lim_\nu g^\nu(\bx)\in\mathcal{X}\}$, i.e., the set of initial points such that $\{\bx^{\nu}\}$ converges to a member of $\mathcal{X}$.} of strict saddles has zero measure), implying almost sure convergence of GD iterates to SoS points \cite[Corollary 2]{Jordan_FOM_AvoidsSaddles}. It is known that the DGD iterates (\ref{DGD_intro}) can be interpreted as instances of the GD applied to the following auxiliary function \cite{WYin_DGD_SIAM2016, WYin_ncvxDGD_SIAM2018}: denoting $\bx\triangleq [\bx_1^\top,\ldots \bx_n^\top]^\top$, \vspace{-0.2cm}
\begin{equation}
L_\alpha (\mathbf{x})\triangleq \underbrace{\sum_{i=1}^n f_i(\bx_i)}_{ {\triangleq F_c(\bx)}}+\frac{1}{2\alpha}\sum_{i=1}^n\sum_{i=j}^n(e_{ij}-D_{ij})\bx_i^\top\bx_j,
\label{LyapunovFunc_intro}\vspace{-0.2cm}
\end{equation}
where $e_{ij}=1$ if there is an edge in the graph between agent $i$ and agent $j$; and $e_{ij}=0$ otherwise. Using \eqref{LyapunovFunc_intro}, (\ref{DGD_intro}) can be rewritten as: denoting $\bx^\nu\triangleq [\bx_1^{\nu\,\top},\ldots \bx_n^{\nu\,\top}]^\top$,\vspace{-0.2cm}
\begin{equation}
\mathbf{x}^{\nu+1}=\mathbf{x}^{\nu}-\alpha\nabla L_\alpha(\mathbf{x}^\nu).\vspace{-0.2cm}
\label{DGD_update_GDupdate_intro}
\end{equation}
One can then apply the above argument (cf. \cite[Corollary 2]{Jordan_FOM_AvoidsSaddles}) to \eqref{DGD_update_GDupdate_intro} and readily establish the following result (see Theorem \ref{L_criticalPoints_are_alpha_statPoints} for the formal statement)
\begin{description}
\item[Fact 1 (informal):] For sufficient small $\alpha>0$, randomly initialized DGD \eqref{DGD_update_GDupdate_intro} [and thus (\ref{DGD_intro})] converges almost surely to a second-order critical point of $L_\alpha$.
\end{description}
Unfortunately, this result alone is not satisfactory, as no connection is known between the critical points of $L_\alpha$ and those of $F$ (note that $L_\alpha:\mathbb{R}^{n\cdot m}\to \mathbb{R}$ whereas $F:\mathbb{R}^{m}\to \mathbb{R}$). To cope with this issue we prove the following two facts.
\begin{description}
\item[Fact 2 (informal):] Every limit point $\overline{\bx}^\infty$ of the average sequence $\overline{\bx}^\nu=1/n\sum_{i=1}^n \bx_i^\nu$ can be made arbitrarily close to a critical point of $F$ by using a sufficiently small $\alpha>0$ (Theorem \ref{prop_DGD_local_convergence});
\item[Fact 3 (informal):] Whenever the limit point $\bar{\mathbf{x}}^\infty=1/n\,\sum_{i=1}^n\mathbf{x}^\infty_i$ belongs to a sufficiently small neighborhood of a strict saddle of $F$, ${\mathbf{x}}^\infty=[\bx_1^{\infty\top},\ldots,\bx_n^{\infty\top}]^\top$ must be a {strict saddle of $L_\alpha$} (Proposition \ref{DGD_SSP_L_and_F_connect} and Corollary \ref{DGD_SSP_L_and_F_connect_2_AMIREM}).
\end{description}
The above three facts will then ensure that, for sufficiently small $\alpha>0$, with almost complete certainty,
$\{\bar{\mathbf{x}}^\nu\}$ will not get trapped in a neighborhood of a strict saddle of $F$--as $\mathbf{x}^\infty$ would be a strict saddle of $L_\alpha$--thus landing in a neighborhood of a SoS solution of \eqref{eq:P}.
Facts 2 \& 3 above are proved under a regularity condition on $F$ which recalls (albeit slightly weaker than) \cite{Gelfand_et_al_1991}. Roughly speaking, the gradient flow over some \emph{annulus} must be uniformly positive correlated with any outward (from the origin) direction (cf. Assumption \ref{Strip_assumption}). This condition is quite mild and is satisfied by functions arising, e.g., from several machine learning applications, including distributed PCA, matrix sensing, and binary classification problems; see Sec. \ref{ProblemSetting_SubSe} for more details. Furthermore, this condition is also sufficient to prove convergence of DGD without assuming the objective function to be globally $L$-smooth (but just locally L-smooth, $LC^1$ for short), a requirement that instead is common to existing (first-order) convergence conditions of DGD. Notice that the loss functions arising from many of the aforementioned machine learning problems are not globally $L$-smooth.
\subsubsection{DOGT \eqref{eq:DOGT_x_update_intro}-\eqref{eq:DOGT_track_update_intro}} For DOGT, we establish the following three results.
\begin{description}
\item[(i)] { When $F$ is nonconvex and the graph is either undirected or directed, it is proved that every limit point of the sequence generated by DOGT is a critical point of $F$. Furthermore, a merit function, measuring distance of the iterates from stationarity and consensus disagreement is introduced, and proved to vanish at a sublinear rate--see {Theorem \ref{Th_asympt_convergence}}. This extends convergence results \cite{ANedich_GeoConvAlg_arxive_2018, xin2018linear}, established only for convex functions. To deal with nonconvexity, our analysis builds on a novel Lyapunov-like function {[cf. \eqref{eq:def_L}]}, which properly combines optimization error dynamics, consensus and tracking disagreements. While these three terms alone do not ``sufficiently'' decrease along the iterates--as local optimization and consensus/tracking steps might act as competing forces--a suitable combination of them, as captured by the Lyapunov function, does monotonically decrease.}
\item[(ii)] When $F$ satisfies the Kurdyka-{\L}ojasiewicz (K\L) property \cite{MR0160856,Kurdyka1998} at any of its critical points, convergence of the entire sequence to a critical point of $F$ is proved (cf. {Theorem \ref{main_global_conv_thm}}), and a convergence rate is provided (cf. {Theorem \ref{Rate_theorem}}). Although inspired by \cite{Attouch2013}, establishing similar convergence results (but no rate analysis) for centralized first-order methods, our proof follows a different path building on the descent of the Lyapunov function introduced in (i), which does not satisfy {\cite[conditions H1-H2]{Attouch2013}});
see Sec. \ref{Global_Conv_Subsec} for details.
\item[(iii)] The sequence of iterates generated by DOGT is shown to converge to SoS solutions of \eqref{eq:P} almost surely, when initial points are randomly drawn from a suitably chosen linear subspace--see {Theorem \ref{Main_SOS_guarantee_long}}. This result is proved for undirected and directed networks. The proofs build on the stable manifold theorem, based upon the interpretation of DOGT dynamics as fixed-point iterates of a suitably defined map.
The challenge in finding such a map is ensuring that the stable set of its undesirable fixed-points--those associated with the strict saddles of $F$--has measure zero in the subspace where the initialization of DOGT takes place. Note that this subspace is not full dimensional.
\end{description}
\smallskip
While our paper was under review after its initial arXiv posting \cite{Danarxiv2018} and its companion conference version \cite{8636044}, we became aware of a followup line of related works \cite{vlaski2019distributed_part1,vlaski2019distributed_part2}.
These schemes study second-order guarantees of variations of the DGD algorithm (\ref{DGD_intro}). Specifically,
\cite{vlaski2019distributed_part1,vlaski2019distributed_part2} studied the behavior of (the ATC version of) DGD wherein exact gradients are replaced by stochastic approximations; the algorithm is proved to return approximate second-order stationary points in polynomial number of iterations. Finally, \cite{li2019landscape} proposed a variant of DGD to solve the distributed low-rank matrix factorization problem and they prove almost sure convergence to global minima of the problem.
\vspace{-0.1cm}
\subsection{Paper organization}
The rest of the paper is organized as follows. The main assumptions on the optimization problem and network are introduced in Sec.~\ref{ProblemSetting_SubSe}. Sec.~\ref{sec:DGD} studies guarantees of DGD over undirected graphs, along the following steps: i) existing convergence results are discussed in Sec. \ref{sec:DGD_convergence_preliminary}; ii) Sec. \ref{sec:critical-DGD} studies convergence to a neighborhood of a critical point of $F$; and iii) Sec. \ref{sec:SOS-DGD} establishes second-order guarantees.
\algname/ algorithms are studied in Sec.~\ref{sec:DOGT} along the following steps: i) Sub-sequence convergence is proved in {Sec.}~\ref{Subsequence_Conv_Subsec}; ii) Sec.~\ref{Global_Conv_Subsec} establishes global convergence under the {K\L} property of $F$; and iii) Sec.~\ref{Guarantees_Sec} derives second-order guarantees over undirected and directed graphs. Finally, Sec.~\ref{numerical_experiments} presents some numerical results. \vspace{-0.1cm}
\subsection{Notation}\label{sec:notation}
The set of nonnegative integers is denoted by $\mathbb{N}_+$ and we use $[n]$ as a shorthand for $\{1,2,\ldots,n\}$. All vectors are denoted by bold letters and assumed to be column vectors; given a vector $\bx$,
$||\mathbf{x}||$ denotes the $\ell_2$ norm of $\mathbf{x}$; any other specific vector norm is subscripted accordingly.
$\mathbf{x}$ is called \emph{stochastic} if all its components are nonnegative and sum to one; and $\mathbf{1}$ is the vector of all ones (we write $\mathbf{1}_m$ for the $m$--dimensional vector, if the dimension is not clear from the context). Given sets $\mathcal{X},\mathcal{Y}\subseteq \mathbb{R}^m$, we denote $\mathcal{X}\setminus \mathcal{Y}\triangleq \{x\in\mathcal{X}:x\notin\mathcal{Y}\}$, $\overline{\mathcal{X}}\triangleq \mathbb{R}^m\setminus\mathcal{X}$ (complement of $\mathcal{X}$), and $\mathbf{x}+\mathcal{X}=\{\mathbf{x}+\mathbf{z}:\mathbf{z}\in\mathcal{X}\}$. { $\mathcal{V}_{\mathbf{x}}$ and $\mathcal{B}(\mathbf{x},r)^d$ denote a neighborhood of $\mathbf{x}$ and the $d$-dimensional closed ball of radius $r>0$ centered at $\mathbf{x}$, respectively; when the ball is centered at $\mathbf{0}$, we will write $\mathcal{B}^d_r$. We further define an \emph{annulus} by $\mathcal{S}_{r,\epsilon}\triangleq \mathcal{B}^d_r\setminus\mathcal{B}^d_{r-\epsilon}$, with some $r>\epsilon>0$.}
The Euclidean projection of $\mathbf{x}\in\mathbb{R}^m$ onto the convex closed set $\mathcal{X}\subseteq \mathbb{R}^m$ is $\mathrm{proj}_\mathcal{X}(\mathbf{x})\triangleq \arg\min_{\mathbf{y}\in \mathcal{X}}||\mathbf{x}-\mathbf{y}||$. The sublevel set of a function $U$ at $u$ is denoted by $\mathcal{L}_{U}(u)\triangleq \{\mathbf{x}:U(\mathbf{x})\leq u\}$.
Matrices are denoted by capital bold letters; $A_{ij}$ is the the $(i,j)$-th element of $\mathbf{A}$; $\mathcal{M}_m(\mathbb{R})$ is the set of all $m\times m$ real matrices; $\mathbf{I}$ is the identity matrix (if the dimension is not clear from the context, we write $\mathbf{I}_m$ for the $m\times m$ identity matrix); $\mathbf{A}\geq 0$ denotes a nonnegative matrix; and $\mathbf{A}\geq \mathbf{B}$ stands for $\mathbf{A}-\mathbf{B}\geq 0$.
The spectrum of a square real matrix $\mathbf{M}$ is denoted by $\text{spec}(\mathbf{M})$ and its spectral radius is $\text{spradii}(\mathbf{M})\triangleq\max\{|\lambda|:\lambda\in\text{spec}(\mathbf{M})\}$; the spectral norm is $||\mathbf{M}||\triangleq \max_{||\mathbf{x}||\neq 0}||\mathbf{Mx}||/||\mathbf{x}||$, and any other matrix norm is subscripted accordingly.
Finally, the minimum (resp. maximum) singular value are denoted by $\sigma_\mathrm{min}(\mathbf{M})$ (resp. $\sigma_\mathrm{max}(\mathbf{M})$) and minimum (resp. maximum) eigenvalue by $\lambda_\mathrm{min}(\mathbf{M})$ (resp. $\lambda_\mathrm{max}(\mathbf{M})$).
{ The sequence generated by DGD (and DOGT) depends on the step-size $\alpha$ and the initialization $\bx^0$. When necessary,
we write $\{\bx^\nu(\alpha,\bx^0)\}$ for $\{\bx^\nu\}$.
Throughout the paper, we assume that all the probability measures are absolutely continuous with respect to the Lebesgue measure.}
\section{Problem \& network setting}
\label{ProblemSetting_SubSe}\vspace{-0.1cm}
In this section, we introduce the various assumptions on the functions $f_i$ and the graph, under which our results are derived.
\begin{assumption}[{On Problem \ref{eq:P}}]\label{P_assumption} Given Problem~\eqref{eq:P},
\begin{enumerate}[leftmargin=1.5cm,label=(\roman*)]
\item $f_i$ ($\forall i$) is $r+1$ times continuously differentiable for some $r\geq 1$, and $\nabla f_i$ is $L_{i}$-Lipschitz continuous. Denote $L_\mathrm{max}\triangleq \max_i~L_i$;
\item $F$ is
coercive.
\end{enumerate}
\end{assumption}
For some convergence results of DGD we need the following slightly stronger condition.
\hypertarget{P_assumption_prime}{\emph{Assumption} 2.1'.} Assumption \ref{P_assumption}-(i) is satisfied and (ii) each $f_i$ is coercive.\smallskip
We also make the blanket assumption that each agent $i$ knows only its own $f_i$ but not the rest of the objective function.
Note that Assumption \ref{P_assumption}, particularly the global Lipschitz gradient continuity of $f_i$, is quite standard in the literature. Motivated by some applications of interest (see examples below), we will also prove convergence of DGD under $LC^1$ only and the mild condition \eqref{Strip_assumption_eq} below (cf. Assumption \ref{Strip_assumption}). Although strictly not necessary, coercivity in Assumptions \ref{P_assumption} \& \hyperlink{P_assumption_prime}{2.1'} simplifies some of our derivations; our results can be extended under the weaker assumption that \eqref{eq:P} has a solution.
{Some of the convergence results of DGD and DOGT are established under the assumption that $F$ satisfies the Kurdyka-\L ojasiewicz ({K\L}) inequality \cite{Kurdyka1998, MR0160856}
\begin{definition}[{K\L} property]\label{KL_def}
Given a function $U:\mathbb{R}^N\rightarrow\mathbb{R}\cup\{+\infty\}$, we set $[a<U<b]\triangleq \{\mathbf{z}\in \mathbb{R}^N\,:\, a<U( \mathbf{z})<b\}$, and
\begin{enumerate}[label=(\alph*)]
\item The function $U$ has {K\L} property at $\acute{\mathbf{z}}\in\mathrm{dom}~\partial U$ if there exists $\eta\in(0,+\infty]$, a neighborhood $\mathcal{V}_{\acute{\mathbf{z}}}$, and a continuous concave function $\phi:[0,\eta)\rightarrow \mathbb{R}_+$ such that:
\begin{enumerate}[label=(\roman*)]
\item $\phi(0)=0$,
\item $\phi$ is $\mathcal{C}^1$ on $(0,\eta)$,
\item for all $s\in(0,\eta)$, $\phi'(s)>0$,
\item for all $\mathbf{z}\in\mathcal{V}_{\acute{\mathbf{z}}}\cap[U(\acute{\mathbf{z}})<U<U(\acute{\mathbf{z}})+\eta]$, the {K\L} inequality holds:\vspace{-0.1cm}
\begin{equation}
\phi'\left(U(\mathbf{z})-U(\acute{\mathbf{z}})\right)\mathrm{dist}(0,\partial U(\mathbf{z}))\geq 1.
\label{KL_ineq}\vspace{-0.1cm}
\end{equation}
\end{enumerate}
\item A proper lower-semicontinuous function $U$ is called {K\L} if it satisfies the {K\L} inequality at every point in $\mathrm{dom}~\partial U$.
\end{enumerate}
\end{definition}
Many problems involve functions satisfying the {K\L} inequality; real semi-algebraic functions provide a very rich class of functions satisfying the K\L, see \cite{attouch2010proximal} for a thorough discussion.
}
Second-order guarantees of DGD are obtained under the following two extra assumptions below; Assumption \ref{Lipschitz_Hessian} is quite standard and widely used in the literature to establish second-order guarantees of centralized algorithms (e.g., \cite{pmlr-v40-Ge15, pmlr-v70-jin17a, Nesterov2006, Cartis2011_part1, Cartis2011_part2, Agarwal:2017:FAL:3055399.3055464,Curtis2017}) as well as of distributed algorithms \cite{Mingyi_SOAlgs_AvoidsSaddles,vlaski2019distributed_part1,vlaski2019distributed_part2}. Assumption \ref{Strip_assumption} is introduced for this paper and commented below.
\begin{assumption}
\label{Lipschitz_Hessian}
Each $f_i:\mathbb{R}^m\rightarrow\mathbb{R}$ is twice differentiable and $\nabla^2 f_i$ is $L_{\nabla_i^2}$-Lipschitz continuous.
The Lipschitz constant of $\nabla^2 F$ and $\nabla^2 F_c$ are $L_{\nabla^2}=\sum_{i=1}^n L_{\nabla^2_i}$ and $L_{\nabla^2_c}= \max_{i}~L_{\nabla^2_i}$, respectively, {where $F_c$ is defined in eq. \eqref{LyapunovFunc_intro}.}
\end{assumption}
{
\begin{assumption}
\label{Strip_assumption}
(i) Each $f_i$ is $LC^1$; and (ii) there exist $0<\epsilon<R$ and $\delta>0$ such that \vspace{-0.2cm}
\begin{equation}
\label{Strip_assumption_eq}
\inf_{\bt\in\mathcal{S}_{R,\epsilon}}\left\langle\nabla f_i(\bt),\bt/\norm{\bt}\right\rangle\geq \delta,\quad\forall i\in[n].
\end{equation}
\end{assumption}
Roughly speaking, the condition above postulates that the gradient $\nabla f_i(\bt)$ is positively correlated with any radial direction $\bt/\norm{\bt}$, for all $\bt$ in the annulus $\mathcal{S}_{R,\epsilon}$. A slightly more restrictive form of the above assumption has appeared in \cite[Assumption A3]{Gelfand_et_al_1991}.
Many functions of practical interest satisfy this assumption; some examples arising from machine learning applications are listed below.
\begin{description
\item[Distributed PCA \cite{Eckart1936}:] Given matrices $\mathbf{M}_i\in\mathbb{R}^{m\times m}$, $i\in [n]$, the distributed PCA problem is to find the leading eigenvector of $\sum_{i=1}^n\mathbf{M}_i$ by solving
\begin{equation}
\label{Dist_PCA}
\min_{\bt\in\mathbb{R}^{m}}\quad \frac{1}{4}\Big\|\boldsymbol{\theta\theta}^\top-\sum_{i=1}^n\mathbf{M}_i\Big\|_F^2,
\end{equation}
which can be rewritten in the form \eqref{eq:P}
\item \textbf{Phase retrieval { \cite{chi2019nonconvex}}}: Let $\{(\mathbf{a}_i,y_i)\}_{i=1}^n$, with $\mathbf{a}_i\in\mathbb{R}^{m}$ and $y_i\in\mathbb{R}$ such that
$y_i=\mathbf{a}_i^\top\mathbf{M}^\ast\mathbf{a}_i=(\mathbf{a}_i^\top\bt^\ast)^2$, and $\mathbf{M}^\ast=\bt^{\ast}\bt^{\ast\top}\in\mathbb{R}^{m\times m}$.
The phase retrieval problem reads
\begin{equation}\label{PR_problem}
\min_{\bt\in\mathbb{R}^{m}}\quad \frac{1}{4}\sum_{i=1}^n\left(||\mathbf{a}_i^\top\bt||^2-y_i\right)^2+\frac{\lambda}{2}||\bt||^2,
\end{equation}
where $\lambda>0$ is a given parameter.
\item \textbf{Matrix sensing {\cite{chi2019nonconvex}}: } Let $\{(\mathbf{A}_i,y_i)\}_{i=1}^n$, with $\mathbf{A}_i\in\mathbb{R}^{m\times m}$ and $y_i\in\mathbb{R}$ such that $y_i=\langle \mathbf{A}_i,\mathbf{M}^\ast\rangle$, and $\mathbf{M}^\ast=\bT^{\ast}\bT^{\ast\top}\in\mathbb{R}^{m\times m}$, $\bT^\ast\in\mathbb{R}^{m\times r}$. The matrix sensing problem reads \vspace{-0.2cm}
\begin{equation}\label{MS_problem}
\min_{\bT\in\mathbb{R}^{m\times r}}\quad \frac{1}{4}\sum_{i=1}^n\Big(\Big\langle \mathbf{A}_i,\bT\bT^\top\Big\rangle -y_i\Big)^2+\frac{\lambda}{2}\norm{\bT}^2_F,
\end{equation}
where $\lambda>0$ is a given parameter.
\item \textbf{Gaussian mixture model {\cite{li2019landscape}}}:
Let $\{\mathbf{z}_i\}_{i=1}^n$ be $n$ points drawn from a mixture of $q$ Gaussian distributions, i.e., $\mathbf{z}_i\sim \sum_{d=1}^q\mathcal{N}(\boldsymbol{\mu}_d^\ast,\boldsymbol{\Sigma})$, where $\mathcal{N}(\boldsymbol{\mu}_d^\ast,\boldsymbol{\Sigma})$ is the Gaussian distribution with mean $\boldsymbol{\mu}_d^\ast\in\mathbb{R}^{m}$ and covariance $\boldsymbol{\Sigma}\in\mathbb{R}^{m\times m}$. The goal is to estimate the mean values $\boldsymbol{\mu}_1^\ast, \ldots, \boldsymbol{\mu}_q^\ast$ by solving the maximum likelihood problem \vspace{-0.2cm}
\begin{equation}
\min_{\{\bt_d\in \mathbb{R}^{m}\}_{d=1}^q}\, -\sum_{i=1}^n\log{\bigg(\sum_{d=1}^q\phi_m(\mathbf{z}_i-\bt_d)\bigg)}+\frac{\lambda}{2}\norm{\bt_d}^2,
\end{equation}
where $\phi_m(\bt)$ is the multivariate normal distribution with $\mathbf{0}$ mean and covariance ${\boldsymbol{\Sigma}}$;
\item \textbf{Bilinear logistic regression} \cite{Dyrholm_JMLR_2007}: The description of the problem along with some numerical results can be found in Sec. \ref{BLR_simulations};
\item \textbf{Artificial neuron { \cite{auer1996exponentially,zhao2010convex}}}: Let $\{(\mathbf{s}_i,\xi_i)\}_{i=1}^n$ be $n$ samples, with $\mathbf{s}_i\in\mathbb{R}^m$, $\xi_i\in\mathbb{R}$, and measurement model $\xi_i=\sigma (\mathbf{s}_i^\top\bt^\ast)$, where $\bt^\ast$ is the optimal weights and $\sigma(\cdot)$ is a \emph{transfer} function; e.g., the logistic regression function $\sigma(\theta)=1/(1+\exp(-\theta))$. The goal is to estimate $\bt^\ast$ by solving
\begin{equation}
\min_{\bt\in\mathbb{R}^m}\quad \sum_{i=1}^n ~\frac{1}{2n}\left[\left(\xi_i-\sigma(\mathbf{s}_i^\top\bt)\right)^2+\frac{\lambda}{2}\norm{\bt}^2\right],
\end{equation}
where $\lambda>0$ is a given parameter. Further binary classification models satisfying Assumption \ref{Strip_assumption} include $f_i$ functions such as \cite{zhao2010convex}
\vspace{-0.2cm}\begin{equation}
\begin{aligned}
f_i(\bt)=&1-\tanh{\xi_i\mathbf{s}_i^\top\bt}+\frac{\lambda}{2}\norm{\bt}^2,
\\
f_i(\bt)=&\left(1-\sigma(\xi_i\mathbf{s}_i^\top\bt)\right)^2+\frac{\lambda}{2}\norm{\bt}^2,
\\
f_i(\bt)=&-\ln{\sigma(\xi_i\mathbf{s}_i^\top\bt)}+\ln{\sigma(\xi_i\mathbf{s}_i^\top\bt+\mu)}+\frac{\lambda}{2}\norm{\bt}^2,
\end{aligned}
\end{equation}
where $\lambda>0$ and $\mu>0$ are given parameters.
\end{description}
In all these examples, Assumption \ref{Strip_assumption} is satisfied for any sufficiently large $R$ and $R-\epsilon$; the proof can be found in Appendix \ref{Strip_assumption_appendix}.
Note that many of the functions listed above are not $L$-smooth on their entire domain, violating thus (part of) Assumption \ref{P_assumption}(i). Motivated by these examples, we will extend existing convergence results of DGD, replacing Assumption \ref{P_assumption}(i) with Assumption \ref{Strip_assumption}.
}
\textbf{Network model:}
The network is modeled as a (possibly) directed graph $\mathcal{G}=(\mathcal{V},\mathcal{E})$, where the set of vertices $\mathcal{V}=[n]$ coincides with the set of agents, and the set of edges $\mathcal{E}$ represents the agents' communication links: $(i,j)\in\mathcal{E}$ if and only if there is link directed from agent $i$ to agent $j$. The {in-neighborhood} of agent $i$ is defined as $\mathcal{N}_i^{\rm in}=\{j|(j,i)\in\mathcal{E}\}\cup\{i\}$ and represents the set of agents that can send information to agent $i$ (including agent $i$ itself, for notational simplicity). The {out-neighborhood} of agent $i$ is similarly defined $\mathcal{N}_i^{\rm out}=\{j|(i,j)\in\mathcal{E}\}\cup\{i\}$. When the graph is undirected, these two sets coincide and we use
$\mathcal{N}_i$ to denote the neighborhood of agent $i$ (with a slight abuse of notation, we use the same symbol $\mathcal{G}$ to denote either directed or undirected graphs). Given a nonnegative matrix $\mathbf{A}\in\mathcal{M}_n(\mathbb{R})$, the directed graph induced by $\mathbf{A}$ is defined as $\mathcal{G}_{A}=(\mathcal{V}_A,\mathcal{E}_A)$, where $\mathcal{V}_A\triangleq [n]$ and $(j,i)\in \mathcal{E}_A$ if and only if $A_{ij}>0$. The set of roots of all the directed spanning trees in $\mathcal{G}_{A}$ is denoted by $\mathcal{R}_A$.
We make the following blanket standard assumptions on $\mathcal{G}$.
\begin{assumption}[On the network] \label{Net-Assump} The graph (resp. digraph) $\mathcal{G}$ is connected (resp. strongly connected).
\end{assumption}
\section{The DGD algorithm}\label{sec:DGD}
{ Consider Problem \eqref{eq:P} and assume that the network is modeled as an undirected graph $\mathcal{G}$. As described in Sec. \ref{Sec_intro}, the DGD algorithm is based on a decentralization of GD as described in \eqref{DGD_intro}.
It is convenient to rewrite the update \eqref{DGD_intro} in the matrix/vector form: Using the definition of \emph{aggregate} function $F_c(\mathbf{x})$ [cf. \eqref{LyapunovFunc_intro}] and
$\bx^\nu\triangleq [\bx_1^{\nu\top},\ldots \bx_n^{\nu\top}]^\top$,
we have
\begin{equation}
\mathbf{x}^{\nu+1}=\mathbf{W}_D\,\mathbf{x}^{\nu}-\alpha\nabla F_c(\mathbf{x}^\nu),
\label{DGD_update}
\end{equation}
given $\bx^0\in\mathbb{R}^{mn}$, where
$\mathbf{W}_D\triangleq \mathbf{D}\otimes \mathbf{I}_m$, and $\mathbf{D}\in\mathcal{M}_n(\mathbb{R})$ satisfying
the following assumption.}
\begin{assumption}
\label{DoublyStochastic_D}
$\mathbf{D}\in\mathcal{M}_n(\mathbb{R})$ is nonnegative, doubly-stochastic, and compliant to $\mathcal{G}$, i.e., $D_{ij}>0$ if and only if $(j,i)\in\mathcal{E}$, and $D_{ij}=0$ otherwise.
\end{assumption}
\subsection{Existing convergence results}\label{sec:DGD_convergence_preliminary}
Convergence of DGD applied to the nonconvex problem \eqref{eq:P} has been established \cite{WYin_DGD_SIAM2016,WYin_ncvxDGD_SIAM2018}, and summarized below.
\begin{theorem}[\cite{WYin_DGD_SIAM2016,WYin_ncvxDGD_SIAM2018}]
\label{DGD_conv_cons_thm}
Let Assumptions \hyperlink{P_assumption_prime}{2.1'}, \ref{Net-Assump} hold. Given arbitrary $\mathbf{x}^0\in\mathbb{R}^{mn}$ and $0<\alpha< \alpha_{\max}\triangleq \sigma_\mathrm{min}(\mathbf{I}+\mathbf{D})/L_c$, let $\{\mathbf{x}^\nu\}$ be the sequence generated by the DGD algorithm \eqref{DGD_update} under Assumption \ref{DoublyStochastic_D}. Then $\{\mathbf{x}^\nu\}$ is bounded and
\begin{enumerate}[label=(\roman*)]
\item[(i)] \emph{[almost consensus]:} for all $i\in[n]$ and $\nu\in \mathbb{N}_+$,\vspace{-0.2cm}
\begin{equation*}
\|\mathbf{x}_{i}^\nu-\bar{\mathbf{x}}^\nu\|\leq (\sigma_2)^\nu||\mathbf{x}_i^0|| +\frac{\alpha H}{1-\sigma_2}, \vspace{-0.2cm}
\end{equation*}
where $\sigma_2<1$ is the second largest singular value of $\mathbf{D}$, and $H$ is a universal upper-bound of $\{||\nabla F_c(\mathbf{x}^\nu)||\}$;
\item[(ii)] \emph{[stationarity]:} every limit point $\mathbf{x}^\infty$ of $\{\mathbf{x}^\nu\}$ is such that $\mathbf{x}^\infty\in\mathrm{crit}~L_{\alpha}$.
\end{enumerate}
In addition, if $L_\alpha$ is a {K\L} function, then $\{\mathbf{x}^\nu\}$ is globally convergent to some $\mathbf{x}^\infty\in\mathrm{crit}~L_{\alpha}$.
\end{theorem}
}
Although $L$-smoothness of $f_i$'s is a common assumption in the literature, above convergence results can also be established without this condition but under Assumption \ref{Strip_assumption}--see Remark \ref{local_lip_extension_remark} and Appendix \ref{DGD_conv_cons_thm_extended_proof_appendix} for details.
{Since \eqref{DGD_update} is the gradient update applied to $L_\alpha$ (cf. \eqref{DGD_update_GDupdate_intro}), non-convergence of the DGD algorithm to strict saddle points of $L_\alpha$ can be established by applying
\cite[Corollary 2]{Jordan_FOM_AvoidsSaddles} to \eqref{DGD_update_GDupdate_intro}; the statement is given in Theorem \ref{DGD_naive_SOG} below. The following extra assumption on the weight matrix $\mathbf{D}$ is needed.}
\begin{assumption}
\label{nonsingular_W}
The matrix $\mathbf{D}\in\mathcal{M}_n(\mathbb{R})$ is nonsingular.
\end{assumption}
\vspace{-0.2cm}
{
\begin{theorem}
\label{DGD_naive_SOG}
Consider Problem \eqref{eq:P}, under Assumptions \hyperlink{P_assumption_prime}{2.1'}, \ref{Net-Assump}, and further assume that each $f_i$ is a {K\L} function.
Let $\{\mathbf{x}^\nu\}$ be the sequence generated by the DGD algorithm with step-size $0<\alpha<\frac{\sigma_\mathrm{min}(\mathbf{D})}{L_c}$ and weight matrix $\mathbf{D}$ satisfying Assumptions \ref{DoublyStochastic_D} and \ref{nonsingular_W}. Then, the stable set of strict saddles has measure zero. Therefore, $\{\mathbf{x}^\nu\}$ convergences almost surely to a SoS solution of $L_\alpha$, where the probability is taken over the random initialization $\mathbf{x}^0\in\mathbb{R}^{mn}$.
\end{theorem}
}
As anticipated in Sec. \ref{sec:DGD_intro_contribution}, the above second-order guarantees are not satisfactory as they do not provide any information on the behavior of DGD near critical points of $F$, including the strict saddles of $F$. In the following, we fill this gap. We first show that the DGD algorithm convergences to
neighborhood of the critical points of $F$, whose size is controlled by the step-size $\alpha>0$ (cf. Section \ref{sec:critical-DGD}). Then, we prove that, for sufficiently small $\alpha>0$, such critical points are almost surely SoS solutions of \eqref{eq:P}, where the randomization is taken on the initial point (cf. Section \ref{sec:SOS-DGD}).
\subsection{DGD converges to a neighborhood of critical points of $F$} \label{sec:critical-DGD}
Let us begin with introducing the definition of $\epsilon$-critical points of $F$.
\begin{definition}
\label{eps_critical_points_def}
A point $\boldsymbol{\theta}\in\mathbb{R}^m$ such that
$||\nabla F(\boldsymbol{\theta})||\leq \varepsilon$, with $\varepsilon> 0$, is called $\varepsilon$-critical point of $F$. The set of $\varepsilon$-critical points of $F$ is denoted by $\mathrm{crit}_\varepsilon F$.
\end{definition}
In this section, we prove that when the step-size is sufficiently small and DGD is initialized in a compact set, the iterates $\{\bx_i^\nu\}$, $i\in[n]$, converge to an arbitrarily small neighborhood of critical points of $F$--the result is formally stated in Theorem \ref{prop_DGD_local_convergence}. Roughly speaking, this is proved chaining the following intermediate results: \begin{description}
\item[i) Lemma \ref{L_criticalPoints_are_alpha_statPoints}:] Every limit point of DGD is an $\mathcal{O}(\alpha)$-critical point of $F$
\item[ii) Lemma \ref{Y_set_found}:] Every sequence generated by DGD for given $\alpha>0$ and initialization in a compact set, is enclosed in some compact set, for all $\alpha\downarrow 0$; and \item[iii) Lemma \ref{approximate_critical_points_lemma}:] Any $\epsilon$-critical point of $F$ {achievable} by DGD
is arbitrarily close to a critical point of $F$, when $\epsilon$ is sufficiently small.
\end{description}
Lemma \ref{L_criticalPoints_are_alpha_statPoints} implies that, for any given $\epsilon>0$, one can find arbitrarily small $\alpha>0$ so that every limit point of each $\{\bx_i^\nu\}$ (whose existence is guaranteed by Lemma \ref{Y_set_found}) is an $\epsilon$-critical point of $F$. Finally, Lemma \ref{approximate_critical_points_lemma} guarantees that every such $\epsilon$-critical point can be made arbitrarily close to a critical point of $F$ as $\epsilon \downarrow 0$. The proof of the above three lemmata follows.
\begin{lemma}
\label{L_criticalPoints_are_alpha_statPoints}
Let Assumptions \hyperlink{P_assumption_prime}{2.1'} and \ref{Net-Assump} hold. Given arbitrary $\mathbf{x}^0\in\mathbb{R}^{mn}$ and $0<\alpha<\sigma_\mathrm{min}(\mathbf{I}+\mathbf{D})/L_c$,
every limit point
$\mathbf{x}^\infty=[\mathbf{x}_1^{\infty\top},\ldots,\mathbf{x}_n^{\infty\top}]^\top$
of $\{\mathbf{x}^\nu\}$ generated by the DGD algorithm satisfies $\bar{\mathbf{x}}^\infty\in \mathrm{crit}_{K'\alpha}F$, with $\bar{\mathbf{x}}^\infty\triangleq (1/n)\,\sum_{i=1}^n\mathbf{x}^\infty_i$ and $K'= n\sqrt{n}L_cH/$ $(1-\sigma_2)$, where $H$ and $\sigma_2$ are defined in Theorem \ref{DGD_conv_cons_thm}.
\end{lemma}
\begin{proof}
By Theorem \ref{DGD_conv_cons_thm}(ii), $({\mathbf{1}\otimes\mathbf{I}})^\top \nabla L_\alpha(\mathbf{x}^\infty)=\mathbf{0}$, which using (\ref{LyapunovFunc_intro}) and the column stochasticity of $\mathbf{D}$ yields
$({\mathbf{1}\otimes\mathbf{I}})^\top$ $\nabla F_c(\mathbf{x}^\infty)= \mathbf{0}$. Hence,
\begin{equation}
\label{nablaD_epsCriticalbound}
\begin{aligned}
\norm{\nabla F(\bar{\mathbf{x}}^\infty)}= & \norm{({\mathbf{1}\otimes\mathbf{I}})^\top \left(\nabla F_c(\mathbf{1}\otimes\bar{\mathbf{x}}^\infty)- \nabla F_c(\mathbf{x}^\infty)\right)}
\\
\leq & L_c\sqrt{n}\norm{\mathbf{x}^\infty-\mathbf{1}\otimes\bar{\mathbf{x}}^\infty}\overset{(a)}{\leq} \alpha\cdot \frac{n\sqrt{n}L_cH}{1-\sigma_2},
\end{aligned}\vspace{-0.3cm}
\end{equation}
where
in (a) we used Theorem \ref{DGD_conv_cons_thm}(i).
\end{proof}
To proceed, we limit DGD initialization to $\mathbf{x}_i^0\in\mathcal{X}_i$, $i\in[n],$ where $\mathcal{X}_i^0\subseteq \mathbb{R}^{m}$ is some compact set with positive Lebesgue measure.
\begin{lemma}
\label{Y_set_found}
Consider Problem \eqref{eq:P}, under Assumptions \hyperlink{P_assumption_prime}{2.1'}, \ref{Strip_assumption} and \ref{Net-Assump}.
Let $\{\mathbf{x}^\nu(\alpha,\bx^0)\}$ be any sequence generated by DGD under Assumption \ref{DoublyStochastic_D}, with step-size $\alpha$ and initialization $\mathbf{x}^0$. Then, there exists a bounded set $\mathcal{Y}$ such that $\{\mathbf{x}^\nu(\alpha,\bx^0)\}\subseteq \mathcal{Y}$, for all $0<{\alpha}\leq\alpha_{\max}= \sigma_\mathrm{min}(\mathbf{I}+\mathbf{D})/L_c$ and $\mathbf{x}_i^0\in\mathcal{X}_i\subseteq\mathcal{B}^m_R, i\in[n]$, where $R$ is defined in Assumption \ref{Strip_assumption}.
\end{lemma}
\begin{proof}
We proceed by induction. For the sake of notation, throughout the proof, we will use for $\bx^\nu(\alpha,\bx^0)$
the shorthand $\bx^\nu$.
Define $h\triangleq \max_{i\in[n],\boldsymbol{\theta}\in\mathcal{B}^m_R}||\nabla f_i(\boldsymbol{\theta})||$. By assumption, there holds $\mathbf{x}_i^0\in\mathcal{B}^m_R$, for all $i$. Suppose $\mathbf{x}_i^\nu\in\mathcal{B}^m_R$, for all $i$. If $\mathbf{x}_i^{\nu}\in\mathcal{B}^m_{R-\epsilon}$ and $\alpha\leq \epsilon D_{ii}/h$, then $\mathbf{x}_i^\nu-\frac{\alpha}{D_{ii}}\nabla f_i(\mathbf{x}_i^\nu)\in\mathcal{B}^m_R$, since \vspace{-0.2cm}
\begin{equation}
\big\|\mathbf{x}_i^\nu-\frac{\alpha}{D_{ii}}\nabla f_i(\mathbf{x}_i^\nu)\big\|\leq \norm{\mathbf{x}_i^\nu}+\frac{\alpha}{D_{ii}} \norm{\nabla f_i(\mathbf{x}_i^\nu)}
\leq R-\epsilon+\frac{\alpha h}{D_{ii}}.
\end{equation}
If $\mathbf{x}_i^{\nu}\in\mathcal{S}_{R,\epsilon}$ and $\alpha\leq 2D_{ii}\delta(R-\epsilon)/h^2$, then $\mathbf{x}_i^\nu-\frac{\alpha}{D_{ii}}\nabla f_i(\mathbf{x}_i^\nu)\in\mathcal{B}^m_R$, since
\begin{equation}
\begin{aligned}
\big\|\mathbf{x}_i^\nu-\frac{\alpha}{D_{ii}}\nabla f_i(\mathbf{x}_i^\nu)\big\|^2 & =
\norm{\mathbf{x}_i^\nu}^2-\frac{2\alpha||\mathbf{x}_i^\nu||}{D_{ii}}\left\langle \frac{\mathbf{x}_i^\nu}{||\mathbf{x}_i^\nu||},\nabla f_i(\mathbf{x}_i^\nu)\right\rangle+\frac{\alpha^2}{D^2_{ii}} \norm{\nabla f_i(\mathbf{x}_i^\nu)}^2\\ & \leq R^2-\frac{2\alpha \delta(R-\epsilon)}{D_{ii}}+\frac{\alpha^2 h^2}{D^2_{ii}}.
\end{aligned}
\end{equation}
By agents' updates $\mathbf{x}_i^{\nu+1}=\sum_{j\neq i}D_{ij}\mathbf{x}_j^\nu+D_{ii}(\mathbf{x}_i^\nu-\frac{\alpha}{D_{ii}}\nabla f_i(\mathbf{x}_i^\nu))$ and convexity of the norm, we conclude that if $\mathbf{x}_i^{\nu}\in\mathcal{B}^m_{R}$, for all $i$, and $0<\alpha\leq \alpha_b\triangleq \min_i\min\{\epsilon D_{ii}/h,2D_{ii}\delta(R-\epsilon)/h^2\}$, then $\mathbf{x}_i^{\nu+1}\in\mathcal{B}^m_R$. This proves that, for $\alpha\in(0,\alpha_b]$, any sequence $\{\bx^{\nu}_i\}$ initialized in $\mathcal{B}^m_R$ lies in $\mathcal{B}^m_R$, for all $i$.
We prove now the same result for $\alpha\in[\alpha_b,\sigma_\mathrm{min}(\mathbf{I}+\mathbf{D})/L_c]$. Note that since each $f_i$ is coercive (cf. Assumption \hyperlink{P_assumption_prime}{2.1'(ii)}), any sublevel set of $L_\alpha$ is compact. Also, since $\{L_\alpha(\mathbf{x}^\nu)\}$ is non-increasing for all $\alpha\in(0,\sigma_\mathrm{min}(\mathbf{I}+\mathbf{D})/L_c]$ (cf. \cite[lemma 2]{WYin_DGD_SIAM2016}), then $\{\mathbf{x}^\nu\}\subseteq\mathcal{L}_{L_\alpha}(F_c(\mathbf{x}^0)+\frac{1}{2\alpha}||\mathbf{x}^0||^2_{\mathbf{I-W}})$, and furthermore,
\begin{equation}
\begin{aligned}
& \mathcal{L}_{L_\alpha}\left(F_c(\mathbf{x}^0)+\frac{1}{2\alpha}||\mathbf{x}^0||^2_{\mathbf{I-W}}\right)\subseteq \mathcal{L}_{L_\alpha}\left(F_c(\mathbf{x}^0)+\frac{1}{2\alpha_b}||\mathbf{x}^0||^2_{\mathbf{I-W}}\right)
\\
\subseteq & \mathcal{L}_{F_c}\left(F_c(\mathbf{x}^0)+\frac{1}{2\alpha_b}||\mathbf{x}^0||^2_{\mathbf{I-W}}\right) \subseteq \mathcal{L}_{F_c} \left(\max_{\mathbf{x}_i^0\in\mathcal{B}^m_R, i\in [n]} \left\{ F_c(\mathbf{x}^0)+\frac{1}{2\alpha_b}||\mathbf{x}^0||^2_{\mathbf{I-W}}\right\}\right).
\end{aligned}
\end{equation}
Since $||\mathbf{x}^0||^2_{\mathbf{I-W}}\leq 2||\mathbf{x}^0||^2$, it follows
\begin{equation}\label{L_0_Y_inclusion}
\mathcal{L}_{L_\alpha}\left(F_c(\mathbf{x}^0)+\frac{1}{2\alpha}||\mathbf{x}^0||^2_{\mathbf{I-W}}\right)\subseteq \underbrace{\mathcal{L}_{F_c}\left(\max_{\mathbf{x}_i^0\in\mathcal{B}^m_R, i\in [n]} \left\{\sum_{i=1}^n f_i(\mathbf{x}_i^0)\right\}+\frac{R^2}{\alpha_b}\right)}_{\triangleq \bar{\mathcal{L}}}.
\end{equation}
The statement of the lemma hods with $\mathcal{Y}=\bar{\mathcal{L}}\cup\prod_{i=1}^n\mathcal{B}^m_{R}$.
\end{proof}
The following lemma shows that any $\epsilon$-critical point of $F$ achievable by DGD (i.e., any point in $\mathrm{crit}_\varepsilon F\cap\bar{\mathcal{Y}}$) can be made arbitrarily close to a critical point of $F$, when $\epsilon>0$ (and thus $\alpha>0$) is sufficiently small.
\begin{lemma}
\label{approximate_critical_points_lemma}
Suppose $F:\mathbb{R}^{m}\rightarrow\mathbb{R}$ is continuously differentiable. For any given compact set $\bar{\mathcal{Y}}\subseteq \mathbb{R}^m$, there holds
\vspace{-0.3cm}
\begin{equation}\label{eq:vanishing_crit_f_error_statement}
\lim_{\varepsilon\rightarrow 0}~\max_{\mathbf{q}\in\mathrm{crit}_\varepsilon F\cap\bar{\mathcal{Y}}}~\mathrm{dist}(\mathbf{q},\mathrm{crit}~F)=0.\vspace{-0.1cm}
\end{equation}
\end{lemma}
\begin{proof}
We prove the lemma by contradiction. Suppose \vspace{-0.2cm}
\begin{equation}\label{eq:vanishing_crit_f_error}
\limsup_{\varepsilon\rightarrow 0}~\max_{\mathbf{q}\in\mathrm{crit}_\varepsilon F\cap\bar{\mathcal{Y}}}~\mathrm{dist}(\mathbf{q},\mathrm{crit}~F)=\gamma>0.\vspace{-0.2cm}
\end{equation}
Then, one can construct $\{\mathbf{q}^\nu\}$ with $\mathbf{q}^\nu \in\mathrm{crit}_{1/\nu} F\cap\bar{\mathcal{Y}}$ such that $\mathrm{dist}(\mathbf{q}^\nu,\mathrm{crit} F)\geq \gamma$ for all $\nu\in\mathbb{N}$. Since $\nabla F$ is continuous,
$\mathrm{crit_{1}} F$ is closed and $\mathrm{crit_{1}} F\cap\bar{\mathcal{Y}}$ is compact.
Note that $\{\mathbf{q}^\nu\}\subseteq\mathrm{crit_{1}} F\cap\bar{\mathcal{Y}}$, which ensures
$\{\mathbf{q}^\nu\}$ is bounded. Let $\{\mathbf{q}^{t_\nu}\}$ be a convergent subsequence of $\{\mathbf{q}^\nu\}$; its limit point $\mathbf{q}^\infty$ satisfies $\mathrm{dist}(\mathbf{q}^\infty,\mathrm{crit}~F)\geq \gamma$. By construction, for any $\acute{\nu}\in\mathbb{N}$, $\{\mathbf{q}^{t_\nu}\}$ eventually settles in $\mathrm{crit}_{1/\acute{\nu}}~F\cap\bar{\mathcal{Y}}$, thus $\mathbf{q}^\infty\in\mathrm{crit}_{1/\acute{\nu}}~F\cap\bar{\mathcal{Y}}$. This means that $||\nabla F(\mathbf{q}^{\acute{\nu}})||\leq 1/\acute{\nu}$, for all $\acute{\nu}\in\mathbb{N}$, implying $||\nabla F(\mathbf{q}^\infty)||=0$. Hence $\mathrm{dist}(\mathbf{q}^\infty,\mathrm{crit}~F)=0$, which contradicts \eqref{eq:vanishing_crit_f_error}.
\end{proof}
We can now combine Lemmas \ref{L_criticalPoints_are_alpha_statPoints}-\ref{approximate_critical_points_lemma} with Theorem \ref{DGD_conv_cons_thm}(i) and state the main result of this section.
\begin{theorem}
\label{prop_DGD_local_convergence}
Let Assumptions \hyperlink{P_assumption_prime}{2.1'}, \ref{Strip_assumption} and \ref{Net-Assump} hold. Let $\epsilon>0$. There exists $\bar{\alpha}>0$ (which depends on $\epsilon$)
such that with any initialization $\mathbf{x}_i^0\in\mathcal{X}_i^0\subseteq \mathcal{B}^m_R$ ($R>0$ is defined in Assumption \ref{Strip_assumption}), $i\in [n]$, and any step-size $0<\alpha\leq \bar{\alpha}$, all the limit points
$\mathbf{x}^\infty(\alpha,\bx^0)=[\mathbf{x}_1^{\infty}(\alpha,\bx^0)^\top,\ldots,\mathbf{x}_n^{\infty}(\alpha,\bx^0)^\top]^\top$
of the sequence $\{\mathbf{x}^\nu(\alpha,\bx^0)\}$, generated by DGD satisfies
\begin{equation}\label{eq:crit_neig}
\mathrm{dist}\big(\bar{\mathbf{x}}^\infty(\alpha,\bx^0),\mathrm{crit\,F}\big)<\epsilon \quad \text{and} \quad \norm{\mathbf{x}^\infty(\alpha,\bx^0)-\mathbf{1}\otimes \bar{\mathbf{x}}^\infty(\alpha,\bx^0)}<\epsilon,\vspace{-0.1cm}
\end{equation}
where $\bar{\mathbf{x}}^\infty(\alpha,\bx^0)\triangleq (1/n)\,\sum_{i=1}^n\mathbf{x}^\infty_i(\alpha,\bx^0)$.
\end{theorem}
\begin{proof}
Combining Lemmata \ref{L_criticalPoints_are_alpha_statPoints}-\ref{approximate_critical_points_lemma} proves that there exists some $\alpha_1>0$ such that $\mathrm{dist}(\bar{\mathbf{x}}^\infty(\alpha,\bx^0), \mathrm{crit\,F})<\epsilon$, for all $\alpha\leq \alpha_1$. In addition, Theorem \ref{DGD_conv_cons_thm}(i), with $H=\sup_{\bx\in\mathcal{Y}} F_c(\bx)$, implies that there exists some $\alpha_2>0$ such that $||\mathbf{x}^\infty(\alpha,\bx^0)-\mathbf{1}\otimes \bar{\mathbf{x}}^\infty(\alpha,\bx^0)||<\epsilon$, for all $\alpha\leq \alpha_2$. Hence, choosing $\bar{\alpha}= \min\{\alpha_1,\alpha_2\}$ proves \eqref{eq:crit_neig}.
\end{proof}
\subsection{DGD likely converges to a neighborhood of SoS solutions of $F$} \label{sec:SOS-DGD}
We study now second-order guarantees of DGD.
Our path to prove almost sure convergence to a neighborhood of SoS solutions of \eqref{eq:P} will pass through the non-convergence of DGD to strict saddles of $L_\alpha$ (cf. Theorem \ref{L_criticalPoints_are_alpha_statPoints}). Roughly speaking, our idea is to show that whenever $\bar{\mathbf{x}}^\infty=1/n\,\sum_{i=1}^n\mathbf{x}^\infty_i$ belongs to a sufficiently small neighborhood of a strict saddle of $F$ inside the region (\ref{eq:crit_neig}),
$\mathbf{x}^\infty=[\mathbf{x}_1^{\infty\top},\ldots,\mathbf{x}_n^{\infty\top}]^\top$
must be a \emph{strict saddle of $L_\alpha$}.
The escaping properties of DGD from strict saddles of $L_\alpha$ will then ensure that it is unlikely that $\{\bar{\mathbf{x}}^\nu= 1/n\,\sum_{i=1}^n\mathbf{x}^\nu_i\}$ gets trapped in a neighborhood of a strict saddle of $F$, thus ending in a neighborhood of a SoS solution of \eqref{eq:P}.
Proposition \ref{DGD_SSP_L_and_F_connect} makes this argument formal; in particular, conditions (i)-(iii) identify the neighborhood of a strict saddle of $F$ with the mentioned escaping properties.\vspace{-0.1cm}
\begin{proposition}
\label{DGD_SSP_L_and_F_connect}
Consider the setting of Lemma \ref{Y_set_found} and further assume that Assumption \ref{Lipschitz_Hessian} hold. Let $\bar{\mathcal{Y}}$ be the image of the compact set $\mathcal{Y}$ (defined in Lemma \ref{Y_set_found}) through the linear operator $(\mathbf{1}_n\otimes \mathbf{I}_m)^\top$. Suppose that the limit point
$\mathbf{x}^\infty=[\mathbf{x}_1^{\infty\top},\ldots,\mathbf{x}_n^{\infty\top}]^\top$
of $\{\mathbf{x}^\nu\}$, along with $\bar{\mathbf{x}}^\infty= 1/n\sum_{i=1}^n\mathbf{x}_{i}^\infty$, satisfy
\begin{enumerate}[label=(\roman*)]
\item $\mathrm{dist}(\bar{\mathbf{x}}^\infty,\mathrm{crit\,F})<\dfrac{\delta}{2L_{\nabla^2}}$,
\item $\norm{\mathbf{x}^\infty-\mathbf{1}\otimes \bar{\mathbf{x}}^\infty}<\dfrac{\delta}{2nL_{\nabla^2_c}}$,
\item There exists $\boldsymbol{\theta}^\ast\in\mathrm{proj}_{\mathrm{crit\, F}}(\bar{\mathbf{x}}^\infty)\cap \Theta^\ast_{ss}$,
\end{enumerate}
\vspace{.25cm}
for some $\delta$ such that $\delta\leq -\lambda_\mathrm{min}\left(\nabla^2 F(\boldsymbol{\theta}^\ast)\right),\forall \boldsymbol{\theta}^\ast\in\Theta_{ss}^\ast\cap\bar{\mathcal{Y}}$. Then, ${\mathbf{x}^\infty}$ is a strict saddle point of $L_\alpha$.
\end{proposition}
\begin{proof}
Given $\boldsymbol{\theta}\in\mathbb{R}^{m}$, let $\boldsymbol{\upsilon}(\boldsymbol{\theta})$ denote the unitary eigenvector of $\nabla^2 F(\boldsymbol{\theta})$ associated with the smallest eigenvalue, and define $\tilde{\boldsymbol{\upsilon}}(\boldsymbol{\theta})\triangleq \mathbf{1}\otimes \boldsymbol{\upsilon}(\boldsymbol{\theta})$. Then, we have
\begin{equation}\label{LHessian_bound0}
\hspace{-0.4cm}\begin{aligned}
&\tilde{\boldsymbol{\upsilon}}(\boldsymbol{\theta})^\top\nabla^2 L_\alpha (\mathbf{x}^\infty)\tilde{\boldsymbol{\upsilon}}(\boldsymbol{\theta})\overset{(a)}{=} \tilde{\boldsymbol{\upsilon}}(\boldsymbol{\theta})^\top\nabla^2 F_c (\mathbf{x}^\infty)\tilde{\boldsymbol{\upsilon}}(\boldsymbol{\theta})
\\
& \qquad {\leq}\, {\boldsymbol{\upsilon}}(\boldsymbol{\theta})^\top\nabla^2 F (\boldsymbol{\theta}){\boldsymbol{\upsilon}}(\boldsymbol{\theta})
\\
&\qquad \quad +||\nabla^2 F (\bar{\mathbf{x}}^\infty)-\nabla^2 F (\boldsymbol{\theta})||\norm{\boldsymbol{\upsilon}(\boldsymbol{\theta})}^2
+||\nabla^2 F_c (\mathbf{x}^\infty)-\nabla^2 F_c (\mathbf{1}\otimes\bar{\mathbf{x}}^\infty)||\norm{\tilde{\boldsymbol{\upsilon}}(\boldsymbol{\theta})}^2
\\
& \qquad \overset{(b)}{\leq}\, {\boldsymbol{\upsilon}}(\boldsymbol{\theta})^\top\nabla^2 F (\boldsymbol{\theta}){\boldsymbol{\upsilon}}(\boldsymbol{\theta})
+L_{\nabla^2}\norm{\bar{\mathbf{x}}^\infty-\boldsymbol{\theta}}+n\,L_{\nabla^2_c} \norm{\mathbf{x}^\infty-\mathbf{1}\otimes \bar{\mathbf{x}}^\infty}
\end{aligned}
\end{equation}
where (a) follows from $\tilde{\boldsymbol{\upsilon}}(\boldsymbol{\theta})\in\mathrm{null}(\mathbf{W}_D-\mathbf{I})$;
and (b) is due to Assumption \ref{Lipschitz_Hessian}.
Let us now evaluate \eqref{LHessian_bound0} at some $\boldsymbol{\theta}^\ast$ as defined in condition (iii) of the proposition; using ${\boldsymbol{\upsilon}}(\boldsymbol{\theta}^\ast)^\top\nabla^2 F (\boldsymbol{\theta}^\ast){\boldsymbol{\upsilon}}(\boldsymbol{\theta}^\ast)\leq -\delta$ and conditions (i) and (ii), yields
$\tilde{\boldsymbol{\upsilon}}(\boldsymbol{\theta}^\ast)^\top\nabla^2 L_\alpha (\mathbf{x}^\infty)$ $\tilde{\boldsymbol{\upsilon}}(\boldsymbol{\theta}^\ast)
<0.$
By the Rayleigh-Ritz theorem,
it must be $\lambda_\mathrm{min}(\nabla^2 L_\alpha (\mathbf{x}^\infty))<0$. This, together with $\mathbf{x}^\infty\in\mathrm{crit}~L_{\alpha}$ (cf. Theorem \ref{DGD_conv_cons_thm}(ii)), proves the proposition.
\end{proof}
Invoking now Theorem \ref{prop_DGD_local_convergence}, we infer that there exists a sufficiently small $\alpha>0$ such that conditions (i) and (ii) of Proposition \ref{DGD_SSP_L_and_F_connect} are always satisfied, implying that ${\mathbf{x}^\infty}$ is a strict saddle of $L_\alpha$ if there exists a strict saddle of $F$ ``close'' to $\bar{\mathbf{x}}^\infty$ [in the sense of (iii)].
This is formally stated next.
\begin{corollary}
\label{DGD_SSP_L_and_F_connect_2_AMIREM}
Consider the setting of Theorem \ref{prop_DGD_local_convergence} and Proposition \ref{DGD_SSP_L_and_F_connect}.
There exists a sufficiently small $\alpha>0$ such that, if $\mathrm{proj}_{\mathrm{crit\, F}}(\bar{\mathbf{x}}^\infty)\cap \Theta^\ast_{ss}\neq \emptyset$, then ${\mathbf{x}^\infty}$ is a strict saddle of $L_\alpha$.
\end{corollary}
To state our final result, let us introduce the following merit function: given
$\mathbf{x}=[\mathbf{x}_1^{\top},\ldots,\mathbf{x}_n^{\top}]^\top$
let\vspace{-0.3cm}
\begin{equation*}
M(\mathbf{x})\triangleq \max\Big(\mathrm{dist}(\bar{\mathbf{x}},\mathcal{X}_{SOS}),\norm{\mathbf{x}-\mathbf{1}\otimes \bar{\mathbf{x}}}\Big),\vspace{-0.1cm}
\end{equation*}
where $\mathcal{X}_{SOS}$ denotes the set of SoS solutions of \eqref{eq:P}, and $\bar{\mathbf{x}}= 1/n\sum_{i=1}^n \mathbf{x}_i$. $M(\mathbf{x})$ capture the distance of the average $\bar{\mathbf{x}}$ from the set of SoS solutions of \eqref{eq:P} and well as the consensus disagreement of the agents' local variables $\bar{\mathbf{x}}_i$. \vspace{-0.1cm}
\begin{theorem}
\label{SOG_of_DGD_thm}
Consider Problem \eqref{eq:P} under Assumptions \hyperlink{P_assumption_prime}{2.1'}, \ref{Lipschitz_Hessian}, \ref{Strip_assumption}, and \ref{Net-Assump}; further assume that {each $f_i$ is a {K\L} function}. For every $\epsilon>0$, there exists sufficiently small $0<\bar{\alpha}<\frac{\sigma_\mathrm{min}(\mathbf{D})}{L_c}$ such that \vspace{-0.2cm}
\begin{equation*}
\mathbb{P}_{\mathbf{x}^0}\big(M(\mathbf{x}^\infty)\leq \epsilon\big)=1,
\end{equation*}
where
$\mathbf{x}^\infty=[\mathbf{x}_1^{\infty\top},\ldots,\mathbf{x}_n^{\infty\top}]^\top$
is the limit point of the sequence $\{\mathbf{x}^\nu\}$ generated by the DGD algorithm \eqref{DGD_update} with $\alpha\in (0,\bar{\alpha}]$, the weight matrix $\mathbf{D}$ satisfying Assumptions \ref{DoublyStochastic_D} and \ref{nonsingular_W}, and initialization $\mathbf{x}^0\in \prod_{i=1}^n \mathcal{X}^0_i\subseteq \prod_{i=1}^n \mathcal{B}^m_R$; $R$ is defined in Assumption \ref{Strip_assumption} and each $\mathcal{X}_i^0$ has positive Lebesgue measure;
and the probability is taken over the initialization $\mathbf{x}^0\in \prod_{i=1}^n \mathcal{X}^0_i$. Furthermore, any $\boldsymbol{\theta}^\ast\in\mathrm{proj}_{\mathrm{crit~F}}(\bar{\mathbf{x}}^\infty)$ is almost surely a SoS solution of $F$ where $\bar{\mathbf{x}}^\infty=(1/n)\sum_{i=1}^n \mathbf{x}^\infty_i$.
\end{theorem}
\begin{proof}
For sufficiently small $\alpha<\bar\alpha_1$, if $\mathrm{proj}_{\mathrm{crit~F}}(\bar{\mathbf{x}}^\infty)$ contains a strict saddle point of $F$, then $\mathbf{x}^\infty$ is also a strict saddle point of $L_\alpha$ (by Corollary \ref{DGD_SSP_L_and_F_connect_2_AMIREM}). Let also $\bar\alpha_2$ be a sufficiently small step-size such that every limit point $\mathbf{x}^\infty$ satisfies $\mathrm{dist}(\bar{\mathbf{x}}^\infty,\mathrm{crit}~F)\leq \epsilon$ and $\norm{\mathbf{x}^\infty-\mathbf{1}\otimes \bar{\mathbf{x}}^\infty}\leq\epsilon$ (by Theorem~\ref{prop_DGD_local_convergence}). Now consider DGD update \eqref{DGD_update} with $\alpha<\min\{\bar\alpha_1,\bar\alpha_2\}$ and $\mathbf{x}^0$ being drawn randomly from the set of probability one measure $\prod_{i=1}^n \mathcal{X}^0_i$ for which the algorithm converges to a SoS solution of $L_\alpha$ (by Theorem \ref{DGD_naive_SOG}\footnote{Note that the conclusion of Theorem \ref{DGD_naive_SOG} is valid also when the set of initial points is restricted to $\prod_{i=1}^n\mathcal{X}_i^0$, as $\prod_{i=1}^n\mathcal{X}_i^0$ has positive measure (the Cartesian product of sets with positive measure has positive measure-- cf. \cite[Sec. 35]{halmos1976measure}).}). Finally, by the above properties of $\alpha$, it holds that $M(\mathbf{x}^\infty)\leq\epsilon$ and $\mathrm{proj}_{\mathrm{crit~F}}(\bar{\mathbf{x}}^\infty)$ must contain only SoS solutions of $F$. Therefore, there exists a $\boldsymbol{\theta}^\ast\in \mathrm{crit}~F$ such that $\boldsymbol{\theta}^\ast\in\mathcal{X}_{SoS}$ and $\|\bar{\mathbf{x}}^\infty-\boldsymbol{\theta}^\ast\|\le \epsilon$.
\end{proof}
\begin{remark}\label{local_lip_extension_remark}
All (first- and second-order) convergence results of DGD established in this section remain valid when $\nabla f_i$'s are not globally Lipschitz continuous [Assumption \ref{P_assumption}(i)] but Assumption \ref{Strip_assumption} holds. Specifically, Theorems \ref{DGD_conv_cons_thm}, \ref{DGD_naive_SOG}, \ref{prop_DGD_local_convergence} and Lemmata \ref{L_criticalPoints_are_alpha_statPoints}-\ref{Y_set_found} hold if one replaces Assumption \ref{P_assumption}(i) with Assumption \ref{Strip_assumption} and the global Lipschitz constant $L_c$ with the Lipschitz constant of $\nabla F_c$ \emph{restricted} to the compact set $\tilde{\mathcal{Y}}$, defined in Appendix \ref{DGD_conv_cons_thm_extended_proof_appendix}, where we refer to for the technical details.
\end{remark}
\section{DOGT Algorithms} \label{sec:DOGT}
{
The family of DOGT algorithms is introduced in Sec. \ref{DOGT_method_intro}. We begin here rewriting \eqref{eq:DOGT_x_update_intro}-\eqref{eq:DOGT_track_update_intro} in matrix/vector form. Denoting $\bx^\nu\triangleq [\bx_1^{\nu\top},\ldots \bx_n^{\nu\top}]^\top$ and $\by^\nu\triangleq [\by_1^{\nu\top},\ldots \by_n^{\nu\top}]^\top$, we have
\begin{equation}\label{NEXT_generic}
\left\lbrace\begin{aligned}
\mathbf{x}^{\nu+1} &=\mathbf{W}_R\, \mathbf{x}^{\nu}-\alpha\,\mathbf{y}^\nu,
\\
\mathbf{y}^{\nu+1} &=\mathbf{W}_C\, \mathbf{y}^{\nu}+\nabla F_c\big(\mathbf{x}^{\nu+1}\big)-\nabla F_c\big( \mathbf{x}^\nu\big),
\end{aligned}\right.
\end{equation}
where $\mathbf{W}_R\triangleq \mathbf{R} \otimes \mathbf{I}_m$ and $\mathbf{W}_C\triangleq \mathbf{C} \otimes \mathbf{I}_m$ with $\mathbf{R}\triangleq (R_{ij})_{i,j=1}^n$ and $\mathbf{C}\triangleq (C_{ij})_{i,j=1}^n$ being some \emph{column-stochastic} and \emph{row-stochastic} matrices (respectively) compliant to the graph $\mathcal{G}$ (cf. Assumption \ref{matrix_C_R} below).
The initialization of (\ref{NEXT_generic}) is set to $\mathbf{x}^{0}\in \mathbb{R}^{mn}$ and $\mathbf{y}^0\in\nabla F_c(\mathbf{x}^0)+\text{span}\left(\mathbf{W}_C-\mathbf{I}\right)$. Note that the latter condition is instrumental to preserve the \emph{total-sum} of the ${y}$-variables, namely $\sum_i\mathbf{y}_i^\nu=\sum_if_i(\bx_i^\nu)$ (which holds due to the column-stochasticity of matrix $\mathbf{C}$--cf. Assumption \ref{matrix_C_R}). This property is imperative for the ${y}$-variables to track the sum-gradient. Notice that the condition used in the literature \cite{NEXT,Harnessing_Na_Li, SONATA_technical_report, nedich2016achieving,xin2018linear}--$\mathbf{y}^0=\nabla F_c(\mathbf{x}^0)$--is a special case of the proposed initialization. On the practical side,
this initialization can be enforced in a distributed way, with minimal coordination. For instance, agents first choose independently a vector $\mathbf{y}_i^{-1}\in \mathbb{R}^m$; then they run one step of consensus on the $y$-variables using the values ${y}_i^{-1}$'s and weights matrix $\mathbf{C}$, and set $\mathbf{y}_i^0 = \nabla f_i(\bx_i^0) +\sum_{j\in \mathcal{N}_i^{in}} C_{ij} \mathbf{y}_j^{-1}- \mathbf{y}_i^{-1}$, resulting in $\mathbf{y}^0\in \nabla F_c(\mathbf{x}^0)+ \text{span}(\mathbf{W}_C-\mathbf{I})$.
}
Different choices for $\mathbf{R}$ and $\mathbf{C}$ are possible, resulting in different existing algorithms. For instance, if $\mathbf{R}=\mathbf{C}\in\mathcal{M}_n(\mathbb{R})$ are doubly-stochastic matrices compliant to the graph $\mathcal{G}$, \eqref{NEXT_generic} reduces to the NEXT algorithm \cite{dilorenzo2015distributed,NEXT} (or the one in \cite{xu2015augmented}, when \eqref{eq:P} is convex). If $\mathbf{R}$ and $\mathbf{C}$ are allowed to be time-varying (suitably chosen)
(\ref{NEXT_generic}) reduces to the SONATA algorithm applicable to (possibly time-varying) digraphs \cite{sun2016distributed,Scutari_Ying_LectureNote,SONATA_technical_report, YingDanScuConvergence18} [or the one later proposed in \cite{nedich2016achieving} for strongly convex instances of \eqref{eq:P}]. Finally, if $\mathbf{R}$ and $\mathbf{C}$ are chosen according to Assumption~\ref{matrix_C_R} below, the scheme (\ref{NEXT_generic}) becomes the algorithm proposed independently in \cite{ANedich_GeoConvAlg_arxive_2018} and \cite{xin2018linear}, for strongly convex objectives in \eqref{eq:P}, and implementable over fixed digraphs.
\begin{assumption}(On the matrices $\mathbf{R}$ and $\mathbf{C}$)\label{matrix_C_R} The weight matrices $\mathbf{R},\mathbf{C}\in \mathcal{M}_n(\mathbb{R})$ satisfy the following:
\begin{itemize}
\item[(i)] $\mathbf{R}$ is nonnegative row-stochastic and ${R}_{ii}>0$, for all $i\in [n]$;
\item[(ii)] $\mathbf{C}$ is nonnegative column-stochastic and ${C}_{ii}>0$, for all $i\in [n]$;
\item[(iii)] The graphs $\mathcal{G}_R$ and $\mathcal{G}_{C^\top}$ each contain at least one spanning tree; and $\mathcal{R}_R\cap\mathcal{R}_{{C^\top}}\neq \emptyset$.
\end{itemize}
\end{assumption}
It is not difficult to check that matrices $\mathbf{R}$ and $\mathbf{C}$ above exist if and only if the digraph $\mathcal{G}$ is strongly connected; however, $\mathcal{G}_R$ and $\mathcal{G}_{C^\top}$ need not be so. Several choices for such matrices are discussed in \cite{ANedich_GeoConvAlg_arxive_2018,xin2018linear}. Here, we only point out the following property of $\mathbf{R}$ and $\mathbf{C}$, as a consequence of Assumption \ref{matrix_C_R}, which will be used in our analysis. The result is a consequence of \cite[Lemma 1]{xin2019distributed}.
{
\begin{lemma} \label{R_C_norms_plus_realAnalyticity}
Given $\mathbf{R}$ and $\mathbf{C}$ satisfying Assumption \ref{matrix_C_R} with stochastic left eigenvector $\mathbf{r}$ (resp. right eigenvector $\mathbf{c}$ ) of $\mathbf{R}$ (resp. $\mathbf{C}$) associated with the eigenvalue one, then there exist matrix norms\vspace{-0.1cm}
\begin{align}
||\mathbf{X}||_R&\triangleq ||\diag(\sqrt{\mathbf{r}})\mathbf{X}\diag(\sqrt{\mathbf{r}})^{-1}||_2,\label{R_matNorm}
\\
||\mathbf{X}||_C&\triangleq ||\diag(\sqrt{\mathbf{c}})^{-1}\mathbf{X}\diag(\sqrt{\mathbf{c}})||_2,\label{C_matNorm}
\end{align}
such that $\rho_R\triangleq \|\mathbf{R}-\mathbf{1}\mathbf{r}^\top\|_R<1$ and $\rho_C\triangleq \|\mathbf{C}-\mathbf{c}\mathbf{1}^\top\|_C<1$. Furthermore, $\mathbf{r}^\top\mathbf{c}>0$.
\end{lemma}
Using Lemma \ref{R_C_norms_plus_realAnalyticity}, it is not difficult to check that the following properties hold:\vspace{-0.1cm}
\begin{align}
&\rho_R=\sigma_2\left(\diag(\sqrt{\mathbf{r}})\mathbf{R}\diag(\sqrt{\mathbf{r}})^{-1}\right),
\label{explicit_rho_R}
\\
&\rho_C=\sigma_2\left(\diag(\sqrt{\mathbf{c}})^{-1}\mathbf{C}\diag(\sqrt{\mathbf{c}})\right),\label{explicit_rho_C}
\\
&||\mathbf{R}||_R=||\mathbf{1r}^T||_R=||\mathbf{I}-\mathbf{1r}^T||_R=1,
\label{explicit_Rnorm_Eqs}
\\
&||\mathbf{C}||_C=||\mathbf{c1}^T||_C=||\mathbf{I}-\mathbf{c1}^T||_R=1.
\label{explicit_Cnorm_Eqs}
\end{align}
The vector norms associated with above matrix norms are\vspace{-0.1cm}
\begin{align}
||\mathbf{x}||_R&=||\diag(\sqrt{\mathbf{r}})\mathbf{x}||_2,\label{R_vecNorm}
\\
||\mathbf{x}||_C&=||\diag(\sqrt{\mathbf{c}})^{-1}\mathbf{x}||_2;\label{C_vecNorm}
\end{align}
and $||\cdot ||_a\leq K_{a,b}||\cdot||_b$ holds for $a,b\in\{R,C,2\}$ with\vspace{-0.1cm}
\begin{equation}\label{eq:def_K_1}
\begin{aligned}
&K_{R,2}=\sqrt{r_{\max}},\quad K_{2,R}=1/\sqrt{r_{\min}},
\\
&K_{C,2}=1/\sqrt{c_{\min}},\quad K_{2,C}=\sqrt{c_{\max}},
\\
&K_{R,C}=\sqrt{r_{\max}c_{\max}},\quad K_{C,R}=1/\sqrt{c_{\min}r_{\min}},
\end{aligned}
\end{equation}
where $r_{\min}$ (resp. $c_{\min}$) and $r_{\max}$ (resp. $c_{\max}$) are minimum and maximum elements of $\mathbf{r}$ (resp. $\mathbf{c}$).
}
Convergence of DOGT algorithms in the form \eqref{NEXT_generic} (with $\mathbf{R}$ and $\mathbf{C}$ satisfying Assumption~\ref{matrix_C_R}) has not been studied in the literature when $F$ is nonconvex. In next subsection we fill this gap and provide a full characterization of the convergence behavior of DOGT including its second-order guarantees.
\subsection{First-order convergence \& rate analysis}
\label{Subsequence_Conv_Subsec}
In this section, we study asymptotic convergence to first-order stationary solutions; we assume $m=1$ (scalar optimization variables); while this simplifies the notation, all the conclusions hold for the general case $m>1$.
As in \cite{ANedich_GeoConvAlg_arxive_2018}, define the weighted sums \vspace{-0.1cm} \begin{equation}\label{eq:def_w_vectors}
\bar{{x}}^{\nu}\triangleq \mathbf{r}^\top\mathbf{x}^\nu,\quad \bar{{y}}^{\nu}\triangleq \mathbf{1}^\top\mathbf{y}^\nu,\quad \text{and}\quad \bar{{g}}^{\nu}\triangleq \mathbf{1}^\top\nabla F_c(\mathbf{x}^\nu),\vspace{-0.1cm}
\end{equation}
where we recall that $\mathbf{r}$ is the Perron vector associated with $\mathbf{R}$ (cf.~Lemma~\ref{R_C_norms_plus_realAnalyticity}). { Note that $\nabla F_c$ is $L_c$-Lipschitz continuous with $L_c\triangleq L_\mathrm{max}$.
}
Using \eqref{NEXT_generic}, it is not difficult to check that the following holds
\vspace{-0.1cm}
\begin{equation}
\label{x_bar_eq}
\begin{aligned}
\bar{{x}}^{\nu+1
=\bar{{x}}^{\nu}-\zeta\alpha\bar{y}^\nu-\alpha\mathbf{r}^\top\left(\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu \right)\quad \text{and}\quad \bar{y}^\nu=\bar{g}^\nu,
\end{aligned}\vspace{-0.1cm}
\end{equation}
where $\mathbf{c}$ is the Perron vector associated with $\mathbf{C}$, and
$\zeta\triangleq \mathbf{r}^\top\mathbf{c}>0$ (cf.~Lemma \ref{R_C_norms_plus_realAnalyticity}).
\subsubsection{Descent on $F$} Using the descent lemma along with \eqref{x_bar_eq} yields
\begin{equation*}
\begin{aligned}
F(\bar{{x}}^{\nu+1})
& = F\left(\bar{{x}}^{\nu}-\zeta\alpha\bar{y}^\nu-\alpha\mathbf{r}^\top\left(\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu \right)\right)
\\
& \leq F(\bar{{x}}^{\nu})-\zeta\alpha\left\langle\nabla F({\bar{x}}^{\nu}),\bar{{y}}^{\nu}\right\rangle-\alpha\left\langle\nabla F({\bar{x}}^{\nu}),\mathbf{r}^\top\left(\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu \right)\right\rangle
\\
&\quad +\frac{L}{2}\norm{\zeta\alpha\bar{y}^\nu+\alpha\mathbf{r}^\top\left(\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu \right)}^2.
\end{aligned}
\end{equation*}
Adding/subtracting suitably chosen terms we obtain
\begin{equation}
\begin{aligned}
F(\bar{{x}}^{\nu+1})
\leq & F(\bar{{x}}^{\nu})-\zeta\alpha\left\langle\nabla F({\bar{x}}^{\nu})-\bar{y}^\nu,\bar{{y}}^{\nu}\right\rangle-\zeta\alpha|\bar{y}^\nu|^2
\\
&-\alpha\left\langle\nabla F({\bar{x}}^{\nu})-\bar{y}^\nu,\mathbf{r}^\top\left(\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu \right)\right\rangle-\alpha \left\langle \bar{y}^\nu,\mathbf{r}^\top\left(\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu \right)\right\rangle
\\
&+L\zeta^2\alpha^2|\bar{y}^\nu|^2+L\alpha^2\norm{\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu}^2
\\
\leq & F(\bar{{x}}^{\nu})+\frac{\zeta\alpha}{2\epsilon_1}|\nabla F({\bar{x}}^{\nu})-\bar{y}^\nu|^2+\frac{\zeta\alpha\epsilon_1}{2}|\bar{y}^{\nu}|^2
-\zeta\alpha|\bar{y}^\nu|^2
\\
&+\frac{\alpha}{2}|\nabla F({\bar{x}}^{\nu})-\bar{y}^\nu|^2+\frac{\alpha}{2}\norm{\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu}^2
+\frac{\alpha \epsilon_2}{2}|\bar{y}^{\nu}|^2+\frac{\alpha }{2\epsilon_2}\norm{\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu}^2
\\
&+L\zeta^2\alpha^2|\bar{y}^\nu|^2+L\alpha^2\norm{\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu}^2
\\
=& F(\bar{{x}}^{\nu})+\left(\frac{\zeta\alpha\epsilon_1}{2}-\zeta\alpha+\frac{\alpha \epsilon_2}{2}+L\zeta^2\alpha^2\right)|\bar{y}^{\nu}|^2
\\
&+\left(\frac{\zeta\alpha}{2\epsilon_1}+\frac{\alpha}{2}\right)|\nabla F({\bar{x}}^{\nu})-\bar{y}^\nu|^2
+\left(\frac{\alpha}{2}+\frac{\alpha }{2\epsilon_2}+L\alpha^2\right)\norm{\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu}^2,
\end{aligned}
\label{descent1}
\end{equation}
where $\epsilon_1$ and $\epsilon_2$ are some arbitrary positive quantities (to be chosen). {
By $\bar{y}^\nu=\bar{g}^\nu$ [cf.~\eqref{x_bar_eq}], it holds that \vspace{-0.1cm}
\begin{equation}
|\nabla F({\bar{x}}^{\nu})-\bar{y}^\nu|= \left|\sum_{i=1}^n \nabla f_i(\bar{{x}}^\nu)-\sum_{i=1}^n \nabla f_i({x}_i^\nu)\right|
\leq L_c\sqrt{n} \norm{\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu}.
\label{grad_error_bound}
\end{equation}
}
Combining \eqref{descent1} and \eqref{grad_error_bound} yields
{
\begin{equation}
\begin{aligned}
& F(\bar{{x}}^{\nu+1})
\\
\leq &\,
F(\bar{{x}}^{\nu})+\left(\frac{\zeta\alpha\epsilon_1}{2}-\zeta\alpha+\frac{\alpha \epsilon_2}{2}+L\zeta^2\alpha^2\right)|\bar{y}^{\nu}|^2
\\
& +nL_c^2K_{2,R}^2\left(\frac{\zeta\alpha}{2\epsilon_1}+\frac{\alpha}{2}\right)\norm{\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu}_R^2
+K_{2,C}^2\left(\frac{\alpha}{2}+\frac{\alpha }{2\epsilon_2}+L\alpha^2\right)\norm{\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu}_C^2,
\end{aligned}
\label{descent2}
\end{equation}
where $K_{2,R}=1/\sqrt{r_{\min}}$ and $K_{2,C}=\sqrt{c_{\max}}$ [cf. \eqref{eq:def_K}].}
\subsubsection{Bounding the consensus and gradient tracking errors}
Let us bound the consensus error $\|\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu\|_R$. Using $\|\mathbf{z}+\mathbf{w}||^2_R\leq (1+\epsilon)\,\|\mathbf{x}\|^2_R+(1+{1}/{\epsilon})\,\|\mathbf{y}\|^2_R$, for arbitrary $\mathbf{z}, \mathbf{w}\in \mathbb{R}^m$ and $\epsilon>0$, along with Lemma~\ref{R_C_norms_plus_realAnalyticity}, yields
{
\begin{equation}
\begin{aligned}
& \norm{\mathbf{x}^{\nu+1}-\mathbf{1}\bar{x}^{\nu+1}}^2_R
= \norm{\left(\mathbf{R}-\mathbf{1}\mathbf{r}^\top\right)\left(\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu\right)-\alpha\left(\mathbf{I}-\mathbf{1}\mathbf{r}^\top\right)\left(\mathbf{y}^{\nu}-\mathbf{1}\bar{y}^\nu\right)}^2_R
\\
&\leq (1+\epsilon_x)\norm{\left(\mathbf{R}-\mathbf{1}\mathbf{r}^\top\right)\left(\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu\right)}^2_R+\alpha^2 \left(1+\frac{1}{\epsilon_x}\right)\norm{\left(\mathbf{I}-\mathbf{1}\mathbf{r}^\top\right)\left(\mathbf{y}^{\nu}-\mathbf{1}\bar{y}^\nu\right)}^2_R
\\
&\leq \rho_R^2 (1+\epsilon_x) \norm{\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu}^2_R+\alpha^2 \left(1+\frac{1}{\epsilon_x}\right)\|\mathbf{I}-\mathbf{1}\mathbf{r}^\top \|_R^2\norm{\mathbf{y}^{\nu}-\mathbf{1}\bar{y}^\nu}^2_R
\\
&\overset{\eqref{explicit_Rnorm_Eqs}}{\leq} \rho_R^2 (1+\epsilon_x) \norm{\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu}^2_R+2\alpha^2 \left(1+\frac{1}{\epsilon_x}\right)\,\norm{\mathbf{y}^{\nu}-\mathbf{c}\bar{y}^\nu}^2_R
\\
& \qquad +2\alpha^2 (1+\frac{1}{\epsilon_x})\norm{(\mathbf{1}-\mathbf{c})\bar{y}^\nu}^2_R
\\
& \leq \rho^2_R (1+\epsilon_x) \norm{\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu}^2_R+\alpha^2 K_2\norm{\mathbf{y}^{\nu}-\mathbf{c}\bar{y}^\nu}^2_C+\alpha^2 K_3|\bar{y}^\nu|^2_2,
\end{aligned}
\label{x_bound}
\end{equation}
where $\epsilon_x>0$ is arbitrary and we defined
\begin{equation}\label{K_consts_1}
K_2\triangleq 2K_{R,C}^2 \left(1+\frac{1}{\epsilon_x}\right),\quad K_3\triangleq
2n(1+\frac{1}{\epsilon_x}).
\end{equation}}
Similarly, the tracking error can be bounded as
{
\begin{equation}\hspace{-.2cm}
\label{y_bound}
\begin{aligned}
&\norm{\mathbf{y}^{\nu+1}-\mathbf{c}\bar{y}^{\nu+1}}^2_C
= \norm{\left(\mathbf{C}-\mathbf{c}\mathbf{1}^\top\right)\mathbf{y}^{\nu}+\left(\mathbf{I}-\mathbf{c}\mathbf{1}^\top\right)\left(\nabla F_c(\mathbf{x}^{\nu+1})-\nabla F_c(\mathbf{x}^{\nu})\right)}^2_C
\\
& \leq (1+\epsilon_y)\norm{\left(\mathbf{C}-\mathbf{c}\mathbf{1}^\top\right)\left(\mathbf{y}^{\nu}-\mathbf{c}\bar{y}^\nu\right)}^2_C
\\
&\quad +(1+\frac{1}{\epsilon_y})\norm{\left(\mathbf{I}-\mathbf{c}\mathbf{1}^\top\right)\left(\nabla F_c(\mathbf{x}^{\nu+1})-\nabla F_c(\mathbf{x}^{\nu})\right)}^2_C
\\
& \overset{\eqref{explicit_Cnorm_Eqs}}{\leq} \rho_C^2(1+\epsilon_y)\norm{\mathbf{y}^{\nu}-\mathbf{c}\bar{y}^\nu}^2_C+K_{C,2}^2L_{c}^2\left(1+\frac{1}{\epsilon_y}\right)\,\norm{\mathbf{x}^{\nu+1}-\mathbf{x}^{\nu}}^2
\\
& \overset{(a)}{=} \rho_C^2(1+\epsilon_y)\norm{\mathbf{y}^{\nu}-\mathbf{c}\bar{y}^\nu}^2_C
\\
&+3K_{C,2}^2L_{c}^2\left(1+\frac{1}{\epsilon_y}\right)\,\left[\norm{(\mathbf{R}-\mathbf{I})(\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu)}^2+\alpha^2\norm{\mathbf{y}^{\nu}-\mathbf{c}\bar{y}^\nu}^2 +\alpha^2|\bar{y}^\nu|^2\norm{\mathbf{c}}^2\right]
\\
& \overset{\eqref{explicit_Rnorm_Eqs}}{\leq} \rho_C^2(1+\epsilon_y)\norm{\mathbf{y}^{\nu}-\mathbf{c}\bar{y}^\nu}^2_C
\\
&+3K_{C,2}^2L_{c}^2\left(1+\frac{1}{\epsilon_y}\right)\,\left[K_{2,R}^2\norm{\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu}^2+K_{2,C}^2\alpha^2\norm{\mathbf{y}^{\nu}-\mathbf{c}\bar{y}^\nu}_C^2 +\alpha^2|\bar{y}^\nu|^2\right]
\\
& = \rho_C^2(1+\epsilon_y)\norm{\mathbf{y}^{\nu}-\mathbf{c}\bar{y}^\nu}^2_C
\\
&+3K_{C,2}^2L_{c}^2\left(1+\frac{1}{\epsilon_y}\right)\,\left[K_{2,R}^2\norm{\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu}^2+K_{2,C}^2\alpha^2\norm{\mathbf{y}^{\nu}-\mathbf{c}\bar{y}^\nu}_C^2 +\alpha^2|\bar{y}^\nu|^2\right]
\\
& \leq \left(\rho^2_C+\frac{\alpha^2 K_4}{\epsilon_y}\right)(1+\epsilon_y) \norm{\mathbf{y}^{\nu}-\mathbf{c}\bar{y}^\nu}^2_C+\alpha^2 K_5|\bar{y}^\nu|^2_2+K_6\left(1+\frac{1}{\epsilon_y}\right)\norm{\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu}^2_R,
\end{aligned}
\end{equation}
where in (a) we used $\bx^{\nu+1}-\bx^\nu=(\mathbf{R}-\mathbf{I})(\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu)-\alpha(\mathbf{y}^{\nu}-\mathbf{c}\bar{y}^\nu)-\alpha\mathbf{c}\bar{y}^\nu$ and the Jensen's inequality; and in the last inequality we defined
\begin{equation}\label{K_consts_2}
K_4=3K_{C,2}^2K_{2,C}^2L_c^2,\quad K_5=3K_{C,2}^2L_c^2,\quad K_6=3K_{C,2}^2K_{2,R}^2L_c^2.
\end{equation}
}
\subsubsection{Lyapunov function}
{
Let us introduce now the candidate Lyapunov function: denoting $\mathbf{J}_R\triangleq \mathbf{1}\mathbf{r}^\top$ and $\mathbf{J}_C\triangleq \mathbf{c}\mathbf{1}^\top$, define
\begin{equation}\label{eq:def_L}
L(\mathbf{x},\mathbf{y})\triangleq F_c(\mathbf{J}_R\mathbf{x})+\norm{(\mathbf{I}-\mathbf{J}_R)\mathbf{x}}^2_R+\varkappa\norm{(\mathbf{I}-\mathbf{J}_C)\mathbf{y}}^2_C,
\end{equation}
where $\varkappa>0$ is a positive constant (to be properly chosen). Combining \eqref{descent2}-\eqref{y_bound} and using $\bar{y}^\nu=\bar{g}^\nu=\sum_{i=1}^n\nabla f_i(x_i^\nu)$ [cf. \eqref{x_bar_eq}] leads to the following descent property for $L$:
\begin{equation}
L(\mathbf{x}^{\nu+1},\mathbf{y}^{\nu+1})\leq L(\mathbf{x}^{\nu},\mathbf{y}^{\nu})-d(\mathbf{x}^{\nu},\mathbf{y}^{\nu})^2,
\label{L_descent}
\end{equation}
where
\begin{equation} \label{d_def_func}
d(\bx,\by)\triangleq \sqrt{(1-\tilde{\rho}_R)\norm{(\mathbf{I}-\mathbf{J}_R)\mathbf{x}}^2_R+\varkappa(1-\tilde{\rho}_C)\norm{(\mathbf{I}-\mathbf{J}_C)\by}^2_C+\Gamma~\Big|\sum_{i=1}^n\nabla f_i(x_i)\Big|^2}
\end{equation}
and
\begin{equation}
\begin{aligned}
\tilde{\rho}_R\triangleq &\rho^2_R (1+\epsilon_x)+\frac{\alpha nL_c^2K_{2,R}^2}{2}\left(1+\frac{\zeta}{\epsilon_1}\right)+\varkappa K_6\left(1+\frac{1}{\epsilon_y}\right),
\\
\tilde{\rho}_C\triangleq &\rho^2_C(1+\epsilon_y)+ \frac{\alpha K_{2,C}^2}{2\varkappa}\left(1+\frac{1}{\epsilon_2}\right)+\alpha^2\left(\frac{LK_{2,C}^2+K_2}{\varkappa}+K_4\left(1+\frac{1}{\epsilon_y}\right)\right),
\\
\Gamma\triangleq &\left(\zeta-\frac{\epsilon_1\zeta}{2}-\frac{\epsilon_2 }{2}\right)\alpha-\left(L\zeta^2+K_3+K_5\varkappa\right)\alpha^2.
\end{aligned}
\end{equation}
Note that the function $d(\bullet,\bullet)$ is a valid measure of optimality/consensus for DOGT: i) it is continuous; and ii) $d(\bx,\by)=0$ implies $x_i=x_j=x^\ast$, for all $i,j\in[n]$ and some $x^\ast$ such that $\sum_{i=1}^n\nabla f_i(x^\ast)=0$, meaning that all $x_i$ are consensual and equal to a critical point of $F$. }
{
To ensure $\tilde{\rho}_R<1$, $\tilde{\rho}_C<1$, and $\Gamma>0$ in $d(\bx,\by)$, we choose the free parameters $\epsilon_x$, $\epsilon_y$, $\epsilon_1$, $\epsilon_2$, and $\varkappa$ as follows:
\begin{equation}
\begin{aligned}
& 0<\epsilon_x<\frac{1-\rho_R^2}{2\rho_R^2}, \quad & 0<\epsilon_y<\frac{1-\rho_C^2}{\rho_C^2},\quad &
\\
& \epsilon_1=\epsilon_2=\epsilon,\quad & 0<\epsilon<\frac{2\zeta}{1+\zeta},\qquad & 0<\varkappa\leq \frac{\rho^2_R\epsilon_x}{K_6(1+1/\epsilon_y)},
\end{aligned}
\label{parameters_assymptotic_conv}
\end{equation}
and finally, $\alpha>0$ must satisfy
\begin{equation}
\begin{aligned}
\alpha & <\frac{2}{nL_c^2K_{2,R}^2\left(1+\frac{\zeta}{\epsilon}\right)}\left(1-\rho^2_R (1+2\epsilon_x)\right),
\\
\alpha & <\frac{1-\rho^2_C (1+\epsilon_y)}{\frac{1}{2\varkappa}K_{2,C}^2\left(1+\frac{1}{\epsilon}+2L\right)+\frac{K_2}{\varkappa}+K_4\left(1+\frac{1}{\epsilon_y}\right)},
\\
\alpha & <\frac{\zeta-\frac{\epsilon}{2}\left(\zeta+1\right)}{L\zeta^2+K_3+K_5\varkappa}.
\end{aligned}
\label{alpha_conds_assymptotic_conv}
\end{equation}
Substituting \eqref{eq:def_K_1}, \eqref{K_consts_1}, and \eqref{K_consts_2} in (\ref{alpha_conds_assymptotic_conv}) and setting for simplicity
\begin{equation}\label{alpha_conds_assymptotic_conv_2}
\epsilon_x=\frac{1-\rho_R^2}{4\rho_R^2},\quad \epsilon_y=\frac{1-\rho_C^2}{2\rho_C^2},\quad\epsilon=\frac{\xi}{1+\xi},\quad \varkappa= \frac{c_{\min}r_{\min}}{24L_c^2}(1-\rho_R^2)(1-\rho_C^2),
\end{equation}
we obtain the following sufficient conditions for (\ref{alpha_conds_assymptotic_conv}):
\begin{equation}
\begin{aligned}
\alpha & \leq \tilde{\alpha}_1\triangleq \frac{r_{\min}(1-\rho_R^2)}{3nL_c^2},
\\
\alpha & \leq \tilde{\alpha}_2\triangleq\frac{(1-\rho_R^2)^2(1-\rho_C^2)^2r_{\min}^2c_{\min}^2}{1152L_c^2(2+L)},
\\
\alpha & \leq \tilde{\alpha}_3\triangleq \frac{r_{\min}c_{\min}(1-\rho_R^2)}{2(L+16n)}.
\end{aligned}
\label{alpha_conds_assymptotic_conv_3}
\end{equation}
A further simplification, leads to the following final more restrictive condition on $\alpha$:
\begin{equation}
\begin{aligned}
0<\alpha & \leq \frac{(1-\rho_R^2)^2(1-\rho_C^2)^2r_{\min}^2c_{\min}^2}{1152L_c^2(L+16n)}.
\end{aligned}
\label{alpha_conds_assymptotic_conv_final}
\end{equation}
}
The descent property (\ref{L_descent}) readily implies the following convergence result for $\{L(\mathbf{x}^\nu,\mathbf{y}^\nu)\}$ and $\{d(\bx^\nu,\by^\nu)\}$.
\begin{lemma}
\label{sequence_props_a}
Under Assumptions \ref{P_assumption}, \ref{Net-Assump}, and \ref{matrix_C_R}, and the above choice of parameter, there hold:
\begin{enumerate}[label=(\roman*)]
\item The sequence $\{L(\mathbf{x}^\nu,\mathbf{y}^\nu)\}$ converges;
\item $\sum_{\nu=0}^\infty d(\bx^\nu,\by^\nu)^2<\infty$, and thus $\lim_{\nu\to \infty}d(\bx^\nu,\by^\nu)= 0$.
\end{enumerate}
\end{lemma}
{ We conclude this subsection by lower bounding $d(\bx^\nu,\by^\nu)$ by the magnitude of the gradient of the Lyapunov function $L$. This will allow us to transfer the convergence properties of $\{d(\bx^\nu,\by^\nu)\}$ to $\{||\nabla L(\bx^\nu,\by^\nu)||\}$. The lemma below will also be useful to establish global convergence of DOGT under the {K\L} property (cf. Sec. \ref{sec:KL-main-results}).}
{
\begin{lemma}
\label{nablaL_bound_lemma} Let $\nabla L(\mathbf{x}^\nu,\mathbf{y}^\nu)\triangleq (\nabla_\mathbf{x} L(\mathbf{x}^\nu,\mathbf{y}^\nu),\nabla_\mathbf{y} L(\mathbf{x}^\nu,\mathbf{y}^\nu))$, where $\nabla_\mathbf{x} L$ (resp. $\nabla_\mathbf{y} L$) are the gradient of $L$ with respect to the first (resp. second) argument. In the setting above, there holds\vspace{-0.2cm}
\begin{equation}
\norm{\nabla L(\mathbf{x}^\nu,\mathbf{y}^\nu)}\leq Md(\bx^\nu,\by^\nu),\quad \nu\geq 0,
\label{nablaL_bound_original}
\end{equation}
with \vspace{-0.3cm}
\begin{equation}
M=\sqrt{2}\max\left(\frac{(2r_{\max}+L_c\sqrt{n})^2}{r_{\min}(1-\tilde{\rho}_R)},\frac{2\varkappa c_{\max}}{c_{\min}^{2}(1-\tilde{\rho}_C)}, \frac{1}{\Gamma}\right)^{\frac{1}{2}}.
\end{equation}
\end{lemma}
\begin{proof}
Recall that $\mathbf{J}_R=\mathbf{1r}^\top$ and $\mathbf{J}_C=\mathbf{c1}^\top$.
By definition \eqref{eq:def_L} and Lemma \ref{R_C_norms_plus_realAnalyticity}, we can write\vspace{-0.1cm
\begin{equation}
\begin{aligned}
\nabla_\mathbf{x} L(\mathbf{x}^\nu,\mathbf{y}^\nu) &=\mathbf{J}_R^\top\nabla F_c(\mathbf{J}_R\mathbf{x}^\nu)+2(\mathbf{I}-\mathbf{J}_R)^\top\diag({\mathbf{r}})(\mathbf{I}-\mathbf{J}_R)\mathbf{x}^\nu
\\
& \overset{(a)}{=} \mathbf{r}\,\bar{y}^\nu+\mathbf{J}_R^\top\left(\nabla F_c(\mathbf{J}_R\mathbf{x}^\nu)-\nabla F_c(\mathbf{x}^\nu)\right)
\\
&\quad +2(\mathbf{I}-\mathbf{J}_R)^\top\diag({\mathbf{r}})(\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu),\smallskip
\\
\nabla_\mathbf{y} L(\mathbf{x}^\nu,\mathbf{y}^\nu) &=2\varkappa(\mathbf{I}-\mathbf{J}_C)^\top\diag({\mathbf{c}})^{-1}(\mathbf{I}-\mathbf{J}_C)\mathbf{y}^\nu\\
& = 2\varkappa(\mathbf{I}-\mathbf{J}_C)^\top\diag({\mathbf{c}})^{-1}(\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu),
\end{aligned}\label{nablaL_expression}\vspace{-0.1cm}
\end{equation}
where (a) is due to $\bar{y}^\nu=\bar{g}^\nu$ (cf. \eqref{x_bar_eq}). Thus there holds
\begin{equation}
\begin{aligned}
||\nabla_\mathbf{x} L(\mathbf{x}^\nu,\mathbf{y}^\nu)|| &\leq ||\mathbf{r}||~ |\bar{y}^\nu|+||\mathbf{J}_R^\top\left(\nabla F_c(\mathbf{J}_R\mathbf{x}^\nu)-\nabla F_c(\mathbf{x}^\nu)\right)||
\\
&\quad +2||(\mathbf{I}-\mathbf{J}_R)^\top\diag({\mathbf{r}})(\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu)||
\\
& \overset{(b)}{\leq} |\bar{y}^\nu|+K_{2,R}\left(2r_{\max}+L_c\sqrt{n}\right)||\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu||_R,
\\
||\nabla_\mathbf{y} L(\mathbf{x}^\nu,\mathbf{y}^\nu)|| &\overset{(c)}{\leq} 2\varkappa K_{2,C}c_{\min}^{-1}||\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu||_C,
\end{aligned}
\label{nablaL_expression2}
\end{equation}
where (b) holds due to $||\diag(\mathbf{r)}||_R=||\diag(\mathbf{r})||_2=r_{\max}$, $||\mathbf{r}||\leq 1$, $||\mathbf{J}_R||_2\leq \sqrt{n}$ and \eqref{explicit_Rnorm_Eqs}; (c) is due to $||\diag(\mathbf{c})^{-1}||_C=||\diag(\mathbf{c})^{-1}||_2=c_{\min}^{-1}$ and \eqref{explicit_Cnorm_Eqs}.
Eq.~\eqref{nablaL_bound_original} follows readily from \eqref{nablaL_expression2}.
\end{proof}}
\subsubsection{Main result}
We can now state the main convergence result of DOGT to critical points of $F$.
{ \begin{theorem}\label{Th_asympt_convergence}
\label{sequence_props} Consider Problem \eqref{eq:P}, and suppose that Assumptions \ref{P_assumption} and \ref{Net-Assump} are satisfied.
Let $\{(\mathbf{x}^\nu,\mathbf{y}^\nu)\}$ be the sequence generated by the DOGT Algorithm \eqref{NEXT_generic}, with $\mathbf{R}$ and $\mathbf{C}$ satisfying Assumption \ref{matrix_C_R}, and $\alpha$ chosen according to \eqref{alpha_conds_assymptotic_conv_final} [or \eqref{alpha_conds_assymptotic_conv_2}]; let $\{\bar{x}^\nu\}$ and $\{\bar{y}^\nu\}$ be defined in \eqref{eq:def_w_vectors}; and let $\{d(\bx^\nu,\by^\nu)\}$ be defined in \eqref{d_def_func}. Given $\epsilon>0$, let $T_\epsilon=\min\{\nu\in \mathbb{N}_+\,:\, d(\bx^\nu,\by^\nu)\leq \epsilon\}$. Then, there hold
\begin{enumerate}[label=(\roman*)]
\item[(i)] \emph{[consensus]:} $\lim_{\nu\to \infty} \norm{\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu}=0$ and $\lim_{\nu\to \infty} \bar{y}^\nu =0$;
\item[(ii)] \emph{[stationarity]:} let $\mathbf{x}^\infty$ be a limit point of $\{\mathbf{x}^\nu\}$; then, $\mathbf{x}^\infty=\theta^\infty \,\mathbf{1}$, for some $\theta^\infty\in\mathrm{crit}~F$;
\item[(iii)] \emph{[sublinear rate]:} $T_\epsilon=o(1/\epsilon^2)$.
\end{enumerate}
\end{theorem}
\begin{proof}
(i) follows readily from Lemma~\ref{sequence_props_a}(ii).
We prove (ii). Let $(\mathbf{x}^\infty, \mathbf{y}^\infty)$ be a limit point of $\{(\mathbf{x}^\nu, \mathbf{y}^\nu)\}$. By (i), it must be $(\mathbf{I}-\mathbf{J}_R){\mathbf{x}}^\infty=\mathbf{0}$, implying $ {\mathbf{x}}^\infty=\mathbf{1}\theta^\infty$, for some $\theta^\infty\in\mathbb{R}$. Also, $\lim_{\nu\to \infty} \mathbf{1}^\top\nabla F_c(\mathbf{x}^\nu)= \lim_{\nu\to \infty} \bar{g}^\nu=\lim_{\nu\to \infty} \bar{y}^\nu=0$, which together with the continuity of $\nabla F_c$, yields $0=\mathbf{1}^\top\nabla F_c(\mathbf{1}\theta^\infty)=\nabla F(\theta^\infty)$. Therefore, $\theta^\infty\in\mathrm{crit}~F$.
We prove now (iii). Using \eqref{L_descent} and the definition of $T_\epsilon$, we can write \vspace{-0.3cm}
\begin{equation}\label{T_eps_bound}
\frac{T_\epsilon}{2}\epsilon^2 \leq \sum_{t=\floor{\frac{T_\epsilon}{2}}+1}^{T_\epsilon} d(\bx^t,\by^t)^2\leq l^{\floor{\frac{T_\epsilon}{2}}+1}-l^{T_\epsilon+1},\vspace{-0.2cm}
\end{equation}
where we used the shorthand $l^\nu\triangleq L(\mathbf{x}^\nu,\mathbf{y}^\nu)$. Consider the following two cases: (1) $T_\epsilon\rightarrow \infty$ as $\epsilon\rightarrow 0$, then $l^{\floor{\frac{T_\epsilon}{2}}+1}-l^{T_\epsilon+1}\rightarrow 0$ (recall that $\{l^\nu\}$ converges, cf. Lemma \ref{sequence_props_a}(i)); and (2) $T_\epsilon<\infty$ as $\epsilon\rightarrow 0$, then $\{l^\nu\}$ converges in a finite number of iterations. Therefore, by \eqref{T_eps_bound}, we have $T_\epsilon=o(1/\epsilon^2)$. \end{proof}}
Note that, as a direct consequence of Lemma~\ref{nablaL_bound_lemma}, one can infer the following further property of the limit points $(\mathbf{x}^\infty, \mathbf{y}^\infty)$ of the sequence $\{(\mathbf{x}^\nu, \mathbf{y}^\nu)\}$: any such a $(\mathbf{x}^\infty, \mathbf{y}^\infty)$ is a critical point of $L$ [defined in (\ref{eq:def_L})].
\subsection{Convergence under the {K\L} property} \label{Global_Conv_Subsec}
{
We now strengthen the subsequence convergence result in Theorem~\ref{Th_asympt_convergence}, under the additional assumption that $F$ is a {K\L} function \cite{MR0160856,Kurdyka1998}: We prove that the entire sequence $\{\mathbf{x}^\nu\}$ converges to a critical point of $F$ (cf. Theorem \ref{main_global_conv_thm}), and establish asymptotic convergence rates (cf. Theorem \ref{Rate_theorem}). We extend the analysis developed in \cite{Attouch2009, Attouch2013} for centralized first-order methods to our distributed setting and complement it with a rate analysis. The major difference with \cite{Attouch2013} is that the \emph{sufficient decent} condition postulated in \cite{Attouch2013} is neither satisfied by the objective value sequence $\{F(\mathbf{x}^\nu)\}$ (as requested in \cite{Attouch2013}), due to consensus and gradient tracking errors, nor by the Lyapunov function sequence $\{L(\mathbf{x}^\nu,\mathbf{y}^\nu)\}$, which instead satisfies \eqref{L_descent}.
A key step to cope with this issue is to establish necessary connections between $\nabla L(\bx,\by)$ and $d(\bx,\by)$ (defined in \eqref{eq:def_L} and \eqref{d_def_func}, respectively)--see Lemma \ref{nablaL_bound_original} and Proposition \ref{pre_conv_result}.
\subsubsection{Convergence analysis} \label{sec:KL-main-results}
We begin proving the following abstract intermediate results similar to \cite{Attouch2013} but extended to our distributed setting, which is at the core of the subsequent analysis; we still assume $m=1$ without loss of generality.
\begin{proposition}
\label{pre_conv_result}
In the setting of Theorem~ \ref{Th_asympt_convergence}, let $L$ defined in (\ref{eq:def_L}) is {K\L} at some $\acute{\mathbf{z}}\triangleq (\acute{\mathbf{x}},\acute{\mathbf{y}})$. Denote by $\mathcal{V}_{\acute{\mathbf{z}}}$, $\eta$, and $\phi:[0,\eta)\rightarrow \mathbb{R}_+$ the objects appearing in Definition \ref{KL_def}. Let $\rho>0$ be such that $\mathcal{B}(\acute{\mathbf{z}},\rho)^{2mn}\subseteq \mathcal{V}_{\acute{\mathbf{z}}}$.
Consider the sequence $\{\mathbf{z}^\nu\triangleq (\mathbf{x}^\nu,\mathbf{y}^\nu)\}$ generated by the DOGT Algorithm \eqref{NEXT_generic}, with initialization $\mathbf{z}^0\triangleq(\mathbf{x}^0,\mathbf{y}^0)$; and define $\acute{l}\triangleq L(\acute{\mathbf{z}})$ and $l^\nu\triangleq L(\mathbf{z}^\nu)$. Suppose that
\begin{equation}
\acute{l}<l^\nu<\acute{l}+\eta,\quad \forall \nu\geq 0,
\label{nonImprovingIterates_cond}\vspace{-0.2cm}
\end{equation}
and\vspace{-0.2cm}
\begin{equation}
K \, M\, \phi(l^0-\acute{l})+\norm{\mathbf{z}^{0}-\acute{\mathbf{z}}}<\rho,
\label{proximity_cond}
\end{equation}
where \vspace{-0.2cm}\begin{equation} K= \sqrt{3}(1+L_c)\max\left(\frac{4nK_{||}^2}{1-\tilde{\rho}_R},\frac{K_{||}^2}{\varkappa(1-\tilde{\rho}_C)}\left(\alpha+\frac{2\sqrt{n}}{1+L_c}\right)^2, \alpha^2/\Gamma\right)^{1/2},
\label{eq:def_K}
\end{equation}
and $M>0$ is defined in \eqref{nablaL_bound_original} (cf.~Lemma \ref{nablaL_bound_lemma}).
Then, $\{ \mathbf{z}^\nu\}$ satisfies:
\begin{enumerate}[label=(\roman*)]
\item $\mathbf{z}^\nu\in\mathcal{B}(\acute{\mathbf{z}},\rho)^{2mn}$, for all $\nu\geq 0$;
\item $\sum_{t=k}^{\nu}\norm{\mathbf{z}^{t+1}-\mathbf{z}^{t}} \leq KM\left(\phi(l^k-\acute{l})-\phi(l^{\nu+1}-\acute{l})\right)$ for all $\nu,k\geq 0$ and $\nu\geq k$;
\item $l^\nu\rightarrow\acute{l},\quad\mathrm{as}~\nu\rightarrow \infty$.
\end{enumerate}
\end{proposition}
\begin{proof}
{ Throughout the proof, we will use the following shorthand $d^\nu\triangleq d(\bx^\nu,\by^\nu)$. Let $d^\nu>0$, for all integer $\nu\geq 0$; otherwise, $\{\mathbf{x}^{\nu}\}$ converges in a finite number of steps, and its limit point is $\mathbf{x}^\infty=\mathbf{1}\theta^\infty$, for some $\theta^\infty\in\mathrm{crit}~F$.
}
We first bound the ``length" $\sum_{t=k}^{\nu}\norm{\mathbf{z}^{t+1}-\mathbf{z}^{t}}$. By \eqref{NEXT_generic}, there holds
\begin{equation*}
\begin{aligned}
\mathbf{x}^{\nu+1}-\mathbf{x}^{\nu}=&\left(\mathbf{R}-\mathbf{I}\right)\left(\mathbf{x}^\nu-\mathbf{1}\bar{x}^\nu\right)-\alpha\left(\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu\right)-\alpha\mathbf{c}\bar{y}^\nu,
\\
\mathbf{y}^{\nu+1}-\mathbf{y}^{\nu}=&\left(\mathbf{C}-\mathbf{I}\right)\left(\mathbf{y}^\nu-\mathbf{c}\bar{y}^\nu\right) +\nabla F_c(\mathbf{x}^{\nu+1})-\nabla F_c(\mathbf{x}^{\nu}).
\end{aligned}
\end{equation*}
Using $
||\mathbf{A}||_2\leq \sqrt{n}||\mathbf{A}||_\infty$ and $
||\mathbf{A}||_2\leq \sqrt{n}||\mathbf{A}||_1$, with $\mathbf{A}\in\mathcal{M}_n(\mathbb{R})$; and $||\mathbf{R}-\mathbf{I}||_\infty\leq 2 $ and $||\mathbf{C}-\mathbf{I}||_1\leq 2 $,
we get\vspace{-0.1cm}
\begin{equation*}
\begin{aligned}
\sum_{t=k}^{\nu}\norm{\mathbf{x}^{t+1}-\mathbf{x}^{t}}
\leq & \sum_{t=k}^{\nu}2\sqrt{n}\norm{\mathbf{x}^t-\mathbf{1}\bar{x}^t}+\alpha\norm{\mathbf{y}^t-\mathbf{c}\bar{y}^t}+\alpha |\bar{y}^t|,\\
\sum_{t=k}^{\nu}\norm{\mathbf{y}^{t+1}-\mathbf{y}^{t}}
\leq & \sum_{t=k}^{\nu}2\sqrt{n}\norm{\mathbf{y}^t-\mathbf{c}\bar{y}^t}+L_c \sum_{t=k}^{\nu}\norm{\mathbf{x}^{t+1}-\mathbf{x}^{t}},
\end{aligned}
\end{equation*}
where $L_c$ is the Lipschitz constant of $\nabla F_c$. The above inequalities imply\vspace{-0.2cm}
\begin{equation}
\begin{aligned}
&\sum_{t=k}^{\nu}\norm{\mathbf{z}^{t+1}-\mathbf{z}^{t}}
\\
\leq & \sum_{t=k}^{\nu}2(1+L_c)\sqrt{n}K_{||}\norm{\mathbf{x}^t-\mathbf{1}\bar{x}^t}_R+K_{||}\left(\alpha(1+L_c)+2\sqrt{n}\right)\norm{\mathbf{y}^t-\mathbf{c}\bar{y}^\nu}_C
\\
&+\alpha (1+L_c)|\bar{y}^t|
\leq K\sum_{t=k}^{\nu}d^t,
\end{aligned}
\label{length_bound}
\end{equation}
where
$K$ is defined in \eqref{eq:def_K}.
We prove now the proposition, starting from statement (ii). Multiplying both sides of \eqref{L_descent} by $\phi'(l^\nu-\acute{l})$ and using $\phi'(l^\nu-\acute{l})>0$ [due to property (iii) in Definition~\ref{KL_def} and \eqref{nonImprovingIterates_cond}] and the concavity of $\phi$, yield
\begin{equation}
(d^\nu)^2~ \phi'(l^\nu-\acute{l})\leq \phi'(l^\nu-\acute{l})~ \left(l^{\nu}-l^{\nu+1}\right)\leq \phi(l^\nu-\acute{l})-\phi(l^{\nu+1}-\acute{l}).
\label{deltaPHI_bound1}
\end{equation}
For all
$\mathbf{z}\in\mathcal{V}_{\acute{\mathbf{z}}}\cap[ \acute{l} <L<\acute{l}+\eta]$, the {K\L} inequality \eqref{KL_ineq} holds; hence, assuming
$\mathbf{z}^{t}\in\mathcal{B}(\acute{\mathbf{z}},\rho)^{2mn}$ for all $t=0,\ldots,\nu$, yields
\begin{equation}
\phi'(l^t-\acute{l})||\nabla L(\mathbf{z}^t)||\geq 1,\quad t=0,\ldots,\nu,
\label{KL_ineq_restated}
\end{equation}
which together with \eqref{deltaPHI_bound1} and \eqref{nablaL_bound_original} (cf. Lemma \ref{nablaL_bound_lemma}), gives
\begin{equation*}
M\left(\phi(l^t-\acute{l})-\phi(l^{t+1}-\acute{l})\right)\geq d^t,\quad t=0,\ldots,\nu,
\end{equation*}
and thus \vspace{-0.3cm}
\begin{equation}
M\left(\phi(l^k-\acute{l})-\phi(l^{\nu+1}-\acute{l})\right)\geq \sum_{t=k}^\nu d^t.
\label{sum_d_nu_ineq}
\end{equation}
Combining \eqref{sum_d_nu_ineq} with \eqref{length_bound}, we obtain \begin{equation}
\sum_{t=k}^{\nu}\norm{\mathbf{z}^{t+1}-\mathbf{z}^{t}} \leq KM\left(\phi(l^k-\acute{l})-\phi(l^{\nu+1}-\acute{l})\right).
\label{length_bound_2}
\end{equation}
{Ineq. \eqref{length_bound_2} proves (ii) if $\mathbf{z}^{\nu}\in\mathcal{B}(\acute{\mathbf{z}},\rho)^{2mn}$ for all $\nu\geq 0$, which is shown next.}
Now let us prove statement (i). Letting $k=0$ in \eqref{length_bound_2}, by \eqref{proximity_cond}, we obtain\vspace{-0.2cm}
\begin{equation*}
\norm{\mathbf{z}^{\nu+1}-\acute{\mathbf{z}}}\leq KM\left(\phi(l^0-\acute{l})-\phi(l^{\nu+1}-\acute{l})\right)+\norm{\mathbf{z}^{0}-\acute{\mathbf{z}}}<\rho.
\end{equation*}
Therefore, $\mathbf{z}^\nu\in\mathcal{B}(\acute{\mathbf{z}},\rho)^{2mn}$, for all $\nu\geq 0$.
We finally prove statement (iii). Inequalities \eqref{nablaL_bound_original} (cf. Lemma \ref{nablaL_bound_lemma}) and \eqref{KL_ineq_restated} imply \vspace{-0.2cm}
\begin{equation}
\phi'(l^\nu-\acute{l})~ d^\nu\geq 1/M,\quad \nu\geq 0.
\label{KL_plus_nablaLineq}
\end{equation}
On the other hand, by Lemma \ref{sequence_props_a}-(i), as $\nu\rightarrow\infty$, we have $l^\nu\rightarrow p$, for some $p\geq \acute{l}$. In fact, $p =\acute{l}$, otherwise $p-\acute{l}>0$, which would contradict \eqref{KL_plus_nablaLineq} (because $d^\nu\rightarrow 0$ as $\nu\rightarrow\infty$ and $\phi'(p-\acute{l})<\infty$).
\end{proof}
Roughly speaking, Proposition~\ref{pre_conv_result} states that, if the algorithm is initialized in a suitably chosen neighborhood of a point at which $L$ satisfies the {K\L} property, then it will converge to that point.
Combining this property with the subsequence convergence proved in Theorem \ref{main_global_conv_thm} we can obtain global convergence of the sequence to critical points of $F$, as stated next.
{
\begin{theorem
\label{main_global_conv_thm}
Consider the setting of Theorem~\ref{sequence_props}, and furthermore assume that $F$ is real-analytic. Any sequence $\{(\mathbf{x}^\nu,\mathbf{y}^\nu)\}$ generated by the DOGT Algorithm \eqref{NEXT_generic} converges to some $(\mathbf{x}^\infty,\mathbf{y}^\infty)\in\mathrm{crit}~L$. Furthermore, $\mathbf{x}^\infty=\mathbf{1}\otimes\mathbf{\theta}^\infty$, for some $\mathbf{\theta}^\infty\in\mathrm{crit}~F$.\vspace{-0.1cm}
\end{theorem}
\begin{proof}
Let $\mathbf{z}^\infty\triangleq (\mathbf{x}^\infty,\mathbf{y}^\infty)$ be a limit point of $\{\mathbf{z}^\nu\triangleq (\mathbf{x}^{\nu},\mathbf{y}^{\nu})\}$. Since $\{l^\nu\triangleq L(\mathbf{z}^\nu)\}$ is convergent (cf. Lemma \ref{sequence_props_a}) and $L$ is continuous, we deduce $l^\nu \to l^\infty\triangleq L(\mathbf{z}^\infty)$. Since $F$ is real-analytic, $L$ is real analytic (due to Lemma \ref{R_C_norms_plus_realAnalyticity} and the fact that summation/composition of functions preserve real-analytic property \cite[Prop. 2.2.8]{krantz2002primer}) and thus {K\L} at at $\mathbf{z}^\infty$ \cite{MR0160856}. Set $ \acute{\mathbf{z}}=\mathbf{z}^\infty$ and $\acute{l}=l^\infty$; denote by $\mathcal{V}_{\acute{\mathbf{z}}}$, $\eta$, and $\phi:[0,\eta)\rightarrow \mathbb{R}_+$ the objects appearing in Definition \ref{KL_def}; and let $\rho>0$ be such that $\mathcal{B}(\acute{\mathbf{z}},\rho)^{2mn}\subseteq \mathcal{V}_{\acute{\mathbf{z}}}$. By the continuity of $\phi$ and the properties above, we deduce that there exists an integer $\nu_0$ such that i) $l^\nu\in (\acute{l}, \acute{l}+\eta)$, for all $\nu \geq \nu_0$; and ii)
$
K ~ M~ \phi(l^{\nu_0}-\acute{l})+\norm{\mathbf{z}^{\nu_0}-\acute{\mathbf{z}}}<\rho,
$ with $K$ and $M$ defined in (\ref{eq:def_K}) and \eqref{nablaL_bound_original}, respectively. Global convergence of the sequence $\{\mathbf{z}^\nu\}$ follows by applying Proposition~\ref{pre_conv_result} to the sequence $\{\mathbf{z}^{\nu+\nu_0}\}$.
Finally, by Lemma \ref{sequence_props_a}(ii), $d(\bx^\nu,\by^\nu)\rightarrow 0$ as $\nu\rightarrow\infty$. Invoking the continuity of $\nabla L$ and Lemma \ref{nablaL_bound_lemma}, we have $\nabla L(\mathbf{x}^\infty,\mathbf{y}^\infty)=\mathbf{0}$, thus $(\mathbf{x}^\infty,\mathbf{y}^\infty)\in\mathrm{crit}~L$. By Theorem \ref{Th_asympt_convergence}(ii), $\mathbf{x}^\infty=\mathbf{1}\otimes \mathbf{\theta}^\infty $, with $\mathbf{\theta}^\infty\in\mathrm{crit}~F$.
\end{proof}
}
In the following theorem, we provide some convergence rate estimates.
{\begin{theorem}
\label{Rate_theorem}
In the setting of Theorem \ref{main_global_conv_thm}, let $L$ be a {K\L} function with $\phi(s)=cs^{1-\theta}$, for some constant $c>0$ and $\theta\in[0,1)$. Let $\{\mathbf{z}^\nu\triangleq (\mathbf{x}^\nu,\mathbf{y}^\nu)\}$ be a sequence generated by DOGT Algorithm \eqref{NEXT_generic}. Then, there hold:
\begin{enumerate}[label=(\roman*)]
\item If $\theta=0$, $\{\mathbf{z}^\nu\}$ converges to $ {\mathbf{z}}^\infty$ in a finite number of iterations;
\item If $\theta\in(0,1/2]$, then $||\mathbf{z}^\nu-{\mathbf{z}}^\infty||\leq C\tau^\nu$, for all $\nu\geq \bar{\nu}$ for some $\tau\in[0,1)$, $\bar{\nu}\in\mathbb{N}_+$, $C>0$;
\item If $\theta\in(1/2,1)$, then $||\mathbf{z}^\nu- {\mathbf{z}}^\infty||\leq C\nu^{-\frac{1-\theta}{2\theta-1}}$, for all $\nu\geq \bar{\nu}$ for some $\bar{\nu}\in\mathbb{N}_+$, $C>0$.
\end{enumerate}
\end{theorem}}
\begin{proof}
{ For sake of simplicity of notation, denote $d^\nu\triangleq d(\bx^\nu,\by^\nu)$ and define $D^\nu\triangleq \sum_{t=\nu}^\infty d^t$.} By \eqref{length_bound},
we have\vspace{-0.2cm}
\begin{equation}
\norm{\mathbf{z}^{\nu+1}-\mathbf{z}^\infty}\leq \sum_{t=\nu}^\infty\norm{\mathbf{z}^{t+1}-\mathbf{z}^{t}} {\leq}K D^\nu.\vspace{-0.2cm}
\end{equation}
It is then sufficient to establish the convergence rates for the sequence $\{D^\nu\}$.
By {K\L} inequality \eqref{KL_ineq} and \eqref{nablaL_bound_original}, we have\vspace{-0.1cm}
\begin{equation}
Md^\nu\phi'(l^\nu-l^\infty)\geq 1
\implies \tilde{M} (d^\nu)^{(1-\theta)/\theta}\geq (l^\nu-l^\infty)^{1-\theta},\quad \forall\nu\geq \bar{\nu}
\label{rate_bound1}
\end{equation}
for sufficiently large $\bar{\nu}$, where $\tilde{M}= \left(Mc(1-\theta)\right)^{(1-\theta)/\theta}$, $l^\nu\triangleq L(\mathbf{z}^\nu)$, and $l^\infty\triangleq L({\mathbf{z}}^\infty)$. In addition, by \eqref{sum_d_nu_ineq} (setting $\acute{l}=l^\infty$), we have
$D^\nu\leq M\phi(l^\nu-l^\infty)=Mc(l^\nu-l^\infty)^{1-\theta}$,
which together with \eqref{rate_bound1}, yields\vspace{-0.2cm}
\begin{equation}
D^\nu\leq \tilde{M}Mc(d^\nu)^{(1-\theta)/\theta}=\tilde{M}Mc(D^{\nu}-D^{\nu+1})^{(1-\theta)/\theta},\quad \forall\nu\geq \bar{\nu}.
\label{rate_bound3}
\end{equation}
The convergence rate estimates as stated in the theorem can be derived from \eqref{rate_bound3}, using the same line of analysis introduced in \cite{Attouch2009}. The remaining part of the proof is provided in Appendix \ref{Rate_theorem_proof_appendix} for completeness.
\end{proof}
\subsection{Second-order guarantees}
\label{Guarantees_Sec}
We prove that the DOGT algorithm almost surely converges to SoS solutions of \eqref{eq:P}, under a suitably chosen initialization and some additional conditions on the weight matrices $\mathbf{R}$ and $\mathbf{C}$. Following a path first established in \cite{pmlr-v49-lee16} and further developed in \cite{Jordan_FOM_AvoidsSaddles}, the key to our argument for the non-convergence to strict saddle points of $F$ lies in formulating the DOGT algorithm as a dynamical system while leveraging an instantiation of the stable manifold theorem, as given in \cite[Theorem 2]{Jordan_FOM_AvoidsSaddles}.
The nontrivial task is finding a self-map representing DOGT so that the stable set of the strict saddles of $F$ is zero measure with respect to the domain of the mapping; note that the domain of the map--which is the set of initialization points--is not full dimensional and is the same as the support of the probability measure.
Our analysis is organized in the following three steps: 1) Sec.~\ref{sec:prelim_manifold} introduces the preparatory background;
2) Sec.~\ref{DOGT_dynamicalSys} tailors the results of Step 1 to the DOGT algorithm; and 3) finally, Sec. \ref{sec:SOS_main} states our main results about convergence of the DOGT algorithm to SoS solutions of \eqref{eq:P}.
\subsubsection{The stable manifold theorem and unstable fixed-points}\label{sec:prelim_manifold}
{Let $g:\mathcal{S}\to \mathcal{S}$ be a mapping from $\mathcal{S}$ to itself,
where $\mathcal{S}$ is a manifold without boundary.}
Consider the dynamical system $\mathbf{u}^{\nu+1}=g(\mathbf{u}^\nu)$, with $\mathbf{u}^0\in \mathcal{S}$; we denote by $g^\nu$ the {$\nu$-fold} composition of $g$. Our focus is on the analysis of the trajectories of the dynamical system
around the fixed points of $g$; in particular we are interested in the set of unstable fixed points of $g$. We begin introducing the following definition.
\begin{definition}[Chapter 3 of \cite{Absil:2007:OAM:1557548}]
\label{Differential_def}
The differential of the mapping $g:\mathcal{S}\to \mathcal{S}$, denoted as $\mathrm{D}g(\mathbf{u})$, is a linear operator from $\mathcal{T}(\mathbf{u})\rightarrow\mathcal{T}(g(\mathbf{u}))$, where $\mathcal{T}(\mathbf{u})$ is the tangent space of $\mathcal{S}$ at $\mathbf{u}\in \mathcal{S}$. Given a curve $\gamma$ in $\mathcal{S}$ with $\gamma(0)=\mathbf{u}$ and $\frac{d\gamma}{d t}(0)=\mathbf{v}\in\mathcal{T}(\mathbf{u})$, the linear operator is defined as $\mathrm{D}g(\mathbf{u})\mathbf{v}=\frac{d(g\circ\gamma)}{d t}(0)\in \mathcal{T}(g(\mathbf{u}))$. The determinant of the linear operator $\det(\mathrm{D}g(\mathbf{u}))$ is the determinant of the matrix representing $\mathrm{D}g(\mathbf{u})$ with respect to a standard basis.\footnote{This determinant may not be uniquely defined, in the sense of being completely invariant to the basis used for the geometry. In this work,
we are interested in properties of the determinant that are independent of scaling, and thus the potentially arbitrary choice of a standard basis does not affect our conclusions.}
\end{definition}
We can now introduce the definition of the set of unstable fixed points of $g$.
\begin{definition}[Unstable fixed points]
\label{UFP_def}
The set of unstable fixed points of $g$ is defined as\vspace{-0.3cm}
\begin{equation}\label{def_unstable_fixpoint}
\mathcal{A}_g=\Big\{\mathbf{u}: g(\mathbf{u})=\mathbf{u},~\mathrm{spradii}\big(\mathrm{D} g(\mathbf{u})\big)>1\Big\}.
\end{equation}
\end{definition}
{The theorem below, which is based on the stable manifold theorem \cite[Theorem III.7]{shub1987global}, provides tools to let us connect $\mathcal{A}_g$ with the set of limit points which $\{\mathbf{u}^\nu\}$ can escape from.}\vspace{-0.1cm}
\begin{theorem}[{\cite[Theorem 2]{Jordan_FOM_AvoidsSaddles}}]
\label{StableSetofUnstable_fixed_point_is_zeroM}
Let $g:\mathcal{S}\rightarrow\mathcal{S}$ be a $\mathcal{C}^1$ mapping and\vspace{-0.2cm}
\begin{equation*}
\det\left(\mathrm{D}g(\mathbf{u})\right)\neq 0,\quad \forall\mathbf{u}\in\mathcal{S}.\vspace{-0.2cm}
\end{equation*}
Then, the set of initial points that converge to an unstable fixed point (termed stable set of $\mathcal{A}_g$) is zero measure in $\mathcal{S}$.
Therefore, \vspace{-0.2cm}
\begin{equation*}
\mathbb{P}_{\mathbf{u}^0} \left( \lim_{\nu\to \infty}g^\nu(\mathbf{u}^0) \in \mathcal{A}_g \right)=0,\vspace{-0.2cm}
\end{equation*}
where the probability is taken over the starting point $\mathbf{u}^0\in \mathcal{S}$.
\end{theorem}
\subsubsection{DOGT as a dynamical system}
\label{DOGT_dynamicalSys}
Theorem \ref{StableSetofUnstable_fixed_point_is_zeroM} sets the path to the analysis of the convergence of the DOGT algorithm to SoS solutions of $F$: it is sufficient to describe the DOGT algorithm by a proper mapping $g:\mathcal{S}\to \mathcal{S}$ satisfying the assumptions in the theorem and such that the non-convergence of $g^\nu(\mathbf{u}^0)$, $\mathbf{u}^0\in \mathcal{S}$, to $\mathcal{A}_g$ implies the non-convergence of the DOGT algorithm to strict saddles of $F$.
We begin rewriting the DOGT in an equivalent and more convenient form. Define $\mathbf{h}^\nu\triangleq \mathbf{y}^{\nu}-\nabla F_c(\mathbf{x}^{\nu})$; \eqref{NEXT_generic} can be rewritten as
\begin{equation}
\left\lbrace
\begin{aligned}
& \mathbf{x}^{\nu+1}=\mathbf{W}_R\mathbf{x}^{\nu}-\alpha\left(\mathbf{h}^\nu+\nabla F_c(\mathbf{x}^{\nu})\right);
\\
& \mathbf{h}^{\nu+1} =\mathbf{W}_C\mathbf{h}^{\nu}+\left(\mathbf{W}_C-\mathbf{I}\right)\nabla F_c(\mathbf{x}^{\nu}),
\end{aligned}
\right.
\label{NEXT_generic4}
\end{equation}
with arbitrary $\mathbf{x}^0\in\mathbb{R}^{nm}$ and $\mathbf{h}^0\in\text{span}(\mathbf{W}_C-\mathbf{I})$. By Theorem \ref{Th_asympt_convergence}, every limit point $(\mathbf{x}^\infty,\mathbf{h}^\infty)$ of $\{(\mathbf{x}^\nu,\mathbf{h}^\nu)\}$ has the form $\mathbf{x}^\infty=\mathbf{1}_n\otimes\boldsymbol{\theta}^\infty$ and $\mathbf{h}^\infty=-\nabla F_c(\mathbf{1}_n\otimes \boldsymbol{\theta}^\infty)$, for some $\boldsymbol{\theta}^\infty\in \mathrm{crit}~F$. We are interested in the non-convergence of (\ref{NEXT_generic4}) to such points whenever $\boldsymbol{\theta}^\infty\in \mathrm{crit}~F$ is a strict saddle of $F$. This motivates the following definition.
\begin{definition}[Consensual strict saddle points]
\label{SSP_def}
Let $\Theta^\ast_{ss}=\{\boldsymbol{\theta}^\ast\in \mathrm{crit}~F\,:\,\lambda_{\text{min}}(\nabla^2 F(\boldsymbol{\theta}^\ast))<0\}$ denote the set of strict saddles of $F$. The set of \emph{consensual strict saddle points} is defined as
\begin{equation}\label{def_U_g}
\mathcal{U}^\ast\triangleq \left\{\left[\begin{array}{cc}
\mathbf{1}_n\otimes \boldsymbol{\theta}^\ast\\ -\nabla F_c(\mathbf{1}_n\otimes \boldsymbol{\theta}^\ast)
\end{array}
\right]\,:\, \boldsymbol{\theta}^\star\in \Theta^\ast_{ss}\right\}.
\end{equation}
\end{definition}
Roughly speaking, $\mathcal{U}^\ast$ represents the candidate set of ``adversarial'' limit points which any sequence generated by \eqref{NEXT_generic4} should escape from. The next step is then to write (\ref{NEXT_generic4}) as a proper dynamical system whose mapping satisfies conditions in Theorem \ref{StableSetofUnstable_fixed_point_is_zeroM} and its set of unstable fixed points $\mathcal{A}_g
$ is such that $\mathcal{U}^\ast \subseteq \mathcal{A}_g$.\smallskip
\noindent {\textbf{Identification of $g$ and $\mathcal{S}$.}}
Define $\mathbf{u}\triangleq (\mathbf{x},\mathbf{h})$, where
$\mathbf{x}\triangleq [\mathbf{x}_1^\top,\ldots, \mathbf{x}^\top_n]^\top$,
$\mathbf{h}\triangleq [\mathbf{h}_1^\top,\ldots, \mathbf{h}^\top_n]^\top$,
and each $\mathbf{x}_i,\mathbf{h}_i\in \mathbb{R}^{m}$; its value at iteration $\nu$ is denoted by $\mathbf{u}^\nu\triangleq (\mathbf{x}^\nu,\mathbf{h}^\nu)$. Consider the dynamical system \vspace{-0.2cm}
\begin{equation}
\mathbf{u}^{\nu+1}=g(\mathbf{u}^{\nu}),\quad \text{with}\quad g\left(\mathbf{u}\right)\triangleq
\begin{bmatrix}
\mathbf{W}_R\mathbf{x}-\alpha\nabla F_c\left(\mathbf{x}\right)-\alpha\mathbf{h}
\\
\mathbf{W}_C\mathbf{h}+\left(\mathbf{W}_C-\mathbf{I}\right)\nabla F_c\left(\mathbf{x}\right)
\end{bmatrix},
\label{NEXT_generic_dyn}
\end{equation}
and $\mathbf{u}^0\in\mathbb{R}^{nm}\times \text{span}(\mathbf{W}_C-\mathbf{I})$. The fixed-point iterate \eqref{NEXT_generic_dyn} describes the trajectory generated by the DOGT algorithm (\ref{NEXT_generic4}).
However, the initialization imposed by DOGT leads to a $g$ that maps $\mathbb{R}^{n m}\times \text{span}(\mathbf{W}_C-\mathbf{I})$ into $\mathbb{R}^{n m}\times \mathbb{R}^{n m}$. We show next how to unify the domain and codomain of $g$ to a subspace $\mathcal{S}\subseteq \mathbb{R}^{nm}\times \mathbb{R}^{nm}$ as in form of the mapping in Theorem \ref{StableSetofUnstable_fixed_point_is_zeroM}.
Applying \eqref{NEXT_generic4} telescopically to the update of the $h$-variables yields:
$\mathbf{h}^{\nu}=\mathbf{W}_C^{\nu}\mathbf{h}^0+\left(\mathbf{W}_C-\mathbf{I}\right)\mathbf{g}_{\text{acc}}^\nu,$ for all $\nu\geq 1$, where $\mathbf{g}_{\text{acc}}^\nu\triangleq \sum_{t=0}^{\nu-1}\mathbf{W}_C^t\nabla F_c\left(\mathbf{x}^{\nu-t-1}\right)$. Denoting $\bar{\mathbf{h}}^\nu\triangleq (\mathbf{1}_n^\top \otimes \mathbf{I}_m) \mathbf{h}^\nu$, we have
{\begin{equation}
\bar{\mathbf{h}}^\nu=\cdots =\bar{\mathbf{h}}^0,\quad\text{and}\quad \mathbf{h}^\nu\in\mathbf{W}_C^{\nu}\mathbf{h}^0+\text{span}\left(\mathbf{W}_C-\mathbf{I}\right)\quad \forall \nu\geq 1.
\label{h_embedded_eq}\vspace{-0.4cm}
\end{equation}}
{
The initialization $\mathbf{h}^0\in\text{span}\left(\mathbf{W}_C-\mathbf{I}\right)$ in \eqref{NEXT_generic4}
naturally suggests the following $(2n-1)m$-dimensional linear subspace as candidate set $\mathcal{S}$:
\begin{equation}\label{def_S}
\mathcal{S}\triangleq\mathbb{R}^{nm}\times \text{span}\left(\mathbf{W}_C-\mathbf{I}\right)
\end{equation}
Such an $\mathcal{S}$ also ensures that $g:\mathcal{S}\to \mathcal{S}$. In fact, by (\ref{h_embedded_eq}), $\mathbf{h}^\nu\in\text{span} (\mathbf{W}_C-\mathbf{I})$, for all $\nu\geq 1$, provided that $\mathbf{h}^0\in\text{span}\left(\mathbf{W}_C-\mathbf{I}\right)$. Therefore, $\{g^\nu(\mathbf{u}^0)\}\subseteq\mathcal{S}$, for all $\mathbf{u}^0\in\mathcal{S}$.
}
Equipped with the mapping $g$ in (\ref{NEXT_generic_dyn}) and $\mathcal{S}$ defined in (\ref{def_S}), we check next that the condition in Theorem \ref{StableSetofUnstable_fixed_point_is_zeroM} is satisfied; we then prove that $\mathcal{U}^\ast \subseteq \mathcal{A}_g$.
\textbf{1) ${g}$ is a diffeomorphism:} To establish this property, we add the following extra assumption on the weight matrices $\mathbf{R}$ and $\mathbf{C}$, which is similar to Assumption \ref{nonsingular_W} for the DGD scheme.
\begin{assumption}
\label{nonsingular_RC}
Matrices $\mathbf{R}\in\mathcal{M}_n(\mathbb{R})$ and $\mathbf{C}\in\mathcal{M}_n(\mathbb{R})$ are nonsingular.
\end{assumption}
The above condition is not particularly restrictive and it is compatible with Assumption \ref{matrix_C_R}. A rule of thumb is to choose $\mathbf{R}=(\tilde{\mathbf{R}}+\mathbf{I})/2$ and $\mathbf{C}=(\tilde{\mathbf{C}}+\mathbf{I})/2$, with $\tilde{\mathbf{R}}$ and $\tilde{\mathbf{C}}$ satisfying Assumption \ref{matrix_C_R}. The new matrices still satisfy Assumption \ref{matrix_C_R} due to the following fact:
given two nonnegative matrices $\mathbf{A},\mathbf{B}\in\mathcal{M}_n(\mathbb{R})$, if the directed graph associated with matrix $\mathbf{A}$ has a spanning tree and $\mathbf{B}\geq \rho\mathbf{A}$, for some $\rho>0$, then the directed graph associated with matrix $\mathbf{B}$ has a spanning tree as well.
We build now the differential of $g$. {Let $\tilde{g}$ be a smooth extension of \eqref{NEXT_generic_dyn} to $\mathbb{R}^{mn}\times \mathbb{R}^{mn}$, that is $g=\tilde{g}|_\mathcal{S}$. The differential $\mathrm{D}\tilde{g}(\mathbf{u})$ of $\tilde{g}$ at $\mathbf{u}\in \mathcal{S}$ reads
\begin{equation}
\mathrm{D}\tilde{g}(\mathbf{u})= \begin{bmatrix}
\mathbf{W}_R-\alpha\nabla^2 F_c(\mathbf{x}) &-\alpha\mathbf{I}
\\
\left(\mathbf{W}_C-\mathbf{I}\right)\nabla^2 F_c(\mathbf{x}) & \mathbf{W}_C
\end{bmatrix};
\label{Jacobian}
\end{equation}
$\mathrm{D}\tilde{g}(\mathbf{u})$ is related to the differential of $g$ by $\mathrm{D}{g}(\mathbf{u})=\mathrm{D}\tilde{g}(\mathbf{u})\mathbf{P}_{\mathcal{T}(\mathbf{u})}$ \cite{ReimannianHess_Absil}, where $\mathbf{P}_{\mathcal{T}(\mathbf{u})}$ is the orthogonal projector onto $\mathcal{T}(\mathbf{u})$. Using $\mathcal{T}(\mathbf{u})=\mathcal{S}$, for all $\mathbf{u}\in\mathcal{S}$ (recall that $\mathcal{S}$ is a linear subspace) and denoting by $\mathbf{U}_h\in{\mathbb{R}^{mn\times m(n-1)}}$ an orthonormal basis of $\mathrm{span}(\mathbf{W}_C-\mathbf{I})$, $\mathrm{D}{g}(\mathbf{u})$ reads\vspace{-0.1cm}
\begin{equation}
\label{Reimannian_Jacob}
\mathrm{D}g(\mathbf{u})= \begin{bmatrix}
\mathbf{W}_R-\alpha\nabla^2 F_c(\mathbf{x}) &-\alpha\mathbf{I}
\\
\left(\mathbf{W}_C-\mathbf{I}\right)\nabla^2 F_c(\mathbf{x}) & \mathbf{W}_C
\end{bmatrix} {\mathbf{U}\mathbf{U}^\top},\quad\text{with}\quad
\mathbf{U}\triangleq \begin{bmatrix}
\mathbf{I} &\mathbf{0}
\\
\mathbf{0}&\mathbf{U}_h
\end{bmatrix}.\vspace{-0.1cm}
\end{equation}}
Note that $\mathbf{P}_\mathcal{S}=\mathbf{U}\mathbf{U}^\top$.
We establish next the conditions for $g$ to be a $\mathcal{C}^1$ diffeomorphism, as stated in Theorem \ref{StableSetofUnstable_fixed_point_is_zeroM}.
\begin{proposition}
\label{g_is_diffeomorphism}
Consider the mapping $g:\mathcal{S}\to \mathcal{S}$ defined in (\ref{NEXT_generic_dyn}), under Assumptions \ref{P_assumption}-(i), \ref{matrix_C_R}, and \ref{nonsingular_RC}, with $\mathcal{S}$ defined in (\ref{def_S}). If the step-size is chosen according to \vspace{-0.2cm} \begin{equation}0<\alpha<\frac{\sigma_\mathrm{min}(\mathbf{C}\mathbf{R})}{L_c},\label{eq:step_diff}\vspace{-0.1cm} \end{equation} where $L_c= L_\mathrm{max}$, then $\det\left(\mathrm{D}g(\mathbf{u})\right)\neq 0$, for all $\mathbf{u}\in\mathcal{S}$.
\end{proposition}
\begin{proof}
Since $\mathrm{D}g(\mathbf{u}):\mathcal{S}\rightarrow\mathcal{S}$, it is sufficient to verify that $\mathrm{D}g(\mathbf{u})$ is an invertible linear transformation for every $\mathbf{u}\in \mathcal{S}$. Using the definition of $\mathbf{U}$, this is equivalent to show that $\mathbf{U}^T\mathrm{D}g(\mathbf{u})\mathbf{U}$ is invertible, for all $\mathbf{u}\in\mathcal{S}$.
Invoking \eqref{Reimannian_Jacob}, $\mathbf{U}^\top{\mathrm{D}}g(\mathbf{u})\mathbf{U}$ reads
\begin{equation}
\begin{aligned}
\mathbf{U}^\top{\mathrm{D}}g(\mathbf{u})\mathbf{U}= &\mathbf{U}^\top
\mathrm{D}\tilde{g}(\mathbf{u}) \mathbf{U}
= \begin{bmatrix}
\mathbf{W}_R-\alpha\nabla^2 F_c(\mathbf{x}) &-\alpha\mathbf{U}_h
\\
\mathbf{U}_h^\top\left(\mathbf{W}_C-\mathbf{I}\right)\nabla^2 F_c(\mathbf{x}) & \mathbf{U}_h^T\mathbf{W}_C\mathbf{U}_h
\end{bmatrix}.
\label{Projected_Jacobian0}
\end{aligned}
\end{equation}
Since $\mathbf{U}_h^\top\mathbf{W}_C\mathbf{U}_h$ is non-singular, we can use the Schur complement of $\mathbf{U}^\top{\mathrm{D}}g(\mathbf{u})\mathbf{U}$ with respect to $\mathbf{U}_h^\top\mathbf{W}_C\mathbf{U}_h$ and write
\begin{equation}
\begin{aligned}
\mathbf{U}^\top{\mathrm{D}}g(\mathbf{u})\mathbf{U}=\mathbf{S}_1\begin{bmatrix}
\mathbf{W}_R-\alpha\nabla^2 F_c(\mathbf{x})+\alpha \boldsymbol{\Phi}\left(\mathbf{W}_C-\mathbf{I}\right)\nabla^2 F_c(\mathbf{x}) &\mathbf{0}
\\
\mathbf{0}& \mathbf{U}_h^\top\mathbf{W}_C\mathbf{U}_h
\end{bmatrix}\mathbf{S}_2,
\end{aligned}
\label{Projected_Jacobian2}
\end{equation}
where $\boldsymbol{\Phi}\triangleq \mathbf{U}_h\left(\mathbf{U}_h^\top\mathbf{W}_C\mathbf{U}_h\right)^{-1}\mathbf{U}_h^\top$, and $\mathbf{S}_1$ and $\mathbf{S}_2$ are some nonsingular matrices.
By \eqref{Projected_Jacobian2}, it is sufficient to show that
\begin{equation}
\begin{aligned}
\mathbf{S} \triangleq \,&\mathbf{W}_R-\alpha\nabla^2 F_c(\mathbf{x})+\alpha \boldsymbol{\Phi}\left(\mathbf{W}_C-\mathbf{I}\right)\nabla^2 F_c(\mathbf{x})
\\
=&\mathbf{W}_R-\alpha\mathbf{W}_C^{-1}\nabla^2 F_c(\mathbf{x})+\alpha\left(\boldsymbol{\Phi}-\mathbf{W}_C^{-1}\right)\left(\mathbf{W}_C-\mathbf{I}\right)\nabla^2 F_c(\mathbf{x})
\end{aligned}
\label{Eqq}
\end{equation}
is non-singular.
Using
$\mathbf{W}_C-\mathbf{I}=\mathbf{U}_h\boldsymbol{\Delta}$, for some
$\boldsymbol{\Delta}\in\mathbb{R}^{m(n-1)\times mn}$ (recall that $\mathbf{U}_h$ is an orthonormal basis of $\mathrm{span}(\mathbf{W}_C-\mathbf{I})$), we can write
\begin{equation}
\label{Phi_expression}
\begin{aligned}
\boldsymbol{\Phi} =\,&\mathbf{U}_h\left(\mathbf{U}_h^\top\mathbf{W}_C\mathbf{U}_h\right)^{-1}\mathbf{U}_h^\top
= \mathbf{U}_h \left(\mathbf{I}+\boldsymbol{\Delta}\mathbf{U}_h\right)^{-1}\mathbf{U}_h^\top
\\
\overset{(a)}{=} \,& \mathbf{U}_h\mathbf{U}_h^\top-\mathbf{U}_h\boldsymbol{\Delta}\left(\mathbf{I}+\mathbf{U}_h\boldsymbol{\Delta}\right)^{-1}\mathbf{U}_h\mathbf{U}_h^\top
\\
= \,& \mathbf{U}_h\mathbf{U}_h^\top-\left(\mathbf{W}_C-\mathbf{I}\right)\mathbf{W}_C^{-1}\mathbf{U}_h\mathbf{U}_h^\top
\\
=\, &\mathbf{W}_C^{-1}\mathbf{U}_h\mathbf{U}_h^\top,
\end{aligned}
\end{equation}
where (a) we used the Woodbury identity of inverse matrices. Using \eqref{Phi_expression} in \eqref{Eqq}, we obtain\vspace{-0.2cm}
\begin{equation*}
\begin{aligned}
\mathbf{S}
=\,&\mathbf{W}_R-\alpha\mathbf{W}_C^{-1}\nabla^2 F_c(\mathbf{x})-\alpha\mathbf{W}_{C}^{-1}\underset{=\mathbf{{0}}}{\underbrace{\left(\mathbf{I}-\mathbf{U}_{h}\mathbf{U}_{h}^{\top}\right)\left(\mathbf{W}_{C}-\mathbf{I}\right)}}\nabla^2 F_c(\mathbf{x}).
\end{aligned}
\label{Eqq2}\vspace{-0.2cm}
\end{equation*}
Therefore, if $\alpha<\frac{\sigma_{\min}(\mathbf{CR})}{L_c}$, $\mathbf{S}$ is invertible, and consequently, so is $\mathbf{U}^\top{\mathrm{D}}g(\mathbf{u})\mathbf{U}$.
\end{proof}
\textbf{2) The consensual strict saddle points are unstable fixed points of $g$ ($\mathcal{U}^\ast \subseteq \mathcal{A}_g$): } First of all, note that every limit point of the sequence generated by \eqref{NEXT_generic4} is a fixed point of $g$ on $\mathcal{S}$; the converse might not be true. The next result establishes the desired connection between the set $\mathcal{A}_g$ of unstable fixed points of $g$ (cf.~Definition \ref{UFP_def}) and the set $\mathcal{U}^\ast$ of consensual strict saddle points (cf.~Definition \ref{SSP_def}). This will let us infer the instability of $\mathcal{U}^\ast$ from that of $\mathcal{A}_g$.
\begin{proposition}
\label{CSS_is_in_unstable_set}
{ Suppose that Assumptions \ref{P_assumption}-(i) and \ref{matrix_C_R} hold along with one of the following two conditions }
\begin{itemize}
\item[(i)] The weight matrices $\mathbf{R}$ and $\mathbf{C}$ are symmetric;
\item[(ii)] $m=1$.
\end{itemize}
Then, any consensual strict saddle point is an unstable fixed point of $g$, i.e.,
\begin{equation}
\mathcal{U}^\ast\subseteq\mathcal{A}_g,
\end{equation}
with $\mathcal{A}_g$ and $\mathcal{U}^\ast$ defined in \eqref{def_unstable_fixpoint} and \eqref{def_U_g}, respectively.
\end{proposition}
\begin{proof}
Let $\mathbf{u}^\ast\in\mathcal{U}^\ast$; $\mathbf{u}^\ast$ is a fixed point of $g$ defined in \eqref{NEXT_generic_dyn}. It is thus sufficient to show that $\mathrm{D}g(\mathbf{u}^\ast)$ has an eigenvalue with magnitude greater than one.
To do so, we begin showing that the differential $\mathrm{D}\tilde{g}(\mathbf{u}^\ast)$ of $\tilde{g}$ at $\mathbf{u}^\ast$ has an eigenvalue greater than one. Using \eqref{Jacobian},
$\mathrm{D}\tilde{g}(\mathbf{u}^\ast)$ reads
\begin{equation}
\mathrm{D}\tilde{g}(\mathbf{u}^\ast)= \begin{bmatrix}
\mathbf{W}_R-\alpha\nabla^2 F_c^\ast &-\alpha\mathbf{I}
\\
\left(\mathbf{W}_C-\mathbf{I}\right)\nabla^2 F_c^\star & \mathbf{W}_C
\end{bmatrix},
\label{Jacobian_star}
\end{equation}
where we defined the shorthand $\nabla^2 F_c^\ast\triangleq \nabla^2 F_c\left(\mathbf{1}\otimes \boldsymbol{\theta}^\ast\right)$, and $\boldsymbol{\theta}^\ast\in\Theta^\ast_{ss}$.
We need to prove \vspace{-0.2cm}
\begin{equation}
\det\left(\mathrm{D}\tilde{g}(\mathbf{u}^\ast)-\lambda_u\mathbf{I}\right)=0,\quad\text{ for some }\quad |\lambda_u|>1.
\label{det_eq1}
\end{equation}
If $|\lambda_u|>1$, $\mathbf{W}_C-\lambda_u\mathbf{I}$ is nonsingular (since $\mathrm{spradii}(\mathbf{C})=1$). Using the Schur complement of $\mathrm{D}\tilde{g}(\mathbf{u}^\ast)-\lambda_u\mathbf{I}$ with respect to $\mathbf{W}_C-\lambda_u\mathbf{I}$, we have
\begin{equation}
\begin{aligned}
\mathrm{D}\tilde{g}(\mathbf{u}^\ast)-\lambda_u\mathbf{I}= \tilde{\mathbf{S}}_1\begin{bmatrix}
\left(\mathrm{D}\tilde{g}(\mathbf{u}^\ast)-\lambda_u\mathbf{I}\right)/\left(\mathbf{W}_C-\lambda_u\mathbf{I}\right)&\mathbf{0}
\\
\mathbf{0} & \mathbf{W}_C-\lambda_u\mathbf{I}
\end{bmatrix}
\tilde{\mathbf{S}}_2,
\end{aligned}
\label{NEXT_generic_dyn_Jacobian_Decomposition}
\end{equation}
for some $\tilde{\mathbf{S}}_1,\tilde{\mathbf{S}}_2\in\mathcal{M}_{2mn}(\mathbb{R})$, with $\det(\tilde{\mathbf{S}}_1)=\det(\tilde{\mathbf{S}}_2)=1$.
Given \eqref{NEXT_generic_dyn_Jacobian_Decomposition}, \eqref{det_eq1} holds if and only if
\begin{equation*}
\det
\begin{bmatrix}
\mathbf{W}_R-\lambda_u\mathbf{I}-\alpha\nabla^2 F_c^\star+\alpha \left(\mathbf{W}_C-\lambda_u\mathbf{I}\right)^{-1} \left(\mathbf{W}_C-\mathbf{I}\right) \nabla^2 F_c^\ast & \mathbf{0}
\\
\mathbf{0} & \mathbf{W}_C-\lambda_u\mathbf{I}
\end{bmatrix}=0,
\end{equation*}
or equivalently \vspace{-0.2cm}
\begin{equation}
\det
\left(
\mathbf{W}_R-\lambda_u\mathbf{I}-\alpha\nabla^2 F_c^\ast+\alpha \left(\mathbf{W}_C-\lambda_u\mathbf{I}\right)^{-1} \left(\mathbf{W}_C-\mathbf{I}\right) \nabla^2 F_c^\ast\right)=0.
\label{det_eq3}
\end{equation}
Multiplying both sides of \eqref{det_eq3} by $\det(\mathbf{W}_C-\lambda_u\mathbf{I})$ yields \vspace{-0.2cm}
\begin{equation}
Q(\lambda_u)\triangleq \det
\bigg(
\underbrace{\left(\mathbf{W}_C-\lambda_u\mathbf{I}\right)\left(\mathbf{W}_R-\lambda_u\mathbf{I}\right)+\alpha(\lambda_u-1)\nabla^2 F_c^\ast}_{\triangleq\mathbf{T}(\lambda_u)}\bigg)=0.
\label{det_eq4}
\end{equation}
Trivially $Q(\lambda_u)>0$, if $\lambda_u\gg 1$. Therefore, to show that \eqref{det_eq1} holds, it is sufficient to prove that there exists some $\lambda_u>1$ such that $Q(\lambda_u)\leq 0$. Next, we prove this result under either condition (i) or (ii).
Suppose (i) holds; $\mathbf{R}$ and $\mathbf{C}$ are symmetric. Define $\tilde{\boldsymbol{\upsilon}}\triangleq \mathbf{1}\otimes \boldsymbol{\upsilon}$, where $\boldsymbol{\upsilon}$ is the unitary eigenvector associated with {a negative eigenvalue of $\nabla^2 F(\boldsymbol{\theta}^\ast)$, and let $\lambda_\mathrm{min}(\nabla^2 F(\boldsymbol{\theta}^\ast))=-\delta$;} we can write \vspace{-0.2cm}
\begin{equation}
\tilde{\boldsymbol{\upsilon}}^\top\mathbf{T}(\lambda_u) \tilde{\boldsymbol{\upsilon}}=n(\lambda_u-1)\left(\lambda_u-1- \alpha\delta/n\right)<0,
\end{equation}
for all $1<\lambda_u<1+\alpha\delta/n$. By Rayleigh-Ritz theorem, $\mathbf{T}(\lambda_u)$ has a negative eigenvalue, implying that there exists some real value $\bar{\lambda}_u>1$ such that $Q(\bar{\lambda}_u)=0$.
Suppose now that conditions (ii) holds; $\mathbf{W}_R$ and $\mathbf{W}_C$ reduce to $\mathbf{R}$ and $\mathbf{C}$, respectively. Note that $\mathbf{R}$ and $\mathbf{C}$ are now not symmetric. Let $\lambda_u=1+\epsilon$, and consider the Taylor expansion of \vspace{-0.2cm}
\begin{equation}
Q(1+\epsilon)=\det
\bigg(
\left(\mathbf{C}-\mathbf{I}\right)\left(\mathbf{R}-\mathbf{I}\right)+\epsilon\left(\alpha\nabla^2 F_c^\ast+2\mathbf{I}-\mathbf{C}-\mathbf{R}\right)+\epsilon^2\mathbf{I}\bigg),\vspace{-0.2cm}
\end{equation}
around $\epsilon=0$. Define $\mathbf{M}\triangleq \left(\mathbf{C}-\mathbf{I}\right)\left(\mathbf{R}-\mathbf{I}\right)$ and $\mathbf{N}\triangleq \alpha\nabla^2 F_c^\ast+2\mathbf{I}-\mathbf{C}-\mathbf{R}$.
It is clear that $Q(1)=0$; then, by the Jacobi's formula, we have \vspace{-0.2cm}
\begin{equation}
Q(1+\epsilon)=\text{tr}\Big(\text{adj}\left(\mathbf{M}\right) \mathbf{N}\Big)\epsilon+O(\epsilon^2).\vspace{-0.2cm}
\label{Q_function_decomposition1}
\end{equation}
Expanding \eqref{Q_function_decomposition1} yields \vspace{-0.2cm}
\begin{equation}
\begin{aligned}
Q(1+\epsilon)= & \text{tr}\Big(\text{adj}\left(\mathbf{R}-\mathbf{I}\right)\text{adj}\left(\mathbf{C}-\mathbf{I}\right) \mathbf{N}\Big)\epsilon+O(\epsilon^2)
\\
=&\text{tr}\Big(\mathbf{1}\tilde{\mathbf{r}}^\top\tilde{\mathbf{c}}\mathbf{1}^\top \mathbf{N}\Big)\epsilon+O(\epsilon^2)
=(\tilde{\mathbf{r}}^\top\tilde{\mathbf{c}})\mathbf{1}^\top \mathbf{N}\mathbf{1}\epsilon+O(\epsilon^2),
\end{aligned}
\label{Q_function_decomposition2}\vspace{-0.2cm}
\end{equation}
where $\tilde{\mathbf{r}}$ and $\tilde{\mathbf{c}}$ are the Perron vectors of $\mathbf{R}$ and $\mathbf{C}$, respectively. The second equality in \eqref{Q_function_decomposition2} is due to the following fact: a rank-$(n-1)$ matrix $\mathbf{A}\in\mathcal{M}_n(\mathbb{R})$ has rank-$1$ adjugate matrix $\text{adj}\left(\mathbf{A}\right)=\mathbf{a}\mathbf{b}^\top$, where $\mathbf{a}$ and $\mathbf{b}$ are non-zero vectors belonging to the 1-dimensional null space of $\mathbf{A}$ and $\mathbf{A}^\top$, respectively \cite[Sec. 0.8.2]{Horn-Johnson_book}.
We also have $\tilde{\zeta}\triangleq \tilde{\mathbf{r}}^\top\tilde{\mathbf{c}}>0$, due to Lemma \ref{R_C_norms_plus_realAnalyticity}.
Furthermore, since $\boldsymbol{\theta}^\ast\in\Theta_{ss}^\ast$, $\mathbf{1}^\top \nabla^2F_c^\ast\mathbf{1}\leq -\delta$, for some $\delta>0$, and \vspace{-0.2cm}
\begin{equation}
Q(1+\epsilon)\leq -\delta\tilde{\zeta}\alpha\epsilon+O(\epsilon^2),\vspace{-0.2cm}
\end{equation}
which implies the existence of a sufficiently small $\epsilon>0$ such that $Q(1+\epsilon)<0$. Consequently, there must exist some $\bar{\lambda}_u>1$ such that \eqref{det_eq1} holds. Moreover, such $\bar{\lambda}_u$ is a real eigenvalue of $\mathrm{D}\tilde{g}(\mathbf{u}^\ast)$.
To summarize, we proved that there exists an eigenpair $(\bar{\lambda}_u,\mathbf{v}_u)$ of $\mathrm{D}\tilde{g}(\mathbf{u}^\ast)$, with $\bar{\lambda}_u>1$.
Next we show that $(\bar{\lambda}_u,\mathbf{v}_u)$ is also an eigenpair of $\mathrm{D}{g}(\mathbf{u}^\ast)$. Let us partition $\mathbf{v}_u\triangleq (\mathbf{v}^x_u,\mathbf{v}^h_u)$ such that\vspace{-0.2cm}
\begin{equation}
\begin{bmatrix}
\mathbf{W}_R-\alpha \nabla^2 F_c\left(\mathbf{x}^\ast\right) &-\alpha\mathbf{I}
\\
\left(\mathbf{W}_C-\mathbf{I}\right) \nabla^2 F_c\left(\mathbf{x}^\ast\right) & \mathbf{W}_C
\end{bmatrix}
\begin{bmatrix}
\mathbf{v}^x_u
\\
\mathbf{v}^h_u
\end{bmatrix}
=
\bar{\lambda}_u\begin{bmatrix}
\mathbf{v}^x_u
\\
\mathbf{v}^h_u
\end{bmatrix}. \end{equation}
In particular, we have
$\left(\mathbf{W}_C-\mathbf{I}\right)\left( \nabla^2 F_c\left(\mathbf{x}^\ast\right)\mathbf{v}^x_u+\mathbf{v}^h_u\right)=(\bar{\lambda}_u-1)\mathbf{v}^h_u$,
which implies $\mathbf{v}_u^h\in\text{span}(\mathbf{W}_C-\mathbf{I})$, since $\bar{\lambda}_u-1\neq 0$. Therefore, $\mathbf{v}_u\in \mathcal{S}$.
Now, let $\mathbf{P}_\mathcal{S}$ be the orthogonal projection matrix onto $\mathcal{S}$. Since $\mathbf{v}_u\in \mathcal{S}$, we have
\begin{equation}
\label{eigenpair_equivalence}
\mathrm{D}\tilde{g}(\mathbf{u}^\ast)\mathbf{v}_u=\bar{\lambda}_u\mathbf{v}_u \implies \mathrm{D}\tilde{g}(\mathbf{u}^\ast)\mathbf{P}_\mathcal{S}^\top\mathbf{v}_u=\bar{\lambda}_u\mathbf{v}_u \overset{(a)}{\implies}
\mathrm{D}{g}(\mathbf{u}^\ast)\mathbf{v}_u=\bar{\lambda}_u\mathbf{v}_u ,
\end{equation}
where (a) is due to $\mathrm{D}{g}(\mathbf{u}^\ast)=\mathrm{D}\tilde{g}(\mathbf{u}^\ast)\mathbf{P}_\mathcal{S}^\top$ [cf. \eqref{Reimannian_Jacob}]. Hence $(\bar{\lambda}_u,\mathbf{v}_u)$ is also an eigenpair of $\mathrm{D}{g}(\mathbf{u}^\ast)$, which completes the proof.
\end{proof}
\begin{remark}
Note that condition (i) in Proposition \ref{CSS_is_in_unstable_set} implies that $\mathcal{G}_C$ and $\mathcal{G}_R$ are undirected graphs. Condition (ii) extends the network model to directed topologies under assumption $m=1$. {For sake of completeness, we relax condition (ii) in Appendix \ref{CSS_is_in_unstable_set_extension_Appendix} to arbitrary $m\in\mathbb{N}$, under extra (albeit mild) assumptions on the set of strict saddle points and the weight matrices $\mathbf{R}$ and $\mathbf{C}$.}
\end{remark}
\subsubsection{DOGT likely converges to SoS solutions of \eqref{eq:P}}
\label{sec:SOS_main}
Combining Theorem \ref{StableSetofUnstable_fixed_point_is_zeroM},
Proposition \ref{g_is_diffeomorphism}, and Proposition \ref{CSS_is_in_unstable_set}, we can readily obtain the following second-order guarantees of the DOGT algorithms.
{\begin{theorem}
\label{Main_SOS_guarantee_long} Consider Problem \eqref{eq:P}, under Assumptions \ref{P_assumption} and \ref{Net-Assump}; and let $\{\mathbf{u}^\nu\triangleq (\bx^\nu, \mathbf{h}^\nu)\}$ be the sequence generated by the DOGT Algorithm (\ref{NEXT_generic4}) under the following tuning: i) the step-size $\alpha$ satisfies \eqref{alpha_conds_assymptotic_conv} [or \eqref{alpha_conds_assymptotic_conv_final}] and \eqref{eq:step_diff}; the weight matrices $\mathbf{C}$ and $\mathbf{R}$ are chosen according to Assumptions \ref{matrix_C_R} and \ref{nonsingular_RC}; and the initialization is set to $\mathbf{u}^0\in \mathcal{S}$, with $\mathcal{S}$ defined in (\ref{def_S}). Furthermore, suppose that either (i) or (ii) in Proposition \ref{CSS_is_in_unstable_set} holds. Then, we have \vspace{-0.1cm}
\begin{equation}
\mathbb{P}_{\mathbf{u}^0}\left(\lim_{\nu\rightarrow\infty}\mathbf{u}^\nu\in\mathcal{U}^\ast\right)=0,\vspace{-0.1cm}
\label{zero_probability_to_ssp}
\end{equation}
where the probability is taken over $\mathbf{u}^0\in \mathcal{S}$.
{In addition, if} $F$ is a {K\L} function, then $\{\mathbf{x}^\nu\}$ converges almost surely to $\mathbf{1}\otimes \boldsymbol{\theta}^\infty$ at a rate determined in Theorem \ref{Rate_theorem}, where ${\boldsymbol{\theta}}^\infty$ is a SoS solution of \eqref{eq:P}.
\end{theorem}}
{Note that \eqref{zero_probability_to_ssp} implies the desired second-order guarantees only when the sequence $\{\mathbf{u}^\nu\}$ convergences [i.e., the limit in \eqref{zero_probability_to_ssp} exists]; otherwise \eqref{zero_probability_to_ssp} is trivially satisfied, and some limit point of $\{\mathbf{u}^\nu\}$ can belong to $\mathcal{U}^\ast$ with non-zero probability. A sufficient condition for the required global convergence of $\{\mathbf{u}^\nu\}$ is that $F$ is a {K\L} function, which is stated in the second part of the above theorem.}
{
\begin{remark}[Comparison with \cite{Mingyi_SOAlgs_AvoidsSaddles}]\label{DOGT_vs_ProxPDA_remark}
As already discussed in Sec. \ref{sec:intro_distributed}, the primal-dual methods in \cite{Mingyi_SOAlgs_AvoidsSaddles} is applicable to \eqref{eq:P}; it is proved to almost surely converge to SoS solutions.
Convergence of \cite{Mingyi_SOAlgs_AvoidsSaddles} is proved under stricter conditions on the problem than DOGT, namely: i) the network must be {\it undirected}; and ii) the Hessian of each local $f_i$ must be Lipschitz continuous. It does \emph{not} seem possible to extend the analysis of \cite{Mingyi_SOAlgs_AvoidsSaddles} beyond this assumptions.
\end{remark}
}
{
\section{Numerical Results}
\label{numerical_experiments}
In this section we test the behavior of DGD and DOGT around strict saddles on three classes of nonconvex problems, namely: i) a quadratic function (cf. Sec. \ref{sec:sim-quadratic}); ii) a classification problem based on the cross-entropy risk function using sigmoid activation functions (cf. Sec. \ref{BLR_simulations}); and iii) a two Gaussian mixture model (cf. Sec. \ref{GMM_simulations}).}
\subsection{Nonconvex quadratic optimization}\label{sec:sim-quadratic}
Consider
\vspace{-0.3cm}
\begin{equation}
\min_{\boldsymbol{\theta}\in \mathbb{R}^{m}}\, F\left(\boldsymbol{\theta}\right)= \frac{1}{2}\sum_{i=1}^{n} \left(\boldsymbol{\theta}-\mathbf{b}_i\right)^\top\mathbf{Q}_i\left(\boldsymbol{\theta}-\mathbf{b}_i\right),\vspace{-0.2cm}
\label{eq:P_test}
\end{equation}
where $m=20$; $n=10$; $\mathbf{b}_i$'s are i.i.d Gaussian zero mean random vectors with standard deviation $10^3$; and $\mathbf{Q}_i$'s are $m\times m$ randomly generated symmetric matrices where $\sum_{i=1}^n\mathbf{Q}_i$ has $m-1$ eigenvalues $\{\lambda_i\}_{i=1}^{m-1}$ uniformly distributed over $(0,n]$, and one negative eigenvalue $\lambda_m=-n\delta$, with $\delta=0.01$. Clearly \eqref{eq:P_test} is an instance of Problem \eqref{eq:P}, with
$F$ having a unique strict saddle point $\boldsymbol{\theta}^{\ast}=\left(\sum_i\mathbf{Q}_i\right)^{-1}\sum_i\mathbf{Q}_i\mathbf{b_i}$. The network of $n$ agents is modeled as a ring; the weight matrix $\mathbf{W}\triangleq\{w_{ij}\}_{i,j=1}^{n}$, compliant to the graph topology, is generated to be doubly stochastic.
To test the escaping properties of DGD and DOGT from the strict saddle of $F$, we initialize the algorithms in a randomly generated neighborhood of $\boldsymbol{\theta}^{\ast}$. More specifically, every agent's initial point is $\mathbf{x}_i^0=\boldsymbol{\theta}^{\ast}+\boldsymbol{\epsilon}_{x,i}$, $i\in[n]$. In addition, for the DOGT algorithm, we set $\mathbf{y}^0_i=\nabla f_i(\mathbf{x}_i^0)+(w_{ii}-1)\boldsymbol{\epsilon}_{y,i}+\sum_{j\neq i}w_{ij}\boldsymbol{\epsilon}_{y,j}$, where $\boldsymbol{\epsilon}_{x,i}$'s and $\boldsymbol{\epsilon}_{y,i}$'s are realizations of i.i.d. Gaussian random vectors with standard deviation equal to 1. {Both algorithms use the same step-size $\alpha=0.99\,\sigma_\mathrm{min}(\mathbf{I}+\mathbf{W})/L_c$, with $L_c=\max_i\{|\lambda_i|\}$; this is the largest theoretical step-size guaranteeing convergence of the DGD algorithm (cf. Theorem \ref{DGD_conv_cons_thm}).}
\begin{figure}[t]
\center
\includegraphics[scale=0.22]{figs/NoncvxQuadratic/1.eps}\includegraphics[scale=0.22]{figs/NoncvxQuadratic/2.eps}\vspace{-0.2cm}
\caption{Escaping properties of DGD and DOGT, applied to Problem (\ref{eq:P_test}). Left plot: distance of the average iterates from $\boldsymbol{\theta}^{\ast}$ projected onto the unstable manifold ${E}_u$ versus the number of iterations. Right plot: distance of the average iterates from $\boldsymbol{\theta}^{\ast}$ versus the number of iterations. }
\label{DGD_DOGT_escaping_Figs}\vspace{-0.3cm}
\end{figure}
In the left panel of Fig. \ref{DGD_DOGT_escaping_Figs}, we plot the distance of the average iterates $\bar{\mathbf{x}}^\nu=(1/n)\sum_{i=1}^n\mathbf{x}_i^\nu$ from the critical point $\boldsymbol{\theta}^{\ast}$ projected on the unstable manifold ${E}_{u}=\mathrm{span}(\mathbf{u}^u)$, where $\mathbf{u}^u$ is the eigenvector associated with the negative eigenvalue $\lambda_m=-n\delta$. In the right panel, we plot $\|\bar{\mathbf{x}}^\nu - \boldsymbol{\theta}^{\ast}\|$ versus the number of iterations. All the curves are averaged over 50 independent initializations. Figure in the left panel shows that, as predicted by our theory, both algorithms almost surely escapes from the unstable subspace $E_u$, at an indistinguishable practical rate.
The right panel shows that DOGT gets closer to the strict saddle; this can be justified by the fact that, differently from DGD, DOGT exhibits \emph{exact} convergence to critical points.\vspace{-0.1cm}
{
\subsection{Bilinear logistic regression \cite{Dyrholm_JMLR_2007}}\label{BLR_simulations}
Consider a classification problem with distributed training data set $\{\mathbf{s}_i,\xi_i\}_{i=1}^n$, where $\mathbf{s}_i\in\mathbb{R}^d$ is the feature vector associated with the binary class label $\xi_i\in\{0,1\}$. The bilinear logistic regression problem aims at finding the bilinear classifier $\zeta_i(\mathbf{Q},\mathbf{w};\mathbf{s}_i)= \mathbf{s}_i^\top\mathbf{Q}\mathbf{w}$, with $\mathbf{Q}\in\mathbb{R}^{d\times p}$ and $\mathbf{w}\in\mathbb{R}^p$ that best separates data with distinct labels. Let $(\mathbf{s}_i,\xi_i)$ be private information for agent $i$. Using the sigmoid activation function $\sigma(x)\triangleq 1/(1+e^{-x})$ together with the \emph{cross-entropy risk} function, the optimization problem reads
\begin{equation}
\min_{\mathbf{Q},\mathbf{w}}\quad -\frac{1}{n} \sum_{i=1}^n \Big[\xi_i \ln\big(\sigma(\mathbf{s}_i^\top\mathbf{Q}\mathbf{w})\big)+(1-\xi_i) \ln\big(1-\sigma(\mathbf{s}_i^\top\mathbf{Q}\mathbf{w})\big)\Big]+\frac{\tau}{2}\left(\norm{\mathbf{Q}}^2_F+\norm{\mathbf{w}}^2\right).
\label{eq:P_BilinearLogisticReg}
\end{equation}
It is not difficult to show that $\eqref{eq:P_BilinearLogisticReg}$ is equivalent to the following instance of \eqref{eq:P}: \vspace{-0.2cm}
\begin{equation}
\min_{\mathbf{Q},\mathbf{w}}\quad F(\mathbf{Q},\mathbf{w})= \sum_{i=1}^n ~\underbrace{\frac{1}{n}\left[-\ln\big(\sigma(\tilde{\xi}_i\mathbf{s}_i^\top\mathbf{Q}\mathbf{w})\big)+\frac{\tau}{2}\left(\norm{\mathbf{Q}}^2_F+\norm{\mathbf{w}}^2\right)\right]}_{= f_i(\mathbf{Q},\mathbf{w})},
\label{eq:P_BilinearLogisticReg_EquivalentForm}
\vspace{-0.3cm}
\end{equation}
with\vspace{-0.4cm}
\begin{equation*}
\tilde{\xi_i}\triangleq \left\{
\begin{array}{l l}
-1, & \mathrm{if}\quad \xi_i=0;
\\
1, & \mathrm{if}\quad \xi_i=1.
\end{array}
\right.
\end{equation*}
To visualize the landscape of $F(\mathbf{Q},\mathbf{w})$ (2D plot), we consider the following setting for the free parameters. We set $d=p=1$; $\tau=0.2$; $n=5$; and we generate uniformly random $\tilde{\xi}_i\in\{0,1\}$, and we draw $s_i$ from a normal distribution with mean $\xi_i$ and variance $1$.
The gradient of the local loss $f_i$ reads
\begin{equation*}
\begin{bmatrix}
\nabla_{Q} f_i(Q,w)
\\
\nabla_{w} f_i(Q,w)
\end{bmatrix}=\frac{1}{n}
\begin{bmatrix}
\tau Q-\tilde{\xi}_is_iw\sigma(-\tilde{\xi}_is_iQw)
\\
\tau w-\tilde{\xi}_is_iQ\sigma(-\tilde{\xi}_is_iQw)
\end{bmatrix}.
\end{equation*}
A surface plot of $F({Q},{w})$ in the above setting is plotted in the right panel of Fig. \ref{Contour_plots}.
Note that such $F$ has three critical points, two of which are local minima (see the location of minima in the left or middle panel of Fig. \ref{Contour_plots} marked by $\times$) and one strict saddle point at $(0,0)$--the Hessian at $(0,0)$,\vspace{-0.2cm}
\begin{equation*}
\nabla^2 F(0,0)=
\begin{bmatrix}
\tau & -\frac{1}{2n}\sum_i\tilde{\xi}_is_i
\\
-\frac{1}{2n}\sum_i\tilde{\xi}_is_i & \tau
\end{bmatrix},\vspace{-0.15cm}
\end{equation*}
has an eigenvalue at $\tau-\frac{1}{2n}\sum_i\tilde{\xi}_is_i=-0.26$.
\begin{figure}[t!]
\hspace{-.3cm}\includegraphics[scale=0.32]{figs/BilLogReg/Contour_Directed.eps}\hspace{-.2cm}\includegraphics[scale=0.32]{figs/BilLogReg/Contour_Undirected.eps}\hspace{-.15cm}\includegraphics[scale=0.43]{figs/BilLogReg/surface.eps}\vspace{-0.2cm}
\caption{Escaping properties of the DGD and DOGT, applied to the bilinear logistic regression problem \eqref{eq:P_BilinearLogisticReg}. Left (resp. middle) plot: directed (resp. undirected) network; trajectory of the average iterates on the contour of $F$ ($(0,0)$ is the strict saddle point and $\times$ are the local minima); DGD and DOGT are initialized at $\Box$ and terminated after $100$ iterations at $\ast$. Right plot: plot of $F$.\vspace{-0.5cm}}
\label{Contour_plots}
\end{figure}
We test DGD and DOGT over a network of $n=5$ agents; for DGD we considered undirected graphs whereas we run DOGT on both undirected and directed graphs. Both algorithms are initialized at the same random point and terminated after $100$ iterations; the step-size is set to $\alpha=0.9$. We denote by $Q^\nu_i$ and $w^\nu_i$ the agent $i$'s $\nu$-the iterate of the local copies of $Q$ and $w$, respectively. The trajectories of the average iterates $(\bar{Q}^\nu,\bar{w}^\nu)\triangleq \frac{1}{n}(\sum_iQ^\nu_i,\sum_iw^\nu_i)$ are plotted in Fig. \ref{Contour_plots}; the left panel refers to the directed graph while the middle panel reports the same results for the undirected network. As expected, the DOGT algorithm converges to an exact critical point (local minimum) avoiding the strict saddle $(0,0)$ while DGD converges to a neighborhood of the local minimum. The consensus error is $1/n\sqrt{\sum_{i=1}^n||(Q_i^\nu,w_i^\nu)-(\bar{Q}^\nu,\bar{w}^\nu)||^2}$; at the termination, it reads $2.33\times 10^{-4}$ for DOGT over the directed network, and $2.18\times 10^{-4}$ and $9.74\times 10^{-2}$ for DOGT and DGD, respectively, over undirected networks.
\subsection{Gaussian mixture model}\label{GMM_simulations}
Consider the Gaussian mixture model defined in Sec. \ref{ProblemSetting_SubSe}. The data $\{\mathbf{z}_i\}_{i=1}^n$ where $\mathbf{z}_i\in\mathbb{R}^m$ are realizations of the mixture model $\mathbf{z}_i\sim \frac{1}{2}\mathcal{N}(\boldsymbol{\mu}_1,\boldsymbol{\Sigma}_1)+\frac{1}{2}\mathcal{N}(\boldsymbol{\mu}_2,\boldsymbol{\Sigma}_2)$. Let each agent $i$ own $\mathbf{z}_i$. Both parameters $(\boldsymbol{\mu}_1,\boldsymbol{\mu}_2)$ and $(\boldsymbol{\Sigma}_1,\boldsymbol{\Sigma}_2)$ are unknown. The goal is to approximate $(\boldsymbol{\mu}_1,\boldsymbol{\mu}_2)$ while $(\boldsymbol{\Sigma}_1,\boldsymbol{\Sigma}_2)$ is set to an estimate $(\tilde{\boldsymbol{\Sigma}},\tilde{\boldsymbol{\Sigma}})$. The problem reads\vspace{-0.2cm}
\begin{equation}
\min_{\boldsymbol{\theta}_1,\boldsymbol{\theta}_2\in \mathbb{R}^{m}}\, -\sum_{i=1}^n\log{\left(\phi_m(\mathbf{z}_i-\boldsymbol{\theta}_1)+\phi_m(\mathbf{z}_i-\boldsymbol{\theta}_2)\right)},\vspace{-0.2cm}
\label{eq:GaussianMixtureProblem}
\end{equation}
where $\phi_m(\boldsymbol{\theta})$ is the $m$-dimensional normal distribution with mean $\mathbf{0}$ and covariance $\tilde{\boldsymbol{\Sigma}}$.
Consider the case of mixture of two scalar Gaussians, i.e., $m=1$. We draw $\{\mathbf{z}_i\}_{i=1}^5$ from the the this model, with means $\mu_1=0$, $\mu_2=-5$ and variance $\sigma_1=\sigma_2=25$. The estimate of variance in problem \eqref{eq:GaussianMixtureProblem} is pessimistically set to $\tilde{\sigma}=1$. A surface plot of a random instance of above problem is depicted in right panel of Fig. \ref{GausMix_Contour_plots}. Note that this instance of problem has 2 global minima (marked by $\times$) and multiple local minima.
We test DGD and DOGT on the above problem over the same networks as described in Sec. \ref{BLR_simulations}. Both algorithms are initialized at the same random point and terminated after $250$ iterations; the step-size is set to $\alpha=0.1$. In Fig. \ref{GausMix_Contour_plots}, we plot the trajectories of the average iterates $(\bar{\theta}_1^\nu,\bar{\theta}_2^\nu)\triangleq \frac{1}{n}(\sum_i\theta^\nu_{1,i},\sum_i\theta^\nu_{2,i})$, where $\theta^\nu_{1,i}$ and $\theta^\nu_{2,i}$ are the agent $i$'s $\nu$-the iterate of the local copies of $\theta_1$ and $\theta_2$, respective; the left (resp. middle) panel refers to the undirected (resp. directed) network. DOGT converges to the global minimum while DGD happens to converge to neighborhood of a local minima. The consensus error is measured by $(1/n)\sqrt{\sum_{i=1}^n||(\theta_{1,i}^\nu,\theta_{2,i}^\nu)-(\bar{\theta}_1^\nu,\bar{\theta}_2^\nu)||^2}$ and at the termination it is equal to $1.9\times 10^{-3}$ for DOGT on the directed graph; and $2.8\times 10^{-3}$ and $1.135$ for DOGT and DGD, respectively over the undirected graph.
\begin{figure}[t!]
\hspace{-.15cm}\includegraphics[scale=0.32]{figs/GausMix/contour_directed.eps} \hspace{-.45cm}\includegraphics[scale=0.32] {figs/GausMix/contour_undirected.eps}\hspace{-.32cm}\includegraphics[scale=0.38]{figs/GausMix/surface2.eps}
\vspace{-0.2cm}
\caption{Escaping properties of the DGD and DOGT applied to the Gaussian mixture problem (\ref{eq:GaussianMixtureProblem}). Left (resp. middle) plot: directed (resp. undirected) network; trajectory of the average iterates on the contour of $F$ (the global minima are marked by $\times$); DGD and DOGT are initialized at $\Box$ and terminated after $250$ iterations at $\ast$. Right plot: plot of $F$.\vspace{-0.5cm}}
\label{GausMix_Contour_plots}
\end{figure}
}
\vspace{-0.1cm}
|
\section{Introduction}\label{sec:introduction}
The study of time series has a long history and the literature for it covers many different methods (\cite{hamilton}).
The study of asynchronous time series is an important subset of this.
Asynchronous time series are series for which features are sampled at irregular time intervals, and at any given time step new values of any subset of features may be present.
When a feature does not change values often it can be treated as being present only at times of change.
For example, in industrial IoT several sensors can be monitored, each one with its own sampling rate.
We can have a feature that corresponds to the production rate that seldom changes and it impacts predictions.
This clearly presents difficulties from both a modeling and data input perspective.
Fixed sampling, repeating missing values, and other types of data imputation have all been used to varying degrees of success.
More recent attempts using machine learning for asynchronous time series use Gaussian processes (\cite{cunningham,ligaussian}), which are hard to scale in the presence of many features.
Deep learning is a much more recent development, and it is achieving great success in many fields, e.g. machine vision (\cite{alexnet, inception, resnet}) and natural language processing (\cite{googletranslate, nlp}).
Recurrent neural networks (RNNs) are a type of a neural network that use shared weights applied to sequences.
These types of networks achieve state of the art performance for sequential data.
Time series can be thought of as sequences in the RNN context, and thus they are amenable to RNN-based solutions (\cite{malhotra2015long,che2018recurrent,review}(.
However, many of the most widely used RNN cells, e.g. the Long Short Term Memory (LSTM) cell (\cite{lstm}), are built for problems where either time does not matter or there is a constant time step size.
The adaptation of these cells for asynchronous sequences is a more recent development, varying from combining traditional methods with deep learning (\cite{binkowski, amazon}) to pure neural network approaches (\cite{phasedlstm, tlstm}).
These attempts may take into account time between samples and the patterns of missing values, but they do not fundamentally treat features that are present with varying levels of frequency any differently.
Thus, a feature that occurs at every time step and a feature that almost never occurs at any given time step update the cell in the same manner.
We address this shortcoming by introducing a new recurrent cell that uses features that are present frequently, called dense features, differently from features that are rarely present, termed sparse features, when updating the cell's hidden and memory states.
In particular, our cell's hidden state is split into two parts, one part for dense features and one part for sparse features.
Additionally, each sparse feature maintains its own hidden and memory state that are only updated at time steps when that feature is present.
Thus, the update for the sparse part of the overall hidden state depends on the subset of sparse features present at any given time step, while the update for the dense part of the overall hidden state happens in the standard LSTM manner.
In addition to handling these two feature types, our cell also accounts for the irregular time between time steps, expanding on the work of \citet{tlstm} by allowing more flexibility in how time features are used.
Specifically, we modify the decay function to handle an arbitrary number of delta features, not just the time elapsed since the last time step.
The model also allows for sequence level features, for which time related features, called static delta features, are handled in the same manner as within the recurrent cell, and the non time related features, referred to as static dense features, are embedded and concatenated with the sequence output in the standard manner.
Overall, we have five feature types.
The main contributions of this work are as follows.
First, the introduction of a new recurrent cell that accounts for asynchronous feature sampling and treats features in a variable manner dependent on the frequency with which their values are present.
Second, an expansion of the time-aware LSTM (TLSTM) to handle multiple time features, and the extension of this method applied to the output of the recurrent cell itself.
Third and finally, a general recurrent framework to incorporate the five aforementioned feature types.
\section{Related work}
When dealing with an irregularly sampled multivariate sequence, one can consider the variables not present at any given time step to be missing.
There is some history of work done on recurrent networks with missing data (\cite{bengio1996recurrent, tresp1998solution}).
\citet{lipton2016directly} test the performance of different types of imputation and missing value indicator input features.
This work inputs the missing values patterns into the recurrent network by concatenating with the feature vectors while using a standard LSTM network.
They find zero filling missing features and using indicator features perform better than data imputation, which may actually interfere with the network's ability to pick out missing values.
\citet{che2018recurrent} investigate a similar approach and improve upon this by modifying a GRU (\cite{gru}) cell to separately incorporate the missing value patterns as well as the time between successive measurements for each feature.
It does this by introducing two decay mechanisms.
The first decay acts to bring a missing value toward the baseline mean value from its last measured value so the old value is not input repeatedly.
This work treats all features in the same manner, regardless of the frequency a missing value is present for any given feature.
The work in \cite{che2018recurrent} does not differentiate between features that never have missing values and features that are almost completely composed of missing values.
The second decay mechanism uses the time feature to decay the hidden state before calculating the new hidden state.
We note that missing data at a time is not the same as features not being present (sampled) or a feature changing value seldomly.
We focus on the latter while \cite{che2018recurrent,bengio1996recurrent,tresp1998solution} focus on the former.
\citet{tlstm} incorporate the time between events for event based sequences with irregular intervals in a similar fashion to the second decay mechanism in \cite{che2018recurrent}.
This is done by using decomposing the memory state into short and long term components, and then multiplying the short term component by a decay factor that decreases as the time between events increases.
We use this in our own recurrent cell, and also expand upon the idea by incorporating multiple time features, not just limited to the time between adjacent events.
We also use the same decomposition and decay method outside of our recurrent cell since the prediction time occurs some time after the final event.
Their work does not consider sparse features.
\citet{phasedlstm} modify the standard LSTM cell to address its issues with asynchronous time series.
They do this by incorporating a time gate into the network that controls when and to what extent the hidden and memory states can be updated.
This means that there are time steps where no updates are made, and so it is as if the network is accumulating changes in the feature vector before updating the hidden and memory states.
While the time gate has learnable parameters, they still do not differentiate between features that occur at every time step versus those that rarely occur in the cell updates.
In summary, there is no prior work that considers the notion of sparse features and a single model for all five feature types.
\section{Models}
The main contribution of this paper is the development of a recurrent cell which handles five distinct feature types.
The following subsections review each of the feature types and describes how they are handled in the Sparse Time LSTM (STLSTM model).
The five types are referred to as the following: dense features, sparse features, delta features, static dense features, and static delta features.
The dense features are the standard RNN sequence input, such as the embedding of an event type or the sequence state features at a given time step, and are present at any given time in a sequence.
The sparse features change values infrequently, and thus they can be considered as being present only at select times.
Stating that the value is not present means the feature value has not changed from the previous time.
Another requirement is that the feature values are not time related.
The delta features must be related to measure of time since some specified condition, such as the number of days since that last event in an event driven sequence.
These features are particularly useful when events in the sequence are clustered, meaning that we observe many events in close proximity, and then a long gap until additional events occur.
Both the static dense and the static delta features give information about the sequence as a whole, rather than being tied to a specific time step.
Static dense features are not time related, e.g. the gender of a sequence's user, whereas static delta features are time related, e.g. the number of days since the last feature vector of the sequence.
Figure \ref{fig:urnn_concepts} depicts these concepts.
\begin{figure}[!htb]
\centering
\begin{minipage}{.46\textwidth}
\centering
\includegraphics[width=\textwidth]{feature_concepts.png}
\caption{The framework of the five feature types.}
\label{fig:urnn_concepts}
\end{minipage}%
\qquad
\begin{minipage}{0.46\textwidth}
\centering
\includegraphics[width=\textwidth]{delta_features.png}
\caption{The memory state is decomposed into short and long term components, and then the standard LSTM updates are performed.}
\label{fig:urnn_delta}
\end{minipage}
\end{figure}
We consider a neural network to be a mapping $f(X)$, where $f$ represents the network and $X$ is the input to the network.
We split $X$ first into two parts, $X = (x^{static}, x)$ where $x^{static}$ represents the sequence level static features and $x$ represents the sequence features.
The static features can be split into two parts $x^{static} = (x^{static, d}, x^{static, \Delta})$, for static dense and static delta features respectively.
The sequence features can further be broken down into three parts, $x = (x^d, x^{\Delta}, x^{sp})$ that represent the dense, delta, and sparse features respectively.
For notational convenience, each of these features has a vector at each time step of the sequence (even though in our implementation the sparse features would only be used when the value changes).
The dense features are denoted as $x^d = (x^d_1, x^d_2, ..., x^d_T)$ for a sequence with $T$ time steps.
The delta features are represented as $x^d = (x^{\Delta}_1, x^{\Delta}_2, ..., x^{\Delta}_T)$ where $x^{\Delta}_{ti}$ is proportional to $t - \overline{t}$ for some $1 \leq \overline{t} \leq t$.
The sparse features, however, are represented slightly differently than the previous two types of features, because their changes are not always present.
We use the term present for a sparse feature when its value changes.
A single sparse feature $k$ can be represented as a sequence of tuples, $x^{sp}_k = ((m_{1k}, x^{sv}_{1k}), ..., (m_{Tk}, x^{sv}_{Tk}))$, where $m_{tk} \in \{0,1\}$ is a mask value that is $1$ when the feature is present and $0$ if the feature is missing.
If $m_{tk}=1$, then $x^{sv}_{tk}$ is the actual new feature value, and otherwise $x^{sv}_{tk}$ is set to equal anything, e.g. $x^{sv}_{tk} = 0$.
If a sparse feature is present at a time, then this input is passed into and used in the cell updates.
If there is no sparse input at a given time step, the sparse part of the cell is untouched at that time step.
Alternatively, we can represent a sparse feature by a set of pairs $(t_{k}, x^{sv}_{k})$.
Our recurrent cell is built off of the commonly used LSTM \cite{lstm}, as well as its time-aware variant (TLSTM) \cite{tlstm}.
For LSTM-type cells, there is a memory state, $C$, and a hidden state, $h$, which pass from the cell at the previous time step and into the cell along with the features at the current time step.
The memory state boosts performance for long term dependencies, and the gate structure of the LSTM allows selective parts of the memory to be forgotten and updated.
The standard LSTM architecture does not address irregularly sampled events, however, and the TLSTM variant addresses this by using the time elapsed between events to modify the memory state.
In the TLSTM cell, at each time step, first a fully connected layer is used to decompose the cell's memory state into short and long term components.
The time elapsed is passed as the input to a non-increasing decay function that maps to a scalar decay factor.
The short term component of the memory state is multiplied by the decay factor that decreases as more time passes.
The equations for a single delta feature $x^{\Delta}_{t1}$ for this process are below on the left,
where $g$ is a decay function, $x^{\Delta}_{t1}$ is the delta input, and $C_{t-1}$ is the previous cell memory state.
After the memory state is modified, the cell update is performed in the standard LSTM manner, with the exception that the modified cell memory state calculated in Equation \ref{eqn:urnn_1} is used to update the current memory state.
The equations are as follows on the right where $x^d_{t}$ is the feature vector to the current time step.
A diagram of this base cell is given in Figure \ref{fig:urnn_delta}.
\begin{eqnarray}
\begin{split}
\label{eqn:urnn_1}
C^{S}_{t-1} &= \tanh \left( W^{\Delta} C_{t-1} + b^{\Delta} \right) \nonumber \\
\hat{C}^{S}_{t-1} &= C^{S}_{t-1} \cdot g\left( x^{\Delta}_{t1} \right) \\
C^{L}_{t-1} &= C_{t-1} - C^{S}_{t-1} \nonumber \\
C^{*}_{t-1} &= C^{L}_{t-1} + \hat{C}^{S}_{t-1}, \nonumber
\end{split}
\quad \quad \quad \quad \quad
\begin{split}
\label{eqn:urnn_2}
f^d_t &= \sigma ( W^d_{fh} h_{t-1} + W^d_{fx} x^{d}_{t} + b^d_f ) \nonumber \\
i^d_t &= \sigma ( W^d_{ih} h_{t-1} + W^d_{ix} x^{d}_{t} + b^d_i ) \nonumber \\
\tilde{C^d_t} &= \sigma ( W^d_{Ch} h_{t-1} + W^d_{Cx} x^{d}_{t} + b^d_C ) \\
C^d_t &= f^d_t * C^{*}_{t-1} + i^d_t * \tilde{C^d_t}, \nonumber \\
o^d_t &= \sigma ( W^d_{oh} h_{t-1} + W^d_{ox} x^{d}_{t} + b^d_o ) \nonumber \\
h^d_t &= o^d_t * tanh(C^d_t) \nonumber
\end{split}
\end{eqnarray}
\subsection{Dense features}
The dense features are the standard features used for recurrent models.
It is assumed that the features are not time related, e.g. they encode attributes of events, but not times between events.
At each time step, the hidden state from the previous time step, the memory vector, and the feature vector for the current time step are input into the cell, and the hidden state and memory vector are updated for the next time step.
\subsection{Delta features}
Delta features are not relevant or applicable for some RNN applications such as text or video, but for a sequence with measurements or events that occur with varying frequency, time information can add predictive power.
In the beginning of this section, a model expressed by Equation \ref{eqn:urnn_1}
is presented that allows a single delta feature.
Here we extend this to multiple delta features.
Incorporating multiple deltas requires a simple modification to the decay function $g$.
Instead of mapping from a scalar to a scalar, it maps from a vector to a scalar.
The proposed $g$ function is $g(x^{\Delta}_t) = \frac{1}{\log \left( e + \alpha^T x^{\Delta}_t \right)}$,
where $\alpha$ is a trainable vector.
\subsection{Sparse features}
Sparse features are similar in nature to the dense features, with the exception that their values change very rarely throughout the sequence.
Like the dense features, the sparse features describe the state of the sequence at a particular time step.
However, since they rarely change from one step to another, when treated in the same manner as dense features, the sparse features would just repeat inputs the vast majority of time steps.
There is no clear cutoff point in new value frequency for which a dense feature turns into a sparse feature.
We propose a new recurrent cell, called the Sparse Time Long Short Term Memory (STLSTM) cell, that attempts to more strongly capture the change in sparse features.
The STLSTM cell has a hidden state split into two parts, one corresponding to the dense features, and the other corresponding to the sparse features, $h_t = (h^d_t, h^{sp}_t)$.
Further, each sparse feature has its own memory state, and this memory state is only updated when there is a change in the feature value.
At each time step, the dense part of the hidden state is updated based on Equation \ref{eqn:urnn_1} and Equation 2.
For the sparse part of the hidden state, there is a proposed hidden state from each sparse feature, and these proposed hidden states are aggregated together.
Let us denote the memory state of sparse feature $k$ as $C^{sp}_{tk}$ and the corresponding hidden state as $h^{sp}_{tk}$.
The main idea is to not change $C^{sp}_{tk}$ if $m_{tk} = 0$, i.e. to simply carry it over, and otherwise if $m_{tk} = 1$ to use the feature $x^{sv}_{tk}$ within LSTM-like update equations.
Formally, if $m_{tk} = 0$, then there is no change to the memory state or the hidden state, or
\begin{equation}
C^{sp}_{tk} = C^{sp}_{t-1,k} , \hspace{2mm} h^{sp}_{tk} = h^{sp}_{t-1,k}. \nonumber
\end{equation}
If there is a change in the feature value ($m_{tk} = 1$), the new memory state is determined by the following equations:\begin{eqnarray}
f^{sp}_{tk} &=& \sigma ( W^{sp}_{fh} h_{t-1} + W^{sp}_{fx} x^{sv}_{tk}+ b^{sp}_f ) \nonumber \\
i^{sp}_{tk} &=& \sigma ( W^{sp}_{ih} h_{t-1} + W^{sp}_{ix} x^{sv}_{tk} + b^{sp}_i ) \nonumber \\
\tilde{C}^{sp}_{tk} &=& \sigma ( W^{sp}_{Ch} h_{t-1} + W^{sp}_{Cx} x^{sv}_{tk} + b^{sp}_C ) \nonumber \\
C^{sp}_{tk} &=& f^{sp}_{tk} * C^{sp}_{t-1,k} + i^{sp}_{tk} * \tilde{C}^{sp}_{tk} \nonumber.
\end{eqnarray}
Once the memory state is calculated, the hidden state parts are calculated using standard output gates.
The equations for this are as follows:
\begin{eqnarray}
o^{sp}_{tk} &=& \sigma ( W^{sp}_{oh} h_{t-1} + W^{sp}_{ox} x^{sv}_{tk} + b^{sp}_o ) \nonumber \\
h^{sp}_{tk} &=& o^{sp}_{tk} * tanh(C^{sp}_{tk}) \nonumber \\
h_t &=& \left[ h^d_t, \mathcal{L} (h^{sp}_{t1}, ..., h^{sp}_{tm}; W_{ah})\right] \nonumber \\
o_t &=& \left[ o^d_t, \mathcal{L} (o^{sp}_{t1}, ..., o^{sp}_{tm}; W_{ao})\right] \nonumber
\end{eqnarray}
Here $\mathcal{L}$ is an aggregation function that yields a vector of the same dimension as $h^{sp}_{tk}$ and $W_{ah}, W_{ao}$ are trainable parameters.
For example, $\mathcal{L}$ can be simply averaging the hidden state proposals or we can place a fully connected layer over all of the sparse hidden states, and use this layer to compute a single hidden state for all sparse features.
A diagram for the STLSTM cell is depicted in Figure \ref{fig:urnn_stlstm}.
\begin{figure}[!htb]
\centering
\begin{minipage}{.5\textwidth}
\centering
\includegraphics[width=\textwidth]{sparse_features.png}
\caption{Full wiring for the STLSTM cell.}
\label{fig:urnn_stlstm}
\end{minipage}%
\qquad
\begin{minipage}{0.43\textwidth}
\centering
\includegraphics[width=\textwidth]{static_features.png}
\caption{Static features are embedded and then concatenated to RNN output.}
\label{fig:urnn_static}
\end{minipage}
\end{figure}
\subsection{Static dense features}
All previous discussions dealt with features that can be assigned to a specific time step in a sequence, but we can also have features that are sequence level instead of time step level.
These features, which are not time related, give information about the sequence as a whole, so they should not serve as inputs to the recursive part.
The incorporation of these features is fairly simple.
First the features are passed through a fully connected layer to get an embedding, and then this embedding is concatenated with the output, $h_T$, of the recurrent cell at the final time steps.
A diagram of this can be seen in Figure \ref{fig:urnn_static}.
\subsection{Static delta features}
Similar to the static dense features, the static delta features apply to the sequence as a whole rather than an individual time step.
Their values are related to the time since the last feature vector.
These features are used in the same manner as the delta features within the STLSTM cell.
The output of the recurrent network, $h_T$, is decomposed into short and long term components, and the decay factor is applied to the short term component before being added back to the long term component.
Formally, this is given as\begin{eqnarray}
h^{S}_{T} &=& \tanh \left( W^{static, \Delta} h_{T} + b^{static, \Delta} \right) \nonumber \\
\hat{h}^{S}_{T} &=& h^{S}_{T} \cdot g\left( x^{static, \Delta} \right) \nonumber \\
h^{L}_{T} &=& h_{T} - h^{S}_{T} \nonumber \\
h^{*}_{T} &=& h^{L}_{T} + \hat{h}^{S}_{T} \nonumber
\end{eqnarray}
This modified output is then concatenated with the embedding of the static dense features before input to the decoder.
\section{Computational study}
For all experiments in this section we use Keras with a Tensorflow backend on a single GeForce GTX 1080 GPU card.
We use ADAM as the optimization algorithm, and keep track of the validation F1 score to control the number of training epochs.
If the validation F1 score does not increase for 15 epochs, then training is terminated.
We use standard weight initalization, Glorot uniform for weights connected to the inputs and orthogonal initialization for the recurrent weights.
To test out the STLSTM cell with all five feature types, we use two data sets, one public and one proprietary.
\subsection{Power consumption data set}
A modified version of the UCI household electric power consumption dataset (\cite{ucirvine}) is the public data set.
This dataset contains measurements for seven different electrical quantities and sub-metering values with a sampling rate of one minute taken over the course of nearly four years.
All of the features are sampled every minute, with the exception that approximately $1.25\%$ of records have no measurements.
We simply fill in these missing values by repeating the record that immediately precedes it.
Since the data set does not have natural sparse and delta features, we have to artificially create them, which is described in the Appendix.
For the experiments described in this section, we use sequences covering two hours worth of power data.
The recurrent part of our network is composed of two stacked RNN cells.
We compare using two architectures: one has STLSTM at the bottom layer and LSTM at the second layer, and the other has TLSTM at the bottom layer and LSTM at the second layer.
Experiments have shown that using two layers is optimal.
We use a standard LSTM cell as the second layer instead of STLTSM or TLSTM cells because we assume the time and sparsity information is encoded in the output of the base layer that is fed into the second layer.
These all have a hidden dimension of 64, which is also the size of the single embedding layer for the static dense features.
For all results the static features are present, unless it is stated otherwise.
The first metric we investigate is the relative performance of the STLSTM cell versus the standard TLSTM cell versus sparsity of individual sparse features.
For this we use a dense layer as the aggregation function in the STLSTM cell.
For each of the seven features, we set that feature to be sparse with a sparse ratio ranging from $0.01$ to $0.15$ and the remaining six to be dense.
Figure \ref{fig:urnn_relative} shows the relative performance on F1 scores between the TLSTM and STLSTM models for both types of record subsampling, with percentages greater than $0$ meaning the STLSTM performs better.
The F1 score of STLSTM, which corresponds to the denominator, is approximately 0.7 but depends on which feature is made to be sparse.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=3.42cm]{rel_1.png}
\includegraphics[width=3.42cm]{rel_2.png}
\includegraphics[width=3.42cm]{rel_3.png}
\includegraphics[width=3.42cm]{urnn_churn.png}
\caption{Relative change in F1 score between STLSTM and TLSTM cell. For single sparse features, random sampling is first and group sampling second. Third is 2, 3, and 4 sparse features using group sampling. Fourth is relative performance on different sparse subsets of features for the churn data set.}
\label{fig:urnn_relative}
\end{center}
\end{figure}
A common characteristic for all series is that the relative performance of the STLSTM cell increases as the sparsity increases, but only to a certain point where it then maintains or slightly drops.
Further, STLSTM offers a larger improvement in the case of the group sampling in most cases.
As can be seen in Figure \ref{fig:urnn_relative}, the relative performance of each model depends on the feature we are treating as sparse.
It is not surprising that the feature with the largest dependence on sparsity is the voltage variable itself (represented by blue), since this is the feature we are making predictions on.
The features with the smallest relative performance change are found to have the smallest response to sparsity as well, which may indicate that they have less predictive power and thus it is not surprising that the difference between the TLSTM and STLSTM is also small.
Further than looking at individual sparse features, we also investigate groups of sparse features with different levels of sparsity.
We report the results on groups of sparse features with size 2, 3, and 4 since it was shown that 3 of the features are not affected by sparsity.
Figure \ref{fig:urnn_relative} also shows the relative performance for STLSTM and TLSTM models for groups of sparse features that were found to have the largest effect at a given group size and sparsity.
Unsurprisingly, the characteristics of the series are the same as in the analysis of single features
Using more sparse features does increase the relative performance of the STLSTM cell to a point, but for this dataset where the features are closely related to one another the effectiveness is limited.
In addition to focusing on the relative performance between the TLSTM and STLSTM cells, we also consider the performance of different aggregation methods $\mathcal{L}$ within the STLSTM cell.
In particular, we compare using a dense layer, an averaging aggregation, and a maximum aggregation at four different levels of sparsity.
The aggregation method clearly does not matter if there is only one sparse feature, and so for this comparison we must use sets of sparse features.
In particular, we report results using the set of 3 sparse features that was found to give the largest relative improvement for STLSTM.
The results are summarized in Table \ref{tab:urnn_agg} based on group sampling datasets.
As can be seen in Table \ref{tab:urnn_agg}, the dense layer aggregation performs as well or better than the other two methods at every level of sparsity.
The average and maximum aggregation have similar performance, perhaps slightly leaning toward the average method at low sparsity and the maximum method for higher sparsity.
\begin{table}[htp]
\caption{Aggregation and static feature analysis.}
\fontsize{10}{12}\selectfont
\begin{subtable}[t]{.45\columnwidth}
\caption{F1 scores for STLSTM model with different aggregation methods for different levels of sparsity.}
\label{tab:urnn_agg}
\raggedright
\resizebox{\columnwidth}{1cm}{
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Sparsity & Dense layer & Average & Max \\
\hline
0.03 & \bf{0.668} & 0.658 & 0.659 \\
\hline
0.07 & \bf{0.675} & 0.664 & 0.666 \\
\hline
0.11 & \bf{0.683} & 0.674 & 0.672 \\
\hline
0.15 & \bf{0.693} & 0.685 & 0.681 \\
\hline
\end{tabular}
}
\end{subtable}
\quad
\begin{subtable}[t]{.45\columnwidth}
\caption{Averaged relative F1 score change after incorporating one or both types of static features on group sampling datasets.}
\label{tab:urnn_static_table}
\raggedleft
\resizebox{\columnwidth}{1cm}{
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Sparsity & Static dense & Static delta & Both \\
\hline
0.03 & 1.56 & 0.67 & \bf{2.01} \\
\hline
0.07 & 1.48 & 0.58 & \bf{1.92} \\
\hline
0.11 & 1.45 & 0.51 & \bf{1.76} \\
\hline
0.15 & 1.44 & 0.48 & \bf{1.69} \\
\hline
\end{tabular}
}
\end{subtable}
\end{table}%
Finally, we study the effect of adding both types of static features to the model.
Table \ref{tab:urnn_static_table} shows the average relative improvements in F1 score for adding in one or both types of static features for different subsets of sparse features with varying levels of sparsity.
It is clear from Table \ref{tab:urnn_static_table} that both types of features improve the performance of the model, with the static dense features having a larger effect.
This larger effect may be specific to the features of the power consumption data, but generally both types of features improve performance when properly integrated into the model.
With respect to sparsity, both types of static features improve the model more as the features become more sparse, with the static delta features being more dependent on sparsity.
It is worth discussing the impact of the added complexity of the STLSTM cell on training time.
With STLSTM, as the number of features grows, so too does the number of parameters in the model.
This makes training slower per epoch, but it is mitigated in wall-clock training time by the fact that it takes fewer epochs to reach convergence.
Using standard TLSTM, an epoch of training time takes about 10 seconds with the given sequence parameters for the datasets.
When using four sparse features, this increases to about 18 seconds per epoch.
However, the STLSTM cell generally reaches convergence in 40-50 epochs while the TLSTM cell requires 70-80 epochs.
\subsection{Real world churn data set}
In addition to this constructed dataset, we also use a real world proprietary dataset with naturally sparse features for churn prediction within a future time span.
At each time step, there is an event embedding that is treated as a dense feature.
In addition to this, there are approximately 60 additional sequence state features.
These state features have varying levels of sparse ratios, ranging from 0.03 to 0.6.
We experiment with different subsets of dense and sparse features using the sequence state features. Additionally, we use two delta features, one static dense feature, and one static delta feature.
For this dataset we use a training set with 700,000 samples and validation/test sets with 200,000 samples each.
Sequence length ranges from 1 to 150, but the samples are not evenly distributed by sequence length. The distribution is exponential, with a large portion of sequences on the lowest end of that range and the mean is approximately 40.
For this data set, we compare two stacked architectures composed of three stacked RNN cells.
As before, the bottom layer uses either STLSTM or TLSTM, and the higher levels use standard LSTM cells.
When multiple sparse features are present, we use a dense layer as the aggregation method since it performs best.
All results are given relative to the standard TLSTM model with all 60 additional sequence state features treated as dense features, as this was found to outperform an all LSTM model by approximately 2\% under the all dense setting.
Since this data set is naturally sparse, we do not tune the sparsity, but instead investigate the relative performance for different subsets of sparse and dense features.
There are four qualitative feature subsets we primarily study.
The first sparse subset contains features that relate to relative changes, e.g. a customer can upgrade/downgrade his service to a different tier.
The second sparse subset includes true counts for feature values, e.g. the total number of people associated with an account.
These two subsets are the most sparse and have average sparsity of 0.065.
The other two feature subsets contain features relating customer specific (e.g. a customer relocates) and sequence specific information (e.g. the number of emails to customer service).
The relative performance for these sparse feature subsets is found in Figure \ref{fig:urnn_relative}.
We observe that using only relative change features gives the smallest performance gain.
This is not surprising because in this case a repeated value of $0$ has meaning and the features are specifically designed to capture changes, somewhat mitigating the intended effect of the STLSTM cell.
Using the raw counts does give a larger performance increase.
Both customer and sequence specific subsets show the largest improvement, indicating that these sub groups are responsive to STLSTM architecture.
It is interesting that these two subgroups do not necessarily contain only the most sparse features, but a mix between sparsity ratios roughly from 0.03 to 0.3.
This suggests that in real applications with naturally sparse features there is a large range of sparsity that can occur in the sparse sub group of features.
\subsection{Future work}
For future work, it would be interesting to incorporate even more feature types than the five covered in this work.
One in particular is a feature type that gives time information looking forward in the sequence.
All features in this work use time information related to past events, but there are cases that can benefit from the utility of incorporating future knowledge when available.
One example of this is the time to the prediction from the current time step so the network can have direct knowledge of its absolute time location in the sequence.
|
\section{ Introduction}
In this paper we generalize part of the higher-degree smoothness results in perturbation theory in \cite {3} from the case that the stable maps have the fixed domain $S^2$ to the general genus zero case. Note that genus zero case already captures all analytic difficulties related to the lack of differentiability of transition functions between local slices (see \cite {6} for the discussion on this ).
The results in this paper and \cite {4} together give one of the two methods for the infinite dimensional set-up used in \cite {5}
(compare the other infinite dimensional set-up in \cite{1}).
The main result of this paper is the following theorem (see the relevant definitions in the later sections).
\begin{theorem}
Let $K\simeq K_t$ with $t\in { \bar W}(\Sigma)$ be the fixed part of the local universal family of stable curves ${\cal S}\rightarrow N(\Sigma)$, where
$N(\Sigma)$ is a small open neighborhood of $[{\Sigma}]$ in ${\overline {\cal M}}_{0, k}$ with the local coordinate chart ${\bar W}(\Sigma)$. Consider the local uniformizer (slice)
${ W}(f, {\bf H}_f)$
centered at $f:\Sigma\rightarrow M$ of stable $L_k^p$-maps with domains ${\cal S}_t, t\in {\bar W}(\Sigma)$ and the corresponding space ${\widetilde W}(f_{K})$ of $L_k^p$ maps with domain $K$ with associated bundle ${\cal L}^{K}\rightarrow {\widetilde W}(f_{K})$.
Let ${\xi}^{K}:{\widetilde W}(f_{K})\rightarrow {\cal L}^{K}$ be a smooth section satisfying the condition $C_1$ and $C_2$. Then ${\xi}^{K}$ gives rise a stratified $C^{m_0}$-smooth section $\xi: { W}(f, {\bf H}_f)\rightarrow {\cal L}$ on the local slice ${ W}(f, {\bf H}_f)$, which, viewed in any other local slice,
is still stratified $C^{m_0}$-smooth on their "common intersections
" (=the fiber product over the space of unparametrized stable maps).
\end{theorem}
Here the conditions $C_1$ and $C_2$ are defined in \cite {3} using the bi-grading there
(see the definition in \cite {3} ) as follows.
$C_1:$ The section $\eta:S_f\rightarrow {\cal L}|_{S_f}$ can be extended into a $C^{m_0}$-smooth section
$\eta_{-m}:(S_f)_{-m}\rightarrow {\cal L}_0$ for some $m\geq m_0$
$C_2:$ The image of $\eta(h)$ is lying in $L_{k+m}^p(\Sigma, E_h)=:({\cal L}_h)_{m}$ with $m\geq m_0$ so that $\eta_{-m, m}:(S_f)_{-m}\rightarrow {\cal L}_m$ is smooth.
Recall that $m_0=[k-2/p]$. We will assume that $p>2$ and $m_0>1$ throughout this paper as in \cite {3}.
This theorem is proved in Sec.5. The two kinds of Banach neighborhoods on an end
near a stable nodal map are defined in section 3. The corresponding (stratified ) smooth structures on each of such neighborhoods are defined in section 4.
Only elementary facts on Sobolev spaces and standard calculus on Banach spaces are used in this paper, for which we refer to \cite{2, 7}.
\section {Local universal family of stable curves}
The starting point of this paper is the local deformation theory of stable maps. To this end, we
need recall the local deformation of the stable curves first.
\subsection{Stable curves}
Given an "initial" stable curve $\Sigma^0$, let $T_0$ be the tree associated to the domain of the stable curve $\Sigma^0=(S^0, {\bf d}^0, {\bf x}^0)$ so that the underlying curve $S^0$ is a nodal surface with desingularization ${\hat S^0}=\coprod_{v\in T_0}S^0_v$ as the disjoint union of its components labeled by the vertices $v\in T_0$.
Here the double points ${\bf d}^0=\cup_{v\in T_0}{\bf d}^0_v$ with each ${\bf d}^0_v=\{ d^0_{vw}, [vw]\in E(T_0)\}$, where each double point $ d^0_{vw}$ on $S^0_v$ is labeled by an edge in the set of edges $E(T_0)$ of $T_0$; the marked points ${\bf x}^0=\cup_{v\in T_0}{\bf x}^0_v$ with ${\bf x}^0_v$ to be the marked points on $S^0_v$. Clearly the nodal surface $S^0$ is obtained form ${\hat S^0}$ by identifying the double points.
In above, we have abused notations using ${\bf d}$ to denote both double points and the set of their collections. Similarly for ${\bf x}$, and we will continue do so for other similar notations.
The distinguished points (=the double points and marked points) of $\Sigma^0$ on $S^0_v$/$S^0$, will be denoted by ${\bf p}^0_v$/${\bf p}^0$. Note that the stable curve $\Sigma^0$ determines and is determined by
${\bf p}^0$ upto the actions of $G(=\Pi_{v\in T_0}SL(2, {\bf C})_v)$.
Let $N^{T_0}(\Sigma^0)$ be a small neighborhood of $[\Sigma^0]$ in the moduli space ${\cal M}^{T_0}_{0, k}$ with fixed topological type given by $T_0$, where $[\Sigma^0]\in {\cal M}^{T_0}_{0, k}$ is the "moduli point" of $\Sigma^0$. Then $N^{T_0}(\Sigma^0)$ parametrizes
the stable curves near $\Sigma_0$, or equivalently the nearby distinguished points ${\bf p}$ on the same fixed ${\hat S^0}$. Hence we may introduce the parameters
$b=\{b_{vu}; v\in T_0, p^0_{vu}\in {\bf p}^0_v\}$ with $b_{vu}$ in a small disc $D_{vu}({ p}^0_{vu})$ on $S^0_v$ centered at ${ p}^0_{vu}$. In order to quotient out the (local) actions of $SL(2, {\bf C})_v$, for each $v\in T_0$, we fix the last three parameters in $b_{vu}$. Note this also selects the three corresponding distinguished pints on $S_v$ that makes it {\bf marked} so that the identification $S_v\simeq S^2$ is specified.
Then the parameter
$b$ is corresponding to the stable curve $\Sigma_b=(S_b, {\bf p}_b)$. The collection of such parameters $b$ will be denoted by $W^{T_0}(\Sigma^0)$, considered as one of the natural holomorphic coordinate charts of $N^{T_0}(\Sigma^0)$. Of course, different choices of fixing three elements in each set $b_v, v\in T_0$ above give other but
same kind of coordinate charts of $N^{T_0}(\Sigma^0)$.
In this notation, the initial surface, $\Sigma^0=\Sigma_0$ or $\Sigma_{b} $ with $b=0.$ In the following, the notations $\Sigma^0$ are used interchangeably with $ \Sigma_0$ so that $N^{T_0}(\Sigma^0)$, $W^{T_0}(\Sigma^0)$ etc. will also be denoted by
$N^{T_0}(\Sigma_0)$, $W^{T_0}(\Sigma_0)$ accordingly. Similarly
${\bf p}^0_v$/${\bf p}^0={\bf p}_{0;v}/{\bf p}_0$.
Note that the desingularization ${{\hat S_b}}$ is same as ${{\hat S^0}}=
{\hat S_0}$ so that
$S_b$ has the same components as $S=S_0$ has. As before, $ {\bf p}_b=\cup_{v\in T_0}{\bf p}_{b;v}$, and ${\bf p}_{b;v}={\bf x}_{b;v}\cup
{\bf d}_{b;v}$ lying on $S^0_v$.
Let ${\bar N}^{T_1}(\Sigma_0)$ be a small (full) neighborhood of $\Sigma_0$ in the moduli space ${\overline {\cal M}}^{T_1}_{0, k}$, and ${ N}^{T_1}(\Sigma_0)$ be its top stratum as an open set in stratum ${ {\cal M}}^{T_1}_{0, k}$ with fixed topological type given by $T_1$. Then ${\bar N}^{T_1}(\Sigma_0)$ parametrizes
the stable curves $\Sigma_t$ near $\Sigma_0$ whose topological types are "bewteen $T_0$ and $T_1$". In particular,
when $T_1$ is the top stratum of ${\cal M}_{0, k}$, ${\bar N}^{T_1}(\Sigma_0)$ parametrizes
the stable curves $\Sigma_t$ of all types near $\Sigma_0. $ Here $t\in
{\bar W}^{T_1}(\Sigma_0)$ where $
{\bar W}^{T_1}(\Sigma_0)$ is one of the natural coordinate charts of ${\bar N}^{T_1}(\Sigma_0)$ extending $
{ W}^{T_0}(\Sigma_0)$.
Thus each parameter $t\in {\bar N}^{T_1}(\Sigma_0)$ has the form $t=(b, a)$ with $b\in { W}^{T_0}(\Sigma_0)$. Here $a=\{a_{vw}; [vw]\in E(T_0)\}$ is the collection of the gluing parameters describes the gluing pattern from $S_b$ to the glued surface $S_t$ defined below. The non-zero entries of $a$, denoted by
$a^e=\{a_{vw}; [vw]\in E(T_0), C(v, w)=u\in T_{1}\}$ are the effective parameters $a_{vw}$ gluing the components $S_v$ and $S_w$.
Here
the map $C:T_0\times T_0\rightarrow T_{1}$ is partially defined on $ T_0\times T_0$, and for each $(v, w)$ with
$(v, w)\in E(T_0)$ it is defined by $C(v, w)=u\in T_{1}$ if the component
$S_{t; u}$ of $S_t$ is obtained from the components $S_{b;v}$ and $S_{b;w}$.
Recall the definition of $\Sigma_t=(S_t; {\bf p}_t)$ with ${\bf p}_t={\bf d}_{t}\cup {\bf x}_{t}$ as follows.
(1) $S_{t}$ is obtained from $S_b$ by gluing at those double points $d_{vw}$ with
$a_{vw}\not=0$. Hence a component
$S_{t; u}=\#_{\{a_u=\{a_{v_i, v_j}\};C(v_i, v_j)=u\in T_1\}} (S_{v_1}, \cdots S_{v_{k(u)}})$. Here the right-hand side above is the gluing of the components
$S_{v_1}, \cdots S_{v_{k(u)}}$ in $\Sigma_b$ with the gluing parameter $a_u$.
For each $a_{v_i, v_j}\not =0,$ denote $(v_i, v_j)$ by $(u_+, v_{-})$ and
$a_{v_i, v_j}$ by $a$ temporarily. Let $D_{\pm}$ be the small discs on $S_{\pm}\simeq S^2$ with complex coordinate $w_{\pm}$. Then the gluing $\#_{a}(D_{-}, D_+)=D_{-}\coprod D_+$ quotient out the relation that $w_{-}\cdot w_+=a.$ Applying this to each nonzero $a_{v_i, v_j}$ above gives the desired gluing.
(2) The double points on
$\Sigma_t=\Sigma_{b,a}$ are exactly the part of the double points $ {\bf d}_b=\{{\bf d }_{b;uu'}\}$ such that
$a_{uu'}=0.$ Since we only consider the local deformations, we may assume that
$|a|$ is sufficient small so that the marked points ${\bf x}_b$ become the corresponding ones, denoted by ${\bf x}_t$ through the gluing.
\medskip
\noindent
${\bullet }$ ${\bullet }$ ${\bullet }$ Fixed part $K_t\simeq K_0$ in $S_t$.
The "fixed part" $K_{\epsilon, t,u}$ of $S_{t;u}$ defined by $$K_{\epsilon,t;u}=S_u\setminus \{\cup_{a_{v_i, v_j}\not = 0}N_{\epsilon}(b, a_{v_i, v_j})\cup_{a_{v_i, v_j}= 0}D_{\epsilon}( d_{b:v_i, v_j})\}.$$ Here $N_{\epsilon}(b, a_{v_i, v_j})$ is the "neck" part near the double point $d_{b;v_i, v_j}$ obtained by gluing the two corresponding small discs $D_{\epsilon}( d_{b;v_i, v_j})$ and $D_{\epsilon}( d_{b;v_j, v_i})$ with gluing parameter
$a_{v_i, v_j}\not = 0$. Thus $K_{\epsilon, t,u} \subset S_{t;u}$ becomes a {\bf fixed } subset of $S_{b;u}$ independent of $a$ with $t=(b, a)$.
To get rid of the $b$-dependency of the fixed part, go back to the desingularization ${\hat S}$ of initial underlying curve $S(=S^0)$ of $\Sigma^0(=\Sigma_0)$. For each double point $d^0_{vu}$ on the component $S_v$, choose another small disc $D_{\epsilon_2}(d_{vu})$ of radius $\epsilon_2>\epsilon$ such that $D_{\epsilon}(d_{vu}(b))\subset D_{\epsilon_2}(d_{vu})$ for all $b\in W^{T_0}(\Sigma_0)$. Then define the ( "smallest") fixed part $K$ to be the complement of the union of all the discs $D_{\epsilon_2}(d_{vu})$ in $S$. Then first of all, $K$ can be considered as a subset in $S_b$, denote by $K_b$, since both $S$ and $S_b$ have the same components; secondly by the construction of the gluing, it can also be considered as a subset of $S_t$, denoted by $K_t$. Note that for $|a|$ small enough, the marked points ${\bf x}_t$ are lying on $K_{t}$.
Thus for $|a|$ small enough with the types between $T_0$ and $T_1$, $K_{t}\simeq K_{b}
\simeq K_{0}\subset S$ as the fixed part independent of $t$ while $\{S_{t}\}$ is a family of curves that are deforming.
\subsection{Local universal family of the first kind}
It is well-known that the total family obtained from the gluing construction above, ${\cal S}=:{\cal S}(\Sigma_0)\rightarrow {\bar N}^{T_1}(\Sigma_0)$ with the fiber ${\cal S}_t=S_{t}$ is a proper morphism of complex manifolds/orbifords.
\begin{lemma}
Given $T_1\geq T_0$, let $t_0=(b_0, a_0)$ be the center of $W^{T_1}(\Sigma_{t_0}).$ Then there is a smooth but non-holomorphic family of identifications
$\{\lambda^{t_0}_t:(S_t, {\bf x}_t)\rightarrow (S_{t_0}, {\bf x}_{t_0})\}$ for $t\in W^{T_1}(\Sigma_{t_0})$. Under this identification, the smallest fixed part
$K_t$ is identified with $K_{t_0}$ so that $K_{t}\simeq K_{t_0}\simeq K_b\simeq K_{b_0}\simeq K_0$, the small disks or "neck" areas at or near double points on $S_t$ are identifies with the correspoding ones on $S_{t_0}$. Away from the small annuli of the tubular neighborhoods of the boundaries of $K_t$, $\lambda^{t_0}_t$ is holomorphic and preserves any of the natural metrics.
Moreover, these maps together give rise a smooth map
$\lambda^{t_0}:{\cal S}|_{N^{T_1}(\Sigma_{t_0})}\rightarrow {\cal S}_{t_0}=S_{t_0}$ and hence the induced smooth the map ${\hat \lambda^{t_0}}=( \lambda^{t_0}, \pi):{\cal S}|_{N^{T_1}(\Sigma_{t_0})}\rightarrow {\cal S}_{t_0}\times N^{T_1}(\Sigma_{t_0})$. Here $\pi:{\cal S}|_{N^{T_1}(\Sigma_{t_0})}\rightarrow N^{T_1}(\Sigma_{t_0})$ is the projection map.
\end{lemma}
The proof of this lemma is the immediate consequence of the construction of these diffeomorphisms below.
The map $\lambda^{t_0}$ above will be used to define the smooth structure of the {\bf first kind} on the corresponding neighborhood of the {\bf first kind} of a stable map in Sec. 4. Thus $W^{T_1}(\Sigma_{t_0})$ together with the map ${\hat \lambda^{t_0}}$ will be refer to as a local model of the {\bf first kind} for the local universal family of stable curves.
It is the precise version of the intuitive notion that the parameter $t\in W^{T_1}(\Sigma_0)$ is considered as a point near the ends of the stratum $T_1$ representing
the stable curve $\Sigma_{t}=(S_t; {\bf d}_t, {\bf x}_t )$ whose underlying surface $S_t$ is deforming and degenerating along the ends, while its fixed part
$K_t$ remains fixed such that the relative locations of ${\bf x}_t$ in $K_t$ are the same as the ones of ${\bf x}$ in $ K_0=K$. Note that in the case used in \cite {lt}, the initial curve $\Sigma_0$ is "minimally'' stabilized. In this case, ${\bf x}_t$ is indeed {\bf fixed } in the above model.
Now we define the required diffeomorphisms.
In the case that $T_1=T_0$, the lowest stratum with $t_0=(b_0, 0)\in W^{T_0}(\Sigma_{t_0})$, $\lambda^{t_0}$ is just $\lambda^{b_0}=:\{\lambda^{b_0}_{b, v}, v\in T_0, b\in W^{T_0}(\Sigma_0)\}$
defined as follows.
It is more convenient to define the inverse map of $\lambda^{b_0}_{b, v}$. For $v\in T_0, b\in W^{T_0}(\Sigma_0)$, $(\lambda^{b_0}_{b, v})^{-1}:(S_{b_0,v}, {\bf d}_{b_0, v} )\rightarrow (S_{b,v}, {\bf d}_{b, v} ) $ is defined by
the following conditions:
(i) It is the "identity" map on the complement of the union of all disks of radius $\epsilon_1$ centered
at the double points of the component $S_{b_0,v}$, denoted by $K_{b_0,v,\epsilon_1}$ under the identifications of
${\hat S}_{b}\simeq {\hat S}_{b_0}\simeq {\hat S}_{b=0}$ for $|b|$ and $|b_0|$ sufficiently small. In fact under above identifications, we get the corresponding
identification $K_{b,v,\epsilon_1}\simeq K_{b_0,v,\epsilon_1}$ of the {\bf fixed } parts given by $\lambda^{b_0}_{b, v}$.
(ii) On the disks $D_{\epsilon}(d_{b_0;vu})$ centered at the double point $d_{b;vu}$,
It is the translation that brings $d_{b_0;vu}$ to $d_{b;vu}$, and $D_{\epsilon}(d_{b_0;vu})$ to $D_{\epsilon}(d_{b;vu})$.
(iii) Under the above identification ${\hat S}_{b}\simeq {\hat S}_{b_0}$, the image $D_{\epsilon}(d_{b;vu})$ of
$D_{\epsilon}(d_{b_0;vu})$ can be considered as a subset of $D_{\epsilon_0}(d_{b_0;vu})\subset D_{\epsilon_1}(d_{b_0;vu})$ in $S_{b_0, v}$. Here of course, we assume that $\epsilon<\epsilon_0<\epsilon_1$ and $|b|$ and $|b_0|$ sufficiently small. Hence by (i) and (ii) above, using
the above identification again, the definition for the rest of $(\lambda^{b_0}_{b, v})^{-1}$ is reduced to find a self diffeomorphism of $D_{\epsilon_1}(d_{b_0;vu})$ that is the identity map near the boundary extending
the map already defined on $D_{\epsilon}(d_{b_0;vu})$ in (ii). This can be done
by extending the corresponding vector field. Then the desired diffeomorphism is the time-$1$ map of the flow of the extended vector field.
Note that the restriction map $$\lambda^{b_0}_{b, v}: \cup_{[vu]\in E(T_0)}D_{\epsilon}(d_{b;vu})\cup K_{b,v,\epsilon_1}\rightarrow \cup_{[vu]\in E(T_0)}D_{\epsilon}(d_{b_0;vu})\cup K_{b_0,v,\epsilon_1}$$ is holomorphic and preserves the "natural" metrics (Spheric, cylindrical or 'flat' ones).
To move to the higher stratum $T_1$, recall that for given $t=(b,a)\in W^{T_1}(\Sigma_0)$ with $t$ close $t_0=(b_0, a_0)$, at a
double point $d_{b, vu}=d_{b, uv}$ where the gluing parameter $a_{vu}=a_{uv}\not = 0$, the gluing of the pair of disks of radius $\epsilon $ on the components $S_{b, v}$ and $S_{b, u}$ at the double points, $D_{\epsilon}(d_{b, vu})\#_{a_{vu}=a_{uv}}D_{\epsilon}(d_{b, uv})$ was defined in this section.
The identification of the pair of disks $D_{\epsilon}(d_{b, vu})$ and $ D_{\epsilon}(d_{b, uv})$ with
$D_{\epsilon}(d_{b_0, vu})$ and $ D_{\epsilon}(d_{b_0, uv})$
by $\lambda^{b_0}_{b, v}$ and $\lambda^{b_0}_{b, u}$ induces
the corresponding identification $$D_{\epsilon}(d_{b, vu})\#_{a_{vu}=a_{uv}}D_{\epsilon}(d_{b, uv})\simeq D_{\epsilon}(d_{b_0, vu})\#_{a_{vu}=a_{uv}}D_{\epsilon}(d_{b_0, uv}).$$ Applying this to each double point with $a_{uv}\not = 0, $
We get a family of identifications $S_{b_0,a}\simeq S_{b, a}$ smooth in $b$.
Thus the construction of $ \lambda^{t_0}_{t}$ with $t_0=(b_0, a_0)$, $t=(b,a)$ and $|t_0|$ and $|t|$ small can be obtained by using the following two families of identifications of finite cylinders.
(A) When $l$ is close to $l_0$, there is a family of identifications
$[-l_0, l_0]\times S^1\rightarrow [-l, l]\times S^1$ smooth in $l$ and induced from the corresponding identifications $[-l_0, l_0]\rightarrow [-l, l]$.
We may assume that the restriction identifications to $[-l_0=1, l_0-1]\times S^1$
is the identity map.
Applying these identifications, we get a family of identifications
$S_{b_0,a'}\simeq S_{b, a}$ with $|a'|=|a_0|$ and $arg \, a'=arg \, a$.
(B) When $\theta\in S^1$ close to $1$, there is a smooth $\theta$-dependent family of identifications $[-l_0, l_0]\times S^1\rightarrow [-l_0, l_0]\times S^1$ that is the identity map on $[-l_0, l_0-1]\times S^1$ and is the rotation of angle $\theta$ on $\{l_0\}\times S^1.$ The effect of these identifications is to untwist the angular twisting in the gluing construction.
Applying this to the identifications obtained above so far, we finally get
the desire family of identifications $\lambda^{t_0}=\{\lambda^{t_0}_{ t}:S_t\rightarrow S_{t_0}\}$.
Let $\epsilon_2>\epsilon_1$ and assume that ${ D}_{\epsilon _1}({\bf d}_{b}\subset { D}_{\epsilon _2}({\bf d}_{b_0})$. Then define the ( "smallest")
fixed part $K=:K_{\epsilon_2}$ to be the complement of the union of all disks
centered at double points of radius $\epsilon_2$ on $S_0$. Then we have the fixed parts
$K_{t}\simeq K_{t_0}\simeq K_b\simeq K_{b_0}\simeq K$ in the corresponding
surfaces, on which the maps $\lambda^{t_0}_{ t}$ above are the "identity" map.
Note that the marked points ${\bf x}_t$ are lying on $K_{t}$.
It follows from the construction above, the family of identification has the properties in the above lemma.
\subsection{Local universal family of the second kind}
An important property the restriction above local universal family ${\cal S}(\Sigma_0)$, a fixed stratum of type $T_1$ with
$T_1\geq T\geq T_0$, can be considered as a set of "moving" marked points ${\bf x}_t$ on a fixed reference surface ${\hat S}_{t_0}$.
Then the parameter $t\in \in W^{T_1}(\Sigma_{t_0})$, as part of the local coordinates of the moduli space ${\cal M}^{T_1}_{0,k}$ also describes the
local moduli of the ("moving") distinguished points ${\bf p}_t$ on the ({\bf fixed }) reference curve $ {\hat S}_{t_0}$ of type $T_1$.
The local universal family considered this way will be regarded as the second model.
Thus in the second model, we are in the exact the same situation as we were at the beginning of this section: to parametrize stable curves ( considered as moving
distinguished points on a fixed reference surface) in a fixed stratum $T$, but with the initial curve $\Sigma_{t_0}$ instead of $\Sigma_0$.
In particular in this model the metric on $ {\hat S}_{t_0}$ is fixed by the marking $ {\hat S}_{t_0}\simeq S^2$.
To distinguish the situation here with the one before, instead of using the corresponding coordinate charts of the form $W^{T}(\Sigma_{t_0})$ with the local parameter $b=b(t)\in W^{T}(\Sigma_{t_0})$ representing the moduli point of $\Sigma_t$, we will denote $b=b(t)$ by $u:=u(t)$, and accordingly ${\bf p}_{b(t)}$ by ${\bf p}_{u}=:{\bf p}_{u(t)}$,
$W^{T}(\Sigma_{t_0})$ by $ U^{T}(\Sigma_{u_0})$ with $u_0=u(t_0)$, etc.
We state this more formally as following: the two models of local universal family over the open set $N^{T}(\Sigma_{t_0})$ of ${\cal M}^{T}_{0, k}$ centered at $\Sigma_{t_0}=\Sigma_{u_0}$, denoted by
${\cal S}|_{W^{T}(\Sigma(t_0))}$ and $ {\cal S}|_{U^{T}(\Sigma(u_0))}$
are identified by a fiber-wise analytic map $\phi^{-1}=\{\phi^{-1}_t\}$ with $\phi^{-1}_{t_0}:(S_{t_0},{\bf p}_{t_0})\rightarrow (S_{u_0}; {\bf p}_{u_0})$ being the identification of the two (but the "same") central fibers.
\begin{lemma}
The identification $\phi^{-1}_{t_0}:(S_{t_0},{\bf p}_{t_0})\rightarrow (S_{u_0}; {\bf p}_{u_0})$ induces a family of component-wise biholomorphic identifications, denoted by $\phi^{-1}_{t}: (S_{t},{\bf p}_{t}))\rightarrow (S_{u}; {\bf p}_{u})$ with ${\hat S}_{u}$ being the fixed ${\hat S}_{u_0}$ such that the induced map $ {\underline \phi }^{-1}:W^{T}(\Sigma_{t_0})\rightarrow U^{T}(\Sigma_{u_0})$
on the coordinate charts given by $t\rightarrow u$
is a holomorphic identification.
Moreover these maps $\{\phi^{-1}_t\}$ fit together to form an analytic identification between the corresponding universal families, denoted by $\phi^{-1}:{\cal S}|_{W^{T}(\Sigma_{t_0})}\rightarrow {\cal S}|_{U^{T}(\Sigma_{u_0})}$ that are holomorphic on the desingularizations ${\hat{\cal S}}|_{W^{T}(\Sigma_{t_0})}$ and $ {\hat {\cal S}}|_{U^{T}(\Sigma _{u_0})}$.
\end{lemma}
The above lemma is essentially a tautology in the situation above.
However, it also follows from the local universal property of the family ${\cal S}$.
\section {Two types of neighborhoods near ends}
\subsection {"Base" deformations of a stable map $f=f_0:\Sigma_0\rightarrow M$}
\noindent
${\bullet }$ "Base" deformations of a initial stable map $f=f_0:\Sigma_0\rightarrow M$ within the same stratum of type $T_0$.
\medskip
There are two kinds of such "base" deformations of $f_0: \Sigma_0\rightarrow M$, denoted by $\{{\tilde f}_b:\Sigma_b\rightarrow M, b \in W^{T_0}(\Sigma_0)\}$ and $\{f_b:\Sigma_b\rightarrow M, b \in W^{T_0}(\Sigma_0) \}$ respectively.
The first one is simply defined to be ${ \tilde f}_{b}=:\cup_{v\in T_0}{ \tilde f}_{b;v}$ with ${ \tilde f}_{b;v}=f_v$ under the assumption that $f$ is constant on
all the small disks near double points.
The last identity makes sense as the domains of the two maps are the same $S^0_v$ after forgetting the distinguished points ${\bf p}_{b;v}$ and ${\bf p}_v^0$.
To define the second deformation,
let $\lambda_{b;v}^{b_0}:S_{b; v}\rightarrow S_v=:S_v^{b_0}$ in last section.
Then we define ${ f}_{b; v}:S_{b;v}\rightarrow M$ to be ${ f}_{b; v}=f_v\circ \lambda_{b;v}^{b_0}$
and ${ f}_b=\cup_{v\in T_0}{ f}_{b; v}:S_b\rightarrow M$.
Next we extend $\{f_b, b\in W^{T_0}(\Sigma_0)\}$ to a higher stratum $T_1>T_0$.
${\bf \bullet}$ "Base" deformation $\{f_t, t\in {\bar N}(\Sigma_0)\}.$
Given $f=f_{-}\bigvee f_{+};(D_{-}, d_{-})\bigvee_{d_{-}=d_{+}}(D_{+}, d_{+})\rightarrow M$ and a gluing parameter $a_0=exp\{-(s_0+t_0i)\}\not = 0$, to define the gluing $\#_{a_0}(f_{-}, f_{+}): \#_{a_0}(D_{-}, D_+)\rightarrow M$ below, we introduce the cylindrical coordinate $(s_{\pm}, t_{\pm})\in {\bf R}^{\pm}\times S^1 $ on $D_{\pm}$ by the
identification of $D_{\pm}\simeq {\bf R}^{\pm}\times S^1$, $D_{\pm}$. Then $\#_{a_0}(D_{-}, D_+)$ is obtained by cutting of the part of $D_{\pm}$ with $|s_{\pm}|>-log|a_0|$ and glue the rest
along the boundaries twisted with an angel $arg a_0.$ Thus $\#_{a_0}(D_{-}, D_+)\simeq [-log|a_0|, log|a_0|]\times S^1$ with the induced cylindrical coordinate $(s, t)$ with $s=0$ corresponding to the "middle circle" and the other two
cylindrical coordinates $(s_{\pm}, t_{\pm})$ with $s_{\pm}=0$ corresponding to the two boundary circles. Note that $s_{\pm}=s{\mp}log|a_0|.$
Then $\#_{a_0}(f_{-}, f_{+})$ is defined to be $\#_{a_0}(f_{-}, f_{+})(s,t)=exp_{f_{\pm}(d_{\pm})}(\beta_{-}(s){\hat f}_{-}(s, t)+(\beta_{+}(s){\hat f}_{+}(s, t)).$ Here ${\hat f}_{\pm}$ is a vector field over $D_{\pm}$ such that
$f_{\pm}=exp_{f_{\pm}(d_{\pm})} {\hat f}_{\pm}$, and $\beta_{\pm}$ are cut-off functions supported on $[-1,1]$ with $\beta_{-}+\beta_{+}=1.$
The deformation $f_t=(f_b)_a$ is then defined by implanting above construction to each double points $d_{vw}$ of $f_b$ with $a_{vw\not= 0.}.$
\subsection{Neighborhoods of the first kind }
Let $f_t, t\in {\bar W}(\Sigma_0)$ be the "base" deformation of $f_0$ constructed above over a small coordinate/uniformizer ${\bar W}(\Sigma_0)$ of ${\bar N}(\Sigma_0)$ centered at $\Sigma_0$.
A neighborhood $W^{\nu({\bf a})}_{\epsilon}(f_0, {\bf H}_{f_0})$ of $[f_0]$ in the space of unparametrized stable maps, as a slice, is defined to be a family of Banach manifolds over the "base" deformation $\{f_t\},$
$$W^{\nu({\bf a})}_{\epsilon}(f_0, {\bf H}_{f_0})=\cup_{t\in {\bar W}(\Sigma_0)}W_{\epsilon}^{ t; \nu(a)}(f_t, {\bf H}_{f_0}).$$
Here for each fixed $t$,
$$W_{\epsilon}^{\nu(a) ;t}(f_t, {\bf H}_{f_0})
$$
$$ =\{h_t:(S_t, {\bf x}_t)\rightarrow
(M, {\bf H}_{f_0})\,|\, \|h_t-f_t\|_{k, p; \nu(a)}<\epsilon\}.$$
Here $"a" $ is gluing parameter in $t=(b,a)$.
\medskip
\noindent
${\bf \bullet }$ The $ \nu(a)$-exponentially weighted norm
The norm $\|-\|_{k, p; \nu(a)}$ used in the definition above is the $\nu$-exponentially weighted norm
along the all "neck" areas $N(b, a_{v_i, v_j})$ of $S_t$ with $a_{v_i, v_j}\not = 0 $ obtained by gluing from $S_0$ at the double points $d_{v_i, v_j}(b)$; on the rest of $S_t$, the norm is just
the usual $L_k^p$-norm. More specifically, recall that "neck" areas of $S_t$ with $a_{v_i, v_j}\not = 0 $ and $t=(b, a)$, $N(b, a_{v_i, v_j})\simeq (-|log |a_{v_i, v_j}||, |log |a_{v_i, v_j}||)\times S^1$. The weight
function $\nu(a_{i, j})$ is equal to $exp\{\nu |s_{\pm}|\}$ for points in $N(b, a_{v_i, v_j})$ with
$s_{\pm}\in (0, |log |a_{v_i, v_j}|\,|-2)$ and is a smooth function equal to 1 on two ends of the neck. Outside these necks, $\nu(a)=1$ so that these weight $\nu(a_{i, j})$
functions together defines a smooth weight function $\nu(a)$.
Here $\nu$ is a fixed positive constant with $\nu< (p-2)/p.$
The $L^p_{k, \nu(a)}$-norm $\|h_t\|_{k, p, \nu(a)}$ then is defined to be
$\|\nu(a)\cdot h_t\|_{k, p}.$
For each fixed $t$, the norm so defined makes $W_{\epsilon}^{\nu(a) ;t}(f_t, {\bf H}_{f_0})
$ became a Banach manifold. However, on $W^{\nu({\bf a})}_{\epsilon}(f_0, {\bf H}_{f_0})$, this $t$ -dependent family of norms does not necessarily define
a topology without further conditions
as in [L6?]. The reason for this is that the mixed $L_{k}^p$/$L_{k, \nu(a)}^p$-norm used here is not continuous when $h_t$ is moving from higher stratum to the lower ones.
On the other hand, on each fixed stratum, near any given point the $L_{k}^p$ and $L_{k, \nu(a)}^p$ norms are "locally" equivalent so that the resulting space is at least a (topological ) Banach manifold.
More specifically, consider the decomposition of $W^{\nu({\bf a})}_{\epsilon}(f_0, {\bf H}_{f_0})$ into its open strata,
$$W^{\nu({\bf a})}_{\epsilon}(f_0, {\bf H}_{f_0})=: \cup_{T\geq T_0}W^{ \nu (a_T),T}_{\epsilon}(f_0, {\bf H}_{f_0})$$ with each $$W^{ \nu (a_T), T}_{\epsilon}(f_0, {\bf H}_{f_0})=:\cup_{t\in W^{T}_{\epsilon}(\Sigma_0)}W_{\epsilon}^{ \nu(a),t}(f_t, {\bf H}_{f_0}).$$ Then the collection of all such neighborhoods $W^{ \nu (a_T), T}_{\epsilon}(f, {\bf H}_{f_0})$ of a fixed stratum $T$ generate a topology by the above mentioned local equivalent equivalence of the $L_{k, \nu(a)}^p$-norm with standard $L_k^p$-norm.
Since in order to prove the main results of this paper on the higher smoothness of the admissible perturbations, we need to localize further by using small neighborhoods of
any given element $g_{t_0}$ in $W^{ \nu (a_T), T}_{\epsilon}(f, {\bf H}_{f_0})$. We need spell out more on the existence of such neighborhoods.
${\bullet}$ Neighborhoods of the first kind on $W^{\nu({\bf a})}_{\epsilon}(f_0, {\bf H}_{f_0})$:
Fix a stratum of type $T=T_1\geq T_0.$ Recall that $T_0$ is the tree associated with the lowest stratum that $f_0$ lies on.
Give $\{g_{t_0}:(S_{t_0}, {\bf x}_{t_0})\rightarrow (M, {\bf H}_{f_0})\}\in
W_{\epsilon}^{\nu (a_{T_{1}}), T_{1} }(f_0, {\bf H}_{f_0})$ with $t_0=(b_0, a_0)\in W^{T_1}(\Sigma_0)$ and $\Sigma_{t_0}=(S_{t_0}, {\bf p}_{t_0})$,
a neighborhood of $g_{t_0}$ of the {\bf first kind} in
$ W_{\epsilon}^{ \nu (a_{T_{1}}),T_{1}}(f_0, {\bf H}_{f_0})$, denoted by
$$W^{\nu(a_{T_{1}}), T_{1}}_{\epsilon '}(g_{t_0}, {\bf H}_{f_0})=:W^{\nu(a_{T_{1}}), T_{1}}_{\epsilon '}(g_{t_0}, {\bf H}_{f_0}; g_t)=\{h_t:S_t\rightarrow M|\,\, \|h_t-g_t\|_{k, p, {\nu(a)}}<\epsilon'\}$$ with ${\epsilon '}<<\epsilon ,$
can be defined in a few equivalent ways that we describe now.
Here $g_t:S_t\rightarrow M$ is the "base" deformation of $g_{t_0}$ inside $ W_{\epsilon}^{ \nu (a_{T_{1}}),T_{1}}(f_0, {\bf H}_{f_0})$,
similar to the initial deformation $f_t$. Since topological type of $S_t$ is fixed,
we require that
the deformation has a form
$g_t=g_{t_0}\circ T^{t_0}_t$, where $$T=:T^{t_0}=\{T_t^{t_0}\}: {\cal S }|_{W^{T_1}_{\epsilon'}(\Sigma_{t_0})}=\cup_{t\in {W^{T_1}_{\epsilon'}}(\Sigma_{t_0})} S_t \rightarrow S_{t_0}$$
is a smooth family of diffeomorphisms.
Thus we need establish the existence of the required deformation $g_t$.
The key step is the following lemma.
\begin{lemma}
Given the base deformation $\{f_t\}$ of $f_0$ defined earlier in this section, fix a member
$f_{t_0}:S_{t_0}\rightarrow M$ of type $T_1$ in the deformation, there is a smooth family of diffeomorphisms $T=:T^{t_0}=\{T_t^{t_0}\}: {\cal S }|_{W^{T_1}_{\epsilon'}(\Sigma_{t_0})}\rightarrow S_{t_0}$ such that $\lim_{t\rightarrow t_0}\|f_t-f_{t_0}\circ T^{t_0}_{t}\|_{k, p, \nu(a)}=0$ in an uniform manner in $t_0$ for $t_0$ varying in a compact set.
\end{lemma}
Above lemma implies the following two lemmas
\begin{lemma}
Given $g_{t_0}\in W_{\epsilon}^{ \nu (a_{T_{1}}),T_{1}}(f_0, {\bf H}_{f_0})$. There exists a "base" deformation $g_t$ inside $ W_{\epsilon}^{ \nu (a_{T_{1}}),T_{1}}(f_0, {\bf H}_{f_0})$ for $|t-t_0|<<\epsilon$ with
$g_t=g_{t_0}\circ T^{t_0}_t$.
\end{lemma}
\begin{lemma}
These $W^{\nu(a_{T_{1}}), T_{1}}_{\epsilon '}(g_{t_0}, {\bf H}_{f_0})$ generate a topology on $ W_{\epsilon}^{ \nu (a_{T_{1}}),T_{1}}(f_0, {\bf H}_{f_0})$.
\end{lemma}
We note that the required $T^{t_0}_t$ can be taken as the particular family of diffeomorphisms $\lambda^{t_0}_t$ defined before.
Since our main concerns of this paper is the stratified smoothness of generic perturbations, we will not give the proofs of above lemmas. They will be given in [L?].
The neighborhoods $W_{\epsilon'}^{\nu(a_{T_{1}}), T_{1}}(g_{t_0}, {\bf H}_{f_0})$ here will be call the ones of the {\bf first} kind.
Here are some variations or related constructions:
(1) In the definition of $W^{\nu(a_{T_{1}}), T_{1}}_{\epsilon '}(g_{t_0}, {\bf H}_{f_0})$, the conditions that $|t-t_0|< \epsilon'$ and $\|h_t-g_t\|_{k, p; \nu(a)}< \epsilon'$ can be replaced by $|t-t_0|+\|h_t-g_t\|_{k, p; \nu(a)}<\epsilon'.$
(2) In the above definition of the $L_{k, \nu(a)}^p$-norm, a $t$-dependent metric $m_t$ on the domain $S_t$.
is used. Using the diffeomorphisms $T_{t}^{t_0}:S_t\rightarrow S_{t_0}$
to pull-back the fixed metric $m_{t_0}$, we get a family of metrics $ (T_{t}^{t_0})^*(m_{t_0})$ on $S_t$ and the corresponding $L_{k, \nu(a)}^p$-norms and neighborhoods $W^{\nu(a_{T_{1}}), T_{1}}_{\epsilon '}(g_{t_0}, {\bf H}_{f_0})$. Since these families of the metrics are uniformly equivalent for
$|t-t_0|\leq \epsilon''$, the resulting neighborhoods defined this way are equivalent to the previous one.
Thus, upto the effect of $T_t^{t_0}$, we can use a fixed reference metric on $m_{t_0}$ to define the norm.
This implies that the second type of the neighborhoods defined below is (topologically) equivalent to the ones above.
(3) Consider the deformations
of $h_t=h_{t_0}\circ T_t^{t_0}$ for all $h_{t_0}:S_0\rightarrow M $ with $h_{t_0}$ in the central slice $ W^{\nu(a_{0}),t_0, T_{1}}_{\epsilon '}(g_{t_0}, {\bf H}_{f_0})$. Denote the collection of such $h_t$ with $|t-t_0|<\epsilon'_1$ by
$W'^{\nu(a_{T_{1}}), T_{1}}_{\epsilon '_1}(g_{t_0}, {\bf H}_{f_0})$.
\begin{lemma}
The neighborhoods $W'^{\nu(a_{T_{1}}), T_{1}}_{\epsilon '_1}(g_{t_0}, {\bf H}_{f_0})$
so defined are equivalent to the ones defined before.
\end{lemma}
For the proof of this lemma we refer to [L?] again.
\medskip
\subsection { Neighborhoods
of the second kind}
Recall that there is a family of
biholomorphic identifications $\phi^{-1}_{t}: (S_{t},{\bf p}_{t})\rightarrow (S_{u}; {\bf p}_{u})$ which transforms the family of varying curves $\{(S_{t},{\bf p}_t)\}$ with {\bf fixed} $(K_t,{\bf x}_t)=(K_0,{\bf x}_0)$ with
the curve with {\bf fixed} components but with a family of varying distinguished points, $\{({\hat S}_{u_0}, {\bf p}_u)\}$.
It induces a bijection $$\Phi= \cup_{t}\Phi_t: W_{\epsilon'}^{\nu(a_{T_1}), T_{1}}(g_{t_0}, {\bf H}_{f_0})=:\cup_{t\in W^{T_1}(\Sigma_{t_0})} W_{\epsilon'}^{\nu(a), t}(g_{t_0}, {\bf H}_{f_0}) $$ $$ \rightarrow U_{\epsilon'}^{ T_{{1}}}(g_{u_0}, {\bf H}_{f_0})=\cup_{u\in U^{T_1}(\Sigma_{u_0})}
U^{u}_{\epsilon'}(g_{u_0}, {\bf H}_{f_0}) $$ by pull-backs given by $\Phi_t(h_t)=
h_{t}\circ \phi_{t}$ denoted by $h_u.$
The subspace $U^{u}_{\epsilon'}(g_{u_0}, {\bf H}_{f_0}) $ here consists those stable maps $h_u:(S_u, {\bf x}_u)\rightarrow
(M, {\bf H}_{f_0})$ with {\bf fixed} marked points ${\bf x}_u$ on the fixed domain ${\hat S_u}={\hat S_{u_0}}$ such that $\|h_u-g_u\|_{k, p}<{\epsilon'}$. Here we give
each component of ${ \hat S}_{u_0}$ the spherical metric and use it to defined the above $L_k^p$-norm, and $\{g_u=g_t\circ \phi_t\}$ is the transformed base family by $\{\phi_t\} $.
Since for $|t-t_0|<\epsilon'$, the norms of the above two spaces are equivalent, $\Phi$ is a homeomorphism to its image.
The neighborhoods $U_{\epsilon'}^{ T_{{1}}}(g_{u_0}, {\bf H}_{f_0})$ will be refereed as of the {\bf second} kind.
By the remark (2) above, the neighborhoods here are equivalent to the ones of the first kind above.
Note that the lowest stratum, $W^{\nu(a_{T_0}, T_0 )}_{\epsilon}(f_0, {\bf H}_{f_0})$
is the same as $U^{T_0}_{\epsilon}(f_{u=0}, {\bf H}_{f_0})$ since the norm used here is just the usual $L_k^p$-norm without exponential weight.
It is easy to see that
$\Phi_t$ is a diffeomorphism.
\begin{lemma}
For each fixed $t$ and hence $u$, $$\Phi_t: W_{\epsilon'}^{\nu(a), t}(g_{t_0}, {\bf H}_{f_0}) $$ $$ \rightarrow U^{u}_{\epsilon'}(g_{u_0}, {\bf H}_{f_0})
$$ is a diffeomorphism.
\end{lemma}
\proof
The only difference between $U_{\epsilon'}^{ T_1}(g_{u_0}, {\bf H}_{f_0})$ and $W_{\epsilon'}^{\nu(a_{T_1}), T_1}(g_{t_0}, {\bf H}_{f_0})$ is that the domains of stable maps in $W_{\epsilon'}^{\nu(a_{T_1}), T_1}(g_{t_0}, {\bf H}_{f_0})$ are varying depending on the parameter $t$ while the domains of elements in $U^{ T_{1}}(g({t_0}), {\bf H}_{f_0})$ are the fixed ${\hat S}_{u_0}$ but with moving distinguished points parametrized by $u=u(t)$. When $t$ is fixed, the elements
in $U_{\epsilon'}^{u}(g_{u_0}, {\bf H}_{f_0})$ or $W_{\epsilon'}^{\nu(a), t}(g_{t_0}, {\bf H}_{f_0})$ have the same domain with fixed components and distinguished points under the identification map $\phi_t$. Moreover the Banach norm on these two spaces are equivalent. Hence the induced map $\Phi_{t}$ is a diffeomorphism.
\hfill$\Box$\medskip
From now on, if there is no confusion, we will drop the subscript $\epsilon$ that describes the size of a neighborhood.
Next we define the neighborhoods that are still second type obtained from $U^{ T_{1}}(g_{u_0}, {\bf H}_{f_0})$ by
dropping some marked points.
Given $\{h:S_u\rightarrow M\} \in U^{ T_{1}}(g_{u_0}, {\bf H}_{f_0})$ with fixed ${\hat S_u}={\hat S}_{u_0}$, among the $k$ marked points ${\bf x}_u$ on ${\hat S}_{u_0}$ we select $m$ points, denoted by ${\bf x}^r_u$ such that $ (S_u,{\bf x}^r_u) $ is still stable. Denote the resulting stable curve with $m$ marked points
by $\Sigma^r_u=:(S^r_u, {\bf x}^r_u)$ with fixed ${\hat S}^r_u={\hat S}^r_{u_0}$, which is the same as ${\hat S}_{u_0}$ as a surface.
Then the map $ h^r:(S^r_u, {\bf x}^r_u)\rightarrow (M, {\bf H}^r_{f_0})$ is defined to be the same map $ h$ as before but forgetting the rest of the marked points, denoted by ${\bf x}^c_u$, in ${\bf x}_u$, where ${\bf H}^r_{f_0}$ is the corresponding
selection of local hypersurfaces. Note that $h^r_u({\bf x}^c_u)\in {\bf H}^c_{f_0}.$
Then the collection of all such $ h^r$ obtained from $h\in U^{ T_{1}}(g_{u_0}, {\bf H}_{f_0})$ by dropping $k-m$ marked points will be denoted by $U^{ T_{1}}(g^r_{u_0}, {\bf H}^r_{f_0})$ with the centered
$ g^r_{u_0}$.
The process above of course depends on the following choices: (i) the selection of ${\bf x}^r_u$, (ii) an order for ${\bf p}_u\subset {\hat S}_u={\hat S}_{u_0}$ that induces an order for ${\bf p}^r_u\subset S^r_u={\hat S^r}_{u_0}$.
In the following, we fix one of such choices, labeled by the superscript $r$ in the notations here. For each $v\in T_1$, by identifying the first three points in $({\bf p}^r_u)_v$ and $({\bf p}_u)_v$, we get the holomorphic identifications $\psi_v:({\hat S^r}_u)_v=({\hat S^r}_{u_0})_v\rightarrow ({\hat S}_u)_v=({\hat S}_{u_0})_v $ and
$\psi=\cup_{v \in T_1}\psi_{v}:{\hat S}^r_{u_0}\rightarrow {\hat S}_{u_0}, $ which induces the above drop-marking map, denoted by $$\Psi =\cup_{ v\in T_1}\Psi_v:U^{ T_{1}}(g_{u_0}, {\bf H}_{f_0})\rightarrow U^{ T_{1}}(g^r_{u_0}, {\bf H}^r_{f_0})$$ given by
pull-back by $\psi$,
$h\rightarrow h\circ \psi$ denoted by $h^r $.
In next section we will show that this map is a diffeomorphism.
Note that the identifications $\psi_v:({\hat S^r}_u)_v=({\hat S^r}_{u_0})_v\rightarrow ({\hat S}_u)_v=({\hat S}_{u_0})_v $
above also give canonical identifications of these surfaces with the fixed $(S^2,; 0, 1, \infty)$ so that they become "marked" surfaces.
\section{Stratified smooth structures on neighborhoods}
The smooth structures on $U^{ T_{1}}(g^r_{u_0}, {\bf H}^r_{f_0})$ and $U^{ T_{1}}(g_{u_0}, {\bf H}_{f_0})$ are defined similarly,
obtained as $C^{m_0}$ submanifolds $E_m^{-1}({\bf H}^r_{f_0})$ of ${\widetilde U}^{ T_{1}}(g_{u_0})\times {\hat S}_{u_0}^{m }$ and $E_k^{-1}({\bf H}_{f_0})$ of ${\widetilde U}^{ T_{1}}(g_{u_0})\times {\hat S}_{u_0}^{k }$ respectively. Here $E_l:{\widetilde U}^{ T_{1}}(g_{u_0})\times {\hat S}_{u_0}^{l }\rightarrow M^l$ is the $l$-fold total evaluation map at $l$ selected
marked points among the $k$ marked points on the domain $( {\hat S}_{u}, {\bf x}_u)=({\hat S}_{u_0}, {\bf x}_u)$.
Since $E_l$ is of class $C^{m_0}$ (see \cite {L8} ), it is easy to see that
it is a $C^{m_0}$-submersion so that above two subsets have $C^{m_0}$-smooth structures.
\begin{pro}
$\Psi :U^{ T_{1}}(g_{u_0}, {\bf H}_{f_0})\rightarrow U^{ T_{1}}(g^r_{u_0}, {\bf H}^r_{f_0})$ is a local diffeomorphism at $g_{u_0}$.
\end{pro}
\proof
Note that $\Psi :U^{ T_{1}}(g_{u_0}, {\bf H}_{f_0})\rightarrow U^{ T_{1}}(g^r_{u_0}, {\bf H}^r_{f_0})$ is a bijection.
Indeed since $g^r_{u_0}({\bf x}^c_{u_0})\in {\bf H}^c_{f_0}$ and for any $h\in U^{ T_{1}}(g^r_{u_0}, {\bf H}^r_{f_0})$, we already have $h({\bf x}^r_u)\in {\bf H}^r_{f_0}$, when $U^{ T_{1}}(g^r_{u_0}, {\bf H}^r_{f_0})$ is small enough and ${\bf x}^r_u$ as above is fixed, by implicit function theorem, the equation on ${\bf x}^c_u$, $h({\bf x}^r_u, {\bf x}^c_u)\in {\bf H}^c_{f_0}$ has an unique solution that is close to ${\bf x}^c_{u_0}$. This proves that $\Psi=:\Psi^r$ is a bijection with $\Psi^{-1} $ sending $\{h^r:({\hat S}^r_u={\hat S^r}_{u_0}, {\bf x}^r_u)\rightarrow M \}$ to $\{h:({\hat S}_u={\hat S}_{u_0}, {\bf x}^r_u, {\bf x}^c_u )\rightarrow M \}$.
Thus the map $\Psi$ is the restriction to a $C^{m_0}$ submanifold of the obvious smooth projection $\pi:{\widetilde U}^{ T_{1}}(g_{u_0})\times {\hat S}_{u_0}^{k}\rightarrow {\widetilde U}^{ T_{1}}(g_{u_0})\times {\hat S}_{u_0}^{m }$ sending
$k$ marked points to the corresponding $m$ marked points.
\hfill$\Box$\medskip
{\bf Note}: This process of dropping-adding marked points given by $\Psi$ is used in \cite {5} by requiring that domains of the elements in $W^{ T_1}(g^r_{u_0}, {\bf H}^r_{f_0})$ are stabilized with minimal number of marked points.
Next we defined a smooth structure centered at $g_{t_0}$
for $W^{\nu(a_{T_1}), T_1}(g_{t_0}, {\bf H}_{f_0}) $
or $W^{\nu(a_{T_1}), T_1}(g_{t_0}) $. This can be done by using the family of smooth
identifications $\lambda^{t_0}=\{\lambda^{t_0}_{ t}:\Sigma_t\rightarrow \Sigma_{t_0}, t\in W^{T_1}(\Sigma_{t_0})\}$ that is the identity map on the "small fixed part" $K_{t,\epsilon_2}({\bf d}(b_0))$ "centered" at ${\bf d}(b_0)$ defined before in Sec. 2. These maps give rise a smooth trivialization of the local universal family $({\cal S}|_{W^{T_1}(\Sigma_{t_0})}\rightarrow W^{T_1}(\Sigma_{t_0}))\simeq \Sigma_{t_0}\times W^{T_1}(\Sigma_{t_0})$.
Now the smooth structure on $W^{\nu(a_{T_1}), T_1}(g_{t_0}, {\bf H}_{f_0}) $ can be defined by the identification $\Lambda^{t_0}: W^{\nu(a_0), t_0}(g_{t_0}, {\bf H}_{f_0}) \times N^{T_1}(\Sigma_{t_0}) \rightarrow W^{\nu(a), T_1}(g_{t_0}, {\bf H}_{f_0})$ defined by $\Lambda^{t_0}(h, t)=h\circ \lambda^{t_0}_t.$ Note that the norms are equivalent under the map $\Lambda^{t_0}$ by the remark/note (3) in last section.
This gives a smooth structure on each $W^{\nu(a), T_1}(g_{t_0}, {\bf H}_{f_0}) .$ Of course transition functions between two such neighborhoods of type $T_1$
are only continuous. The end $W^{\nu(a_{T_1}), T_1}(f_{0}, {\bf H}_{f_0}) $ then is covered by such neighborhoods.
\section{Higher-degree stratified smoothness of \\ the perturbations }
In this section we give a proof that the ${\xi}$ on
$W^{\nu(a_0), T_1}(g_{t_0}, {\bf H}_{f_0})$ defined below is of class $C^{m_0}$ viewed in any other local slices.
We assume that ${\xi}$ is obtained from ${\xi}^K=\oplus_{v\in T_0}{\xi}^{K_v}$ defined below.
Here
$K=K_0=\cup_{v\in T_0}K_v\subset S_0$ is the fixed part lying on the initial curve $S_0$ with subscript "$0$" corresponding to $t=(a, b)=(0, 0)$.
For each $v\in T_0$, let ${\widetilde W}(f_{K_v})$ be the collection of $L_k^p$-maps $g=g_v:K_v\rightarrow M$ such that $\|g-f_{K_v}||_{k, p}<\epsilon$. Here $f_{K_v}=f|_{K_v} $, the restriction of the initial map $f$ to $K_v$. We give
$K_v$ the induced metric from $S_v$. The bundle $({\cal L}^{K_v}, {\widetilde W}(f_{K_v}))$ is defined as following: for any $h_v\in {\widetilde W}(f_{K_v})$, the fiber ${\cal L}^{K_v}|_{h_v}=(L_{k-1}^p)_0(K_v,h_v^*(E) )$ consists of all $L_{k-1}^p$-sections with compact support in the interior of $K_v$.
Let ${\widetilde W}(f_{K})=\prod_{v\in T_0}{\widetilde W}(f_{K_v})$ and ${\cal L}^{K}
=\oplus_{v\in T_0}{\cal L}^{K_v}$.
Now for each $v\in T_0$, fix a section ${\xi}^{K_v}:{\widetilde W}(f_{K_v})\rightarrow {\cal L}^{K_v}$ of class $C^{\infty}$ satisfying the conditions $C_1$ and $C_2$. Then the section ${\xi}^{K}=:\oplus_{v\in T_0}{\xi}^{K_v}:{\widetilde W}(f_{K})\rightarrow {\cal L}^{K}.$
By the identification of $K_t\simeq K=K_0$, for any fixed
$t\in W^{T_1}(\Sigma_{t_0})$, and $g_t\in W^{\nu(a), t}(g_{t_0}, {\bf H}_{f_0})\subset W^{\nu(a_{T_1}), T_1}(g_{t_0}, {\bf H}_{f_0}) $,
we get the induced section $\xi^t: W^{\nu(a), t}(g_{t_0}, {\bf H}_{f_0})\rightarrow {\cal L}^t$ defined by $\xi^t(g_t)=:\xi^K(g_t)|_{K_t}.$
It is easy to see that the standard local trivializations for
the bundle $({\cal L}^{K}\rightarrow {\widetilde W}(f_{K}))$ and $({\cal L}^{t}\rightarrow {\widetilde W}^{\nu(a), t}(g_{t_0}, {\bf H}_{f_0}))$ are compatible with respect to the above identifications
of $K_t\simeq K_0$ so that for each fixed $t$, $\xi^t$ is still smooth
and satisfies the condition $C_1$ and $C_2$.
It follows
that
${\xi}^{K}$ becomes
a section ${\xi}$ on $W^{\nu(a_{T_1}), T_1}(g_{t_0}, {\bf H}_{f_0}) $, defined by ${\xi}=\cup_{t\in W(\Sigma_{t_0})}\xi^t$.
In fact it becomes a section
${\widetilde\xi}$ on the larger space ${\widetilde W}^{\nu(a_0), T_1}(g_{t_0} )$ without the constraints given by ${\bf H}_{f_0} $.
Recall that these identifications $\lambda^{t_0}=:\{\lambda^{t_0}_t:S_t\rightarrow S_{t_0}\}$,
${\cal S}|_{W^{T_1}(\Sigma_{t_0})}\rightarrow S_{t_0}$,
induce a product structure $${\widetilde W}^{\nu(a_{T_1}), T_1}(g_{t_0} )\simeq {\widetilde W}^{\nu(a_0), t_0}(g_{t_0} ) \times W^{T_1}(\Sigma_{t_0}) ,$$ which in turn gives a smooth structure on ${\widetilde W}^{\nu(a_{T_1}), T_1}(g_{t_0} )$.
Moreover, for $t$ is sufficiently close to $t_0$ the map $\lambda_{t}^{t_0}:S_t\rightarrow S_{t_0}$ induces an identification of
${\cal L}^t_{g_t}$ with ${\cal L}^{t_0}_{g_{t_0}}$ for
${\hat g}_t={\hat g}_{t_0}\circ (\lambda_{t}^{t{_0}})$ essentially by the pull-back of $\lambda_{t}^{t_0}$. Indeed, in the case, that $E=TM$ and ${\cal L}^t_{g_t}=L_{k-1}^p(S_t, {g_t}^*(E))$ or ${\cal L}^t_{g_t}=L_{k-1}^p(S_t, {g_t}^*(E)\otimes \Lambda^ {1})$, it is exactly
given by the pull-back.
For ${\cal L}^t_{g_t}=L_{k-1}^p(S_t, {g_t}^*(E)\otimes \Lambda^{0, 1})$ and ${\cal L}^{t_0}_{g_{t_0}}=L_{k-1}^p(S_t, {g_{t_0}}^*(E)\otimes \Lambda^{0, 1}),$
the identification is given by composition of the pull-back by $\lambda_{t}^{t_0}$ with the map induced by the projection $ \Lambda^{ 1}\rightarrow \Lambda^{0, 1}$ since $\lambda_{t}^{t_0}$ is not holomorphic away from the fixed part $K$ as already observed in \cite {lt}.
Combing this with the
local trivializations of the bundles $({\cal L}^t\rightarrow {\widetilde W}^{\nu(a_0), t}(g_{t}))$ above, this gives a
trivialization of
the bundle $\{{\cal L}^{T_1}=\cup_{t\in W^{T_1}(\Sigma_{t_0})}{\cal L}^t
\rightarrow {\widetilde W}^{\nu(a_{T_1}), T_1}(g_{t_0})=
\cup_{t\in W^{T_1}(\Sigma_{t_0} )}{ \widetilde W}^{\nu(a_0), t}(g_{t})\}$ centered at $g_{t_0}$.
Using the fact that $\lambda_{t}^{t_0}$ is just the identity map on $K=K_{t}=K_{t_0}$ (and hence holomorphic) it is easy to check that with respect to this product smooth structure and local trivialization, $\xi /{\tilde \xi}$ so defined is
"constant" along $W^{T_1}(\Sigma_{t_0}) $-directions in the sense that
for $(h, t)\in { W}^{\nu(a_0), t_0}(g_{t_0} ) \times W^{T_1}(\Sigma_{t_0})\simeq { W}^{\nu(a_{T_1}), T_1}(g_{t_0} ) ,$
$\xi(h, t)=\xi^{t_0}(h)(=\xi^K(h|_K))$ ( similarly for ${\hat \xi}$). Hence it is a smooth section.
Clearly the condition $C_1$ and $C_2$ still hold for ${\tilde \xi}$.
The discussion here can be reformulated in the lemma.
\begin{lemma}
On ${ \widetilde W}^{\nu(a_{T_1}), T_1}(g_{t_0})$,
${\tilde \xi}=(\lambda^{t_0})^*({\tilde \xi}^{t_0}).$
Here the section ${\tilde \xi}|_{ \widetilde W^{\nu(a_{0}), t_0}(g_{t_0})}$ along the central slice ${ \widetilde W^{\nu(a_{0}), t_0}(g_{t_0})}$ is denoted by ${\tilde \xi}^{t_0}$.
\end{lemma}
\proof
This essentially is a tautology.
\hfill$\Box$\medskip
Thus ${\tilde \xi}$ is the $G_e^1$-extension of ${\tilde {\xi}}^{t_0}$. Here $G_e^1=W^{T_1}(\Sigma_{t_0})$ considers as a family of diffeomorphisms $\{\lambda^{t_0}_t:S_t\rightarrow S_{t_0}\}$ parametrized by $t\in W^{T_1}(\Sigma_{t_0})$. This interpretation proves the above lemma again. Similar interpretations using $G_e^i$-extensions with $i>1$ prove the two main theorems
below.
The first main theorem is the following.
\begin{theorem}
The section ${\tilde \xi}^{\Phi}$ is of class $C^{m_0}$.
Consequently its
restriction ${ \xi}^{\Phi}$ to the $C^{m_0}$-submanifold ${ U}^{ T_{{1}}}(g_{u_0}, {\bf H}_{f_0}) $ with respect to the second smooth structure is of class $C^{m_0}$. Here ${\xi}^{\Phi}$ is obtained
from the section ${\xi}$ on
$W^{\nu(a_0), T_1}(g_{t_0}, {\bf H}_{f_0})$ by the transformation $\Phi$, similarly for ${\tilde \xi}^{\Phi}$.
The section ${ \xi}^{\Phi, \Psi}$ on
$U^{ T_{1}}(g^r_{u_0}, {\bf H}^r_{f_0})$
induced by the diffeomorphism $\Psi :U^{ T_{1}}(g_{u_0}, {\bf H}_{f_0})\rightarrow U^{ T_{1}}(g^r_{u_0}, {\bf H}^r_{f_0})$ is of class $C^{m_0}$ as well.
\end{theorem}
\proof
The first statement follows from the fact that ${\tilde \xi}^{\Phi}$ is the $G_e^2$-extension of ${\xi}^{t_0}$ by pull-backs of the elements in $G_e^2$. Here $G_e^2=W^{T_1}(\Sigma_{t_0})(=U^{T_1}(\Sigma_{u(t_0)}))$ considered as the
family of diffeomorphisms $\{\lambda^{t_0}_t\circ \phi_t:S_{u(t)}\rightarrow S_t\rightarrow S_{t_0}\}$ parametrized by $t\in W^{T_1}(\Sigma_{t_0})$.
Since ${\xi}^{t_0}$ satisfies the conditions $C_1$ and $C_2$, a obvious generalization of the main theorem in the first paper of these sequel implies that the $C^{m_0}$-smoothness of the $G_e^2$-extension of ${\xi}^{t_0}$ above so that
${\tilde \xi}^{\Phi}$ is of class $C^{m_0}$.
To prove the last statement, we note that
in addition to diffeomorphism $\Psi:U^{ T_{1}}(g_{u_0}, {\bf H}_{f_0})\rightarrow U^{ T_{1}}(g^r_{u_0}, {\bf H}^r_{f_0})$ of that identified the bases, the standard local trivializations of the bundles ${\cal L}$ centered at $ g_{u_0}$ and $g^r_{u_0}$ using parallel transport along shortest connecting geodesics are also the "same"
in the sense that the trivialization for ${\cal L}$ on $U^{ T_{1}}(g^r_{u_0}, {\bf H}^r_{f_0})$ automatically give the one for ${\cal L}$ on the other by our assumption that each local hypersurfaces in ${\bf H}_{f_0}$ is geodesic submanifold.
\hfill$\Box$\medskip
The idea above can be used to proof the smoothness of ${\xi}$ viewed in any other chart $W'^{T_1}(f_0', {\bf H}')$: that is to define the corresponding $G_e^i$-extension of the same section ${\xi}^{t_0}$.
We start with the neighborhood $U'^{T_1}(g'_{u'_0}, {\bf H}')$ with the class $[g'_{u'_0}]=[g_{u_0}]$. Here $g_{u_0}=g_{u(t_0)}:(S_{u_0}, {\bf x}(u))\rightarrow (M,{\bf H}_{f_0})$ is the center of $U^{T_1}(g_{u_0}, {\bf H}_{f_0})$, similarly for $g'_{u'_0}=g_{u'(t'_0)}:(S_{u'_0}, {\bf x}'(u'))\rightarrow (M,{\bf H}'_{f'_0})$. Since adding-dropping making points does not affect the smoothness, we may assume that
the number od marked points of $g_u$ and $g'_{u'}$ are the same.
Now fix a dropping marking map $r=r_k^m$ that selects $m$ elements ${\bf x}^r(u)$ from the $k$ marked points ${\bf x}(u)$ satisfies the condition that each free component of an (hence any ) element $g_{u_0}\in U^{T_1}(g_{u_0})$ is minimally stabilized. This gives a "new" marking, an identification $(S_u)_v\simeq S^2_v=S^2$ of a free component $(S_u)_v, v\in T_1$. Choose a "compatible marking" $r'$ for an (hence any ) element $g'_{u'_0}\in U'^{T_1}(g'_{u'_0})$ accordingly.
Let $\Gamma=\Pi_{v\in T_1} \Gamma_v$ acting on the free components of $S_u$. Here each $\Gamma_v$ is a subgroup of $PSL(2, {\bf C})$ depending on
the number of doubles points on the component $S^2_v$. In particular, if $S^2_v$ is stable $\Gamma_v=\{e\}.$
Now the assumption $[g'_{u'_0}]=[g_{u_0}]$ of the centers above implies that there is a
${\widetilde \gamma_0}:S'_{u'_0}\rightarrow S_{u_0}$ such that $g_{u_0}\circ {\widetilde \gamma_0}=g'_{u'_0}$. Using the identifications (markings) of the free components with (a collection of ) $S^2$(s ) given by $r$ and $r'$, and denoted by
$\psi^r$ and $\psi^{r'} $, the map ${\widetilde \gamma_0}=\psi^r\circ \gamma_0\circ (\psi^{r'})^{-1}$ for an element
${\widetilde \gamma_0}\in \Gamma$, unique upto the finite isotropies of $g_{u_0}$.
Thus we can define the $\Gamma$-action on $U^{T_1}(g_{u_0})$ by a similar formula,
$\gamma\cdot g_{u}=:g_{u}\circ (\psi^r\circ \gamma\circ (\psi^{r'})^{-1})$. Here
$\psi^r=\{\psi^r_v, v\in T_1\}$ with $\psi^r_v:(S_u)_v\rightarrow S^2$, similarly for
$\psi^{r'}.$ Note that for all elements in $ g_{u}\in U^{T_1}(g_{u_0})$ the domains
are all the "same"; the parameter $u$ in te notation $\Sigma_u$ or $S_u$ only describes the locations of the distinguished points. In term of this "action" of $\Gamma$, we have $g'_{u'_0}=\gamma\cdot g_{u_0}$ between the two centers.
Denote $(\psi^r\circ \gamma\circ (\psi^{r'})^{-1}:S_{u'}\rightarrow S_u$ by ${\widetilde \gamma}$. Let ${\bf x}'^{\gamma_0}(u'_0):= {\widetilde \gamma}_0^{-1}({\bf x}(u)\in S_{u'}$ be the "new" marked points. Note that $g'_{u'_0}({\bf x}'^{\gamma_0}(u'_0))=g_{u_0}\circ {\widetilde \gamma}_0({\widetilde \gamma}_0^{-1}({\bf x}(u))=g_{u_0}({\bf x}(u))\in {\bf H}_{f_0}$. Now consider the collection of the $L_k^p$ maps $g'_{u'}$ of type $T_1$ near $g'_{u'_0}$ with the constrains on the new marked points: $g'_{u'}({\bf x}'^{\gamma_0}(u'))\in {\bf H}_{f_0}, $
denoted by
$ U'^{T_1}(g'_{u_0}, {\bf H}^{\gamma_0}_{f_0})$.
\begin{theorem}
Let ${\xi}'^{\Phi, \gamma_0} $ be the section ${\xi}^{\Phi} $ viewed in $ U'^{T_1}(g'_{u_0}, {\bf H}^{\gamma_0}_{f_0})$. Then ${\xi}'^{\Phi, \gamma_0} $ is of class $C^{m_0}$.
\end{theorem}
\proof
Let $G_e^3=\Gamma\times W^{T_1}(f_{t_0}, {\bf H}_{f_0})(=\Gamma\times G_e^1=\Gamma\times G_e^2$ with action of $G_e^2$ first then composing wit the action of $\Gamma$. Then the discussion above shows that ${\xi}'^{\Phi, \gamma_0} $ is the restriction to the slice $ U'^{T_1}(g'_{u_0}, {\bf H}^{\gamma_0}_{f_0})$ of the $G_e^3$-extension of $\xi^{t_0}$. Hence it is of class $C^{m_0}$.
\hfill$\Box$\medskip
Applying implicit function theorem in a similar way to the proof that $\Psi$ induced by dropping makings is a differomorphism implies the next corollary.
\begin{cor}
Let ${\xi}'^{\Phi} $ be the section ${\xi}^{\Phi} $ viewed in $ U'^{T_1}(g'_{u_0}, {\bf H}'_{f'_0})$. Then ${\xi}'^{\Phi} $ is of class $C^{m_0}$.
\end{cor}
This proves the half of the main theorem below.
\begin{theorem}
Let ${\xi}^{K}:{\widetilde W}(f_{K})\rightarrow {\cal L}^{K}$ be a smooth section satisfying the condition $C_1$ and $C_2$. Then the smooth section ${\xi}$ on $W^{\nu(a_{T_1}), T_1}(g_{t_0}, {\bf H}_{f_0})$
is of class $C^{m_0}$ viewed in any other such local slices $W^{\nu(a_{T_1}), T_1}(g'_{t'_0}, {\bf H}_{f'_0})$
or $U^{ T_1}(g'_{u'_0}, {\bf H}_{f'_0})$ with respect their own smooth structures. Here $[g'_{t'_0}]=[g_{t_0}]$ as unparametrized maps.
\end{theorem}
\proof
The other half essentially follows from the obvious "extension " of the action of $G_e^3$ defined above by composing further the identifications given by inverse of
$\phi'_{t'}$ (with $t'$ and $t$ corresponding to each other) first, then the inverse of $\lambda^{t_0'}_{t'}$. Indeed, let ${ {\xi}}'$ be the section $\xi$ viewed in
$W^{T_1}(f_{t_0}, {\bf H}_0)$.
Then upto the effect of using different (but fixed ) markings, ${ {\xi}}'$ is the restriction to $W^{T_1}(f_{t_0}, {\bf H}_0)$ of the new $G_e^3$-extension
of $\xi^{t_0}$. A similar argument to the proof of the corollary above will eliminate the effect of different makings so that ${ {\xi}}'$ is of class $C^{m_0}.$
\hfill$\Box$\medskip
|
\section{Introduction}
Weyl semimetal is a nontrivial topological gapless state and exhibits lots of exotic novel and robust properties, including chiral anomaly and Fermi arc etc. It has been a research focus recently as they are both experimentally important and theoretically interesting \cite{Witten:2015aoa, vishwanath,burkov0}. On the one hand, they are ideal systems to test the
macroscopic effects due to quantum anomaly \cite{Landsteiner:2016led}, including chiral magnetic effect and transport effects induced by the mixed axial-gravitational anomaly.
On the other hand, Weyl semimetal is a novel kind of topological quantum matter which goes beyond the Landau-Ginzburg's paradigm for classification of states of matter.
Similar to graphene systems \cite{jan}, Weyl semimetal systems could be strongly coupled and do not possess well-defined
quasiparticles. It is theoretically challenging and important to study strongly interacting Weyl semimetals, to go beyond the conventional approach on topological states of matter based on topological band theory or weakly coupled theory.
\para
The holographic correspondence maps the difficult question of strongly interacting field theory to a tractable weakly coupled gravitational problem. There have been lots of remarkable applications of holography to tackle the strongly interacting condensed matter questions \cite{Zaanen:2015oix,book0,review}. In particular,
holographic models for strongly interacting Weyl semimetals have been constructed in \cite{Landsteiner:2015pdh, Landsteiner:2015lsa} in which the anomalous Hall conductivity is an order parameter to characterize the quantum topological phase transition. The effects of the surface state \cite{Ammon:2016mwa} and topological invariants \cite{Liu:2018djq} in this holographic model exhibit key features of topological Weyl semimetals. Therefore with strong interaction topological Weyl semimetal still exits and holography is a practical tool to explore its property. Moreover, the nontrivial topological structure in the strongly interacting system can be revealed from the gravitational bulk physics \cite{Liu:2018bye, Liu:2018djq}. There exist two bulk matter fields in which one field is to generate a gap in the dual theory while the other matter field is to deform the Fermi points to a topologically nontrivial configuration (Weyl points or nodal lines). The different topological phases arise due to the different IR solutions in the bulk which are adiabatically disconnected and only one of the matters fields dominates in each solution. From the holographic model a nontrivial prediction is that the presence of odd viscosity is due to mixed axial-gravitational anomaly \cite{Landsteiner:2016stv}. Other various interesting aspects of holographic Weyl semimetals have been explored, including optical conductivity \cite{Grignani:2016wyz}, axial Hall conductivity \cite{Copetti:2016ewq}, disorder effect on topological phase transition \cite{Ammon:2018wzb} and the butterfly velocity \cite{Baggioli:2018afg}.
\para
In condensed matter systems, from weakly coupled theory Weyl semimetal can go through a quantum phase transition to a normal band insulator \cite{burkov1, Roy:2016rqw} or to Chern insulator \cite{burkov1,cm-1,Roy:2016amv} etc.\footnote{See appendix \ref{app:a} for examples from field theoretical approach.} It would be extremely interesting to explore the phase diagram of strongly interacting topological Weyl semimetal from holography. In the previous holographic models \cite{Landsteiner:2015pdh, Landsteiner:2015lsa} only a portion of degrees of freedom are gapped in the trivial phase and Weyl semimetal phase goes to a trivial semimetal phase after the phase transition. This paper aims to provide a holographic model to describe a quantum phase transition from Weyl semimetal to a phase in which all the degrees of freedom are gapped, namely, the trivial phase is instead an insulating phase. In doing so we start from the most generic holographic Weyl semimetal model by using the Stueckelberg trick to replace the complex scalar field in \cite{Landsteiner:2015pdh} by two real scalar fields and introduce the most general dilatonic coupling. Writing the equations for fluctuations of gauge fields into a Schrodinger equation, we can get the condition for the dilatonic couplings to produce the insulating phase.
With a proper choice of dilatonic coupling and potential terms, we could realise a holographic topological quantum phase transition between strongly interacting Weyl semimetal phase and Chern insulator (3+1D anomalous Hall state) phase. Then we show the evidences of the phase transition from the perspectives of free energy and conductivities.
\para
Our paper is organized as follows. In section \ref{sec2}, we introduce a generalized holographic model with dilatonic coupling to realise the quantum phase transition from Weyl semimetal phase to insulator phase and show that it is a first order phase transition. In section \ref{sec3}, the conductivities of the dual theory are explored by studying the vector gauge field fluctuations above the background geometry. Evidence for the insulating phase being a Chern insulator is discussed. In section \ref{sec4}, we conclude and discuss the open problems. Appendices \ref{app:a}, \ref{app:b}, \ref{app:c} are devoted to the details of the field theory model, the finite temperature equations of motion for holographic model and the Schrodinger potential approach for conductivities.
\section{Holographic setup}
\label{sec2}
\para
We shall start from the most general holographic system which duals to an anomalous system with $U(1)_V\times U(1)_A$. This $U(1)_A$ will be explicitly broken by turning on a source term which plays the similar role of mass effect in the dual field theory.
With the dilatonic coupling,
the generic holographic model is
\begin{eqnarray}
\mathcal{S}&=&\int d^5x\sqrt{-g}\bigg[\frac{1}{2\kappa^2}\big(R+12\big)-\frac{Y(\phi)}{4}\mathcal{F}^2-\frac{Z(\phi)}{4}F^2+
\frac{\alpha}{3}\epsilon^{abcde}A_a\Big(F_{bc}F_{de}+3\mathcal{F}_{bc}\mathcal{F}_{de}\Big)\nonumber\\
&&~~~~~~~~~~~~\label{eq:action}
-\frac{1}{2}(\partial\phi)^2-\frac{W(\phi)}{2}(A_a-\partial_a\theta)^2-V(\phi)
\bigg
\end{eqnarray}
with the vector gauge field strenght
$\mathcal{F}_{ab}=\partial_a V_b-\partial_b V_a$ and axial gauge field strength $F_{ab}=\partial_a A_b-\partial_b A_a.
$ Here $V_a$ and $A_a$ correspond to vector and axial current respectively. Note that the scalar fields $\phi$ and $\theta$ are real which are dual to operators $\bar{\psi}\psi$ and $\bar{\psi}\gamma^5\psi$. The action is invariant under the gauge transformation $\theta\to \theta+\chi, ~A_a\to A_a+\partial_a\chi$. One can recover the holographic Weyl semimetal model in \cite{Landsteiner:2015pdh} via $\Phi=\frac{1}{\sqrt{2}}\phi e^{i\theta}$, and choose $Y(\phi)=Z(\phi)=1, ~W(\phi)=q^2\phi^2, ~ V(\phi)=\frac{m^2}{2}\phi^2$. The model (\ref{eq:action}) is a generic holographic model for Weyl semimetal.\footnote{Note that
$\epsilon_{abcde}=\sqrt{-g}\varepsilon_{abcde}$ with
$\varepsilon_{0123r}=1$. Similar generalisation has been made in \cite{Grignani:2016wyz} to study the optical conductivity in the quantum critical regime.
}
\para
From now on we set $2\kappa^2=1.$
The equations of motion of the system are
\begin{eqnarray}
R_{ab}-\frac{1}{2}g_{ab}\big(R+12\big)-\frac{1}{2}T_{ab}&=&0 \,,\nonumber\\
\nabla_b\big(Y(\phi) \mathcal{F}^{ba}\big)+2\alpha\epsilon^{abcde}F_{bc}\mathcal{F}_{de}&=&0 \,,\nonumber\\
\nabla_b\big(Z(\phi) F^{ba}\big)+\alpha\epsilon^{abcde}\big(F_{bc}F_{de}
+\mathcal{F}_{bc}\mathcal{F}_{de}\big)-W(\phi) (A^a-\nabla^a \theta)&=&0\,,\nonumber\\
\nabla_{a}\nabla^{a}\phi-\frac{\partial_\phi Y(\phi)}{4}\mathcal{F}^2-\frac{\partial_\phi Z(\phi)}{4}F^2
-\frac{\partial_\phi W(\phi)}{2}(A_a-\partial_a\theta)^2-\partial_\phi V(\phi)
&=&0 \,,\nonumber\\
\nabla_a\left[W(\phi)(A^a-\nabla^a\theta)\right]&=&0 \,,\nonumber
\end{eqnarray}
where
\begin{eqnarray}
T_{ab}&=& Y(\phi)\bigg[\mathcal{F}_{ac}\mathcal{F}_{b}^{~c}-\frac{1}{4}g_{ab}\mathcal{F}^2\bigg]+Z(\phi) \bigg[F_{ac}{F}_{b}^{~c}-\frac{1}{4}g_{ab} F^2\bigg]+W(\phi)\bigg[ (A_a-\partial_a\theta) (A_b-\partial_b\theta)
\nonumber\\
&&~~~~~~
-\frac{1}{2}g_{ab} (A_c-\partial_c\theta)^2\bigg]+
\nabla_{a}\phi\nabla_{b}\phi-\frac{1}{2}g_{ab}(\partial\phi)^2-g_{ab}V(\phi)
\,.\nonumber
\end{eqnarray}
\para
The dual consistent currents can be obtained through the variation of the on-shell action with respect to the gauge fields,
\begin{eqnarray}
J^\mu&=&\lim_{r_c\to\infty}\Big[\sqrt{-g}Y \mathcal{F}^{\mu r}+4\alpha\sqrt{-g} \varepsilon^{r\mu\nu\rho\lambda}A_\nu \mathcal{F}_{\rho\lambda}+\frac{\delta S_\text{c.t.}}{\delta v_\mu}\Big]\,,\\
J_5^\mu&=&\lim_{r_c\to\infty}\Big[\sqrt{-g}Z F^{\mu r}+\frac{4}{3}\alpha\sqrt{-g} \varepsilon^{r\mu\nu\rho\lambda}A_\nu F_{\rho\lambda}+\frac{\delta S_\text{c.t.}}{\delta a_\mu}\Big]\,,
\end{eqnarray}
with
\begin{eqnarray}
\nabla_\mu J^\mu&=&0\,,\\
\nabla_\mu J_5^\mu&=&\lim_{r_c\to\infty}\Big[\sqrt{-g}W\left(A^r-\nabla^r\theta\right)-
\frac{\alpha}{3}\sqrt{-g}\varepsilon^{r\mu\nu\rho\lambda}
(F_{\mu\nu}F_{\rho\lambda}+3\mathcal{F}_{\mu\nu}\mathcal{F}_{\rho\lambda})\Big]+\text{c.t.}\,.
\end{eqnarray}
For simplicity the countertem part is not shown here. Note that the above equations are the dual Ward identities at the operator level. One can always choose the radial gauge $A_r=0$. For a particular state of the dual field theory, i.e. the fluctuation state around the background in the bulk, the term $-\sqrt{-g}Wg^{rr}\partial_r\theta$ plays the role of the explicit breaking term as in the weakly coupled theory which can be found in appendix \ref{app:a}. Since the Ward identity of conserved currents should not depend on the coupling constant of the system, it is expected that this holographic model describes a strongly interacting Weyl semimetal model.\footnote{There are also other holographic models for Weyl semimetal, e.g. from the pespective of fermionic spectral function \cite{Gursoy:2012ie} and top-down models \cite{Hashimoto:2016ize}. }
\para
We shall focus on the zero temperature physics. The
ansatz for the background fields at zero temperature is
\begin{eqnarray}\label{eq:bgzeroT}
ds^2=u (-dt^2+ dx^2+ dy^2)+\frac{dr^2}{u}+h dz^2\,,~~A=A_z dz\,,~~\phi=\phi(r)
\,,
\end{eqnarray}
where fields $u, h,A_z, \phi$ are functions of the radial coordinate $r$. Note that according to the equation of motion for $\theta$, a constant solution of $\theta$ will be found and we have set it to be zero.
The corresponding equations of motion are
\begin{eqnarray}\label{eq:1steom}
\frac{3u''}{u}+\phi'^2-\frac{3h'u'}{2hu}-\frac{W A_z^2}{hu}&=&0\,,\\
\frac{1}{4}\phi'^2+\frac{6}{u}-\frac{3u'}{4u}\Big(\frac{u'}{u}
+\frac{h'}{h}\Big)-\frac{V}{2u}
-\frac{W A_z^2}{4uh}+\frac{Z A_z'^2}{4h}
&=&0\,,\\
A_z''+A_z'\left(\frac{2u'}{u}-\frac{h'}{2h}+\frac{\phi' \partial_\phi Z}{Z}\right)
-\frac{A_z W}{u Z}&=&0\,,\\
\label{eq:4steom}
\phi''+\phi'\left(\frac{2u'}{u}+\frac{h'}{2h}\right)
-\frac{\partial_\phi V}{u}-\frac{A_z^2 \partial_\phi W}{2h u}-\frac{A_z'^2\partial_\phi Z}{2h}&=&0\,,
\end{eqnarray}
where the prime is the derivative with respect to the radial coordinate $r$.
We have four independent ODEs for four unknown fields.
\para
In this paper
we will choose
\begin{eqnarray}\label{eq:modelpara}
Z(\phi)=1 \,,~~~W(\phi)=-q_0 \Big[1-\cosh\big[\sqrt{\frac{2}{3}}\phi\big]\Big]\,,~~~V(\phi)= \frac{9}{2}\Big[1-\cosh\big[\sqrt{\frac{2}{3}}\phi\big]\Big]\,.
\end{eqnarray}
Note that the system is invariant under the transformation $\phi\to -\phi.$ When $\phi\to 0$, we have $W(\phi)\simeq\frac{q_0}{3}\phi^2$ and $V(\phi)\simeq -\frac{3}{2}\phi^2.$ It is obvious that $q_0$ plays a similar role as axial charge and we restrict to $q_0>0$. Close to the boundary (i.e. $r\to\infty$), $\phi\to 0$, the potential in (\ref{eq:modelpara}) has the form of
$V(\phi)=\frac{1}{2}m^2 \phi^2+\dots$
with $m^2=-3$.
The coupling $Y$ does not play any role in the background solution while it plays an important role for computing the conductivities. We set
\begin{eqnarray}\label{eq:formY}
Y(\phi)=\cosh\Big[\sqrt{\frac{2}{3}}\phi \Big]\,.
\end{eqnarray}
\para
Close to the UV boundary we have
\begin{eqnarray}
\phi=\frac{M}{r}+\dots\,,~~~~A_z=b+\dots
\end{eqnarray}
and a detailed expansion will be shown in subsection \ref{subs:aefe}. $M$ and $b$ corresponds to the sources of the dual scalar operator $\bar{\psi}\psi$ and chiral current $\bar{\psi}\gamma^5\gamma^z\psi$. Turning on these two sources, the dual field theory has the same structure as the weakly coupled field theory described in appendix \ref{app:a}. In the following we shall study the bulk geometry and its free energy by tunning the parameter $M/b$ in the UV.
\subsection{Zero temperature solutions}
\label{subs:zeroT}
To study the groundstate of the system, we focus on the zero temperature solutions.\footnote{At finite temperature the ansatz of the background fields, the corresponding equations of motion and asymptotic expansions can be found in appendix \ref{app:b}.} We will first find the near horizon solutions and then turn on irrelevant perturbations to generate the full solutions. At zero temperature, we find three different kinds of IR solutions.
\para
\noindent{\em The insulating phase.} For the insulating phase, the near horizon solution is\footnote{The Ricci scalar for the near horizon geometry at the leading order is $-3/r$ and therefore the geometry is singular near the horizon. Nevertheless the scalar potential in this solution is bounded above and satisfy the
the Gubser criterion \cite{Gubser:2000nd, Charmousis:2010zz}. Thus the singularity is acceptable and
the dual field theory is not pathological.}
\begin{eqnarray}
\label{eq:in1}
u&=&r(1+r)\,,\\
\label{eq:in2}
h&=&r(1+r)\,,\\
\label{eq:in3}
A_z&=&a_1 r^{\frac{1}{4}(\sqrt{1+8q_0}-1)}\,,\\
\label{eq:in4}
\phi&=&-\sqrt{\frac{3}{2}}\log \frac{r}{1+r}\,,
\end{eqnarray}
where $a_1$ is a free parameter. Note that $a_1$-term is the subleading term and it sources higher oder terms in $\phi$, thus different $a_1$ will flow the geometry to different $M/b$. The leading order of metric fields takes the form of $ds^2=r (-dt^2+dx^2+dy^2+dz^2)+\frac{dr^2}{r}$. This particular metric is known as the GPPZ gapped geometry \cite{Girardello:1999hj} and the properties of entanglement entropy and behavior of dual scalar operators have been studied in e.g. \cite{Liu:2013una, Bianchi:2001de}. The difference is that a nontrivial $A_z$ will generate an anisotropic geometry. As we will show in subsection \ref{subsec3.1}
there is a hard gap in the real part of diagonal optical conductivities while the anomalous Hall conductivity is nonzero at zero frequency, therefore this phase corresponds to a Chern insulator phase.
\para
\noindent{\em The Weyl semimetal phase.} The near horizon solution is
\begin{eqnarray}\label{eq:wsm1}
u&=&r^2 \,,\\
\label{eq:wsm2}
h&=&r^2\,,\\
\label{eq:wsm3}
A_z&=&a_0+\frac{\phi_0^2}{4a_0 r}e^{-\frac{2a_0\sqrt{q_0}}{\sqrt{3}r}}\,,\\
\label{eq:wsm4}
\phi&=&\frac{\phi_0}{r^{3/2}}e^{-\frac{a_0\sqrt{q_0}}{\sqrt{3}r}}\,.
\end{eqnarray}
The leading order of the IR geometry is an AdS$_5$ geometry with a constant $A_z$ and we can always rescale $a_0$ to $1$. The exponential terms are the irrelevant perturbations. With the free parameter $\phi_0$, this IR geometry could flow to the whole spacetime asymptotic to AdS$_5$ with different $M/b$. This kind of near horizon also shows up in the groundstate of the holographic superconductor \cite{Horowitz:2009ij} and
the holographic Weyl semimetal phase studied in \cite{Landsteiner:2015pdh}.
\para
{\em The critical point.} The near horizon solution is
\begin{eqnarray}\label{eq:cp1}
u&=&u_0 r^2 \big(1+\delta u r^{\alpha_c} \big)\,,\\
\label{eq:cp2}
h&=&\frac{q_0}{9}r^{2\beta}\big(1+\delta h r^{\alpha_c}\big)\,,\\
\label{eq:cp3}
A_z&=&r^\beta\big(1+\delta a r^{\alpha_c} \big)\,,\\
\label{eq:cp4}
\phi&=&\sqrt{\frac{3}{2}} (\log\phi_1) \Big(1+\delta \phi r^{\alpha_c} \Big)\,.
\end{eqnarray}
In the case of $q_0=15$, we have $(u_0,\beta,\phi_1,\alpha_c)\simeq (1.150,0.769,1.797,1.230)$ and
$(\delta u,\delta h,\delta a)\simeq(0.147,-1.043,0.591)\delta \phi$. At the leading order there is a Lifshitz symmetry $(t, x, y, r^{-1})\to c (t, x, y, r^{-1})$, $z\to c^\beta z$ which can set $\delta\phi=-1$ to flow the Lifshitz geometry to AdS$_5$. In the UV we have $(M/b)_c\simeq0.986$. Note that for $q_0>0$ other relevant perturbations around the Lifshitz fixed point are always complex which indicates the ciritical point is unstable \cite{Hartnoll:2011pp, Donos:2012js} and we will confirm this by studying the free energy.
\para
Integrating from the above near horizon solution to the boundary, we could obtain the full solution. Different from the previous studies on holographic semimetals \cite{Landsteiner:2015pdh, Liu:2018bye}, we find that the near horizon behavior (\ref{eq:wsm1} - \ref{eq:wsm4}) flows to $M/b$ whose value runs from zero to $(M/b)_c$, and keeps increasing to a finite value of $(M/b)_{t+}$ with $(M/b)_{t+}>(M/b)_c$ and then turns back to $(M/b)_{c}$. While the near horizon behavior (\ref{eq:in1} - \ref{eq:in4}) flows to $M/b$ whose value runs from infinity to $(M/b)_c$, and keeps decreasing to a finite $(M/b)_{t-}$ with $(M/b)_{t-}<(M/b)_c$ and increases to reach $(M/b)_{c}$ finally. Examples for the bulk profiles of the matter fields at different values of $M/b$ are shown in Fig. \ref{fig:bg}. The axial gauge field and the scalar field configurations in the topological phase and the insulating phase are generally separated by the bulk profiles (dashed black) at the critical point. In the topological phase, from UV to IR the axial gauge field $A_z$ decreases monotonically and ends at a finite value in the deep IR. The scalar field is not monotonic and it first increases, then decreases to zero in the deep IR. In the insulating phase, the axial gauge field decreases from UV to zero in the IR while the scalar field increases monotonically until it hits the IR singularity. Near the critical value of $M/b$, we observe oscillatory behavior of the matter fields (dashed color lines), which is due to the complex irrelevant deformations around the Lifshtiz fixed point. This can be taken as a signature of unstable critical solution, indicating that the phase transition is not continuous, which will be confirmed from the free energy in the next subsection.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.46\textwidth]{plot-bgaz.pdf}
\includegraphics[width=0.445\textwidth]{plot-bgphi.pdf}
\end{center}
\vspace{-0.6cm}
\caption{\small{The bulk profiles of $A_z$ and $\phi$ for different values of $M/b=0.941$ (green), $0.983$ (blue),
$0.987$ (dashed cyan), $0.986$ (dashed black), $0.984$ (dashed brown), $0.987$ (orange), $1.019$ (purple). The solid lines are profiles in the stable phase while dashed lines are for the unstable phase.}}
\label{fig:bg}
\end{figure}
\subsection{Asymptotic expansions and free energy}
\label{subs:aefe}
In order to study the stability of the background, we shall study the free energy of the bulk geometry. The asymptotic behavior and free energy for the finite temperature case can be found in \ref{app:af} and the zero temperature results can be obtained straightforwardly by setting $f=u$. At zero temperature, we have the following behaviour of fields near the UV boundary
\begin{eqnarray}
u&=& r^2-\frac{M^2}{6}+\frac{u_2}{r^2}+...\,,\\
h&=&r^2-\frac{M^2}{6}+\frac{b^2 q_0 M^2 }{12}\frac{\log r}{r^2}+\frac{h_2}{r^2}+...\,,\\
A_z &=&b-\frac{b q_0 M^2}{6}\frac{\log r}{r^2}+\frac{\eta}{r^2}+...\,,
\end{eqnarray}
\begin{eqnarray}
\phi &=&\frac{M}{r}-\frac{b^2 w_0 M }{6}\frac{\log r}{r^3}+\frac{O}{r^3}+...\,,
\end{eqnarray}
with
$u_2=\frac{1}{6}(b\eta-MO)+\frac{1}{72}q_0 b^2 M^2+\frac{M^4}{108}$ and $h_2=-\frac{1}{3}b\eta -\frac{1}{6}MO-\frac{1}{144}q_0 b^2 M^2+\frac{M^4}{108}.$
The free energy density can be obtained from the on-shell action to be
\begin{eqnarray}
\frac{\Omega}{V}=-\frac{1}{24} b^2 M^2 q_0 -\frac{b \eta}{2} +\frac{M^4}{48}-\frac{M O}{2}\,.
\end{eqnarray}
\para
With the bulk solution found in the previous subsection, we can obtain the free energy numerically.
Fig. \ref{fig:free} shows the free energy as a function of $M/b$ close to the phase transition. The critical point generated by IR geometry (\ref{eq:cp1} - \ref{eq:cp4}) is unstable and the system undergoes a first order quantum phase transition from the Weyl semimetal phase to an insulator phase.\footnote{Note that these quantum phases are not distinguished by symmetry breaking.} This behavior exists for any $q_0>0$. Notably this is quite different from the previous holographic model \cite{Landsteiner:2015pdh} in which a continuous holographic phase transition happens between the topological Weyl semimetal phase and a trivial semimetal phase.
The different order of the phase transition may imply the different underlying mechanics for these two types of phase transitions.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.7\textwidth]{plot-free.pdf}
\end{center}
\vspace{-0.7cm}
\caption{\small The free energy density as a function of $M/b$ for $q_0=15$. The blue (dashed) lines are solutions generated from the Weyl semimetal phase while the red (dashed) line are from the insulator phase. The black dot represents the free energy at the unstable critical point. The system undergoes a first order quantum phase transition from the Weyl semimetal phase to an insulating phase. }
\label{fig:free}
\end{figure}
\para
The phase transitions for interacting Weyl semimetals were studied in \cite{Roy:2016rqw} from the field theoretical approach, and it was found that for sufficiently strong interactions there exists a first order quantum phase transition between the Weyl semimetal and a normal band insulator. Our holographic study
shows that the quantum phase transition from strongly interacting Weyl semimetal to a Chern insulator (as we will show in the next section) is also of first order. Thus it broads our understanding on the phase structure of strongly interacting Weyl semimetals.
\section{Transport properties of the dual theory}
\label{sec3}
To figure out the exact nature of the stable phases,
we should study the conductivities.
In the following we will compute the full frequency dependent longitudinal and transverse electric conductivities. We will also study the phase diagram from the behavior of anomalous Hall conductivity at zero frequency.
\para
The conductivities of a quantum many body system can be computed via the Kubo formula
\begin{eqnarray}
\sigma_{ij}=\lim_{\omega\to 0}\frac{1}{i\omega}\langle J_i J_j\rangle_R (\omega, {\bf k}=0)\,.
\end{eqnarray}
In holography, the current-current retarded correlators can be computed by studying the fluctuations of the gauge fields dual to the currents around the background with infalling boundary conditions.
\subsection{Longitudinal conductivities}
\label{subsec3.1}
\para
We perturb the background (\ref{eq:bgzeroT}) by the fluctuation $\delta V_z=v_z(r) e^{-i\omega t}$, and obtain
the equation
\begin{eqnarray}\label{eq:fluvz}
v_z''+\bigg(\frac{2u'}{u}-\frac{h'}{2h}+\frac{\partial_\phi Y}{Y}\phi'\bigg)v_z'+\frac{\omega^2}{u^2}v_z=0\,.
\end{eqnarray}
The electric conductivities depend on the form of dilatonic coupling $Y(\phi)$ in the action (\ref{eq:action}) which is chosen to be (\ref{eq:formY}). Near the conformal boundary we have
\begin{eqnarray}
v_z=v_z^{(0)}+\frac{v_z^{(2)}}{r^2}+\frac{v_z^{(0)}\omega^2\log\Lambda r}{2r^2}+\cdots\,.
\end{eqnarray}
With proper boundary conditions in the IR, the optical longitudinal conductivity is then
\begin{eqnarray}
\sigma_{zz}=\frac{1}{i\omega}\bigg(2\frac{v_z^{(2)}}{v_z^{(0)}}-\frac{\omega^2}{2}\bigg)
\end{eqnarray}
where we have considered the counterterm to cancel the $\log \Lambda r$ term.
\para
In the phase with IR geometry (\ref{eq:in1} - \ref{eq:in4}), there are two linearly independent solutions for $v_z$ in IR
\begin{eqnarray}\label{eq:bc1}
v_{z1}&\simeq& c_1r^{\frac{1}{4}(1+ \sqrt{1-16\omega^2})}\Big(1+\mathcal{O}(r)\Big)\,,\\
\label{eq:bc2}
v_{z2}&\simeq& c_2
r^{\frac{1}{4}(1- \sqrt{1-16\omega^2})}\Big(1+\mathcal{O}(r)\Big)\,.
\end{eqnarray}
Both these two solutions are real and normalisable when $\omega<\Delta=1/4$. The unique boundary condition can be fixed by the analyticity condition of the correlator in $\omega$ \cite{Kiritsis:2015oxa}. When $\omega> \Delta$, the solutions become complex and we can choose the infalling boundary condition
\begin{eqnarray}
\label{eq:nhbc1}
v_z\simeq r^{\frac{1}{4}(1- i \sqrt{16\omega^2-1})}\Big(1+\mathcal{O}(r)\Big)\,.
\end{eqnarray}
The solution in (\ref{eq:bc1}) and (\ref{eq:bc2}) to produce the above infalling boundary conditon under the $\omega\to \omega+i\epsilon$ prescription is the first one, i.e. we choose the boundary condition for $v_z$ when $\omega<\Delta$
\begin{eqnarray}\label{eq:inphase-bc}
v_z\simeq r^{\frac{1}{4}(1+ \sqrt{1-16\omega^2})}\Big(1+\mathcal{O}(r)\Big)\,.
\end{eqnarray}
Since this boundary condition is real for $\omega<\Delta$, this leads to the result that the real part of the conductivity $\sigma_{zz}(\omega)$ vanishes. There is a continuum above a gap in the optical longitudinal conductivity. Note that in this calculation we have set the unit in which the IR horizon geometry is of the form (\ref{eq:in1} - \ref{eq:in4}). As this solution flows to a specific value of $A_z\simeq b_0$ at the UV boundary, this indicates that if we set the unit $b=1$, the width of the hard gap $\Delta/b$ is $1/4b_0$. As $b_0$ depends on the parameter $a_1$ in (\ref{eq:in3}) which will generate different $M/b$, we shall have different width of the gap for different $M/b$ in the insulating phase. Another equivalent way to see that there is indeed a hard gap in the optical conductivity is from the Schrodinger potential approach in appendix \ref{app:c}.
\para
In the phase with IR geometry (\ref{eq:wsm1} - \ref{eq:wsm4}), the IR infalling boundary condition for $v_z$ is
\begin{eqnarray}\label{eq:nhbc2}
v_z\simeq \frac{-i\omega}{r}K_1\big[\frac{-i\omega}{r}\big]\,.
\end{eqnarray}
In this case there is always a continous gapless spectrum for
$\text{Re}[\sigma_{zz}(\omega)].$
\para
With the boundary conditions (\ref{eq:nhbc1}) and (\ref{eq:nhbc2}), we could obtain the retarded Green function.
In Fig. \ref{fig:czz}, we plot the real part of the longitudinal conductivity in the topological phase and the insulating phase for different values of $M/b$. In the topological phase, the longitudinal conductivity is linear in $\omega$ at both small and large frequency regimes, which is similar to the holographic results in \cite{Grignani:2016wyz}. In the insulating phase there is a hard gap in the conductivity which confirms the nature of holographic insulating phase. There exists a continuum gapless spectrum above the gap and the conductivity eventually becomes linear in $\omega$ at large frequency.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.6\textwidth]{plot-czz.pdf}
\end{center}
\vspace{-0.5cm}
\caption{\small The real part of longitudinal conductivity as a function of the frequency $\omega/b$ for different values of $M/b=0.941$ (green), $0.983$ (blue), $0.987$ (orange), $1.019$ (purple) in the topological phase and the insulating phase.}
\label{fig:czz}
\end{figure}
\para
The dependance of the width of the hard gap as a function of $M/b$ is shown in Fig. \ref{fig:gap}. Similar to the weakly coupled case, it is monotonically increasing when we increase $M/b$ in the insulator phase and for large enough $M/b$, we have $\Delta/b \propto 0.22 (M/b-0.3)$.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.58\textwidth]{plot-gap.pdf}
\end{center}
\vspace{-0.5cm}
\caption{\small The dependance of the width of the hard gap as a function of $M/b$ in the holographic insulating phase (red) and weakly coupled field theory (dashed green).}
\label{fig:gap}
\end{figure}
\subsection{Transverse conductivities}
\label{subsec3.2}
The transverse conductivities can be studied by
considering fluctuations
$\delta V_x=v_x(r) e^{-i\omega t},
\delta V_y=v_y(r) e^{-i\omega t}.$
The corresponding equations for $v_x$ and $v_y$ are
\begin{eqnarray}\label{eq:fluvx1}
v_x''+\bigg(\frac{u'}{u}+\frac{h'}{2h}+\frac{\partial_\phi Y}{Y}\phi'\bigg)v_x'+\frac{\omega^2}{u^2}v_x+
8i\alpha\omega\frac{A_z'}{Yu\sqrt{h}}v_y &=&0\,,\\
\label{eq:fluvy1}
v_y''+\bigg(\frac{u'}{u}+\frac{h'}{2h}+\frac{\partial_\phi Y}{Y}\phi'\bigg)v_y'+\frac{\omega^2}{u^2}v_y-
8i\alpha\omega\frac{A_z'}{Yu\sqrt{h}}v_x&=&0\,.
\end{eqnarray}
Define $v_\pm=v_x\pm i v_y$, we obtain
\begin{eqnarray}
v_\pm''+\bigg(\frac{u'}{u}+\frac{h'}{2h}+\frac{\partial_\phi Y}{Y}\phi'\bigg)v_\pm'+\frac{\omega^2}{u^2}v_\pm \pm
8\alpha\omega\frac{A_z'}{Yu\sqrt{h}}v_\pm&=&0\,.
\end{eqnarray}
For the last two terms, in the deep IR $r\to 0$, the term $\frac{\omega^2}{u^2}$ always dominates. Thus the near horizon boundary conditions are the same as the case for the longitudinal conductivities. More explicitely,
for IR geometry (\ref{eq:wsm1} - \ref{eq:wsm4}), the IR infalling boundary conditions
$
v_\pm= \frac{-i\omega}{r}K_1\big[\frac{-i\omega}{r}\big],
$
while for IR geometry (\ref{eq:in1} - \ref{eq:in4}), the IR boundary conditions are
$
v_\pm\simeq r^{\frac{1}{4}(1+\sqrt{1-16\omega^2})}\big(1+\mathcal{O}(r)\big)
$
when $\omega<\Delta=1/4$ and $
v_\pm \simeq r^{\frac{1}{4}(1- i \sqrt{16\omega^2-1})}\Big(1+\mathcal{O}(r)\Big)
$ when $\omega> \Delta$.
\para
With the Green functions $G_\pm$ from the new variables $v_\pm$, we can compute $G_{xx}$, $G_{yy}$ and $G_{xy}$. We have $\sigma_{xy}\pm i\sigma_{xx}=\pm \frac{G_\pm}{\omega}$, i.e.
\begin{eqnarray}\label{eq:trancond}
\sigma_T=\sigma_{xx}=\sigma_{yy}=\frac{G_++G_-}{2i \omega}\,,~~~\sigma_{xy}=\frac{G_+-G_-}{2\omega}\,.
\end{eqnarray}
The Chern-Simons term in the consistent current contributes to the anomalous Hall conductivity. We define ${\bf J}_\text{cons}=\sigma_{\text{AH}}{\bf e_b}\times {\bf E}$ and we have
$\sigma_\text{AH}=8\alpha b-\sigma_{xy}=8\alpha b-\frac{G_+-G_-}{2\omega}$ \cite{Landsteiner:2015lsa}.
\para
The numerical results for the full frequency dependent transverse conductivities are shown in Fig. \ref{fig:cxy}.
The left plot in Fig. \ref{fig:cxy} is for the real part of optical transverse conductivities $\sigma_{xx}$ and $\sigma_{yy}$. We have a gapless spectrum in the Weyl semimetal phase. In the insulating phase, there is a continuous gapless spectrum above a hard gap $\Delta/b=1/4b_0$. This behavior is similar to the longitudinal component.
Different from the longitudinal one, in the Weyl semimetal phase if we increase $M/b$, the ratio of the transverse conductivity $\text{Re}\sigma_T$ to the frequency increases at low frequency. This difference is caused by the emergent Lifshitz symmetry in the critical point which leads to the result that $\text{Re}[\sigma_{zz}(\omega)]\propto \omega^{2-\beta}$ while both $\text{Re}[\sigma_{T}(\omega)]$ and $\text{Re}[\sigma_\text{AH}(\omega)]$ are proportional to $\omega^{\beta}$ when $M/b$ is approaching the (unstable) critical value \cite{Landsteiner:2016stv,Grignani:2016wyz}. The right plot in Fig. \ref{fig:cxy} is the real part of optical anomalous Hall conductivity. In the insulating phase different from the diagonal component, the anomalous Hall conductivity approaches a nonzero value at zero frequency although there is an emergent time reversal symmetry in the deep IR. This is because $\sigma_\text{AH}$
depends on the real part of $G_{\pm}$, it is nonvanishing when $\omega<\Delta$ and there is no hard gap. The nonvanishing $\sigma_\text{AH}$ crucially depends on the IR boundary condition for the fluctuations which is fixed by the $\omega\to \omega+i\epsilon$ prescription. Furthermore, we observe that there is smooth change at $\omega=\Delta$ for the optical anomalous Hall conductivity in the insulating phase.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.48\textwidth]{plot-cxy.pdf}
\includegraphics[width=0.48\textwidth]{plot-cah.pdf}
\end{center}
\vspace{-0.6cm}
\caption{\small {The real part of the transverse conductivity (left) and the anomalous Hall conductivity (right) as a function of frequency at different values of $M/b=0.941$ (green), $0.983$ (blue), $0.987$ (orange), $1.019$ (purple) in the topological phase and the insulator phase.}}
\label{fig:cxy}
\end{figure}
The behavior of conductivities in the insulating phase resembles that of a Chern insulator, indicating that our holographic model realises a quantum phase transition from a topological Weyl semimetal to a Chern insulator. In weakly coupled field theory, there are models to describe the phase transition from a Weyl semimetal to a Chern insulator \cite{burkov1,cm-1,Roy:2016amv}. Our holographic study confirms that a similar phase structure exists for strongly interacting Weyl semimetal.
\subsection{Phase diagram}
\label{subsec3.3}
The order parameter of the quantum phase transition between the topological phase and the insulating phase is the DC anomalous Hall conductivity. The DC anomalous Hall conductivities can be computed using a near-far matching method following \cite{Landsteiner:2015pdh}. We show here the explicit procedure of the calculations in the topological phase, and also comment on the calculations in the insulating phase.
\para
In the topological phase, near horizon the solution of $v_\pm$ with infalling boundary condition is ${v^{(n0)}_\pm}= \frac{-i\omega}{r}K_1\big[\frac{-i\omega}{r}\big].$ In the matching regime $\omega\ll r\ll \text{min} \{M, b\}$, this solution can be expanded as
\begin{eqnarray}
v^{(n0)}_\pm=1-\frac{\omega^2}{4r^2}\Big(-1+2\gamma+2\ln\Big[\frac{-i\omega}{2r}\Big]\Big)\,,
\end{eqnarray}
where $\gamma$ is the Euler-Mascheroni constant.
From this expansion, we know that at the matching region the infalling solution corresponds to the solution $1$ while the $\omega^2$ term can be ignored since we are interested in the $\omega\to 0$ result.
The linear order correction in $\omega$ to the near region solution is sourced by the infalling leading order solution. Thus at matching region the full linear order in $\omega$ boundary condition is
\begin{eqnarray}\label{eq:zeroTbc}
v^{(n)}_\pm=1+\omega v^{(n1)}_\pm,
\end{eqnarray} where $
{v^{(n1)}_\pm}'=\mp\frac{8\alpha (A_z(r)-A_z(0))}{ r^3}\,.
$
\para
In the far region $\omega\ll r$, we have
\begin{eqnarray}\label{eq:farreg}
{v^{(f)}_\pm}''+\Big(\frac{h'}{2h}+\frac{u'}{u}+\frac{\partial_\phi Y}{Y}\phi'\Big){v^{(f)}_\pm}'\pm\frac{8\omega\alpha }{Yu\sqrt{h}}A_z'{v^{(f)}_\pm}+\frac{\omega^2}{u^2}{v^{(f)}_\pm}=0\,.
\end{eqnarray}
Its solution can be expanded according to $\omega$ and we will solve the equation (\ref{eq:farreg}) up to the first order in $\omega$. Note that the last term in (\ref{eq:farreg}) can be ignored at order $\omega$. With the near horizon boundary condition (\ref{eq:zeroTbc}), we obtain the solution $v^{(f)}_\pm=1+ \omega v^{(f1)}_\pm$ where
$
{v^{(f1)}_\pm}' = \mp\frac{8\alpha (A_z(r)-A_z(0))}{Yu\sqrt{h}}.
$
\para
With the far region solutions, we obtain $G_\pm=\omega\big(\pm 8\alpha (b-A_z(0))\big)$. From (\ref{eq:trancond}) we obtain the DC conductivities
\begin{eqnarray}\label{eq:sigxy}
\sigma_{xy}=\frac{G_+-G_-}{2\omega}=8\alpha \big(b-A_z(0)\big)\,,~~~\sigma_{xx}=\sigma_{yy}= 0\,.
\end{eqnarray}
Note that in the computations above, the result of (\ref{eq:sigxy}) is for the anomalous Hall conductivity defined from the covariant currents. In the following, we will obtain the anomalous Hall conductivity for the consistent currents which is more close to results in real experimental systems \cite{Landsteiner:2016led, Landsteiner:2015lsa}. The final result for the zero frequency anomalous Hall conductivity in the holographic Weyl semimetal phase is
\begin{eqnarray}\label{eq:ahe}
\sigma_\text{AHE}=\text{Re}[\sigma_\text{AH}(\omega\to 0)]=8\alpha b-\sigma_{xy}=8\alpha A_z(0)
\end{eqnarray}
which is completely determined by the near horizon value of the axial gauge field.
\para
In the gapped phase, one could attempt to repeat the above near-far matching method to compute the DC anomalous Hall conductivity. The solution at the matching regime should be modified to be $v^{(n)}_\pm=\sqrt{r}+\omega v^{(n1)}_\pm$, which leads to the observation that $\omega v^{(n1)}_\pm$ can not be determined analytically.
In the insulating phase we do not have a simple formula as (\ref{eq:ahe}). This is also intuitively correct. The boundary condition (\ref{eq:inphase-bc}) is determined by the analytical continuation of the large frequency condition which reflects necessary information beyond the near horizon behavior. Therefore we have to compute the DC anomalous Hall conductivity numerically by taking $\omega\to 0$ limit of $\text{Re}[\sigma_\text{AH}(\omega)]$ obtained from the last subsection.
\para
Fig. \ref{fig:ahe} shows the anomalous Hall conductivity at zero frequency as a function of $M/b$. The solid line is for the stable phase while the dashed line is for the unstable phase, similar to the free energy plots in Fig. \ref{fig:free}. This figure shows that when we increase $M/b$ the non-zero anomalous Hall conductivity decreases and jump directly at a phase transition point to a nonzero value which seems to be insensitive to $M/b$.
The blue lines is for the background from IR geometry (\ref{eq:wsm1} - \ref{eq:wsm4}) while the red line is for the solutions from (\ref{eq:in1} - \ref{eq:in4}). The discontinuity of the zero frequency anomalous Hall conductivity further supports that this holographic phase transiton is of first order, which is consistent with the result from the free energy analysis. Moreover, in the holographic insulating phase, in the diagonal components of the optical conductivities there is a continuous gapless spectrum above a hard gap, and the zero frequency anomalous Hall conductivity is nonzero. These are the signals of a Chern insulator. Therefore, our holographic model describes a first order quantum phase transition from a strongly interacting Weyl semimetal to a Chern insulator.
\begin{figure}[h!]
\begin{center}
\includegraphics[height=6.7cm, width=0.65\textwidth]{plot-ahe.pdf}
\hspace{0.1in}
\includegraphics[height=6.6cm, width=0.3\textwidth]{plot-ahezoom.pdf}
\end{center}
\vspace{-0.6cm}
\caption{\small Both plots are for the zero temperature anomalous Hall conductivity at zero frequency of the holographic system. The right plot is a zoomed in version of the left plot close to the phase transiton point. In both plots, the blue line is for the topological Weyl semimetal phase and the red line is for the Chern insulator phase. The solid and dashed lines are for stable and unstable phases respectively. We see that under the phase transition, the anomalous Hall conductivity is discontinuous and the system undergoes a first order phase transition from a Weyl semimetal phase to a Chern insulator phase.}
\label{fig:ahe}
\end{figure}
\section{Conclusion and discussion}
\label{sec4}
We have provided a holographic model to charaterize the quantum phase transition between the strongly interacting Weyl semimetal and the Chern insulator, by tunning the ratio between the mass parameter and time reversal symmetry breaking parameter in the dual field theory. We established that this quantum phase transition is of first order. We also computed the frequency dependent conductivities numerically in each phase. In the holographic Weyl semimetal phase, we found that there is a nontrivial DC anomalous Hall conductivity and the diagonal components of optical conductivities are linear in frequency in both small and large frequency regimes. In the holographic Chern insulator phase, we found that there is a hard gap in the real part of the diagonal components of the frequency dependent conductivities and there is also a nonvanishing DC anomalous Hall conductivity. This is a first example of Chern insulator from holography signified by a nontrivial anomalous Hall conductivity in a gapped state.
\para
Our holographic model reveals the interesting phase diagram for strongly interacting Weyl semimetal and provides a novel framework to explore further problems of strongly coupled topological states. There are many open questions. Firstly, in the particular holographic model we studied, the phase transition is of first order. It would be interesting to see if it is still first order for more general holographic phase transition models between Weyl semimetal and insulating phase with different dilatonic couplings. It is also important to have a better understanding of the essential physics at the first order holographic quantum phase transition point.
Secondly, note that in field theory, there are studies of the disorder effects on the quantum phase transitions between Weyl semimetal and Chern insulators \cite{cm-1,Roy:2016amv,WSMCI}. It would be worthwhile to explore the disorder effects or other momentum dissipation effects on this holographic quantum phase transition to understand the similarities or differences to the weakly coupled field theoretical results. Meanwhile, the transport properties of the holographic system at finite temperature is to be further explored. Finally, the insulating phase we found in this work is a Chern insulator with nontrivial anomalous Hall conductivity. It would be very interesting to study the topological invariants of this holographic Chern insulator following \cite{Liu:2018djq}, to explore effects of surface states, to realise the phase transition to a normal insulator and so on. These studies should be helpful to build holographic models for topological insulators towards more complicated topological states of mater. We hope to explore some of these questions further in the future.
\vspace{.8cm}
\subsection*{Acknowledgments}
We thank Rong-Gen Cai, Karl Landsteiner, Francisco Pena-Benitez, Jie Ren, Shun-Qing Shen, Ya-Wen Sun, Zhong Wang for helpful discussions. This work was supported by the National Thousand Young Talents Program of China, NFSC Grant No.11875083 and a grant from Beihang University. Y.L. would also like to thank Hanyang University for the hospitality during the APCTP focus program ``Holography and Geometry of Quantum Entanglement'' where this work was presented
\vspace{.3 cm}
|
\subsection{Linear SIPLOGs}
In this section, we compute a KKT point of the following LSIPLOG for various values of $\mu$.
This problem is obtained by slightly modifying
the semi-infinite eigenvalue optimization problem solved in \cite{li2004solution}:
\begin{align}
\begin{array}{rcl}
\displaystyle{\mathop{\rm Maximize}_{X\in S^m}}& &A_0\bullet X-\mu\log\det X\\
\mbox{subject to}& & A(\tau)\bullet X\ge 0\ (\tau\in T)\\
& & I\bullet X = 1\\
& & X\in S^m_{++},
\end{array}\label{eig_semi}
\end{align}
where $A_0\in S^m$ and $A:T\to S^m$ is a symmetric matrix valued function
whose elements are $q$-th order polynomials in $\tau$.
We set $T=[0,1]$, i.e., $T_{\rm min}=0$ and $T_{\rm max}=1$, $(A(\tau))_{i,j}=\sum_{l=0}^qa_{i,j,l}\tau^l$ for $1\le i,j\le m$, and $q=9$.
We choose all entries of $A_0$ and the {coefficients} $a_{i,j,l}$ {in $A(\tau)$} from the interval $[-1,1]$ randomly.
Among {those} generated data {sets}, we use only data such that the semi-infinite constraint
{includes at least one active constraint}
at {an} optimum of \eqref{eig_semi}.
Specifically, for each data set, we compute an optimum, say $\tilde{X}$, of the SIPLOG obtained by removing the semi-infinite constraints.
If
$\min_{1\le i\le 21}A\left(T_{\rm min} + \frac{i-1}{20}(T_{\rm max}-T_{\rm min})\right)\bullet \tilde{X}\le -10^{-3}$, which implies that $\tilde{X}$ does not satisfy the semi-infinite constraints, we adopt it as a valid data set.
In the above manner, we generated 10 test problems for each $(m,\mu)\in \{10,25\}\times \{1,10^{-5}\}$ and
applied the algorithm for solving the generated problems.
All instances were successfully solved.
{We show the obtained results in Tables~\ref{ta1} and \ref{ta2}, where
``time(s)'', ``$R^{\ast}$'', ``{$\sharp$QP}'', and ``$\sharp$ite'' {stand for}
the average running time in seconds, the average value of $R$ at the solution output by the algorithm, the average number of QPs solved per run,
and the average number of iterations, respectively.
Moreover, ``H.K.M (resp. NT)''
{means} that the H.K.M (resp. NT) matrix is used as a scaling matrix $P_r$ in Step~2.
We observe that $\sharp\mbox{ite}$ {tends} to increase as $m$ {increases}.
For example, Table\,\ref{ta1} {shows that}, when the H.K.M scaling matrix is used,
$\sharp \mbox{ite}$ is 9.60 for $m=10$ while it is 13.40 for $m=25$.
A similar tendency can be {observed} between the values of $\sharp\mbox{ite}$ and $\mu$.
Actually, for the case of $10(\mbox{H.K.M})$, we find that
$\sharp\mbox{ite}$ {is} 9.6 for $\mu=1$
while it grows up to 33.2 for $\mu = 10^{-5}$.
This phenomenon might be caused because a solution of the LSIPLOG
approaches the boundary of the semi-definite region as $\mu$ decreases.
As the next observation, we see that $\frac{\sharp{\rm QP}}{\sharp{\rm ite}}$ lies between 1 and 2 in each row of Tables~\ref{ta1} and \ref{ta2}.
This indicates that, in the exchange method used in Step~2,
a solution satisfying the conditions\,\eqref{al:opt} {were} found after solving only one or two QPs on average.
{Finally, it may be worth mentioning that,
for many instances,
we observed superlinear-like convergence of
the value of the function $R$ to 0
in a last stage of iterations.}
\begin{table}[h]
\centering
\small
\begin{tabular}{|c|c|c|c|c|}\hline
$m$ (H.K.M./NT) & time(s)& $R^{\ast}$& {$\sharp$ QP}&$\sharp$ ite \\ \hline
10 (H.K.M.) & 0.15 & $6.35 \cdot 10^{-8}$ &11.7 &9.60 \\
25 (H.K.M.) & 0.53 & $ 1.77\cdot 10^{-7}$ & 15.9 & 13.40 \\ \hline
10 (NT) & 0.09 & $ 8.40 \cdot 10^{-8}$ &13.0 &10.20 \\
25 (NT) & 0.28 & $1.76 \cdot 10^{-7}$ &17.8 & 14.00 \\ \hline
\end{tabular}
\caption{Results for the LSIPLOG with $\mu=1$}
\label{ta1}
\end{table}
\begin{table}
\centering
\small
\begin{tabular}{|c|c|c|c|c|}\hline
$m$ (H.K.M./NT) & time(s) & $R^{\ast}$& {$\sharp$ QP}&$\sharp$ ite \\ \hline
10 (H.K.M.) & 0.54 & $1.44\cdot 10^{-7}$ & 45.9 &33.2 \\
25 (H.K.M.) & 1.83 & $7.88\cdot 10^{-8}$ & 42.4 & 37.9 \\ \hline
10 (NT) & 0.48 & $1.22\cdot 10^{-7}$ & 26.7 & 18.3 \\
25 (NT) & 0.91 & $1.62\cdot 10^{-7}$ & 23.4 & 19.1 \\ \hline
\end{tabular}
\caption{Results for the LSIPLOG with $\mu=10^{-5}$}
\label{ta2}
\end{table}
}
\subsection{Nonconvex SIPLOGs}
Next, we solve the following SIPLOG:
\begin{align}
\begin{array}{rcl}
\displaystyle{\mathop{\rm Minimize}_{x\in \R^{\frac{m(m+1)}{2}}}}& &\frac{1}{2}x^{\top}Mx+c^{\top}x+\omega{\|x\|^4}-\mu\log\det (X+\kappa I )\\
\mbox{subject to}& & a(\tau)^{\top}x\le b(\tau)\ (\tau\in T)\\
& & X + \kappa I\in S^m_{++}
\end{array}\label{eig_semi}
\end{align}
where
$\kappa >0$,
$a(\tau) : = (1,\tau,\tau^2,\tau^3,\ldots,\tau^{n-1})^{\top}\in \R^n$ with $n:=m(m+1)/2$ and
$b(\tau):=\sum_{i=1}^n\tau^{2i} + \sin(9\pi \tau)+2$.
All elements of $M\in S^n$ and $c\in \R^n$ are randomly generated from the interval $[-1,1]$.
The objective function is not convex in general but coercive in the sense that
$f(x)\to \infty$ as $\|x\|\to \infty$, and thus the problem is guaranteed to have at least one local optimum.
We set $T=[0,1]$ and $\kappa = \omega=0.01$, and generated 10 problems for each $(m,\mu)\in \{10, 25\}\times \{1,10^{-3}\}$.
We applied the algorithm for solving those problems.
We show the obtained results in Tables~\ref{ta3} and \ref{ta4}, where each column means the same as in Tables~\ref{ta1} and \ref{ta2}.
Compared with the LSIPLOG, there are more
significant differences between the results for $\mu=1$ and $\mu=10^{-3}$.
Specifically, when using the H.K.M scaling matrix with $m=25$,
$(\sharp {\rm ite},\sharp {\rm QP})$ is
$(329.9,424.4)$ for $\mu =10^{-3}$, while
$(\sharp {\rm ite},\sharp {\rm QP})$ is $(16.0,23.4)$ for $\mu =1$.
For the case where the NT scaling matrix is used, we also observe big differences between the results for
$\mu = 1$ and $\mu=10^{-3}$.
However, the NT scaling seems to exhibit more stable behavior than the H.K.M scaling.
In fact, when $m=25$,
time(s) for the H.K.M changes drastically from $1.29$ to $28.4$, while time(s) for the NT increases from $1.19$ to $7.32$.
As for $\sharp {\rm ite}$ and $\sharp {\rm QP}$, a similar tendency is observed.
\begin{table}[h]
\centering
\small
\begin{tabular}{|c|c|c|c|c|}\hline
$m$ (H.K.M./NT) & time(s) & $R^{\ast}$& {$\sharp$ QP}&$\sharp$ ite \\ \hline
10 (H.K.M.) & 0.22 & $8.24\cdot 10^{-7}$ &16.3 &9.90 \\
25 (H.K.M.) & 1.29 & $2.02\cdot 10^{-7}$ & 23.4 &16.0 \\ \hline
10 (NT) & 0.19 & $9.28\cdot 10^{-7}$ &14.9 & 10.0 \\
25 (NT) & 1.19 & $4.04\cdot 10^{-7}$ & 24.3 & 13.2 \\ \hline
\end{tabular}
\caption{Results for the NSIPLOG with $\mu =1$}
\label{ta3}
\end{table}
\begin{table}
\centering
\small
\begin{tabular}{|c|c|c|c|c|}\hline
$m$ (H.K.M./NT) & time(s) & $R^{\ast}$& {$\sharp$ QP}&$\sharp$ ite \\ \hline
10 (H.K.M.) & 0.98 & $1.78\cdot 10^{-7}$ & 71.8 &54.5 \\
25 (H.K.M.) & 28.40 & $4.69\cdot 10^{-7}$ & 422.4 & 329.9 \\ \hline
10 (NT) & 0.45 & $1.77\cdot 10^{-7}$ & 40.2 & 22.6 \\
25 (NT) & 7.32 & $3.48\cdot 10^{-7}$ & 162.8 & 75.5 \\ \hline
\end{tabular}
\caption{Results for the NSIPLOG with $\mu = 10^{-3}$}
\label{ta4}
\end{table}
\section{}
\section{Introduction}
In this paper, we consider the following semi-infinite
program that minimizes
a nonlinear function including a log-determinant (logdet) function over
an infinite number of convex inequality constraints and a positive-semidefinite constraint, called SIPLOG for short:
{\begin{align}
\begin{array}{ll}
\displaystyle{\mathop{\rm Minimize}} &f(x)-\mu\log\det F(x)
\vspace{0.5em}\\
{\rm subject~to} &g(x,\tau)\le 0\ \mbox{ for all } {\tau}\in T, \\
&F(x)\in S^m_{++},\\
&Gx=h,
\end{array}\label{lsisdp}
\end{align}
where $\mu\in \R$
is a positive constant,
$f:\R^n\to \R$ is a continuously differentiable function, and $T$ is a compact metric space.
In addition, $g:\R^n\times T\to \R$ is a continuous function with $g(\cdot,\tau)$ being convex and continuously differentiable.
Moreover, $S^m$ and $S^m_{++} (S^m_{+})$ denote the sets of
$m\times m$ symmetric matrices
and symmetric positive (semi-)definite matrices, respectively, and
$F(\cdot):\R^n\to S^{m}$ is an affine function, i.e.,
$$
F(x):=F_0+\sum_{i=1}^nx_iF_i
$$
with $F_i\in S^m$ for $i=0,1,\ldots,n$ and $x=(x_1,x_2,\ldots,x_n)^{\top}$.
Finally, $G\in \R^{s\times n}$ and $h\in \R^s$.
{Throughout the paper,
we assume that SIPLOG\,\eqref{lsisdp} has a Slater point, i.e., a vector $\bar{x}\in \R^n$ such that
$$
F(\bar{x})\in S^m_{++},\ g(\bar{x},\tau)<0\ (\tau\in T),\
G\bar{x}=h.
$$
This assumption implies that the set of feasible points taking finite objective values is not empty.
When $\mu{\searrow} 0$, the SIPLOG reduces to a semi-infinite semi-definite program (SISDP):
\begin{equation}
{\rm min}\ f(x)\ \ \mbox{s.t. }\ g(x,\tau)\le 0\ \mbox{ for all } {\tau}\in T,\ {F}(x)\in S^m_{+},\ Gx=h.\label{lsisdp2}
\end{equation}
FIR filter design problem\,\cite{spwu1996} and
robust envelop-constrained filter design with orthonormal bases\,\cite{li2007robust} can be formulated as an SISDP whose functions are all affine with respect to $x$.
For solving linear SISDPs, a discretization-type method and relaxed cutting plane method were proposed
in \cite{li2004solution} and \cite{li2006relaxed}, respectively.
In the absence of the semi-definite constraint and the logdet function, the SIPLOG becomes a nonlinear semi-infinite program (NSIP) with an infinite number of convex inequality constraints. For an overview of the NSIP, see \cite{sip2,sip-recent,Reem} and references therein.
On the other hand, in the absence of the semi-infinite constraints, the SIPLOG becomes a nonlinear semi-definite program (SDP).
We can find
some important applications for the nonlinear SDP in finance\,\cite{leibfritz2009successive,konno2003cutting}
and optimal control\,\cite{freund2007nonlinear,leibfritz2002interior}.
For solving the nonlinear SDP, several existing algorithms for nonlinear programs such as a primal dual interior point method and a sequential quadratic programming (SQP) method were extended\,\cite{freund2007nonlinear,yabe,yamashita2012local}. See the survey article\,\cite{yamashita2015survey} for other algorithms of the nonlinear SDP.
The logdet function plays a crucial role in
various fields such as statistics, experimental design, and information and communication theory.
In continuous optimization, it has a close connection with the interior point method for the SDP\,\cite{wolkowicz2012handbook}.
Accordingly, many algorithms
for solving optimization problems including the logdet function
have been studied extensively so far. For example, see \cite{vandenberghe1998determinant,yang2013proximal,wang2016solve}.
It makes sense to study the SIPLOG itself.
Indeed, the D-optimal experimental design problem can be formulated as an SIPLOG straightforwardly\,\cite{vandenberghe1998connections}.
Moreover,
in the spirit of the primal-dual interior point method for the nonlinear SDP\,\cite{yabe},
we can expect that a sequence of Karush-Kuhn-Tucker (KKT) points for the SIPLOG
with $\mu>0$ decreasing to 0 converges to a KKT point of the SISDP\,\eqref{lsisdp2}.
Hence, development of algorithms for solving the SIPLOG can be connected to efficient interior point methods for the SISDP.
In this paper, we focus on computing a KKT point of the SIPLOG\,\eqref{lsisdp}.
More specifically, we propose a new interior point SQP-type algorithm combined with an exchange method\,\cite{
Reem, Lai, okuno2012regularized,okuno2016exchange}.
In the method, we inexactly solve a semi-infinite (convex) quadratic program approximating the SIPLOG\,\eqref{lsisdp} to compute a search direction in the primal space.
Furthermore, to compute a search direction in the dual matrix space associated with the semi-definite constraint $F(x)\in S^m_+$,
we solve certain
scaled Newton equations that yield
the family of Monteiro-Zhang directions\,\cite[Chapter~10]{wolkowicz2012handbook}.
The proposed method can be regarded as an extension of the primal-dual interior point method \cite{yabe} for computing a barrier KKT point for the nonlinear SDP. However, the extension is not straightforward due to the presence of semi-infinite constraints.}
The paper is organized as follows:
In Section~\ref{sec:2}, we describe the KKT conditions for the SIPLOG.
In Section~\ref{sec:3}, we
propose an interior point SQP-type method for finding a KKT point and establish its convergence.
In Section~\ref{sec:4}, we conduct some numerical experiments to demonstrate the efficiency of the proposed method.
Finally, we conclude the paper with some remarks.
\subsection*{Notations}
The identity matrix of order $m$ is denoted by $I$.
For any $P\in \R^{m\times m}$, ${\rm Tr}(P)$ denotes the trace of $P$.
For any symmetric matrices $X,Y\in S^m$, we define
the Jordan product of $X$ and $Y$ by $X\circ Y:=(XY+YX)/2$ and the inner product of $X$ and $Y$ by
$X\bullet Y={\rm Tr}(XY)$.
Also, we denote the Frobenius norm of $X\in S^m$ by $\|X\|_F:=\sqrt{X\bullet X}$.
We define the linear operator $\mathcal{L}_X:S^m\to S^m$ by $\mathcal{L}_X(Z):=X\circ Z$ for any $X\in S^m$.
We also denote $({\zeta})_+:=\max({\zeta},0)$ for any $\zeta\in \R$.
For sequences $\{y^k\}$ and $\{z^k\}$ of vectors, if $\|y^k\|\le M\|z^k\|$
for any $k$ with some $M>0$,
we write $\|y^k\|=O(\|z^k\|)$.
Moreover, if there exists a positive sequence $\{\alpha_k\}$ with $\lim_{k\to\infty}\alpha_k=0$ and $\|y^k\|\le \alpha_k\|z^k\|$ {for any $k$},
we write $\|y^k\|=o(\|z^k\|)$.
{For matrices $X_1,X_2,\ldots,X_p\in S^m$ and $Y\in S^m$,
we denote
$
(X_{i}\bullet Y)_{i=1}^n
:=
(X_1\bullet Y,
X_2\bullet Y,
\ldots,
X_{n}\bullet Y)^{\top}.
$}
{Let $\mathcal{C}(T)$ be the set of real-valued continuous functions defined on $T$ endowed with the supremum norm $\|h\|:=\max_{\tau\in T}|h(\tau)|$.
Let $\mathcal{M}(T)$ be the dual space of $\mathcal{C}(T)$
that can be identified with the space of (finite signed) regular Borel measures with the Borel sigma algebra $\mathcal{B}$ on $T$ equipped with the total variation norm, i.e.,
$\|y\|:=\sup_{A\in \mathcal{B}}y(A)-\inf_{A\in \mathcal{B}}y(A)$ for $y\in \mathcal{M}(T)$,
and denote by $\mathcal{M}_+(T)$ the set of all the nonnegative Borel measures of $\mathcal{M}(T)$. }
{\section{KKT conditions for the SIPLOG}\label{sec:2}
In this section, we present the Karush-Kuhn-Tucker (KKT) conditions for the SIPLOG\,\eqref{lsisdp}.
We say that the KKT conditions for SIPLOG\,\eqref{lsisdp} hold at $x^{\ast}\in \R^n$ if
there exists some finite Borel-measure $y\in \M(T)$ such that
\begin{align*}
&\nabla f(x^{\ast})+\int_T\nabla_xg(x^{\ast},\tau)\ydt-
(F_{i}\bullet \mu F(x^{\ast})^{-1})_{i=1}^n
+G^{\top}z=0,\notag \\
&\int_Tg(x^{\ast},\tau)\ydt=0,\ g(x^{\ast},\tau)\le 0\ (\tau\in T),\ y\in\M_+(T),\\
&Gx^{\ast}=h, \notag
\end{align*}
where
$z\in \R^s$ is a Lagrange multiplier vector {associated with} the equality constraints $Gx=h$.
If $x^{\ast}$ is a local optimum of the SIPLOG, under Slater's constraint qualification, the KKT conditions hold at $x^{\ast}$.
In particular, there exists some discrete measure $y\in \M_+(T)$ satisfying the KKT conditions and $|{\rm supp}(y)|\le n$, where ${\rm supp}(y):=\{\tau\in T\mid y(\{\tau\})\neq 0\}$.
Conversely, when $f$ is convex, if the KKT conditions hold at $x^{\ast}$, then $x^{\ast}$ is an optimum of SIPLOG\,\eqref{lsisdp}.
Let $V\in S^m$.
Since $F(x^{\ast})\in S^m_{++}$,
$\mu F(x^{\ast})^{-1} = V$ if and only if $F(x^{\ast})\circ V=\mu I$ and $V\in S^m_{++}$.
Then, using the matrix $V$ as a slack matrix variable, we can rewrite the KKT conditions as
\begin{align}
&\nabla f(x^{\ast})+\int_T\nabla_xg(x^{\ast},\tau)\ydt-(F_i\bullet V)_{i=1}^n+G^{\top}z=0,\label{e1}\\
&{F}(x^{\ast})\circ V=\mu I,\ F(x^{\ast})\in S^m_{++},\ V\in S^m_{++},\label{e2}\\
&\int_Tg(x^{\ast},\tau)\ydt=0,\ g(x^{\ast},\tau)\le 0\ (\tau\in T),\ y\in\M_+(T),\label{e3}\\
&Gx^{\ast}=h.\label{e4}
\end{align}
Hereafter, we call a quadruple $(x,y,z,V)\in \R^n\times \M(T)\times \R^s\times S^m$ satisfying the conditions \eqref{e1}--\eqref{e4} a KKT point of the SIPLOG\,\eqref{lsisdp}.
Note that the conditions\,\eqref{e1}, \eqref{e2}, and \eqref{e4} can be cast as the perturbed KKT conditions for the nonlinear SDP which is obtained by removing the semi-infinite constraint $g(x,\tau)\le 0\ (\tau\in T)$ from SISDP\,\eqref{lsisdp2}.
Yamashita et al.\,\cite{yabe} proposed a primal-dual interior point method to find a solution satisfying those perturbed (barrier) KKT conditions for the nonliear SDP\footnote{
{Yamashita et al.\,\cite{yabe} considered the nonlinear SDP of the form:
$
\min\ f(x)\ \mbox{s.t. }\hat{h}(x)=0,\ \hat{G}(x)\in S^m_{++},
$
where the functions $\hat{h}:\R^n\to \R^s$ and $\hat{G}:\R^n\to S^{m}$ are continuously differentiable.} }.
In the next section, we will propose an interior point SQP-type algorithm for computing a KKT point $(x,y,z,V)$ satisfying the conditions\,\eqref{e1}--\eqref{e4}.
This algorithm can be regarded as an extension of the algorithm proposed by Yamashita et al.\,\cite{yabe}.
Nevertheless, the way of extension is not straightforward because we must handle the semi-infinite constraint efficiently.}
\section{Interior point SQP-type algorithm for finding a KKT point}\label{sec:3}
In this section, we give an interior point SQP-type method for getting a KKT point of SIPLOG\,\eqref{lsisdp}.
Throughout the section, we use the following notations:
\begin{align*}
y_i^r&:=y(\tau_i^r)\ \ \mbox{for }{\rm supp}(y^r)=\{\tau_1^r,\tau_2^r,\ldots,\tau_{p_r}^r\}\\
w^r &:= (x^r,y^r,z^r,V_r)\in \R^n\times \mathcal{M}(T)\times \R^s\times S^m.
\end{align*}
The proposed algorithm composes iteration points $\{w^r\}_{r\ge 0}$ sequentially
by
$$
\left(x^{r+1},
V_{r+1}
\right)
= \left(x^r +s_r\Delta x^r, V_{r}+s_r\Delta V_r\right),
$$
where
$\left(\Delta x^r,\Delta V_r\right)\in \R^n\times S^m$ denotes a search direction and
$s_r>0$ is a step size chosen so that the interior point constraints
\begin{equation*}
F(x^{r+1})\in S^m_{++}\ \mbox{and }\ V_{r+1}\in S^m_{++}
\end{equation*}
are satisfied.
In addition, we produce a sequence $\{y^r\}\subseteq \mathcal{M}_+(T)$ with $\left|{\rm supp}(y^r)\right|<\infty$ for any $r\ge 0$.
Hereafter, we often drop super- or sub-scripts from those symbols for simplicity of expression.
\subsection{Search direction $(\Delta x,\Delta V)$ and Lagrange multipliers $(y^+, z^+)$}
In what follows, we explain how to generate a search direction $\left(\Delta x,\Delta V\right)$
together with Lagrange multiplier measure $y^+\in \M_+(T)$ and vector $z^+\in \R^s$
at the current point
$w=(x,y,z,V)$.
In applying an SQP-like method to \eqref{lsisdp}, it is natural to think of the following semi-infinite quadratic program, called SIQP for short, with infinitely many {\it linear} constraints:
\begin{align}
\begin{array}{ll}
\displaystyle{\mathop{\rm Minimize}_{\Delta x}} &\nabla f(x)^{\top}\Delta x +\frac{1}{2}\Delta x^{\top}B\Delta x-\mu \xi(x)^{\top}\Delta x
\vspace{0.5em}\\
{\rm subject~to} &g(x,\tau)+\nabla_xg(x,\tau)^{\top}\Delta x\le 0\hspace{1.0em}(\tau\in T),\\
&G(x+\Delta x)=h,
\end{array}\label{SIQP}
\end{align}
where the coefficient matrix $B\in S^n$ is chosen to be positive definite and
the function $\xi:\R^n\to \R^n$ is defined by
\begin{equation}
\xi(x):=\nabla\log\det F(x)=(F_i\bullet F(x)^{-1})_{i=1}^n.\label{eq:wx}
\end{equation}
Solving the above problem is still difficult since it contains the semi-infinite constraints
$
g(x,\tau)+\nabla_xg(x,\tau)^{\top}\Delta x\le 0\ (\tau\in T).
$
To relax the difficulty, we propose to make use of its inexact solution
$\Delta x\in \R^n$ together with Lagrange multiplier measure $y^+\in \mathcal{M}_+(T)$ satisfying $\left|{\rm supp}(y^+)\right|<\infty$ and vector $z^+\in \R^s$
such that
\begin{align}\label{al:opt}
&\nabla f(x)+B\Delta x-\mu \xi(x)+\int_T\nabla_xg(x,\tau)\ydtp+G^{\top}z^+=0,\notag \\
&g(x,\tau)+\nabla_xg(x,\tau)^{\top}\Delta x\le 0\ \ (\tau\in {\rm supp}(y^+)),\\
&\int_T\left(g(x,\tau)+\nabla_xg(x,\tau)^{\top}\Delta x\right)\ydtp=0,\ G(x+\Delta x)=h,\notag \\
&\max_{\tau\in T}\left(
g(x,\tau)+\nabla_xg(x,\tau)^{\top}\Delta x
\right)_+\le \gamma,\notag
\end{align}
where $\gamma>0$ is a relaxation parameter controlled in the algorithm.
Thanks to $|{\rm supp}(y^+)|<\infty$, the above integral forms can be replaced with simple finite summations:
\begin{align*}
&\int_T\nabla_xg(x,\tau)\ydtp=\sum_{j=1}^p\nabla_xg(x,\tau_j)y^+({\tau_j}),\\
&\int_T\left(g(x,\tau)+\nabla_xg(x,\tau)^{\top}\Delta x\right)\ydtp=
\sum_{j=1}^p\left(g(x,\tau_j)+\nabla_xg(x,\tau_j)^{\top}\Delta x\right)y^+({\tau_j})
\end{align*}
with $p:=|{\rm supp}(y^+)|$ and ${\rm supp}(y^+)=\{\tau_1,\tau_2,\ldots,\tau_p\}$.
Notice here that, if $\gamma=0$, then
the conditions\,\eqref{al:opt} are noting but the KKT conditions for SIQP\,\eqref{SIQP}.
We should further remark that exchange-type methods \cite{Reem,Lai, okuno2012regularized,okuno2016exchange} work effectively in finding vectors satisfying those conditions.
Below, an exchange method for finding $\left(\Delta x,y^+,z^+\right)$ satisfying the conditions \eqref{al:opt} is described:
\begin{center}
{\underline{Exchange method}}
\end{center}
\begin{description}
\item{Step~0:} Choose the initial {index} set $T_0\subsetneq T$ such that $|T_0|<\infty$. Set $k:=0$.
\item{Step~1:} Solve SIQP\,\eqref{SIQP} with $T$ replaced by $T_k$ to obtain an optimum $\Delta x^k$ and Lagrange multipliers $\zeta_{\tau}\ge 0\ (\tau\in T_k)$ corresponding to the inequality constraints.
\item{Step~2:} Set $\tilde{T}_{k}:=\{\tau\in T_k\mid \zeta_{\tau}>0\}$.
\item{Step~3:} Find an index $\tau\in T$ such that $g(x,\tau)+\nabla g(x,\tau)^{\top}\Delta x>\gamma$ and let $T_{k+1}:=\tilde{T}_k\cup \{\tau\}$.
If such an index $\tau\in T$ does not exist, i.e., $\max_{\tau\in T}\left(g(x,\tau)+\nabla g(x,\tau)^{\top}\Delta x\right)\le \gamma$ holds, stop the algorithm. Otherwise, go to Step~4.
\item{Step~4:} Set $k:=k+1$ and return to Step~1.
\end{description}
In Step~2, we drop indices corresponding to the inequality constraints with zero Lagrange multipliers, which contain inactive constraints.
Particularly,
it can be proved in a manner similar to \cite[Theorem~3.2]{okuno2012regularized}
that
under the positive-definiteness of the matrix $B$,
the above exchange method stops in finitely many iterations.
We next consider how we derive $\Delta V$ by means of scaling techniques.
Choose a nonsingular matrix $P\in \R^{m\times m}$ arbitrarily and scale $F(x)$ and $V$ by
\begin{align}
&\tilde{F}(x):=PF(x)P^{\top},\ \tilde{V}:=P^{-\top}VP^{-1}.\label{scal}
\end{align}
Note that the barrier shifted complementarity conditions
$F(x)\circ V=\mu I,\ F(x)\in S^m_+,\ V\in S^m_+$ are equivalent to the scaled ones
$\tilde{F}(x)\circ \tilde{V}=\mu I,\ \tilde{F}(x)\in S^m_+,\ \tilde{V}\in S^m_+$.
Then, the Newton equations for $\tilde{F}(x)\circ \tilde{V}=\mu I$ are written as
\begin{equation}
\left(\tilde{F}(x)+\sum_{i=1}^n\Delta x_i\tilde{F}_i\right)\circ\tilde{V}+
\tilde{F}(x)\circ\Delta \tilde{V}=\mu I. \label{eq:scaled_new}
\end{equation}
Here, $\tilde{F}_i:=PF_iP^{\top}$ for $i=1,2,\ldots,n$.
In terms of the linear operator $\mathcal{L}_{\tilde{F}(x)}:S^m\to S^m$,
\eqref{eq:scaled_new}
is rephrased as
\begin{equation}
\sum_{i=1}^n\Delta x_i{\tilde{F}_i}\circ\tilde{V}+
\mathcal{L}_{\tilde{F}(x)}\Delta \tilde{V}=\mu I-\mathcal{L}_{\tilde{F}(x)}\tilde{V}.\label{eq:scaled_new_liap}
\end{equation}
Under the condition that ${F}(x)\in S^m_{++}$,
$\mathcal{L}_{\tilde{F}(x)}$ is invertible, and hence \eqref{eq:scaled_new_liap} is uniquely solvable for $\Delta\tilde{V}$. Actually, we have
\begin{equation}
\Delta \tilde{V}=\mu\tilde{F}(x)^{-1}-\sum_{i=1}^n\Delta x_i\mathcal{L}_{\tilde{F}(x)}^{-1}\mathcal{L}_{\tilde{V}}\tilde{F}_i-\tilde{V}.\label{eq:0326}
\end{equation}
Now, we propose to derive $\Delta V$ by the inverse-scaling of $\Delta \tilde{V}$. Specifically,
$\Delta V$ is computed {as}
\begin{equation}
\Delta V=P^{\top}\Delta \tilde{V}P=
\mu{F}(x)^{-1}-V-\sum_{i=1}^n\Delta x_iP^{\top}\mathcal{L}_{\tilde{F}(x)}^{-1}\mathcal{L}_{\tilde{V}}\tilde{F}_iP.
\label{eq:dv}
\end{equation}
The direction $\Delta V$ obtained by \eqref{eq:dv} may be seen as a member of the family of Monteiro-Zhang (MZ) directions\,\cite[Chapter~10]{wolkowicz2012handbook}.
Depending on the choice of $P$, generated directions admit different properties.
In particular, the following selections of $P$ and the correspondingly obtained directions are significant in the context of the LSDPs and NSDPs.
\begin{description}
\item[AHO direction ($P=I$):]
$\Delta V=\mu{F}(x)^{-1}-V-\sum_{i=1}^n\Delta x_i\mathcal{L}_{{F}(x)}^{-1}({F}_i\circ{V}).$
\item[HRVW/KSH/M direction ($P=F(x)^{-\frac{1}{2}}$ ):]
In this case, $\tilde{F}(x)=I$ and $\Delta V=\mu{F}(x)^{-1}-V-\left(F(x)^{-1}\left(\sum_{i=1}^n\Delta x_iF_i\right)V+V\left(\sum_{i=1}^n\Delta x_iF_i\right)F(x)^{-1}\right)/2$.
\item[NT direction ($P=W^{-\frac{1}{2}},\ W:=F(x)^{\frac{1}{2}}(F(x)^{\frac{1}{2}}VF(x)^{\frac{1}{2}})^{-\frac{1}{2}}F(x)^{\frac{1}{2}})$:]
In this case, $\tilde{F}(x)=\tilde{V}$ and
$\Delta {V}=\mu{F}(x)^{-1}-V-W^{-1}\left(\sum_{i=1}^n\Delta x_i{F}_i\right)W^{-1}$.
\end{description}
As for the HRVW/KSH/M and NT directions,
we should note that $\tilde{F}(x)$ and $\tilde{V}$ commute, namely, $\tilde{F}(x)\tilde{V}=\tilde{V}\tilde{F}(x)$.
Hereafter, we call the scaling matrices for making AHO, HRVW/KSH/M, and NT directions
AHO, HRVW/KSH/M, and NT matrices, respectively.
\subsection{Step size along $(\Delta x,\Delta V)$}
To find a step size $s\in (0,1]$ along the obtained search direction $\Delta W:=(\Delta x, \Delta V)$, we use an Armijo-like line search technique along with the merit function $\Phi_{\rho}:\R^n\times S^m\to \R$ defined {by}
\begin{equation}
\Phi_{\rho}(x,V): = \chi_{\rho}(x)+\nu\psi(x,V),\label{merit}
\end{equation}
where $\nu>0$ is a positive parameter, $\rho>0$ is a penalty parameter, and
\begin{align}
\chi_{\rho}(x)&:= f(x) - \mu\log\det F(x)+\rho\max_{\tau\in T}\left(g(x,\tau)\right)_+ +\rho\|Gx-h\|_{1}, \label{al:chi}\\
\psi(x,V)&:=F(x)\bullet V-\mu\log\det F(x)V.\notag
\end{align}
The function $\Phi_{\rho}$ is a straightforward extension of the primal-dual barrier merit function for getting the BKKT point of the NSDP\,\cite{yabe}.
The function $\psi(\cdot,\cdot)$ can be regarded as a merit function for the barrier shifted complementarity condition
$
F(x)\circ V=\mu I, F(x)\in S^m_+,\ V\in S^m_+.
$
Actually, when $F(x)\in S^m_+$ and $V\in S^m_+$, it holds that
\begin{equation}
\nabla \psi(x,V)=0
\Longleftrightarrow
F(x)\circ V=\mu I.
\end{equation}
A step size $s>0$ is then determined using the Armijo-like line search method, so that the value of $\Phi_{\rho}$ approximately decreases while keeping $F(x)\in S^m_{++}$ and $V\in S^m_{++}$:
More specifically, we first choose parameters $\alpha$, $\beta\in (0,1)$ arbitrarily.
Then, we find the smallest nonnegative integer $\ell$ such that
\begin{align*}
\Phi_{\rho}(x+\bar{s}\beta^{\ell}\dx,V+\bar{s}\beta^{\ell}\dV)&\\
&\hspace{-5em}\le \Phi_{\rho}(x,V)-\alpha \bar{s}\beta^{\ell}\left(\dx^{\top}B\dx+\nu {\psi}^{\prime}(x,V;\dx,\dV)\right)+\bar{s}\beta^{\ell}\rho\gamma,
\end{align*}
where
$B\in S^n_{++}$ is the matrix prescribed in SIQP\,\eqref{SIQP},
$\psi^{\prime}(x,V;\dx,\dV)$ denotes the directional derivative of function $\psi$ at $(x,V)$ in the direction $(\Delta x,\Delta V)$ and it is explicitly represented as
\begin{equation}\label{direct}
\psi^{\prime}(x,V;\Delta x,\Delta V)={\rm Tr}\left(\sum_{i=1}^n\Delta x_iF_iV+F(x)\Delta V-\mu F(x)^{-1}\sum_{i=1}^n\Delta x_iF_i-\mu V^{-1}\Delta V\right).
\end{equation}
Also, $\gamma>0$ is the constant prescribed in \eqref{al:opt}.
Moreover, $\bar{s}\in (0,1]$ is the initial step size chosen so that
\begin{equation}
F(x)+s\sum_{i=1}^n\Delta x_iF_i\in S^m_{++}\ \mbox{and }V+s\Delta V\in S^m_{++}\notag
\end{equation}
hold for any $s\in [0,\bar{s}]$.
For example, we set
\begin{equation}
\bar{s}=\min(s_x,s_V,1),\label{eq:s0}
\end{equation}
where
\begin{align*}
s_x&:=
\begin{cases}
\displaystyle{- \frac{\sigma}{\lambda_{\rm min}(F(x)^{-1}\sum_{i=1}^n\dx_iF_i)}} &\mbox{if }\lambda_{\rm min}(F(x)^{-1}\sum_{i=1}^n\dx_iF_i)<0\\
1 &\mbox{otherwise},
\end{cases}\\
s_V&:=
\begin{cases}
\displaystyle{- \frac{\sigma}{\lambda_{\rm min}(V^{-1}\Delta V)}} &\mbox{if }\lambda_{\rm min}(V^{-1}\Delta V)<0\\
1 &\mbox{otherwise}
\end{cases}
\end{align*}
with $\sigma\in (0,1)$ a positive parameter.
Furthermore, to ensure that the generated direction $\left(\dx,\dV\right)$ is a decent direction for $\Phi_{\rho}$,
the value of the penalty parameter $\rho$ must be chosen sufficiently large.
Specifically,
we set $\rho$ so that
\begin{equation*}
\rho>\max\left(\|y\|,\|z\|_{\infty}\right).
\end{equation*}
Now, we describe the algorithm for getting BKKT points.\vspace{1.0em}
\begin{center}
\underline{Algorithm~1}\vspace{1.0em}
\end{center}
\begin{description}
\item[Step~0 (Initialization):]
Set $(\ru[x][0],\ru[z][0],\rl[V][0])\in\R^n\times \R^s\times S^m_{++}$ and
$\ru[y][0]\in \mathcal{M}_+(T)$ such that $|{\rm supp}(y^0)|<\infty$.
Choose parameters
$
\alpha,\beta_1,\beta_2,\sigma\in (0,1)$, and $\delta, \nu, \rl[\rho][0], \rl[\gamma][0]>0.$
Set $r:=0$.
\item[Step~1 (Stopping condition):]
If
{$(x^r,y^r,z^r,V_r)$
satisfies the KKT conditions\,\eqref{e1}--\eqref{e4}},
then
stop the algorithm. Otherwise, go to Step~2.
\item[Step~2 (Select $P_{r}$ and $B_{r}$):]
Choose a nonsingular matrix $\rl[P][r]\in \R^{n\times n}$ and
positive definite matrix $\rl[B][r]\in S^n$.
\item[Step~3 (Generate $(\Delta x^{r}$, $\Delta V_{r}, y^{r+1},z^{r+1})$):]
Find a search direction $\rl[\dx][r]$,
Lagrange multiplier {measure} $\ru[y][r+1]\in\mathcal{M}_+(T)$
{such that $|{\rm supp}(y^{r+1})|<\infty$},
and vector $\ru[z][r+1]\in\R^s$ satisfying \eqref{al:opt} with
$x=x^r$, $B=B_r$, and $\gamma=\gamma_r$.
Compute $\rl[\Delta V][r]$ from \eqref{eq:dv} with $x=x^r$.
\item[Step~4 (Update $\rho_r$):] If
\begin{equation*}
\rl[\rho][r]>\max\left(\|\ru[y][r+1]\|,\ \|\ru[z][r+1]\|_{\infty}\right),
\end{equation*}
set $\rl[\rho][r+1]:=\rl[\rho][r]$. Otherwise, set
\begin{equation}
\rl[\rho][r+1]:=\delta+
\max\left(\|\ru[y][r+1]\|,\ \|\ru[z][r+1]\|_{\infty}\right).
\end{equation}
\item[Step~5 (Armijo-like line search):]
Compute $\bar{s}$ by \eqref{eq:s0} and let $\rl[s][r]=\bar{s}\beta_1^{\ell}$ with
the smallest nonnegative integer $\ell\ge 0$ satisfying
\begin{equation}\label{eq:armijo-rule}
\Phi_{\rl[\rho][r+1]}(\ru[x][r]+\rl[s][r]\ru[\dx][r],V_r+\rl[s][r]\rl[\dV][r])\le \Phi_{\rl[\rho][r+1]}(\ru[x][r],\rl[V][r])-\alpha \rl[s][r]\Delta\Phi_r\ +\rl[\rho][r+1]s_r\rl[\gamma][r],
\end{equation}
where $\Delta\Phi_r:=(\ru[\dx][r])^{\top}\rl[B][r]\ru[\dx][r]-\nu \psi^{\prime}(\ru[x][r],\rl[V][r];\ru[\dx][r],\rl[\dV][r])$.
\item[Step~6 (Update $x^{r}$, $V_{r}$ and $\gamma_{r}$):] Set
\begin{equation*}
(\ru[x][r+1],\rl[V][r+1]):=(\ru[x][r]+\rl[s][r]\ru[\dx][r],\rl[V][r]+\rl[s][r]\rl[\dV][r]),
\end{equation*}
$\gamma_{r+1}:=\beta_2\gamma_{r}$, and $r:=r+1$.
Return to Step~1.
\end{description}
\subsection{Choice of the coefficient matrix $B_r$}\label{subsec:Br}
In this section, we consider a particular choice of the coefficient matrix $B_r$ in SIQP\,\eqref{SIQP}.
In the conventional SQP, one of the basic selections for $B_r$ is the Hessian of the Lagrangian for
the SIPLOG\,\eqref{lsisdp}, i.e.,
\begin{equation}
\nabla_{xx}^2L(x^r,y^r,z^r)=\nabla^2f(x^r)+\int_T\nabla^2_{xx}g(x^r,\tau)dy^r(\tau)-\mu\nabla^2\log\det {F(x^r),}\label{eq:hessian}
\end{equation}
where
$L(x,y,z):=f(x)+\int_Tg(x,\tau)dy(\tau)+(Gx-h)^{\top}z-\mu\log\det F(x)$.
To explore other candidates for $B_r$,
we consider
the matrix valued function $B:\R^n\times \mathcal{M}(T)\times S^m_{++}\to S^n_{++}$ of the following form:
\begin{equation}
B(x,y,V):={M}(x,y)+{H}_{P}(x,V),\label{eq:0913}
\end{equation}
where ${M}: \R^n\times \mathcal{M}(T)\to S^n$
is some positive definite matrix valued function and ${H}_P:\R^n\times S^m_{++}\to S^n$
is defined by
\begin{equation}
\left({H}_P(x,V)\right)_{i,j}:=\frac{\tilde{F}_i\bullet\left(\mathcal{L}_{\tilde{F}(x)}^{-1}\mathcal{L}_{\tilde{V}}
+\mathcal{L}_{\tilde{V}}\mathcal{L}_{\tilde{F}(x)}^{-1}
\right)\tilde{F}_j}{2}\label{eq:HP}
\end{equation}
for $i,j=1,2,\ldots,n$ and $(x,V)\in \R^n\times S^m$. Recall here that $\tilde{F}(x)$ and $\tilde{V}$ are positive definite matrices
obtained by scaling $F(x)$ and $V$ with the matrix $P$. See \eqref{scal}.
When $\tilde{F}(x)$ and $\tilde{V}$ commute,
{so} do $\mathcal{L}^{-1}_{\tilde{F}(x)}$ and
$\mathcal{L}_{\tilde{V}}$.
Then,
by noting \eqref{eq:0326},
the KKT conditions\,\eqref{al:opt}
of SIQP\,\eqref{SIQP} with $B=B(x,y,V)$ can be represented in terms of $\Delta\tilde{V}$ as
\begin{align*}\label{al:opt2}
&M(x,y)\Delta x+\nabla f(x)-
(\tilde{F}_i\bullet (\tilde{V}+\Delta\tilde{V}))_{i=1}^n
+\int_T\nabla_xg(x,\tau)d(y+\Delta y)(\tau)+G^{\top}(z+\Delta z)=0,\notag \\
&\tilde{F}(x)\circ \tilde{V}+\tilde{F}(x)\circ \Delta \tilde{V}
+\sum_{i=1}^n\Delta x_i\tilde{F}_i\circ\tilde{V} =\mu I,\notag\\
&g(x,\tau)+\nabla_xg(x,\tau)^{\top}\Delta x\le 0\ \ (\tau\in {\rm supp}(y+\Delta y)),\\
&\int_T\left(g(x,\tau)+\nabla_xg(x,\tau)^{\top}\Delta x\right)d(y+\Delta y)(\tau)=0,\ y+\Delta y\in\mathcal{M}_+(T),\ G(x+\Delta x)=h,\notag \\
&\max_{\tau\in T}\left(
g(x,\tau)+\nabla_xg(x,\tau)^{\top}\Delta x
\right)_+\le 0.\notag
\end{align*}
Actually, by substituting $\tilde{V}+\Delta\tilde{V}=\mathcal{L}_{\tilde{F}(x)}^{-1}(\mu I - \sum_{i=1}^n\Delta x_i\tilde{F}_i\circ\tilde{V})$, which is obtained from the above second condition, into the first condition, we can regain the KKT conditions\,\eqref{al:opt} with $B=B(x,y,V)$.
Furthermore, if the function $M(x,y)$ is the Hessian of the function
\begin{equation}
L_2(x,y):=f(x)+\int_Tg(x,\tau)dy(\tau),\label{eq:0615-0916}
\end{equation}
namely,
$\nabla^2_{xx}L_2(x,y) = \nabla^2f(x)+\int_T\nabla^2_{xx}g(x,\tau)dy(\tau)$, then solving the above system is regarded as the scaled Newton iteration for the KKT conditions.
From these observations, we can expect that employing such $B(x^r,y^r,V_r)$ as $B_r$ accomplishes a rapid convergence.
{For reference, we list the formulas of $({H}_P(x,V))_{ij}$
below for the case where the HRVW/KSH/M and NT matrices are selected as the scaling matrix $P$.
\begin{description}
\item[HRVW/KSH/M matrix:] $({H}_P(x,V))_{ij}={\rm Tr}\left(F(x)^{-1}F_jVF_i\right)$,
\item[NT matrix:] $({H}_P(x,V))_{ij}={\rm Tr}\left(W^{-1}F_iW^{-1}F_j\right)$ with $W=F(x)^{\frac{1}{2}}(F(x)^{\frac{1}{2}}VF(x)^{\frac{1}{2}})^{-\frac{1}{2}}F(x)^{\frac{1}{2}}$.
\end{description}}
\subsection{Convergence analysis of Algorithm~1}
In the subsequent analysis,
we make the following assumptions:
\begin{description}
\item{\bf Assumption~A:}
\end{description}
\begin{enumerate}
\item\label{B2}
The scaling matrix $P_r$ is the HRVW/KSH/M or NT matrix for any $r=0,1,2,\ldots$.
\item\label{B3} The sequence of penalty parameters $\{\rho_{r}\}$ is bounded.
\item\label{B4} The generated sequence $\{x^{r}\}$ is bounded.
\item\label{B5} The initial point $\ru[x][0]$ is chosen to satisfy $G\ru[x][0]=h$.
\end{enumerate}
Although similar assumptions are made in many existing works on interior point methods for nonlinear programs or NSDPs,
Assumption~A-\ref{B3} may seem rather strong{.}
To relax it is one of future subjects that should be settled.
As for Assumption~A-\ref{B4}, we can show its validity under a certain hypothesis; see Proposition~\ref{bound_x} below.
Assumption A-\ref{B5} is made for simplicity of expression in some subsequent proofs.
The proofs can be extended straightforwardly to the general case {where} Assumption~A-\ref{B5} {is absent}.
\begin{Prop}\label{bound_x}
Suppose that Assumption~A-\ref{B5} hold{s} and $g(x,\tau)$
is an affine function with respect to $x$, i.e., $g(x,\tau)=a(\tau)^{\top}x-b(\tau)$,
where $a:T\to\R^n$ and $b:T\to \R$ are continuous functions.
Furthermore, assume that the feasible set of the SISDP\,\eqref{lsisdp} is nonempty and compact.
Then, the generated sequence $\{\ru[x][r]\}$ is bounded.
\end{Prop}
\begin{proof}
We can show that for each $r\ge 0$
\begin{equation*}
F(x^{r})\in S^m_{++},\ a(\tau)^{\top}x^{r}-b(\tau)\le \delta_0\ \ (\tau\in T),\ Gx^{r}=h,
\end{equation*}
where $\delta_0:=\max\left(\max_{\tau\in T}\left(a(\tau)^{\top}x_0-b(\tau)\right),\gamma_0\right)$.
We prove only the second relation by mathematical induction.
It holds for $r=0$ obviously. Next, suppose that it holds true for some $r>0$. Then, by noting $\max_{\tau\in T}\left(a(\tau)^{\top}(x^r+\Delta x^r)-b(\tau)\right)\le \gamma_r\le \gamma_0$ and $s_r\in (0,1]$, we have
\begin{align*}
a(\tau)^{\top}x^{r+1}-b(\tau)&=a(\tau)^{\top}(x^r+s_r\Delta x^r)-b(\tau)\\
&\le s_r\gamma_0+(1-s_r)(a(\tau)^{\top}x^r-b(\tau))\\
&\le \max\left(a(\tau)^{\top}x^r-b(\tau),\gamma_0\right)\\
&\le \delta_0
\end{align*}
for any $\tau\in T$. Therefore, we ensure the targeted inequality holds for all $r\ge 0$.
Denote the feasible set of the SISDP\,\eqref{lsisdp} by $\mathcal{F}$ and
define a proper closed convex function ${\pi}:\R^n\to \R$ by
$$
{\pi}(x):={\max}\left(-\lambda_{\min}(F(x)),\ \max_{\tau\in T}a(\tau)^{\top}x-b(\tau),\ \|Gx-h\|\right).
$$
Since the level set $\{x\in \R^n \mid {\pi}(x)\le 0\}(=\mathcal{F})$ is compact from the assumption that $\mathcal{F}$ is nonempty and compact, any level set $\{x\in \R^n \mid {\pi}(x)\le \eta\}$ with $\eta>0$ is also compact. Then,
we can see that
$\{x^r\}\subseteq \{x\in \R^n \mid {\pi}(x)\le \delta_0\}$ and thus $\{x^k\}$ is bounded.
\end{proof}
Now, we enter the essential part of the global convergence of the algorithm.
The following lemmas play key roles in establishing the well-definedness of the Armijo-like linesearch in Step~5.
\begin{Lem}\label{lem:0819}
For any $x\in \R^n$ with $F(x)\in S^m_{++}$ and any $V\in S^m_{++}$, it holds that
\begin{equation*}
\psi(x,V)\ge m\mu (1 -\log \mu),
\end{equation*}
where the equality holds if and only if $F(x)V=\mu I$.
\end{Lem}
\begin{proof}
Denote the eigenvalues of $F(x)V$ by $\lambda_i>0\ (i=1,2,\ldots,m)$.
Then,
$\psi(x,V)={\rm Tr}(F(x)V)-\mu\log\det F(x)V=\sum_{i=1}^m(\lambda_i-\mu\log\lambda_i)\ge m\mu(1-\log\mu)$.
The equality holds if and only if $\lambda_1=\lambda_2=\cdots=\lambda_m=\mu$, that is, $F(x)V=\mu I$.
\end{proof}
\begin{Lem}\label{lem_bp}
\begin{enumerate}
\item\label{lem_bp1}
It holds that
\begin{equation}
\psi^{\prime}(x^{r},V_{r};\Delta x^{r},\Delta V_{r})\le 0.\label{eq:0326-1}
\end{equation}
In particular, the equality holds if and only if $F(x^r)V_r=\mu I$. \vspace{0.5em}\\
\item
Let
$\theta(x):=\max_{\tau\in T}\left(g(x,\tau)\right)_+$
and $\theta^{\prime}(x;\Delta x)$ be
the directional derivative of $\theta$ at $x$ in the direction $\dx$.
Then, $\theta(\ru[x][r])+\theta^{\prime}(\ru[x][r];\ru[\dx][r])\le \rl[\gamma][r]$ holds.
\end{enumerate}
\end{Lem}
\begin{proof}
{\rm 1.} Although the proof can be given in a fashion similar to \cite[Lemma~3]{yabe}, we show it here for completeness.
We first prove $\psi^{\prime}(\ru[x][r],\rl[V][r];\ru[\dx][r],\rl[\dV][r])\le 0$.
Note first that
\begin{equation}
\mu{\rm Tr}(\tilde{F}(x^r)^{-1}\sum_{i=1}^n\Delta x_i^r\tilde{F}_i+\tilde{V}_r^{-1}\Delta\tilde{V}_r)={\rm Tr}(\mu^2\tilde{F}(x^r)^{-1}\tilde{V}_r^{-1}-\mu I),\label{eq:0601}
\end{equation}
which is implied by the scaled Newton equation $\tilde{F}(\ru[x][r])\circ\rl[\tilde{V}][r]+\sum_{i=1}^n\ru[\dx_i][r]\tilde{F}_i\circ\rl[\tilde{V}][r]+\tilde{F}(\ru[x][r])\circ\rl[\tilde{\dV}][r]=\mu I$ together with $\tilde{F}(x^r)\tilde{V}_r=\tilde{V_r}\tilde{F}(x^r)$.\\
Then, using $\tilde{F}(x^r)\tilde{V}_r=\tilde{V}_r\tilde{F}(x^r)$ again,
we have from \eqref{direct}
\begin{align}
&\hspace{1em}\psi^{\prime}(\ru[x][r],\rl[V][r];\ru[\dx][r],\rl[\dV][r])\notag \\
&={\rm Tr}\left(\sum_{i=1}^n\ru[\dx_i][r]F_i\rl[V][r]+F(\ru[x][r])\rl[\dV][r]-\mu F(\ru[x][r])^{-1}\sum_{i=1}^n\ru[\dx_i][r]F_i-\mu \rl[V][r]^{-1}\rl[\dV][r]\right)\notag \\
&={\rm Tr}\left(\sum_{i=1}^n\ru[\dx_i][r]\tilde{F}_i\rl[\tilde{V}][r]+\tilde{F}(\ru[x][r])\rl[\Delta\tilde{V}][r]-\mu \tilde{F}(\ru[x][r])^{-1}\sum_{i=1}^n\ru[\dx_i][r]\tilde{F}_i-\mu \rl[\tilde{V}][r]^{-1}\Delta \tilde{V}_r\right)\notag \\
&={\rm Tr}\left(\mu I-\tilde{F}(\ru[x][r])\rl[\tilde{V}][r]-\mu \tilde{F}(\ru[x][r])^{-1}\sum_{i=1}^n\ru[\dx_i][r]\tilde{F}_i-\mu \rl[\tilde{V}][r]^{-1}\rl[\Delta\tilde{V}][r]\right)\notag\\
&={\rm Tr}\left(2\mu I-\tilde{F}(\ru[x][r])\rl[\tilde{V}][r]-\mu^2\rl[\tilde{V}][r]^{-1}\tilde{F}(x^r)^{-1}\right)\notag\\
&=-\left\|\mu\rl[\tilde{V}][r]^{-\frac{1}{2}}\tilde{F}(\ru[x][r])^{-\frac{1}{2}}-\tilde{F}(\ru[x][r])^{\frac{1}{2}}\rl[\tilde{V}][r]^{\frac{1}{2}}\right\|_F^2\label{al:0819-direct}\\
&\le 0,
\end{align}
where the third equality holds because
$$
{\rm Tr}\left(
\mu I -\tilde{F}(x^r)\tilde{V}_r
\right)
={\rm Tr}\left(
\sum_{i=1}^n\Delta x_i^r\tilde{F}_i\tilde{V}_r+\Delta\tilde{V}_r\tilde{F}(x^r)\right)
$$
from the scaled Newton equations,
the fourth equality follows from \eqref{eq:0601}.
Thus, we get \eqref{eq:0326-1}.
By \eqref{al:0819-direct}, we observe that
$\psi^{\prime}(x,V;x^r,V_r)=0$ if and only if
$
\mu\rl[\tilde{V}][r]^{-\frac{1}{2}}\tilde{F}(\ru[x][r])^{-\frac{1}{2}}-\tilde{F}(\ru[x][r])^{\frac{1}{2}}\rl[\tilde{V}][r]^{\frac{1}{2}}=O,
$
whch is equivalent to $\tilde{F}(x^r)\tilde{V}_r=\mu I$, i.e.,
${F}(x^r){V}_r=\mu I$.
We therefore obtain the latter claim. \vspace{0.5em}\\
{\rm 2.} Let $T(x):={\rm argmax}_{\tau\in T}g(x,\tau)$.
Consider the three cases where the value of $\max_{\tau\in T}g(\ru[x][r],\tau)$ is (i) $<0$, (ii) $>0$, and (iii) {$=0$}:
\begin{enumerate}
\item[(i)] In this case, $\theta(\ru[x][r])=0$ and $\theta^{\prime}(\ru[x][r];\ru[\dx][r])=0$ holds.
Then $\theta(\ru[x][r])+\theta^{\prime}(\ru[x][r];\ru[\dx][r])=0<\rl[\gamma][r]$ readily follows.
\item[(ii)] In this case, we have
$\theta(\ru[x][r])=\max_{\tau\in T(\ru[x][r])}g(\ru[x][r],\tau)>0$.
We then get $\theta^{\prime}(\ru[x][r];\ru[\dx][r])=\max_{\tau\in T(\ru[x][r])}\nabla_xg(\ru[x][r],\tau)^{\top}\ru[\dx][r]$ and therefore
\begin{align*}
\theta(\ru[x][r])+\theta^{\prime}(\ru[x][r];\ru[\dx][r])
&=\max_{\tau\in T(\ru[x][r])}\left(\theta(x^r)+\nabla_xg(\ru[x][r],\tau)^{\top}\ru[\dx][r]\right)\\
&=\max_{\tau\in T(\ru[x][r])}
\left(g(\ru[x][r],\tau)+\nabla_xg(\ru[x][r],\tau)^{\top}\ru[\dx][r]\right)\le\rl[\gamma][r].
\end{align*}
\item[(iii)] In this case, we have $\theta(\ru[x][r])=g(x^r,\tau)=0$
for each $\tau\in T(x^r)$ and $\theta^{\prime}(\ru[x][r];\ru[\dx][r])=\max_{\tau\in T(\ru[x][r])}\left(\nabla_xg(\ru[x][r],\tau)^{\top}\ru[\dx][r]\right)_+$.
Then,
$$
\theta(\ru[x][r])+\theta^{\prime}(\ru[x][r];\ru[\dx][r])=\theta^{\prime}(\ru[x][r];\ru[\dx][r])=\max_{\tau\in T(\ru[x][r])}\left(g(\ru[x][r],\tau)+\nabla_xg(\ru[x][r],\tau)^{\top}\ru[\dx][r]\right)_+\le\rl[\gamma][r].
$$
\end{enumerate}
We have obtained the desired conclusion.
\end{proof}
\begin{Lem}\label{lem:0820}
Suppose $\Delta x=0$. Then, ${\psi^{\prime}}(x,V;\Delta x,\Delta V)=0$ implies $\Delta V=O$.
\end{Lem}
\begin{proof}
From Lemma\,\ref{lem_bp}\eqref{lem_bp1} and ${\psi^{\prime}}(x,V;\Delta x,\Delta V)=0$,
we have $F(x)V=\mu I$, i.e., $\tilde{F}(x)\tilde{V}=\mu I$.
This together with $\Delta x=0$ implies that the scaled Newton equations\,\eqref{eq:scaled_new}
yield $\tilde{F}(x)\circ \Delta \tilde{V}=O$, i.e., $\mathcal{L}_{\tilde{F}(x)}\Delta\tilde{V}=O$.
It then follows that $\Delta\tilde{V}=O$ since
$\mathcal{L}_{\tilde{F}(x)}$ is invertible by $\tilde{F}(x)\in S^m_{++}$. Consequently, we have $\Delta V=O$.
\end{proof}
\begin{Prop}\label{Prop:0329}
For any sufficiently small $s\in (0,1]$,
we have
\begin{equation}
\Phi_{\rho_r}(\ru[x][r]+s\ru[\dx][r],\rl[V][r]+s\rl[\dV][r])\le \Phi_{\rho_r}(
\ru[x][r],\rl[V][r])-s\alpha (\Delta \ru[x][r])^{\top}B_{r}\Delta \ru[x][r]+\Delta\psi+
s\rho_r \rl[\gamma][r],
\end{equation}
where
$
\Delta\psi:=\psi^{\prime}(\ru[x][r],\rl[V][r];\Delta \ru[x][r],\Delta \rl[V][r]).
$
\end{Prop}
\begin{proof}
If $\left(\ru[\dx][r],\rl[\dV][r]\right)=(0,O)$, then $\psi^{\prime}(x^r,V_r;\Delta x^r,\Delta V_r)=0$ holds and the desired conclusion obviously holds for any $s>0$. So, we provide with the proof
by assuming $\left(\ru[\dx][r],\rl[\dV][r]\right)\neq (0,O)$.
For simplicity of expression, we abbreviate $\rho_r$, $x^{r}$, $\Delta x^{r}$, and $B_{r}$ as $\rho$, $x$, $\Delta x$, and $B$, respectively.
In addition, we write ${\rm supp}(y)=\{\tau_1,\tau_2,\ldots,\tau_p\}$ and represent $y(\tau_i)$ as $y_i$ for each $i$.
Note that
$\|y\|=\sum_{i=1}^py_i$ by $y\in \mathcal{M}_+(T)$.
Let $f_{\rm bp}(x):=f(x)-\mu\log\det F(x)$ and
$\theta(x):=\max_{\tau\in T}\left(g(x,\tau)\right)_+$.
Moreover, define $\Delta\psi:=\psi^{\prime}(x,V;\Delta x,\Delta V)$, $w:=(x,V)$, and $\Delta w:=(\dx,\dV)$.
We then have
{\begin{align}
\Phi_{\rho}(w+s\Delta w)-\Phi_{\rho}(w)&=s\nabla f_{\rm bp}(x)^{\top}\Delta x +s\nu\Delta\psi+
s\rho\theta^{\prime}(x;\dx)+o(s)\notag \\
&=-s\Delta x^{\top}B\Delta x
+
s\sum_{i=1}^pg(x,\tau_i)y_i
+s\rho\theta^{\prime}(x;\dx)+s\nu\Delta\psi+o(s)\notag \\
&\le -s\Delta x^{\top}B\Delta x+s\left(\sum_{i=1}^py_i\right)\theta(x)+
s\rho\theta^{\prime}(x;\dx)+s\nu\Delta\psi+o(s)\notag\\
&\le -s\Delta x^{\top}B\Delta x+s\rho(\theta(x)+\theta^{\prime}(x;\dx))+s\nu\Delta\psi+o(s)\notag\\
&\le -s\Delta x^{\top}B\Delta x+s\rho\gamma+s\nu\Delta\psi+o(s),\label{a3}
\end{align}}
where the second equality follows from the KKT conditions
\begin{align}
&B\Delta x+\nabla f_{\rm bp}(x)+\sum_{i=1}^p\nabla_xg(x,\tau_i)y_i+G^{\top}z=0,\ G\Delta x = 0,\notag\\
&0\ge g(x,\tau_i)+\nabla_xg(x,\tau_i)^{\top}\Delta x,\
y_i\ge 0,\ \left(g(x,\tau_i)+\nabla_xg(x,\tau_i)^{\top}\Delta x\right)y_i=0\ (i=1,2,\ldots,p),\notag
\end{align}
and the first inequality follows from $y_i\ge 0$ for $i=1,2,\ldots,p$ and $g(x,\tau)\le \theta(x)\ (\tau\in T)$.
Moreover, the second inequality is obtained from $\rho>\|y\|$ and $\theta(x)\ge 0$, and the last inequality is due to Proposition\,\ref{Prop:0329}.
We hence obtain
\begin{align*}
\Phi_{\rho}(w+s\Delta w)-\left(\Phi_{\rho}(w)+\rho s\gamma\right)&\le -s\Delta x^{\top}B\Delta x+s\nu\Delta\psi+o(s).
\end{align*}
Now, since we can show $-\Delta x^{\top}B\Delta x+\nu\Delta\psi<0$ whenever $(\dx,\dV)\neq (0,O)$,
it holds that
\begin{equation}
\Phi_{\rho}(w+s\Delta w)\le \Phi_{\rho}(w)+\alpha s\left(-\Delta x^{\top}B\Delta x+\nu\Delta\psi\right)+\rho s\gamma \notag
\end{equation}
for all sufficiently small $s>0$.
The proof is complete.
\end{proof}
\begin{Prop}\label{penal}
Suppose that Assumption~A-\ref{B3} holds.
There exist some $\bar{\rho}>0$ and $\bar{r}>0$
such that $\rho_r=\bar{\rho}$ for any $r\ge \bar{r}$.
\end{Prop}
\begin{proof}
The proof is easily obtained, and so omitted.
\end{proof}
Let $\bar{r}>0$ and $\bar{\rho}>0$ be as given in the above proposition.
The following proposition shows that the produced sequences are bounded.
\begin{Prop}\label{property_bound}
Suppose that Assumption~A holds.
Then, we have the following:
\begin{enumerate}
\item\label{prop:bound}
$\left\{\Phi_{{\rho}_r}(x^r,V_r)\right\}$ is bounded from above.
\item\label{rnum:infF} $\liminf_{r\to \infty}\det F(x^r)>0$ and $\liminf_{r\to \infty}\det V_r>0$.
\item\label{prop:0820-w}
$\{\xi(\ru[x][r])\}$ with $\xi(\cdot)$ defined by \eqref{eq:wx} is bounded.
\item\label{bound_v} $\{(y^r,z^r,V_r)\}$ is bounded.
\item\label{pr_bound} $\{\rl[P][r]\}$ and $\{\rl[P][r]^{-1}\}$ are bounded.
\end{enumerate}
\end{Prop}
\begin{proof}
{\rm 1.} By the linesearch {rule},
we have, for $r\ge \bar{r}$,
\begin{align*}
\Phi_{{\bar{\rho}}}(x^r,V_r)&\le \Phi_{{\bar{\rho}}}(x^{r-1},V_{r-1})+\bar{\rho}\gamma_{r-1}\\
&= \Phi_{{\bar{\rho}}}(x^{\bar{r}},V_{\bar{r}})+\bar{\rho}\gamma_{\bar{r}}\sum_{i=\bar{r}}^r\beta_2^{i-\bar{r}}\\
&\le \Phi_{{\bar{\rho}}}(x^{\bar{r}},V_{\bar{r}})+\bar{\rho}\gamma_{\bar{r}}\sum_{i=\bar{r}}^{\infty}\beta_2^{i-\bar{r}}\\
&<\infty,
\end{align*}
where
the last inequality follows from $0<\beta_2<1$.
We thus get the boundedness of $\left\{\Phi_{\bar{\rho}}(x^r,V_r)\right\}$. \vspace{0.5em} \\
{\rm 2.} We first prove $\liminf_{r\to \infty}\det F(x^r)>0$. Assume to the contrary that
$\liminf_{r\to \infty}\det F(x^r)=0$.
Notice that $\{\psi(x^r,V_r)\}$ is bounded from below by Lemma\,\ref{lem:0819}.
Also, notice that ${\chi}_{\bar{\rho}}(x^r)\to \infty$ as $r\to \infty$, where $\chi_{\bar{\rho}}(\cdot)$
is defined in \eqref{al:chi} with $\rho=\bar{\rho}$. This is because $\{x^r\}$ is bounded by assumption.
We then see that $\Phi_{\bar{\rho}}(x^r,V_r)\to \infty$ as $r\to \infty$, which contradicts item-\ref{prop:bound}.
Hence we obtain $\liminf_{r\to \infty}\det F(x^r)>0$.
We next show $\liminf_{r\to \infty}\det V_r>0$.
For contradiction,
we assume without loss of generality {that} $\lim_{r\to\infty}\det V_r=0$. Notice that $\{-\mu\log\det F(x^r)\}$ is bounded as
$\{x^r\}$ is bounded and $\liminf_{r\to \infty}\det F(x^r)>0$.
In addition, $F(x^r)\bullet V_r>0$ follows from $F(x^r),V_r\in S^m_{++}$.
In view of these facts, we have {$\lim_{r\to\infty}\Phi_{\bar{\rho}}(x^r,V_r)=\infty$}. This contradicts item-\ref{prop:bound} again.
Hence, we conclude $\liminf_{r\to \infty}\det V_r>0$. \vspace{0.5em}\\
{\rm 3.} Note that $\liminf_{r\to \infty}\det F(x^r)>0$ by item-\ref{rnum:infF} and $\{F(x^r)\}\subseteq S^m_{++}$
is bounded since $\{x^r\}$ is bounded. Then,
$\{F(x^r)^{-1}\}$ is also bounded and
there exists some $M>0$ such that
$\|F(x^r)^{-1}\|_F\le M$ for any $r\ge 0$.
We then have
\begin{equation*}
\|\xi(x^r)\|\le \sqrt{\sum_{i=1}^n\|F(x^r)^{-1}\|_F^2\|F_i\|_F^2}
=\|F(x^r)^{-1}\|_F\sqrt{\sum_{i=1}^n\|F_i\|_F^2}\le M\sqrt{\sum_{i=1}^n\|F_i\|_F^2}
\end{equation*}
for all $r\ge 0$.
We thus obtain the desired result. \vspace{0.5em}\\
{\rm 4.} The boundedness of $\{\left(y^r,z^r\right)\}$ can be obtained from the boundedness of penalty parameters {$\rho_r$} (Assumption~A-\ref{B3}). We have only to show the boundedness of $\{V_r\}$.
For contradiction, suppose that $\{V_r\}$ is unbounded.
We may assume
without loss of generality that
$\|V_r\|\to \infty$ as $r\to\infty$.
Let $X_r:=F(x^r)^{-\frac{1}{2}}V_rF(x^r)^{-\frac{1}{2}}$.
Denote
the eigenvalues of $X_r\in S^m_{++}$ by $0<\lambda_1^r\le\lambda_2^r\le \cdots\le\lambda_m^r$.
Then, by the positive definiteness of $X_r$ and the boundedness of $F(x^r)^{-\frac{1}{2}}$
derived from item-\ref{rnum:infF}, it follows that $\lim_{r\to \infty}\lambda^r_m=\infty$.
We then obtain
\begin{align}
\psi(x^r,V_r)&= {\rm Tr}(X_r)-\mu\log\det X_r\notag \\
& = \sum_{i=1}^m\left(\lambda_i^r-\mu\log\lambda_i^r\right)\rightarrow \infty \ (r\to \infty),\notag
\end{align}
which together with the boundedness of $\{{\chi}_{\rho}(x^r)\}$ implies the unboundedness of $\{\Phi_{\rho}(x^r,V_r)\}$.
This contradicts item-\ref{prop:bound}.
We thus conclude that $\{V_r\}$ is bounded.\vspace{0.5em}\\
{\rm 5.} Since $\liminf_{r\to\infty}\det \rl[V][r]>0$ by item-\ref{rnum:infF}, we see that $\{\rl[V][r]^{-1}\}$ is bounded.
By Assumption~A-\ref{B2}, $\rl[P][r]$ is set to be $\rl[P][r]=F(\ru[x][r])^{-\frac{1}{2}}$ or $\rl[P][r]=W_r^{-\frac{1}{2}}$ with
$$
W_r=F(\ru[x][r])^{\frac{1}{2}}(F(\ru[x][r])^{\frac{1}{2}}\rl[V][r]F(\ru[x][r])^{\frac{1}{2}})^{-\frac{1}{2}}F(\ru[x][r])^{\frac{1}{2}}.
$$
Using these facts, it is not difficult to verify that $\{\rl[P][r]\}$ and $\{\rl[P][r]^{-1}\}$ are bounded.
\end{proof}
Below, we additionally impose the following assumption on the sequence $\{B_r\}$:
\begin{description}
\item[Assumption~B:]\label{B1}$\{B_{r}\}\subseteq S^n_{++}$ is a bounded sequence such that
$$
\eta_2 I \succeq B_{r} \succeq \eta_1 I,\ \ r=0,1,2,\ldots
$$
for some $\eta_2>\eta_1>0$.
\end{description}
Assumption~B holds true if we choose the identity matrix as $B_r$ for all $r\ge 0$.
However, it is not obvious when the sequence $\{B_r\}$ suggested in Section~\ref{subsec:Br} becomes bounded.
In the next proposition, we prove that $\{B_r\}$ is bounded if $f$ is convex and Assumption~A holds.
\begin{Prop}
Suppose that $f$ and $g(\cdot,\tau)\ (\tau\in T)$ are twice continuously differentiable convex functions and Assumption~A holds.
Further, assume that the matrices $F_i\ (i=1,2,\ldots,n)$ are linearly independent in $S^m$.
Then, the sequence of matrices $B_r$ defined by either of the following formulas satisfies Assumption~B:
\begin{enumerate}
\item $B_r:=\nabla_{xx}^2L(x^r,y^r,z^r)$ for any $r\ge 0$,
\item $B_r:=B(x^r,y^r,V_r)$ for any $r\ge 0$,
\end{enumerate}
where $\nabla^2_{xx}L(x,y,z)$ and $B(x,y,V)$ are defined by \eqref{eq:hessian} and \eqref{eq:0913}, respectively.
\end{Prop}
\begin{proof}
Denote $M_r:=\nabla^2f(x^r)+\int_T\nabla^2_{xx}g(x^r,\tau)dy^r(\tau)$ for any $r\ge 0$.
Note that $M_r\in S^m_{+}$ follows for each $r\ge 0$ from the convexity of $f$ and $g(\cdot,\tau)\ (\tau\in T)$
together with $y^r\in\M_+(T)$, and moreover $\{M_r\}$ is bounded, since $\{(x^r,y^r)\}$ is bounded by Proposition\,\ref{property_bound}. \vspace{0.5em}\\
{\rm 1.} Let $H_r:=-\nabla^2\log\det F(x^r)$ and denote $d{F}:=\sum_{i=1}^nd_iF_i$ for $d\in \R^n$.
Then, the $(i,j)$-th entry of $H_r$ is represented as
$
F_i\bullet F(x^r)^{-2}F_j
$
for $1\le i,j\le m$, and we have
\begin{equation*}
d^{\top}H_rd=\sum_{1\le i,j\le m}d_iF_i\bullet F(x^r)^{-2}d_jF_j=dF\bullet F(x^r)^{-2}dF.
\end{equation*}
By the boundedness of $\{F(x^r)^{-1}\}\subseteq S^m_{++}$ and $\{F(x^r)\}\subseteq S^m_{++}$
from Proposition\,\ref{property_bound}
together with the linear independence of $F_1,F_2,\ldots,F_n$,
there exist some $c_1,c_2>0$ such that
$$
c_1\le d^{\top}H_rd\le c_2
$$
for any $r\ge 0$ and $d\in \R^n$ with $\|d\|=1$.
Therefore, $\{H_r\}$ is uniformly positive definite and bounded.
Since $\nabla^2_{xx}L(x^r,y^r,z^r)=M_r+\mu H_r$, the sequence $\left\{\nabla^2L(x^r,y^r,z^r)\right\}$ satisfies Assumption~B.\vspace{0.5em} \\
{\rm 2.} Let $d\tilde{F}:=\sum_{i=1}^nd_i\tilde{F}_i$ for $d\in \R^n$.
Recall that ${H}_P(x,V)$ is defined by \eqref{eq:HP} for $P\in \R^{m\times m}$.
It then holds that
\begin{align}
d^{\top}{H}_{P_r}(x^r,V_r)d
&=\sum_{1\le i,j\le m}d_i\tilde{F}_i\mathcal{L}^{-1}_{\tilde{F}(x^r)}\mathcal{L}_{\tilde{V_r}}\tilde{F}_jd_j\notag \\
&=d\tilde{F}\bullet \mathcal{L}^{-1}_{\tilde{F}(x^r)}\mathcal{L}_{\tilde{V}}d\tilde{F}\notag \\
&=\mathcal{L}^{-1}_{\tilde{F}(x^r)}d\tilde{F}\bullet \mathcal{L}_{\tilde{V}_r}\mathcal{L}_{\tilde{F}(x^r)}\mathcal{L}^{-1}_{\tilde{F}(x^r)}d\tilde{F}\notag \\
&=\mathcal{L}^{-1}_{\tilde{F}(x^r)}d\tilde{F}\bullet(\tilde{F}(x^r)\circ \tilde{V}_r)\mathcal{L}^{-1}_{\tilde{F}(x^r)}d\tilde{F}\label{al:0916},
\end{align}
where the third equality is due to the symmetry of the linear operator $\mathcal{L}_{\tilde{F}(x^r)}$ and the last one holds since $\tilde{F}(x^r)$ and $\tilde{V}_r$ commute.
When the scaling matrix $P_r$ is the NT matrix,
we obtain
\begin{equation}
\eqref{al:0916}
=d\tilde{F}\bullet \tilde{V}_rd\tilde{F}\label{eq:0916-1}
\end{equation}
from $\tilde{F}(x^r)=I$.
On the other hand, when $P_r$ is the HRVW/KSH/M matrix, we get
\begin{equation}
\eqref{al:0916}
=
\mathcal{L}^{-1}_{\tilde{F}(x^r)}d\tilde{F}\bullet
\tilde{F}(x^r)^2\mathcal{L}^{-1}_{\tilde{F}(x^r)}d\tilde{F}
\label{eq:0916-2}
\end{equation}
since $\tilde{F}(x^r)=\tilde{V}_r$.
By noting the boundedness of $\{P_r\}$ and $\{P_r^{-1}\}$ from Proposition\,\ref{property_bound},
the sequences
$\{\tilde{F}(x^r)\}$, $\{\tilde{F}(x^r)^{-1}\}$, and $\{\tilde{V}_r\}$ are bounded.
Using these facts together with the linear independence of $F_1,F_2,\ldots,F_n$,
the above expressions \eqref{eq:0916-1} and \eqref{eq:0916-2} yield that there exist some $c_3,c_4>0$ such that
$$
c_3\le d^{\top}{H}_{P_r}(x^r,V_r)d\le c_4
$$
for any $r\ge 0$ and $d\in \R^n$ with $\|d\|=1$.
Therefore, $\{{H}_{P_r}(x^r,V_r)\}$ is uniformly positive definite and bounded.
Since $B(x^r,y^r,V_r)=M_r+{H}_{P_r}(x^r,V_r)$, the sequence $\left\{B(x^r,y^r,V_r)\right\}$ satisfies Assumption~B.
\end{proof}
We next present the following proposition concerning $\{(\Delta x^r,\Delta V_r)\}$.
\begin{Prop}\label{prop:bd:dx}
Suppose that Assumptions~A and B hold.
Then, we have the following:
\begin{enumerate}
\item\label{wpx} $\{\Delta x^r\}$ is bounded;
\item $\{\Delta V_r\}$ is bounded.
\end{enumerate}
\end{Prop}
\begin{proof}
{\rm 1.} Let $x_{\rm f}\in \R^n$ be a feasible point for \eqref{lsisdp}, i.e., a point satisfying $Gx_{\rm f}=h$ and $g(x_{\rm f},\tau)\le 0\ (\tau\in T)$.
Then,
according to Assumptions~A-\ref{B4}, B, and item-\ref{prop:0820-w}
of Proposition\,\ref{property_bound},
there exists some positive constant $M>0$ by which the following three sequences are bounded from above:
$\left\{\|\xi(x^r)\|\right\}$, $\left\{\|\nabla f(\ru[x][r])\|\right\}$, and
$$
\left\{\nabla f(x^r)^{\top}(x_{\rm f}-x^r)+\frac{1}{2}(x_{\rm f}-x^r)^{\top}B_r(x_{\rm f}-x^r)-\mu \xi(x^r)^{\top}(x_{\rm f}-x^r)\right\}.$$
Then,
since $x_{\rm f}-x^r$ is feasible to SIQP\,\eqref{SIQP} with $x=x^r$ and $B=B_r$ for any $r\ge 0$ and
$B_r\succeq \eta_1 I$ (Assumption~B),
we have
\begin{align*}
&-M\|\Delta x^r\|+\frac{\eta_1}{2}\|\Delta x^r\|^2-\mu M\|\Delta x^r\|\\
&\le \nabla f(\ru[x][r])^{\top}\Delta x^r+\frac{1}{2}(\Delta x^r)^{\top}B_r\Delta x^r-\mu \xi(x^r)^{\top}\Delta x^r\\
&\le \nabla f(\ru[x][r])^{\top}(x_{\rm f}-x^r)+\frac{1}{2}(x_{\rm f}-x^r)^{\top}B_r(x_{\rm f}-x^r)-\mu \xi(x^r)^{\top}(x_{\rm f}-x^r)\\
&\le M,
\end{align*}
from which it is easy to see the boundedness of $\{\|\Delta x^r\|\}$.\vspace{0.5em}\\
{\rm 2.} Recall that
$\{F(x^r)\}$, $\{V_r\}$, $\{P_r\}$, and the sequences of their inverse matrices are bounded
by the previous statements and assumptions.
Then,
the scaled sequences
$\{\tilde{F}(x^r)\}$,
$\{\tilde{V}_r\}$,
$\{\tilde{F}(x^r)^{-1}\}$, and $\{\tilde{V}_r^{-1}\}$ are all bounded.
We thus find that the sequences of linear operators $\{\mathcal{L}_{\tilde{F}(x^r)}\}$ and $\{\mathcal{L}_{\tilde{F}(x^r)}^{-1}\}$ are bounded.
By these observations and the Newton equations $\tilde{V}_r+\Delta \tilde{V}_r=\mu \tilde{F}(x^r)^{-1}
-\mathcal{L}_{\tilde{F}(x^r)}^{-1}\left(\sum_{i=1}^n\Delta x_i \tilde{F}_i\circ \tilde{V}_r\right)$ (see \eqref{eq:scaled_new}),
$\{\Delta \tilde{V}_r\}$ is bounded.
Then, by using
$\Delta V_r=P_r\Delta \tilde{V}_rP_r^{\top}$
and the boundedness of $\{P_r\}$ again,
we see that $\{\Delta {V}_r\}$ is bounded.
\end{proof}}
\begin{Prop}\label{prop:conv}
Suppose that Assumptions~A and B hold.
Then, $\{\Phi_{\rl[\rho][r]}(\ru[x][r],\rl[V][r])\}$ is convergent.
\end{Prop}
\begin{proof}
By the line search procedure, we have, for any $r\ge \bar{r}$, $\rho_r=\bar{\rho}$ and
\begin{equation}
\Phi_{{\bar{\rho}}}(\ru[x][r+1],\rl[V][r+1])\le \Phi_{\bar{\rho}}(\ru[x][r],\rl[V][r])+s\bar{\rho}\rl[\gamma][r].\notag
\end{equation}
Notice that $\sum_{r=0}^{\infty}\rl[\gamma][r]=\gamma_0\sum_{r=0}^{\infty}\beta^r<\infty$.
Also, notice that $\{\Phi_{\bar{\rho}}(x^r,V_r)\}$ is bounded below, since $\left\{\left(x^r,V_r\right)\right\}$ is bounded by Assumption~A-\ref{B4}
and
item\,\ref{bound_v} in Proposition\,\ref{property_bound}.
In view of these observations, $\{\Phi_{\bar{\rho}}(x^r,V_r)\}$ is a convergent sequence.
\end{proof}
\begin{Prop}\label{prop:0719}
Suppose that Assumptions~A and B hold.
Then, $\Delta x^r\to 0$ and $\Delta V_r\to O$ as $r\to \infty$.
\end{Prop}
\begin{proof}
From Propositions\,\ref{property_bound}, \ref{prop:bd:dx}, and Assumption~A-\ref{B4},
$\left\{\left(x^r,V_r,\Delta x^r,\Delta V_r,B_r\right)\right\}$ is bounded and has at least one accumulation point, say $(x^{\ast},V_{\ast},\Delta x^{\ast},\Delta V_{\ast},B_{\ast})\in \R^n\times S^m_{++}\times \R^n\times S^m\times S^n_{++}$.
Without loss of generality, we may suppose that
$$
\lim_{r\to \infty}(x^r,V_r,\Delta x^r,\Delta V_r,B_r)=\left(x^{\ast},V_{\ast},\Delta x^{\ast},\Delta V_{\ast},B_{\ast}\right).
$$
Since $\{\Phi_{\bar{\rho}}(x^r,V_r)\}$ is convergent according to Proposition\,\ref{prop:conv},
\begin{equation}
s_r\left(-(\Delta x^r)^{\top}B_r\Delta x^r
+
\psi^{\prime}(x^r,V_r;\Delta x^r,\Delta V_r)\right)\to 0
\end{equation}
as $r\to \infty$, from which together with $B_r\in S^n_{++}$ and
$\psi^{\prime}(x^r,V_r;\Delta x^r,\Delta V_r)\le 0$
we have
\begin{equation}
\lim_{r\to \infty}s_r(\Delta x^r)^{\top}B_r\Delta x^r=0,\ \lim_{r\to \infty}s_r\psi^{\prime}(x^r,V_r;\Delta x^r,\Delta V_r)=0\label{sono0}
\end{equation}
as $r\to \infty$.
If $\liminf_{r\to\infty}s_r>0$ holds, we can easily derive that $\Delta x^{\ast}=0$ and $\Delta V_{\ast}=O$, and the proof is complete.
Suppose $\liminf_{r\to\infty}s_r=0$.
Without loss of generality, we may assume that
$\lim_{r\to\infty}s_r=0$.
It follows
from the linesearch rule that
\eqref{eq:armijo-rule} does not hold with $s=s_r/\beta$, namely,
\begin{align}
&\beta\left(\Phi_{\bar{\rho}}\left(x^r+\frac{s_r}{\beta}\Delta x^r,V_r+\frac{s_r}{\beta}\Delta V_r\right)-\Phi_{\bar{\rho}}\left(x^r,V_r\right)\right)\notag \\
>&-\alpha s_r(\Delta x^r)^{\top}B_r\Delta x^r+\nu {\alpha}s_r\psi^{\prime}(x^r,V_r;\Delta x^r,\Delta V_r)+
\bar{\rho}s_r\gamma_r\label{sono1}
\end{align}
for any $r$.
Dividing both sides of \eqref{sono1} by $s_r>0$ and letting $r\to \infty$ yield
\begin{align}
\Phi^{\prime}_{\bar{\rho}}(x^{\ast},V_{\ast};\Delta x^{\ast},\Delta V_{\ast})
&=\chi^{\prime}_{\bar{\rho}}(x^{\ast};\Delta x^{\ast})+\nu\psi^{\prime}(x^{\ast},V_{\ast};\Delta x^{\ast},\Delta V_{\ast})\notag\\
&\ge\alpha\left(-(\Delta x^{\ast})^{\top}B_{\ast}\Delta x^{\ast}+\nu\psi^{\prime}(x^{\ast},V_{\ast};\Delta x^{\ast},\Delta V_{\ast})\right).\label{sono2}
\end{align}
On the other hand, it follows from \eqref{a3} that
\begin{align*}
&{\Phi_{\bar{\rho}}(x^r+s_r\Delta x^r,V_r+s_r\Delta V_r)-\Phi_{\bar{\rho}}(x^r,V_r)}\\
&\hspace{5em}\le -s_r(\Delta x^r)^{\top}B_{r}\Delta x^{r}+
\nu s_r\psi^{\prime}(x^r,V_r;\Delta x^r,\Delta V_r)
+s_r\bar{\rho}\gamma_r+o(s_r).
\end{align*}
Dividing both sides by $s_r$ and letting $r\to\infty$
yield
\begin{equation}
\Phi^{\prime}_{\bar{\rho}}(x^{\ast},V_{\ast};\Delta x^{\ast},\Delta V_{\ast})
\le -(\Delta x^{\ast})^{\top}B_{\ast}(\Delta x^{\ast})^{\top}+
\nu\psi^{\prime}(x^{\ast},V_{\ast};\Delta x^{\ast},\Delta V_{\ast}).\label{sono3}
\end{equation}
Combining \eqref{sono2} and \eqref{sono3}, we have
\begin{equation}
\alpha\left(-(\Delta x^{\ast})^{\top}B_{\ast}\Delta x^{\ast}
+\nu\psi^{\prime}(x^{\ast},V_{\ast};\Delta x^{\ast},\Delta V_{\ast})
\right)\le -(\Delta x^{\ast})^{\top}B_{\ast}\Delta x^{\ast}+\nu\psi^{\prime}(x^{\ast},V_{\ast};\Delta x^{\ast},\Delta V_{\ast}).\label{eq:sono4}
\end{equation}
By the positive definiteness of $B_{\ast}$, we have $(\Delta x^{\ast})^{\top}B_{\ast}\Delta x^{\ast}\ge 0$.
From Lemma\,\ref{lem_bp},
we can deduce $\psi^{\prime}(x^{\ast},V_{\ast};\Delta x^{\ast},\Delta V_{\ast})\le 0$.
Hence, the relation\,\eqref{eq:sono4} together with $\alpha\in (0,1)$ yields that $\Delta x^{\ast}=0$ and $\psi^{\prime}(x^{\ast},V_{\ast};\Delta x^{\ast},\Delta V_{\ast})=0$. Moreover, we obtain $\Delta V_{\ast}=O$ from Lemma\,\ref{lem:0820}.
\end{proof}
Using the above propositions, we have the following convergence theorem.
\begin{Thm}
Suppose that Assumptions~A and B hold.
Then, the generated sequence
$\left\{\left(x^r,y^r,z^r,V_r\right)\right\}$ is bounded.
Furthermore, {any $\wast$-accumulation point $(x^{\ast},y^{\ast},z^{\ast},V_{\ast})$ of $\{(x^r,y^r,z^r,V_r)\}$} is a KKT point for the SIPLOG\,\eqref{lsisdp}.
\end{Thm}
{\begin{proof}
The first claim follows from Assumption~A\,\ref{B4} and item\,\ref{bound_v} of Proposition\,\ref{property_bound} immediately.
We show the second-half.
Recall that any bounded sequence in $\M(T)$ has at least one weak* accumulation point
and one can extract a subsequence weakly* converging to that point.
Hence,
we may assume that the entire sequence $\{(x^r,y^r,z^r,V_r)\}$ weakly* converges to $(x^{\ast},y^{\ast},z^{\ast},V_{\ast})$ without loss of generality.
Since
\begin{equation}
\Delta x^{r-1}\to 0,\ \Delta V_{r-1}\to O\label{eq:0719-1356}
\end{equation}
as $r\to \infty$ by Proposition\,\ref{prop:0719} and $s_{r-1}\in [0,1]$ for each $r$,
we see that
\begin{equation}
\lim_{r\to \infty}(x^{r-1},V_{r-1}) = \lim_{r\to \infty}
\left(x^{r}-s_{r-1}\Delta x^{r-1},V_{r-1}-s_{r-1}\Delta V_{r-1}\right) =(x^{\ast}, V_{\ast}),\label{eq:0719-1352}
\end{equation}
which together with ${\rm w}^{\ast}\mbox{-}\lim_{r\to \infty}y^r = y^{\ast}$ implies
\begin{equation}
\lim_{r\to \infty}
\int_T\left(g(x^{r-1},\tau)+\nabla_xg(x^{r-1},\tau)^{\top}\Delta x^{r-1}\right)dy^r(\tau)=
\int_Tg(x^{\ast},\tau)dy^{\ast}(\tau).\label{eq:0719-1353}
\end{equation}
Moreover, since $\{P_r\}$ and $\{P_r^{-1}\}$ are both bounded by item~\ref{pr_bound} of Proposition\,\ref{property_bound},
\eqref{eq:0719-1356} implies
\begin{equation}
\lim_{r\to \infty}\Delta \tilde{V}_{r-1}=
\lim_{r\to \infty}
P_{r-1}^{\top}\Delta {V}_{r-1}P_{r-1}=O.\label{eq:0719_1455}
\end{equation}
In addition, there exists an accumulation point $P_{\ast}\in \R^{m\times m}$ of $\{P_r\}$.
Without loss of generality, we can suppose that $\lim_{r\to \infty}P_r = P_{\ast}$.
Then, it holds that
\begin{equation}
\lim_{r\to \infty}\left(
\tilde{F}(x^r),\tilde{V}_r
\right)
=\left(P_{\ast}F(x^{\ast})P_{\ast}^{\top},
P_{\ast}^{-\top}V_{\ast}P_{\ast}^{-1}
\right).\label{eq:0719-1706}
\end{equation}
From \eqref{al:opt} and \eqref{eq:scaled_new}, we have, for each $r$,
\begin{align}\label{al:opt2}
&\nabla f(x^{r-1})+B\Delta x^{r-1}-\mu \xi(x^{r-1})+\int_T\nabla_xg(x^{r-1},\tau)dy^r(\tau)+G^{\top}z^{r}=0,\notag \\
&g(x^{r-1},\tau)+\nabla_xg(x^{r-1},\tau)^{\top}\Delta x^{r-1}\le 0\ \ (\tau\in {\rm supp}(y^r)),\\
&\int_T\left(g(x^{r-1},\tau)+\nabla_xg(x^{r-1},\tau)^{\top}\Delta x^{r-1}\right)dy^r(\tau)=0,\ G(x^{r-1}+\Delta x^{r-1})=h,\notag \\
&\max_{\tau\in T}\left(
g(x^{r-1},\tau)+\nabla_xg(x^{r-1},\tau)^{\top}\Delta x^{r-1}
\right)_+\le \gamma_{r-1},\ y^{r}\in \mathcal{M}_+(T),\notag \\
&\left(\tilde{F}(x^{r-1})+\sum_{i=1}^n\Delta x_i^{r-1}\tilde{F}_i\right)\circ\tilde{V}_{r-1}+
\tilde{F}(x^{r-1})\circ\Delta \tilde{V}_{r-1}=\mu I,\ \tilde{F}(x^{r-1})\in S^m_{+}, \tilde{V}_{r-1}\in S^m_{+} \notag
\end{align}
Note \eqref{eq:0719-1356}--\eqref{eq:0719-1706}, and $\gamma_{r-1}\to 0\ (r\to \infty)$.
Then, by letting $r\to \infty$ in \eqref{al:opt2}, we conclude that
$(x^{\ast},y^{\ast},z^{\ast},V_{\ast})$ is a KKT point of the SIPLOG\,\eqref{lsisdp}.
\end{proof}}
\section{Numerical experiments}\label{sec:4}
\input{numeric_bsqp.tex}
\section{Conclusion}\label{sec:conclusion}
In this paper, we have proposed the interior point SQP-type method (Algorithm~1) for finding a KKT point for the SIPLOG\,\eqref{lsisdp}.
In this method, we generate a sequence of inexact KKT points of semi-infinite quadratic programs approximating the SIPLOG\,\eqref{lsisdp}.
We further solve scaled Newton equations to generate a
NT or HRVW/KSH/M search direction in the dual matrix space.
We have shown that any weak* accumulation point of a produced sequence is a KKT point for the SIPLOG under some assumptions.
To examine the efficiency of the proposed algorithm,
we conducted some numerical experiments, in which we solve the SIPLOG\,\eqref{lsisdp} for various values of the barrier parameter $\mu$.
From the numerical results, we observed that the proposed algorithm performed well for finding a KKT point of the SIPLOG.
As a future work, we will develop a path-following method for solving the SISDP\,\eqref{lsisdp2} based on Algorithm~1.
\input{bib_jcam.tex}
\end{document}
\endinput
|
\section{Introduction}
\begin{figure}[t]
\centering
\includegraphics[width=8.5cm]{ARViewer_Image.png}
\caption{Stereoscopic AR Predictive Display (SARPD) showing the predicted slave-tools in real-time (highlighted with the dotted red lines), overlaid on the camera feedback under 1sec of round trip delay. SARPD can help users coordinate motor tasks preemptively to save time.}
\label{fig:AR_Example}
\end{figure}
\begin{figure*}[b]
\centering
\includegraphics[width=16cm]{FlowChart.png}
\caption{Architecture for SARPD with a round trip delay of $d$ where $m$, $s_m$, and $s$ are all poses, $j$ are the joint angles, $T^c_b$ are the hand-eye transforms between the base of the slave arm and camera, and $I$ are stereo images. For the sake of simplicity, the entire system is assumed to have a sampling rate of $f_s$, and let $n_d = f_sd$}
\label{fig:FlowChart}
\end{figure*}
Teleoperation brings the advantage of remote control and manipulation to distant locations or harsh or constrained environments. The system allows operators to send commands from a remote console, traditionally called a $master$ device, to a robot, traditionally called a $slave$ device, and offers synchronization of movements. This allows the remote user to operate as if on-site, making teleoperational systems an ideal and often only solution to a wide range of applications such as underwater exploration, space robotics, mobile robots, and telesurgery \cite{yipDasJournal,TeleoperationHistory}.
Perhaps the most well known and still widely used teleoperation system to date is the da Vinci\textregistered{} Surgical System, which is deployed in thousands of hospitals. However, the master and slave systems are operated in the same room, providing some advantages of tremor reduction and stereoscopic vision \cite{reduceTremors} but none of the remote teleoperation functionality. The main technical challenge in realizing remote telesurgery (and similarly, all remote teleoperation) is the latency from the communication distance between the master and slave. Latency can reach beyond a second when linking sites between continents \cite{surgeryDelayMeasurement}, and may have higher ranges when using satellite or space communications. This delay causes overshoot and oscillations in the commanded positions, and are observable and statistically significant in as little as 50msec \cite{SteadyHand} \cite{TreatmentPlanning} of round trip communication delay. Anvari et al. reported on 21 remote laparoscopic surgeries where a distance of over 400km had delay of 135-140msec \cite{firstTeleRobotic}. Furthermore, 300msec has been stated as the largest delay where surgeons feel safe \cite{TransatlanticSurgery, LatencyEffects}, and when using satellite communication for operations between London and Toronto, a delay of 560msec was observed \cite{reasonForDelay}.
Teleoperational systems under delay is not specific to telesurgery and its history is too vast to be covered in this paper. Please refer to Hokayem and Spong's historical survey for a more extensive background on prior work \cite{TeleoperationHistory}. Nonetheless, early work by Ferrel and Sheridan suggested supervisory control to mitigate the negative effects of delay \cite{supervisoryControl}. This gives the robotic system the capability of making limited decisions on its own while being supervised by the operator; however, the master loses explicit control of the motions of the slave, and thus supervisory controlled can only be implemented practically at this time in structured environments.
The largest area of research in delayed teleoperation has been for haptic feedback. When haptic feedback is given to the operator, it has been experimentally measured and theoretically shown that delay causes instability \cite{experimentalInstability, theoryInstability}. While there are some techniques to dampen the unstable overshoot and oscillatory behavior such as wave variables \cite{waveVariables}, these techniques have been shown to increase task completion time when under delay \cite{yipBadHaptic}, \cite{yipBadHaptic_2}, and teleoperating without haptic feedback can be often the better alternative.
Predictive displays are virtual reality renderings, generally designed for space operations, that show a prediction of the events to follow in a short amount of time. It can be used to overcome the negative effects of delay by giving the operator immediate feedback from a predicted environment. Furthermore, it does not suffer stability issues that arise with delayed haptic feedback. Early predictive displays included manipulation of the Engineering Test Satellite 7 from ground control where the round trip delay can be up to 7sec \cite{spaceTeleopDelay2, spaceTeleopDelay3} and Augmented Reality (AR) rendering where the prediction is overlaid on raw image data \cite{AR_pd}. These strategies can be applied to telesurgery, but require overcoming the unique challenges in calculating and tracking the 3D environment for a full environment prediction, which includes non-rigid material such as tissue. Furthermore, prior work in the surgical robotics community highlights the need for active tracking rather than only relying on kinematic calibrations to localize the slave due to the millimeter scale of a surgical operation and the often utilized cable driven actuation \cite{need_tracking1, need_tracking2, EKF, need_tracking3, need_tracking4}.
In this paper we propose the first predictive display designed for teleoperated surgical robots. Most crucially, we calculate the predicted behavior of the robot arms and display this prediction in real-time to the operator to anticipate the delayed slave feedback video. The novel contributions of the paper are:
\begin{enumerate}
\item a real-time strategy for kinematically estimated AR registration, rendering, and lens-distorted image overlays for stereoscopic left and right streams of laparoscopic stereo cameras,
\item an Extended Kalman Filtering (EKF) strategy to address the challenges of visual-mismatch between the prediction and the actual movements that arise from imperfect kinematic calibrations with on-the-fly corrections, and
\item an adaptive transparency filter that prevents confusion arising from overlapping virtual and real visuals of the robot arms.
\end{enumerate}
We present a complete system, and show through a user study that the result is an efficient AR rendering architecture for streaming stereoscopic displays. We call this a Stereoscopic AR Predictive Display (SARPD). We show that over 30fps for the stereoscopic AR rendering and 24Hz for the slave-tool tracking from the EKF can be achieved while running simultaneously on a commodity GPU and using the robot operating system (ROS). A user study is carried out to demonstrate its ability to improve the speed of procedures without affecting error rates. Beyond telesurgery, this solution can be deployed to any teleoperated robot with visual feedback once calibrated. Furthermore, it does not require stereo cameras or displays since the real-time slave-tool tracking is done on a single monocular camera data stream.
\section{Methods}
A block diagram in Fig. \ref{fig:FlowChart} shows the architecture for SARPD, and the variables shown will be used throughout this paper. In surgical systems, such as the da Vinci\textregistered{} Surgical System, translational motions from the operator are scaled down to improve accuracy. This constant scaling relationship to set the slave's pose, $s$, from the operators input pose, $m$, in a teleoperation system under delay is described as follows:
\begin{equation}
p_{s_m}[n] = scale(p_{m}[n] - p_{m}[n-1]) + p_{s}[n-1]
\end{equation}
\begin{equation}
q_{s_m}[n] = q_{m}[n]
\end{equation}
\begin{equation}
s[n] = s_m[n-n_d/2]
\end{equation}
where $p_x$ and $q_x$ are the translational component and quaternion of pose $x$. Equations (1) and (2) give the target pose for the slave, $s_m$, through the constant scaling for the position and mirroring the operators rotational input. Note that both $s_m$ and $s$ are poses in the corresponding slave arm base frame. The rotation is mirrored because master consoles such as the a Vinci\textregistered{} Surgical System use wrist orientation as an input. Equation (3) simply highlights the delay channel.
The predictive display proposed here has two major components: slave-tool tracking by correcting the hand-eye transform in real-time to calculate an accurate prediction and stereoscopic AR rendering to display the prediction. Both are running asynchronously, and ROS is used to pass the data from the slave-tool tracking to the stereoscopic AR rendering.
\subsection{EKF to Correct Hand-Eye}
Ye et al. previously developed a tacking algorithm for the slave-tools and showed it to be accurate and robust to surgical environments \cite{EKF}. It successfully tracked the slave-tools in real-time by estimating the error of the initial hand-eye calibrations between the base of both slave arms and the left camera in real-time by using:
\begin{enumerate}
\item virtual slave-tool rendering to generate part-based templates online,
\item template matching between virtual slave-tool parts and image data,
\item geometric context to extract the best location estimates of the slave-tool parts, and
\item EKF to track the error of the hand-eye transform using the previously found 2D estimates for the update.
\end{enumerate}
The hand-eye correction from \cite{EKF} is implemented on the master side, so the predicted AR slave-tools use the correction corresponding to the image data they will be displayed with. This means the joints angles, $j[n]$, from encoder readings and image data, $I[n]$, are passed through the communication channel before calculating the correction. This differs from \cite{EKF} which is not concerned with delayed teleoperation. Each slave-tool part is given its own thread in step 1 and 2 to improve the real-time performance. Fig. \ref{fig:EKF} is an example photo showing the detected features and rendered slave-tools that use the hand-eye correction.
\begin{figure}[t]
\vspace{2mm}
\centering
\includegraphics[width=8.5cm]{EKF_Image.png}
\caption{Implementation of hand-eye correction inspired by \cite{EKF} with adaptation for round trip delay of 1sec and parallelization to improve performance. Colored points represent visual features used in the EKF tracker that are related to tracking of kinematic robot links.}
\label{fig:EKF}
\end{figure}
\begin{figure*}[t]
\centering
\vspace{2mm}
\begin{subfigure}{.245\textwidth}
\centering
\includegraphics[width=1\linewidth]{ARViewer_Image2.png}
\end{subfigure}%
\hspace{0.0001\textwidth}
\begin{subfigure}{.245\textwidth}
\centering
\includegraphics[width=1\linewidth]{ARViewer_Image3.png}
\end{subfigure}%
\hspace{0.0001\textwidth}
\begin{subfigure}{.245\textwidth}
\centering
\includegraphics[width=1\linewidth]{ARViewer_Image4.png}
\end{subfigure}%
\hspace{0.0001\textwidth}
\begin{subfigure}{.245\textwidth}
\centering
\includegraphics[width=1\linewidth]{ARViewer_Image5.png}
\end{subfigure}
\caption{Example photos of AR prediction and dynamic transparency under 1sec of round trip delay. From left to right, a ring hand-off task is shown.}
\label{fig:AR_Viewer_Fading}
\end{figure*}
To initialize the EKF, the hand-eye transform from calibration completes steps 1 through 3, and instead of step 4, a Perspective-n-Point (PnP) solver on the detected features is applied \cite{EKF}. To improve robustness, this is repeated when the distance or absolute difference in roll, pitch, or yaw between the corrected hand-eye and initial hand-eye from calibration is greater than a set threshold. This and the EKF updates to the corrected transform, $\hat{T}_b^c$, cause large steps in the hand-eye transform relative to the scale of surgical environments. These large steps will make the predicted AR slave-tool appear jumpy and therefore untrustworthy to operators. To smooth this, the following filter is applied to get the outputted hand-eye transform, $T_b^c$:
\begin{align}
q_b^c[n] &= \frac{\text{sin}\big((1-a)\Omega[n]\big)}{\text{sin } \big(\Omega[n] \big)} q_b^c[n-1] + \frac{\text{sin}(a\Omega[n])}{\text{sin } \big(\Omega[n] \big)} \hat{q}_b^c[n] \\
\Omega[n] &= \text{cos}^{-1}\big(q_b^c[n-1] \cdot \hat{q}_b^c[n] \big) \\
p_b^c[n] &= (1-a) p_b^c[n-1] + a \hat{p}_b^c[n]
\end{align}
where $q_b^c$ and $p_b^c$ are the quaternion and translation representations of the outputted, smoothed hand-eye transform, $T_b^c$. Likewise, $\hat{q}_b^c$ and $\hat{p}_b^c$ are the corrected transforms from the EKF's output, $\hat{T}_b^{c}$. This is simply a first order infinite impulse response filter with parameter $a$ and spherical linear interpolation to average the rotations.
\subsection{Stereoscopic AR Rendering}
SARPD is rendered using The Visualization Toolkit (VTK) and OpenGL. To render the predicted slave-tools, their 3D CAD models are loaded as VTK Actors and uses the most recent filtered hand-eye transform to be in the left camera frame. The predicted joint angles for the slave-tools, $\hat{j}[n]$, are calculated through inverse kinematics from target pose $s_m[n]$, as shown in Fig. \ref{fig:FlowChart}. For the right display rendering, the slave-tools are additionally transformed using the baseline from stereo camera calibration to be in the right camera frame.
Distortion must be applied to the rendered slave-tools since the master console shows the distorted images to the operator. Camera calibration procedures in MATLAB and OpenCV find the coefficients for equations that un-distort barrel and tangential camera distortions. Solving for the inverse of these equations gives multiple solutions. Therefore, bilinear interpolation on the inverse of the discretized un-distort mapping is used to find the distortion map for both the left and right cameras
Similar to how mappings are applied in OpenCV, the distortion map for each camera is split into two separate mappings, columns and rows. The four mappings are uploaded as textures onto the GPU once at the beginning and then applied to the slave-tool renderings using fragment shaders.
Two virtual cameras, left and right, render the stereoscopic display. Each virtual camera display overlays the predicted and distorted slave-tool renderings on top of the corresponding image frame which are uploaded to the GPU as textures. Both the left and right rendering pipeline are done in parallel and the VTK slave-tool actors and textures are shared resources on the GPU. Through this implementation optimization and utilizing fragment shaders to apply the distortion mapping, the rendering pipeline can be run on consumer grade GPUs. Fig. \ref{fig:AR_Example} shows an example of the predicted slave-tools and image data rendering.
Even with the real-time hand-eye correction, there are other kinematic inaccuracies that are unaccounted for. Joint angle inaccuracies on the wrist of the slave-tool will cause inconsistency between the AR rendering and the image. This is noticeable when the slave-tools are making no motion. Even small, sub-millimeter inaccuracies were observed to cause confusion to users during initial studies. To overcome this, the opacity, $\alpha$, of the rendered slave-tools is dynamically set with the following equation:
\begin{align}
l &= \big| \big| p_{s_m}[n] - p_{s} [n-n_d/2] \big| \big| \\
\alpha &= \text{min} \Big( \alpha_{max}, r\big(\text{max} ( l_{thresh}, l) - l_{thresh}\big) \Big)
\end{align}
The opacity will increase proportionally with the distance $l$, which is the distance between the new target pose that the prediction shows and the pose in the delayed image. The distance threshold, $l_{thresh}$, is the minimum value $l$ must have to not be fully transparent, and $\alpha_{max}$ simply sets the maximum opacity. The dynamic opacity allows for the operator to use the image data undisturbed when making precise, slow motions where the kinematic inaccuracies would be apparent and naturally use the AR prediction when making larger motions. Fig. \ref{fig:AR_Viewer_Fading} shows an example of the dynamic transparency during a ring hand-off.
\section{Experimental Setup}
To initialize the hand-eye transform for the EKF, we provide a calibration method that only requires a rigidly mounted checkerboard on the slave-tools (only during the calibration phase). This is the only additional hardware required to implement the proposed predictive display, and calibration is only required once before a procedure. Performance and latency tests were also conducted to show the efficiency of SARPD.
To measure the effectiveness SARPD, a user study with ten participants was conducted on the da Vinci\textregistered{} Surgical System. The participants were seven novices and three surgeons who use the da Vinci for their practice. Errors and time to complete task are used as metrics, and statistical testing was done to draw conclusions of SARPD's performance. A recent study found that when using satellite communication for telesurgery between London and Toronto, a round trip delay of $560.7 \pm 16.5$mecs was measured \cite{reasonForDelay}. So a round trip delay of 1sec was used for the delayed environment in the user study to ensure SARPD performs in realistically high latency remote operations.
SARPD is rendered to the master console at 30fps. The operator manipulates two master arms in the console to get pose $m$. There are two 7-DOF slave arms which go to pose $s$. Modifications were made to the daVinci Research Toolkit (dVRK \cite{dVRK}) to support the architecture proposed in Fig. 2 for the da Vinci\textregistered{} Surgical System. The computer used to run the modified dVRK and SARPD has an Intel\textregistered{} Core\texttrademark{} i9-7940X Processor and NVIDIA's GeForce GTX 1060.
\begin{figure*}[b]
\centering
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1\linewidth]{Error1.png}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1\linewidth]{Error2.png}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1\linewidth]{Error3.png}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1\linewidth]{Error4.png}
\end{subfigure}
\caption{Example errors participants made. From left to right: touch peg, stretch ring on peg, drop ring, and stretch ring during handoff.}
\label{fig:exampleErrors}
\end{figure*}
\subsection{Initial Hand-Eye Calibration}
To find the initial hand-eye transform, a checkerboard is rigidly attached to the end-effector of the slave-tool. The arm is moved around and photos are taken of the checkerboard and the corresponding joint angles are recorded. The $i$-th corner on the checkerboard in the checkerboard frame, $p_e(i)$, with side length $s$, is projected onto the image plane with the following equation:
\begin{equation}
p_c (i, j) = \frac{1}{d} K T_b^c T_{-e}^b (j) T_e^{-e} p_e(i)
\end{equation}
$\frac{1}{d}K$ projects a point onto the image plane using the camera matrix. $T_{-e}^b(j)$ is the pose of the checkerboard which is calculated by using forward kinematics from the recorded joint angles at image $j$, and $T_e^{-e}$ is the constant offset error of the calculated $T_{-e}^b (j)$ which accounts for mounting errors of the checkerboard. To solve for the hand-eye transform in a robust manner, the following optimization problem is solved when there are $n$ corners on the checkerboard and $m$ recorded images and joint angles:
\begin{equation}
\text{ }\underset{T_b^c, T_e^{-e},s}{\operatorname{argmin}}\text{ } \frac{1}{m} \sum \limits_{j=1}^m \sum \limits_{i=1}^n \Big|\Big| p_c(i,j) - z(i,j) \Big|\Big|^2
\end{equation}
where $z(i,j)$ is the pixel position of the checkerboard corner on the recorded image. Optimizing the constant offset of the checkerboard and side length, $s$, will help account for alignment errors and printing inaccuracies. $T_e^{-e}$ is initially set to the identity matrix
\begin{figure}[t]
\vspace{2mm}
\centering
\includegraphics[width=8.5cm]{HandEye-Checkerboard.png}
\caption{Calibration tool on the instruments to find the transform from the slave arms base to the camera frame. This calibration is carried out once, as needed to attain an initial hand-eye transform for the EKF. The circles highlight the re-projected checkerboard corners.}
\label{fig:hand-eye}
\end{figure}
To initialize the hand-eye transform, a PnP solver is used on each recorded image to find the pose of the camera in the checkerboard frame. This is then further transformed with the corresponding $T_{-e}^b(j)$ to get image $j$'s individual estimate of the hand-eye transform. The initial value for the hand-eye is simply set to the arithmetic average of the positional component and roll, pitch, and yaw of all of the images individual estimates. The optimization is solved by using MATLAB's fmincon function, and Fig \ref{fig:hand-eye}. shows an example result of the re-projected corners on a recorded image
\subsection{Task for User Study}
A peg transfer task, similar to a task in the Fundamentals of Laparoscopic Surgery, is the sample task used in the user study due to the complex motions involved with it. A photo of the environment from the endoscope is provided in Fig. \ref{fig:task_environment}. To complete the task, the operator must pick up the ring from the front right or left peg with the corresponding arm, pass the ring to the other arm, place the ring on to opposite front peg, and then repeat with the back pegs and ring.
Time to complete task and weighted error are used to evaluate performance of completing the task. To find the weighted error, all errors that occurred during a trial are counted and weighted according to Table 1. Example errors are shown in Fig. \ref{fig:exampleErrors}. The weightings were chosen such that the severity of the errors would be reflected properly when evaluating the performance of completing the peg transfer. Each participant was also shown a video of a complete task with no errors and examples of the errors before starting.
\begin{figure}[t]
\vspace{2mm}
\centering
\includegraphics[width=8.5cm]{taskSpace.png}
\caption{Environment setup used for the peg transfer task. Transfer of rings are carried out between instruments and between rows of pegs to characterize single and dual arm control in pickup, hand-off, and place-down task.}
\label{fig:task_environment}
\end{figure}
\begin{table}[h]
\caption{\\Weights associated with type of error}
\begin{center}
\begin{tabular}{ |c|c| }
\hline
\textbf{Error} & \textbf{Weight} \\ \hline
Touch peg & 1 \\ \hline
Touch ground & 2 \\ \hline
Stretch ring during handoff for a second or less & 2 \\ \hline
Drop ring & 3 \\ \hline
Stretch ring on peg for a second or less & 4 \\ \hline
Stretch ring for an additional second & 4 \\ \hline
Stretch/move peg & 10 \\ \hline
Knock down peg & 20 \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{figure*}[b]
\centering
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width=1\linewidth]{weightedErrorPlot.png}
\caption{Weighted Error}
\end{subfigure}
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width=1\linewidth]{timePlot.png}
\caption{Time to Complete Task}
\end{subfigure}
\caption{Weighted error and completion time results from the user study. No statistically significant change in weighted errors is observed with SARPD, but a statistical decrease in completion time is observed when teleoperating under delay.}
\label{fig:userStudyPlots}
\end{figure*}
\subsection{Procedure for User Study}
After introducing the participant to the study and showing the video, the participant goes through the following procedure:
\begin{enumerate}
\item Practice: complete the task twice under no delay and with no SARPD
\item Record: complete the task once under no delay and with no SARPD
\item Repeat step 1 and 2 under delay and with no SARPD
\item Repeat step 1 and 2 under delay and with SARPD
\end{enumerate}
Two practice trials are always completed before every recording, so that the participant can overcome the learning curve for the new environment. To further ensure this, participants are also offered additional practice before recording. The order was chosen so that participants only encounter one new environmental effect at a time. Through initial experimentation we set the following values: $scale = 0.2$ (the default value for dVRK) for equation (1), $a = 0.8$ for equation (4) and (6), and $l_{thresh} = 5.3$mm, $\alpha_{max} = 0.8$, and $r = 100$ for equation (8).
Note that SARPD without using the active tracking from the EKF is not included in the experiment. This was elected due to the the highly inaccurate prediction from the AR rendering when relying on the kinematic calibration.
\section{Results}
The original work for the slave-tool tracking \cite{EKF} was measured to run at 26Hz and 13Hz for the part-based template generation and overall tracking respectively. Through the parallelization efforts, the slave-tool tracking now runs at 50Hz and 24Hz for the part-based template generation and overall tracking respectively. During this measurement, the stereoscopic AR rendering ran asynchronously at 36fps.
We also measured the latency of the entire image pipeline, which is the time from when an image is captured to when it is displayed to the operator under no delay. When using the SARPD, the latency was measured to be $100 \pm 20$ms. To compare, we ran the same experiment with ROS's $image\_view$ to display the raw image data and measured a latency of $110 \pm 20$ms. ROS's $image\_view$ is the recommended way to view images in dVRK, so our solution also improves on the latency of the system.
\begin{table}[h]
\centering
\caption{\\Average and standard deviation from the user study results}
\begin{tabular}{|c|c|c|}
\hline
& Weighted Error & Completion Time (sec) \\ \hline
\textbf{No SARPD, d=0sec} & $2.8\pm3.4$ & $29.6\pm8.8$ \\ \hline
\textbf{No SARPD, d=1sec} & $4.9\pm4.7$ & $60.6\pm14.8$ \\ \hline
\textbf{SARPD, d=1sec} & $3.7\pm3.2$ & $49.2\pm6.3$ \\ \hline
\end{tabular}
\end{table}
The results of the user study comparing delayed teleoperation with and without SARPD is shown in Fig. \ref{fig:userStudyPlots} and the statistics are given in Table 3. To identify a statistically significant ($p<0.05$) change in weighted error or time to complete task under the different conditions, two sided paired t-test's were done. No statistically significant difference was measured in weighted errors when comparing SARPD under delay with not using SARPD under both no delay ($p=0.55$) and delay ($p=0.50$). Repeating for time to complete task, statistical significance was measured when comparing SARPD under delay with not using SARPD under both no delay ($p=1.7e-4$) and delay ($p=0.02$). To conclude, we measured that using SARPD on average decreases the time to complete task by 19\% when under delay while having no statistically significant change in errors.
\section{Discussions and Conclusion}
SARPD is the first predictive display developed for teleoperated surgical robots and has been shown to be an efficient system. On a consumer grade GPU, the tool-tracking is computed in real-time and the AR rendering pipeline is able to support two 1080p displays, each running above 30fps. By utilizing AR to show the prediction, no assumptions need to be made about the environment so SARPD can be applied to any calibrated teleoperated system under delay. SAPRD also on average decreases time to complete task while not affecting the number of errors. This is an expected result since the AR rendering should not assist when small, precise motions are made which is when the errors in our user study occur.
For future work, more studies need to be done on different surgical tasks with rigorous accuracy measurements and subjective data to understand cognitive load when operating with and without SARPD under various levels of delay. In particular, tasks with more environmental interaction. Furthermore, additions to the tracking algorithm must be done to incorporate endoscopic motions which is part of normal operating procedure. The rendering pipeline also needs to account for obstructions of the the slave-tools from environmental obstacles.
\section{Acknowledgements}
We thank all the participants from the user study for their time, all the members of the Advanced Robotics and Controls Lab at University of California San Diego for the intellectual discussions and technical help, and the UCSD Galvanizing Engineering in Medicine (GEM) program and the US Army AMEDD Advanced Medical Technology Initiative (AAMTI) for the funding.\\
\balance
\bibliographystyle{ieeetr}
|
\section{Device design and simulation}
Waveguide dispersion diagrams and mode profiles are numerically calculated using a commercial Finite Difference Eigenmode (FDE) solver (Lumerical, Mode Solutions). Numerical simulation shows that, for the current device layer thickness of 600 nm, air cladding is necessary for anomalous dispersions. For the filter ring, however, a SiO$_2$ cladding gives rise to a better electro-optic tuning efficiency \cite{wang_nanophotonic_2018}. Therefore in the final chip, the SiO$_2$ cladding in the comb generator area is intentionally removed, while the rest of the chip, including the filter ring, is cladded (Fig. \ref{fig1}). The filter tuning efficiency of 2.4 pm/V is lower than our previous results \cite{wang_nanophotonic_2018} since only one arm of the ring resonator is modulated.
\section{Device fabrication}
Devices are fabricated from a commercial x-cut LN-on-insulator (LNOI) wafer (NANOLN) with a 600-nm device layer thickness. Electron-beam lithography (EBL) and Ar+-based reactive ion etching (RIE) are used to create optical waveguides and microring resonators in the LN film, using a similar process as described in our previous work \cite{zhang_monolithic_2017}. A 1.5-$\mu$m-thick PMMA EBL resist is spun coated and exposed using a second EBL with alignment, to produce the microelectrodes of the filter ring via a lift-off process. The structures are then cladded by an 800-nm-thick SiO$_2$ layer using plasma-enhanced chemical vapor deposition (PECVD). The oxide cladding in the comb generation areas are then removed through a photolithography step followed by hydrofluoric acid (HF) wet etching to realize air-cladded devices with the required anomalous dispersions. Finally, the chip edges are diced and polished to improve the fiber-chip coupling.
\section{Characterization of the comb generation, filtering and modulation}
For frequency comb characterization, continuous-wave (CW) light from a tuneable telecom laser (Santec TSL-510) is amplified using an erbium-doped fiber amplifier (EDFA, Amonics). A 3-paddle fibre polarization controller is used to control the polarization of input light. Tapered lensed fibres are used to couple light into and out from the waveguide facets of the LN chip. The output light is sent into an optical spectrum analyzer (OSA, Yokogawa) for analysis. For filter control and manipulation, TE polarized modes are used to exploit the highest electro-optic tuning efficiency. DC signals from a voltage supply (Keithley) and AC signals from an arbitrary waveform generator (AWG, Tektronix 70001A) are combined using a bias T, before being sent to the filter electrodes using a high-speed ground-signal (GS) probe (GGB Industries). The output optical signal from the drop port is sent to a 12-GHz photodetector (Newport 1544A), and analysed using a 1-GHz real-time oscilloscope (Tektronix).
\section{Optical parametric oscillation threshold}
The threshold power of the OPO process can be estimated as \cite{kippenberg_kerr-nonlinearity_2004}
\begin{equation}
\label{eq:th}
P_{\textrm{th}}\approx 1.54\frac{\pi}{2}\frac{Q_c}{2 Q_L}\frac{n_{\textrm{eff}^2} V}{n_2 \lambda_p Q_L^2}
\end{equation}
where $Q_C$ and $Q_L$ are the coupling and loaded $Q$ factors of the resonator, $n_\textrm{eff}$ is the effective refractive index of the LN waveguide, $n_2$ is the nonlinear refractive index, $\lambda_p$ is the pump wavelength and $V$ is the resonator mode volume.
From the measured $Q_L = 6.6 \times 10^5$ and the optical transmission depth $T = 4.6\%$ for TE mode (Fig. 1b), we estimate the intrinsic quality factor $Q_i=(2Q_L)/(1+\sqrt{T})=1.1\times 10^6$ assuming the device is under-coupled. Since $Q_L^{-1} = Q_C^{-1} + Q_i^{-1}$, the coupling $Q$ is calculated to be $Q_C = 1.7 \times 10^6$. We numerically calculate the effective index of our LN waveguide to be $n_\textrm{eff}$ = 1.91 and the mode area $A = 0.875 \mu$m$^2$. The resonator mode volume is $V = 2\pi RA = 440 \mu$m$^3$, where $R = 80 \mu$m is the radius of the ring resonator. The nonlinear refractive index $n_2 = 0.91 \times 10^{-15}$ cm$^2$/W. From these parameters we calculate the optical parametric oscillation threshold to be $P_{\textrm{th}} = 80$ mW.
\section{Filter transmission spectra}
The filter ring resonator is designed to have a free spectral range (FSR) that is $1.2\%$ larger than the comb-generating resonator. Supplementary Figure \ref{figs1} shows the optical transmission spectra at both through and drop ports of the filter, measured using a tunable telecom laser. The transmission dips at the through port correspond to resonances of both ring resonators, while the transmission peaks at the drop port correspond to the filter passband. The spectra show that the two ring resonances are aligned near 1606 nm, and have increasing mismatch when the wavelength is away from 1606 nm.
\begin{figure*}
\centering
\includegraphics[angle=0,width=\textwidth]{figs1.pdf}
\caption{\label{figs1} \textbf{Optical transmission spectra of the device at both through and drop ports of the filter.} The resonances of the comb generator and the filter are aligned near 1606 nm, and have a $1.2\%$ mismatch in the free spectral range.}
\end{figure*}
\section{Filter transfer function}
Supplementary Figure \ref{figs2} shows the theoretical transfer function of the filter, which has a FSR of 2.19 nm and a linewidth of 3 pm. The calculated suppression ratio for the next comb line, which has a 27-pm resonance mismatch, is 25 dB. The calculated suppression ratio for the pump light (resonance mismatch by 730 pm) is 52 dB.
\begin{figure*}
\centering
\includegraphics[angle=0,width=0.7\textwidth]{figs2.pdf}
\caption{\label{figs2} \textbf{Numerically calculated filter transfer function.} The calculated suppression ratios for the next line and the pump are 25 dB and 52 dB respectively.}
\end{figure*}
\end{document}
|
\section{Introduction}
\hspace{0.5cm}Exoplanetary discovery has widened our perspective and knowledge of planetary science in the past two decades. The space-based mission \textit{Kepler} used transit photometry to detect and characterise exoplanets (\citealt{2010Sci...327..977B,2011ApJ...736...19B,2010ApJ...713L..79K}), with one of its key objectives being the determination of the frequency of terrestrial planets in the habitable zones of stars. From their sample of over 4000 transiting planet candidates, it was revealed that small planets ($R_P < 4.0 R_{\oplus}$) are by far the most common in our Galaxy (\citealt{2012ApJS..201...15H,2013ApJS..204...24B,2013ApJ...767...95D,2013PNAS..11019273P}), a result that is also supported by radial-velocity surveys (e.g. \citealt{2013A&A...549A.109B,2011arXiv1109.2497M}). While the \textit{Kepler} sample provided an insight into the planet occurrence rate (e.g. \citealt{2014PNAS..11112647B}), only a few dozen host stars were bright enough for follow-up characterisation. With the loss of two reaction wheels on the \textit{Kepler} spacecraft, the \textit{K2} mission was adopted to extend the transiting exoplanet discoveries \citep{2014PASP..126..398H}. \textit{K2} has observed nineteen fields so far, and supplied precise photometry of approximately $20,000$ bright stars per campaign. This has yielded hundreds of transiting planet candidates (e.g. \citealt{2016ApJS..222...14V,2016A&A...594A.100B,2016MNRAS.461.3399P}), over 300 of which have been statistically validated (e.g. \citealt{2015ApJ...809...25M,2015MNRAS.454.4267B,2016ApJS..226....7C}).
Super-Earths are absent in our own Solar system. Therefore, they are of particular interest in the study of planet formation and evolution. To probe the formation histories of these small planets, it is necessary to derive the planetary masses and radii with precision better than a few percent in order to differentiate their internal compositions in the context of planet evolution models (e.g. \citealt{2013PASP..125..227Z,2017ApJ...850...93B}). Recent theories have proposed a distinct transition in the composition of small exoplanets (\citealt{2014ApJ...783L...6W,2015ApJ...801...41R}). Planets with $\rm R_P \lesssim 1.6 ~R_{\oplus}$ typically have high densities and are predominantly rocky. On the other hand, planets with larger radii typically have lower densities and possess extended H/He envelopes. In fact, planets such as Kepler-10~b ($\rm R_P = 1.42 \pm 0.03 ~R_{\oplus}$, $ \rho_{P} = 8.8 \pm 2.5\ \rm g\ cm^{-3}$; \citealt{2011ApJ...729...27B}), LHS1140~b ($\rm R_P = 1.43 \pm 0.10 ~R_{\oplus}$, $\rho_P = 12.5 \pm 3.4\ \rm g\ cm^{-3}$; \citealt{2017Natur.544..333D}), Kepler-20~b ($\rm R_P = 1.87 \pm 0.05 ~R_{\oplus}$, $ \rho_P = 8.2 \pm 1.4\ \rm g\ cm^{-3}$; \citealt{2016AJ....152..160B}), and K2-38~b ($\rm R_P = 1.55 \pm 0.02 ~R_{\oplus}$, $ \rho_P = 17.5 \pm 7.35\ \rm g\ cm^{-3}$; \citealt{2016ApJ...827...78S}) all have densities higher than that of the Earth ($\rho_{\oplus} = 5.5 \rm g\ cm^{-3}$) and compositions consistent with a rocky world, whereas low density planets such as GJ~1214~b ($\rm R_P = 2.68 \pm 0.13 ~R_{\oplus}$, $\rho_P = 1.87 \pm 0.40\ \rm g\ cm^{-3}$; \citealt{2009Natur.462..891C}), the Kepler-11 system ($\rm R_P = 1.97$--$4.52 ~R_{\oplus}$, $\rho_P = 0.5$--$3.1\ \rm g\ cm^{-3}$; \citealt{2011Natur.470...53L}), and HIP~116454~b ($\rm R_P = 2.53 \pm 0.18 ~R_{\oplus}$, $ \rho_P = 4.17 \pm 1.08\ \rm g\ cm^{-3}$; \citealt{2015ApJ...800...59V}) have solid cores, and substantial gaseous envelopes.
Recent efforts by the California-\textit{Kepler} Survey (CKS) (\citealt{2017AJ....154..108J,2017AJ....154..109F}) have refined the physical characteristics of \textit{Kepler} short-period planets ($\rm P < 100 ~days$) and their host stars for an in-depth study of the planet size distribution. Their results show a significant lack of planets with sizes between $1.5\ \rm R_{\oplus}$ and $2.0\ \rm R_{\oplus}$. The gap in the radius distribution can be explained by the `photoevaporation' model (\citealt{2013ApJ...775..105O,2013ApJ...776....2L}), where the gaseous envelopes of planets are stripped away as a result of exposure to high incident flux from their host stars. The CKS also highlighted the importance of obtaining precise measurements of planet masses and radii in order to perform statistically significant studies of the radius distribution.
In this paper, we report the detection of a $2.37$-day transiting super-Earth, K2-265 b. \textit{K2} photometry and HARPS radial velocity measurements were used to constrain the radius and mass measurements of this planet with a precision of 6\% and 13\%, respectively. In Section \ref{observations}, we describe the observations made from \textit{K2}, data reduction, and spectroscopic follow up. Our analyses and results are presented in Section \ref{results}, and we conclude the paper with a summary and discussion in Section \ref{conclusion}.
\section{Observations\label{observations}}
\subsection{K2 Photometry\label{photometry}}
\hspace{0.5cm}
K2-265 was observed during \textit{K2} Campaign 3 in long cadence mode. The photometry was obtained between November 2014 and January 2015. The target was independently flagged as a candidate from two transit searches; the first made use of the POLAR pipeline \citep{2016A&A...594A.100B}, and the second used the methods described in \cite{2015A&A...579A..19A} and \cite{2015A&A...582A..33A}, where human input was involved to identify high priority candidates. K2-265 was also independently identified as a planet-hosting candidate by other search algorithms \citep{2016ApJS..222...14V, 2016ApJS..226....7C,2018AJ....155..136M}.
The \textit{K2} lightcurve generated from the POLAR pipeline \citep{2016A&A...594A.100B} has less white noise than that of \cite{2015A&A...579A..19A,2015A&A...582A..33A}, hence the former was used in the planetary system analysis. The POLAR pipeline is summarised as follows:
The \textit{K2} pixel data was downloaded from the Mikulski Archive for Space Telescopes (MAST)\footnote{\url{http://archive.stsci.edu/kepler/data_search/search.php}}. The photometric data was extracted using the adapted CoRoT imagette pipeline \cite{2014A&A...569A..74B} which uses an optimal aperture for the photometric extraction. In this case, the optimal aperture was found to be close to circular and comprised of 44 pixels. The Modified Moment Method developed by \citet{1989AJ.....97.1227S} was used to determine the centroid positions for systematic corrections. Flux and position variations of the star on the CCD can lead to systematics in the data. These were corrected following the self-flat-fielding method similar to \citet{2014PASP..126..948V}. Figure \ref{k2lc} shows the final extracted lightcurve, and Table \ref{PhotometricProperties} gives the photometric properties of K2-265.
\begin{figure*}[htbp!]
\centering
\includegraphics[width=\textwidth]{susana_detrend_lc.pdf}
\caption{\label{k2lc}The detrended \textit{K2} transit light curve of K2-265 with positions of transits marked with blue dashed lines. }
\end{figure*}
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.5\textwidth]{EPIC-1496_TransitB.pdf}
\caption{\label{k2foldedlc}The phase-folded \textit{K2} lightcurve of K2-265 is shown in the top panel with the ephemeris from our analysis. The best-fit transit model is plotted in a red solid line, and the residuals of the fit are plotted in the bottom panel.}
\end{figure}
\begin{table}
\begin{threeparttable}
\caption{\label{PhotometricProperties}Properties of K2-265. K2-265 has a nearby bound companion (see text for detailed description), hence values presented in this table are for the blended photometry. The photometric magnitudes listed were used in deriving the SED as described in Section \ref{MCMC}.}
\begin{tabular}{lcc}
\toprule
Parameter & Value and uncertainty & Source\\
\midrule
K2 Campaign & 3 & a\\
EPIC & 206011496 & a\\
2MASS ID & 2MASS J22480755$-$1429407 & b\\
RA(J2000) & 22:48:07.56 & c\\
Dec(J2000) & $-$14:29:40.84 & c\\
$\mu_{RA}$ (mas/yr) & $30.20 \pm 0.09$ & c\\
$\mu_{DEC}$ (mas/yr) & $-23.34 \pm 0.06$ & c\\
Parallax (mas) & $7.18 \pm 0.05$ & c\\
\hline
\multicolumn{3}{l}{\textit{Photometric magnitudes}} \\
Kp & $10.92$ & a\\
Gaia G & $10.928$ & c\\
Johnson B & $11.845\pm0.029$ & d\\
Johnson V & $11.102\pm0.037$ & d\\
Sloan g$^{\prime}$ & $11.419\pm0.042$ & d\\
Sloan r$^{\prime}$ & $10.879\pm0.047$ & d\\
Sloan i$^{\prime}$ & $10.689\pm0.084$ & d\\
2-MASS J & $9.726\pm0.026$ & b\\
2-MASS H & $9.312\pm0.022$ & b\\
2-MASS Ks & $9.259\pm0.027$ & b\\
WISE W1 & $9.178\pm0.022$ & e\\
WISE W2 & $9.213\pm0.020$ & e\\
WISE W3 & $9.162\pm0.040$ & e\\
\bottomrule
\end{tabular}
\begin{tablenotes}
\small
\item a. EXOFOP-K2: \url{https://exofop.ipac.caltech.edu/k2/}
\item b. The Two Micron All Sky Survey (2MASS)
\item c. \textit{Gaia} DR2
\item d. The AAVSO Photometric All-Sky Survey (APASS)
\item e. AllWISE
\end{tablenotes}
\end{threeparttable}
\end{table}
\subsection{Spectroscopic Follow Up\label{RV}}
\hspace{0.5cm}We obtained radial velocity (RV) measurements of K2-265 with the HARPS spectrograph ($\rm R\sim110,000$), mounted on the $3.6$ m Telescope at ESO La Silla Observatory \citep{2003Msngr.114...20M}. A total of 153 observations were made between 2016 October 29 and 2017 November 22 as part of the ESO-K2 large programme\footnote{Based on observations made with ESO Telescopes at the La Silla Paranal Observatory under programme ID 198.C-0169.}. An exposure time of 1800\,s was used for each observation, giving a signal-to-noise ratio of about 50 per pixel at 5500\,\AA. The data were reduced using the HARPS pipeline \citep{1996A&AS..119..373B}. RV measurements were computed with the weighted cross-correlation function (CCF) method using a G2V template (\citealt{1996A&AS..119..373B,2002Msngr.110....9P}), and the uncertainties in the RVs were estimated as described in \citet{2001A&A...374..733B}. The line bisector (BIS), and the full width half maximum (FWHM) were measured using the methods of \citet{2011A&A...528A...4B} and \citet{2015MNRAS.451.2337S}.
Ten observations that were obtained when the target was close to a bright Moon exhibit a significant anomaly in their FWHM, up to 500 \ms. We removed these data completely from the analyses described in the later sections. The remaining 143 RV measurements and their associated uncertainties are reported in Table.~\ref{RVtable}. The time-series RVs and the phase-folded RVs of K2-265 are shown in Figures \ref{RVtimeseries} and \ref{RVfold} respectively. Following the calibrations of \cite{1984ApJ...287..769N}, we derived the activity index of $\log R'_{HK} = -4.90 \pm 0.12$. The activity index is used in Sect.~\ref{K2rotation} to derive the stellar rotational period.
\begin{figure*}[htbp!]
\centering
\includegraphics[width=\textwidth]{EPIC-1496_RVtime.pdf}
\caption{\label{RVtimeseries}Time-series HARPS radial velocity measurements (red circles) of K2-265. The best-fit Keplerian orbit of K2-265 b is plotted in black. The stellar activity is fitted with a Gaussian process. The grey region show the 1-$\sigma$ confidence interval of the Gaussian process.}
\end{figure*}
\begin{figure}[htbp!]
\includegraphics[width=0.5\textwidth]{EPIC-1496_RVphaseB.pdf}
\caption{\label{RVfold} \textit{Upper panel}: Phase-folded HARPS radial velocity measurements (black circles) of K2-265 as a function of the orbital phase. The black solid line is the best-fit RV curve. The binned RV measurements are denoted as red open circles.}
\end{figure}
\subsection{Direct imaging observations\label{astrometry}}
Shallow imaging observations were obtained with the NIRC2 instrument at Keck on 2015-08-04 in the narrow-band $Br_{\gamma}$ filter at 2.169~$\mu$m (programme N151N2, PI: Ciardi). Several images were acquired with a dithering pattern on-sky and they were simply realigned and median-combined. In the combined image, a candidate companion was clearly detected at close separation from the star. Figure \ref{KECKAO} shows the K-bank Keck AO image of K2-265 and the near-by companion, where the contrast of the objects is measured to be $\Delta \rm mag = 8.12$ in the K-band. The relative astrometry of the candidate was estimated using a simple Gaussian fitting on both the star and the candidate. The error on the measurement is conservatively estimated to $\sim$0.5~pixel, i.e. $\sim$5~mas. The relative Keck astrometry was derived following methods described in \cite{Vigan2016}, and the following parameters were obtained: $\Delta\alpha = -910 \pm 5$ mas, $\Delta\delta = -363 \pm 5$ mas, separation $= 979 \pm 5$ mas, and position angle $= 248.27 \pm 0.29$ deg.
The target was further observed with the SPHERE/VLT instrument in the IRDIFS mode \citep{2010MNRAS.407...71V,2014A&A...572A..85Z}. More details on these observations, together with the data reduction are presented in \cite{2018arXiv180903848L}. The relative astrometry of the candidate companion with respect to the star were derived from SPHERE/IRDIFS, and the results are shown in Table \ref{tab:astrometry}. The combined astrometry confirms that the companion is bound with the target star. The SPHERE/IFS data was used to derive a low-resolution NIR spectrum \citep{2018arXiv180903848L} which we used to characterise the companion star and estimate its contamination in the \textit{K2} photometry (see section \ref{companion_star}.
\subsection{GAIA Astrometry}
The \textit{Gaia} Data Release 2 (DR2) has surveyed over one billion stars in the Galaxy \citep{2016A&A...595A...1G,2018arXiv180409365G,2018arXiv180409366L} and provided precise measurements of the parallaxes and proper motions for the sources. K2-265 has a measured parallax of $7.18 \pm 0.05$ mas, corresponding to a distance of $139 \pm 1$ pc. The proper motion of K2-265 is $\mu_{RA} = 30.20 \pm 0.09$ mas, $\mu_{DEC} = -23.34 \pm 0.06$ mas. As part of the \textit{Gaia} DR2, the stellar effective temperature of K2-265 was derived from the three photometric bands \citep{2018arXiv180409374A} as $\rm T_{eff} = 5390_{-53}^{+194}$ K. The G-band extinction $A_g = 0.101$ and the reddening $E(BP-RP) = 0.065$ estimated from the parallax and magnitudes were used to determine the stellar luminosity, which in turn provides an estimate of the stellar radius as $R_s = 0.914_{-0.06}^{+0.02}~R_{\odot}$ \citep{2018arXiv180409374A}. The stellar parameters from the results of \textit{Gaia} DR2 are consistent with the distance estimate, effective temperature and stellar radius which are derived in the joint Bayesian analysis in section \ref{MCMC}. However, \textit{Gaia} DR2 does not detect the companion star in the system and K2-265 is registered as a single object.
\section{Analysis and Results\label{results}}
\subsection{Spectral Analysis\label{stellarparam}}
\hspace{0.5cm}The spectral analysis of the host star was performed by co-adding all the individual (Doppler corrected) spectra with IRAF\footnote{IRAF is distributed by National Optical Astronomy Observatories, operated by the Association of Universities for Research in Astronomy, Inc., under contract with the National Science Foundation, USA.}.
We first derived the stellar parameters following the analysis of \citet{2008A&A...487..373S} by measuring the equivalent widths (EW) of Fe~{\sc i} and Fe~{\sc ii} lines with version 2 of the ARES code\footnote{The ARES code can be downloaded at http://www.astro.up.pt/~sousasag/ares/} \citep{2015A&A...577A..67S}, and the chemical abundances were derived using the 2014 version of the code MOOG \citep{sneden} which used the iron excitation and ionization balance. We obtained the following parameters: T$_{\rm eff}$\,=\,5457\,$\pm$\,29\,K, log~{\it g}\,=\,4.42\,$\pm$\,0.05\,dex, [Fe/H]\,=\,0.08\,$\pm$\,0.02\,dex, microturbulence $\xi_t$\,=\,0.81\,$\pm$\,0.05\,km\,s$^{-1}$. The errors provided here for the stellar parameters are precision errors which are intrinsic to the method \citep{2011A&A...526A..99S}.
The chemical abundances of the host star are found in Table.~\ref{abundance}. For more details on this analysis and the complete list of lines we refer the reader to the following works:
\citet{adibekyan12}, \citet{santos15}, and \citet{delgadomena17}. Li and S abundances were derived by spectral synthesis as performed in \citet{delgadomena14} and \citet{ecuvillon04}, respectively.
\subsection{Characterisation of the Companion Star \label{companion_star}}
To determine the physical parameters of the bound companion, we used the same approach as in \cite{2016ApJ...824...55S}. We fit the magnitude difference between the target and companion star, as observed by SPHERE IRDIFS, with the BT-Settl stellar atmosphere models \citep{2012RSPTA.370.2765A}. The two stars are bound companions (see sections \ref{astrometry}), hence they have the same distance to Earth and age, and they are assumed to have the same iron abundance. We used an MCMC method to derive the companion mass, using the results of the spectral analysis of the target star as priors on the analysis. We used the Dartmouth stellar evolution tracks to convert the companion mass (at a given age and metallicity) into spectroscopic parameters. Our final derivation gives: $\rm T_{eff} = 3428 \pm 22 ~K$, $\rm \log g = 4.870 \pm 0.017 ~[cgs]$, $\rm M_{star B} = 0.40 \pm 0.01 ~M_{\odot}$, $\rm R_{star B} = 0.391_{-0.010}^{+0.006} ~R_{\odot}$, corresponding to a star of spectral type M2 \citep{2000asqu.book.....C}. Using this result, we integrated the SED models in the Kepler band, and derived the contribution of flux contamination in the light curve of star A from star B to be $0.952 \pm 0.024 \%$. The derived contamination of the companion star was taken into account in the joint Bayesian analysis in section \ref{MCMC} to determine the system parameters of K2-265. The parameters of the companion star and their corresponding uncertainties are reported in Table \ref{EPIC1496Param}.
\subsection{Stellar Rotation\label{K2rotation}}
Rotational modulation is observed in the detrended \textit{K2} lightcurve as shown in Fig.~\ref{k2lc}. We derived the rotational period of K2-265 using multiple methods to determine the origin of the periodic variation.
We first calculated the stellar rotational period with the auto-correlation-function (ACF) method as described in \citet{2013ApJ...775L..11M,2014ApJS..211...24M}, and found the stellar rotational period as $15.14 \pm 0.38$ d, with a further peak observed at $30.48 \pm 0.28$ d.
The Lomb-Scargle periodogram \citep{1976Ap&SS..39..447L,1982ApJ...263..835S} analysis was performed to determine the periodicity in the RV data. Fig.~\ref{LSplot} shows the periodogram of the bisector analysis (BIS), the full width at half maximum (FWHM), the RV measurements, and the S index. A clear peak is measured in all four periodograms at $32.2 \pm 0.6$ d, which is larger than but marginally consistent with the ACF period of 30.48 d at a 2-$\sigma$ level. The timescale of lightcurve variation measures the changing visibility of starspots. We attribute the discrepancy between the two rotation periods to latitude variation of the magnetically active regions.
An upper limit of the sky-projected stellar rotational velocity was derived from the FWHM of the HARPS spectra ($v \sin{i} < 1.9 \pm 0.2$ km s$^{-1}$). Using the stellar radius in Table \ref{EPIC1496Param}, we estimate a rotation period $P_{rot} > 26.02 \pm 3.08$ d (assuming an aligned system, $i = 90^{\circ}$), which agrees with the $\sim30$ d period derived from the photometry and the RV data.
Furthermore, the stellar rotation period was also derived following the method of \cite{2008ApJ...687.1264M}. In summary, we used the $\rm B-V$ colour from APASS\footnote{https://www.aavso.org/apass} to find the convective turnover time $\tau_c$ using calibrations from \cite{1984ApJ...287..769N}. We then used the measured Mount Wilson index $S_{MW} = 0.195 \pm 0.025$ to derive $\rm \log R'_{HK} = -4.90 \pm 0.12$, from which we determine the Rossby number $\rm R_o = 1.94$ using calibrations from \cite{2008ApJ...687.1264M}. Finally, using the relation $\rm P_{rot} = R_o \ \times$ $\tau_{c}$, we calculated the stellar rotation period as $32 \pm 10$ d.
\begin{figure*}
\centering
\subfigure[t!][]{\label{periodogramBIS}\includegraphics[width=0.45\textwidth]{Periodogram_BIS.pdf}}
\subfigure[t!][]{\label{periodogramFWHM}\includegraphics[width=0.45\textwidth]{Periodogram_FWHM.pdf}}
\subfigure[t!][]{\label{periodogramRV}\includegraphics[width=0.45\textwidth]{Periodogram_RV.pdf}}
\subfigure[t!][]{\label{periodogramSindex}\includegraphics[width=0.45\textwidth]{Periodogram_Sindex.pdf}}
\caption{\label{LSplot}Lomb-Scargle periodogram of (a) Bisector Span (BIS); (b) Full Width Half Maximum (FWHM); (c) Radial velocity (RV); (d) S index (S$_{MW}$). The peak position is marked by the purple line and corresponds to a period of $32.2\pm0.6$ d. The orbital period of K2-265 b is indicated by the yellow line, and the planet signal is only significant in the RV at the 1\% FAP level.}
\end{figure*}
\subsection{Joint Bayesian Analysis With \texttt{PASTIS} \label{MCMC}}
We employed a Bayesian approach to derive the physical parameters of the host star and the planet. We jointly analysed the \textit{K2} photometric light curve, the HARPS RV measurements and the spectral energy distribution (SED) observed by the APASS, 2-MASS, and WISE surveys (\citealt{2014AJ....148...81M,2014yCat.2328....0C}; a full list of host star magnitudes can be found in Table \ref{PhotometricProperties}) using the \texttt{PASTIS} software \citep{2014MNRAS.441..983D,2015MNRAS.451.2337S}. The light curve was modelled using the \texttt{jktebop} package \citep{2008MNRAS.386.1644S} by taking an oversampling factor of 30 to account for the long integration time of the K2 data \citep{2010MNRAS.408.1758K}. The RVs were modelled with Keplerian orbits. Following similar approaches to \cite{2017A&A...608A..25B} and \cite{2018NatAs...2..393S}, a Gaussian process (GP) regression was used to model the activity signal of the star. The SED was modelled using the BT-Settl library of stellar atmosphere models \citep{2012RSPTA.370.2765A}. \newline
The Markov Chain Monte Carlo (MCMC) method was used to derive the system parameters. The spectroscopic parameters of K2-265A were converted into physical stellar parameters using the Dartmouth evolution tracks \citep{2008ApJS..178...89D} at each step of the chain. The quadratic limb darkening coefficients were also computed using the stellar parameters and tables of \cite{2011A&A...529A..75C}.
For the stellar parameters, we used normal distribution priors centred on the values derived in our spectral analysis. We chose a normal prior for the orbital ephemeris centred on values found by the detection pipeline. Furthermore, we adopted a sine distribution for the inclination of the planet. Uninformative priors were used for the other parameters. The priors of the fitted parameters used in the model can be found in Table \ref{MCMCprior}.
Twenty MCMC chains of $3 \times 10^5$ iterations were run during the MCMC analysis, where the starting points were randomly drawn from the joint prior. The Kolmogorov-Smirnov test was used to test for convergence in each chain. We then removed the burn-in phase and merged the converged chains to derive the system parameters.
\begin{table}[!t]
\caption{\label{EPIC1496Param}System parameters of K2-265 obtained from \texttt{PASTIS}. Stellar parameters of Star B were derived as described in section \ref{companion_star}.}
\small
\centering
\begin{tabular}{ l c}
\hline
Parameter & Value and uncertainty \\
\hline
\textit{Stellar parameters} \\
\textit{Star A} \\
Effective temperature T$_{\rm eff}$\ [K] & $5477 \pm 27$ \\
Surface gravity log~{\it g}\ [cgs] & $4.419\pm0.053$ \\
Iron abundance [Fe/H]\ [dex] & $0.078\pm0.020$ \\
Distance to Earth $D$ [pc] & $145\pm8$\\
Interstellar extinction $E(B-V)$ [mag] & $0.009^{_{+0.011}}_{^{-0.007}}$ \\
Systemic radial velocity $\gamma$ [\kms] & $-18.186\pm0.002$ \\
Stellar density $\rho_{\star}/\rho_{\astrosun}$ & $0.98\pm0.19$ \\
Stellar mass M$_{\star}$\ [\Msun] & $0.915\pm0.017$ \\
Stellar radius R$_{\star}$\ [\hbox{$\mathrm{R}_{\astrosun}$}] & $0.977\pm0.053$ \\
Stellar age $\tau$\ [Gyr] & $9.7\pm3.0$ \\
&\\
\textit{Star B}\\
Effective temperature T$_{\rm eff}$\ [K] & $3428 \pm 22$ \\
Surface gravity log~{\it g}\ [cgs] & $4.870\pm0.017$ \\
Stellar mass M$_{\star}$\ [\Msun] & $0.40\pm0.0.01$ \\
Stellar radius R$_{\star}$\ [\hbox{$\mathrm{R}_{\astrosun}$}] & $0.391_{-0.010}^{0.006}$ \\
&\\
\hline
\textit{Planet Parameters}\\
Orbital Period $P$ [d] & $2.369172\pm8.9\times10^{-5}$ \\
Transit epoch $T_{0}$ [BJD - 2456000] & $981.6431\pm1.6\times10^{-3}$ \\
Radial velocity semi-amplitude $K$ [\ms] & $3.34\pm0.43$ \\
Orbital inclination $i$ [$^{\circ}$] & $87.7\pm1.6$ \\
Planet-to-star radius ratio $k$ & $0.01604\pm0.00041$ \\
Orbital eccentricity $e$ & $0.084\pm0.079$ \\
Impact parameter $b$ & $0.30\pm0.20$ \\
Transit duration T$_{14}$ [h] & $2.266\pm0.050$ \\
Semi-major axis $a$ [AU] & $0.03376\pm0.00021$ \\
Planet mass M$_{p}$ [\hbox{$\mathrm{M}_{\oplus}$}] & $6.54 \pm 0.84$ \\
Planet radius R$_{p}$ [\hbox{$\mathrm{R}_{\oplus}$}] & $1.71\pm0.11$ \\
Planet bulk density $\rho_{p}$ [\gcm3] & $7.1\pm1.8$\\
&\\
\hline
\end{tabular}
\end{table}
\subsection{Stellar Age\label{age}}
\hspace{0.5cm}From the joint analysis of the observational data, together with the Dartmouth stellar evolution tracks, the age of K2-265 was determined as $\tau_{iso} = 9.7\pm3.0$ Gyr. The stellar rotation analysis in Section \ref{K2rotation} found that K2-265 has a rotation period of $\sim 30$ d. We adopted a rotational period of $32.2 \pm 0.6$ d, and followed the methods by \cite{2010ApJ...722..222B} to find that K2-265 has a gyrochronological age of $\tau_{gyro} = 5.34 \pm 0.19$ Gyr. We further derived the age of K2-265 using the relation between the [Y/Mg] abundance ratio and the stellar age \citep{2016A&A...590A..32T,2015A&A...579A..52N}, and found an age of $\tau_{[Y/Mg]} = 3.97 \pm 2.59$ Gyr. $\tau_{[Y/Mg]}$ agrees with $\tau_{gyro}$ within 1-$\sigma$ uncertainty but is lower than the derived isochronal age. The low lithium abundance A(Li~{\sc ii})$ < 0.45$ of the host star obtained from spectral analysis (Section \ref{stellarparam}) suggests that the host is not young. Hence it is likely that the host is of at least an intermediate age.
\section{Discussion \& Conclusion\label{conclusion}}
\hspace{0.5cm}K2-265 b has a mass of $\rm 6.54 \pm 0.84~M_{\oplus}$ and a radius of $\rm 1.71 \pm 0.11~R_{\oplus}$. This corresponds to a bulk density of $7.1 \pm 1.8~ \rm g\ cm^{-3}$, which is slightly higher than that of the Earth's density. We applied a number of theoretical models to investigate the planet's interior composition.
\citet{2007ApJ...659.1661F} modelled the radii of planets with a range of different masses at various compositions, and derived an analytical function which allows an estimate of the rock mass fraction (rmf) of ice-rock-iron planets. We find a rmf of 0.84 for K2-265 b which is equivalent to a rock-to-iron ratio of $0.84/0.16$, a rock fraction that is higher than the Earth. \citet{2007ApJ...669.1279S} also used interior models of planets to study the mass-radius relation of solid planets. By assuming the planets are composed primarily of iron, silicates, water, and carbon compounds, \citet{2007ApJ...669.1279S} showed that masses and radii of terrestrial planets follow a power law. Using the derived best-fit mass and radius of K2-265 b, the bulk composition of the planet was determined to be predominantly rocky with $>70\%$ of silicate mantle by mass.
\begin{figure}[htbp!]
\includegraphics[width=0.5\textwidth]{MRplot.pdf}
\caption{\label{MRplot} A mass-radius diagram of confirmed Earth-sized planets with masses up to $20\ \rm M_{\oplus}$. Data were taken from NASA Exoplanet archive \protect\footnotemark . The mass-radius relations were taken from \cite{2016ApJ...819..127Z}. From top to bottom, the black solid lines denotes a pure water, pure rock and pure iron composition. The grey dashed lines between the solid lines are mass-radius relations for water-rock and rock-iron composites. The red solid line is the lower limit of a planet radius as a result of collisional stripping \citep{2010ApJ...712L..73M}. K2-265 b has a composition consistent with a rocky terrestrial planet. }
\end{figure}
\footnotetext{\url{https://exoplanetarchive.ipac.caltech.edu/index.html}}
We performed a more detailed investigation of the composition of K2-265 b using the interior model of \cite{2017ApJ...850...93B}. This model considers planets made out of three differentiated layers: core (metals), mantle (rocks), and a liquid water envelope. Figure \ref{ternary} shows the possible compositions of K2-265 b inferred from the 1-$\sigma$ uncertainties on the planet’s mass and radius. By focusing on terrestrial compositions only (i.e. without any water), we show that the central mass and radius of the planet are best fitted with a rock mass fraction of 81\%, consistently with other theoretical predictions. However, given the uncertainties on the fundamental parameters, the rmf remains poorly constrained, namely within the 44–100\% range. If we assume that the stellar Fe/Si ratio (here $0.90\pm0.41$) can be used as a proxy for the bulk planetary value \citep{2015A&A...577A..83D,2017ApJ...850...93B}, this range is reduced to 60–83\%. In the case of a water-rich K2-265 b, the model only allows us to derive an upper limit on the planet’s water mass fraction (wmf). Indeed, given the high equilibrium temperature of the planet ($\sim1300$ K assuming an Earth-like albedo), water would be in the gaseous and supercritical phases, which are less dense than the liquid phase. From the uncertainties on the mass, radius, and bulk Fe/Si ratio of K2-265 b, we infer that this planet cannot present a wmf larger than 31\%.
\begin{figure}[htbp!]
\includegraphics[width=0.5\textwidth]{OD_composition.pdf}
\caption{\label{ternary}Ternary diagram showing the possible composition of K2-265 b. The thick black line is the allowed composition of the planet inferred from the central values of the planet’s mass and radius, and the dashed lines denote the deviations from this line allowed by the 1-$\sigma$ uncertainties on the fundamental parameters. The red line and area show the compositions compatible with the planetary bulk Fe/Si ratio derived for K2-265 b from the stellar value. Compositions of the Earth and Mercury are shown for reference.}.
\end{figure}
The California-\textit{Kepler} Survey (CKS) measured precise stellar parameters of Kepler host stars using spectroscopic follow-up \citep{2017AJ....154..108J}, and refined the planetary radii to study the planet size distribution and planet occurrence rate \citep{2017AJ....154..109F}. The survey has revealed a bimodal distribution of small planet sizes. Planets tend to have radii of either $\approx1.3 \rm ~R_{\oplus}$ or $\approx2.4 \rm ~R_{\oplus}$, with a deficit of planets at $\approx1.8 \rm ~R_{\oplus}$. The survey confirms the prediction by \citet{2013ApJ...775..105O}, whereby a gap in the planetary radius distribution exists as a consequence of atmospheric erosion by the photoevaporation mechanism. Alternatively, the core-powered mass loss mechanism could also drive the evaporation of small planets \citep{2016ApJ...825...29G, 2018MNRAS.476..759G}.
Due to its close proximity to the host star, the super-Earth K2-265 b is exposed to strong stellar irradiation. The planet’s gas envelope could be evaporated as a result. This process was observed in a number of systems (e.g. HD209458 b ; \citealt{2003Natur.422..143V}, GJ 436 b; \citealt{2015Natur.522..459E}). The present irradiance of the planet is $\rm S = S_{\oplus}(L_s/L_{\odot})(AU/a)^2 = 9.32 \times 10^5~W m^{-2}$, where $\rm L_s$ and $\rm L_{\odot}$ are the luminosity of the star and the Sun, $\rm S_{\oplus}$ is the Solar irradiance on Earth, and $\rm a$ is the semi-major axis of the planet. The equilibrium temperature of the K2-265 b can be estimated using Equ.~1 of \cite{2007ApJ...667L.191L}: $\rm T_{eq} = T_{eff}(R_s/a)^{1/2}[f(1-A_B)]$, where f and $\rm A_B$ are the reradiation factor and the Bond albedo of the planet. Assuming an Earth-like Bond albedo $\rm A_B=0.3$ and that the incident radiation is redistributed around the atmosphere (i.e. $\rm f=1/4$), the equilibrium temperature of K2-265 b is $\rm T_{eff} \approx 1300$ K.
Indeed, K2-265 b lies below the lower limit of the photoevaporation valley as shown in the 2D radius distribution plot in Figure \ref{evaporationvalley}. This implies that the planet could have been stripped bare due to photoevaporation, revealing its naked core. This atmospheric stripping process is presumed to have occurred in the first $\approx100$ Myr since the birth of the planet when X-ray emission is saturated \citep{2012MNRAS.422.2024J}, after which the X-ray emission decays. We estimated the total X-ray luminosity of K2-265 over its lifetime, $E_x^{tot}$, using the X-ray-age relation of \cite{2012MNRAS.422.2024J}. Using the results of section \ref{age}, we adopted a mean age of 6.32 Gyr for the host star. The X-ray-to-bolometric luminosity ratio in the saturated regime for a $B-V = 0.743$ star is $\log{(L_x/L_{bol})} = -3.71 \pm 0.05 \pm0.47$. The corresponding turn-off age is $\log{\tau_{sat}} = 8.03 \pm 0.06 \pm 0.31$, where the decrease in X-ray emission follows a power law ($\alpha = 1.28 \pm 0.17$). Over the lifetime of the star, $E_x^{tot} = 6.70 \times 10^{45}$ ergs (assuming efficiency factor $\eta = 0.25$) and K2-265 b is expected to have lost 2.7\% of its mass under the constant-density assumption. K2-265 b has a predominantly rocky interior as shown in Figure \ref{ternary}. This indicates that the planet was likely formed inside the ice-line, and could have either migrated to its current orbital separation well before $\approx100$ Myr or accreted its mass locally \citep{2017ApJ...847...29O}.
\begin{figure}[htbp!]
\includegraphics[width=0.5\textwidth]{EPIC1496_evaporationvalley.pdf}
\caption{\label{evaporationvalley} Planet radius distribution as a function of orbital period. The grey circles denote the planet sample obtained from the CKS sample \citep{2017AJ....154..109F}. The blue dot-dashed line and the green dashed line indicate the peak of the bimodal distribution of the planet radius distribution, where planets tend to favour radii of $\sim 1.3~R_{\oplus}$ and $\sim 2.4~R_{\oplus}$ due to the photoevaporation mechanism. The red dotted line indicates the lower limit of the photoevaporation valley derived from \cite{2017ApJ...847...29O}.}
\end{figure}
K2-265 b is among the denser super-Earths below the photoevaporation gap. In addition to photoevaporation, giant impact between super-Earths could drive mass loss in the planetary atmosphere. Super-Earths are thought to have formed via accretion in gas discs, followed by migration and eccentricity damping due to their interactions with the gas disc (e.g. \citealt{2015ApJ...811...41L}), leading to densely packed planetary systems. As the gas disc disperses, secular perturbation between planets excites their eccentricity, triggering giant impacts between the bodies before the system becomes stable \citep{2014A&A...569A..56C}. Two planets of comparable sizes could collide at a velocity beyond the surface escape velocity \citep{2004ApJ...613L.157A,2009ApJ...700L.118M}. The impact could lead to a reduction in the planet envelope-to-core-mass ratio, hence an increase in the mean density and alteration of the bulk composition of the planet \citep{2015ApJ...812..164L, 2016ApJ...817L..13I}.
Discoveries of super-Earths have shown a diversity of small planets in the mass-radius diagram. Precise RV and photometric measurements with an accuracy of a few percent are necessary to put strong constraints on the planetary mass and radius, and provide a precise bulk composition. The core composition of the planet can be derived as a result. In particular, the mass fraction of a planetary core can inform us of the formation and evolution history of the planet. K2-265 b has a precisely determined mass (13\%) and radius (6\%), and the composition of the planet is consistent with a rocky planet. Its small radius and short orbital period suggest that K2-265 b could have been photoevaporated to a bare rocky core. Its high rock-to-mass fraction implies a planet formation within the ice line. Studying planets with an exposed core could provide valuable insight to planet formation via the core accretion mechanism. The increasing sample of small planets will help distinguish planet origins, identify types of mass loss mechanism, and probe the efficiency of atmospheric evaporation processes.
\begin{acknowledgements}
We thank the anonymous referee for the helpful comments which improved the manuscript.
DJAB acknowledges support by the UK Space Agency. DJA gratefully acknowledges support from the STFC via an Ernest Rutherford Fellowship (ST/R00384X/1).
This work was funded by FEDER - Fundo Europeu de Desenvolvimento Regional funds through the COMPETE 2020 - Programa Operacional Competitividade e Internacionalização (POCI), and by Portuguese funds through FCT - Fundação para a Ciência e a Tecnologia in the framework of the projects POCI-01-0145-FEDER-028953 and POCI-01-0145-FEDER-032113.
N.S. and O.D. also acknowledge the support from FCT and FEDER through COMPETE2020 to grants UID/FIS/04434/2013 \& POCI-01-0145-FEDER-007672, PTDC/FIS-AST/1526/2014 \& POCI-01-0145-FEDER-016886 and PTDC/FIS-AST/7073/2014\& POCI-01-0145-FEDER-016880.
S.G.S acknowledge support from FCT through Investigador FCT contract nr. IF/00028/2014/CP1215/CT0002.
SCCB also acknowledges support from FCT through Investigador FCT contracts IF/01312/2014/CP1215/CT0004.
E.D.M. acknowledges the support by the Investigador FCT contract IF/00849/2015/CP1273/CT0003.
RL has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement n.664931.
FF acknowledges support from PLATO ASI-INAF contract n.2015-019-R0.
This research was made possible through the use of the AAVSO Photometric All-Sky Survey (APASS), funded by the Robert Martin Ayers Sciences Fund. This publication makes use of data products from the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation. This publication makes use of data products from the Wide-field Infrared Survey Explorer, which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute of Technology, funded by the National Aeronautics and Space Administration. This research has made use of NASA's Astrophysics Data System Bibliographic Services.
This paper includes data collected by the K2 mission. Funding for the K2 mission is provided by the NASA Science Mission directorate.
This research has made use of the Exoplanet Follow-up Observation Program website, which is operated by the California Institute of Technology, under contract with the National Aeronautics and Space Administration under the Exoplanet Exploration Program.
This work has made use of data from the European Space Agency (ESA) mission
{\it Gaia} (\url{https://www.cosmos.esa.int/gaia}), processed by the {\it Gaia}
Data Processing and Analysis Consortium (DPAC,
\url{https://www.cosmos.esa.int/web/gaia/dpac/consortium}). Funding for the DPAC
has been provided by national institutions, in particular the institutions
participating in the {\it Gaia} Multilateral Agreement.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
The recent surge in the study of scalar field condensate dark matter (DM) is in part driven by the failure to detect individual Weakly Interacting Massive Particles (WIMPs) at the Large Hadron Collider, or in various direct detection experiments. There is another avenue for dark matter to manifest, as condensates of macroscopic size \cite{Kaup,RB,BreitGuptaZaks,CSW,SS,Friedberg,SS2,Liddel,Lee}. Electrically neutral boson particles, if they are a component of dark matter, can naturally form gravitationally bound Bose-Einstein condensate (BEC) bubbles below a critical temperature, due to quantum statistical effects. These entities are known either as \emph{oscillons} or \emph{boson stars}. It is quite interesting and important to understand the physical properties of such condensates to determine if they are a viable alternative to the more popular and widely investigated WIMPs. Thus it is important to develop analytical and numerical methods to study condensate formation, as well as their stability, evolution, and possible decay into component particles.
One widely studied example of scalar field DM forming BECs is the axion. Axion stars were considered first around 30 years ago, originally suggested to form from collapse of overdense miniclusters in the early universe \cite{HoganRees,KolbTkachev} (see also more recent simulations \cite{TkachevSim}). Since then, many properties of axion stars have been studied extensively; these include structural stability \cite{SikivieYang,ChavanisMR,ChavanisMR2,BB,BarrancoNS,ESVW,GHPW,Braaten,WilczekASt,Hertzberg} (including nonzero angular momentum \cite{Davidson,Sarkar,Hertzberg2}), the process of gravitational collapse \cite{Harko,ChavanisCollapse,ELSW,ELSW2,Levkov,Helfer,Michel}, and their decay through emission of relativistic particles \cite{ESW,MTY,BraatenEmission,EMSW,Hertzberg3,ELSWFDM}. There has recently been a significant amount of work regarding relativistic corrections more generally to the classical field description of axion stars \cite{GuthRelativistic,GRB,BraatenRelativistic,EMTWY}. Other authors have investigated possible connections to astrophysical radio sources \cite{Iwazaki,TkachevFRB}. In some scalar field models, boson stars can be extremely heavy and (if they are stable) could give rise to gravitational wave signatures \cite{GW1,GW2,GW3,GW4,GW5,GW6}. Clearly this is a field booming with new and interesting results. (For boson star reviews, see e.g. \cite{Jetzer,MielkeSchunck,Liebling}.)
Besides DM, there are other classes of problems in cosmology where condensate formation is important. Elementary Hermitian boson fields known as inflatons are postulated to drive cosmological inflation, the hypothesized early-time exponential expansion of the universe. In addition to this, a bosonic degree of freedom (termed quintessence) is posited to generate the energy density which causes the observed late time acceleration of the universe. Both inflatons and quintessence can fragment and form BEC bubbles in the same way as dark matter candidates discussed previously \cite{AminInf1,AminInf2,AminQuint}. If sufficiently long lived, these entities can play a crucial role during inflation and also at later stages of cosmological evolution.
Theoretical studies of Bose-Einstein condensation gained prominence in 1990’s, after the experimental
discovery of atomic BECs in systems of cold atoms \cite{Stoof,Arovas,PS}. Atomic condensates are described by the Gross-Pit\"aevskii (GP) equation, which is a form of the non-linear Schr\"odinger equation. The interatomic interactions could be attractive or repulsive, and the atoms could be placed on external potentials. Analytic solutions of the GP equation are difficult to obtain, and various approximation methods had to be employed. Static problems were more amenable to numerical studies, but dynamical questions like expansion, collapse, and decay of condensates required the employment of approximation methods.
The variational method is one approximation technique widely utilized in studies of atomic condensates. Recently, the variational method was also adopted for the study of gravitationally bound condensates of
bosons by Chavanis \cite{ChavanisMR}; he used a Gaussian ansatz to approximate the wavefunction and compared the results he obtained thus with numerical solutions of the GP equation \cite{ChavanisMR2}. This comparison of ans\"atze to the numerical solutions is imperative for the study of dynamical problems that are much more difficult to solve numerically. Some such dynamical problems that have been analyzed using the variational method are the collapse \cite{Harko,ChavanisCollapse,ELSW,ELSW2} and collisions \cite{Cotner,ELLSW} of BECs. In subsequent years, numerous authors have presented various ans\"atze for both static and dynamical problems, either to improve numerical agreement or computational efficiency \cite{GHPW,Hertzberg,Hertzberg2,Hertzberg3}.\footnote{A third approach, in which the exact wavefunction can be computed using an analytical expansion, was developed in \cite{Kling1,Kling2}. This approach has the dual advantage of arbitrary precision and analyticity, though it is still more computationally intensive than the variational approach.}
To our knowledge, the relative efficacy of one ansatz compared to another has not previously been considered in a rigorous fashion. Some ans\"atze are simpler computationally, others are more complicated. Some match numerical results for boson star masses but do not reproduce as well the radius. As the literature on boson stars becomes more complex, it becomes increasingly important to have at hand a tractable analytic approach, and to understand the benefits and weaknesses of different choices of approximate wavefunctions. This is the basic goal of the current project.
In this work, we will provide background information for both the time-independent and time-dependent variational methods (Section \ref{sec:VarMethod}). The latter is crucial for understanding dynamical processes, like collapse and decay. We will then analyze various classes of approximate wavefunctions used in the literature, providing comparisons across different ans\"atze (Section \ref{sec:Ansatze}) as well as comparison to exact numerical results (Section \ref{sec:NumMethod}). In this discussion, we will also propose a new ansatz with a free parameter, which can be varied to give excellent agreement by whichever measure is relevant to a particular scenario. Finally, we compare the nonrelativistic formulation to the underlying relativistic theory in Section \ref{sec:RB}, and examine where the Thomas-Fermi approximation (often used in the literature) breaks down. We conclude in Section \ref{sec:Conclusions}.
We will use natural units throughout, where $\hbar = c = 1$.
\section{Variational Method} \label{sec:VarMethod}
\subsection{General Formulation}
The variational method~\cite{Ritz} for finding approximate analytic solutions to eigenvalue problems was utilized as an important calculational tool in the early development of quantum theory. Its success in describing the ground state of Helium~\cite{Kellner} played an important role in establishing modern quantum mechanics as a viable theory.
As explained in quantum mechanics texts, the variational method involves the extremization of the expectation value of the Hamiltonian of the system $\bra{\Psi} H \ket{\Psi} $ with respect to variations of a class of candidate wave functions $\ket{\Psi}$. In the calculation, the norm $ \vev{\Psi|\Psi} $ is held fixed, and in the end one obtains approximate analytic solutions to the time-independent Schr\"odinger equation. A corresponding method to obtain solutions to the time-dependent Schr\"odinger equation is to extremize the action
$$
S= \int dt\,d^3r \left<\Psi(t)\left| i \frac{ \partial}{\partial t}-H \right|\Psi(t)\right>
$$
with respect to a class of time-dependent variational wave functions $ \left|\Psi(t)\right>$. This time-dependent variational method, first introduced by Dirac in 1934 \cite{Dirac}, is not usually described in standard texts of quantum mechanics, but it has found wide use in fields like nuclear physics~\cite{Kerman}, chemistry~\cite{Langhoff}, and quantum field theory
~\cite{Jackiw}.
In condensed matter discussions, the GP formalism is developed starting from an $N$-particle wave function where each particle is in the state $\phi(\vec{r})$. Following the derivations in \cite{PS}, the expectation value of the Hamiltonian
\begin{align}
H = \sum_i \left(\frac{p_i^2}{2m} + V(\vec{r})\right) + U_0 \sum_{i<j} \delta^3\left(\vec{r}_i - \vec{r}_j\right)
\end{align}
in the state $\chi \left(\vec{r}_1, ... , \vec{r}_n\right) = \prod_i \phi(\vec{r}_i)$ yields the energy functional
\begin{align}
E = N \int d^3 r \left( \frac{\left| \nabla \phi \right|^2}{2m} + V(\vec{r}) \left|\phi \right|^2 + \frac{U_0}{2} \left(N-1 \right) \left|\phi \right|^4 \right)
\end{align}
where $V(\vec{r})$ is an external potential, $U_0$ represents some contact interaction among the scalars, and $\phi$ is normalized to unity. Defining $\psi = \sqrt{N} \phi(\vec{r})$ and taking $N\gg1$ yields,
\begin{align} \label{GPenergy}
E = \int d^3 r \left( \frac{\left|\nabla \psi \right|^2}{2m} + V(\vec{r}) \left|\psi \right|^2 + \frac{U_0}{2} \left|\psi \right|^4 \right).
\end{align}
Extremizing the above equation, subject to the normalization constraint
\begin{equation} \label{Norm}
\int d^3 r \left|\psi \right|^2 = N
\end{equation}
yields
\begin{align} \label{GPmu}
\mu \,\psi = -\frac{1}{2m} \nabla^2 \psi + c\,V(\vec{r})\, \psi + U_0 \left|\psi \right|^2 \psi,
\end{align}
where $\mu$, the Lagrange multiplier introduced to maintain the normalization, is defined as the chemical potential. The constant $c=1$ if $V(\vec{r})$ does not depend on $\psi^*$ (true for an external trap), but can take other values elsewhere (e.g. for condensates bound by self-gravity). By multiplying eq. (\ref{GPmu}) by $\psi^*$ and performing a volume integral, we get
\begin{align} \label{muint}
\mu N = \int d^3 r \left(\frac{\left|\nabla\psi\right|^2}{2m}+ c\, V(\vec{r})\,\left|\psi\right|^2
+ U_0 \left|\psi \right|^4\right).
\end{align}
Comparing eqs. (\ref{GPenergy}) and (\ref{muint}), we can conclude the chemical potential is not equal to the energy per particle given in eq. (\ref{GPenergy}) if the interaction strength $U_0 \neq 0$ or if the trapping potential $V(\vec{r})$ depends on $\psi^*$ (i.e. if $c\neq1$).
In the time-dependent case, the GP formalism starts with the action $S = \int dt L$, where
\begin{align} \label{L(t)}
L = \int d^3 r \left( i \psi^* \dot{\psi} - i \psi \dot{\psi^*} - \frac{\left| \nabla \psi \right|^2}{2m} - V(\vec{r})\left| \psi \right|^2 - \frac{U_0}{2} \left|\psi \right|^4\right).
\end{align}
The Hamiltonian is
\begin{align}
H &= p_{\psi} \dot{\psi} + p_{\psi^*} \dot{\psi^*} - L \\
&= \int d^3 r \left( \frac{\left|\nabla \psi \right| ^2}{2m} + V(\vec{r}) \left| \psi \right|^2 + \frac{U_0}{2} \left|\psi \right|^4 \right)
\end{align}
which is of course identical to the energy functional defined in equation (2.3). Variation of eq. (\ref{L(t)}) with respect to $\psi^*$ gives the Gross-Pit\"aevskii (GP) equation
\begin{equation} \label{GPtime}
i\,\frac{\partial \psi}{\partial t} = -\frac{1}{2m} \nabla^2 \psi + c\,V(\vec{r})\, \psi + U_0 \left|\psi \right|^2 \psi,
\end{equation}
which reduces to the time-independent case of eq. (\ref{GPmu}) if the time dependence of $\psi$ is given by the factor $\exp(-i\,\mu\,t)$. This is also in agreement with the the identification of $\mu$ as the chemical potential.
The variational method has been applied to study time-independent as well as time-dependent solutions to the GP equation. As we will describe in Section \ref{sec:Ansatze}, in this case one describes the wavefunction by some ansatz $\psi(\vec{r})$, so that, for a given physical system, eq. (\ref{GPenergy}) can be extremized analytically. Of course, for many physical applications the solutions to the time-independent equations can be found numerically to an arbitrary precision; therefore the variational method offers qualitative understanding of the systems, but falls short of the numerical accuracy.
On the other hand, dynamical problems are more difficult to solve using numerical methods. Time-dependent solutions to the GP equation can be approximated using the variational method formulation of Pethick and Smith \cite{PS}, which we have described above. A good example of such an application is the collapse of BECs \cite{Harko,ChavanisCollapse,ELSW,ELSW2}. In this case, the variational parameter $\sigma$ is taken to be a function of time, $\sigma(t)$, while the wavefunction is multiplied by some phase. This phase depends on the velocity field of the system, which for spherical collapse is taken to be in the radial direction and proportional to $r$. It also depends on some parameter (analogous to the Hubble parameter in cosmology) which is related to $\sigma(t)$ by $H(t) = \dot{\sigma}(t)/\sigma(t)$. It is remarkable that using a simple analytic description of this type, one can assess many of the relevant features of BEC collapse. This illustrates the power of the time-dependent variational formalism.
\subsection{Application to Boson Stars}
Dynamics of boson stars are described by the Klein-Gordon equation in the presence of self-gravity. This action,
\begin{align}
S = \int \sqrt{-g} \,dt \,d^3 r \left[\frac{1}{2} g_{\mu \nu} \left(\partial^{\mu} \phi\right) \left(\partial^{\nu} \phi \right) - \frac{m^2}{2} \phi^2 - \frac{\lambda}{4!} \phi^4\right]
\end{align}
in the nonrelativistic limit and with the replacement ${\phi = \frac{1}{\sqrt{2m}} \left(e^{-i m t} \psi + h.c.\right)}$ gives
\begin{align} \label{Action}
S = \int dt\, d^3r \Big[&i \psi \frac{ \partial \psi^*}{\partial t}-i \psi^* \frac{\partial \psi}{\partial t} + \frac{|\nabla \psi|^2}{2m} \nonumber \\
&+ \frac{1}{2} V_{g}\,|\psi|^2
+ \frac{\l}{16\,m^2}\left|\psi\right|^4 \Big].
\end{align}
The gravitational potential
\begin{equation} \label{Poisson}
\nabla^2 V_g = 4\pi\,G\,m^2\,|\psi|^2
\end{equation}
originates from the nonrelativistic limit of the Einstein field equations with $G=M_P{}^{-2}$ Newton's gravitational constant. (See Section \ref{sec:RB} for a more thorough description of the nonrelativistic limit.) In this work, we will consider the case of a quartic self-interaction, where self-coupling $\lambda$ can be positive (giving rise to a repulsive interaction) or negative (attractive interaction). Variation of eq. (\ref{Action}) with respect to $\psi^*$ yields a GP-type equation comparable to eq. (\ref{GPtime}).
\begin{equation} \label{GPtime2}
i \frac{ \partial \psi}{\partial t} = - \frac{1}{2m}\nabla^2 \psi
+ V_{g}\,\psi
+ \frac{\l}{8\,m^2}\left|\psi\right|^2 \psi.
\end{equation}
Finding analytic solutions to the above non-linear equations (\ref{Poisson}) and (\ref{GPtime2}), known as the Gross-Pit\"aevskii+Poisson (GPP) system, is a challenging task. In order to find an approximate analytic solution one could extremize the action $S$ within a class of variational wave functions as described previously in condensed matter/atomic physics applications.
Assuming a harmonic time dependence for the wavefunction, the chemical potential of eq. (\ref{muint}) is given by
\begin{equation} \label{muint2}
\mu\,N = \int d^3 r \left(\frac{\left|\nabla\psi\right|^2}{2m}+ V_{g}\,\left|\psi\right|^2
+ \frac{\l}{8\,m^2} \left|\psi \right|^4\right).
\end{equation}
while the energy of eq. (\ref{GPenergy}) is
\begin{equation} \label{Energy}
E = \int d^3r \left(\frac{|\nabla \psi|^2}{2m}
+ \frac{1}{2} V_{g}\,|\psi|^2
+ \frac{\l}{16\,m^2}\left|\psi\right|^4 \right).
\end{equation}
The latter can be extremized by assuming some $r$ dependence for $\psi$, giving rise to an approximate bound state solution. We present a number of prevalent choices for such an ansatz in the next section. We will focus here on the time-independent case, because it will allow us to analyze the radial dependence of the wavefunction; any relevant time-dependent factors will depend on the particular application one considers (e.g. collapse) and can be added on later.
In order to begin, we will need to impose some conditions on the classes of wavefunction ans\"atze we consider. Assuming that the wave function $\psi(r) $ is differentiable to all orders at $r=0$ and the gravitational potential $V_g$ has a Laurent expansion in $r$, the GPP equations (\ref{Poisson}) and (\ref{GPtime2}) imply that the gravitational potential is an analytic function in the variable $ r^2 $ at the origin; consequently, the first derivative of the wave function $\psi(r) $ vanishes there. At large $r$, the interaction term in the GP equation becomes negligibly small, and the GP equation reduces to the linear Schr\"odinger equation for a particle moving in the gravitational potential $V_g$. The potential at large $r$ takes the $G M/ r $ Newtonian form, and the wave functions asymptotically are nearly hydrogen-like, with a $\exp(-kr) $ behavior (though see below). We impose these boundary conditions when numerically solving the GPP equations. A successful variational ansatz should also exhibit similar behavior.
Actually, the behavior of the wavefunction at large $r$ is not exactly exponential, but can be calculated in a straightforward way. First, at large $r$ we throw away the terms in the equation of motion that are higher order in the fields than the gravity term; this gives eq. (\ref{GPtime2}) with $\lambda=0$ and $V_g = -G M m / r$. This equation has an exact solution in terms of hypergeometric functions, but the leading order solution is proportional to
\begin{equation} \label{asymptotic}
\psi \propto e^{-\sqrt{-2\,\mu\,m}\,r}
r^{\frac{G\,m^2\,M}{\sqrt{-2\,\mu\,m}} - 1}.
\end{equation}
This is in agreement with the results of \cite{Kling1,Kling2}.
Compare and contrast this case to that of atomic Bose condensates trapped by external harmonic oscillator potentials. In the latter situation, the GP wave function at large $r$, when interaction terms are negligible, should approach Gaussian form, making such functions ideal candidates for variational ans\"atze. This makes clear the fact that the optimal ansatz for the wavefunction should depend on the potential.
\section{Ans\"atze for Boson Star Wavefunctions} \label{sec:Ansatze}
\subsection{Generalized Ansatz}
In \cite{ELSW,ELSW2} we performed an analysis using a general ansatz for the wavefunction of an axion star, under only the (weak) assumptions that the wavefunction is finite at the origin, spherically symmetric, and decreases monotonically with the radial coordinate $r$. Here, we apply the same method to a general boson star with a 4-point coupling $\l$.
The starting point is the rescaling of the macroscopic quantities
\begin{equation} \label{scaling}
\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda = \sqrt{|\l|}\frac{M_P}{m^2}\,\rho \qquad N = \frac{M_P}{m\,\sqrt{|\l|}}\,n.
\end{equation}
where $\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda$ is a variational distance parameter explained in detail in the next section. Then a general ansatz for a boson star can be written in the form
\begin{equation}\label{ansatz}
\psi(r)= w\, F\left(\frac{r}{\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda}\right)\equiv \psi(0)\,F(\xi),
\end{equation}
where at fixed $\xi$ the function $F(\xi)$ is independent of $\rho$, $n$, and $\l$. Because we can fix $F(0)=1$ without loss of generality, we are able to identify $w = \psi(0)$, the central value of the wavefunction. Substituting the ansatz into the normalization condition (\ref{Norm}) gives the central value as
\begin{equation}
\psi(0) = \frac{m^{5/2}}{M_P\,|\l|}\,\sqrt{\frac{n}{\rho^3\,C_2}},
\end{equation}
where we introduced the notation
\begin{equation}
C_k = 4\,\pi\int_0^\infty d\xi\,\xi^2 \,F(\xi)^k.
\end{equation}
Using eq. (\ref{Energy}) and the general ansatz of eq. (\ref{ansatz}), we obtain for the energy functional
\begin{equation}\label{binding}
\frac{E(\rho)}{m\,N} = \frac{m^2}{M_P{}^2\,|\l|}\left(\frac{D_2}{2\,C_2}\frac{1}{\rho^2}-\frac{B_4}{2\,C_2{}^2}\,\frac{n}{\rho}
+ \text{sgn}(\l) \frac{C_4}{16\,C_2{}^2}\frac{n}{\rho^3}\right),
\end{equation}
and, using eq. (\ref{muint2}), we obtain for the chemical potential
\begin{equation}\label{Chemical}
\frac{\mu(\r)}{m} = \frac{m^2}{M_P{}^2\,|\l|}\left(\frac{D_2}{2\,C_2}\frac{1}{\rho^2}-\frac{B_4}{C_2{}^2}\,\frac{n}{\rho}
+ \text{sgn}(\l) \frac{C_4}{8\,C_2{}^2}\frac{n}{\rho^3}\right).
\end{equation}
For simplicity, we have defined the dimensionless parameters
\begin{align}
D_2&=4\,\pi\int_0^\infty d\xi\,\xi^2 \,F'(\xi)^2,\\
B_4&=32\,\pi^2\int_0^\infty d\xi\,\xi \,F(\xi)^2\int_0^\xi d\eta\, \eta^2\,F(\eta)^2.
\end{align}
Recall once more that $\l>0$ ($\l<0$) will correspond to repulsive (attractive) self-interactions.
Because the GP equation can be derived from the variation of the total energy, a solution will be a stationary point of (\ref{Energy}). Given an ansatz for the wavefunction, we approximate this exact solution by minimizing eq. (\ref{binding}) with respect to $\rho$. This procedure gives the dilute boson star radius
\begin{equation} \label{rhodilute}
\rho_d = \frac{C_2\,D_2}{B_4\,n}\left[1 + \sqrt{1 + \text{sgn}(\l) \frac{n^2}{\bar{n}^2}}\right].
\end{equation}
where $\bar{n}$ is an important scaled particle number given by
\begin{equation} \label{nCrit}
\bar{n} = \sqrt{\frac{8}{3}}\,\frac{C_2\,D_2}{\sqrt{B_4\,C_4}}.
\end{equation}
For $\l<0$, $\bar{n}$ determines the mass of the maximum stable configuration, and so we will denote it by $\bar{n}=n_c$. For $\l>0$, there is no maximum mass in the nonrelativistic limit, so we will instead use the notation $\bar{n}=n_*$; in this case, self-interactions become increasingly important for $n\sim n_*$, eventually approaching the region in which the Thomas-Fermi approximation is relevant. We will discuss these points in detail below.
Eq. (\ref{rhodilute}) is a stable minimum of the energy classically and a metastable solution in quantum theory. There exists another root of the energy at a radius of
\begin{equation} \label{rhounstable}
\rho_{u} = \frac{C_2\,D_2}{B_4\,n}\left[1 - \sqrt{1 + \text{sgn}(\l) \frac{n^2}{\bar{n}^2}}\right],
\end{equation}
which we will discuss later. In the case of attractive self-interactions, $\r_u>0$ is an unstable maximum of the energy; for repulsive interactions, $\r_u < 0$ is an unphysical configuration.
Above, we have used the formulation of our previous works \cite{ELSW,ELSW2}, in which the dimensionful parameters are scaled out; then the numerical constants $B_4$, $C_k$, and $D_2$ are dimensionless and depend only on the shape of the ansatz being employed. Other authors, e.g. \cite{ChavanisMR,EKNW,ESVW2}, use a different formulation; for ease of comparison we provide the expressions for translating between the two in Appendix \ref{AppB}.
This formulation has a number of useful applications. For example, in the limit of no self-interactions ($\lambda\to0$)\footnote{Of course, in the $\lambda\to0$ limit the scaling of eq. (\ref{scaling}) is not appropriate. In that case one should use $\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda = \rho/m$ and $N = M_P{}^2\,n/m^2$, which gives rise to the analogous equations for $E$ and $\mu$ without the interaction term. In the end, one finds the first two terms in eqs. (\ref{binding}) and (\ref{Chemical}) do not change, which is all that is needed here.}, there is a simple relation between the energy and chemical potential:
\begin{align} \label{muoverE_NI}
\frac{N\,\m(\r_d)}{E(\r_d)} &= \frac{\frac{D_2}{2\,C_2}\frac{1}{\rho_d^2}-\frac{B_4}{C_2{}^2}\,\frac{n}{\rho_d}}
{\frac{D_2}{2\,C_2}\frac{1}{\rho_d^2}-\frac{B_4}{2C_2{}^2}\,\frac{n}{\rho_d}} \nonumber \\
&= \frac{-3/8}{-1/8} \nonumber \\
&= 3,
\end{align}
which was derived using other methods in \cite{Membrado,ChavanisMR}.
For attractive interactions, we of course recover the standard result that the local minimum and maximum of the energy become equal at $n=n_c$.
For $n>n_c$, no stable solutions exist.
The value $n=n_c$ also corresponds to a critical minimum radius,
\begin{equation} \label{rhoCrit}
\rho_c = \sqrt{\frac{3 C_4}{8B_4}}.
\end{equation}
It is useful to define a parameter
\begin{align}
\delta = \sqrt{1 - \left(\frac{N}{N_c}\right)^2} = \sqrt{1 - \left(\frac{n}{n_c}\right)^2}
\end{align}
which parameterizes the closeness of $N$ to $N_c$. Then substituting $\rho_d$ from eq. (\ref{rhodilute}) into eqs. (\ref{binding}) and (\ref{Chemical}), the energy and chemical potential at the minimum for attractive self-interactions can be written in the form
\begin{align}
\frac{E(\rho_d)}{m \,N} &= -\frac{1-\delta}{1+\delta}(1+2\,\delta) \frac{m^2}{M_P{}^2 |\lambda|} \frac{4D_2}{9C_2}\frac{B_4}{C_4}\nonumber\\
\frac{\mu(\rho_d)}{m} &= -\frac{1-\delta}{1+\delta}(5+4\,\delta) \frac{m^2}{M_P{}^2 |\lambda|} \frac{4D_2}{9C_2}\frac{B_4}{C_4}.
\end{align}
This implies a simple relationship between the energy and the chemical potential that is independent of the choice of ansatz,
\begin{equation}
\frac{N\,\mu(\rho_d)}{E(\rho_d)}=\frac{5+4\,\delta}{1+2\,\delta},
\end{equation}
which is an analogue of the non-interacting result in eq. (\ref{muoverE_NI}), but applied to attractive self-interactions.
We can see that when $\delta\to0$ ($N\to N_c$), the ratio goes exactly to $5$. Even though the derivation assumed some ansatz for the variational approach, the result does not depend on what form the wavefunction takes, and so it holds even in the exact case.
In the case of repulsive self-interactions, there also exists a critical mass, though it is not at $n=n_*$ as defined above; it arises due to relativistic effects which we do not consider here \cite{CSW}. Relativistic effects can be taken into account using the Ruffini-Bonazzola (RB) formalism for analyzing boson stars \cite{RB}. One could in principle formulate a variational method which approximated the relativistic equations of motion, in which case these effects would become apparent. We leave such an analysis for future work. Of course, at weak gravity and small binding energy, the RB equations of motion reduce to the GPP system, as we will describe in Section \ref{sec:RB}.
\subsection{Non-Compact Ans\"atze}
It is well-known by direct solution of eqs. (\ref{Poisson}) and (\ref{GPtime2}) that the wavefunction of a boson star does not typically have compact support; it is nonzero for all $r\geq 0$ though decreases extremely fast at large $r$. The standard definition for the ``size" of a boson star is $R_{99}$, the radius inside which $0.99$ of the mass is contained. A number of non-compact ans\"atze have appeared in the literature to approximate the exact solution; a few of the most popular ones are
\begin{widetext}
\begin{align} \label{eq:ansatze}
\psi_A(r) =
\begin{dcases}
\sqrt{\frac{N}{\pi^{3/2}\,\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda^{3}}}\,e^{-r^2/2\,\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda^2} & \text{(Gaussian (G) \cite{ChavanisMR,ChavanisCollapse,ELSW,ELSW2,ELLSW,Hertzberg2,Hertzberg3})} \\
\sqrt{\frac{N}{\pi\,\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda^3}}\,e^{-r/\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda} & \text{(Exponential (E) \cite{GHPW,Hertzberg})} \\
\sqrt{\frac{N}{7\,\pi\,\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda^3}}\left(1 + \frac{r}{\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda}\right)\,e^{-r/\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda} & \text{(Linear + Exponential (LE) \cite{Hertzberg})} \\
\sqrt{\frac{3\,N}{\pi^3\,\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda^3}}\,\text{sech}\left(\frac{r}{\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda}\right) & \text{(Sech (S) \cite{Hertzberg,Hertzberg3})} \\
\sqrt{\frac{N}{\pi\,\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda^3}}\left(1 + \frac{1}{a^5} - \frac{16}{a(1+a)^3}\right)^{-1/2}\,
\left(e^{-r/\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda} - \frac{1}{a}e^{-a\,r/\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda}\right) & \text{(Double Exponential (DE$_a$))}
\end{dcases}
\end{align}
\end{widetext}
On the right we give the long name (e.g. ``Gaussian") and short abbreviation (e.g. ``G") for each ansatz, and cite a collection of previous works which utilize them in the study of boson stars. We will use the notation that, for example, $\psi_G(r)$ is the Gaussian wavefunction, $\psi_{LE}(r)$ is the linear+exponential wavefunction, etc. The first four functions listed here are popular choices in the literature; the last one (the double exponential) is a proposal of ours with some constant parameter $a$ which can be fixed by matching to the exact solutions. We will show in this work that the double exponential can be optimized for a given numerical result.
Note the appearance of $\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda$ in each of the ans\"atze. While this parameter has units of distance, it should not be confused with the radius of a boson star. For each ansatz, the parameter $\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda_d$ of the solution is related by some constant factor to the radius of the boson star $R_{99}$; we will label this by a real number $\kappa$, i.e.
\begin{equation}
\kappa\ = \frac{R_{99}}{\sigma_d},
\end{equation}
where both $\sigma_d$ and $\kappa$ depend on the ansatz under consideration.
Other reasonable distance scales in a calculation like this one include the expectation values
\begin{align}
\vev{r} &\equiv \frac{1}{N}\int d^3r\,r\,|\psi(r)|^2 \nonumber \\
\vev{r^2} &\equiv \frac{1}{N}\int d^3r\,r^2\,|\psi(r)|^2,
\end{align}
which will be useful for comparing to exact solutions later. Because the translation to a physical length is different for a different ansatz, the parameter $\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda$ means something different depending on the ansatz in which it is employed; said a different way, $\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda$ itself is unphysical and should not be compared across ans\"atze.
\begin{table*}
\centering
\begin{tabular}{| c || c | c | c | c | c | c | c |}
\hline
& G
& E
& LE
& S
& DE$_2$
& DE$_{3\pi/8}$ \\
\hline \hline
$F(\xi)$ & $\dis{e^{-\xi^2/2}}$
& $\dis{e^{-\xi}}$
& $\dis{\left(1 + \xi \right)\,e^{-\xi}}$
& $\dis{\text{sech}\left(\xi\right)}$
& $\dis{2 e^{-\xi} - e^{-2\xi}}$
& $\dis{\approx{6.6e^{-\xi} - 5.6e^{-1.2\xi}}}$
\\ \hline
$B_4$ & $\dis{\sqrt{2\pi}\,\pi^2}$
& $\dis{\frac{5\,\pi^2}{8}}$
& $\dis{\frac{5373\,\pi^2}{256}}$
& $\dis{8\pi\left(2\pi\zeta(3) - \frac{\pi^3}{6}\right)}$
& $\dis{\approx{47}}$
& $\dis{\approx{61}}$
\\ \hline
$C_2$ & $\dis{\pi^{3/2}}$
& $\dis{\pi}$
& $\dis{7\pi}$
& $\dis{\frac{\pi^3}{3}}$
& $\dis{\approx{9.2}}$
& $\dis{\approx{17}}$
\\ \hline
$C_4$ & $\dis{\left(\frac{\pi}{2}\right)^{3/2}}$
& $\dis{\frac{\pi}{8}}$
& $\dis{\frac{437\pi}{256}}$
& $\dis{\frac{2\pi}{9}\left(\pi^2-6\right)}$
& $\dis{\approx{2.1}}$
& $\dis{\approx{4.2}}$
\\ \hline
$D_2$ & $\dis{\frac{3\,\pi^{3/2}}{2}}$
& $\dis{\pi}$
& $\dis{3\pi}$
& $\dis{\frac{\pi(12+\pi^2)}{9}}$
& $\dis{\approx{6.7}}$
& $\dis{\approx{8.7}}$
\\ \hline
$\bar{n}$ & $\dis{2\pi\,\sqrt{3}}$
& $\dis{\sqrt{\frac{512\pi}{15}}}$
& $\dis{\frac{3584}{3}\sqrt{\frac{2\pi}{86963}}}$
& $\dis{\frac{2}{3}\frac{\left(\frac{1}{6} + \frac{2}{\pi^2}\right)\pi^{9/2}}{\sqrt{[\pi^2 - 6][12\zeta(3) - \pi^2]}}}$
& $\dis{\approx{10.1255}}$
& $\dis{\approx{10.1518}}$
\\ \hline
$\r_c$ & $\dis{\sqrt{\frac{3}{32\pi}}}$
& $\dis{\sqrt{\frac{3}{40\pi}}}$
& $\dis{\frac{1}{6}\sqrt{\frac{437}{398\pi}}}$
& $\dis{\frac{\sqrt{[\pi^2 - 6][12\zeta(3) - \pi^2]}}{\sqrt{\pi}[48\zeta(3) - 4\pi^2]}}$
& $\dis{\approx{0.13}}$
& $\dis{\approx{0.11}}$
\\ \hline
$F'(0)$ & 0
& $<0$
& 0
& 0
& 0
& 0
\\ \hline
$\dis{\k}$ & $\dis{2.8}$
& $\dis{4.2}$
& $\dis{5.4}$
& $\dis{3.3}$
& $\dis{3.4}$
& $\dis{4.7}$
\\ \hline
\hline
\end{tabular}
\caption{Table of various parameters for the ans\"atze under consideration. Exact solutions for the double-exponential are given in Appendix \ref{AppC} for arbitrary values of the parameter $a$.}
\label{tab1}
\end{table*}
\subsection{Compact Ans\"atze} \label{compact_ansatze}
It is sometimes advantageous to employ a compact function to approximate the wavefunction. An example of such a case is the Thomas-Fermi (TF) limit of repulsive interactions, where one neglects the kinetic energy term, and the resulting wavefunction is very close to having an exact radius $R$. We also found a compact ansatz advantageous when discussing collapse of an axion star through collisions with astrophysical sources \cite{ELLSW}. Here, we include a few compact ans\"atze for completeness:
\begin{align}
\psi_A(r) =
\begin{dcases}
\sqrt{\frac{4\,\pi\,N}{(2\pi^2 - 15)\,R^3}}\,\cos^2\left(\frac{\pi\,r}{2\,R}\right) & \text{(Cos$^2$ \cite{ELSW,ELLSW})} \\
\sqrt{\frac{\pi N}{4 R^3} \frac{\sin\left(\pi r/ R\right)}{\left(\pi r/R\right)}} & \text{(TF \cite{BH_TF})}
\end{dcases}
\end{align}
However, we do not include them in our analysis, partly because it is somewhat problematic to compare $R_{99}$ for non-compact ans\"atze to the exact radius for compact ans\"atze. It should also be noted that one must require a compact ansatz to vanish above $r = R$. The TF ansatz given above is the exact solution to the GPP equations in the Thomas-Fermi limit \cite{BH_TF}. However, given that the solution has an infinite derivative at $r = R$, we do not recommend its use as an ansatz.
\section{Comparison with Numerical Calculation} \label{sec:NumMethod}
\subsection{Numerical Algorithm}
To investigate the efficacy of these ans\"atze further, we have to solve the full equations of motion (\ref{Poisson}) and (\ref{GPtime2}) in the stationary limit. We employ the following scaling to dimensionless (``tilde-d") variables:
\begin{align} \label{eq:numscale}
\psi &= \frac{m^{5/2}}{M_P\,|\l|}\,\tilde{\psi} \nonumber \\
V_g &= \m + \frac{m^3}{M_P{}^2\,|\l|}\,\tilde{V} \nonumber \\
r &= \frac{\sqrt{|\l|}\,M_P}{m^2}\,\tilde{r}
\end{align}
Writing the equations of motion in terms of the rescaled quantities, we solve the resulting system numerically:
\begin{align}
\tilde{\nabla}^2 \tilde{\psi} &= 2\,\tilde V\,\tilde{\psi} + \frac{\text{sgn}(\l)}{4} \tilde{\psi}^3, \nonumber \\
\tilde{\nabla}^2 \tilde{V} &= 4\pi\,\tilde{\psi}^2.
\end{align}
We employ a shooting method to determine correct boundary conditions $\tilde{\psi}\to 0$ and $\tilde{V}\to constant$ as $\tilde{r}\to\infty$. In practice, in the numerical routine $\tilde\psi$ and $\tilde{V}$ converge up to some finite radius $\tilde{r}_0$, which can be as large as the precision of the calculation requires.
After solving for $\tilde\psi$ and $\tilde V$, one can calculate the dimensionless macroscopic quantities using
\begin{align}
n &= \int_0^{\tilde{r}_0} d\tilde{r}\,4\pi\,\tilde r^2\,\tilde{\psi}^2 \nonumber \\
.99\,n &= \int_0^{\tilde{R}_{99}} d\tilde{r}\,4\pi\,\tilde{r}^2\,\tilde{\psi}^2 \nonumber \\
\tilde\mu &= - \lim_{\tilde{r}\to\tilde{r}_0} \tilde{V},
\end{align}
which determine the rescaled number, radius, and chemical potential (respectively). To convert back into standard physical units, we use
\begin{align} \label{macro}
M &= \frac{M_P}{\sqrt{|\l|}}\,n , \nonumber \\
R_{99} &= \frac{M_P}{m^2}\sqrt{|\l|}\,\tilde{R}_{99} , \nonumber \\
\mu &= \frac{1}{|\l|}\frac{m^3}{M_P{}^2}\,\tilde\mu.
\end{align}
Before describing the solutions, note that the GPP energy can be written in terms of tilde-d quantities,
\begin{align} \label{DimEnergy}
E(\tilde{\psi}) = \frac{m^2}{M_P\,|\l|^{3/2}} \int d^3\tilde r \Big[\frac{1}{2} \tilde\nabla & \tilde\psi|^2
+ \frac{1}{2} \left(\tilde V + \tilde\mu\right)\,\tilde\psi^2 \nonumber \\
&+ \frac{\text{sgn}(\l)}{16}\tilde\psi^4 \Big].
\end{align}
Using $M = m\,N$ and eq. (\ref{macro}) for $\tilde N$, we can write the energy per particle as
\begin{align}
\frac{E(\tilde{\psi})}{m\,N} = \frac{m^2}{M_P{}^2\,|\l|} \frac{1}{n}\int d^3\tilde r\,\Big[
&\frac{1}{2}|\tilde\nabla \tilde\psi|^2
+ \frac{1}{2} \left(\tilde V + \tilde\mu\right)\,\tilde\psi^2 \nonumber \\
&+ \frac{\text{sgn}(\l)}{16}\tilde\psi^4\Big].
\end{align}
\begin{comment}
We can then identify the contributions to the energy from the kinetic, gravitational, and interaction components term by term
\begin{align}
&\tilde{K} \equiv \frac{1}{2}\frac{\int\,d\tilde r\,\tilde r^2\,|\tilde\nabla \tilde\psi|^2}{\int d\tilde r\,\tilde r^2\,\tilde\psi^2}
\qquad \qquad \qquad \leftrightarrow \qquad \qquad
\frac{D_2}{2\,C_2}\,\frac{1}{\r_d{}^2} \equiv K \nonumber \\
&\tilde{G} \equiv \frac{1}{2}\frac{\int d\tilde r\,\tilde r^2\,\left(\tilde V + \tilde\mu\right)\,\tilde\psi^2}{\int d\tilde r\,\tilde r^2\,\tilde\psi^2}
\qquad \quad \leftrightarrow \qquad \qquad
\frac{B_4}{2\,C_2{}^2}\,\frac{n}{\r_d} \equiv G \nonumber \\
& \tilde{\Ical} \equiv \frac{1}{16}\frac{\int\,d\tilde r\,\tilde r^2\,\tilde\psi^4}{\int d\tilde r\,\tilde r^2\,\tilde\psi^2}
\qquad \qquad \qquad \quad ~ \leftrightarrow \qquad \qquad
\frac{C_4}{16\,C_2{}^2}\,\frac{n}{\r_d{}^3} \equiv \Ical.
\end{align}
The LHS can be computed using the numerical solutions to arbitrary precision; the RHS is ansatz dependent. Note that the comparison must be performed at fixed $N$, which is not difficult since $\rho_d$ depends only on $n$.
\end{comment}
A similar procedure for the chemical potential gives
\begin{align}
N\,\mu(\tilde\psi) = \frac{m^2}{M_P\,|\l|^{3/2}} \int d^3\tilde r \Big[\frac{1}{2}|\tilde\nabla& \tilde\psi|^2
+ \left(\tilde V + \tilde\mu\right)\,\tilde\psi^2 \nonumber \\
&+ \,\frac{\text{sgn}(\l)}{8}\,\tilde\psi^4 \Big],
\end{align}
so that, after dividing by $N$ and $m$, we get
\begin{align}
\frac{\mu(\tilde{\psi})}{m} = \frac{m^2}{M_P{}^2\,|\l|} \frac{1}{n}\int d^3\tilde r\,\Big[
&\frac{1}{2}|\tilde\nabla \tilde\psi|^2
+ \left(\tilde V + \tilde\mu\right)\,\tilde\psi^2 \nonumber \\
&+ \frac{\text{sgn}(\l)}{8}\tilde\psi^4 \Big].
\end{align}
Of course, because $\mu$ appears on both the LHS and RHS, this equation can be used as a consistency check on the numerical solutions. We have verified for every numerical solution that the eigenvalue $\mu$ satisfies this constraint.
The important physical quantities describing a boson star are its mass and radius. In Figure \ref{fig:MassRadius}, we show the exact numerical relationship between these quantities, along with the result from various ans\"atze. For attractive interactions, we observe that $\psi_E$ tends to be a good fit for small $R_{99}$ (on the unstable branch of solutions) but a poor fit for large $R_{99}$ (the stable branch). It is also interesting to note that $\psi_G$ is used quite often in the literature, but is a poor fit for small $R_{99}$ and badly approximates the position of the maximum mass. The double exponential ansatz has a free parameter $a$, and we find that we can match the value of $n_c$ almost exactly by taking the value $a = 3\pi/8$.
One interesting feature that is apparent in the right panel of Figure \ref{fig:MassRadius} is that, in the case of repulsive interactions, the radius at large $M$ approaches a constant. This is the limit of the Thomas-Fermi approximation, which will be discussed in Section \ref{sec:RB}. When using the variational approach, the radius $R_{TF}$ to which a given ansatz approaches at large $M$ in the repulsive case is equal to $R_c$, the radius at which one obtains the maximum mass in the attractive case. (This is clear by examination of eq. (\ref{rhodilute}) as well.) This relationship between these two radii in very different limits is a consequence of the fact that we use the same form for the ansatz in both the repulsive and attractive interaction cases. As a result, it does not hold true for the numerical solutions.
\begin{figure*}
\includegraphics[scale=.53]{MvsR99_Comp_v5}
\includegraphics[scale=.55]{MvsR99_Comp_Rep_v2
\caption{The rescaled mass $M$ and radius $R_{99}$ for boson stars. The black lines denote the exact numerical solutions, whereas the ans\"atze are color-coded in the legend. The left panel is the attractive interaction case, where the maximum mass is calculable in a nonrelativistic method; the right panel is the repulsive interaction case.}
\label{fig:MassRadius}
\end{figure*}
It is interesting that the result for the double exponential ansatz almost exactly approximate the sech and linear+exponential ans\"atze for the choices $a=2$ and $a=3\pi/8$, respectively. We will discuss this in more detail in the following section.
\subsection{Fitting the Double Exponential} \label{sec:DE}
Every variational calculation, regardless of the ansatz, can be improved by the addition of an extra parameter. The double exponential ansatz proposed in this paper has an extra variational parameter that can be tuned to fit the numerical solution and is computationally efficient. The extra variational parameter allows one to choose how to fit the ansatz to the numerical solution. For example, one can choose to fit the ansatz and numerical wavefunctions, or to fit various expectation values. In the comparisons above, we found that the choice $a = 3\pi/8$ reproduced the exact value of the maximum mass. This choice also resulted in an ansatz, denoted $\psi_{DE_{3\pi/8}}$, that was in very good agreement with $\psi_{LE}$. We also considered the variational parameter $a = 2$, which resulted in an ansatz $\psi_{DE_2}$ that was in good agreement with $\psi_S$. By varying the parameter $a$, we found we could optimize agreement to the numerical solutions on both branches of attractive solutions, as well as for repulsive interactions.
As noted previously, $\psi_{DE_a}$ can be tuned to fit $\psi_{LE}$ and $\psi_{S}$. We can explain this by looking at the double exponential in different limits of $a$. First of all, it is clear that in the limit $a\gg 1$, we have $\psi_{DE_a} \to \psi_E$, the ordinary exponential ansatz. On the other hand, in the limit $a\to 1$,
\begin{align}
\psi_{DE_a} &= f(a) \left(1 - \frac{1}{a}e^{-(a-1)r/\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda}\right)e^{-r/\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda} \nonumber \\
&\approx f(a) \left(1 - \frac{1}{a}\left[1 - (a-1)\frac{r}{\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda}\right]\right)e^{-r/\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda} \nonumber \\
&= f(a) \left(1 - \frac{1}{a}\right) \left[1 + \frac{r}{\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda}\right] e^{-r/\sigma} \def\o{\omega} \def\t{\tau} \def\l{\lambda},
\end{align}
where $f(a)$ is the original prefactor in eq. (\ref{eq:ansatze}). Because the prefactor in any ansatz is finite and equal to $\psi(0)$ by definition, we find that $\psi_{DE_a} \to \psi_{LE}$ in the limit $a\to1$. In this sense, the double exponential ansatz interpolates between $\psi_E$ and $\psi_{LE}$ as $a$ varies between $1$ and $\gg1$.
Slightly more perplexing is the sech ansatz, which seems to be reproduced approximately when $a=2$. The reason for that is that the second derivatives at $r=0$ coincide for those ans\"atze. The behavior near the origin is the most important for all integrals over the wave functions, because the largest contributions to all integrals come from that region. This explains this coincidence and further illustrates the versatility of $\psi_{DE_a}$.
The double exponential can be easily further generalized, as it is straightforward to add additional exponential functions with further fitting parameters $a_1,a_2,a_3,...$. Because computations on exponential functions are relatively fast, these additional functions do not increase the computation time very much (also see next section). Of course, one could generalize any of the other ans\"atze by introducing a fitting parameter, but the increase in computation time is more problematic for these other ans\"atze.
\begin{figure*} [ht]
\includegraphics[scale=.50]{PsiComp_LOG_n=1_v2
~ \quad
\includegraphics[scale=.50]{PsiComp_LOG_n=10_v2
\includegraphics[scale=.50]{DeltaPsi_v2}
~ \quad
\includegraphics[scale=.48]{DeltaR_v2
\caption{Quantitative comparisons for attractive interactions, on the stable branch of solutions. In the upper two panels, we show the rescaled wavefunctions for two choices of particle number $N = 0.12N_c$ (left) and $N=0.98N_c$ (right). In the lower two panels, we show the corresponding deviations $\Delta_\psi$ (left) and $\Delta_r$ (right) of the ans\"atze from the exact solution, as defined in eqs. (\ref{eq:Deltar}) and (\ref{eq:Deltapsi}).}
\label{fig:StableComp}
\end{figure*}
\subsection{Methods for Quantitative Comparisons} \label{sec:methods}
In this section, we report on the direct comparison of the exact numerical solutions of the previous section with the ans\"atze we presented in Section \ref{sec:Ansatze}. One criterion on which one can compare different ans\"atze is computational efficiency; in our analysis we have seen, for example, that the sech ansatz is significantly more difficult to employ than a function with exponential dependence. This is an important consideration because the purpose of using an ansatz in the first place is to simplify the calculation. The sech ansatz, for example, is orders of magnitude slower than the others (and, as we will show below, it does not pay off sufficiently in numerical accuracy). The double exponential is relatively fast, regardless of the value of $a$.
\begin{table}[t]
\centering
\hspace{-15pt}
\begin{tabular}{| c || c | c | c | c | c |}
\hline
& Gaussian
& Exp.
& Lin $+$ Exp
& Sech
& Double Exp. \\
\hline \hline
Time [sec] & $2.28$
& $2.48$
& $4.67$
& $358.2$
& $\sim 11$
\\ \hline
$\displaystyle{\frac{\text{Time}}{\text{Time}_{G}}}$ & $1$
& $1.09$
& $2.05$
& $157.4$
& $\sim 4.8$
\\ \hline
\hline
\end{tabular}
\caption{The computation time for the parameters in Table \ref{tab1}, for each ansatz. The top row are the absolute times in seconds on a standard laptop; the lower row is the time normalized to the Gaussian case.}
\label{tab_times}
\end{table}
\begin{figure*} [ht]
\includegraphics[scale=.50]{PsiComp_Unstable_LOG_n=6_v2
~ \quad
\includegraphics[scale=.50]{PsiComp_Unstable_LOG_n=10_v2
\includegraphics[scale=.50]{DeltaPsi_Unstable_v2}
~ \quad
\includegraphics[scale=.48]{DeltaR_Unstable_v3
\caption{Quantitative comparisons for attractive interactions, on the unstable branch of solutions. In the upper two panels, we show the rescaled wavefunctions for two choices of particle number $N = 0.59N_c$ (left) and $N=0.98N_c$ (right). In the lower two panels, we show the corresponding deviations $\Delta_\psi$ (left) and $\Delta_r$ (right) of the ans\"atze from the exact solution, as defined in eqs. (\ref{eq:Deltar}) and (\ref{eq:Deltapsi}).}
\label{fig:UnstableComp}
\end{figure*}
Another criterion which is potentially more important is proximity to the exact solution. To make such comparisons simple to understand, we will rewrite the ans\"atze slightly so that the radial variable is the same as the numerical one, given in eq. (\ref{eq:numscale}); that is, we write
\begin{equation}
\psi_A(\tilde{r}) = \frac{m^{5/2}}{M_P\,|\l|}\sqrt{\frac{n}{\rho^3\,C_2}} F_A\left(\frac{\tilde{r}}{\rho}\right),
\end{equation}
where $A$ labels the ansatz being considered, as before. Then
\begin{equation}
\tilde{\psi}_A(\tilde{r}) \equiv \frac{M_P\,|\l|}{m^{5/2}}\psi_A(\tilde{r})
= \sqrt{\frac{n}{\rho^3\,C_2}} F_A\left(\frac{\tilde{r}}{\rho}\right)
\end{equation}
can be compared directly to the numerical result $\tilde{\psi}$ as a function of $\tilde{r}$.
A simple comparison of the shapes of the wavefunctions can be performed as follows: for a given $N$, compute $\vev{r}$, $\vev{r^2}$, and $R_{99}$ for the ansatz and compare it to the exact value computed numerically. These three radial variables should be a reasonable proxy for the shape of the wavefunction. More precisely, we will compute
\begin{equation} \label{eq:Deltar}
\Delta} \def\G{\Gamma} \def\O{\Omega_r(n) \equiv \sqrt{\left[\frac{\vev{r^2} - \vev{r^2}_{A}}{\vev{r^2}}\right]^2 +
\left[\frac{\vev{r} - \vev{r}_{A}}{\vev{r}}\right]^2 +
\left[\frac{R_{99} - R^{A}_{99}}{R_{99}}\right]^2}.
\end{equation}
A potentially more robust method of comparison is to evaluate the numerical solution and a given ansatz at the same value of $N$, and see how large the wavefunction deviations are over the whole range of $r$. To make this quantitative, we compute the difference integral
\begin{equation} \label{eq:Deltapsi}
\Delta} \def\G{\Gamma} \def\O{\Omega_{\psi}(n) \equiv \frac{\int d^3\tilde{r} \left[\tilde{\psi}(\tilde{r}) - \tilde{\psi}_A(\tilde{r})\right]^2}
{\int d^3\tilde{r}\,\tilde{\psi}(\tilde{r})^2}.
\end{equation}
We will use both $\Delta} \def\G{\Gamma} \def\O{\Omega_{\psi}$ and $\Delta} \def\G{\Gamma} \def\O{\Omega_{r}$ to compare how well a given ansatz matches the exact result. In the following sections, we will show the result of these analyses for attractive interactions on both the stable and unstable branches, as well as for repulsive interactions.
\subsection{Attractive Interactions: Comparison for Stable Branch}
We begin with an analysis of the stable branch of solutions for attractive self-interactions. This branch is relevant for so-called dilute boson stars, including axion stars. In the top row of Figure \ref{fig:StableComp}, we show the exact and approximate wavefunctions, with the vertical axis on a log scale. The top row correpsonds to attractive interactions along the stable branch. The most commonly used ans\"atze are the exponential and Gaussian functions, which turn out to be the worst fits to the numerical solutions, and are particularly bad in the large-$r$ tail. The other ans\"atze do reasonably well both near the core and in the tail.
Near the maximum mass, both the Gaussian and the exponential ans\"atze approximate the exact solution exceptionally badly. This is in part clear from Figure \ref{fig:MassRadius}: because these functions do not well-approximate the exact value of the maximum mass, the radius will also be very different, resulting in larger overall deviations (we will quantify these statements in a more precise analysis below). The other ans\"atze provide a much better fit in this region of parameter space. At large $\tilde{r}$, the double exponential does slightly better than the linear+exponential function.
For a more quantitative comparison, we turn to the bottom row of Figure \ref{fig:StableComp}. At every $n$, the smallest $\Delta} \def\G{\Gamma} \def\O{\Omega_\psi$ deviations are found using the LE ansatz. In the limit $a\to1$, the same is found for the double exponential function, as discussed in an earlier section. In fact, we have scanned over values of the parameter $a$ and found that the best agreement to numerical results (i.e. smallest $\Delta} \def\G{\Gamma} \def\O{\Omega_\psi$ and $\Delta} \def\G{\Gamma} \def\O{\Omega_r$) is obtained in this limit. In this sense, the linear+exponential function provides the best fit to the data, and is also computationally efficient. It is also the most similar to the leading-order wavefunction of \cite{Kling1,Kling2}, whose wavefunction is explicitly calculated to have the correct asymptotic behavior as $r\to\infty$.
\begin{figure*} [ht]
\includegraphics[scale=.50]{PsiComp_Rep_LOG_n=p1_v2
~ \quad
\includegraphics[scale=.50]{PsiComp_Rep_LOG_n=100_v2
\includegraphics[scale=.50]{DeltaPsi_Rep_v2}
~ \quad
\includegraphics[scale=.48]{DeltaR_Rep_v2
\caption{Quantitative comparisons for repulsive interactions. In the upper two panels, we show the rescaled wavefunctions for two choices of particle number $N = 0.12N_*$ (left) and $N=40N_*$ (right). In the lower two panels, we show the corresponding deviations $\Delta_\psi$ (left) and $\Delta_r$ (right) of the ans\"atze from the exact solution, as defined in eqs. (\ref{eq:Deltar}) and (\ref{eq:Deltapsi}).}
\label{fig:RepComp}
\end{figure*}
\subsection{Attractive Interactions: Comparison for Unstable Branch}
We move now to an analysis of the unstable branch which is relevant for collapsing boson stars, an important topic both for ordinary QCD axion stars \cite{ELSW,ELSW2}, as well as possible galaxy core collapse induced by astrophysical interactions \cite{ELSWFDM}. In the top row of Figure \ref{fig:UnstableComp}, we show the exact and approximate wavefunctions, with the vertical axis on a log scale for attractive interactions along the unstable branch. It can be seen that for large $\tilde{r}$, $\psi_G$ is a poor fit for both small and large masses, while $\psi_S$ and $\psi_{DE_2}$ tend to be better fits than $\psi_{LE}$ and $\psi_{DE_{3\pi/8}}$. It is also interesting to note that $\psi_E$ does better than all ans\"atze for some range of $n$, though we do not have a satisfactory explanation for this behavior at this time. We do observe in the numerical solutions that the wavefunctions become increasingly ``squeezed" towards $r=0$ on the unstable branch, and while the wavefunctions remain cored, they more closely approach an exponential behavior closer to the origin.
In the bottom row of Figure \ref{fig:UnstableComp}, we show the $\Delta} \def\G{\Gamma} \def\O{\Omega_\psi$ and $\Delta} \def\G{\Gamma} \def\O{\Omega_r$ deviations for the unstable branch. For a range of masses, $\psi_E$ does better than all ans\"atze, while $\psi_S$ and $\psi_{DE_2}$ do better than $\psi_{LE}$ and $\psi_{DE_{3\pi/8}}$ for all masses except near the maximum mass. We conclude that over the most relevant parameter space, an exponential function is the ideal choice for analysis on the unstable branch, though close to $n_c$ it is preferable to use the double exponential with $a=2$. We do not recommend use of the sech function, only because of the computational inefficiency pointed out in Section \ref{sec:methods}.
\subsection{Repulsive Interactions: Comparison} \label{RepComp}
Finally, we move to an analysis of repulsive interactions which is relevant for axion-like particles; for recent model-building of dark matter scalar with repulsive interactions, see \cite{Fan}. For such self-interactions, a maximum mass arises due to relativistic effects, to which our nonrelativistic analysis is not sensitive. This maximum mass \cite{CSW}
\begin{equation}
M_{\text{max}}^{\text{rep}} = 0.22 \sqrt{\frac{\lambda}{4\pi}}\frac{M_P{}^3}{m^2}
\end{equation}
is much greater than the masses resulting from our numerical solutions, and so the comparisons made here are all for physical masses.
For repulsive interactions, as the mass of the condensate increases, it approaches the Thomas-Fermi limit at which its kinetic energy becomes negligible and its radius becomes independent of mass \cite{DGPS_TF}. From eqs. (\ref{rhodilute}) and (\ref{nCrit}), one can see that as the number of particles $n$ increases, the solution for the equilibrium radius does, indeed, approach a constant. This constant radius happens to be the critical radius for attractive self-interactions (eq. \ref{rhoCrit}), given that the ansatz considered is held constant. It is also interesting to note that $N=N_*$ can be seen as a scale at which the condensate begins to approach the Thomas-Fermi limit.
In the top row of Figure \ref{fig:RepComp}, we show the exact and approximate wavefunctions, with the vertical axis on a log scale. One can see that for $n \gg n_*$ all ans\"atze become increasingly worse fits in the tail. As is the case for attractive self-interactions, $\psi_{LE}$ and $\psi_{DE_{3\pi/8}}$ do better than $\psi_S$ and $\psi_{DE_2}$. It is interesting to note that for $n/n_* \sim 40$, the wavefunction of the numerical solution falls off more rapidly at large $\tilde{r}$ than $\psi_G$. This is in agreement with the results of B\"ohmer and Harko \cite{BH_TF} who found an exact solution to the equations of motion for the Thomas-Fermi limit to be of the form $\sqrt{\sin(\xi)/\xi}$. This exact solution has been included as a possible compact ans\"atze to be used for axion-like condensates with repulsive interactions and large numbers of particles (see Section \ref{compact_ansatze}). For $N\ll N_c$, the condensate approaches the noninteracting limit; this is why the top-left panel of Figure \ref{fig:RepComp} is extremely similar to the top-left panel of Figure \ref{fig:StableComp}.
In the bottom row of Figure \ref{fig:RepComp}, we show the $\Delta} \def\G{\Gamma} \def\O{\Omega_\psi$ and $\Delta} \def\G{\Gamma} \def\O{\Omega_r$ deviations for repulsive interactions from which one can see that $\psi_G$ does better than all other ans\"atze at large $n$. It is also interesting to note that the deviations have a considerable change in behavior near $n_*$ (most notably for $\psi_G$), which is presumably due to the condensate approaching the Thomas-Fermi limit. For repulsive interactions at large $n$, the Gaussian function is thus the appropriate choice.
\section{Relativistic Formulation} \label{sec:RB}
In the preceding sections, we have analyzed various approximations to the GPP formalism. However, even the exact GPP equations are themselves an approximation to the underlying relativistic theory. The theory at high energies is defined by the Klein-Gordon equation for a scalar field, which is coupled to a curved spacetime metric described by the Einstein equations. In this section, we describe precisely the limit of the Einstein+Klein-Gordon (EKG) equations in which one recovers the GPP equations. We will then apply the relativistic formulation to the case of repulsive interactions, to analyze the limitations of the TF approximation.
\subsection{Equivalence between GPP and EKG Equations}
A relativistic method of describing boson stars was pioneered by Ruffini and Bonazzola (RB) \cite{RB}. The crucial idea is to take the field ${\mathcal A}} \newcommand{\Ocal}{{\mathcal O}$ describing the boson star to be linear in ground-state creation and annihilation operators $a_0^{(\dagger)}$,
\begin{equation} \label{fieldRB}
{\mathcal A}} \newcommand{\Ocal}{{\mathcal O}(r,t) = R(r) \left[e^{-i\,\mu_0\,t}\,a_0 +
e^{i\,\mu_0\,t}\,a_0^\dagger\right],
\end{equation}
and use these operators to build $N$-particle states defined by
\begin{equation}
\ket{N} = \frac{(a_0^\dagger)^N}{\sqrt{N!}} \ket{0}.
\end{equation}
Here, $\mu_0$ is the energy eigenvalue of the \emph{relativistic} equations; it is related to $\mu$ in the GPP formulation by $\mu = -(m-\mu_0)$.
The proper normalization of the axion wavefunction $R(r)$ is \cite{RB}
\begin{equation}
2\,\frac{\mu_0}{m}\,\int d^3r\,R(r)^2\,\sqrt{\frac{A(r)}{B(r)}} = 1,
\end{equation}
which effectively defines the particle number $N$.
The gravitational metric describing the curvature of spacetime in the presence of the boson star is
\begin{equation}
ds^2 = -B(r)\,dt^2 + A(r)\,dr^2 + d\O^2,
\end{equation}
which is spherically symmetric, as is the ground state wavefunction $R(r)$ in eq. (\ref{fieldRB}). Then the $rr$ and $tt$ components of the Einstein equations (taken as expectation values), along with the Klein Gordon equation, constitute a complete system of equations for $R(r)$, $A(r)$, and $B(r)$:
\begin{align} \label{NEKGN}
\bra{N}G_\m{}^\n\ket{N} &= 8\,\pi\,G\,\bra{N}T_\m{}^\n\ket{N} \nonumber \\
KG[{\mathcal A}} \newcommand{\Ocal}{{\mathcal O}] = \Box\,{\mathcal A}} \newcommand{\Ocal}{{\mathcal O} - V'({\mathcal A}} \newcommand{\Ocal}{{\mathcal O}) &= 0.
\end{align}
In the original RB paper, the self-interaction potential was trivial $V({\mathcal A}} \newcommand{\Ocal}{{\mathcal O})=m^2{\mathcal A}} \newcommand{\Ocal}{{\mathcal O}^2/2$, and only the $\bra{N}KG[{\mathcal A}} \newcommand{\Ocal}{{\mathcal O}]\ket{N-1}=0$ expecation value of the KG equation is nontrivial. In that case the ansatz of eq. (\ref{fieldRB}) is an exact solution. For the $\phi^4$ potential we have been considering, we have
\begin{equation}
V({\mathcal A}} \newcommand{\Ocal}{{\mathcal O}) = \frac{m^2}{2}{\mathcal A}} \newcommand{\Ocal}{{\mathcal O}^2 + \frac{\l}{4!}{\mathcal A}} \newcommand{\Ocal}{{\mathcal O}^4.
\end{equation}
Evaluating the expectation values as in \cite{RB}, we find\footnote{The third equation of (\ref{EKG1}) is the expectation value $\bra{N}KG[{\mathcal A}} \newcommand{\Ocal}{{\mathcal O}]\ket{N-1}=0$, but in the self-interacting theory higher-order expectation values $\bra{N}KG[{\mathcal A}} \newcommand{\Ocal}{{\mathcal O}]\ket{N-(2k+1)}=0$ with $k>0$ will not be satisfied, and so eq. (\ref{fieldRB}) is not an exact solution. However, as argued in \cite{GRB}, taking the leading order is a good approximation for all but the most strongly bound boson star configurations. In the context of the axion potential, we have presented a method of going beyond the RB ansatz to calculate relativistic corrections perturbatively in \cite{GRB}. For other work on relativistic corrections to scalar field theory, see \cite{GuthRelativistic,BraatenRelativistic,EMTWY}.}
\begin{align} \label{EKG1}
&\frac{A'}{A^2\,r} + \frac{A-1}{A\,r^2}
= \frac{8\,\pi}{M_P{}^2}\Big[
\frac{\mu_0{}^2\,N\,R^2}{B} + \frac{N\,R'^2}{A} \nonumber \\
&\hspace{4cm} + N\,R^2 + \frac{N^2\,\l}{4\,m^2}R^4\Big],
\nonumber \\
&\frac{B'}{A\,B\,r} - \frac{A-1}{A\,r^2}
= \frac{8\,\pi}{M_P{}^2}\Big[
\frac{\mu_0{}^2\,N\,R^2}{B} + \frac{N\,R'^2}{A} \nonumber \\
&\hspace{4cm} - N\,R^2 - \frac{N^2\,\l}{4\,m^2}R^4\Big],
\nonumber \\
&R'' + \left(\frac{2}{r} + \frac{B'}{2\,B}
- \frac{A'}{2\,A}\right)R' \nonumber \\
&\hspace{2cm} + A\left[\frac{\mu_0{}^2}{B}R
-m^2\,R - \frac{N\,\l}{2}R^3\right]=0.
\end{align}
One can simplify the expression by defining \cite{ESVW}
\begin{equation}
X(y) = \frac{2\,\sqrt{|\l|\,N}}{m}R(r), \qquad y = m\,r,
\end{equation}
in which case the EKG equations simplify to
\begin{align} \label{EKG2}
&\frac{A'}{A^2\,y} + \frac{A-1}{A\,y^2}
= \frac{2\,\pi\,m^2}{M_P{}^2\,|\l|}\Big[
\left(\frac{\mu_0}{m}\right)^2 \frac{X^2}{B} + \frac{X'^2}{A} \nonumber \\
&\hspace{4cm}
+ X^2 + \frac{\text{sgn}(\l)}{16}X^4\Big],
\nonumber \\
&\frac{B'}{A\,B\,y} - \frac{A-1}{A\,y^2}
= \frac{2\,\pi\,m^2}{M_P{}^2\,|\l|}\Big[
\left(\frac{\mu_0}{m}\right)^2\frac{X^2}{B} + \frac{X'^2}{A} \nonumber \\
&\hspace{4cm}
- X^2 - \frac{\text{sgn}(\l)}{16}X^4\Big],
\nonumber \\
&X'' + \left(\frac{2}{y} + \frac{B'}{2\,B}
- \frac{A'}{2\,A}\right)X' \nonumber \\
&\hspace{2cm}
+ \frac{A}{2}\left[\left(\frac{\mu_0}{m}\right)^2\frac{X}{B}
- X - \frac{\text{sgn}(\l)}{4}X^3\right]=0.
\end{align}
Note that primes in eq. (\ref{EKG2}) now indicate derivatives with respect to $y$ (rather than $r$).
In \cite{ESVW}, our original investigation of axion stars, we identified two small parameters and expanded the EKG equations in both. The analogue of these small parameters for a general theory with $\l\,\phi^4$ interactions is
\begin{equation}
\delta = \frac{8\,\pi\,m^2}{M_P{}^2\,|\l|} \ll 1 \qquad \text{and} \qquad \Delta} \def\G{\Gamma} \def\O{\Omega = \sqrt{1 - \frac{\mu_0{}^2}{m^2}} \ll 1.
\end{equation}
The requirement $\delta\ll1$ corresponds to the weak gravity limit, as the $\delta\to0$ limit recovers the Einstein equations for the vacuum. It is valid only when $|\l| \gg 8\,\pi\,m^2/M_P{}^2$. As an example, axions have $\l = -m^2/f^2$ and so this condition is equivalent to $8\,\pi\,f^2 \ll M_P{}^2$, a condition that is easily satisfied in nearly all applications. In particular, for QCD axions, $\delta\approx10^{-14}$, and even theories of axionic ``fuzzy dark matter'', which have much larger $f$, still $\delta\ll1$ is satisfied easily. The other parameter $\Delta} \def\G{\Gamma} \def\O{\Omega$ is small precisely when the axion star is weakly bound, i.e. when the eigenenergy $\mu_0$ is of the same order as the particle mass $m$.
In applications with weak gravity, we can expand the metric components as
\begin{equation}
A(r) = 1 + \delta\,a(x), \qquad B(r) = 1 + \delta\,b(x).
\end{equation}
We also rescale the axion wavefunction $X(y)$ using
\begin{equation}
Y(x) = \frac{1}{\Delta} \def\G{\Gamma} \def\O{\Omega}\,X(y),
\end{equation}
with $x = y\,\Delta} \def\G{\Gamma} \def\O{\Omega$ (which is to say, the wavefunction and coordinate scale with $\Delta} \def\G{\Gamma} \def\O{\Omega$ as their scaling dimension). The resulting equations of motion, to leading order in $\delta$ and $\Delta} \def\G{\Gamma} \def\O{\Omega$, take the form
\begin{align} \label{EKG}
a'(x) &= \frac{x}{2}\,Y(x)^2 - \frac{a(x)}{x}, \nonumber \\
b'(x) &= \frac{a(x)}{x}, \nonumber \\
Y''(x)&= - \frac{2}{x}Y'(x)-\frac{1}{8}\, Y(x)^3
+ [1 + \kappa\,b(x)]\,Y(x).
\end{align}
The constant $\k\equiv \delta/\Delta} \def\G{\Gamma} \def\O{\Omega^2$ controls the effective coupling to gravity, and is finite even though $\delta,\Delta} \def\G{\Gamma} \def\O{\Omega\ll1$. We have integrated this set of equations previously in \cite{ESVW} to find the spectrum of weakly bound axion stars.
Finally, we can integrate the first two equations to eliminate $a(x)$ and obtain a Poisson-like equation for $b(x)$:
\begin{equation}
\nabla_x^2 b(x) = \frac{1}{2}\,Y(x)^2.
\end{equation}
This implies that $b(x)$ is proportional to the Newtonian gravitational potential $V_g$. The solution of this equation is
\begin{align}
b(x) &= -\frac{1}{8\pi}\left[\int d^3 x'
\frac{Y(x')^2}{|\vec{x}-\vec{x'}|}\right].
\end{align}
Using this expression in the third equation in (\ref{EKG}), we arrive at a self-contained expression for the rescaled wavefunction $Y(x)$:
\begin{equation} \label{KGFinal}
0 = \left[1 - \nabla_x^2 - \frac{1}{8}\,Y(x)^2
-\frac{\kappa}{8\pi} \left( \int d^3 x'
\frac{Y(x')^2}{|\vec{x}-\vec{x'}|}\right)\right] Y(x).
\end{equation}
Because this equation is true only to leading order in $\Delta} \def\G{\Gamma} \def\O{\Omega$, we refer to this equation as the \emph{infrared}, or low energy, limit of the Klein-Gordon equation for the axion.
We may manipulate the GPP equations (\ref{Poisson}) and (\ref{GPtime2}) into a very similar form in a straightforward way. First, rewriting the gravitational potential $V_g$ in integral form, we find a single integro-differential equation of motion
\begin{equation} \label{GP1}
i\,\dot{\psi} = -\frac{1}{2m}\nabla^2 \psi
- \frac{\l}{8\,m^2}|\psi|^2\psi
- G\,m^2\,\left[\int d^3r'\,\frac{|\psi(r')|^2}{|\vec{r}-\vec{r'}|}
\right]\,\psi.
\end{equation}
In the nonrelativistic limit, the wavefunction has a time dependence which is approximately harmonic, such that $i\dot{\psi} = \mu\,\psi = -(m-\mu_0)\psi \approx -m\,\Delta} \def\G{\Gamma} \def\O{\Omega^2\,\psi / 2$. Further, we rescale to the dimensionless wavefunction
\begin{equation}
\phi(r) = \sqrt{\frac{2\,|\l|}{m^3}}\frac{1}{\Delta} \def\G{\Gamma} \def\O{\Omega} \psi
\end{equation}
and dimensionless coordinate $x = m\,\Delta} \def\G{\Gamma} \def\O{\Omega\,r$ as before. Then eq. (\ref{GP1}) takes the form
\begin{align} \label{GPFinal}
0 &= m^2\,\Delta} \def\G{\Gamma} \def\O{\Omega^2\,\phi - m^2\,\Delta} \def\G{\Gamma} \def\O{\Omega^2\,\nabla_x^2\phi
- \frac{m^2\,\Delta^2}{8}\phi^3 \nonumber \\
& \hspace{4cm} - \frac{G\,m^4}{|\l|}\left[\int d^3x'\,
\frac{\phi(x')^2}{|\vec{x}-\vec{x'}|}\right]\phi \nonumber \\
&= m^2\,\Delta} \def\G{\Gamma} \def\O{\Omega^2\left[1 - \nabla_x^2 - \frac{1}{8}\phi^2
- \frac{G\,m^2}{|\l|\,\Delta} \def\G{\Gamma} \def\O{\Omega^2}\left(\int d^3x'\,
\frac{\phi(x')^2}{|\vec{x}-\vec{x'}|}\right)\right]\phi.
\end{align}
Finally, we recognize the prefactor on the gravitational term as
\begin{equation}
\frac{G\,m^2}{|\l|\,\Delta} \def\G{\Gamma} \def\O{\Omega^2} = \frac{\delta}{8\pi\,\Delta} \def\G{\Gamma} \def\O{\Omega^2} \equiv \frac{\kappa}{8\pi}.
\end{equation}
Thus, eq. (\ref{GPFinal}) is \emph{exactly equivalent} to eq. (\ref{KGFinal}) with the identification
\[
Y(x) \leftrightarrow \phi(x).
\]
Thus the GP formalism is equivalent to the RB formalism used in the infrared limit. We summarize the identifications between the two paradigms in Table \ref{tabEquiv}.
\begin{table*}
\begin{center}
\begin{tabular}{| c | c | c |}
\hline
Name & GPP & Leading-Order RB \\
\hline
Equation(s) of Motion & Equation (\ref{GP1}) & Equations (\ref{EKG}) \\
Wavefunction & $\displaystyle{\psi(r)}$ &
$\displaystyle{\sqrt{\frac{m}{2\,|\l|}}\,m\,\Delta} \def\G{\Gamma} \def\O{\Omega\,Y(x)}$ \\
Gravitational Potential & $\displaystyle{V_g(r)}$&
$\displaystyle{\frac{4\,\pi\,G\,m^3}{|\l|}\,b(x)}$ \\
Normalization & $\displaystyle{\int |\psi(r)|^2\,d^3r = N}$ &
$\displaystyle{\frac{\Delta} \def\G{\Gamma} \def\O{\Omega^2}{2|\l|}\int Y(x)^2\,d^3x = N}$\\
\hline
\end{tabular}
\caption{Equivalence relations between the Gross-Pit\"aevskii+Poisson equations and Ruffini-Bonazzola equations governing axion stars in the low-energy limit. The coordinates are related by $x = m\,\Delta} \def\G{\Gamma} \def\O{\Omega\,r$.}
\label{tabEquiv}
\end{center}
\end{table*}
Because these formalisms are precisely equivalent, one can work with whichever is more convenient for the application at hand. For example, in analyzing axion star decay, it is better to use the RB approach, as the transition matrix elements are more directly calculable \cite{ESW}. It is also more straightforward to generalize if higher-order relativistic corrections are needed \cite{GRB}. For the analysis of collapsing axion stars, we instead found it more convenient to use the GP formalism \cite{ELSW}.
\subsection{The Thomas-Fermi Approximation}
We will use the RB formulation to critically analyze the Thomas-Fermi (TF) approximation, which is commonly used in studies of repulsively interacting boson stars. At leading order, the equations of motion are
\begin{align} \label{eom}
\nabla_y^2 b(y) &= \frac{1}{2}\,X(y)^2 \nonumber \\
\nabla^2_y X(y) &- \Delta ^2\,X(y)-\delta \, b(y) \,X(y) - \frac{1}{8}\,X(y)^3=0,
\end{align}
where we have chosen to work with the wavefunction $X(y)$ rather than $Y(x)$ so that the appearance of the small parameters $\delta,\Delta} \def\G{\Gamma} \def\O{\Omega$ is manifest. In the TF approximation, one neglects the kinetic term $\nabla^2_y X(y)$ compared to the other terms. The consistency of that assumption must then be checked after the solution for $b(y)$ and $X(y)$ have been found. Suppose the scale for $r$ is $R$, the radius of the boson star. Then the kinetic term is
\begin{equation}
\nabla^2_y X(y) = \Ocal\left(\frac{X(y)}{m^2\,R^2}\right).
\end{equation}
Consequently, $\nabla_y^2 X$ is negligible compared to $\Delta^2\,X$ if
\begin{equation}
R\gg \frac{1}{\Delta\,m}.
\end{equation}
If the TF approximation is valid, then the Klein-Gordon equation simplifies to
\begin{align}\label{eombZ}
0 &= - \Delta ^2\,X(y)-\delta \, b(y) \,X(y) - \frac{1}{8}\,X(y)^3,
\end{align}
which can be directly solved for $X(y)$ to obtain
\begin{equation} \label{Xeq}
X(y)^2 = -8\,\left(\delta\,b(y) + \Delta} \def\G{\Gamma} \def\O{\Omega^2\right).
\end{equation}
Substituting this back into the Poisson equation gives
\begin{equation}
\nabla_y^2 b(y) = -4\,\left(\delta\,b(y) + \Delta} \def\G{\Gamma} \def\O{\Omega^2\right).
\end{equation}
The solution of this equation which is regular at the origin is
\begin{equation} \label{bsol}
b(y) = -\frac{\Delta} \def\G{\Gamma} \def\O{\Omega^2}{\delta} - \frac{c}{y}\,\sin\left(p_0\,y\right),
\end{equation}
where $p_0 = 2\sqrt{\delta}$ and $c$ is a dimensionless constant which will be determined below. Finally, we can calculate the wavefunction in the TF limit by substituting eq. (\ref{bsol}) into eq. (\ref{Xeq}):
\begin{equation} \label{XTFsol}
X(y) = \sqrt{\frac{8\,\delta\,c}{y}\,\sin\left(p_0\,y\right)}.
\end{equation}
We can now observe directly that the constant $p_0 = \pi / m\,R_{TF}$ defines the radius $R_{TF}$ of the condensate in the TF limit, given by
\begin{equation}
R_{TF} = \frac{\pi}{m\,p_0} = \frac{\pi}{2m\sqrt{\delta}} = \pi \sqrt{\frac{\l}{32\pi}}\frac{M_P}{m^2}
\end{equation}
and matches the standard result \cite{BH_TF}, previously derived by considering a polytropic equation of state. Note that the radius of the condensate in the TF limit does not depend on the particle number $N$; this behavior of the cutoff is characteristic of the TF approximation. The coefficient $c$ can be determined by normalization of the wavefunction (c.f. Table \ref{tabEquiv})
\begin{equation}
N = \frac{1}{m} \int d^3r\,T_{00} \approx \frac{1}{2\,\l}\int d^3y \,X(y)^2 = \frac{16\pi^2\,c\,\delta}{p_0{}^2\l} = \frac{4\pi^2\,c}{\l},
\end{equation}
which implies $c = N\,\l / (4\pi^2)$. Thus, we can write down the final expression for the rescaled wavefunction
\begin{equation}
X(y) = \frac{2\,\sqrt{\delta^{3/2}\,N\,\l}}{\pi} \sqrt{\frac{\sin(p_0\,y)}{p_0\,y}}.
\end{equation}
Finally, we consider the breakdown of the TF approximation. To check the self-consistency of omitting the kinetic term from the equation of motion, (\ref{eom}), we must substitute the solutions of eqs. (\ref{XTFsol}) and (\ref{bsol}) back into (\ref{eom}), to see whether the kinetic term is much smaller than the rest of the terms. Substituting those solutions into the kinetic term gives a complicated expression, which simplifies considerably if we evaluate it just at the boundaries of the physical range for $r$, at $r=0$ and $r=R_{TF}=\pi\,/\,(m\,p_0).$ In both limits we obtain, up to numerical constants,
\begin{equation} \label{KinTF}
\nabla_y^2 X(y) \sim \delta\,\sqrt{\delta^{3/2}\,N\,\l}.
\end{equation}
Similarly, evaluating the interaction term at the the and of the range of $r$ we obtain
\begin{equation} \label{IntTF}
X(y)^3 \sim (\delta^{3/2}\,N\,\l)^{3/2}.
\end{equation}
The validity of the Thomas-Fermi approximation requires that the ratio of eq. (\ref{KinTF}) and (\ref{IntTF}) be small, i.e.
\begin{equation}\label{constraint2}
\frac{\delta}{\delta^{3/2}\,N\,\l} = \frac{1}{\sqrt{\delta}\,N\,\l} \ll 1.
\end{equation}
This implies
\begin{equation}
N \gg N_{\rm min,TF} \sim \frac{1}{\sqrt{\delta}\,\l} \sim \frac{M_P}{m\,\sqrt{\l}},
\end{equation}
and comparing to eq. (\ref{scaling}) we see that this condition simply implies $n\gg 1$.
A further constraint on the Thomas-Fermi solution is the relationship between the Schwarzschild radius, $R_{\rm SCH}$ and $R$. Unless $R>R_{\rm SCH}$ the boson star collapses to a black hole. This constraint leads to the inequality
\begin{equation}
R_{TF} = \pi\,\sqrt{\frac{\l}{32\pi}} \frac{M_P}{m^2} \gg R_{\rm SCH}=2\,N\,\frac{m}{M_P{}^2}
\end{equation}
In other words, there is an upper limit for $N$,
\begin{equation}
N \ll N_{\rm max,TF} = \sqrt{\frac{\pi\,\l}{128}} \frac{M_P{}^3}{m^3}.
\end{equation}
Note that this differs from the exact result of Colpi et al. \cite{CSW} only by a small numerical factor.
Of course, if $N\ll N_{\rm max}$ is not satisfied, then our calculations, based on the Newtonian approximation to gravity, would not be acceptable. Therefore we obtain upper and lower (\ref{constraint2}) bounds for $N$ in the Thomas-Fermi approximation:
\begin{equation}\label{limits}
\frac{M_P}{m\sqrt{\lambda}} \ll N_{\rm TF} \ll \sqrt{\frac{\pi\,\l}{128}} \frac{M_P{}^3}{m^3},
\end{equation}
providing a wide range for the applicability of the approximation, provided
\begin{equation}\label{constraint3}
\lambda\gg \frac {m^2}{M_P^2}.
\end{equation}
The TF approximation has been generalized to describe rotating \cite{Shiraishi1} or charged \cite{Shiraishi2} boson stars, as well as boson stars comprised of $N$ scalar fields \cite{Shiraishi3}.
Note that the behavior of the TF wavefunction near $R_{TF}$ is not physical at finite $N$. One could in principle calculate corrections to the TF equations of motion which take this into account, which should give rise to the standard exponentially falling wavefunction at large $r$. This would have consequences in certain applications where the tail behavior is important, including calculations of the classical decay rate. An analysis of this type is beyond the scope of the present paper.
\section{Conclusions} \label{sec:Conclusions}
In this analysis we have analyzed a number of approximate methods for describing boson stars. We focused on the Gross-Pit\"aevskii+Poisson (GPP) equations, which have a broad range of applicability for weakly-bound, nonrelativistic boson stars.
Using a time-independent variational formalism, we compared various ans\"atze which describe gravitationally bound BECs with self-interactions. These ans\"atze allow the GPP system to be solved analytically, alleviating the obligation to cumbersome numerical solutions. Moreover, numerical solutions exist primarily for stationary BEC configurations, and are much more difficult to use in dynamical applications; ans\"atze are powerful tools for solving dynamic BEC problems such as collapse, collisions, and expansion.
We have treated the numerical solution to the stationary GPP system as a benchmark for comparing different ans\"atze, prioritizing factors such as computational ease, fit of the wavefunction profile, and value of maximum mass (for attractive interactions). We found that a linear+exponential wavefunction is the best fit for attractive self-interactions along the stable branch, as well as for repulsive self-interactions at small $N$. For attractive self-interactions along the unstable branch a single exponential is the best fit for small $N$ while a sech wavefunction fits better for large $N$, though the latter is computationally inefficient. A Gaussian wavefunction, which is used often in the literature, is exceedingly poor across most of the parameter space, with the exception of repulsive interactions for large $N$.
We found that our proposed double-exponential ansatz is much more tunable compared to other ans\"atze in the literature. By choosing various values of the parameter $a$, one may optimize a given fit parameter over others, or near-perfectly replicate more computationally complex ans\"atze. In particular, on the stable branch for attractive interactions, the limit $a\to1$ gives rise to the linear+exponential ansatz, which we found to be in closest agreement to the exact case. On the unstable branch, one can take the $a\gg1$ limit to obtain the exponential ansatz, appropriate at large central densities, while the choice $a=2$ very nearly reproduces the sech ansatz, which is quite inefficient computationally. We also showed how to generalize the addition of free parameters in order to create more computationally efficient ans\"atze using the double exponential.
There remain many unstudied applications of the ans\"atze we have compiled here, and analytic solutions to the time-dependent GPP system will undoubtedly have myriad uses. As mentioned in the introduction, there are many open questions regarding oscillons formed by inflation and quintessence fields, and axion BECs remain the subject of active study. Along with a rigorous, multi-criterion comparison of ans\"atze from the relevant literature, we have presented arguments for the necessity of time-dependent ans\"atze, and have introduced a tunable, computationally simple ansatz.
Finally, we also highlighted the relevant differences between the relativistic Einstein+Klein-Gordon (EKG) equations, as derived using the Ruffini-Bonazzola formalism, and the nonrelativistic GPP system. In particular, we showed explicitly the relevant expansion parameters which characterize the nonrelativistic limit. We used this formulation to critically analyze the Thomas-Fermi approximation, which is a large-$N$ limit for repulsive self-interactions in which the kinetic energy is taken to be small.
\section*{Acknowledgements}
We thank N. Bar, F. Kling, M. Ma, B. Maddock, and C. Vaz, for discussions. The work of J.E. was supported by the Zuckerman STEM Leadership Program. J.E. also thanks the Galileo Galilei Institute for Theoretical Physics for the hospitality and the INFN for partial support during the completion of this work. M.L. thanks the Barry Goldwater Scholarship and Excellence in Education Foundation for scholarship support. M.L. and L.S. thank the Department
of Physics at the University of Cincinnati for financial support in the form of Violet Diller Fellowships.
|
\section{Introduction}
The efficacy of refrigerators at the nanoscale has been made more than obvious in the past half decade \cite{giu}. From being useful in schemes for removal of excess heat in nanosystems to achieving low temperatures or in designing quantum computers they have been one of the most productive areas of research\cite{tan}. {In these nanoscales, refrigerators which utilize quantum properties like wave nature of electrons are particularly useful. Since as systems size decrease, it will invariably lead to mesoscopic lengths where the wave nature of electrons is apparent\cite{heikkila}. In this context, we study a quantum refrigerator(QR) designed with a strained graphene layer, wherein the quantum property of wave nature of electrons is much more apparent than say in 2DEG's}. Even though, the thermoelectric figure of merit $ZT$ of graphene is very small around $0.01-0.1$\cite{yuri}, which is much smaller than some of the most efficient thermoelectric materials, e.g., Bi$_2$Te$_3$, see Refs.~\cite{kim,yuri}. The reason behind this small $ZT$ factor is its large thermal conductance and absence of any band gap. In some recent theoretical works\cite{tran,sadeghi,pohao,anno}, this factor $ZT$ has been predicted to rise to moderate value of around $2.5-3$. This improvement in the thermoelectric figure of merit is attributed to doping graphene with isotopes\cite{tran} or nanopores\cite{sadeghi,pohao} or disorder\cite{anno,tran} or by nano-patterning the graphene surface\cite{kim}. A large $ZT$ factor is required to generate high coefficient of performance (COP) in any refrigerator. A large COP means the refrigerator can use the electrical power to absorb heat energy from the cooler terminal more efficiently. Instead of doping as done in related works, we consider straining the graphene layer in order to generate this high COP. In a previous work of ours\cite{arjun}, we have shown the potential use of a strained graphene sample to operate as a quantum heat engine. In this manuscript, we concentrate only on the refrigeration aspect of a strained graphene monolayer.
Another important aspect of this proposed strained graphene quantum refrigerator is that it operates in the steady state transport regime. Quantum refrigerator's (QR's) which work in the cyclic transport regime are also a major avenue of research. Examples of cyclic quantum refrigeration in literature can be seen in Refs.~\cite{lin, feng, he}. Cyclic QR's are of two types- 1) reversible, 2) irreversible. Advantage of cyclic reversible quantum refrigeration over both steady state quantum refrigeration and cyclic irreversible quantum refrigeration is that cyclic reversible QR's, such as Carnot refrigerators and Otto refrigerators, are independent of the property of the working substance\cite{he}, i.e., the material characteristics, while all steady state QR's and cyclic irreversible QR's are dependent on the working substance. Cyclic QR's do have some disadvantages too. Cyclic reversible QR's like Carnot or Otto refrigerators are based on reversible processes, thus to complete one full cycle they take infinite time, thereby, reducing the practical application of these refrigerators. Cyclic irreversible QR's are dependent on irreversible processes which take much less time to complete one full cycle, however, the efficiency of these refrigerators is much reduced from the Carnot limit due to the dissipation within the system. On the other hand, examples of steady state QR's can be seen in Refs.~\cite{ya, wang, edward, koslof}. Steady state QR's absorb heat from the cooler terminal by moving the microscopic particles like electrons, phonons or photons, rather than moving any microscopic part of the system (like nano cyclic refrigerators). Since, steady state QR's do not depend on the movement of the microscopic body part of the system, they can be much smaller in size than the cyclic QR's. {In both steady state as well as cyclic QR's the efficiency can reach the Carnot limit but always in the limit of zero cooling power. This happens due to the fact that cooling power and efficiency depend on the electrical, thermal conductances and Peltier coefficient in such a way that if one tries to increase the cooling power to its maximum value then the efficiency reduces to a minimum and vice versa.}
The rest of the paper is arranged as follows. In section 2 we describe the theory needed to calculate the thermoelectric figure of merit $ZT$, Onsager coefficients, maximum COP and cooling power. The operating regime where our system works as a quantum refrigerator is also explained. Next in section 3 we describe our model of monolayer strained graphene sample and using the boundary conditions the transmission function for electrons between the reservoirs is derived. In section 4, we describe the results for our model and explained the reason behind getting large COP and cooling power observed in our system. Next in section 5 we discuss the experimental realization of our model. Finally, in section 6 we conclude our paper with a conclusion and a comparison with the results observed in other related proposals.
\section{Theory of the quantum refrigerator}
To design an efficient QR, apart from the fact that we are in length scales wherein the electron wave nature is apparent, we need a large Peltier coefficient along with a large Seebeck coefficient. Large Peltier and Seebeck coefficients are required to increase the $ZT$ factor and a large $ZT$ factor will engender a large COP. However, having large Seebeck and Peltier coefficients is a double edged sword, it reduces the cooling power of the system, i.e., the amount of heat energy which can be absorbed from the colder terminal. The cooling power only increases when the thermal and electrical conductances of the system increase with Seebeck and Peltier coefficients simultaneously decrease. To optimize these quantities in such a way that we get a large COP along with large cooling power is quite difficult, because all these parameters are inter-related. If we want to increase or decrease any one of these parameters, other parameters too are affected. So, it is one of the outstanding challenges in quantum thermoelectrics to design an efficient QR which would optimize these parameters effectively such that we get large COP and large cooling power. The difference between a quantum heat engine and QR is that in a quantum heat engine one always needs a small thermal conductance to get a large efficiency at maximum output power, while for quantum refrigeration one needs a large thermal conductance for large cooling power. Of course here we are talking about electronic contribution to the thermal conductance to be large, phonon contribution to the thermal conductance has to be small otherwise it will decrease both the COP and the cooling power. To calculate the COP and cooling powers and to design a QR, first we need to calculate the themoelectric properties of our strained graphene device, i.e., the Seebeck and Peltier coefficients alongwith the electrical and thermal conductances. In linear transport regime the electrical and heat currents are related to the electric and thermal biases via the Onsager coefficients, which are written as-\cite{ramshetti, benetti, sothmann}-
\begin{equation}\label{current}
\left(\begin{array}{c} J^e \\ J^{Q}\end{array}\right)=\left(\begin{array}{cc} L^{11} & L^{12}\\ L^{21} & L^{22} \end{array}\right) \left(\begin{array}{c} \mathcal{V}\\ \Delta \theta \end{array}\right),
\end{equation}
where $J^e$ and $J^Q$ define the electric and heat currents respectively, $L_{ij}$ with $i,j$ $\in$ $1,2$ denotes the Onsager coefficients, {$\mathcal{V}$ and $\Delta\theta$ are the potential bias and temperature bias applied to the system respectively}. The Seebeck coefficient is defined as the voltage difference generated across the system due to an unit temperature difference applied, while Peltier coefficient is defined as the ratio of the heat current transmitting through the junctions to the electrical current passing through that junction. They are given as follows-
\begin{equation}\label{Seeback}
S=-\frac{L^{12}}{L^{11}}, \quad\text{and}\qquad P=\frac{L^{21}}{L^{11}}.
\end{equation}
The Onsager co-efficient matrix written in Eq.~(\ref{current}) linking electric and heat currents to the temperature difference ($\Delta\theta$) and applied voltage bias ($\mathcal{V}$) thus can be rewritten as \cite{ramshetti,mazamutto}-
\begin{equation}\label{onsager}
\left(\begin{array}{cc} L^{11} & L^{12}\\ L^{21} & L^{22} \end{array}\right) =\left(\begin{array}{cc} \mathcal{L}^{0} & \mathcal{L}^{1}/e\theta\\ \mathcal{L}^{1}/e & \mathcal{L}^{2}/e^2\theta \end{array}\right),
\end{equation}
wherein,
\begin{eqnarray}\label{con}
\mathcal{L}^{\alpha}=G_0\int_{-\pi/2}^{\pi/2}\!\!\!d\phi \cos\phi \int_{-\infty}^{\infty}\!\!\!d\epsilon(-\frac{\partial f}{\partial \epsilon})\frac{|\epsilon|}{\hbar v_f}(\epsilon-\mu)^\alpha T(\epsilon,\phi),
\end{eqnarray}
here $G_0=(e^2/\hbar)(W/\pi^2)$, $\mathcal{L}^0=G$ defines the conductance of system with width of sample-$W$\cite{dolphas}, $\epsilon$-the electron energy, $f$-the Fermi-Dirac distribution function, $\phi$ defines the angle of incidence for electrons, $T(\epsilon,\phi)$ is the electronic transmission probability through strained graphene layer and $\mu$- the Fermi energy. Eq.~(\ref{con}) is for a sheet of monolayer graphene (2D system) with the transmission function($T$) defined between the two terminals, see also Fig.~1. Once we know the transmission probability-$T(\epsilon,\phi)$, we can calculate the Onsager coefficient's- $L^{ij}$'s as shown in Eq.~(\ref{current}). After calculating the Onsager coefficients, other quantities such as cooling power and COP can be calculated as follows. The cooling power \cite{benetti}, is defined as -
\begin{eqnarray}\label{power}
{J^Q}=(L^{21}\mathcal{V}+L^{22}\Delta \theta)
\end{eqnarray}
Our strained graphene mono-layer device is shown in Fig.~\ref{fig1}. It works as a quantum refrigerator(QR) only when a net heat current is flowing from the cooler to hotter terminal making the colder terminal still more cooler, i.e., when it flows against the applied temperature bias $\Delta \theta=\theta_1-\theta_2$ ({$\theta_i$ is the temperature at contact $i$}) and the net electrical current flows from higher potential bias to the lower bias, i.e., the external work is done on the system. Thus, if a temperature bias $\Delta \theta$ is applied at the left terminal ($\theta_1>\theta_2$) and potential bias $\mathcal{V}$ is applied at the right terminal ($V_2>V_1$), then both heat current and electrical current flow from right to left ($J^e<0, J^Q<0$ considering $+x$ direction as the positive direction) for this system to work as a QR. The efficiency of a QR, i.e., how good it is in converting a stream of charged particles into carrying heat energy is called coefficient of performance (COP)\cite{edward,koslof}. COP of a quantum refrigerator is defined as the ratio of heat current absorbed from the hot reservoir to electrical power-$\mathcal{P}$ applied on the system, such that-
\begin{eqnarray}
\eta^r=\frac{J^Q}{\mathcal{P}},
\end{eqnarray}
where $\mathcal{P}$, is the electrical work or power done on the system, is defined as-
\begin{eqnarray}
\mathcal{P}=J^e\mathcal{V}=(L^{11}\mathcal{V}+L^{12}\Delta \theta)\mathcal{V}.
\end{eqnarray}
The COP-$\eta^r$ is maximum when $\frac{\partial \eta^r}{\partial \mathcal{V}}=0$ and considering $J^Q<0$ and $\mathcal{P}<0$, we have-
\begin{eqnarray}
\mathcal{V}&=&\frac{L^{22}}{L^{21}}(-1-\sqrt{\frac{L^{11}L^{22}-L^{12}L^{21}}{L^{11}L^{22}}})\Delta \theta.
\end{eqnarray}
The COP becomes maximum when the above relation Eq.~(8) between the potential bias and thermal bias holds and thus the maximum COP is when-
\begin{eqnarray}
\quad\eta^r_{max}&=&\frac{\eta^r_c}{x}\frac{\sqrt{ZT+1}-1}{\sqrt{ZT+1}+1},\\
\text{with, }\quad J^Q(\eta^r_{max})&=&\sqrt{\frac{L^{22}(L^{11}L^{22}-L^{12}L^{21})}{L^{11}}}\Delta \theta,
\end{eqnarray}
wherein $\eta^r_c=\frac{T}{\Delta \theta}$ is defined as the Carnot efficiency of an ideal refrigerator and $J^Q(\eta^r_{max})$ is the cooling power when COP is maximum. The thermoelectric figure of merit $ZT$ is defined as-
\begin{figure}
\centering {\includegraphics[width=0.48\textwidth]{fig1-ref.pdf}}
\vskip -1.50in
\caption{Top: Monolayer graphene with uniaxial strain applied in the x direction. The middle portion is strained region while the two side portions are normal graphene regions. Voltages $V_1$ and $V_2$ are applied to the two terminals which are at temperatures $\theta_1$ (left side) and $\theta_2$ (right side) respectively. Bottom: An electron is incident on the interface between normal graphene and strained graphene with incident angle $\phi$ and refracted to strained region with refraction angle $\theta$.}
\label{fig1}
\end{figure}
\begin{eqnarray}\label{ZT}
ZT=\frac{G S^{2}}{\kappa} \theta,
\end{eqnarray}
while thermal conductivity is-
\begin{eqnarray}\label{kappa}
\kappa=\frac{L^{11}L^{22}-L^{12}L^{21}}{L^{11}}.
\end{eqnarray}
In multi-terminal systems with broken time reversal(TR) symmetry, the upper bound on the refrigerator efficiency $\eta^r_{max}$ decreases from $\eta^r_c$ as the asymmetric parameter $\mathcal{\chi}=\theta L^{12}/L^{21}$ deviates from one \cite{brandner, bran}. In multi terminal systems with TR symmetry preserved and for all two terminal systems irrespective of whether TR symmetry is broken or not, the asymmetric parameter $x$ is unity, and upper bound on the corresponding maximum efficiency-$\eta^r_{max}$ equals $\eta^r_c$. This is the advantage of multi-terminal systems which preserve TR symmetry and for any two-terminal system, that they can work as highly efficient QR's with almost Carnot efficiency. However, for multi-terminal systems with broken TR symmetry working as QR, the upper bound is always less than $\eta^r_c$(the Carnot limit). To work as a QR, the relation between potential bias $\mathcal{V}$ and thermal bias $\Delta \theta$ has to be such that the electrical current $J^e<0$ and the heat current $J^Q<0$, i.e., both electrical and heat current flow from the cooler to the hotter region. Solving these two equations $J^e<0$ and $J^Q<0$ we get the operating regime for QR's as $\mathcal{V}<-(\kappa/(GP)-S)\Delta\theta$. This defines the parametric space in which our graphene device acts as a QR. In the next section, we give a detailed description of our model.
\section{Model}
\subsection{Hamiltonian}
Graphene is a semi-metal with zero band gap. It is a carbon allotrope with carbon atoms arranged in a single layer of honey comb lattice with inter penetrating triangular sublattices. A sublattice is a non-empty subset of a lattice, thus, a crystal can have several independent sublattices. Due to the presence of these two independent sublattices, graphene does have two non-equivalent set of local minima (Dirac point) in its momentum space at the edge of the first Brillouin zone. These two sets of local minima (Dirac points) are called as $K$ and $K'$ valleys. Dirac points are those points in the momentum space where conduction band meets with the valence band.
An uniaxial strain is introduced in our model of mono-layer graphene sheet lying in the $xy$ plane via stretching or compressing the region between $x=0$ and $x=L$ as shown in Fig.~1. The region to the left and right of this strained region is normal graphene. To design our model as a nano refrigerator we apply thermal bias $\Delta \theta$ at the left contact and a potential bias $\mathcal{V}$ at the right contact. At steady state, electric current $J^e$ and heat currents $J^Q$ flow between the reservoirs. In the strained region, electrons gets refracted away from the normal (perpendicular to the interface) in one valley (say $K$) and refracted towards the normal in the other valley ($K'$). The transmission of electrons gets shifted in the two valley as a function of incident angle in two opposite directions. However, the total transmission which is sum over all the incident angles remains same for both the valleys $K$ and $K'$. The Hamiltonian for the system for $K$ and $K'$ valley is-
\begin{equation}\label{ham}
\mathcal{H}_{K}=\hbar v_f\sigma(k-s'), \quad \mathcal{H}_{K'}=-\hbar v_f\sigma^*(k+s').
\end{equation}
Considering Landau gauge, one can replace the strain with pseudo magnetic vector potential $A=(0, \pm A_y)$, where `+' and `-' signs denote $K$ and $K'$ valleys respectively\cite{castro}. {The parameter $s'$ in Eq.~(13) is related to the strain $s$ by the relation $s=\hbar v_f s'$, and $s$ can be defined as the perturbation to the nearest neighbor hopping amplitudes, $\delta t$. Throughout our manuscript we have defined strain as $s$ in units of $meV$, if we divide the strain $s$ by the nearest neighbour hopping amplitude $t$ (which is a constant and equal to $2.7 meV$) then we will get the strain as a dimensionless quantity. Thus, strain-$s=\hbar v_f s'=eA_y[\Theta(x)-\Theta(x-L)]$, with $v_f$-the Fermi velocity, $\sigma=(\sigma_x, \sigma_y)$ are Pauli matrices operating on the sub-lattices A and B with $\sigma^*$ being complex conjugation, $\Theta$-the Heaviside step function and $k(=\{k_x, k_y\})$-the 2D wave vector.} From Hamiltonian, Eq.~(\ref{ham}), we get the wave equation for $K$ valley-
\begin{eqnarray}\label{wav}
\hbar v_f(-i\partial x-\partial y-i s)\psi_B=E\psi_A,\nonumber\\
\hbar v_f(-i\partial x+\partial y+i s)\psi_A=E\psi_B,
\end{eqnarray}
where $\psi_A$ and $\psi_B$ are the wave functions at $A$ and $B$ sublattices respectively. Using Eq.~(\ref{wav}), one can calculate the transmission probability for ballistic transport in a monolayer of strained graphene sample, which we show in the next subsection.
\subsection{Wave functions and boundary conditions}
In Fig.~1(bottom) an electron is incident from the left side of the interface between two regions (region I and region II) with energy $\epsilon$, then either it can reflect back to region I or it can transmit into region II depending on its energy and angle of incidence. We define three regions I, II and III as normal graphene ($x<0$), strained graphene ($0<x<L$) and normal graphene ($x>L$) respectively. The wave functions of electrons in the three regions for A and B sublattices in $K$ valley are as follows-\\
For $x<0$-
\begin{eqnarray}\label{a}
\left[\begin{array}{c}\psi_A^1(x,y)\\\psi_B^1(x,y)\end{array}\right]=\left[\begin{array}{c} (e^{ik_xx}+r e^{-ik_xx}) \\(e^{ik_xx+i\phi}-r e^{-ik_xx-i\phi})\end{array}\right]e^{ik_yy},
\end{eqnarray}
in strained graphene layer $0<x<L$-
\begin{eqnarray}\label{b}
\left[\begin{array}{c}\psi_A^2(x,y)\\\psi_B^2(x,y)\end{array}\right]=\left[\begin{array}{c} (a e^{iq_xx}+b e^{-iq_xx}) \\ (a e^{iq_xx+i\theta}-b e^{-iq_xx-i\theta}) \end{array}\right]e^{ik_yy},
\end{eqnarray}
and for $x>L$-
\begin{eqnarray}\label{c}
\left[\begin{array}{c} \psi_A^3(x,y)\\\psi_B^3(x,y)\end{array}\right]=\left[\begin{array}{c} t e^{ik_xx} \\t e^{ik_xx+i\phi} \end{array}\right]e^{ik_yy},
\end{eqnarray}
where $k_x=(\epsilon/\hbar v_f)\cos\phi$ and $k_y=(\epsilon/\hbar v_f)\sin\phi$ are the $x$ and $y$ components of momentum wave vector in normal graphene. In strained graphene $k_x$ is replaced with $q_x=\sqrt{(\epsilon/\hbar v_f)^2-(k_y-s)^2}=(\epsilon/\hbar v_f) \cos \gamma$ and $k_y-s=(\epsilon/\hbar v_f) \sin \gamma$, $\gamma$ being the refraction angle in the strained region as shown in Fig. 1(bottom) and also satisfies $\tan \theta=(k_y-s)/q_x$. Using wave functions- Eqs.~(\ref{a}-\ref{c}), and applying boundary conditions at $x=0$-
\begin{eqnarray} \label{ba}
\psi_B^2(x=0)=\psi_B^1(x=0), \quad
\psi_A^2(x=0)=\psi_A^1(x=0),
\end{eqnarray}
and at $x=L$-
\begin{eqnarray} \label{bb}
\psi_A^2(x=L)=\psi_A^3(x=L), \quad \psi_B^2(x=L)=\psi_B^3(x=L).
\end{eqnarray}
Solving Eqs.~(\ref{ba}-\ref{bb}) we derive the electronic transmission probability for $K$ valley-
\begin{eqnarray} \label{t}
T(\epsilon,\phi)=\frac{1}{\cos^2[q_xL]+\sin^2[q_xL](\frac{1-\sin[\gamma]\sin[\phi]}{\cos[\gamma]\cos[\phi]})^2}.
\end{eqnarray}
Similarly, one can derive the transmission function by solving the Hamiltonian for $K'$ valley as in Eq.~(\ref{ham}), using the same boundary conditions and substituting $\phi\rightarrow-\phi$, $s\rightarrow-s$. The total electronic transmission probability $T(\epsilon)$ is the sum of $K$ and $K'$ valley transmissions. It is apparent that although transmission $T(\epsilon,\phi)$ differs in $K$ and $K'$ valley\cite{can}, when integrated over $'\phi'$ this difference disappears. Thus, total transmission probability $T(\epsilon)$ is twice that of $K$ valley transmission.
\begin{figure}
\centering {\includegraphics[width=0.5\textwidth]{ref_fig2.pdf}}
\caption{(a) Conductance in units of $2e^2/h$ at temperature $\theta=30$K for various lengths of strained graphene layer with width $W=20$nm and strain $s=30$meV, (b) Conductance in units of $2e^2/h$ at temperature $\theta=30$K for various lengths of strained graphene layer with width $W=20$nm and Fermi energy $\mu=29.6$meV.}
\end{figure}
\section{Results and Discussion}
To design an efficient quantum refrigerator(QR), we not only need large Seebeck coefficient but also a large Peltier coefficient. A large Peltier coefficient means that a stream of electrons or holes carrying current from one terminal to another not only transport charge but also heat energy along with them. Thus, decreasing the temperature of the colder terminal. In any multi terminal system breaking time reversal symmetry implies an asymmetric Onsager matrix. However, for a two terminal system, Onsager matrix is always symmetric regardless of whether time reversal(TR) symmetry is broken or not\cite{datta}. This implies that the Seebeck and Peltier coefficient, for our system as shown in Fig.~1, are related by the relation $P=\theta S$ (due to the Onsager reciprocity relation $L_{21}=\theta L_{12}$). In Figs.~3 (a, b) we plot Peltier coefficient. The Seebeck coefficient can be inferred from these plots via $S=P/\theta$. Breaking TR symmetry leads to the upper bound of COP reducing from Carnot limit of efficiency, while in our system due to the conservation of TR symmetry the upper bound of COP can in principle reach the maximum Carnot limit. We first (in Fig.~2) plot the electrical conductances and see that it increases with increasing Fermi energy, see Fig.~2(a). We also see in Fig.~2(a), that increasing the length of the strained layer shifts the transport of incident electrons with low energy into the evanescent regime, and thus opens a conduction gap close to the Dirac point, while it remains almost unaffected at higher Fermi energies. Increasing strain can make the transmission of an electron at particular incident angle to be unity due to Klein effects\cite{kat}, however, the overall transmission summed over all the incident angle reduces with strain, see Fig.~2(b).
The Peltier coefficient is the ratio of off-diagonal Onsager matrix element $L_{21}$ to the electrical conductance. Increasing strain reduces the electrical conductance (see Fig.~2(b)) which, in turn, increases the Peltier coefficient (see Fig.~3(a)). However, increasing Fermi energy increases the electrical conductance, this in turn, leads to an increase in the Peltier coefficient near the Dirac point. Near the Dirac point, there is an imbalance between electronic contribution and hole contribution to the Peltier coefficient, which increases the Peltier coefficient initially. However, as one goes away from the Dirac point this effect vanishes and the Peltier coefficient decreases gradually.
\begin{figure}
\centering {\includegraphics[width=0.5\textwidth]{ref_fig3.pdf}}
\caption{(a) Peltier coefficient P in units of ($k_b\theta/e$) at temperature $\theta=30$K for various lengths of strained region with width $W=20$nm and Fermi energy $\mu=29.6$meV, (b) Peltier coefficient P in units of ($k_b\theta/e$) at temperature $\theta=30$K for various lengths of strained layer with width $W=20$nm and strain $s=30$meV.}
\end{figure}
This reduction of electrical conductance is almost independent of the length of the strained region. {In Refs.~\cite{mik, wu} it has been shown that strain can give rise to a resonant tunneling for some particular energy of incident electrons, thus acting like a quantum dot or quantum wire. However, this will not affect the performance of our device since we are taking an average over all the incidents angles. Resonant tunneling occurs only for some particular energy and incident angle of the electron. If we take an average over all the incident angles then the overall transmission probability always reduces with increasing strain, and thus to vanishing resonance effects.} In Figs.~3(a) and (b), we see that increasing either strain or the length of the strained region makes electrical conduction slowly to shift to the evanescent regime, then electrical conduction reduces enabling Peltier coefficient to increase since it is inversely proportional to conductance. This means electrons carry large amount of heat energy along with them and thus we see a corresponding increase in the Peltier coefficient. As strain is increased more, the electrical transmission $T(\epsilon, \phi)$ shifts completely to the evanescent regime, and Peltier coefficient increases much more. This is the reason why, COP of our model is huge for larger strains and large lengths of the strained region, see Fig.~4(a), while the cooling power decreases, see Fig.~4(b). The positive Peltier coefficient for Fermi energies greater than the Dirac point energy observed in Fig.~3(b) is attributed to the increased electronic contribution while for Fermi energies below the Dirac point energy the negative Peltier coefficient is attributed to increasing hole contribution. The peaks in Peltier coefficient observed near the Dirac point for both positive as well as negative energies is due to the imperfect cancellation of electron and hole contributions to the Peltier coefficient, while for energies much higher or much lower than the Dirac point the electronic and hole contribution to the Peltier coefficient almost cancel leading to a vanishing Peltier coefficient.
{
According to Eq.~(9) of our manuscript, coefficient of performance is proportional to the thermoelectric figure of merit $ZT$,which is the product of electrical conductivity $G$ and the square of Seebeck coefficient $S$, while inversely proportional to the thermal conductivity $\kappa$, i.e.,-
\begin{equation}
ZT=\frac{GS^2}{\kappa}\theta
\end{equation}
In the aforewritten equation, $\theta$ is the temperature. According to the Wiedemann-Franz law, for metals, electrical conductivity $G$ is proportional to the thermal conductivity $\kappa$ at finite temperature. Since, Graphene is a semi-metal, for its also $G$ is proportional to the $\kappa$. So, the ratio $G/\kappa$ is almost constant at finite temperature $T$. Thus, thermoelectric figure of merit $ZT$ solely depends on the square of the Seebeck coefficient. Since, graphene is a Dirac material, its Dirac cones shift in the opposite direction in two valleys upon applying strain creating a conduction gap close to the Dirac point, see Fig.~2(a). This conduction gap increases with increasing strain. Thus, the Seebeck (or Peltier) coefficient also increases with strain, since Seebeck (Peltier) coefficient is inversely proportional to the electrical conductivity $G$, see Eq.~(2). So, thermoelectric figure of merit $ZT$ as well as the coefficient of performance both increase with strain. However, this effect is not monotonous, if we keep on increasing strain then the ratio $G/\kappa$ is no more a constant and it reduces the $ZT$ as well as the coefficient of performance, see Fig.~4(a).}
In Fig.~5(a), we see that COP of our model is not temperature independent, the COP reduces as temperature decreases. On the other hand the cooling power is almost temperature independent, see Fig.~5(b). { The reason being that as we decrease temperature, the thermal conductance of the device remains almost unaffected by the strain, and thus the cooling power too, see Eqs.~(10, 12). However, the thermoelectric figure of merit of the device reduces with the temperature due to reduction in Peltier (Seebeck) coefficient and thus the COP reduces with temperature, see Eqs.~(9, 11).} In Fig.~6(a), we see that two peaks appear in the COP as function of Fermi energy. The first peak is present even at zero strain, that means strain is not a source of this peak. This peak appears close to the Dirac point and is due to the imbalance in the contribution of hole and electrons towards the Peltier coefficient as explained above. The second peak increases with increasing strain and vanishes for zero strain, this means strain is the sole reason behind the peak. The reason behind the appearance of the second peak is for the case of Peltier coefficient dominating over the electrical conductance, that's why the second peak increases with strain as Peltier coefficient increases. In Fig.~6(b), we see that the cooling power increases as a function of Fermi energy, but decreases with strain. So, we see that strain helps in increasing the COP but on the other hand it decreases the cooling power. Thus, we have to choose our parameters in such a way that we get the optimum values for COP and cooling power, i.e., the COP can still be large and the cooling power not so small. That is why we have defined a `Q' point for our quantum refrigerator where our device works at optimum values. `Q` point is the optimal operating range of our quantum refrigerator. Operating the refrigerator at the `Q' point, entails good COP as well as cooling power. The `Q' point of our refrigerator is at strain $s=30$ meV, with length of the strained region $L=60$ nm, temperature $\theta=30$ K, Fermi energy $\mu=29.6$ meV and width of graphene sample $W=20$ nm. For the `Q' point parameters, $COP=0.1 \eta_c^r$ and cooling power $J^Q=2 k_B^2\theta\Delta\theta/h$. However, our device exhibits a maximum $COP$ of $0.95\eta_c^r$ as seen in Fig.~7 but cooling power in this parametric regime is very small. Further, maximum cooling power possible in our system is $12k_B^2\Delta\theta/h$ as in Fig.~6(b) but COP at this parameter value is very small as shown in Fig.~6(a). The discussion so far hasn't addressed the width of the sample. From Eq.~(4) of our manuscript we can see that the Onsager coefficients linearly depend on sample width $W$. However, since Seebeck, Peltier coefficients and figure of merit $ZT$ are ratios of these Onsager coefficients, all of them won't depend on the width $W$ of the system. Thus, coefficient of performance will remain unaffected due to the changes in width, however, the cooling power will increase linearly with the width of the system. We have chosen $W=20$ nm, which is a reasonable value for experimental applications.
{ Finally, we address the reasons for neglecting the phonon contribution in the Figures presented in this section. The primary reason for this is because the thermal conductance in graphene is quite small (almost absent) at low temperatures $0-30 K$, see Figs.~2,3 in Ref.~\cite{mazamutto} and Fig.~5 of Ref.~\cite{xu}. Beyond $25-30 K$ range, the phonon contribution increases linearly with temperature, as shown in Refs.~\cite{mazamutto,zuev}. Thus, the phonon contribution to the thermal conductance can be neglected at the temperature range $20-30 K$ discussed in our work.}
\begin{figure}
\centering {\includegraphics[width=0.5\textwidth]{ref_fig4.pdf}}
\caption{(a) COP in units of $\eta_c^r$ at temperature $\theta=30$K for various lengths of strained layer with width $W=20$nm and Fermi energy $\mu=29.6$meV, (b) Cooling power in units of $(k_B^2\theta\Delta\theta)/h$ at $\theta=30 K$ for various lengths of the strained layer with width $W=20$nm and Fermi energy $\mu=29.6$meV.}
\end{figure}
\begin{figure}
\centering {\includegraphics[width=0.5\textwidth]{ref_fig5.pdf}}
\caption{(a) COP in units of $\eta_c^r$ at different temperatures with L=60 nm, width $W=20$ nm and Fermi energy $\mu=29.6$ meV, (b) Cooling power in units of $(k_B^2\theta\Delta\theta)/h$ for various temperatures with L=60 nm, width $W=20$ nm and Fermi energy $\mu=29.6$ meV.}
\end{figure}
\begin{figure}
\centering {\includegraphics[width=0.5\textwidth]{ref_fig6.pdf}}
\caption{(a) COP in units of $\eta_c^r$ at $\theta=30$K at different strains with $L=60$nm, width $W=20$nm, (b) Cooling power in units of $(k_B^2\theta\Delta\theta)/h$ at $\theta=30$K for various strains with $L=60$nm, width $W=20$nm.}
\end{figure}
\section{Experimental Realization}
Our proposal of a quantum refrigerator based on a strained monolayer graphene layer is experimentally realizable. Many theoretical, see Ref.~\cite{castro, nanoscale}, as well as many experimental works\cite{tmg, chhikara}, deal with uniaxial strain in monolayer graphene system. Thus, realizing strain in graphene system would not be much difficult. Also, the amount of strain used in our system is very small. One can apply a maximum $20\%$ strain ($540$ meV) without opening a band gap, while in our paper the optimum strain applied at `Q' point is $30$ meV ($1\%$ strain)\cite{liu,per-tight}. Further confining strain region only to the middle part has been attempted before too by stretching, compressing
or suspending the middle part of the graphene only without affecting the nonstrained region, see Ref.~\cite{balandin}. In Ref.~\cite{balandin}, only the middle part of the graphene sample is suspended across a wide trench in Si substrate, generating a finite strain but limited to the middle part only, see Box. 1(a) (on page 573 of Ref.~\cite{balandin}).
We have considered a sharp strain potential between the two normal regions for convenience only. One can also consider a small slope to the strain potential rather than a sharp strain potential. However, it has been shown that a small slope to the strain at the boundary between strain and normal regions also leads to the same amount of tunneling suppression for the electrons, see Refs.~\cite{castro, ade}. Recently, there have been other works, which show that strain can be reliably and easliy controlled in graphene, see Ref.~\cite{zhni}. Numerical values of all the other parameters used in our paper are also physically realizable and used in other works too, see Refs.~\cite{castro,dolphas}.
\section{Conclusion and Perspective on related works}
In this manuscript a strained monolayer graphene sheet has been designed to work as a quantum refrigerator(QR). The maximum coefficient of performance of our QR is around $0.95 \eta_c^r$ (see Fig.~7). This large COP seen in our paper occurs at a strain $s=300$ meV, which is $11\%$ strain, much less than the maximum 20\% strain ($540$ meV) and thus does not open any band gap within our system. The maximum cooling power possible with our QR is $12(k_B^2\Delta\theta)/h$ (see Fig.~6(b)). However, we note that the maximum values of coefficient of performance and of cooling power seen in our QR do not correspond to same set of parameters, the optimum values of coefficient of performance as well as cooling power are $0.1 \eta_c^r$ and $2(k_B^2\theta)/h$ and these occur at identical parametric values, i.e., the`Q' point. Steady state quantum refrigeration as seen in this manuscript in strained graphene is a nascent topic. Not many works\cite{ya,wang,giazotto} have addressed the topic of steady state quantum refrigeration. In the following we briefly address those works mentioning their advantages and disadvantages vis-a-vis our work.
\begin{figure}
\centering {\includegraphics[width=0.45\textwidth]{fig7_T_30_L_60_1.pdf}}
\caption{COP in units of $\eta_c^r$ at $\theta=30$K at different strains with $L=60$nm, width $W=20$nm.}
\end{figure}
Ref.~\cite{ya} discusses a three terminal quantum dot refrigerator, wherein the maximum COP and cooling power are $0.4\eta_c^r$ and $0.87(k_B^2\theta)/h$ respectively (see Fig.~9 of Ref.~\cite{ya}). Thus, we see while maximum COP delivered is larger than that seen at the `Q' point of our QR, the cooling power at that maximum COP is smaller. In Ref.~\cite{wang}, it has been shown that a magnon driven quantum dot refrigerator has COP $0.2\eta_c^r$ while again the cooling power seen $0.8(k_B^2\theta)/h$ is much smaller than that seen at the `Q' point of our system. Unfortunately, Refs.~\cite{ya, wang} do not discuss the `Q' point for their refrigerators. Further, Refs.~\cite{ya, wang} consist of a three terminal system, which by design has an advantage over a two terminal system since in a three terminal system heat and electric currents can flow between separate terminals, so one can have better control over these parameters. Similarly, we propose that in a quantum refrigerator based on three terminal strained graphene system the performance can be increased further, which can be the subject of another manuscript. {Further, our device works in the linear transport regime. The discussion on the non-linear transport regime is beyond the scope of this work. However, in a future work this aspect can be studied as whether the performance of the refrigerator increases more or not as compared to the linear transport regime.}
\acknowledgments
This work was supported by funds from Science \& Engineering research Board, New Delhi, Govt. of India, Grant No. EMR/20l5/001836.
|
\section{Conclusion}
\label{sec:conclusion}
We have presented a method to improve the availability of time-series
monitoring data for managing service-based systems.
To overcome the limited connectivity of the MANET nodes, the method
transfers the data over intermediate nodes in successive gossip
cycles. The method minimizes the network overhead caused by continuous
and repeated data transfer by storing local information about previous
transfers, imposing limits on the age of the data, and eliminating
irrelevant data. Through an extensive set of emulation-based
experiments, we have evaluated the capacity of the method to transfer
data from monitors to management nodes in two types of MANET
environments. Within the context of a case study, we demonstrated that
the method improves the results of management analysis tasks due to
the increased availability of monitoring data. Moreover, we have shown
how to tune the method to minimize the network overhead in the
resource-constrained MANET environment.
We have observed that the storage and communication overhead of the
method is relatively small. Of course, the actual amount of data
stored and transferred will depend on the nature of the time series of
interest.
Putting this into context, the overhead is
negligible compared to the overhead of the service platform and
services hosted within it.
We have compared the method to four other data harvesting methods and shown that harvesting time-series data in MANETs is a challenging task in which the method performs significantly better in both, data availability and network overhead.
An important issue that we have yet to explore is how to select the
parameters when the fundamental nature of the network dynamics
changes. This is a challenging problem because a good configuration is
dependent upon a variety of factors, including: (i)~network topology
characteristics, such as node degree and network diameter, which
affect the data propagation speed critical to the epidemic process;
(ii)~node mobility characteristics, which affect the patterns of node
groupings and network partitioning; (iii)~urgency of the data, which
determines how quickly the data need to be available at remote nodes;
and (iv)~the communication bandwidth available for the epidemic
process.
Ideally, we would like the method to be self-configuring, such that it
can recognize the factors above and deduce appropriate configuration
parameters. There are several approaches to consider incorporating
into our design for this purpose:
\begin{itemize}
\item
a closed-loop process in which the current network status and the data
propagation rate are piggy-backed within the data synchronization
messages;
\item
an advanced peer-selection process that considers the history of past
dataset transfers and the full or partial knowledge of the current
dataset stored in the peer candidates, rather than a statistically
random selection process, as a way to increase the speed of data
dispersion within the network;
\item
a hybrid approach that combines the ``push'' epidemic data propagation
method with the ``pull'' on-demand method, where the pull would be
initiated only when the data from remote nodes are recognized to have
not yet been pushed all the way to the management node;
\item
a back-pressure-like protocol for data dissemination in which the
synchronization agents can adjust their synchronization rates
according to available communication bandwidth and the data
availability at other nodes; and
\item
a broadcasting based data dissemination approach to increase the rate of data dispersion in suitable cases (such as with high density of one hop-distant neighbors).
\end{itemize}
\noindent
We are currently exploring these approaches in our on-going work,
which should result in an adaptive framework for data propagation and
harvesting in dynamic mobile networks.
\section{Introduction}
\label{sec:intro}
Mobile ad hoc networks (MANETs) are used to establish communication in
difficult environments such as search and rescue, forest-fire
fighting, and war zones. The increasing sophistication of end-user
devices has led to an increase in the richness and complexity of the
systems deployed on MANETs, including those structured as
interconnected and interdependent (micro)services, what are referred
to as \emph{service-based
systems}~\cite{Esfahani+:TAAS:2016,Jalaparti+:SIGCOMM:2013,Lund+:IEEE-CM:2007,Lund+:IEEE-CM:2010}. Our
interest is in developing methods to help manage those systems in
MANET deployment environments.
The foundation of any such management method is the collection of
run-time data describing the ``health'' state of the system. In
service-based systems designed for and deployed in MANETs, these data
change frequently, as the system attempts to adapt to the dynamicity
of the underlying network. For example, the bindings between services
can dynamically change according to the status of the nodes, links,
and paths.
Time-varying, local data are observed by \emph{monitors} situated at
the nodes of the network and captured as a \emph{time series}. The
time-series data are later \emph{harvested} from the nodes and
presented to a management element located somewhere in the network so
that it can perform a global state analysis, such as fault
identification~\cite{Novotny+:IEEE-TNSM:2015,5691315}, service
discovery and composition~\cite{Issarny+:Springer-JISA:2011}, or
service
re-placement~\cite{Novotny+:IEEE-MCC:2015,Silvestri+:ICCCN:2015}. As
it turns out, harvesting is an especially difficult problem in MANETs
due to the inherent resource limitations of wireless devices and the
fact that mobility can lead to network instabilities,
asymmetries,\footnote{An asymmetric network is one in which there is a
path from node $n$ to node $m$, but no path from $m$ to
$n$. Wireless media are subject to such communication asymmetries,
sometimes even between neighbors.} and partitions resulting from
relative node movements (e.g., going in and out of range) and link
properties (e.g., directionality, interference, and
noise)~\cite{Khandani+:IEEE-TWC:2008}.
\begin{figure}[t]
\centering
\includegraphics[width=0.6\columnwidth]{figures/reachability_combined}
\caption{Reachability of MANET nodes in two scenarios.}
\label{fig:reachability}
\end{figure}
To appreciate the seriousness of the problem, consider
Figure~\ref{fig:reachability}, which is a preview of experimental
results presented in this paper. The figure shows how the reachability
of a node, and therefore access to the data stored on it, can degrade
with time in MANET environments. The experiment examines two different
mobility behaviors under the same hypothetical service-based system.
One behavior is characterized by random, independent movements of
nodes, while the other is characterized by collective, grouped
movements. The experiment models a management element, residing at
some node in the network, needing access to the monitoring data stored
on a specific subset of nodes in the network. Approximately 90\% of
those nodes are reachable at time zero. As time progresses and nodes
move about, more nodes in that subset become unreachable, meaning that
less monitoring data are available to the management
element. Interestingly, we can see that random node movements can
cause more problems than grouped node movements.
The general problem of harvesting distributed data is, of course, not
new. Many mature solutions exist in the domain of wired, stable
networks, where brute-force techniques exploit the high capacity,
reliable network environment (e.g., IBM's Tivoli). For example, the
data can be continuously streamed to one or more repositories, or the
individual nodes can be reliably contacted to provide their data on
demand.
In the dynamic, unreliable, and resource-limited domain of MANETs,
however, the key issue is guaranteeing the availability of relevant
and timely data whenever and wherever it is needed, even if a node
whose data is of interest is unreachable. Existing mechanisms for data
dissemination in MANETs either \emph{ignore the problem} of network
asymmetries and partitioning, \emph{assume prior knowledge} of the
intended destination of the data, or \emph{focus on
single aggregated values} rather than the time-series data required
for system management tasks (see Section~\ref{sec:related}).
In general, then, the harvesting method must be sensitive to the
properties of time-series data and management tasks in MANETs. The
data are \emph{time bounded}, meaning that only some portion of a
monitor's stored data will be relevant to a given task, and \emph{time
sensitive}, meaning that their utility in a task degrades over
time. Moreover, only data from some subset of the monitors may be of
interest to a task. Finally, network instability and dynamics mean
that the location of the management node within the network topology
cannot be anticipated. In fact, the monitors may be unaware of
whether, which, and where their data might eventually be used.
Conceptually, our idea is to spread the data in an \emph{epidemic},
\emph{delay tolerant}, and \emph{intelligent} fashion, trading extra
communication and storage for increased availability in the presence
of node reachability problems. In particular, the method is:
\begin{itemize}[leftmargin=*,topsep=0pt]
\item
\emph{epidemic} in that it uses a gossip
protocol~\cite{Birman:OSR:2007} to create a network-wide data
transmission and replication overlay. Instead of requiring a
management node to obtain data directly from a monitored node using an
end-to-end path, the data are transferred to and stored at
intermediate nodes from which the data are then also available. Our
algorithm for selecting peers from among neighbors at each cycle in
the gossip protocol is designed to use connectivity metrics and a
standard random selection process to account for the dynamic MANET
topology and limited connectivity.
\item
\emph{delay tolerant} in that it opportunistically uses whichever
neighbors happen to be in range at a gossip cycle, rather than relying
on a fixed topology to reach some desired (albeit unknown) end
point. This is an approach to moving data across an unstable network
that probabilistically guarantees that the data will be in reach of a
management node.
\item
\emph{intelligent} in that it is sensitive to the time-bounded and
time-sensitive nature of time-series data. Moreover, it minimizes the
data exchanged among peers through a synchronization and aggregation
algorithm that takes account of the history of past encounters and
exchanges.
\end{itemize}
\noindent
We have implemented our method in Java~EE and carried out an extensive
set of validation experiments using the US~Naval Research Laboratory's
Common Open Research Emulator (CORE)
Extendable Mobile Ad-hoc Network Emulator (EMANE)
facilities~\cite{5680218}. We compare the performance of our
harvesting method against
four data harvesting methods,
through a case study: the capture of \emph{service dependence data}~\cite{Novotny+:IEEE-TNSM:2015}.
The experiments explore a
range of network and system dynamics, with the primary dependent
variable being the accuracy of the dependence graphs produced from the
data as compared to the ground truth.
The novelty of the work presented in this paper is two fold: (i)~an
extension of ideas developed in the domain of epidemic protocols to
solve the problem of efficiently spreading \emph{time-series} data and
(ii)~doing so in the context of the \emph{unstable and unreliable
environment} of~MANETs. The method is specifically designed to address the needs of service-based systems, however, in generalized form it can be applied to harvest any type of time-series data in MANETs such as data collected from IoT sensors, system monitors and other sources.
We next review related work. In Section~\ref{sec:technique} we
present the distributed monitoring architecture and gossip protocol
underlying our method. Our experimental setup and evaluation are
detailed in Sections~\ref{sec:setup} and~\ref{sec:results},
respectively. We conclude with a summary of our observations and a
look at on-going and future work in Section~\ref{sec:conclusion}.
\section*{Acknowledgment}
This research was sponsored by the U.S. Army Research Laboratory and
the U.K. Ministry of Defence and was accomplished under Agreement
Number W911NF-06-3-0001. The views and conclusions contained in this
document are those of the author(s) and should not be interpreted as
representing the official policies, either expressed or implied, of
the U.S. Army Research Laboratory, the U.S. Government, the
U.K. Ministry of Defence or the U.K. Government. The U.S. and
U.K. Governments are authorized to reproduce and distribute reprints
for Government purposes notwithstanding any copyright notation hereon.
\section*{References}
\bibliographystyle{model1-num-names}
\section{Related Work}
\label{sec:related}
The problem of maintaining the availability of data in a MANET
environment can be seen as a special case of the more general problem
of data sharing in that environment. From this point of view, there
are three broad areas of related work: peer-to-peer data access, data
replication, and epidemic protocols. While these areas are not
orthogonal, they represent the major foci of relevant work in the
MANET literature.
\para{Peer-to-peer data access.} Several techniques have been proposed
to create peer-to-peer overlays that can compensate for the dynamic
nature of a mobile
network~\cite{Heer+:PerComW:2006,Oliveira+:WoWMoM:2005}. The most
advanced of these techniques are distributed hash tables (DHTs), which
are designed to provide fast access to distributed data in the face of
``churn'' in the set of peers.
A DHT primarily serves as an efficient look-up service (``index''),
not as a storage service \textit{per se}~\cite{Shah:MSN:2011}.
As such, DHTs help improve the availability of data only in terms of
finding where in the network the data, and possibly redundant copies
of those data, can be found, but not in directly addressing what can
happen if the nodes storing the data are themselves
unavailable. Moreover, although designed for churn, DHTs are not well
suited to environments that experience especially high rates of churn,
as can be the case in MANETs.
\para{Data replication.} The goal of data replication is to increase
data availability by creating consistent replicas on multiple network
nodes~\cite{Derhab+:CST:2009, Padmanabhan+:VLDBJ:2008}. Most
techniques make use of information provided by or known in advance
about the users of the
data~\cite{Bellavista+:ISCC:2005,Hara:2003,Hara:2004,916653}, aiming
to create replicas that are topologically near to consumers. We make
the weaker assumption that such information is unknown. Similarly,
caching techniques~\cite{Atsan+:CN:2013,Fiore+:TVT:2011,Hara+:TMC:2006,Hao+:WIRLES:2005}
replicate frequently used parts of data into temporary stores topologically
closer to users. However, since requests for data are infrequent, and
typically only refer to small subsets of the available data, caching
is ineffective.
Some replication techniques try to optimize placement so as to
minimize the amount of data movement, either by limiting the locations
where data can be created~\cite{Pacitti+:VLDB:1999} or by delaying and
bundling updates for transfer in
bulk~\cite{MoonCho:IADIS-AC:2004}. These techniques are essentially
transactional in nature, viewing the data as a database that is
subject to updates, whereas data in our framework is in the nature of
a time series.
Only a small number of data replication techniques explicitly address
the problem of disconnected
nodes. Huang~et~al.~\cite{Huang+:CIKM:2003} and Wang and
Li~\cite{WangLi:INFOCOM:2002} use mobility and data usage patterns to
place and redistribute replicas.
Hauspie~et~al.~\cite{Hauspie+:CNRS:2002} attempt to predict network
partitions, by measuring the quality of links and end-to-end paths, in
order to decide where best to place replicas. Although these
techniques do indeed increase data availability in partitioned
networks, they introduce substantial additional network overhead to
test links and discover paths.
\para{Epidemic protocols.} An epidemic protocol (and the related
technique of gossiping) is a distributed algorithm that uses periodic
interactions between nodes to propagate data through a
network~\cite{Birman:OSR:2007,Dimakis+:IEEE:2010,Friedman+:SIGOPS:2007,Reina:AHN:2015}.
Particularly relevant to the problem addressed in this paper is that
the protocols do not require reliable communication links. The peers
(or neighbors) engaged in the protocol are usually chosen randomly.
In the context of MANETs, epidemic protocols are typically used in
conjunction with in-network computations, such as distributed signal
processing~\cite{RabbatNowak:IPSN:2004,Yu+:INFOCOM:2004}, to decrease
the overall network load by propagating only aggregate data values.
Our method makes heavy use of epidemic protocols and in-network
computations, where those computations are instead meant to
intelligently synchronize the time-series data traversing common
network paths rather than perform simple data aggregations.
\section{Experimental Evaluation}
\label{sec:results}
We focus the evaluation of our delay tolerant harvesting method on
the following key issues:
\begin{enumerate}
\item impact of the synchronization frequency;
\item impact of the number of peers;
\item impact of constraining the number of peers;
\item comparison to four data harvesting methods;
\item tradeoff between overhead and precision.
\end{enumerate}
We study these issues in the context of the dependence discovery use
case.
\subsection{Impact of synchronization frequency}
The frequency at which peers attempt to disseminate monitoring data is
a fundamental tuning parameter in our method, as it induces a trade
off between communication overhead, on the one hand, and increased
availability of monitoring data on the other. To isolate the impact of
this parameter, we fix the delay between the end of the client
conversation and the start of the dependence analysis at five minutes,
and vary the number of gossip cycles within that five-minute period
from zero (no synchronization) to~32.
At each gossip cycle, we have each node select at most one peer node
at random from its neighbors in the network.
\begin{figure}[t]
\centering
\includegraphics[width=0.6\columnwidth]{figures/single_backup}
\caption{Availability of monitoring data, as measured by TP ratio,
with different numbers of gossip cycles.}
\label{fig:ep-single-backup}
\end{figure}
We hypothesize that as the number of cycles increases in the period
between the end of the conversation and the start of the data
analysis, so too should the availability of the monitoring data,
resulting in an improved TP ratio. With no synchronization occurring
before the data harvesting, the only monitoring data available to the
harvesting node are those at its local monitor. However, with every
next cycle, the amount of data arriving from remote monitors should
increase.
The results are reported in
Figure~\ref{fig:ep-single-backup}. Initially, with no gossip cycles,
the only data available are from the client monitors and from the
locally hosted services. As the number of gossip cycles increases, the
nodes receive increasing amounts of monitoring data from remote
monitors via the gossip protocol.
Comparing the two mobility scenarios, the availability of the data
increases faster with an increasing number of gossip cycles under the
military scenario than under firefighting. This is because, in the
military scenario, the services involved in the conversations are with
high probability hosted on nodes within the same, relatively stable,
group of nodes as the client. On the other hand, in the firefighting
scenario, the services involved in conversations in general should be
drawn from greater distances, since the nodes are dispersed relatively
evenly over a large area, requiring the monitoring data to be
transferred between more nodes than in the military scenario and so
requiring more gossip cycles to achieve an equivalent TP ratio.
Eventually, with 32 cycles, the availability reaches 99.8\% in the
military scenario and 98.8\% in the firefighting scenario.
\subsection{Impact of number of peers}
We now investigate the impact of the number of peers selected in each
gossip cycle. Again, dependence analysis is started five minutes after
the conversation ends and the candidate peers are limited to
neighbors. We fix the synchronization frequency to one per minute,
i.e., four gossip cycles in that five-minute period. Our hypothesis is
that increasing the number of peers should increase the availability
of monitoring data.
In practice, the number of selected peers is bounded not only by the
upper-limit parameter, but also by the number of neighbors. Because
the size of the candidate neighbor set depends on the node density and
mobility of the network, the number of selected peers on average will
be somewhat lower than the configured limit. Furthermore, due to the
presence of poor-quality wireless links, the number of peers that
successfully receive the data can be smaller than the number of
selected peers. Especially when many peers are selected, congestion in
the wireless medium hampers message transmission.
This gap can be seen in Figure~\ref{fig:ep-neighbors}, which shows a
varying upper limit, the number of selected peers (``candidates''), and
the number of peers that
successfully receive the data under the two mobility scenarios. The
gap between these numbers widens as the upper limit increases. Notice
the subtle difference between the two scenarios. The gap consistently
increases in the firefighting scenario, where the nodes are evenly
dispersed. On the other hand, in the military scenario, there is
little gap between the upper limit and the number of selected peers
below a certain point (eight nodes), and then quickly widens beyond
that. This is due to the nature of the military scenario, where nodes
move in cohesive groups of certain sizes.
The availability of the monitoring data is reported in
Figure~\ref{fig:ep-multiple-backups}. As expected, with increasing
numbers of peers per gossip cycle, the availability of the data also
increases. However, a threshold of about four peers is reached beyond
which there is little further gain. This is explained by the gap
between the upper-limit parameter and the number of peers successfully
receiving data, as discussed above. In principle, this threshold
should be taken into account in the configuration of the method to
balance overhead against the expected quality of the monitoring data
analysis result.
\begin{figure}[t]
\centering
\begin{tabular}{@{}c@{}}
\includegraphics[width=0.6\columnwidth]{figures/neighbors_c1}
\end{tabular}
\caption{Upper limit on peers, number of candidate neighbors, and
number of peers successfully receiving data.}
\label{fig:ep-neighbors}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.6\columnwidth]{figures/multiple_backup}
\caption{Availability of monitoring data, as measured by TP ratio,
with different upper limit on peers per gossip
cycle.}
\label{fig:ep-multiple-backups}
\end{figure}
\subsection{Impact of constraining the number of peers}
In the next experiment we return to the synchronization frequency, but
remove the constraint on the number of peers.
Unlike in the previous experiments, the number of selected peers is
bounded only by the availability of the neighbors. Hence, the factor
that determines the inclusion of nodes is the node density and
mobility of the network. The dependence analysis is started five
minutes after the conversation ends and we vary the number of gossip
cycles within that period. We expect that increasing the number of
gossip cycles should increase the availability of monitoring data.
When the number of peers is unconstrained, data are propagated one hop
in each cycle from their source in all directions. Thus, the number of
cycles needed to make the data available to an analysis element
reflects the distance from the nodes on which those data are
collected. This distance, however, is impacted by the dynamics induced
by node mobility, as well as by the quality of the wireless
links. Furthermore, congestion in the wireless medium may have a more
significant impact on the transmissions than in the previous
experiments, since the data are sent in contiguous sequences to
multiple peers.
The results are reported in Figure~\ref{fig:ep-unconstrained-backup}.
In the firefighting scenario, where there is a relatively even
dispersion of nodes over a large area, the average number of
1-hop-distant peers is~7.2. In the military scenario, where nodes move
in relatively stable groups, the average number of 1-hop-distant peers
is~10.4. As with the one-peer-per-cycle configuration, increasing the
number of cycles increases the availability of the data. A threshold
of about two cycles is reached beyond which there is only a small gain
from further cycles. The higher number of 1-hop-distant peers in the
military scenario leads to a higher availability with same number of
cycles than in the firefighting scenario. However, to achieve a data
availability of 99\% or more, at least four cycles are needed in
either scenario to reach more distant nodes.
To compare the one-peer-per-cycle and unconstrained-peers
configurations, we evaluate the network overhead imposed by each
configuration to achieve a comparable data availability. In the
firefighting scenario, the one-peer-per-cycle configuration requires
32 cycles to achieve 98.8\% data availability, while the
unconstrained-peers configuration resulting in about 7.2 peers-per-cycle requires four cycles to achieve a
comparable 99.1\% availability. The overhead of the one-peer-per-cycle
configuration is 0.32~KB/s per node and 0.38~KB/s per node in the
unconstrained-peers configuration. In the military scenario, the
one-peer-per-cycle configuration requires 32 cycles to achieve a
99.8\% data availability, while the unconstrained-peers configuration
resulting in about 10.4 peers-per-cycle requires 8 cycles to achieve a comparable 99.3\% availability. The
overhead of the one-peer-per-cycle configuration is 0.19~KB/s per node
and about 0.29~KB/s per node in the unconstrained-peers
configuration. Hence, the unconstrained-peers configuration induces
overall about a 30\% to 50\% higher network overhead than the
one-peer-per-cycle configuration to achieve a comparable data
availability.
\begin{figure}[t]
\centering
\begin{tabular}{@{}c@{}}
\includegraphics[width=0.6\columnwidth]{figures/unconstrained}
\end{tabular}
\caption{Availability of monitoring data after four gossip cycles, as measured by TP ratio,
with different numbers of gossip cycles and an unconstrained number
of peers.}
\label{fig:ep-unconstrained-backup}
\end{figure}
\subsection{Comparison to data harvesting methods}
To place our method in context, we next compare it to
two baseline methods: \emph{Gossip protocol} and \emph{DHT}, and to
two state of the art methods: \emph{DAFN} and \emph{SCALAR}.
\subsubsection{Gossip protocol}
A na\"{i}ve approach to harvesting would make use of a generic gossip
protocol that is not sensitive to the data already possessed by peer
nodes, and thus passes \emph{all} available data in \emph{each}
transmission regardless of past interactions.
We base the na\"{i}ve method on the design of a standard gossip
protocol~\cite{Jelasity+:ACM-TOCS:2007} and use a \emph{uniform
random} selection of peers, \emph{push propagation} of data, with
\emph{no confirmation} of the success or failure of
transmissions. Nevertheless, rationally, we constrain the maximum age
of data to be passed.
\subsubsection{DHT}
A similarly na\"{i}ve method to harvesting would make use of DHT with
locations of time-series within the network, and attempt
to transfer data with direct, on-demand data harvesting approach. In that
approach, the analysis element attempts to communicate \emph{directly}
with individual monitoring nodes to obtain (``pull'') their data, rather
than using a gossip protocol to disseminate (``push'') the data. The
availability of the data is therefore limited by the network reachability of the
nodes.
\subsubsection{DAFN}
Dynamic Access Frequency and Neighborhood (DAFN) method~\cite{Hara+:TMC:2006} increases data availability with collaborative message relaying and placing data replicas on mobile
nodes based on data access frequency. Every node maintains a table of frequency
of accessing data. When node requests data, a request message is broadcasted to
its neighbors. Upon reception of the request message, node either responds with
data from its local cache or forwards the request to its neighbors. Upon
reception of the data, the node stores the data into the local cache. To
eliminate redundant data replicas between neighbor nodes, each connected set of
nodes elects a coordinator responsible for optimizing the allocation of
replicas. The coordinator monitors the allocation of replicas between neighbor nodes
and regularly prunes replicas on nodes with lower access frequencies.
\subsubsection{SCALAR}
Scalable data Lookup And Replication framework (SCALAR)~\cite{Atsan+:CN:2013}
is a scalable method of data lookup and passive replication in MANETs. The
method builds a dynamic virtual backbone between mobile nodes based on an
approximation of minimum connected dominating set within a set of connected
nodes. Thus, every node within a set of connected nodes is directly connected to
at least one node of the virtual backbone. The virtual backbone is used to
minimize the number nodes involved in relaying data request and response
messages within the network. A scalable data lookup protocol uses the virtual
backbone to relay data request and response messages between clients and data
sources. A reactive replication mechanism uses the data lookup protocol to
preload data from sources closer to clients. The mechanism preloads data based on
monitoring of request frequency and the distance between requester and source of
the data.
\subsubsection{Comparison of data availability}
\begin{figure}[t]
\centering
\begin{tabular}{@{}c@{}}
\includegraphics[width=0.6\columnwidth]{figures/comparison_availability_firefighting}
\\
\textbf{(a)} \\[5pt]
\includegraphics[width=0.6\columnwidth]{figures/comparison_availability_military}
\\
\textbf{(b)}
\end{tabular}
\caption{Comparison of TP ratio of dependencies with DHT, DAFN, SCALAR, Gossip protocol and our Harvesting method (HM) in the
firefighting (a) and military (b) scenarios.}
\label{fig:available-dependencies}
\end{figure}
Our aim in this experiment is to compare the data availability achieved by the methods under range of delays of starting data harvesting after the end of conversation.
Our method is configured with a single peer per cycle and a frequency of 32
cycles within the delay of data harvesting after the conversation and
the \emph{Maximum age limit of time slot} parameter is set to the length of the delay of data harvesting after the conversation.
Since the na\"{i}ve method uses the same fundamental gossiping mechanism, we
configure the method same as our harvesting method, with a single peer and
32 cycles within the delay period.
In the DAFN method, we do not constrain the size of cache available for storing the time-series data replicas on nodes. Thus, the method’s decision on
reducing the number of data replicas between neighbor nodes is based solely on
the frequency of data access. Similarly, we do not constrain the size of the
cache available to data replicas in the SCALAR method. To achieve high data
availability, we configure the method to actively replicate incremental changes in
the time-series if there were more than one request
within the delay period.
Figure~\ref{fig:available-dependencies} shows the availability of
monitoring data using DHT, DAFN, SCALAR, Gossip protocol and our Harvesting method (HM), measured in terms of the TP ratio. The x-axis represents the delay between the end of a
conversation and the beginning of the harvesting.
Notice that the
general trend of data availability of DHT, DAFN and SCALAR methods follows the reachability of
nodes in the network shown in Figure~\ref{fig:reachability} of Section~\ref{sec:intro}. Initially, with no delay in issuing data harvesting requests after the conversation, data are highly available because the topology of the network has not significantly changed. Thus, methods which use the (``pull'') approach for obtaining data on request achieve high data availability. However, with increasing delay between end of conversation and the requests for data from clients, movements of nodes decrease the data availability.
The baseline DHT method, harvesting data directly on-demand over multiple hops achieves 90\% and 69\% TP ratios immediately after the conversation and declines to 75\% and 45\% TP ratios when requests for data are issued 16 minutes after the conversation in the military and firefighting scenarios, respectively. The DAFN method uses collaborative relaying of messages coupled with caching of obtained data on multiple replicas within the network. This approach increases the data availability to 93\% and 82\% TP ratios after the conversation and declines to 81\% and 49\% TP ratios at 16 minutes after the conversation. The SCALAR method based on relaying messages over virtual backbone, caching and limited (``push'') based replication yields similar data availability at 93\% and 80\% TP ratios after the conversation and declines to 84\% and 55\% TP ratios at 16 minutes after the conversation.
The baseline Gossip protocol and our Harvesting method both employ a (``push') approach to disseminate time-series data through the network. Both methods achieve very high data availability. Immediately after the conversation, before any data can be disseminated in the network, both methods achieve 34\% TP ratios in both scenarios, providing only data available in local stores. However, once the data start to be disseminated in the network, the availability increases to 99.9\% and 99.6\% TP ratios in the military and firefighting scenarios, respectively. The high availability somewhat declines to 98.5\% and 97.3\% TP ratios when requests are issued 16 minutes after conversation, caused by low frequency of disseminating data from migrating nodes.
\subsubsection{Comparison of network overhead}
\label{sec:comparison-network-overhead}
\begin{figure}[t]
\centering
\begin{tabular}{@{}c@{}}
\includegraphics[width=0.6\columnwidth]{figures/comparison_overhead_firefighting}
\\
\textbf{(a)} \\[5pt]
\includegraphics[width=0.6\columnwidth]{figures/comparison_overhead_military}
\\
\textbf{(b)}
\end{tabular}
\caption{Comparison of transfer overheads of DHT, DAFN, SCALAR, Gossip protocol and our Harvesting method (HM) in the
firefighting (a) and military (b) scenarios.}
\label{fig:comparison-overhead}
\end{figure}
Here we compare the network overhead of the methods. We hypothesize that the DHT based method will induce only a small overhead, while the overhead of the Gossip protocol will be significantly higher due to its sending of the same data multiple times. The DAFN method, flooding the network with request and control messages will also induce significant overhead. The SCALAR method, utilizing caches and limiting the passing of messages to a subset of virtual backbone nodes will likely be more efficient. Our method should ideally induce significantly less overhead than the Gossip protocol and DAFN methods. However, since in our method the same data may be sent along different links at each cycle, the improvement will depend on the topology and connectivity properties of the network, and will certainly induce greater transfer overhead than the DHT.
Note, that in this experiment, the delay between the end of conversation and beginning of harvesting varies among the scenarios, however, the total number of analyzed conversations is constant.
The results are reported in Figure~\ref{fig:comparison-overhead}. The DHT based method yields an overhead of only about 0.04~KB/s and 0.06~KB/s per node in the two scenarios with one minute delay and decreases to 0.02~KB/s and 0.04~KB/s per node with 16 minutes delay. The DAFN method yields significantly higher overhead, about 4.41~KB/s and 7.21~KB/s, and 1.47~KB/s 2.08~KB/s with one and 16 minutes delays respectively, and the SCALAR method yields moderate overhead of about 0.72~KB/s and 0.67~KB/s, and 0.34~KB/s 0.25~KB/s with one and 16 minutes delays respectively.
The Gossip protocol yields substantially higher overhead of about 7.59~KB/s and 13.16~KB/s with one minute delay and decreases to about 1.22~KB/s and 2.21~KB/s with 16 minutes delay. Our method yields low about 0.87~KB/s and 1.24~KB/s and 0.07~KB/s and 0.11~KB/s transfer overhead to achieve the same levels of availability, placing it as expected between the DHT and DAFN as well as below SCALAR in all cases except in scenarios with one minute delay.
\subsubsection{Storage requirements}
The storage requirements are an artifact of the degree of data dissemination in
the network and the number of nodes holding data backups. The time-series data
are stored in the form of regularly pruned series of Boolean values. Overall,
the storage requirements of the harvesting methods are very small. Our method
uses all network nodes to proactively gossip all changes in the monitoring data
throughout the network with each node holding backup of the data received. In
the explored scenarios, the amount of data stored on the individual nodes was in
the order of tens of KB after data older than 20 minutes are pruned. Similarly,
DAFN uses all network nodes to relay data from monitors to clients, yet, with a
lower degree of data dissemination. The method further actively reduces the
number of data backups by eliminating neighbor replicas. The SCALAR method
limits the nodes participating in data transfer and holding data backups to
those of the virtual backbone. The DAFN and SCALAR methods required about 22 to
29\% and 8 to 12\% of storage space compared to our method, respectively. The
Gossip protocol achieves the same degree of data dissemination as our method and requires same storage space, while the DHT based method does not store any
backup data and therefore does not use any storage space.
\subsubsection{Comparison summary}
The transfer of incremental changes of the time-series data is a challenging task for the data transfer methods. The DHT and the Gossiping protocol represent the two extremes of minimum overhead with low data availability and very high overhead with high availability respectively. The DAFN method with very high overhead provides only somewhat higher availability than DHT. The method is repeatedly flooding the network with requests for latest additions of the time-series data and makes low utilization of caches due to data becoming quickly obsolete. The SCALAR method provides somewhat higher availability than DAFN with lower overhead due to utilization of virtual backbone and caching reducing the number of messages requesting and disbursing latest additions of the time-series data. Our Harvesting method, designed specifically to continuously and efficiently transfer incremental changes in the time-series is yielding very high data availability with low network overhead.
\subsection{Tradeoff between overhead and precision}
As a final point of evaluation we look at the important issue of
overhead versus precision.
The amount of data transmitted depends on the size
of the system (i.e., the number of time series), the configuration of
the method (i.e., gossip cycle frequency and number of peers per
cycle), and the length of the time slot within the time series. The
size of the system is an application-specific contextual property,
while the configuration settings reflect operational requirements. The
length of the data transfer time slot impacts overhead imposed on the
network, as well as the resolution (compression) of the transmitted
data. It thus impacts the precision of the resulting analysis applied
to the monitoring data.
To understand this effect for our case study, we use the ratio of
\emph{false positives} (FP)~\cite{Novotny+:IEEE-TNSM:2015,5691315} in
dependence graphs to indicate the impact of (im)precision in the
dependence data, defined as follows:
\begin{equation*}
{\mathit{FP\ ratio}} = \frac{| D(C) - GT(C)|}{|D(C)|}\nonumber
\end{equation*}
where as for the TP ratio defined in Section~\ref{sec:expsetup},
$D(C)$ is the set of discovered dependencies and $GT(C)$ is the set of
ground-truth dependencies, under the assumption that $D(C)$ and
$GT(C)$ are non-empty.
The FP ratio represents the fraction of dependence data not belonging
to a conversation and so erroneously included in the result; a high FP
ratio indicates poor precision. These irrelevant dependencies arise
from a combination of the monitors aggregating the dependence data
into time slots and the inherent behavioral effects of our harvesting
method. We hypothesize that increasing the length of the time slot
used in the transfer of data will decrease the communication overhead,
but increase the FP ratio.
We present our results in Figure~\ref{fig:resolution}, where we show
the impact of the time slot length on the overhead and FP ratio.
Consistent with the previous experiments, the method is configured
with a single peer per cycle and a frequency of 32~cycles within
period of five minutes, yielding a high availability. The results
confirm our hypothesis: with a 0.1~second time slot length, the same
as that used by the monitor, the FP ratio is about 16\% and 18\% in
the military and firefighting scenarios, respectively, while the
overhead is 0.18~KB/s and 0.32~KB/s on average between pairs of
peers. When increasing the length of the time slot to 10~seconds, the
FP ratio increases to 35\% and 36\%, while the overhead decreases to
0.12~KB/s and 0.85~KB/s.
We observe that when the length of the time slot is decreased by 100
times, the FP ratio increases by only about two times, while the
overhead decreases by 33\% and 73\%. This suggests that for
applications able to tolerate a higher FP ratio, trading overhead over
higher FP ratio might be a viable option. Note that on average a node
sends only between 0.09 to 0.32~KB/s of data depending on the scenario
and the length of time slot.
\begin{figure}[t]
\centering
\includegraphics[width=0.6\columnwidth]{figures/resolution_n_overhead}
\caption{Impact of the transfer dataset time slot length on FP ratio
and overhead.}
\label{fig:resolution}
\end{figure}
\subsection{Summary of results}
\label{sec:configuration}
The experimental results presented above establish the relationship
between the configuration parameters of our method (namely, the
synchronization frequency, the number of peers selected at each cycle,
and the length of the transfer dataset time slot), and its ability to
improve the availability of monitoring data.
In general, it is shown that selecting either a higher frequency with
a lower number of peers, or a lower frequency with a higher number of
peers, provide similar outcomes measured in terms of the availability
of the monitoring data at remote nodes. Furthermore, selecting a
shorter length for the transfer time slot provides higher precision
than a longer time slot. However, the data transfer overhead is
significantly higher with a shorter time slot length.
When considering network overhead, configurations using a lower number
of peers and a higher number of cycles achieves comparable data
availability with lower overhead, than configurations with a higher
number of peers and a lower number of cycles.
In comparison to DHT, DAFN, SCALAR and Gossip protocol, our approach achieves
significantly higher data availability and lower overhead, especially in dynamic network
environments such as MANETs, where mobility can easily lead to nodes
becoming unreachable, yet their data remaining critical for analysis.
\section{Evaluation Methodology}
\label{sec:setup}
Our purpose in the evaluation is to understand the performance of the
harvesting method under various conditions. In particular, we are
interested in how well the method can improve the reachability (i.e.,
availability) of monitoring data in the face of network dynamics.
In this section we describe the method we use to evaluate our
approach. The method is based on a case study in which we experiment
with the problem of discovering service dependencies. After describing
the case study, we detail the tools, metrics, and scenarios used to
conduct the experiments.
\subsection{Case study: service dependence data}
\label{sec:casestudy}
In previous work, we developed a technique for discovering dynamic
dependencies among the distributed components of MANET-hosted
applications that are structured as assemblies of
(micro)services~\cite{Novotny+:IEEE-TNSM:2015}. The technique suffered
from the problem that it assumes all nodes of interest are reachable,
on demand, from the node where the dependence analysis is to be
carried out. In fact, it is a common occurrence that not all nodes are
reachable, which significantly reduced the effectiveness of the
technique and inspired the design of our new harvesting method. In
this case study, we evaluate how well the new method can improve the
availability of the monitoring data and, thereby, the effectiveness of
the dependence discovery technique.
In service-based systems, a \emph{dependence} is a relation between
services defined by the message flow, called a \emph{conversation},
induced by a client request and normally ending with a response to
that request. (A dependence is also the relation between a client and
a service. Without loss of generality, we mainly focus here on
relations among services.) When a dependence relation exists between
two services, one is considered the \emph{source} and the other the
\emph{target}. In general, sources issue requests (i.e., method calls)
on targets, thus defining a directionality to the dependence. Targets
are expected to provide replies (i.e., response messages) back to
sources.
A \emph{dependence graph} (DG) captures the run-time dependencies among
services and is the output of a dependence discovery analysis tool. A
DG can be used to represent the full set of dependence relations in the
system, or can be restricted to a subset of those relations.
Figure~\ref{fig:service-assembly} depicts a simple example of several
DGs rooted at clients. Highlighted in the figure are Client~2 and the
service instances it employs, both directly (Service~2) and
transitively (Services 4, 5, and~8). A DG can be combined with network
and service failure data to perform global fault identification
tasks~\cite{Novotny+:SRDS:2012,Tati+:SPIE:2013}. For example, we can
probabilistically identify the root cause of a failed conversation.
To cope with network topology changes, service-based systems deployed
in MANETs make use of dynamic service binding
mechanisms~\cite{Mian+:IPC:2009}. This leads to time-varying
dependencies, which in turn are represented as a time series of data
points, each giving a snapshot of the dependencies at a given instant.
For meaningful use in dependence discovery, relevant and timely
dependence data must be available to the analysis tool.
\begin{figure}[t]
\centering
\includegraphics[width=0.6\columnwidth]{figures/dependence_graph3}
\caption{Service dependence graphs.}
\label{fig:service-assembly}
\end{figure}
Consistent with the generic monitoring architecture
(Section~\ref{sec:architecture}), the dependence data are gathered by
local monitors. In this case, the monitors are deployed within service
containers to observe service-level message traffic
(Figure~\ref{fig:architecture}). The dependence data for a particular
client conversation, covering a specific time period, is provided for
the benefit of a dependence discovery analysis element that then
produces a corresponding~DG. We assume that the dependence discovery
element is a component that can be hosted in any arbitrary node or
nodes (mobile or fixed) in the MANET, but that its specific location or
locations is not globally known.
\begin{figure}[t]
\centering
\includegraphics[width=0.6\columnwidth]{figures/architecture2}
\caption{Monitoring and harvesting dependence data. Note that only two
of possibly many nodes are shown.}
\label{fig:architecture}
\end{figure}
Each time series represents the time-varying dependencies between a
source and target, in which entries for each time slot are Boolean
data about whether or not the given dependence occurred within that
time slot. When the monitor detects the occurrence of a dependence, it
signifies this by setting a 1-bit flag in the corresponding time
slot. It also records identifying information about the source and
target of the dependence. The set of time slots thus represents an
aggregated time series of dependencies. The set of relevant time slots
shifts as new time slots are added and obsolete ones removed,
reflecting the changes in dependencies.
\subsection{Experimental setup}
\label{sec:expsetup}
As mentioned in Section~\ref{sec:intro}, we conduct our experiments in
the widely used
CORE\footnote{\url{http://cs.itd.nrl.navy.mil/work/core/}} network and
EMANE\footnote{\url{http://cs.itd.nrl.navy.mil/work/emane/}} mobility
emulator frameworks. CORE provides a network experimentation
environment using the container-based virtualization facility of the
Linux platform, while EMANE provides real-time modeling of
wireless-link and physical-layer connectivity. This combination
provides high-fidelity, real-time emulation. A detailed description
of the evaluation stack is provided elsewhere~\cite{
NovotnyWolf:DTR-2016-7:2016}.
The application running within the MANET is built as a generic Web
service system based on Java~EE
Glassfish,\footnote{\url{https://glassfish.java.net/}} the reference
implementation of Java~EE's application platform, and Glassfish
Metro,\footnote{\url{https://metro.java.net/}} a reference
implementation of a standard Java Web services stack. The system is
composed of two kinds of configurable components, a generic client
application and a generic Web service, structured as a 2-tiered
system. The first tier consists of client-facing, ``front-end''
services, while the second tier consists of interconnected ``back-end'' services. In our experiments, we use 50~clients, five front-end services, and 20
back-end services. When starting a conversation, each client invokes
a method selected uniformly at random from all methods provided by the
front-end services. The invocation triggers a cascade of message exchanges between the interconnected back-end services yielding a network traffic.
Monitors are implemented as \emph{Tubes} in the Metro framework.
Tubes are chained components, each responsible for part of the
processing of incoming and outgoing service messages. Monitors
intercept incoming and outgoing messages of the clients and services
to extract the necessary information. The monitor extracts the
dependence fields from the intercepted messages and records
occurrences of dependencies.
Synchronization agents are implemented as Java applications that
repeatedly synchronize with other agents, receive data from remote
agents, and maintain in-memory backup stores of the received data. For
purposes of evaluation, synchronization agents also record information
about their activities in a trace file for later performance analysis
(e.g., measurements of network overheads, success rates of
synchronization attempts, and the like).
Finally, dependence discovery analysis elements, implemented as a
library, are used by clients to discover DGs for the conversations
they initiate (i.e., a series of message exchanges), by querying the
dependence information found in their local stores as populated by our
harvesting method.
\begin{table}
\centering
\begin{tabular}{@{}|l|r|@{}}
\hline
Number of nodes & 50 \\
Mobility speed & 3 - 6.6 km/h \\
WiFi standard & 802.11b \\
WiFi unicast rate & 11Mbps \\
WiFi multicast rate & 1Mbps \\
Transmit power & -15 dBm \\
Path loss mode & 2ray \\
Routing protocol & OLSR \\
Protocol stack & TCP/IPv4 \\
\hline
\multicolumn{1}{c}{\mbox{}}\\
\end{tabular}
\caption{Network-layer parameters.}
\label{tab:network-parameters}
\medskip
\begin{tabular}{@{}|l|r|@{}}
\hline
Number of clients & 50 (one per node) \\
Number of services & 25 \\
Size of dependence graph & 4 \\
Invokable methods per service & 2 \\
Workload (client request rate) & 30s \\
Number of service replicas & 5 \\
Response timeout & 60s \\
\hline
\multicolumn{1}{c}{\mbox{}}\\
\end{tabular}
\caption{Service-layer parameters.}
\label{tab:service-parameters}
\medskip
\begin{tabular}{@{}|l|r|@{}}
\hline
Monitor time slot length & 0.1s \\
Transfer dataset time slot length & 0.1s \\
Transfer dataset response time & 60s \\
Maximum age limit of time slot & 300s \\
Maximum peer distance & 1 hop \\
Number of gossip cycles & 0-32 \\
Number of peers & 1-10 \\
\hline
\multicolumn{1}{c}{\mbox{}}\\
\end{tabular}
\caption{Gossip protocol parameters.}
\label{tab:agent-parameters}
\end{table}
Each (virtual) CORE node runs a client and/or a service, along with a
monitor, a synchronization agent, and an analysis element. (The
analysis element would not normally be deployed at all nodes, but we
do so here to give us maximum flexibility in evaluating the
dissemination of monitoring data.) The DGs are constructed on demand
by the discovery element. The graphs are rooted at a given client,
beginning at a given time instant, and for some time window. Each DG
is constructed for a particular conversation; thus, the time window
begins and ends with the start and end of that conversation.
In our experiments, one of the clients is chosen as the component
requiring the dependence data, such as when it needs to perform a
fault analysis as the initiator of a failed
conversation~\cite{Novotny+:SRDS:2012}. This reflects a realistic
use case in which the location of the analysis element cannot be
known \textit{a priori}.
To observe the effects of network dynamics, we use two different node
mobility patterns as scenarios. The first one, ``military'', consists
of a unit of 50 members, each carrying a mobile device. The members
are collected into several subunits, each of which moves as a whole
within an area of 2km~x~2km, exhibiting a group mobility driven by the
\emph{nomadic community mobility model}~\cite{Roy:2011}. The second,
``firefighting'', represents a pattern of independent mobility of 50
members in a 1km~x~2km area, driven by the \emph{random waypoint
mobility model}~\cite{Roy:2011}. Other network and service
parameters used in the experiments are summarized in
tables~\ref{tab:network-parameters}
and~\ref{tab:service-parameters}. Most notably, the mobile nodes move
at a walking-like speed in the range of 3 to 6.6km/h. We use WiFi
standard 802.11b and OLSR as a routing protocol.
The configuration of the gossip protocol for our experiments is
summarized in Table~\ref{tab:agent-parameters}. The lengths of the
monitor and transfer dataset time slots are both set to 0.1~seconds,
except as part of the last experiment presented in
Section~\ref{sec:results} in which we vary the transfer time slot from
0.1 to 10~seconds. The resolution of 0.1~seconds reflects the need for high precision of the captured dependencies necessary to construct accurate DG.
The monitor waits at most 60 seconds for confirmation
of successful receipt of the dataset. The maximum age of a time slot to be transferred is
limited to five minutes and the selection of candidate peers is
limited to network neighbors (i.e., 1-hop distant). The number of
gossip cycles within a given time period determines the frequency at
which each agent synchronizes the data in its local store with that of
its peers. The final parameter, number of peers, allows us to vary the
maximum number of other agents with which each agent shares the data
in its local store during each gossip cycle.
We collect our results from 40~minutes of execution after excluding
10~minutes of warm up. Each combination of parameters results in
thousands of conversations during the 40-minute execution. The results
given in the next section are averages over the data collected from
these conversations, where each conversation is then a statistical
sample subject to the random variables.
The primary evaluation question for our harvesting method is how well
the dependence data of each conversation~$C$ has been propagated
through the network after a certain period of time. More specifically,
we measure the quality of the harvest in terms of the ratio of
\emph{true positives} (TP) in the dependence analysis result, defined
as follows~\cite{Novotny+:IEEE-TNSM:2015,5691315}:
\begin{equation*}
{\mathit{TP\ ratio}} = \frac{| D(C) \bigcap GT(C)|}{|GT(C)|}
\end{equation*}
\noindent where $D(C)$ is the set of discovered dependencies, $GT(C)$
is the set of ground-truth dependencies, and true positives are in the
intersection of these two sets (we assume $D(C)$ and $GT(C)$ are
non-empty). A good result for our method would be that it can transfer
as many dependencies of $C$ as possible, while not decreasing data
resolution due to aggregation.
The secondary evaluation question is the network overhead imposed by
our harvesting method. We measure the overhead of a network node as an average of sum of all data sent and received by the synchronization agent hosted on the node within a unit of time (i.e. KB/s). The metric includes content (i.e. headers and payload) of control and data messages induced by the harvesting method in all peer-to-peer and client-to-peer exchanges. The dominant traffic in the network induced by the conversations between clients and services, as well as, the messages exchanged by the underlying routing protocol, are both excluded from the metric.
\section{The Harvesting Method}
\label{sec:technique}
We now introduce our harvesting method by first describing the basic
monitoring architecture and then detailing how the data are
disseminated through the network using a gossip protocol.
\subsection{Monitoring architecture}
\label{sec:architecture}
At each (participating) node in the network, the harvesting method
makes use of two kinds of components and a local store. One component
is a \emph{monitor}, situated at a node, responsible for gathering
time-series data about the local state of the network and hosted
applications at that node. Monitors store the time-series data
according to \emph{time slots} of some length and use a sliding
expiration window such that only a limited history is maintained. A
time slot represents either a single measurement or an aggregate
series of measurements taken within the period of time represented by
the time slot. The time series thus consists of a continuous sequence
of time slots. What specific data are collected and how monitors
manage to gain access to those data are not of concern here, as this
is largely a domain-specific problem solvable in a variety of
well-studied ways; in the case study of Section~\ref{sec:casestudy} we
describe one particular mechanism and architecture.
Monitoring data are disseminated through the network using the second
kind of component, \emph{synchronization agents}, also situated at
each node of the network. As their name implies, the agents are
responsible for synchronizing the data shared with peer agents. In
particular, the agent at a node passes local monitoring data, together
with data received from other agents, to the node's
peers. Importantly, the agent also uses the local store to maintain a
backup (i.e., history) of data received from other agents. In this
way, the agents collectively form a network-wide data dissemination
overlay.
The ultimate consumer of monitoring data is an \emph{analysis element}
supporting a network and/or service management task of some sort. We
assume that the analysis element is itself located at a (mobile or
fixed) node in the network. In fact, there may be several such
analysis elements located across the network, but we do not assume
that the location of an analysis element is a globally known
property. Notice, however, that since the dissemination overlay formed
by the synchronization agents includes the node hosting the analysis
element, a query to a local store by the analysis element is all that
is required to retrieve network-wide monitoring data.
\subsection{Gossip protocol}
\label{sec:gossip}
Synchronization agents employ a \emph{gossip protocol} to proactively
disseminate the monitoring data throughout the network. Unlike
previous gossip protocols, which are used mainly to compute and
propagate aggregate values~\cite{Birman:OSR:2007,Dimakis+:IEEE:2010},
our method is designed specifically to deal with time-series data
whose utility is both time bounded and time sensitive
(Section~\ref{sec:intro}).
At a high level, each agent repeatedly performs the following three
steps in cycles (Figure~\ref{fig:algorithm}):
\begin{enumerate}
\item \emph{Select peers.} The first step is to select the peers with
which to synchronize. In order to maintain the simplicity, yet
efficiency, of data synchronization under network mobility and
wireless link unreliability, we use a random selection of the peers
from candidates within a certain network distance.
\item \emph{Determine transfer datasets.} For each peer, a dataset to
be sent is calculated based on changes in the data since the last
successful synchronization with that peer. Two criteria are used to
determine what dataset to send to each peer: data
\emph{completeness} and data \emph{freshness}.
\item \emph{Transfer dataset.} The dataset for each peer is placed
into a space-efficient data structure and sent to that peer. Since
network links are unreliable, the agent records whether the transfer
was successful by updating a synchronization timestamp associated
with that peer.
\end{enumerate}
\noindent
We now describe each of these steps in detail.
\begin{figure}[t]
\centering
\includegraphics[width=0.6\columnwidth]{figures/algorithm}
\caption{The gossip cycle.}
\label{fig:algorithm}
\end{figure}
\para{Select peers.}
The goal of peer selection is to ensure an even dispersion of the data
across the network in a simple yet effective manner. We use a
\emph{random}, \emph{memory-less} peer selection process,
in which the peers are selected regardless of which and when
the data were previously sent~\cite{Jelasity+:ACM-TOCS:2007}, since it
is impractical to maintain a structured overlay topology across a
mobile network. Friedman~et~al.~\cite{Friedman+:SIGOPS:2007}
demonstrated that this strategy is effective for information
dissemination in~MANETs.
Using this process, each agent creates a candidate set of peers. The
candidate set is chosen based on a \emph{connectivity metric} such
that it contains only those with a high probability of successful data
transfer. It is well documented that as the hop distance between two
nodes increases in a MANET, the packet delivery ratio between those
nodes decreases
dramatically~\cite{DeCouto:2005:HPM:1150536.1150541}. Hence, we use
the hop distance metric, which can be conveniently obtained from the
local routing table, as the selection criterion. Specifically, a peer
is added to the candidate set if the hop distance is within a certain
threshold value (i.e., upper-limit parameter). A value of ``1'' is a \emph{neighbor} in the
network. Once the candidate set is established, a random subset of
nodes in the set is selected as the actual peers to be sent the
data. The number of selected peers is bounded by a configuration
parameter, as is the hop-distance upper-limit.
An illustrative example of the peer selection mechanism is shown in
Figure~\ref{fig:peer-selection}. In this example, the distance
threshold for nodes to be included in the candidate set is set to
1~hop. The upper bound of the number of peers is set to~2.
In the first cycle, shown in Figure~\ref{fig:peer-selection}a, the
candidate set will contain all nodes directly reachable from the
center node. The direct reachability is shown with the dashed circle.
The candidate set will thus contain nodes~\{1,2,3,4\}. From this
candidate set two peers \{1,2\} are randomly selected. In the next
gossip cycle, shown in Figure~\ref{fig:peer-selection}b, the positions
of the nodes have changed and now the candidate set contains nodes
\{2,3,7,5,8\} and the randomly selected peers are~\{3,8\}.
\begin{figure}[t]
\centering
\begin{tabular}{@{}c@{\hspace{1em}}c@{}}
\includegraphics[width=0.3\columnwidth]{figures/target_selection1}
&
\includegraphics[width=0.3\columnwidth]{figures/target_selection2}
\\
\textbf{(a)}
& \textbf{(b)}
\end{tabular}
\caption{Peer selections in two gossip cycles.}
\label{fig:peer-selection}
\end{figure}
\para{Determine transfer dataset.}
The data stored locally by a synchronization agent are conceptually
organized into a table (Figure~\ref{fig:dataset-calculation}), with
columns for each time series. The rows contain values for each time
slot in the series. Once the peers are selected at the beginning of
the gossip cycle, each agent determines the dataset to
transfer to each peer individually. The goal of this process is to
propagate the data as efficiently as possible, as well as to keep the
data fresh. To do this, for each peer $i$, and for each time series
$d$, the source agent maintains (and updates) the most recent time
slot number, denoted by $t_s(d,i)$, in $d$ that was successfully
transferred to $i$. It then uses the following criteria to determine
which time slots of each time series to send to each peer:
\begin{enumerate}[leftmargin=*,label=(\alph*)]
\item Only the \emph{incremental changes} in each time series since
the last successful synchronization to the peer are sent, i.e., only
data in time slots $t > t_s(d,i)$ in time series~$d$ are sent to
peer~$i$.
\item Time series data have an \emph{aging limit}, i.e., only the
values in time slots no older than a maximum limit $T$ are
transferred. This aging limit is a system configuration parameter,
ensuring that obsolete data are not included in the synchronization
process. Note, however, that eliminating data from a synchronization
does not necessarily mean those data are dropped completely from the
harvest, since agents could have received those data from other
agents via other paths in the synchronization overlay.
\item To reduce communication overhead, certain empty time slots,
which indicate the absence of monitoring data, are not transferred
to peer~$i$. In particular, those that would appear at the beginning
or end of the dataset are dropped, while those appearing between
values are kept. We do this to simplify how a time series can be
reconstructed at a peer. Of course, more sophisticated
\emph{compression} techniques, both lossy and loss-less, could be
considered, depending on how the monitoring data are to be used.
\end{enumerate}
\noindent
To enable the correct reconstruction of a time series, two additional
pieces of meta information are also transferred along with the values
of each time series: (i)~unique \emph{identifying information} for
each time series, defined and provided by the monitors that generate
the time series, and (ii)~the \emph{timestamp} of the last (newest)
time slot included in the dataset.
\begin{figure}[t]
\centering
\includegraphics[width=0.6\columnwidth]{figures/dataset3}
\caption{Selection of time slots to include in a transfer dataset,
sent from the agent to the peer. The transfer dataset (light shade)
is extracted from the original data locally stored on the node of
the source agent.}
\label{fig:dataset-calculation}
\end{figure}
Figure~\ref{fig:dataset-calculation} provides an example of the
time-slot selection process at a given node as it attempts to
synchronize with a peer. The figure shows five time
series,~D1--D5. The symbol ``X'' represents the presence of some data
value in a slot, the dark shading indicates time slots that were
successfully sent to the peer in previous gossip cycles, and the
timings of the current and previous gossip cycles are shown as
horizontal lines.
The light shading indicates the time slots selected for transfer in
the current gossip cycle according to the criteria described
above. For time series D1, three time slots will be transferred; the
first and last are included because they contain some values, with an
empty time slot in between. For~D2, five time slots will be
transferred, but in this case some slots older than the last gossip
cycle are also included. These older slots were either received from
some other node after the last gossip cycle or were not successfully
transferred in a previous cycle.
Notice that the trailing (i.e., the most recent) empty slot is not
included in the transfer dataset. Similarly only two time slots are
included in the transfer dataset for D3, for which only those time
slots older than the last gossip cycle are not empty. Finally, no time
slot will be transferred for either D4 or D5: no new value is present
in D4 since the last successful synchronization, and the time slot of
new values in D5, while it was not previously transferred, is older
than the age limit.
\para{Transfer dataset.}
After the peer and transfer dataset have been determined, the dataset
is sent to that peer. However, since the network links can be
unreliable, the agent waits some predetermined timeout period for
confirmation of successful receipt. Only once receipt is confirmed
does the agent advance the timestamp of the last successfully
synchronized time slot for that peer.
If, due to a network or other failure, the dataset transfer cannot be
completed or confirmed, the agent will abandon the synchronization
process with that peer in the current cycle. The confirmation timeout
period is a system parameter, obviously no longer than the time between
two successive gossip cycles, as this may lead to a race
condition in which the agent repeatedly attempts to send data that
might already have been received.
Note that the length of the time slots used by the gossip protocol may
not correspond to the length of the time slots used by the monitors.
Thus, the data maintained in the time slots by the monitors may be
aggregated and/or split in order to fit into the transfer time slots.
The aggregation of the data in the time slots used for transfer impacts
the overhead imposed on the network. This impact is explored in
Section~\ref{sec:results}.
Upon receipt of a dataset, an agent will reconstruct the transferred
portion of the time series (including any empty slots) and append it to
the time-series data in its local store. The data in the local store
are retained for a limited period of time, with obsolete data pruned
regularly to avoid extraneous use of resources.
|
\section{Introduction}
Language models (LMs) play a crucial role in many speech and language processing tasks such as speech recognition and machine translation. The current state of the art are recurrent neural network (RNN) based LMs \cite{Mikolov10}, and more specifically long short-term memory (LSTM)~\cite{lstm} LMs~\cite{lstm_lm,Joz}.
This work focuses on extending LSTM LMs with a cache~\cite{cache}, which gives the LM two main advantages: firstly, the cache is a simple and computationally cheap manner to adapt the LM to the current topic since it increases the probabilities of words recently seen. Secondly, Kuhn and De Mori~\cite{cache} first proposed a cache model as a means to overcome one of the limitations of n-gram LMs, namely that they are not capable of modeling short-term patterns in word use.
One can argue that LSTMs are better at this because they should be able to remember words for a longer time. However, the effect of a single word might be obscured after a certain amount of time steps since with each time step, a complete update of the cell state and hidden state is done. Explicitly adding previously seen words to a cache and combining the cache and LSTM probabilities can counteract this effect.
Grave et al.~\cite{Grave} recently proposed a \textit{continuous} cache (henceforth referred to as \textit{neural} cache, as opposed to Kuhn and De Mori's \textit{regular} cache~\cite{cache}), that assigns cache probabilities based on the similarity between the current hidden state and the hidden state associated with the word in the cache. As a result, words previously seen in a similar context get a higher cache probability than words seen in a different context. The advantages of the neural cache for neural LMs have also been demonstrated by Merity et al.~\cite{Merity} and Khandelwal et al.~\cite{Khandelwal}, but they use only a neural cache and do not compare with a regular cache, and only evaluate on textual data.
The manner in which the cache probability is combined with the LSTM LM probability is a crucial part in cache LMs. Grave et al.~\cite{Grave} mention linear interpolation and another method called global normalization, but the latter approach does not outperform linear interpolation. One can argue that linear interpolation -- assigning the same weight to the cache probability regardless of the type of word -- is suboptimal since the cache probability will be more helpful for content words whereas the LM probability should receive more weight for function words.
Kuhn and De Mori~\cite{cache} train separate interpolation weights and use separate caches for each part-of-speech (POS), combining the cache model with a class-based n-gram model in which the classes correspond to POS classes. However, we believe that keeping a separate cache for certain classes such as determiners and prepositions does not make much sense because for those classes, the overall statistics captured by the LM should be much more meaningful. For example, having seen the word \textit{in} multiple times before does not increase the probability of seeing it again. On the other hand, certain constructions such as verbs requiring this preposition (e.g. \textit{believe, specialize}) or sentences in which adverbial phrases indicating time or space are plausible (e.g. \textit{On the day of its release \textbf{in} Japan, Work began on the Tower Building \textbf{in} 1840}\footnote{Extracts from WikiText~\cite{Wiki}.}), do increase the probability of seeing \textit{in}, and such dependencies should be captured by the LM.
Another disadvantage of Kuhn and De Mori's approach is that it needs LM training data that has been tagged for POS.
The goal of this paper is two-fold: 1) we propose to weigh the cache model with a measure that is solely based on the frequency statistics of the training data and which hence does not need training data augmented with POS: the \textit{information weight} (IW) of a word (see section~\ref{interp} for a definition). This measure can be automatically calculated, and we investigate whether it can give further improvements for a cache model.
The IW can be used in \textit{information-weighted interpolation} and/or for an \textit{information-weighted selective} cache, to which only words with an IW larger than a certain threshold are added.
Secondly, as far as we know no comparison between regular and neural cache models for automatic speech recognition (ASR), an application for which cache models typically give good improvements, has been made. Thus, 2) the second goal of this paper is to investigate what advantage a (information-weighted) neural cache can have for ASR.
Typically, neural LMs for ASR are used in a multi-pass approach: in a first pass, the search space is narrowed down with the help of a simple (usually n-gram or FST) LM, producing a set of hypotheses in the format of a lattice or an N-best list. We evaluate our models by rescoring N-best lists.
In the remainder of this paper, we will first define the different cache models and different manners of interpolation (section~\ref{cache}) and validate the proposed methods experimentally (section~\ref{exp}). We end with a conclusion and an outlook to future work (section~\ref{concl}).
\section{Cache language models}
\label{cache}
\subsection{Regular cache}
\label{sec:reg-cache}
A regular cache model keeps track of the words previously seen and computes the probability of the next word being $w_t$ as follows:
\begin{equation}
\label{reg-cache}
P_{c}(w_{t}~|~w_{t-|C|-1}\ldots~w_{t-1}) = \frac{\sum\limits_{j = t-|C|-1}^{t-1} I_{\{w_j=w_t\}}}{|C|}
\end{equation}
where $C$ is the cache and $|C|$ its size, $t$ is the current time step and $I$ the indicator function, which is 1 if $w_j = w_t$ and 0 otherwise. An exponential decay can be used to give more weight to words seen more recently: $I_{\{w_j = w_t\}} e^{-\alpha~(t-j)}$, with $\alpha$ the decay rate.
\subsection{Neural cache}
The neural cache model proposed by Grave et al.~\cite{Grave} calculates the cache probability as follows:
\begin{equation}
\begin{aligned}
\label{neur-cache}
\begin{split}
P_{c}(w_{t}~&|~(w_{t-|C|-1},h_{t-|C|-2})\ldots~(w_{t-1},h_{t-2})) = \\
&\frac{\sum\limits_{j=t-|C|-1}^{t-1} I_{\{w_{j} = w_t\}} e^{(\theta~h_t^T~h_{j-1})}}{\sum\limits_{w_i \in C}~\sum\limits_{j=t-|C|-1}^{t-1} I_{\{w_{j} = w_i\}} e^{(\theta~h_i^T~h_{j-1})}}
\end{split}
\end{aligned}
\end{equation}
In the equation above, $h_{j-1}$ is the hidden state of the LSTM cell, which corresponds to the cell's output that is used to predict the next word $w_j$. In other words, the cache $h_c$ contains pairs of hidden states and corresponding target words. $\theta$ is a hyperparameter that controls the flatness of the distribution -- the lower $\theta$, the flatter the distribution of the cache probabilities. Given that $h_t^T~h_{j-1}$ is in the range (-1,1), the minimum value of $e^{(\theta~h_t^T~h_{j-1})}$ will always be $< 1$ and the maximum value $> 1$ (for $\theta > 0$), as opposed to a regular cache, where the value in the denominator is always 1.
If a word is in the cache, its cache probability is calculated based on the inner product between the current hidden state of the LSTM $h_t$ and the hidden state stored in the cache $h_{j-1}$. Hence, the probability is not solely based on whether the word has been seen before or not -- as happens in regular cache models --, but also based on the similarity between the current context and the context in which the cache word occurred, since we assume that the hidden state of the LSTM captures the previous context of the word. For example, in a Wikipedia article about the Chinese poet \textit{Du Fu}, we see that the neural cache assigns a much higher probability (compared to the unigram probability of the regular cache) to \textit{Fu} if \textit{Du} is the input word, since these two words occur together very frequently.
\subsection{Information-weighted interpolation of probabilities}
\label{interp}
Grave et al.~\cite{Grave} describe two ways of combining the cache probability with the standard probabilities output by the network, of which linear interpolation performs slightly better.
However, knowing that a word has been seen before might be more meaningful for certain words than for others.
Hence, we propose to exploit this fact by assigning a larger interpolation weight to words with a large content value than to words with a low content value, such as function words. The information weight $\lambda_i$ of a word is defined as follows:
\begin{equation}
\label{weight}
\lambda_i = 1 + \frac{\sum_{j=1}^{N}~P(j|i)~\log(P(j|i))}{\log(N)}
\end{equation}
where $P(j|i)$ is the probability of document $j$ given that word $i$ occurs in it, and is equal to the frequency of word $i$ in document $j$ divided by the frequency of word $i$ in the whole corpus. $N$ is the number of documents in the corpus; for our experiments we split the corpus in documents of an equal amount of sentences. Equation~\ref{weight} corresponds to 1 -- normalized entropy, hence giving higher importance to words that are sparsely distributed across the corpus. This measure is among others used by Dumais~\cite{dumais} and Bellegarda~\cite{Bellegarda} as a global weighting for the Latent Semantic Analysis word-document matrix. The values for $\lambda$ range between 0, for words that are uniformly distributed among the corpus, and 1, for words that appear in only 1 document. For example, \textit{bilingual} has a weight of $\lambda = 0.910$ whereas \textit{the} has a weight of $\lambda = 0.019$ for the WikiText-2 corpus~\cite{Wiki}.
The final probability for word $i$ is obtained by information-weighted linear interpolation:
\begin{equation}
\label{info-linear}
\begin{split}
&P(w_t~|~history) = \\
&\frac{(1 - \gamma~\lambda_t) P_{LM}(w_t~|~h_t) + (\gamma~\lambda_t) P_{c}(w_t~|~h_c)}{\sum\limits_{w \in V} (1 - \gamma~\lambda_t) P_{LM}(w_t~|~h_t) + (\gamma~\lambda_t) P_{c}(w_t~|~h_c)}
\end{split}
\end{equation}
where $\gamma$ is a hyperparameter that scales the information weights and needs to be set empirically. We only consider values smaller than or equal to 0.5, such that the cache probability will never be assigned a higher weight than the LM probability. Thus, the LM probabilities have more weight for frequent words, while for topical words a combination of LM and cache probabilities is used, which will lead to increased probabilities for topical words in the cache.
\subsection{Information-weighted selective cache}
\label{select}
Another manner to use the information weight of a word is to use it as a hard threshold: only words with an information weight greater than or equal to a certain threshold $\phi$ are added to the cache. This approach has the advantage that no space is `wasted' on words with low content value. It effectively enlarges the scope of the cache. A selective cache can be used in combination both with a regular and neural cache model, and with linear and information-weighted interpolation.
\begin{table*}[t]
\centering
\begin{tabular}{llll|c|c|c|c|c||c|c|c|c|c}
\hline
&\multirow{2}{*}{}&&&\multicolumn{5}{c||}{\textbf{Regular cache}}&\multicolumn{5}{c}{\textbf{Neural cache}}\\
\hline
\textbf{Model}&\textbf{Size}&\textbf{Interp.}&\textbf{IW select.}&\textbf{Valid}&\textbf{Test}&$\lambda~/~\gamma$&$\phi$&$\alpha$&\textbf{Valid}&\textbf{Test}&$\lambda~/~\gamma$&$\phi$&$\theta$ \\
\hline
\hline
Baseline~\cite{Grave}&-&-&-&-&\cellcolor{Gray}99.3&-&-&-&-&\cellcolor{Gray}99.3&-&-&- \\
Cache~\cite{Grave}&100&linear&-&-&\cellcolor{Gray}-&-&-&-&-&\cellcolor{Gray}81.6&-&-&- \\
\hline
Our baseline&-&-&-&102.5&\cellcolor{Gray}97.6&-&-&-&102.5&\cellcolor{Gray}97.6&-&-&- \\
Cache&100&linear&no&92.6&\cellcolor{Gray}87.9&0.10&-&-&85.6&\cellcolor{Gray}81.1&0.10&-&0.3 \\
Cache&100&IW&no&91.9&\cellcolor{Gray}87.1&0.20&-&-&86.1&\cellcolor{Gray}81.9&0.25&-&0.3 \\
\hline
Our baseline&-&-&-&103.0&\cellcolor{Gray}98.6&-&-&-&103.0&\cellcolor{Gray}98.6&-&-&- \\
Cache&100&linear&yes&90.7&\cellcolor{Gray}86.8&0.05&0.2&-&84.4&\cellcolor{Gray}81.3&0.05&0.2&0.3 \\
Cache&100&IW&yes&\underline{86.3}&\cellcolor{Gray}\underline{82.6}&0.25&0.4&-&\underline{79.6}&\cellcolor{Gray}\underline{76.7}&0.35&0.2&0.3 \\
\hline
\hline
Baseline~\cite{Grave}&-&-&-&-&\cellcolor{Gray}99.3&-&-&-&-&\cellcolor{Gray}99.3&-&-&- \\
Cache~\cite{Grave}&2000&linear&-&-&\cellcolor{Gray}-&-&-&-&-&\cellcolor{Gray}68.9&-&-&- \\
\hline
Our baseline&-&-&-&103.4&\cellcolor{Gray}98.1&-&-&-&103.4&\cellcolor{Gray}98.1&-&-&- \\
Cache&2000&linear&no&88.6&\cellcolor{Gray}83.8&0.10&-&0.005&73.4&\cellcolor{Gray}69.6&0.15&-&0.3 \\
Cache&2000&IW&no&87.4&\cellcolor{Gray}83.0&0.30&-&0.004&73.9&\cellcolor{Gray}70.9&0.35&-&0.3\\
\hline
Our baseline&-&-&-&103.3&\cellcolor{Gray}97.5&-&-&-&103.3&\cellcolor{Gray}97.5&-&-&- \\
Cache&2000&linear&yes&88.3&\cellcolor{Gray}84.0&0.05&0.2&0.008&76.6&\cellcolor{Gray}73.4&0.20&0.1&0.3 \\
Cache&2000&IW&yes&\textbf{82.4}&\cellcolor{Gray}\textbf{78.3}&0.35&0.2&0.006&\textbf{72.4}&\cellcolor{Gray}\textbf{66.2}&0.45&0.2&0.3 \\
\hline
\end{tabular}
\caption{Validation and test perplexities for LSTM LMs with a cache trained on WikiText-2, along with their optimal hyperparameters (if applicable). The column $\lambda~/~\gamma$ contains the optimal interpolation weight $\lambda$ for models with linear interpolation and the optimal scale for the IW $\gamma$ for models with IW interpolation. }
\label{tab:all-cache}
\vspace{-1em}
\end{table*}
\section{Experiments}
\label{exp}
\subsection{Setup}
For perplexity experiments, we train our LMs on WikiText-2~\cite{Wiki}. This dataset is also used by Grave et al.~\cite{Grave} and contains 2M words for training, 220k words for validation, 250k words for testing and has a vocabulary of 33k words.
All training and perplexity experiments are implemented with TensorFlow~\cite{tf}.
The LSTM LMs trained on WikiText have 1 layer of 512 LSTM cells, randomly initialized with a uniform distribution between -0.05 and 0.05. We use mini-batches of 20 samples and train the LSTMs with backpropagation through time (35 steps) and 50\% dropout~\cite{dropout}. The norm of the gradients is clipped at 5. We train with stochastic gradient descent: during the first 6 epochs, the learning rate is 1, after which we apply an exponential decay: the learning rate for epoch $i$, $\eta_{i}$, is equal to $\alpha~\eta_{i-1}$, with $\alpha$ the learning rate decay which is set to 0.8. After each epoch, we check how often the validation perplexity has not improved with respect to the previous epochs: if it has not improved for 3 times, we stop training. If the validation perplexity keeps on improving, we stop training anyway after 39 epochs. We train on discourse level, which means that we do not reset the LSTM state at sentence boundaries and predict the first word of the next sentence from the end-of-sentence token (for more information, see Verwimp et al.~\cite{tf-lm}). This model gives a baseline perplexity close the the one reported by Grave et al.~\cite{Grave}.
The hyperparameters specific to the cache model are optimized on the validation set. We provide all values for the hyperparameters together with the results, since it has been shown that these have an important influence~\cite{melis}.
We use Kaldi~\cite{kaldi} for the ASR experiments. We follow the standard \textit{nnet2} recipe for Wall Street Journal (WSJ), generating 1000-best lists for rescoring. The first pass of the recognition makes use of a 3-gram LM with a vocabulary of 120k words. The training data for the LSTM LMs consists of 37M words with a vocabulary of 40k. It contains years 87--89 with non-verbalized punctuation. The new LM probability is calculated as follows:
\begin{equation}
\label{p-new}
\begin{split}
P_{new}(W) &= (1 - \lambda_{LSTM})*P_{ngram}(W) \\
&+ \lambda_{LSTM}*P_{LSTM+C}(W)
\end{split}
\end{equation}
where $P_{ngram}$ is the first-pass LM probability and $P_{LSTM+C}$ is equal to the probability in equation~\ref{info-linear}. A scaled version of $P_{new}(W)$ ($\beta$ = scaling factor) is combined with the score of the acoustic model $P(O|W)$:
\begin{equation}
\label{rescore}
\begin{split}
\log P(W) = \log P(O|W) + \beta~\log (P_{new}(W)) + L ~\omega
\end{split}
\end{equation}
$\omega$ is the word insertion penalty and $L$ the number of words in the hypothesis.
For every model type, hyperparameters such as word insertion penalty $\omega$, LM scale $\beta$, LSTM LM weight $\lambda_{LSTM}$ are optimized on the validation set.
The LSTM LMs used for rescoring are trained with TensorFlow~\cite{tf} with the same hyperparameters as the WikiText models, except that we train on sentence level instead of discourse level.
We use the bootstrap method of Bisani and Ney~\cite{bisani} for significance analysis of the WER results, which is implemented as Kaldi's \textit{compute-wer-bootci} script. This method generates (by default 10k) random bootstrap samples of the data and calculates the percentage of times the bootstrap sample from system B improves over the bootstrap sample from system A: the result is a `probability of improvement' (poi) of system B over system A.
Since the length of the cache is usually longer than the length of the hypotheses in a single N-best list, we would like to be able to transfer the cache across lists. We keep track of the probabilities and caches for every hypothesis in the list, and initialize the cache of the next N-best list with the cache of the most likely hypothesis from the previous list.
\subsection{Perplexity results}
We report the perplexity of the LSTM LMs in table~\ref{tab:all-cache}. Due to implementation requirements, we have to train 4 different models: one with a non-selective cache of size 100, one with a selective cache of size 100 and the same options for a cache size of 2000. The baseline perplexities for LSTMs without cache of these four models do not differ much, but we report them all for the sake of completeness (`Our baseline' in the table). All perplexity results shown below the same baseline can thus be attributed solely to the cache component since they are obtained with the same trained model.
Firstly, we combine the LSTM with a regular cache model with linear interpolation (left part of the table) and see that this simple combination already gives us improvements: 11.4\%/11.6\% relatively on the validation/test set for a cache size of 2000 (with exponential decay). This result confirms our hypothesis, namely that an LSTM in itself is not sufficient in remembering the relevant words in the previous context. Using information-weighted (\textit{IW}) interpolation or IW selective cache usually gives additional improvements. The two IW-based techniques seem to give complimentary improvements, because the combination yields the best perplexity result (relative improvement of 20.2\%/19.7\%) for the regular cache model.
The results for LSTM LMs with a neural cache are reported in the right part of the table.
We observe that, similarly to the regular cache, the combination of both information-based extensions yields perplexity improvements, also with respect to the results reported by Grave et al.~\cite{Grave}. Using only one of the two IW-based techniques does not result in better perplexity results.
Additionally, we observe that the neural cache consistently outperforms the regular cache.
The best result is obtained with a neural selective cache of size 2000 with information-weighted interpolation: a relative improvement of 29.9\%/32.1\% (validation/test) with respect to our own baseline.
\subsection{Analysis}
Taking a closer look at the optimal hyperparameters, we firstly observe that $\gamma$ is larger for the neural cache than for the regular cache. For the best model, $\gamma = 0.45$, resulting in weights close to 0.5 for words with a large IW $\lambda$, which indicates a large reliance on the cache for content words.
Secondly, the optimal threshold $\phi$ for most IW selective cache models is quite low, 0.2. This threshold essentially excludes only 222 word types from the cache, that albeit form more than 50\% of the test set (135k out of 240k).
In figure~\ref{fig:data_distr}, we plot how often the information-thresholded cache assigns a larger interpolated probability to the target word than the original LM probability.
Firstly, since $\phi$ is 0.2 for the selective model, very frequent low-content words (with IW $< 0.2$) will always have a lower interpolated probability than the softmax probability because the cache probability is 0. However, for IWs larger than 0.2, the selective model is more likely to result in a better interpolated probability, and that probability increases if the IW itself increases.
\begin{figure}[ht]
\centering
\resizebox{7.5cm}{7cm}{\input{data_distr.tex}}
\caption{The proportion of times that the interpolated probability is higher than or equal to the baseline LM probability with respect to several ranges of IWs, for the best neural cache models with IW interpolation and with the combination of IW selective cache and IW interpolation.}
\label{fig:data_distr}
\vspace{-1em}
\end{figure}
As a result, if we look at the predictions over all IWs, the non-selective model outperforms the LM probability in 31.9\% of the cases and the selective model only in 25.1\% of the cases. But for IWs $> 0.2$, the selective models outperforms the baseline in 43.5\% of the cases, and the non-selective one only in 33.9\% of the cases.
When the interpolated probability is higher than the softmax probability, the relative improvement of the interpolated probability with respect to the softmax probability, averaged over the test set, is 55.4\% for the non-selective model and 54.3\% for the selective model.
For IWs $> 0.2$, the relative improvement is higher for the selective model (67.8\%) than for the non-selective one (65.8\%), although the difference is small.
Thus, the cache model combining the IW interpolation and the IW selection gives worse estimates for very frequent words than the baseline model and the cache with only IW interpolation since they have a cache probability of 0, but better estimates for topical words, resulting in an overall perplexity improvement.
\subsection{Speech recognition results}
In table~\ref{tab:asr-wsj}, we report the perplexity results on the WSJ validation set and the validation and test WER after rescoring with LSTM LMs. Similarly as for WikiText, we report the results for two different baseline models, a selective one and a non-selective one. Notice that all results within the upper part and respectively the lower part of the table are obtained with the same trained model. Thus, differences in performance can be attributed to the cache parts solely and not to different initialization settings that might lead to better convergence of the model during training.
As a sanity check, we tested whether the differences in WER between the two baseline models were significant: according to the bootstrap method~\cite{bisani}, the probability of the first baseline improving over the second baseline is 39.97.5\%/73.57\% (validation/test set). Thus, we conclude that the difference between the baselines is not significant.
We only test models with a cache size of 100, since preliminary perplexity results with a larger cache size did not give substantial improvements with respect to a cache of 100 for this dataset.
Firstly, we observe that the differences in perplexity are much smaller than for WikiText: this shows that on the WSJ dataset, much less can be gained by using a cache. This is not surprising given the fact that WikiText is specifically designed to contain many long-term dependencies~\cite{Wiki}. Secondly, the neural cache is not consistently better than the regular cache, as opposed to WikiText. The combination of a neural cache with linear interpolation gives the best results, both for a non-selective and a selective model, but the improvement is only 4.6\%/4.1\% (non-selective/selective) relative to the baseline.
Overall, the differences in perplexity between the cache models are quite small, but notice that this dataset is quite small (8k words, compared to 210k/240k words for the validation/test set of WikiText) and hence we should be careful in drawing strong conclusions based on these results.
The cache models give WER improvements between 0.2 and 0.32 absolute for rescoring the 1000-best lists of the validation set. Between brackets we indicate the probability of improvement with respect to the baseline: all results except the neural cache with linear interpolation are significant with a confidence larger than 95\%, but the best results are still quite far from the oracle WER of 2.18. We observe that the best model in terms of perplexity does not yield the best WER results, and that the WER differences between the cache models are not significant.
\begin{table*}[t]
\centering
\begin{tabular}{l|c|c|c|c|c|c|c|c}
\hline
\textbf{Model}&\textbf{Valid PPL}&\textbf{Valid WER (poi)}&\textbf{Test WER (poi)}&$\lambda$ / $\gamma$&$\phi$&$\lambda_{LSTM}$&$\beta$&$\omega$ \\
\hline
baseline&176.9&5.61&2.55&-&-&0.5&18&0 \\
regular linear&170.5&\textbf{5.36} (96.55)&2.34 (93.95)&0.05&-&0.5&15&-0.5 \\
regular IW&171.2&5.41 (96.47)&\textbf{2.25} (\textbf{99.34})&0.4&-&0.5&13&-0.5 \\
neural linear&\textbf{168.8}&5.38 (94.63)&2.43 (80.66)&0.05&-&0.5&15&-0.5 \\
neural IW&170.3&5.38 (\textbf{97.21})&2.39 (89.87)&0.7&-&0.4&13&0 \\
\hline
baseline&175.1&5.59&2.64&-&-&0.5&18&0\\
select. reg. linear&170.2&5.32 (99.10)&2.68 (33.68)&0.05&0.05&0.3&13&0\\
select. reg. IW&168.1&\textbf{5.27} (\textbf{99.8}1)&\textbf{2.61} (\textbf{58.75})&0.1&0.05&0.3&13&0\\
select. neur. linear&\textbf{167.9}&5.32 (99.10)&2.64 (46.33)&0.05&0.05&0.3&13&0 \\
select. neur. IW&168.8&5.31 (99.57)&2.63 (52.32)&0.1&0.005&0.3&13&0 \\
\hline
\end{tabular}
\caption{Perplexities on the validation set of WSJ and WERs after rescoring the 1000-best lists of the validation and test set. All cache models have a cache size of 100. `poi' = probability of improvement~\cite{bisani} of the cache model with respect to its baseline. The 5 last columns contain the hyperparameter values optimized on the validation set by rescoring, which are also used for rescoring the test set. The column $\lambda~/~\gamma$ contains the optimal $\lambda$ for models with linear interpolation and the optimal $\gamma$ for models with IW interpolation. We do not include the decay values since exponential decay did not improve the results. Oracle WER for the 1000-best lists is 2.18 (validation) and 1.21 (test). }
\label{tab:asr-wsj}
\vspace{-1em}
\end{table*}
On the test set, the non-selective models achieve WER reductions between 0.12 and 0.3 absolute; notice that the largest possible reduction is 1.43 (oracle WER 1.21). However, only the regular cache with IW interpolation achieves a significant WER reduction with respect to the baseline.
This model also significantly improves the results for the neural models (both with a confidence $>$ 98\%), but not significantly the result of the regular cache with linear interpolation (poi of 89.63\%). The regular cache with linear interpolation significantly improves its neural counterpart (poi 98.36\%), but not the neural cache with IW interpolation (poi 72.26\%). Given that there are no significant differences between the cache models on the validation set on which the hyperparameters are optimized, this suggests that the regular cache models are more robust against hyperparameter settings that are possibly not optimal for the test set.
The selective models on the other hand, show no or only minor improvement with respect to their baseline model on the test set. The best model only gives 0.03 absolute reduction in WER. We investigated several possible explanations for this large discrepancy between the results of the selective models on the validation and on the test set.
Firstly, it is possible that the selective models are much more sensitive to hyperparameter settings, but tuning the hyperparameters on the test set as an oracle experiment does not improve the results.
A second possible explanation is that the selective models only work well on data with a specific structure/specific patterns, and both WikiText and the validation set have such structure while the test set does not. However, if we calculate perplexity of the test set, the selective neural cache with IW interpolation is among the best models (5.7\% relative improvement). On the other hand, we observe with this dataset that perplexity results do not always extrapolate nicely to WER results. This might be because the perplexity is calculated on the references only, while the WER is the result of rescoring many hypotheses, of which some might be more probable than the reference according to the LM. If the wrong hypothesis is selected, the wrong cache is transferred across N-best lists. Since the selective models have a larger scope because no space is wasted on function words, the wrong words are effectively longer in the cache. The effect of wrong words in the cache might be worse for certain datasets (e.g. the WSJ test set) than for others (e.g. the WSJ validation set).
A very crude measure that provides some evidence for this explanation of the validation versus test set discrepancy, is counting the number of times that a wrongly selected hypothesis is followed by another wrongly selected hypothesis. We observe that for all non-selective models, this number is lower for the cache models than for the baseline model, while for all selective models, this number is higher for the cache models than for the baseline model, providing plausible evidence for the fact that errors are propagated longer for selective models.
This is even the case for the validation set, but on this dataset, the gains that the selective model provides are larger than the losses, such that the overall WER is lower for the cache models than for the baseline model.
\section{Conclusion and future work}
\label{concl}
In this work, we present an extensive evaluation and comparison of regular and neural cache LMs, in terms of perplexity and WER. Additionally, we propose two extensions based on the information weight/content value of words that can be applied to any cache model: IW interpolation of cache and LM probabilities and an IW selective cache.
On the WikiText benchmark, we obtain large improvements in perplexity, the largest one obtained by the combination of the two information-weighted extensions proposed in this paper.
Additionally, we observe that neural cache models consistently outperform regular cache models on this dataset.
We also test the cache models on ASR by rescoring 1000-best lists of WSJ.
We observe significant improvements with respect to a baseline LM on the validation set, while on the test set the regular cache with IW interpolation was the only model giving a significant improvement. As opposed to WikiText, neural cache models are not consistently better than regular cache models, and the combination of the two information-weighted extensions is not the best option. The neural cache seems to be more sensitive to hyperparameter settings, since the regular cache models are significantly better than the neural cache models on the test set, but not on the validation set. The selective models provide significant gains on the validation set, but not on the test set: we hypothesize that those models are more sensitive to the selection of wrong hypotheses in previous segments, because the wrong words stay longer in the cache.
In the future, we would like to test our models on ASR tasks that might give more gains, and test a `smarter' manner of transferring the cache across N-best lists, e.g.~a weighted combination of all hypotheses from the previous list.
|
\section{Introduction}
\label{sec:intro}
Understanding the underlying intent of search queries is a crucial component in virtually every semantic search system, either being a web search engine, a chatbot, or an e-commerce website.
It has been long recognized that knowledge bases such as DBpedia, Freebase, and YAGO are rich sources of information for interpreting and understanding queries.
A large body of efforts in this area is focused on recognizing mentioned entities in the queries and linking them to the corresponding entities in a knowledge base, the so-called task of entity linking in queries~\citep{hasibi:2015:ELQ, Blanco:2015:FSE}.
In this paper, we aim to further the understanding of queries by identifying their entity attributes from a knowledge base; e.g., identifying the attribute \emph{spouse} from DBpedia for the query ``the wife of Lincoln.''
Extracting entity attributes of queries is beneficial for answering the queries in tasks such as question answering and entity retrieval.
It has been shown that joint entity linking and attribute identification of queries improves question answering over knowledge bases~\citep{Xu:2016:QAF, Yao:2014:IES}.
Similarly, entity retrieval approaches can benefit from entity attribute identification by having a focused selection of entity attributes~\citep{Hasibi:2016:ELR} and using them to build fielded representation of entities~\citep{Hasibi:2017:DTC}.
Entity attribute identification can be also employed in the e-commerce websites to improve search results and boost sites' advertising profits and recommendation quality. Consider, for example, the query ``nike shoes size 38'', where the attribute \emph{size} can be used to filter out irrelevant products or advertising similar products from other brands.
Motivated by the above reasons, we set out to focus on identifying entity attributes that help answering a query. We note that this is a highly non-trivial task, mainly due to vocabulary mismatch between query terms and the entity attribute(s) pointed by the query. Take for example the query ``the father of integrated circuits'', which refers to the attribute \emph{inventor}, rather than \emph{father} or \emph{parent}.
We frame the entity identification task as a ranking problem and propose a set of methods to address it. Our first model is based on Markov Random Field (MRF) and incorporates entity annotations of queries as a bridge to rank entity attributes. We further employ a learning to rank approach combining various attribute similarity scores and show significant improvements with respect to our best baseline. We evaluate our results on a purpose-built test collection based on the DBpedia-Entity v2 collection~\citep{Hasibi:2017:DTC} for entity retrieval.
To summarize, the contributions of this work are as follows:
\begin{itemize}
\item We introduce and formulate the task of ``entity attribute identification.''
\item We propose a set of methods (an MRF-based and a learning to rank model) to address the entity attribute identification task, and provide insights into the influence of different contributors of our models.
\item In order to evaluate the task and foster research in this area, we build a test collection, consisting of graded scores for a diverse set of entity oriented queries. The dataset is human-annotated and is made publicly available at \textcolor{blue}{\url{http://tiny.cc/eai}}
\end{itemize}
\section{Related Work}
\label{sec:related}
\subsection{Problem Definition}
\label{sec:en_att:def}
\begin{mydef}[\textbf{Entity attribute identification}]
Given an entity-bearing query, entity attribute identification is defined as the task of returning a ranked list of entity attributes, where the values of those attributes provide answers to the query or help finding the answers.
\end{mydef}
\begin{remark}
In this definition, we focus on entity-bearing queries; i.e., queries that refer to specific entities in a knowledge base. For example the query ``the wife of Lincoln,'' which can be linked to entity \textsc{Abraham Lincoln}. Entity linking in queries is a well studied task, and can be performed using publicly available entity linkers such as TAGME~\cite{tagme} and Nordlys~\cite{Hasibi:2017:NTE}.
\end{remark}
\begin{remark}
Each entity in knowledge base is represented by a list of pairs \texttt{$e = \{\langle a_1,v_1 \rangle, \langle a_2,v_2 \rangle, ..., \langle a_n,v_n \rangle \}$}, where $a_i$ is an attribute and $v_i$ is its associated value. For example, the entity \textsc{Abraham Lincoln} is represented as
\texttt{\{$\langle$spouse, Mary Todd Lincoln$\rangle$, $\langle$death Place, Washington D.C.$\rangle$, ... \}}
\end{remark}
\begin{remark}
The ranked entity attributes belong to the entities that are linked to the query. For example \texttt{\{spouse\}} is the top-ranked attribute for the aforementioned query.
\end{remark}
Here, we relate this problem to the extensive body of work on attribute extraction on (semi-)structured text~\citep{Hoffmann:2010:LRE, Bing:2013:WEE, Zhong:2016:EAN} and highlight that our goal is to further machine-understanding of \emph{queries}, which are short, ambiguous pieces of text (unlike long documents).
Similar efforts have been performed in e-commerce to extract \emph{attribute values} of product titles~\citep{More:2016:AEP}, and further filter out search results based on the matching attribute values. The most similar task to ours is the NTCIR actionable knowledge graph generation (AKGG) task~\citep{Blanco:2017:ONA}, which aims at ranking attributes of a query that are relevant for performing users' actions. In our task, we consider a different (rather broader) context and identify entity attributes that are useful for finding relevant answers to the query. Consequently, the outcome can be incorporated in various other tasks such as entity retrieval and questions answering.
\subsection{MRF-based Model}
\label{sec:prob_model}
Our first model to address entity attribute identification task is based on Markov Random Field (MRF).
Here, our goal is to compute the relevance probability of an attribute $a$ to a given query $q$, which can be estimated by a set of joint probabilities between the attribute, query, and a linked entity to the query:
\begin{equation}
p(a|q) = \frac{p(a,q)}{p(q)} \overset{rank}{=} \sum_{e \in E} p(a,e,q).
\label{eq:first}
\end{equation}
In this equation, $E$ is the set of entities linked to the query $q$ and is obtained by an entity linker system.
In order to estimate $p(a,e,q)$, we follow the idea of~\citet{Metzler:2005:MRF} in using MRF for ad hoc retrieval tasks. MRF is a graphical model, which can be used for estimating joint probability of random variables described by an undirected graph $G$. In this graph, nodes indicate random variables and edges represent dependency between the nodes. The joint probability over variables of the graph $G$ is computed as:
\begin{equation}
\label{eq:mrf_start}
P_{\Lambda}(G) = \frac{1}{Z_\Lambda} \prod_{c \in C(G)} \psi (c; \Lambda),
\end{equation}
where $C(G)$ is the set of cliques in graph $G$ and $\psi (c; \Lambda) = exp[ \lambda_c f(c)]$ is a non-negative potential function, parametrized by the weight $\lambda_c$ and the feature function $f_c$. The parameter $Z_\Lambda$ is a normalization factor, which is generally ignored due to computational infeasibility. Ignoring $Z_\Lambda$ and taking logarithm of the right hand side of Eq. \ref{eq:mrf_start}, the joint probability of $P_{\Lambda}(G)$ is proportional to:
\begin{equation}
\label{eq:mrf2}
P_{\Lambda}(G) \propto \sum_{c \in C(G) } \log [\psi (c; \Lambda)] = \sum_{c \in C(G)} \lambda_c f(c).
\end{equation}
The graph underlying our model consists of independent query terms, an entity, and an attribute; see Figure~\ref{fig:graph}.
In this graph, three types of 2-cliques are defined: (i) cliques involving a query term and the attribute, (ii) a clique involving the entity and the attribute, and (iii) cliques involving a query term and the entity. The 3-cliques involving a query term, an attribute, and an entity are ignored due to computational complexity. Putting all these elements together, the probability $P(a|q)$ is proportional to:
\begin{equation}
\label{eq:ranking}
p(a|q) \overset{rank}{=} \sum_{e \in E} \left(
\lambda_1 \sum_{q_i \in q}{f_1 (q_i,a)} +
\lambda_2 f_2 (a,e) +
\lambda_3 \sum_{q_i \in q}{f_3 (q_i,e)}
\right),
\end{equation}
where the $\lambda$ parameters should meet the constraint of $\sum_{i=1}^{3} \lambda_i = 1$.
We now define the feature functions of our model. The first feature functions is defined as:
\begin{equation}
\label{eq:ff1}
f_1(q_i,a) = log [\frac{1}{|a|} \sum_{w \in a} 1-distance(\vec{q_i},\vec{w})],
\end{equation}
where $w$ is an attribute word and $distance (\vec{q_i},\vec{w})$ indicates the Euclidean distance between the vector representation of words $q_i$ and $w$. We obtain these vector representations form Word2Vec~\cite{word2vec} 300-dimensions vectors, trained on the Google news dataset. Using this feature function, our model is able to capture the semantic similarity between query and attribute terms; e.g. ``spouse'' and ``wife'' in Fig. \ref{fig:graph}.
The second feature function is computed by:
\begin{equation}
\label{eq:ff2}
f_2(a,e) = log [
\mu_1 \frac{|\{\langle t,v \rangle \in e | t=a \}|}{|\{\langle t,v \rangle \in e \}|} + (1-\mu_1) \frac{|\{\langle t,v \rangle \in \mathcal{E} | t=a \}|}{|\{\langle t,v \rangle \in \mathcal{E}\}|}
],
\end{equation}
where $\mu_1$ is the smoothing parameter. Here, $e$ is an entity represented by a set of attribute-value pairs $\langle t,v \rangle$, and $\mathcal{E}$ is the collection of all these pairs from all entities in the knowledge base.
The feature function $f_3(q_i,e)$ measures the similarity between an entity and a query term and is defined as:
\begin{align}
\label{eq:ff3}
f_3(q_i,e) = log [
\mu_2 \frac{|\{\langle t,v \rangle \in e | q_i \in terms(t) \vee q_i \in terms(v) \}|}{|\{\langle t,v \rangle \in e \}|} + \nonumber\\
(1-\mu_2) \frac{|\{\langle t,v \rangle \in \mathcal{E} | q_i \in terms(t) \vee q_i \in terms(v) \}|}{|\{\langle t,v \rangle \in \mathcal{E}\}|}
],
\end{align}
where $terms(.)$ returns a set of terms of a given text, and $\mu_2$ is a smoothing parameter.
\begin{figure}[t]
\vspace*{-0.7\baselineskip}
\includegraphics[width=.55\linewidth]{figs/graph}
\vspace*{-0.7\baselineskip}
\caption{MRF graph for the query ``the wife of U.S. president Lincoln.''}
\label{fig:graph}
\end{figure}
\subsection{Learning to Rank Model}
\label{sec:ltr_model}
In this section, we propose a Learning to Rank (LTR) approach for addressing the entity attribute identification task.
We employ seven features, described in table \ref{tab:features}, and train our learning to rank algorithm. Given the low-dimensional feature space and limited number of training instances, we use Coordinate Ascent (CA)~\citep{metzlerIR07} algorithm for our LTR approach.
The employed features are as follows. Features $f_1$, $f_2$, and $f_3$ capture entity linking probability, entity attribute similarity, and query attribute similarity (cf. Section \ref{sec:prob_model}). For features $f_4$--$f_7$, we partitioned the query terms into two disjoint sets. The first subset includes query terms which are linked to an entity (i.e., linked terms) and the second subset is the set of terms which are not linked to any entity (i.e., not-linked-terms). For example, in the query ``the wife of Lincoln'' linked to entity \textsc{Abraham Lincoln}, the set of linked and not linked terms are \{``Lincoln''\} and \{``the,''``wife,'' ``of''\}, respectively. We then compute the similarity between these terms and concatenation of an attribute-value pair, based on WordNet and Word2Vec~\cite{word2vec} vector representation of words.
\setlength{\abovetopsep}{0pt}
\setlength{\aboverulesep}{0pt}
\setlength{\belowrulesep}{0pt}
\setlength{\belowbottomsep}{0pt}
\begin{table}[t]
\centering
\caption{List of features used in LTR approach.}
\vspace*{-0.7\baselineskip}
\begin{tabular}{|@{~}l@{~}|l@{~}|}
\hline
Feature & Description \\
\midrule
$f_1$ & $\sum_{q_i \in q}{f_3 (q_i,e)}$ \\
$f_2$ & $f_2(a,e)$ \\
$f_3$ & $\sum_{q_i \in q}{f_1 (q_i,a)}$ \\
$f_4$ & WordNet similarity using linked terms of query $q$ \\
$f_5$ & Word2Vec similarity using linked terms of query $q$\\
$f_6$ & WordNet similarity using not linked terms of query $q$ \\
$f_7$ & Word2Vec similarity using not linked terms of query $q$\\
\hline
\end{tabular}%
\label{tab:features}%
\vspace*{-1\baselineskip}
\end{table}%
\section{Entity Attribute Identification}
\label{sec:en_att}
In this section, we formally define the problem of entity attribute identification and describe our proposed methods.
\input{entity_attr_03_01}
\input{entity_attr_03_02}
\input{entity_attr_03_03}
\section{Test Collection Creation}
\label{sec:coll}
In order to evaluate our proposed methods, we created a test collection for the entity attribute identification problem.
We used DBpedia 2015-10 as our knowledge base and built our test collection based on DBpedia-Entity v2 collection~\cite{Hasibi:2017:DTC}.
This dataset consists of 467 queries and their relevant entities from DBpedia 2015-10. Using DBpedia-Entity v2 collection, we generated a new test collection for the attribute identification task
Our test collection was generated in two steps.
In the first step, we identified all entities that could be linked to the query. To improve recall, we used the two publicly available entity linker systems: TAGME~\cite{tagme} and Nordlys~\cite{Hasibi:2017:NTE}.
For each entity $e$ linked to query $q$, all its attributes are obtained and added to the pool of candidate attributes if the value of the attribute is among relevant entities of the query $q$.
In the second step, three information retrieval students were asked to annotate query-attribute pairs. They were all trained about the concepts of entities and asked to grade the entity attributes based on the following definitions. These definitions are intentionally inline with the ones from the DBpedia-Entity collection to keep the consistency of datasets.
\begin{itemize}
\item \textbf{Highly relevant (2): } The attribute holds direct answer to the user's query. That is, the attribute should be put among the top results.
\item \textbf{Relevant (1): } The attribute can guide user to find the exact answer, but does not hold direct answer to the query. In other words, the attribute should not be placed among \emph{top} results.
\item \textbf{Irrelevant (0): } The attribute has no relation to the query and should not be considered as an answer.
\end{itemize}
The collection was annotated by three experts, and in case of disagreement the forth annotator was involved. We measured quality of the obtained labels by computing the inter-annotator agreement using Fleiss' Kappa.
Over all candidates, we got an average Kappa of 0.38, which is considered a fair agreement.
The final test collection includes 167 queries and their relevant attributes.
Query categories of our test collection are similar to ones from DBpedia-Entity collection. Table~\ref{tab:collection} summarizes the statistics of the collection.
\begin{table}[t]
\centering
\small
\vspace*{-0.7\baselineskip}
\caption{Query categories in our test collection, QLen indicates the average number of terms per query. $R_1$ and $R_2$ refer to the average number of relevant and highly relevant attributes per query, respectively.}
\vspace*{-0.7\baselineskip}
\begin{tabular}{|l|l|l|l|l|l|}
\toprule
Category & \#queries & QLen & Type & $R_1$ & $R_2$ \\
\midrule
INEX-LD & 31 & 4.74 & Keyword queries & 2.48 & 1.89 \\
QALD2 & 62 & 7.52 & NL questions & 2.03 & 2.31 \\
SemSearch\_ES & 40 & 2.53 & Named entities & 2.94 & 2.18 \\
ListSerach & 34 & 5.38 & List of entities & 2.38 & 2.38 \\
\midrule
Total & 167 & 5.04 & & 2.46 & 2.19 \\
\bottomrule
\end{tabular}%
\label{tab:collection}%
\vspace*{-1.4\baselineskip}
\end{table}%
\section{Results and Analysis}
\section{Baselines and Settings}
For our baseline methods, we ran BM25, Language Model (LM), and Mixture of Language Models (MLM)~\cite{mlm} on an index built based on DBpedia.
Each document in our index is identified by an entity-attribute pair. Considering the entity $e$ with $k$ attributes $\{a_1,a_2,...,a_k\}$, we create $k$ documents, each represented as $\langle e, a_i \rangle: V_i$, where $V_i$ indicates all values of attribute $a_i$ in entity $e$.
Following~\citep{Balog:2013:TCE}, we set the weights of MLM models to 0.2, and 0.8 for title and content fields (i.e., $a$ and $V$).
We ran BM25 with parameters $k_1 = 1.2$ and $b = 0.8$ and Dirichlet smoothing with $\mu = 2000$ for LM and MLM-tc models.
In the MRF-based model, we set the parameters $\lambda_1 = 0.6$, $\lambda_2 = 0.2$, $\lambda_3 = 0.2$, $\mu_1=0.5$, and $\mu_2=0.5$ (using parameter sweeps).
For LTR experiments, we used the CA implementation provided in the RankLib framework and set the number of random restarts to 3. We obtained the results using 5-fold cross validation, keeping attributes of the each query in the same fold.
We employed a two-tailed paired t-test ($\alpha = 0.05$) to measure statistical significance. Significant improvements over the best baseline model (i.e., MLM-tc) are marked with $^*$ in Table \ref{tab:results}.
\section{Results and Analysis}
Table \ref{tab:results} shows the comparison of the baseline and proposed methods. The NDCG@5, P@5, MRR, and MAP metrics are reported for all methods. The results show that the MRF-based model can significantly improve the baseline methods with respect to all metrics. This improvement can be explained by the fact that the baseline models rely only on exact matching of query and attribute terms, while the MRF-based model tries to score each attribute by considering three similarities: entity-query, entity-attribute, and query-attribute.
The second observation is that the proposed LTR model (i.e., LTR/CA) improves the MRF-based model. This is expected, as
the LTR method uses all the signals used by the MRF-based model (i.e., $f_1$, $f_2$, and $f_3$) as well as other features mentioned in Table~\ref{tab:features}. In addition, the LTR model uses an optimized combination of signals to rank attributes for a given query.
\begin{table}[t]
\vspace*{-0.7\baselineskip}
\centering
\caption{Comparison of baselines and proposed models for attribute identification task.}
\vspace*{-0.7\baselineskip}
\begin{tabular}{|l|cccc|}
\toprule
\textbf{Model} & NDCG@5 & P@5 & MRR & MAP \\
\midrule
BM25 & 0.0467 & 0.0369 & 0.0749 & 0.0503 \\
LM & 0.0527 & 0.0371 & 0.0898 & 0.0618 \\
MLM-tc & 0.0803 & 0.0479 & 0.1168 & 0.0847 \\
\midrule
MRF-based & 0.2844* & 0.1817*& 0.3618* & 0.2167* \\
LTR/CA & \textbf{0.3227*} & \textbf{0.2117*}& \textbf{0.3702*} & \textbf{0.3390*} \\
\bottomrule
\end{tabular}%
\label{tab:results}%
\end{table}%
Figure \ref{fig:cat} indicates the comparison of proposed models for different query categories in our test collection. We observe that the retrieval performance (with respect to NDCG@5) on INEX-LD and ListSearch categories is higher than others. This can be explained by the fact that most of these queries are short and seeking for entities with a direct relation to the mentioned entity in the query. QALD queries, however, are complex and involve further understanding using natural language processing techniques.
We analyze the discriminative power of the features by comparing the ranking performance of each feature in isolation; i.e., using a single feature as a ranker. Table \ref{tab:fi} shows the results.
The third column indicates the NDCG@5 difference between single feature models and the model trained on all features. According to this table, features $f_5$ (Word2Vec similarity for linked-terms in q ) and $f_3$ (query-attribute similarity) have the most discriminative power. Both of these features consider the similarity between query and entity attribute terms.
Comparing features $f_4$ and $f_5$ with $f_6$ and $f_7$, we observe that query terms linked by entity linker systems usually contain more information about entity attributes of the queries than the not-linked terms.
\begin{table}[t]
\centering
\caption{Feature importance analysis.}
\vspace*{-0.7\baselineskip}
\begin{tabular}{|c|cc|}
\toprule
Feature & NDCG@5 & $\Delta \%$ \\
\midrule
$f_1$--$f_7$ & 0.3227 & 0 \\
\midrule
$f_5$ & 0.2876 & -10.88\% \\
$f_3$ & 0.2771 & -14.13\% \\
$f_4$ & 0.2671 & -17.23\% \\
$f_2$ & 0.1761 & -45.43\% \\
$f_6$ & 0.0919 & -71.52\% \\
$f_1$ & 0.0867 & -73.13\% \\
$f_7$ & 0.0816 & -74.71\% \\
\bottomrule
\end{tabular}%
\label{tab:fi}%
\vspace*{-1.1\baselineskip}
\end{table}%
\begin{figure}[t]
\includegraphics[width=.75\linewidth]{figs/cat}
\vspace*{-0.8\baselineskip}
\caption{Performance of MRF-based and LTR models for different query categories.}
\label{fig:cat}
\vspace*{-1.1\baselineskip}
\end{figure}
\section{Conclusions and Future Work}
\label{sec:concl}
In this paper, we proposed the new task of entity attribute identification, which enables better understanding of search queries.
We employed entity annotations of queries as a bridge to identify entity attributes of queries and proposed two methods to address this task.
Since there is no available test collection for this task, we developed a new test collection based on an established test collection for entity retrieval. Using this collection, we examined our methods with a wide range of entity-bearing queries and showed that our models bring significant and substantial improvements over the baseline methods and are most effective for short relational queries.
For future, we plan to improve our model for complex natural language queries, and incorporate identified attributes of the query in the entity retrieval and question answering tasks.
|
\section*{On restricted exceptional Lie (super)algebras for
\setcounter{section}{1}
\ssec{Introduction} Recall that a Lie aglebra $\fg$ over a field $\Kee$ of characteristic $p>0$ is called \textit{restricted} or \textit{with a $p$-structure} $x\mapsto x^{[p]}$, if for any $x\in\fg$, we have $(\ad_x)^p=\ad_{x^{[p]}}$ for some $x^{[p]}\in\fg$. A module $M$ over a restricted Lie algebra $\fg$, and representation $\rho$ defining $M$ are called \textit{restricted} if $\rho(x^{[p]})=\rho(x)^p$. A Lie superalgebra $\fg=\fg_\ev\oplus\fg_\od$ is \textit{restricted} if $\fg_\ev$ is restricted and $\fg$ is a restricted $\fg_\ev$-module. Thanks to \textbf{squaring}, i.e., the map $x\mapsto x^2(=\frac12[x,x]$ if $p\neq 2$) for any $x\in\fg_\od$, any restricted Lie superalgebra has a $2p$-structure, i.e., a map $x\mapsto x^{[2p]}$ for any $x\in\fg_\od$.
In his Appendix to \cite{LL}, P.~Deligne advised us to investigate first of all the \textbf{restricted} Lie (super)algebras and their \textbf{restricted} modules as related to geometry and hence of interest. This note is an addendum to \cite{BLLS}, in which several general statements on restrictedness
valid for any $p>0$ are formulated, to \cite{BGL2}, where the Cartan matrices and Chevalley generators for modular Lie superalgebras are defined, and to \cite{BGL1, GL3, BGLLS, BGLLS1} describing Lie (super)algebras considered here.
The main result of \cite{BLLS} deals with $p=2$; here we give
examples for $p\neq 2$, mainly for $p=3$.
The ground field is algebraically closed.
Classification \cite{BW} is implicit: to explicitly define $p|2p$-structure on a simple Lie superalgebra $\fg$ it suffices to give expressions of $w^{[p]}$ (resp. $w^{[2p]}$) for all even (resp. odd) elements of any basis of $\fg$. We give, at last, the explicit answer in case $\fsvect_{(1+\bar u)}(m;\One|2s)$, see~\eqref{new}; the deforms of series $\fh$ will be considered elsewhere.
No classification of simple Lie superalgebras
is yet available for any $p>0$, or of simple Lie algebras for $p=3$ and $2$, except for Lie (super)algebras with indecomposable Cartan matrix, and their simple subquotients, see \cite{BGL2}, whose $p|2p$-structure, if exists, is given explicitly. These Lie (super)algebras are \lq\lq symmetric\rq\rq, i.e., have a symmetric root system. For the classification of \textit{true deforms}, i.e., \textit{results} of deformations which are neither trivial nor semitrivial, of symmetric Lie (super)algebras whose restrictedness we establish here, see \cite{BLW, BGL3}.
We also consider vectorial Lie (super)algebras.
Following Bourbaki we use Gothic font for Lie (super)algebras; $\One:=(1,\dots,1)$ is the shearing vector with the smallest heights of divided powers.
Proofs of lemmas, Fact~\eqref{Lie3}, formulas \eqref{Lmin1ps3str} and \eqref{new}
are obtained with the help of the \textit{SuperLie} code, see \cite{Gr}.
\ssec{Deforms of Lie (super)algebras with indecomposable Cartan matrix} In Lemmas \ref{br3} and \ref{g(1,6)}, the cocycles $c_k$ and the elements of the Chevalley basis $x_i$ (resp. $y_i$)
corresponding to the positive (resp. negative) roots are given for the Cartan matrix given in \cite{BGL3},
let the $h_j:=[x_j, y_j]$ be the elements of the maximal torus.
\sssec{Deforms of $\fo(5)$ for $p=3$}\label{Exam}
Recall that the contact bracket of two divided powers $f,g\in \cO(p,q,t; \underline{N})$
is defined to be
\begin{equation*}\label{cb}
\{f,g\}_{k.b.}=\triangle f\cdot\partial_t g - \partial_t f\cdot\triangle g
+\partial_p f\cdot\partial_q g - \partial_q f\cdot\partial_p g\text{~~with $\triangle f=2f - p\partial_p f - q\partial_q f$.}
\end{equation*}
A basis of $\mathfrak{L}(\eps, 0,0)$ is expressed in
terms of generating functions of $\mathfrak{k}(3; \One)$ and root vectors of $\fo(5)$ as
follows, see \cite[Prop.~3.2]{BLW}
\begin{equation*}\label{tab}\small
\renewcommand{\arraystretch}{1.4}
\begin{tabular}{|c|l|} \hline
$\deg$&the element${}_{\text{its weight}}\sim$ its generating function(=Chevalley basis vector) \\
\hline \hline
$-2$&$E_{-2\alpha -\beta}=[E_{-\alpha}, E_{-\alpha-\beta}]\sim 1(=y_4);$\\
\hline
$-1$& $E_{-\alpha}\sim p(=y_2);\quad E_{-\alpha-\beta}=[E_{-\beta}, E_{-\alpha}]\sim q(=y_3);$ \\
\hline
$0$& $H_{\alpha}\sim -\eps t+ pq(=h_2);\quad H_{\beta}\sim -pq(=h_1); \quad
E_{\beta}\sim p^2(=y_1); \quad
E_{-\beta}\sim -q^2(=x_1);$ \\
\hline
$1$& $E_{\alpha}\sim -(1+\eps)pq^2+ \eps
qt(=x_2);\quad E_{\alpha+\beta}=[E_\beta, E_\alpha]\sim (1+\eps)p^2q+\eps p t(=x_3);$ \\
\hline
$2$& $E_{2\alpha+\beta}=[E_\alpha,
E_{\alpha+\beta}]\sim \eps(1+\eps)p^2q^2+ \eps^2t^2(=x_4).$ \\
\hline
\end{tabular}
\end{equation*}
\normalsize
Nonzero values of the deformed bracket with parameters $\delta$
and $\rho$ are as follows:
\begin{equation*}\label{croc1}
\renewcommand{\arraystretch}{1.4}\arraycolsep=2pt
\begin{array}{lll}
{}[E_{-2\alpha-\beta}, E_\beta] = \delta E_{\alpha},
&[E_{-2\alpha-\beta}, E_{-\alpha}]= \delta E_{-\beta},\quad{}
&[E_{-\alpha}, E_{\beta}] = - \frac{\delta }{\eps} E_{2 \alpha+\beta} ,\\
{}[E_{-2\alpha-\beta}, E_{-\alpha-\beta}]= \rho E_{\beta},\quad{}
&[E_{-\alpha-\beta}, E_{-\beta}]= -\frac{\rho }{\eps}
E_{2\alpha+\beta},
&[E_{-2\alpha-\beta}, E_{-\beta}]= - \rho E_{\alpha+\beta}.
\end{array}
\end{equation*}
As proved in \cite{Kos}, in the family $\mathfrak{L}(\eps, \delta,\rho)$, only $\mathfrak{L}(2,0,2)$ and Brown algebras $\fbr(2;\eps):=\mathfrak{L}(\eps,0,0)$ for $\eps\neq 0$
represent classes of non-isomorphic Lie algebras up to isomorphisms
$\fbr(2;\eps)\simeq \fbr(2;\eps')$ if and
only if $\eps\eps'=1$ for $\eps\neq \eps'$; observe that $\fbr(2;-1)\simeq\fo(5)\simeq
\fsp(4)$.
\parbegin{Lemma}\label{3strLemm}
The $3$-structure on $\mathfrak{L}(\eps, \delta,\rho)$ is given by the formulas
\begin{equation}\label{Lmin1ps3str}
\renewcommand{\arraystretch}{1.4}
\begin{array}{l}
h_1^{[3]}=h_1, \quad h_2^{[3]}=\eps^2 h_2, \ \
y_2^{[3]}= \delta(1+2 \eps^2)h_1+
\frac{\delta}{\eps}(1+2 \eps^2)h_2,\quad y_3^{[3]}=
\frac{\rho}{\eps}h_2, \\
y_4^{[3]}=\eps \delta\rho(2+\eps^2)y_1,\quad
y_1^{[3]}=
x_1^{[3]}=x_2^{[3]}=x_3^{[3]}=x_4^{[3]}=0.
\end{array}
\end{equation}
\end{Lemma}
In \cite{Kos}, Rudakov's claim \lq\lq $\mathfrak{L}(\eps, \delta,\rho)$ is restricted\rq\rq\ is cited but the explicit formulas \eqref{Lmin1ps3str} were never published, as far as we know.
\sssbegin{Lemma}\label{br3} Let $\fg_{c_k}$ be the deform with even parameter $\lambda$ corresponding to the cocycle $c_k$ of $\fg=\fbr(3)$ or $\fbrj(2;3)$. The Lie (super)algebras $\fg_{c_k}$, and those symmetric to them ($x\longleftrightarrow y$), are restricted. For any $k$, the $p|2p$-maps vanish on all weight vectors, except the following ones: $h_i^{[3]}=h_i$ for all $i$ and
$x_3^{[3]}=- \lambda h_3$ for $\fbr(3)_{c_{-3}}$, and also $x_5^{[3]}=\lambda (h_2+h_3)$ for
$\fbr(3)_{c_{-6}}$; and $x_6^{[3]}=-\lambda (h_1+h_2+h_3)$ for $\fbr(3)_{c_{-9}}$; and $x_{10}^{[3]}=\lambda (h_1-h_2)$ for $\fbr(3)_{c_{-18}}$. Besides, for $\fbrj(2;3)_{c_{-12}}$, we have $x_{6}^{[3]}=2\lambda h_1$ and for $\fbrj(2;3)_{c_{-6}}$, we have $x_{3}^{[3]}=\lambda (h_1- h_2)$.
\end{Lemma}
\sssbegin{Lemma}\label{g(1,6)} Let $\fg_{c_k}$ be the deform with odd parameter $\tau$ corresponding to the cocycle $c_k$ of $\fg=\fg(1,6)$ or $\fg(4,3)$ or $\fg=\fg(2,3)$. The Lie superalgebras $\fg_{c_k}$, corresponding to the cocycles $c_k$, and those symmetric to them ($x\longleftrightarrow y$), are restricted. For any $k$, the $p|2p$-maps vanish on all weight vectors, except the following ones:
$h_i^{[3]}=h_i$ for all $i$ for $\fg=\fg(1,6)$ and $\fg(4,3)$, and also $ d^{[3]} ={}d$ for $\fg=\fg(2,3)$ modulo the central element ~$c=h_1-h_2$.
\end{Lemma}
\ssec{Fact}\label{fact} Let $\fg_0$ be a
restricted Lie (super)algebra and $\fg_{-1}$ an irreducible restricted $\fg_0$-module that generates the Lie superalgebra $\fg_{-} =\mathop{\oplus}_{-d\leq i<0}\fg_{i}$. Let vectorial Lie (super)algebra $\fg(\sdim;\underline{N})$, where $\sdim$ is the superdimension of $\fg_-$, be the
\textit{prolong}, i.e., the result of generalized
Cartan prolongation, see \cite{Shch}, of the pair $(\fg_-, \fg_0)$.
It is easy to see that \textit{the Lie superalgebra
$\fg(\sdim;\underline{N})$ is not restricted if} $\underline{N}\neq\underline{\One}$, see \cite{BLLS}; the proof of this statement for Lie algebras was first published in \cite[Th.2]{KfiD}.
\textbf{Fact}. \textit{If $\Zee$-graded vectorial Lie
(super)algebra $\fg:=\fg(\sdim;\One)$ --- the generalized Cartan prolong of its non-positive components, see \cite{Shch}, --- is restricted, and the $i$-th derived (super)algebra $\fg^{(i)}$
of $\fg$ contains a maximal torus of $\fg$, then $\fg^{(i)}$ is
restricted} \begin{equation}\label{pvert2p}
\begin{minipage}[l]{14cm}
(a) \textit{$\fh^{[p]}\subset \fh$; moreover, if the structure constants lie in
$\Zee/p$, then $h_i^{[p]}=h_i$ for
the basis elements $h_i$ of $\fh$};\\
(b) \textit{$w^{[p]}=0$ (resp. $w^{[2p]}=0$) for the other even (resp. odd)
weight elements $w$ of the basis of $\fg$ with weights relative a maximal torus of $\fder\ \fg$}.
\end{minipage}
\end{equation}
For the simple derived (super)algebra of every vectorial Lie (super)algebra $\fg$ we know for $p=3$, the $3$-structure is given by
expressions~\eqref{pvert2p}, where $\fh$ is a maximal torus of $\fg_0$; for $p>3$, see \cite[Th.7.2.2]{S}.
\sssec{New examples} The left column in \eqref{Lie3} shows where the simple Lie (super)algebras in the right column are described for any $\underline{N}$; \textit{for them eqs.~\eqref{pvert2p} hold}:
\begin{equation}\label{Lie3}\footnotesize
\renewcommand{\arraystretch}{1.4}
\begin{tabular}{|l|l|}
\hline
\cite{GL3}&$\fdy^{(1)}(10; \One)$,\ $\fby^{(1)}(7; \One)$,\ $\fs\fby^{(1)}(7; \One)$,\ $\fmy(6; \One)$,\
$\fs\fmy^{(1)}(6; \One)$,\ $\fer^{(1)}(3; \One)$,\ $\ffr(3; \One)$\\
\hline
\cite{BL}, \cite{BGL1}, &$\mathfrak{Bj}(1; \One|7)$,\
$\mathfrak{Me}(3; \One|3)$,\ $\mathfrak{Brj}^{(2)}(3; \One|5)$,\ $\mathfrak{Bj}(4; \One|5)$,\ $\mathfrak{Bj}(3; \One|3)$, \ $\mathfrak{Bj}^{(1)}(3; \One|4)$, \\
\cite{BGLLS}& $\mathfrak{Brj}^{(1)}(4|3)$ (only $\underline{N}=\One$ is possible, unlike the above 2 lines)
\\
\hline
\end{tabular}
\end{equation}
\sssec{New formulas}
The deforms of simple derived superalgebras of Lie \textbf{super}algebras of the form $\fg(\sdim;\underline{N})$ are not classified, we consider only one example.
Let $\fsvect(m;\underline{N}|2s)$ be the Lie superalgebra preserving the volume element $\vvol$
in even indeterminates $u_1, \dots, u_m$ and odd ones $u_{m+1},\dots , u_{m+2s}$, let $\bar u=u_1^{(p^{\underline{N}_1}-1)}\cdots u_m^{(p^{\underline{N}_m}-1)}u_{m+1}\cdots u_{m+2s}$ and $\fsvect_{(1+\bar u)}(m;\underline{N}|2s)$ be the deform preserving $(1+\bar u)\vvol$. For $\fsvect_{(1+\bar u)}(m;\One|2s)$ and $\fsvect_{(1+\bar u)}^{(1)}(m;\One|2s)$ in characteristic $p>0$, we have (only distinctions with \eqref{pvert2p} are given):
\be\label{new}
((1 - \bar{u})\partial_i)^{[p] }= - (\partial_i^{p-1} \bar{u}) \partial_i, \text{~~where $\partial_i$ is even.}
\ee
|
\section{Introduction}
{\it Introduction.--}
Heat transfer in the far field can be well described by Planck's theory of black-body radiation \cite{Planck} and obeys the Stefan-Boltzmann law, which is independent of gap separation distance.
When the gap separation between two bodies becomes smaller than Wien's wavelength, heat transfer in the near field becomes distance dependent and has been demonstrated to be much larger than that in the far field \cite{near1, review3, review4}.
Within the fluctuational electrodynamics \cite{review3, review4, PvH}, the near-field heat flux typically increases as the two bodies become closer. As such, great efforts have been dedicated to reducing the gap sizes from orders of $1\,\mu$m \cite{XJB, mu-m1, mu-m2} to a few nanometers \cite{nm2, nm3, nm4, nm5-PvH} or even down to few \textup{\AA}ngstr\"{o}ms \cite{angstrom1, angstrom2}, resulting in several folds to several orders of heat transfer enhancement compared to the corresponding far-field results, which may prove useful for near-field thermal management. To our best knowledge, heat flux typically increases with the decrease of gap separation.
Besides by reducing gap separation, several other approaches have been brought forward to enhance near-field heat transfer. Pendry showed that the heat flux can be greatly enhanced by tuning the resistivity of the material \cite{resist1}. Covering both surfaces with adsorbates, so that resonant photon tunneling happens between adsorbate vibrational modes, can also enhance the heat flux \cite{adsorbate}.
Two types of surface waves, which propagate along the material-vacuum interfaces, have been mainly used to increase heat transfer in the near field. One is surface phonon polariton supported in polar dielectrics, such as SiC and SiO$_2$ \cite{SPhP1, SPhP3, SPhP-SPP1, SPhP-SPP2}. The other type is surface plasmon polariton on materials supporting low frequency plasmon \cite{adsorbate, SPhP-SPP1, SPhP-SPP2, plasmon1,plasmon2}, such as graphene \cite{graphene1, graphene2, Yu}, black phosphorus \cite{bp} and silicon \cite{silicon1, silicon2}.
Using hyperbolic metamaterials can help to enhance heat transfer as well \cite{hyperbolic1,hyperbolic2,hyperbolic4}.
In this letter, we report peculiar vacuum gap dependence and enhancement of heat transfer in the near field in the presence of electronic edge states. We consider heat transfer between two carbon-based nano-structures harboring electronic edge states separated by a vacuum gap no further than $20\,$nm, so that (i) the electron-electron interaction dominates the heat transfer, and (ii) an atomistic description is more appropriate. The heat current expression is given in the formalism of the nonequilibrium Green's function (NEGF) within the random phase approximation. Heat current can reach maximum at a finite vacuum gap if the real part of the charge susceptibility is large near zero angular frequency, and this can be realized through the presence of edge states.
The peculiar behaviors are demonstrated using zigzag single-walled carbon nanotubes (SWCNTs) [See Fig.~\ref{fig1}(a)] and graphene nano-triangles forming a bowtie shape [Fig.~\ref{fig1}(b)]. We uncover the the mechanism using the simple Su-Schrieffer-Heeger (SSH) chains [See the inset of Fig.~\ref{fig3}(b)].
\begin{figure}[tb!]
\centering
\includegraphics[width=\columnwidth]{fig1.ps} \\
\caption{Near-field heat transfer between two SWCNTs (a) and graphene nano-triangles forming a bowtie shape (b) which are separated by a vacuum gap with distance $d$. (c) Lattice structure of a zigzag SWCNT corresponding to the left side in panel (a). For the graphene nano-triangles, which are equilateral, we show the case with side length to be $L=4.26\,$nm. }
\label{fig1}
\end{figure}
{\it Theoretical formalism.--}
When two metallic surfaces are separated by a vacuum gap, which is much smaller than the Wien's wavelength $\lambda_{th}$ (several micrometers at room temperature), the contribution to heat transfer from the retarded vector potential can be safely ignored, and the electron-electron interaction dominates the heat transfer \cite{Yu,Mahan,JS1,JS3,JS4,GM1, supp}. If the vacuum gap is below around $1.5\,$nm, electron tunneling process can happen \cite{JS3, crossover}, the picture dominated by electron-electron interaction does not apply. So our formalism below can faithfully describe the heat transfer with the vacuum gap $d$ in the range $1.5\,{\rm nm} < d \ll \lambda_{th}$. The lattice Hamiltonian of a general heat transfer mediated by electron-electron interaction can be written as,
\begin{equation}
H = \sum_{mn} c_m^\dag h_{mn} c_n+ \frac{e_0^2}{2}\sum_{mn} c_m^\dag c_m v_{mn} c_n^\dag c_n , \label{Hamiltonian}
\end{equation}
with $e_0$ the elementary charge. $c_m$ ($c_m^\dag$) is the fermionic annihilation (creation) operator of lattice site $m$ on the left or right side, and $h_{mn}$ is the on-site energy for $m=n$ and hopping parameter for $m\neq n$ locating on the same side. For the situation of indices $m$ and $n$ locating on different sides, $h_{mn}$ vanishes. This implies that electron tunneling from one side to the other is impossible when the vacuum gap is far greater than the spacing between nearest-neighbor atoms. $v_{mn}$ is the Coulomb potential between site $m$ and $n$. Heat transfer occurs via the charge fluctuations between the electronic states sitting at the edges from both sides for the setups considered.
Under random phase approximation, the heat current is expressed as \cite{Yu,JS4},
\begin{equation}
J = \int_0^{\infty} \frac{d\omega}{2\pi} \hbar\omega {\cal T}(\omega)\big[N_L(\omega)-N_R(\omega)\big] ,
\end{equation}
where $N_\alpha(\omega) = 1/[e^{\beta_\alpha \hbar\omega}-1]$ is the Bose-Einstein distribution with $\beta_\alpha = 1/(k_B T_\alpha)$ and $\alpha=L,R$.
The spectral transfer function is given by,
\begin{equation}
{\cal T}(\omega) = 4{\rm Tr} \big\{ \Delta^\dag(\omega) v_{RL} {\rm Im}[\chi_L(\omega)] v_{LR} \Delta(\omega) {\rm Im}[\chi_R(\omega)] \big\} ,
\end{equation}
where the trace is over lattice sites. $\Delta(\omega)$ is expressed as
\begin{equation}
\Delta(\omega) = [{\bf I}- \chi_R(\omega) v_{RL} \chi_L(\omega) v_{LR}]^{-1} , \label{Delta}
\end{equation}
with identity matrix ${\bf I}$ and the charge susceptibility $\chi_{\alpha}(\omega)$ in lattice space obtained through electronic Green's function \cite{supp}.
The entries of the Coulomb potential matrices between left and right sides are $v_{mn} = 1/(4\pi\epsilon_0 d_{mn})$, where $\epsilon_0$ is the dielectric constant of vacuum. $d_{mn}$ is the Euclidean distance between site $m$ and $n$, which sit on different surfaces. The formalism based on NEGF here has been shown \cite{Yu,JS3,JS4} to reduce to that by the fluctuational electrodynamics \cite{PvH,review3, review4} in the non-retardation limit and be equivalent to that given by Mahan \cite{Mahan}.
\begin{figure}[tb!]
\centering
\includegraphics[width=\columnwidth]{current_carbon2.ps}\\
\caption
Gap separation dependences of heat current between zigzag SWCNTs for different $M$ with $\varepsilon =5$ and $\eta=25\,$meV [panel~(a)], and for different dielectric constants $\varepsilon$ with $M=15$ and $\eta=25\,$meV [panel~(b)]. Gap separation dependence of heat current between graphene nano-triangles for different $L$ with the dielectric constant $\varepsilon =5$ [panel (a)], and for $\varepsilon$ with $L=8.52\,$nm [panel (b)].}
\label{fig2}
\end{figure}
{\it Near-field heat transfer between zigzag SWCNTs.--}
For the numerical results presented in this work, temperatures of the left and right sides are set as $T_L=400\,$K and $T_R=300\,$K, respectively.
We first discuss heat transfer between two zigzag SWCNTs in near field. It has been reported that zero-energy localized states (Fujita's edge states) emerge at the edges of zigzag SWCNTs \cite{CNT1, CNT2, CNT3}.
As shown in Fig.~\ref{fig1}(c), the outermost carbon atoms are denoted by $Am$ and the other surface atoms by $Bm$, where $m$ is the site index. The contribution of heat transfer between two bodies are mainly from these surface carbon atoms. The local electronic density of states for the atoms $Am$ are sharply peaked at zero-energy as shown in Fig.~1(a) in the Supplementary Material \cite{supp}.
We use $M$ to denote the total number of the outermost carbon atoms, and it is proportional to the radius of the nanotube. Due to the $O(2)$ rotational invariance of carbon nanotubes, all the carbon atoms of $Am$ and $Bm$ are geometrically equivalent in their respective sides.
The carbon-carbon bond length is $1.42\,$\textup{\AA}. The recursive Green's function technique \cite{Recursive} is used in getting electronic Green's function of the surface sites, where
the nearest-neighbour hopping constant is given by $2.5\,$eV. A damping constant $\eta=25\,$meV is included in calculating Green's function to account for possible dissipations by such as electron-phonon interactions. A dielectric constant $\epsilon$ is included in calculating the intra-side Coulomb interaction.
In Fig.~\ref{fig2}(a), we plot heat currents versus gap distances for zigzag SWCNTs with different radii. The non-monotonic behavior is found for the cases of $M=15$, $20$, and $25$ shown in Fig.~\ref{fig2}(a). The maximal heat currents are identical for these three cases. We find that the larger the nanotube radii, the longer the distances for achieving corresponding maximal heat current. The condition(s), under which the heat current can reach maximum with several nanometers of gap separation for zigzag SWCNTs, can be obtained analytically. The detailed argument is provided in the Supplementary Material \cite{supp}. The critical distance is approximately the summation of all entries of the real parts of the charge susceptibility at zero frequency, that is,
\begin{equation}
d_c \approx -\frac{1}{4\pi\epsilon_0} \sum_{m,n}{\rm Re}[\chi_{mn}(\omega=0)] . \label{d_c}
\end{equation}
Since the terminal sites contribute to the critical distance in an additive way, a small amount of disorder or vacancies will only decrease the critical distance slightly without changing the maximal heat current.
For the case $M=10$ in Fig.~\ref{fig2}(a), the critical distance appears below $1.5\,$nm as predicted by equation~\eqref{d_c}. For such extremely short distance, there may be contribution to the heat transfer from electron tunneling \cite{JS3, crossover}, an aspect which is not included in our model.
Gap separation dependence of heat current for different dielectric constants $\varepsilon$ in zigzag SWCNTs is shown in Fig.~\ref{fig2}(b). One can find that the critical distance disappears for the case with $\varepsilon =3$. With increasing dielectric constant $\varepsilon$, the critical distance shifts to larger values. This is because the amplitudes of the real parts of charge susceptibility at zero frequency increase as a consequence of less screening.
Different dielectric constants can be realized by encapsulating or covering the carbon nanotubes with different dielectric materials. Inserting a dielectric material as a core of the nanotubes can change the dielectric constant as well.
The area formed by a zigzag SWCNT with $M=15$ is $1.1\times 10^{-18}\,{\rm m}^2$, so that the corresponding heat flux, i.e., heat current per area, is about $2.8\times 10^{10}\, {\rm W}/{\rm m}^2$ with heat current reaching maximum $J=30.6\,$nW at critical distance. This heat flux is several orders of magnitude larger than those mediated by surface phonon polaritons or surface plasmon polaritons \cite{silicon2}, and almost comparable to that of heat conduction. Since the structure-controlled growth of SWCNTs is rapidly advanced \cite{CNT-exp1, CNT-exp2}, the results here can be are expected to be experimentally realized in the near future.
{\it Near-field heat transfer between graphene nano-triangles.--}
We further show the non-monotonic behavior of gap separation for near-field heat transfer between two gaphene nano-triangles, which are equilateral. We focus on the vertex-to-vertex geometry forming a bowtie shape as shown in Fig.~\ref{fig1}(b), and length of the nano-triangle's side is $L$. The nearest-neighbour hopping constant is chosen as $2.8\,$eV, and the damping constant $\eta=25\,$meV. The graphene nano-triangles can be maintained at thermal equilibrium through optical pumping or by being attached to additional electrodes. As shown in Fig.~2 in Supplementary Material \cite{supp}, the main contributions to the heat transfer are from sites $Am$ indicated in Fig.~\ref{fig1}(b), which have strongly localized zero-energy states. Gap separation dependent behaviors of heat current by varying side length $L$ and dielectric constant $\varepsilon$ are shown Fig.~\ref{fig2}(c) and (d), respectively.
Similarly to the behaviors found in Fig.~\ref{fig2}(a) for zigzag SWCNTs, we observe non-monotonic behavior for $L=8.52\,$nm, and $L=12.78\,$nm as shown in Fig.~\ref{fig2}(c). From Fig.~\ref{fig2}(d), we see that increasing the dielectric constant in graphene nano-triangles increases the critical distance as well, which is as shown in Fig.~\ref{fig2}(b).
{\it Near-field heat transfer between SSH chains.--}
We bring forth a theoretical understanding for the peculiar phenomenon observed above using the simple one-dimensional SSH chains, which undergo topological phase transitions by varying the hopping parameters \cite{SSH1}.
The Hamiltonian of a SSH chain for side $\alpha$ is expressed as,
\begin{align}
H_{0\alpha} = & -(1+\lambda_\alpha)t \sum_{n=1}^{N} ( c_{An}^\dag c_{Bn} + {\rm H.c.} ) \notag \\
&-(1-\lambda_\alpha)t \sum_{n=1}^{N-1} ( c_{An+1}^\dag c_{Bn} +{\rm H.c.}) ,
\end{align}
with $N$ the number of lattice sites, and $\lambda_\alpha \in [-1,1]$. We consider the case where heat transfer happens only between two end sites, both of which are labeled as $A1$ as shown in the inset of Fig.~\ref{fig3}(a).
We set $\lambda_L=\lambda_R=\lambda$ and the hopping constant as $t=2.2\,$eV in the calculation. A damping constant with $\eta = 22\,$meV is added to each site in calculating electronic Green's function \cite{SSH-Green}.
The energy spectrum of an open SSH chain is shown in the right inset of Fig.~\ref{fig4}(a).
When $\lambda>0$, SSH chain is in a trivial insulator state without in-gap state, and it is in a metallic state for $\lambda=0$ where the gap closes. However, the gap reopens in the topologically nontrivial region with $\lambda<0$, and zero-energy in-gap states appear when open boundary condition is taken.
\begin{figure}[tb!]
\centering
\includegraphics[width=\columnwidth]{current_distance_SSH3.ps} \\
\caption{Gap separation dependence of heat current in the metallic phase (a) and in the topologically nontrivial phase with different $\lambda$ (b). Heat currents with gap separation below $1.5\,$nm are not shown because electron tunneling will no longer be negligible and we wish to exclude this effect. (c) The spectral transfer functions for different gap separations $d$ with $\lambda=-0.3$. $d=4.3\,$nm is the critical gap separation above which the current decays with increasing gap separation. (d) Real and imaginary parts of the charge susceptibility function of the end site $A1$ with temperatures $400\,$K ($\chi_L$, solid lines) and $300\,$K ($\chi_R$, dashed lines).
$\hbar\omega\delta N(\omega)$ is shown as an inset of panel (c). }
\label{fig3}
\end{figure}
The gap separation dependence of heat current in the metallic phase and the topologically nontrivial phase are shown in panels (a) and (b) of Fig.~\ref{fig3}, respectively. For the case of $\lambda=-0.3$, the spectral transfer functions for different gap separations $d$ and the charge susceptibilities $\chi(\omega)$ of the end sites $A1$ are plotted in Fig.~\ref{fig3}(c) and (d), respectively.
There exists a critical gap distance $d_c$ at which the heat flux achieves its maximum point in the presence of edge state. The explanation of this peculiar distance dependence is as follows.
In atomic units, the critical distance is approximately equal to the absolute values of the real parts of $\chi_{L/R}(\omega)$ at zero frequency, i.e., $d_c \approx -{\rm Re}[\chi_{L}(\omega=0)]/(4\pi\epsilon_0)$.
(For $\lambda=-0.3$, the critical distance $d_c$ is near $4.3\,{\rm nm}=81.26\,{\rm a.u.}$. The value in atomic unit is between $|{\rm Re}[\chi_L(\omega=0)]|$ and $|{\rm Re}[\chi_R(\omega=0)|$ which are shown in Fig.~\ref{fig3}(d).)
Near the critical distance $d_c$, ${\rm Re}[\chi_L(\omega\rightarrow 0)]v_{LR} \approx {\rm Re}[\chi_R(\omega\rightarrow 0)]v_{RL} \approx -1$.
One also has ${\rm Im}[\chi_{L/R}(\omega\rightarrow 0)] \approx 0$, so that $v_{LR}\Delta{\rm Im}(\chi_R) |_{\omega\rightarrow 0} \approx i/2$ from equation~\eqref{Delta}, hence we have ${\cal T}(\omega\rightarrow 0)\approx 1$ around the critical distance. Since the function $\hbar\omega \delta N(\omega)$ is a decreasing function with respect to angular frequency $\omega>0$ (shown as an inset of Fig.~\ref{fig3}(c)), the magnitude of the spectral transfer function at low $\omega$ dominates the heat current amplitude. The resonant peak of ${\cal T}(\omega)$ is located close to $\omega=0$ at the critical gap distance at which the heat current achieves its maximum.
The resonant peak shifts towards larger angular frequency with decreasing gap distance (as shown in Fig.~\ref{fig3}(c) for $\lambda=-0.3$), and this results in a suppressed heat current. Above the critical distance, the condition for the appearance of the resonant peak cannot be satisfied. With increasing distance above the critical distance, the magnitude of the spectral transfer function decreases, so does the heat current.
In the metallic phase, we have ${\rm Re}[\chi_L(\omega = 0)] \approx -5$, which means that the critical distance is about $2.6\,$\textup{\AA}. At such a small gap distance, heat conduction due to electron tunneling can happen, and our formalism does not apply \cite{JS3}.
As $\lambda$ approaches $-1$, the critical distance increases due to the fact that edge states become more localized. The maximum heat currents are almost the same in presence of edge states because they share similar spectral transfer function profiles regardless of the critical gap distances. The non-monotonic behaviors with respect to gap separation have been reported for heat radiation between a cylinder and a perforated surface due to dipolar effects \cite{non-monotonic1} and in a multilayer structure due to the interplay of contributions from different surfaces \cite{non-monotonic2}. If the SSH chains experience more dissipation, i.e., larger $\eta$, the edge sates become less localized, and the critical distance becomes shorter [See Fig.~4 in the Supplementary Material \cite{supp}].
\begin{figure}[tb!]
\centering
\includegraphics[width=\columnwidth]{current_lambda4.ps} \\
\caption{(a) Heat current by changing $\lambda$ for a gap separation of $3\,$nm, at which the electron-electron interactions dominate the heat transfer. A log-scale plot for the heat current is shown in the left inset. (b) The spectral transfer functions for different $\lambda$. Energy spectrum of the SSH chain with $160$ lattice sites as a function of $\lambda$ is shown as a right inset of panel (a). }
\label{fig4}
\end{figure}
In Fig.~\ref{fig4}(a), heat current versus $\lambda$ for gap separation $d=3\,$nm is plotted. The heat current for the metallic phase ($\lambda =0$) is several orders smaller than that for the topologically nontrivial phase. In the trivial insulating phase ($\lambda >0$), heat current is extremely small and almost vanishes. A sharp jump occurs with the phase transition point $\lambda <0$, indicating that the presence of edge state can drastically enhance heat current compared to the metallic phase. The spectral transfer functions for different $\lambda$ are shown in Fig.~\ref{fig4}(b). The increase of heat current as $\lambda$ is changed from $-1.0$ to around $-0.22$ can be attributed to the shift of the resonant peak of the spectral transfer function towards $\omega=0$. This is because that $d=3\,$nm is the critical distance corresponding to $\lambda = -0.22$. At $\lambda = -0.22$, the resonant peak of the spectral function ${\cal T}(\omega)$ is at a frequency close to $\omega=0$. For $\lambda<-0.22$, the chosen gap separation $d=3\,$nm is smaller than the corresponding critical distance, and the resonant peak locates at a larger frequency.
By further increasing $\lambda$ from $-0.22$, the SSH chain approaches the metallic phase, and the peak of the spectral transfer function at low angular frequency decreases, thus reducing the heat current.
The fact that heat current can be greatly enhanced in the presence of edge states can be exploited to design a near-field thermal switch, provided that edge states can be tuned. The discussions of near-field heat transfer between SSH chains, with the same chemical potential applied to both sides, are shown in the Supplementary Material \cite{supp}.
{\it Summary and discussion.--}
We have uncovered the peculiar behaviors of near-field heat transfer in the presence of electronic edge states. Our findings are demonstrated using zigzag SWCNTs and graphene nano-triangles forming a bowtie shape. The underlying mechanism is uncovered through the simple SSH chains. In the presence of localized zero-energy edge states, heat current is greatly enhanced and shows a non-monotonic behavior with respect to vacuum gap separation. The maximal heat flux between zigzag SWCNTs are shown to be extremely large, and surpasses near-field heat flux being reported so far.
\begin{acknowledgments}
{\it Acknowledgments.--}
The authors thank Jiebin Peng, Giovanni Vignale, Fuming Xu and Songbo Zhang for discussions and comments. G.T. and J.S.W. acknowledge the financial support from RSB funded RF scheme (Grant No. R-144-000-402-114). J.R. is supported by the NNSFC (No. 11775159), Shanghai Science and Technology Committee (No. 18ZR1442800, No. 18JC1410900), and the Opening Project of Shanghai Key Laboratory of Special Artificial Microstructure Materials and Technology.
\end{acknowledgments}
|
\section{Introduction}
We analyse the stochastic stability of a model of a neural network. Our model is inspired by the \textit{stochastic integrate-and-fire neuron} model. The original model of membrane potentials was introduced by Lapicque (1907) and has been developed over the years (for a review of the model see, e.g., Burkitt (2006a,b)). In this model, at any time $t$, the internal state of a neuron $i$ is given by its membrane potential $Z_i(t)$, which evolves according to a stochastic differential equation
$$dZ_i(t) = F(Z_i(t), I(t), t)dt + \sigma(Z_i(t), I(t), t)dW_i(t),$$
where $F$ is a drift function, $\sigma$ the diffusion coefficient, $I$ is the neuronal input, and $W_i$ is a Brownian motion (see, e.g., Gerstner and Kistler (2002)). The process $W_i(t)$ represents combined internal and external noise. The process $I$ models firings of neurons' potentials (or ''spikes''): whenever a potential $Z_i(t)$ reaches certain threshold, it resets to a base-level, and the neuron sends signals to other neurons.
A large number of experiments have given us an understanding of the dynamics of a single neuron. For example, Hodgkin and Huxley (1952) found three different types of ion current flowing through a neuron's membrane, and introduced a detailed model of a membrane potential. To give a basic description, without any input, the neuron is at rest, corresponding to a constant membrane potential. Given a small change, the membrane potential returns to the resting position. If the membrane potential is given a big enough increase, it reaches a certain threshold, and exhibits a pulse-like excursion that will effect connected neurons. After the pulse, the membrane potential does not directly return to the resting potential, but goes below it. This is connected to the fact that a neuron can not have two spikes one right after another.
As for the neural networks, neurons form a connected graph with synapses between neurons. When a presynaptic neuron fires a spike it sends a signal through a synapse to a postsynaptic neuron. A neuron is called \textit{inhibitory} if its signals predominantly move the membrane potentials away from a threshold; and \textit{excitatory} if they move potentials toward a threshold. In this paper, we consider a model containing inhibitory neurons only. It is important to point out that the effect of a signal depends on the potential of a receiving neuron. For example, if the membrane potential of a postsynaptic neuron is lower than that of a corresponding inhibitory synapse, the effect of a signal will be reversed. Therefore, in the models where signals and potentials are assumed to be independent, it is important to assume that the potentials should not decrease too much.
One of the most classical models, introduced by Stein (1965), is \textit{leaky integrate-and-fire neuron} model, where
$$F(Z_i(t), I(t), t) = -\alpha Z_i(t) + I(t), \ \ \sigma(Z_i(t), I(t), t) = \sigma = \text{const.}$$
There are several variations of this model. For instance, nonlinear models were considered, such as the quadratic model (see, e.g., Latham \textit{et al.} (2000)) where $-\alpha Z_i(t)$ is replaced with\\
$a(Z_i(t) - z_{rest})(Z_i(t) - z_c)$, where $z_c > z_{rest}$. Another direction for generalisation of this model is the Spike Response Model (see, e.g., Gerstner and Kistler (2002), Chapter 4.2). In this model, the relation between the dynamics and the potential is determined by the time of the last spike. This allows one to explicitly forbid spikes to occur one right after another and to write the dynamics in integrated form.
In this paper we consider \textit{perfect integrate-and-fire neuron} model, where $\alpha = 0$ and, therefore, the decay of the membrane potential over time is neglected. This restriction is a stepping stone to achieve more general results and it allows us to write the model in integrated form.
In our model, the spikes and corresponding signals are represented by shifts from a threshold of a random length, independent of everything else. We analyse the system under certain conditions on the distribution of those shifts and prove stability. Instead of considering the recurrence of sets $[-k, H]^N$ (where $H$ is a threshold and $N$ is a number of neurons), we move each coordinate down and reflect the system to work with more convenient sets $[0, k+H]^N$. Thus, in our model, membrane potentials are nonnegative processes that jump to a random positive level after reaching zero. Signals from inhibitory neurons push membrane potentials from the threshold, i.e. they are positive shifts. It is important to note that we assume that travel time of signals between neurons is zero, which in general can cause uncertainty in the order of spikes. However the inhibitory signals do not cause spikes right away, and we assume that the potentials $Z_i(t)$ almost surely do not reach their thresholds at the same time. We refer to Taillefumier \textit{et al.} (2012) for further discussion.
It is often assumed that the studied system of neurons is itself a part of a much larger system of neurons. The corresponding effect on our system is often modelled by a multivariate Brownian motion $W(t)$ with a drift (the drift guaranties the stability of a system of a single neuron). However, we can generalise it to a multivariate spectrally positive (i.e. with positive jumps) L\'evy process $X(t)$ to account for inhibitory signals. It is important for our analysis that the signals do not influence the dynamics of the process $Z(t)$ if it is away from the threshold, i.e. we have $dZ(t) = dX(t)$ if $Z_i(t) > 0$, for $i\in \{1, \ldots, N\}$. Nevertheless, the number of spikes $\eta_i(t)$ before time $t$ is essential to stability analysis. The fact that $\eta_i(t)$ is not pathwise monotone with respect to signal sizes or the initial state brings certain difficulties in proving stability.
As was mentioned, a system of a single neuron is stable, however, for a general distribution of signals between neurons, ''partial stability'' can occur when only a strict subset of neurons (maybe random) stabilises, while membrane potentials of other neurons are ''pushed'' to infinity (which contradicts the physical setup). The latter is of independent mathematical interest and will not be discussed in the main body of the paper.
Under specific conditions on average signals and the drift $\Expect X(1)$, we prove the positive recurrence of the system using the fluid approximation approach, introduced by Rybko and Stolyar (1992) and Dai (1995) (see also Stolyar (1995)). Although this method is usually shown on queueing networks, it is quite universal, and applicable to our model too. Using results from Section 7 of Borovkov and Foss (1992) (see, also, Chapter VII of Asmussen (2003)), we prove \textit{Harris positive recurrence} and convergence to stationary distribution in total variation. We refer to Foss and Konstantopoulos (2004) for an overview of some stochastic stability methods.
The \textit{stochastic integrate-and-fire neuron} model has received an increasing amount of attention in recent years. There are a number of papers considering mean-field limits of such systems. De Masi \textit{et al.} (2015) consider a model with identical inhibitory neurons, where each membrane potential has a drift to the average potential. Inglis and Talay (2015) consider general signals between neurons and describe signal transmissions through the use of the cable equation (instead of instant transmissions). Robert and Touboul (2016) consider a model where neurons do not have a fixed threshold and spikes occur as a inhomogeneous Poisson process, with intensity given as a function of a membrane potential, and prove ergodicity. Several authors studied cases when excitatory signals lead to so-called \textit{blow-up phenomena} (e.g. C\'aceres \textit{et al.}(2011), Delarue \textit{et al.} (2015)).
This paper is structured as follows. In \sectn{model&results} we define our model, introduce auxiliary concepts and notations, and formulate our results. In particular, in \sectn{fluid_model} we introduce the fluid model and formulate related technical results. In \sectn{auxiliary_results} we prove important auxiliary results. In \sectn{lemma_positive_recurrence} we prove positive recurrence. In \sectn{lemma_mixing} we prove that our model satisfies the classical ''minorization'' condition. The Appendix includes the remaining auxiliary results and comments.
\section{Model and results}\label{sect:model&results}
We analyse a network of $N$ stochastic perfect integrate-and-fire inhibitory neurons. At any time $t$, the internal state of all neuron is given by a multidimensional process $Z(t)$ which represents neurons' membrane potential. Let $X(t)$ be a $N$-dimensional spectrally positive left-continuous L\'evy process with a finite mean and its distribution has a non-degenerate absolute continuous component. The process $X(t)$ represents combined internal and external noise. Let $\nu_i = - \Expect X_i(1) > 0$ and $X^0_i(t) = \nu_i t + X_i(t)$. While $Z(t) \in (0, \infty)^N$, membrane potentials evolve as the process $X(t)$, i.e. $dZ(t) = dX(t)$. Let $\{\{\xi^{(k)}_{ij}\}_{i, j=1}^N\}_{k=1}^\infty$ be i.i.d. random matrices, independent of everything else, with a.s. strictly positive elements.
\begin{Remark}
One can allow the absolute continuous component of the distribution of the process $X_i(t)$ to be degenerate (for example, take a sum of a Poisson process and a linear function $-at$) and, instead, condition the distribution of the matrix $\{\xi^{(1)}_{ij}\}_{i, j=1}^N$ to have an absolute continuous component. The main result of the paper would still hold and the proof would need few minor changes.
\end{Remark}
Let $b_{ij} = \Expect \xi^{(1)}_{ij} < \infty$ and $S_{ij}(n) = \sum_{k=1}^{n} \xi^{(k)}_{ij},$ for $i, j \in \{1, 2, \ldots, N\}$. If potential $Z_i(t)$ hits non-positive values for the $k$-th time, then instantaneously it increases to $\xi^{(k)}_{ii}$ and other membrane potentials increase by $\xi^{(k)}_{ij}$. We call this event ''a spike of neuron $i$``.
Let $Z^{\textbf{z}}(t) = (Z^{\textbf{z}}_1(t), \ldots, Z^{\textbf{z}}_N(t)) \in \mathcal{Z} = [0, \infty)^N$ be the membrane potentials at time $t$ with an initial value $\textbf{z} = (z_1, \ldots, z_N)$. Let $T^{\textbf{z}}_{i0} = 0$ and
$$T^{\textbf{z}}_{ik} = \inf\{t> T^{\textbf{z}}_{i(k-1)}: \ Z^{\textbf{z}}_i(t) \leq 0\}, \ \ \text{for $k\geq 1$,}$$
be the times when neuron $i$ reaches its threshold. Let $\eta^{\textbf{z}}_i(0) = 0$ and $\eta^{\textbf{z}}_i(t) = \max\{k: \ T^{\textbf{z}}_{ik} < t\}$ be the number of spikes of $Z^{\textbf{z}}_i(t)$ before time $t$. Then the dynamics of the system is given by
\begin{equation}\label{eq:dynamic_equation}
Z^{\textbf{z}}_i(t) = z_i + X_i(t) + \sum_{j=1}^N \sum_{k=1}^{\eta^{\textbf{z}}_j(t)} \xi^{(k)}_{ji} = z_i + X_i(t) + \sum_{j=1}^N S_{ji}(\eta^{\textbf{z}}_j(t)), \ \ i=1, \ldots, N.
\end{equation}
Before talking about stability of the system it is important to point out that, due to the negative drift, one can easily show that potential of an isolated neuron is stable (this is also a subcase of our main result).
\begin{Remark}\label{rem:partial_stability}
There are examples of parameters $\nu_i$ and $b_{ij}$ such that there exists a subset of neurons which, after reaching stability, can ''push other neurons to infinity``. We do not discuss such cases of partial stability in the main body of the paper, however, we include a few comments in the Appendix.
\end{Remark}
We assume that all potentials have the same drift $\nu$ and that signals from neuron $i = 1, \ldots, N$ to all other neurons have the same mean $w_i$. We have
\begin{equation}\label{eq:simple_network}
\nu_i = \nu >0, \ \ b_{ij} = \Expect \xi^{(1)}_{ij} = w_i > 0 \ \ \text{and} \ \ b_{ii} = H_i > w_i, \ \ \text{for $i=1, \ldots, N$ and $j\neq i$.}
\end{equation}
\begin{Theorem}\label{thr:ergodicity}
Assume condition \textnormal{\eq{simple_network}} to hold. Then the process $(Z^{\textbf{z}}_1(t), \ldots, Z^{\textbf{z}}_N(t))$ is Harris positive recurrent: there is a distribution $\pi$ such that
$$\sup_{A}|\PB\{Z^{\textbf{z}}(t) \in A\} - \pi(A)| \to 0 , \ \text{as $t\to\infty$.}$$
\end{Theorem}
\begin{Remark}\label{rem:spike_rate}
Given \textnormal{\eq{simple_network}}, matrix $B = (b_{ij})_{i, j=1}^N$ is invertible and
$$(\textbf{1}B^{-1})_i = \frac{1}{\left( H_i - w_i \right) \left( 1 + \sum_{k=1}^{N}\frac{w_k}{H_k - w_k} \right)}, \ \ \text{for $i = 1, \ldots, N$,}$$
where $\textbf{1} = (1, \ldots, 1)$. Vector $\nu \textbf{1}B^{-1}$ represents rates of spikes when stability is achieved. In particular, for large $t > \nu^{-1}(1 + \sum_{k=1}^N w_k / (H_k - w_k))$ and for each sequence $\textbf{z}_n$, $\|\textbf{z}_n\| \to \infty$, there exists a subsequence $\textbf{z}_{n_k}$ such that
$$\frac{\eta^{\textbf{z}_{n_k}}(\|\textbf{z}_{n_k}\|(t+\Delta)) - \eta^{\textbf{z}_{n_k}}(\|\textbf{z}_{n_k}\|t)}{\|\textbf{z}_{n_k}\|\Delta} \Rightarrow \nu \textbf{1}B^{-1}, \ \text{for $\Delta >0$.}$$
\end{Remark}
We prove \thr{ergodicity} following two standard steps. For the reader's convenience, we formulate those steps as lemmas. Let $\tau^{\textbf{z}}(\varepsilon, B) = \inf\{t> \varepsilon: \ Z^{\textbf{z}}(t) \in B\}$ be the first hitting time of a set $B$ after time $\varepsilon$. The first step is the proof of positive recurrence which we achieve via the fluid approximation method.
\begin{Lemma}\label{lem:positive_recurrence}
There exists $k_0 >0$ such that for $V = \{\textbf{z}\in \mathcal{Z}: \|\textbf{z}\| < k_0\}$ we have
$$\sup_{\textbf{z}\in V} \Expect\tau^{\textbf{z}}(\varepsilon, V) < \infty.$$
\end{Lemma}
In the second step, we show that our model satisfies the classical ''minorization'' condition.
\begin{Lemma}\label{lem:minorant_measure}
There exist a number $p>0$ and a probability measure $\psi$ such that for a uniformly distributed r.v. $U \in [1,2]$, independent of everything else, we have
$$\inf_{\textbf{z} \in V} \PB\{Z^{\textbf{z}}(U) \in B\} \geq p \psi(B).$$
\end{Lemma}
Using Lemmas $1$ and $2$ we can prove that conditions of Theorem 7.3 from Borovkov and Foss (1992) are satisfied, which gives us the result. The proof of \lemt{positive_recurrence} is based on the fluid approximation. We dedicate the following subsection to formulate corresponding definitions and auxiliary results. We point out that we need to assume condition \eq{simple_network} only in the proof of \lemt{positive_recurrence} and in \rem{spike_rate}.
One of the difficulties of our model is lack of path-wise monotonicity for the number of spikes $\eta^{\textbf{z}}(t)$ with respect to signals $\xi^{(k)}_{ij}$ or initial state $\textbf{z}$. In general, making one neuron firing a spike earlier may lead to other spikes occur later. However, there is a ''partial monotonicity'' which allows us to get an upper bound for process $\eta^{\textbf{z}}(t)$ with useful properties.
Since all neurons are inhibitory, one way to increase the number of spikes is to remove all interactions between neurons. Let the process $\widetilde{Z}^{\textbf{z}}$ be the transformation of the process $Z^{\textbf{z}}$ by replacing signals $\xi^{(k)}_{ji}$, $j\neq i$, by $0$ for $k\geq 1$ (trajectories of $X(t)$ remain the same). The resulting process has a simpler dependence between coordinates and it has a greater number of spikes before any time $t>0$ than that of $Z^{\textbf{z}}$. For our convenience, we want to remove the dependence of the upper bound on $\textbf{z}$ (which is significant because we take $\textbf{z}$ large in the following lemmas) and make the time until the first spike to have the same distribution with the rest of waiting times. Let the process $\bar{Z}$ be the transformation of the process $\widetilde{Z}^{\textbf{z}}$, so that $\bar{Z}_i(0) \overset{d}{=} \xi^{(1)}_{ii}$, $1\leq i \leq N$. Let $\widetilde{\eta}^{\textbf{z}}$ and $\bar{\eta}$ be the number of spikes in processes $\widetilde{Z}^{\textbf{z}}$ and $\bar{Z}$, respectively.
\begin{Lemma}\label{lem:spikes_bound}
We have
$$\eta^{\textbf{z}}_i(t) \leq \widetilde{\eta}^{\textbf{z}}_i(t), \ \ \text{a.s.,}$$
$$\widetilde{\eta}^{\textbf{z}}_i(t) \overset{st.}{\leq} 1 + \bar{\eta}_i(t),$$
and $\bar{\eta}_i(t)$ is an undelayed renewal process, which satisfies the integral renewal theorem and SLLN
$$\frac{\Expect \bar{\eta}_i(t)}{t} \to \frac{\nu_i}{b_{ii}} \overset{a.s.}{\leftarrow} \frac{\bar{\eta}_i(t)}{t}$$
\end{Lemma}
\subsection{Fluid model and corresponding auxiliary results}\label{sect:fluid_model}
Let us define the fluid approximation model. Let $\rho(\textbf{x}, \textbf{y}) = \sum_{i=1}^N |x_i - y_i|$ be the metric on our space $\mathcal{Z}$ and $\|\textbf{x}\| = \rho(\textbf{x}, 0)$, for $\textbf{x}, \textbf{y} \in \mathcal{Z}$. For each $\textbf{z}\in\mathcal{Z}$, introduce a family of scaled processes
$$\widehat{Z}^{\textbf{z}} = \left\lbrace \widehat{Z}^{\textbf{z}}(t) = \frac{Z^{\textbf{z}}(\|\textbf{z}\|t)}{\|\textbf{z}\|}, \ t\geq 0\right\rbrace .$$ We call the family
$$\widehat{Z} = \{\widehat{Z}^\textbf{z}, \ \|\textbf{z}\| \geq 1\}$$
\textit{relatively compact (at infinity)} if, for each sequence $\widehat{Z}^{\textbf{z}_n}$, $\|\textbf{z}_n\| \to \infty$, there exists a subsequence $\widehat{Z}^{\textbf{z}_{n_k}}$ that converges weakly (in Skorokhod topology) to some limit process $\varphi^Z = \{\varphi^Z(t), \ t\geq 0\}$, which is called a \textit{fluid limit}. A family of such limits is called a \textit{fluid model}. The fluid model is \textit{stable} if there exists a finite constant $T$ such that $\|\varphi^Z(T)\| = 0$ a.s. for any fluid limit $\varphi^Z$ (there are several equivalent definitions of stability of a fluid model, see e.g. Stolyar (1995)). Based on stability of a fluid model, one can prove positive recurrence of the original Markov process following the lines of Dai (1995).
Using \lemt{spikes_bound} we prove the next result.
\begin{Lemma}\label{lem:relative_compact}
The family of processes $\{Z^{\textbf{z}}, \ \textbf{z}\in\mathcal{Z}\}$ is such that
\begin{itemize}
\item for all $t>0$ and $\textbf{z}\in\mathcal{Z}$,
$$\Expect\|Z^\textbf{z} (t)\| < \infty$$
and moreover, for any K,
$$\sup_{\|\textbf{z}\|\leq K}\Expect\|Z^\textbf{z} (t)\| < \infty;$$
\item for all $0\leq u < t$, the family of random variables
$$\{\rho(\widehat{Z}^\textbf{z}(u), \widehat{Z}^\textbf{z}(t)), \ \|\textbf{z}\|\geq 1\}$$
is uniformly integrable and there exists a constant $C$ such that
$$\limsup_{\|\textbf{z}\|\to\infty} \PB\{\sup_{u',t'\in [u, t]}\rho(\widehat{Z}^\textbf{z}(u'), \widehat{Z}^\textbf{z}(t')) > C(t-u) \} = 0.$$
\end{itemize}
\end{Lemma}
With this result, one can follow the lines of the proof of Theorem 7.1 from Stolyar (1995) to obtain the following.
\begin{Corollary}
The family of processes $\widehat{Z}$ is relatively compact and every fluid limit $\varphi^Z$ is an a.s. Lipschitz continuous function with Lipschitz constant $C+1$.
\end{Corollary}
Additionally, the function $\varphi^Z(t)$ is a.s. differentiable. We call a time $t_0$ a \textit{regular point} if $\varphi^Z(t)$ is differentiable at $t_0$. Further more, we have
\begin{equation}\label{eq:lipschitz_integration}
\varphi^Z(t) - \varphi^Z(s) = \int_{s}^t \frac{d \varphi^Z}{du}(u) du, \ \ t > s > 0,
\end{equation}
where the derivative is arbitrarily defined (for example, it equals zero) outside regular points.
Let $\widehat{\eta}^{\textbf{z}}(t) = \eta^{\textbf{z}}(\|\textbf{z}\|t) / \|\textbf{z}\|$. Following the lines of the proof of \lemt{relative_compact}, one can prove similar results for the family $\widehat{\eta} = \{\widehat{\eta}^\textbf{z}, \ \|\textbf{z}\| \geq 1\}$. Denote a fluid limit of $\widehat{\eta}$ as $\varphi^{\eta}$. If at time $t$ we have $\varphi^{\eta}_i(t) >0$, then for certain sequence $\textbf{z}_n$ the number of spikes $ \eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t)$ becomes large. If additionally, $ \eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t)$ converges to infinity a.s., then by the law of large numbers
$$\frac{S_{ij}(\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t))}{\|\textbf{z}_n\|} = \frac{\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t)}{\|\textbf{z}_n\|}\frac{S_{ij}(\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t))}{\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t)} \Rightarrow \varphi^\eta_i(t) b_{ij}, \ \text{as $n\to\infty$.}$$
If $\varphi^\eta_i(t) = 0$, then the number of spikes is not as large and, if we prove that the left-hand side of the last equation converges to zero, the resulting convergence will be of the same form.
Using this idea we get the following result.
\begin{Lemma}\label{lem:weak_convergence}
Let $\widehat{\eta}^{\textbf{z}_n}$ converge weakly to a fluid limit $\varphi^{\eta}$ for a sequence $\textbf{z}_n$, $\|\textbf{z}_n\| \to \infty$ as $n\to\infty$. Then we have weak convergence of processes
$$\left(\left(\frac{1}{\|\textbf{z}_n\|}\sum_{i=1}^N S_{ij}(\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t))\right)_{j=1}^N, \ t\geq 0\right) \overset{D}{\Rightarrow} (\varphi^{\eta}(t)B, \ t\geq 0).$$
\end{Lemma}
\section{Proofs of auxiliary results}\label{sect:auxiliary_results}
In this section we prove our auxiliary results for a general matrix $B$ and parameters $\nu_i$.
\subsection{Proof of \lemt{spikes_bound}}
We prove that $T^{\textbf{z}}_{ik} \geq \widetilde{T}^{\textbf{z}}_{ik} $:
\begin{eqnarray*}
T^{\textbf{z}}_{ik} & =&\inf\{t> T^{\textbf{z}}_{i(k-1)}: \ Z^{\textbf{z}}_i(t) \leq 0\} = \inf\{t> T^{\textbf{z}}_{i(k-1)}: \ Z^{\textbf{z}}_i(t) = 0\}\\
& = & \inf\{t> T^{\textbf{z}}_{i(k-1)}: \ z_i + X_i(t) + \sum_{j=1}^N S_{ji}(\eta^{\textbf{z}}_j(t)) = 0\}\\
& = & \inf\{t> T^{\textbf{z}}_{i(k-1)}: \ z_i + X_i(t) + S_{ii}(k-1) + \sum_{j\neq i} S_{ji}(\eta^{\textbf{z}}_j(t)) = 0\}\\
& \geq & \inf\{t> T^{\textbf{z}}_{i(k-1)}: \ z_i + X_i(t) + S_{ii}(k-1) = 0\}.
\end{eqnarray*}
Since $T^{\textbf{z}}_{i0} = \widetilde{T}^{\textbf{z}}_{i0} = 0$, by induction we have
$$T^{\textbf{z}}_{ik} \geq \inf\{t> \widetilde{T}^{\textbf{z}}_{i(k-1)}: \ z_i + X_i(t) + S_{ii}(k-1) = 0\} = \widetilde{T}^{\textbf{z}}_{ik}.$$
Thus, we get $\eta^{\textbf{z}}_i(t) \leq \widetilde{\eta}^{\textbf{z}}_i(t)$. Since $\widetilde{\eta}^{\textbf{z}}_i(t) - 1$ has the same distribution with $\bar{\eta}_i(t-\widetilde{T}^{\textbf{z}}_{i1}) \leq \bar{\eta}_i(t)$, we have the second inequality.
The process $\bar{\eta}_i(t)$ is an undelayed renewal process with waiting times having the same distribution with $\tau_i = \inf\{t> 0: X_i (t) = -\xi^{(1)}_{ii}\}$. Using the strong law of large numbers, one can prove that $\Expect \tau_i = b_{ii}/\nu_i$ (see also Borovkov (1965) for a detailed proof). Therefore, via the standard argument of renewal theory the rest of the proof follows (see e.g. Feller (1971)).
\subsection{Proof of \lemt{relative_compact}}
\textbf{Part 1.} Using \lemt{spikes_bound} and positivity of $\xi^{(k)}_{ij}$, we get
$$
\|Z^\textbf{z} (t)\| = \sum_{i=1}^N |z_i + X_i(t) + \sum_{j=1}^N S_{ji}(\eta^{\textbf{z} }_j(t))| \leq \|\textbf{z}\| + \|X(t)\|
+\sum_{i=1}^N\sum_{j=1}^N S_{ji}(\widetilde{\eta}^{\textbf{z} }_j(t)).
$$
We have
$$
\{\widetilde{\eta}^{\textbf{z} }_i (t) = m\} = \{-\sum_{k=1}^m \xi^{(k)}_{ii} < z_i + \inf_{0\leq s\leq t} X(s) \leq -\sum_{k=1}^{m-1} \xi^{(k)}_{ii}\}
$$
and, therefore,
$$\widetilde{\eta}^{\textbf{z} }_i (t) = \inf\{m \in \mathbb{Z}^+ : \ \sum_{k=1}^m \xi^{(k)}_{ii} > -z_i - \inf_{0\leq s\leq t} X(s)\}.$$
Since $\{\{\xi^{(k)}_{ij}\}_{i, j=1}^N\}_{k=1}^\infty$ and $(X(t), \ t\geq 0)$ are independent, the random variable $\widetilde{\eta}^{\textbf{z} }_i (t)$ is a stopping time for the sequence $\{\{\xi^{(k)}_{ij}\}_{i, j=1}^N\}_{k=1}^\infty$. By Wald's identity,
$$\Expect S_{ji}(\widetilde{\eta}^{\textbf{z} }_j(t)) = \Expect\sum_{k=1}^{\widetilde{\eta}^{\textbf{z} }_j(t)} \xi^{(k)}_{ji} = \Expect\widetilde{\eta}^{\textbf{z} }_j (t)b_{ji} < \infty.$$
\textbf{Part 2.} We have
\begin{equation}\label{eq:distance_t-u}
\begin{split}
\rho(\widehat{Z}^\textbf{z}(u), \widehat{Z}^\textbf{z}(t)) & = \sum_{i=1}^N \frac{|Z^{\textbf{z}}_i(\|\textbf{z}\|t)-Z^{\textbf{z}}_i(\|\textbf{z}\|u)|}{\|\textbf{z}\|} \leq (t-u)\sum_{i=1}^N \nu_i\\
& + \sum_{i=1}^N \frac{|X^0_i(\|\textbf{z}\|t)-X^0_i(\|\textbf{z}\|u)|}{\|\textbf{z}\|}
+ \sum_{i=1}^N\sum_{j=1}^N \frac{S_{ji}(\eta^{\textbf{z}}_j(\|\textbf{z}\|t)) - S_{ji}(\eta^{\textbf{z}}_j(\|\textbf{z}\|u))}{\|\textbf{z}\|}.
\end{split}
\end{equation}
Process $X_i$ is a L\'evy process, from which we have
$$\Expect \frac{|X^0_i(\|\textbf{z}\|t)|}{\|\textbf{z}\|} \leq 2\sup_{0\leq s \leq t}\Expect|X^0_i(s)|, \ \text{for $\|\textbf{z}\| \geq 1$,}$$
and, therefore, the second summand in the right-hand side of \eq{distance_t-u} is uniformly integrable. By \lemt{spikes_bound}, we have
$$S_{ij}(\eta^{\textbf{z}}_j(\|\textbf{z}\|t)) - S_{ij}(\eta^{\textbf{z}}_j(\|\textbf{z}\|u)) \overset{st.}{\leq} S_{ij}(1 + \bar{\eta}_i(\|\textbf{z}\|(t-u))).$$
Since $S_{ij}(n)/n \to b_{ij}$ and $\bar{\eta}_i(\|\textbf{z}\|(t-u)) \to \infty$ a.s., we have
$$\frac{S_{ij}(1 + \bar{\eta}_i(\|\textbf{z}\|(t-u)))}{1 +\bar{\eta}_i(\|\textbf{z}\|(t-u))} \overset{a.s.}{\to } b_{ij},$$
and therefore
\begin{equation*}
\begin{split}
0 & \leq \frac{S_{ji}(\eta^{\textbf{z}}_j(\|\textbf{z}\|t)) - S_{ji}(\eta^{\textbf{z}}_j(\|\textbf{z}\|u))}{\|\textbf{z}\|} \overset{st.}{\leq} \frac{S_{ij}(1 + \bar{\eta}_i(\|\textbf{z}\|(t-u)))}{\|\textbf{z}\|} \\
& = \frac{1 +\bar{\eta}_i(\|\textbf{z}\|(t-u))}{\|\textbf{z}\|}\frac{S_{ij}(1 + \bar{\eta}_i(\|\textbf{z}\|(t-u)))}{1 +\bar{\eta}_i(\|\textbf{z}\|(t-u))} \to (t-u)\frac{\nu_i}{b_{ii}}b_{ij}
\end{split}
\end{equation*}
a.s. and in $L_1$, as $\|\textbf{z}\|\to\infty$. Then the distance on the left-hand side of \eq{distance_t-u} is bounded above by the sum of uniformly integrable random variables and therefore is also uniformly integrable.
Given
$$C>\sum_{i=1}^N\nu_i\left( 1 + \sum_{j=1}^N \frac{b_{ij}}{b_{ii}}\right) ,$$ there exists $\varepsilon >0$ such that for $\|\textbf{z}\|$ large
\begin{equation*}
\begin{split}
\PB\{\sup_{u',t'\in [u, t]}\rho(\widehat{Z}^\textbf{z}(u'), \widehat{Z}^\textbf{z}(t')) > C(t-u) \} & \leq \PB\left\{\sup_{u',t'\in [u, t]}\sum_{i=1}^N \frac{|X^0_i(\|\textbf{z}\|t')-X^0_i(\|\textbf{z}\|u')|}{\|\textbf{z}\|} > \varepsilon(t-u) \right\}\\
& \leq 2N \PB\left\{ \sup_{s\in [0, t-u]}\frac{|X^0_1(\|\textbf{z}\|s)|}{\|\textbf{z}\|} > \frac{\varepsilon}{2N}(t-u) \right\} \to 0,
\end{split}
\end{equation*}
by Theorem 36.8 from Sato (1999).
\subsection{Proof of \lemt{weak_convergence}}
By Skorokhod (1956), it is sufficient to prove that there is a convergence of finite-dimensional distributions on everywhere dense set of times $t$ and that a tightness condition holds. Tightness can be deduced from the second statement of \lemt{relative_compact}. We prove that
\begin{equation}\label{eq:finite_dimensional_convergence}
\PB\left\{\bigcap_{k=1}^K \bigcap_{i, j =1}^N \left\{\frac{S_{ij}(\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t_k))}{\|\textbf{z}_n\| } < y^k_{ij}\right\}\right\} \to \PB\left\{\bigcap_{k=1}^K \bigcap_{i =1}^N \left\{\varphi^{\eta}_i(t_k) < \min_{1\leq j\leq N}\frac{y^k_{ij}}{b_{ij}}\right\}\right\}
\end{equation}
as $n\to\infty$, for appropriate $t\geq 0$ and $\textbf{y}\in (0, \infty)^{K N^2}$.
Define sets
$$C_{ij}^k(n) = \left\{ \frac{S_{ij}(\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t_k))}{\|\textbf{z}_n\| } < y^k_{ij} \right\},$$
$$D^k_i (n, m) = \{\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t_k) > m\},$$
$$E^k_{ij} (n, \delta) = \left\{\left| \frac{S_{ij}(\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t_k))}{\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t_k)} - b_{ij} \right| \leq \delta \right\},$$
$$F^{k \pm}_{i} (n, \delta) = \left\{\frac{\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t_k)}{\|\textbf{z}_n\|} < \min_{1 \leq j \leq N} \frac{y^k_{ij}}{b_{ij} \mp \delta} \right\}, $$
where $\delta \in (0, \min_{i,j} b_{ij})$. We prove that
$$\PB\left\{F^{k -}_{i} (n, \delta)\right\} + o(1) \leq \PB\left\{\bigcap_{j=1}^N C_{ij}^k(n)\right\} \leq \PB\left\{F^{k +}_{i} (n, \delta)\right\} + o(1), \ \text{as $n\to \infty$.}$$
For any $\textbf{y}\in (\mathbb{R}^+)^{K N^2}$ such that $(\min_j (y^k_{ij}/b_{ij}))_{i=1}^N$ is a continuity point of the cdf of $(\varphi^{\eta}(t_k))_{k=1}^k$, there is a neighbourhood $\Delta$ of $\textbf{y}$ such that every point $\textbf{x} \in \Delta$ is also a continuity point. Thus, for $\delta$ small we have
\begin{equation*}
\begin{split}
\PB\left\{\bigcap_{k=1}^K \bigcap_{i=1}^N F^{k \pm}_{i} (n, \delta)\right\} & = \PB\left\{\bigcap_{k=1}^K\bigcap_{i =1}^N\left\{\frac{\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t_k)}{\|\textbf{z}_n\| } < \min_{1 \leq j \leq N} \frac{y^k_{ij}}{b_{ij} \mp \delta}\right\}\right\} \\
& \to \PB\left\{\bigcap_{k=1}^K\bigcap_{i =1}^N \left\{\varphi^{\eta}_i(t_k) < \min_{1\leq j\leq N}\frac{y^k_{ij}}{b_{ij} \mp \delta}\right\}\right\}, \ \text{as $n \to \infty$,}
\end{split}
\end{equation*}
and, therefore, by letting $\delta$ converge to $0$, we get \eq{finite_dimensional_convergence}.
By the law of large numbers, we have
$$\PB\{D^k_i (n, m) \cap \overline{E^k_{ij} (n, \delta)}\} \to 0, \ \ \text{as $m\to\infty$,}$$
and
$$\PB\{\overline{C_{ij}^k(n)} \cap \overline{D^k_i (n, m)}\} \to 0, \ \ \text{as $n\to\infty$,}$$
if $m=o(\|\textbf{z}_n\|)$. Take $m=\sqrt{\|\textbf{z}_n\| }$.
From the definitions we have
\begin{multline*}
\left( \bigcap_{j=1}^N C_{ij}^k(n) \cap D^k_i (n, m) \cap E^k_{ij} (n, \delta)\right) \subseteq \left( F^{k +}_{i} (n, \delta) \cap D^k_i (n, m) \cap E^k_{ij} (n, \delta)\right)\\
= \left( F^{k +}_{i} (n, \delta) \cap D^k_i (n, m)\right) \setminus \left( F^{k +}_{i} (n, \delta) \cap D^k_i (n, m) \cap \overline{E^k_{ij} (n, \delta)}\right)
\end{multline*}
and
$$ \left( F^{k +}_{i} (n, \delta) \cap D^k_i (n, m)\right) = F^{k +}_{i} (n, \delta) \setminus \left( F^{k +}_{i} (n, \delta) \cap \overline{D^k_i (n, m)}\right).$$
Since $m=o(\|\textbf{z}_n\|)$, we have $F^{k +}_{i} (n, \delta) \cap \overline{D^k_i (n, m)} = \overline{D^k_i (n, m)}$ for $n$ large. Combining altogether, we get
\begin{equation*}
\begin{split}
\PB\left\{\bigcap_{j=1}^N C_{ij}^k(n)\right\} & = \PB\left\{\bigcap_{j=1}^N C_{ij}^k(n) \cap D^k_i (n, m)\right\} +
\PB\left\{\bigcap_{j=1}^N C_{ij}^k(n) \cap \overline{D^k_i (n, m)}\right\} \\
& \leq \PB\left\{F^{k +}_{i} (n, \delta)\right\} - \PB\left\{\overline{D^k_i (n, m)}\right\} + \PB\left\{\bigcap_{j=1}^N C_{ij}^k(n) \cap \overline{D^k_i (n, m)}\right\} + o(1),
\end{split}
\end{equation*}
as $n\to\infty$, and
\begin{multline*}
\PB\left\{\bigcap_{j=1}^N C_{ij}^k(n) \cap \overline{D^k_i (n, m)}\right\} - \PB\left\{\overline{D^k_i (n, m)}\right\} = \PB\left\{\bigcup_{j=1}^N \overline{C_{ij}^k(n)} \cap \overline{D^k_i (n, m)}\right\} \to 0,
\end{multline*}
as $n\to\infty$. Following the same lines with replacing a set $F^{k+}_i(n, \delta)$ with a set $F^{k-}_i(n, \delta)$ and relations $\subseteq$ and $\leq$ with relations $\supseteq$ and $\geq$, we get the lower bound.
\section{Proof of \lemt{positive_recurrence}}\label{sect:lemma_positive_recurrence}
We prove that under condition \eq{simple_network} fluid limits $\varphi^Z(t)$ are deterministic and uniquely defined by initial value $\varphi^{Z}(0)$. Further, each coordinate of a fluid limit is a continuous piecewise linear function which tends to zero and then remains there.
Let sequence $\textbf{z}_n$, $\|\textbf{z}_n\| \to \infty$, be such that
$$\widehat{Z}^{\textbf{z}_n} \overset{\mathcal{D}}{\Rightarrow} \varphi^Z \ \text{and} \ \widehat{\eta}^{\textbf{z}_n} \overset{\mathcal{D}}{\Rightarrow} \varphi^\eta.$$
By Corollary $1$, function $\varphi^Z$ is a.s. Lipschitz continuous.
Following the lines of the proof of \lemt{weak_convergence}, one can easily show that
$$\left(\widehat{Z}^{\textbf{z}_n}(t) - \frac{\textbf{z}_n}{\|\textbf{z}_n\|} - \frac{X(\|\textbf{z}_n\|t)}{\|\textbf{z}_n\|}, \ t\geq 0 \right) \overset{\mathcal{D}}{\Rightarrow} (\varphi^Z(t) - \varphi^Z(0) + \nu t \textbf{1} , \ t\geq 0).$$
Now, given \eq{dynamic_equation} and \lemt{weak_convergence}, we have
$$(\varphi^\eta(t) B, \ t\geq 0) \overset{d}{=} (\varphi^Z(t) - \varphi^Z(0) + \nu t \textbf{1} , \ t\geq 0).$$
By \rem{spike_rate}, the matrix $B$ is invertible and we have
$$(\varphi^\eta(t), \ t\geq 0) \overset{d}{=} \left( \left( \varphi^Z(t) - \varphi^Z(0) + \nu t \textbf{1}\right) B^{-1}, \ t\geq 0\right) .$$
Since $\varphi^\eta$ is a weak limit, we can assume without loss of generality $\varphi^\eta(t) = \left( \varphi^Z(t) - \varphi^Z(0) + \nu t \textbf{1}\right) B^{-1}$. Thus, $\varphi^\eta$ is differentiable wherever $\varphi^Z$ is.
Assume that $\|\varphi^Z(t_0)\| > 0$ and $t_0$ is a regular point (see \sectn{fluid_model}). Let $N_0 = \sharp\{i: \ \varphi^Z_i(t_0) = 0\} < N$. Then, with a proper reordering, $ \varphi^Z_i(t_0) = 0$, for $i\in\{1, \ldots, N_0\}$ and $ \varphi^Z_i(t_0) > 0$, for $i\in\{N_0 + 1, \ldots, N\}$. Since $\varphi^{Z}_i(t) \geq 0$ and $t_0$ is a regular point, from $\varphi^{Z}_i(t_0) = 0$ we get $(\varphi^{Z}_i)'(t_0) = 0$. We find the values of
$$(\varphi^Z_i)'(t_0) = -\nu + H_i(\varphi^\eta_i)'(t_0) + \sum_{j\neq i}w_{j}(\varphi^\eta_j)'(t_0).$$
We prove that $(\varphi^\eta_i)'(t_0) = 0$ for $i> N_0$ (if a potential is very far from the threshold then the neuron does not have a spike for a long time) and, therefore,
\begin{equation}\label{eq:zero_level_fluid}
0 = -\nu + (H_i-w_i)(\varphi^\eta_i)'(t_0)+ \sum_{j=1}^{N_0}w_{j}(\varphi^\eta_j)'(t_0), \ \ i=1,\ldots, N_0,
\end{equation}
\begin{equation}\label{eq:positive_level_fluid}
(\varphi^Z_i)'(t_0) = -\nu + \sum_{j=1}^{N_0}w_j(\varphi^\eta_j)'(t_0), \ \ i=N_0+1, \ldots, N.
\end{equation}
Let
$$h = \min_{N_0 + 1 \leq i \leq N} \varphi^Z_i(t_0).$$
We prove that for any $\Delta < h/(4\nu)$ and $i\in[N_0 +1, N]$ equality $\varphi^{\eta}_i(t_0 + \Delta) = \varphi^{\eta}_i(t_0)$ holds. Since $\widehat{Z}_i^{\textbf{z}_n}(t_0) \Rightarrow \varphi^Z_i (t_0)$, we have $\widehat{Z}_i^{\textbf{z}_n}(t_0) > h/2 > 2\nu\Delta$ a.s. for $n$ large. We have
\begin{equation*}
\begin{split}
\PB\{\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|(t_0 + \Delta)) > \eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t_0) \} & \leq \PB\{2\nu\Delta\|\textbf{z}_n\| + \inf_{0\leq s\leq \Delta} (X_i(\|\textbf{z}_n\|(t_0 + s)) - X_i(\|\textbf{z}_n\|t_0)) \leq 0\}\\
& \leq \PB\{\nu\Delta\|\textbf{z}_n\| + \inf_{0\leq s\leq \Delta} (X^0_i(\|\textbf{z}_n\|(t_0 + s)) - X^0_i(\|\textbf{z}_n\|t_0)) \leq 0\}\\
& = \PB\{\sup_{0\leq s\leq \Delta} X^0_i(\|\textbf{z}_n\|s) \geq \nu\Delta\|\textbf{z}_n\|\}.
\end{split}
\end{equation*}
Thus, by Theorem 36.8 from Sato (1999), we have convergence $\eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|(t_0 + \Delta)) - \eta^{\textbf{z}_n}_i(\|\textbf{z}_n\|t_0) \to 0$ in probability and convergence $\widehat{\eta}^{\textbf{z}_n}_i(t_0 + \Delta) - \widehat{\eta}^{\textbf{z}_n}_i(t_0) \to 0$ a.s., as $n\to\infty$. Thus, equality $\varphi^{\eta}_i(t_0 + \Delta) = \varphi^{\eta}_i(t_0)$ holds for $\Delta < h/(4\nu)$ and $(\varphi^{\eta}_i)'(t_0) = 0$.
Using \rem{spike_rate} we solve system \eq{zero_level_fluid} and get
$$(\varphi^\eta_i)'(t_0) = \frac{\nu}{H_i - w_i} \frac{1}{1 + \sum_{k=1}^{N_0}\frac{w_k}{H_k - w_k}}, \ \ i=1,\ldots, N_0,$$
and therefore,
$$(\varphi^Z_i)'(t_0) = -\nu + \frac{\nu}{1 + \sum_{k=1}^{N_0}\frac{w_k}{H_k - w_k}} \sum_{j=1}^{N_0}\frac{w_j}{H_j - w_j} = -\frac{\nu}{1 + \sum_{k=1}^{N_0}\frac{w_k}{H_k - w_k}}, \ \ i=N_0+1, \ldots, N.$$
Therefore, the process $\varphi^Z$ is deterministic and piecewise linear. We have
$$\varphi_i^Z(0)\leq 1 \ \text{and} \ (\varphi^Z_i)'(t_0) \leq -\frac{\nu}{1 + \sum_{k=1}^{N}\frac{w_k}{H_k - w_k}}, \ \ i=1,\ldots, N,$$
for any regular point $t_0$ such that $\varphi_i^Z(t_0) >0$. Thus, from \eq{lipschitz_integration} we have that in time interval\\ $(0, \nu^{-1} (1 + \sum_{k=1}^{N}\frac{w_k}{H_k - w_k}))$ process $\varphi^Z$ reaches zero and stays there.
Let $\tau^{\textbf{z}}(\varepsilon, B) = \inf\{t> \varepsilon: \ Z^{\textbf{z}}(t) \in B\}$. Since fluid limits are stable, there exists $\kappa >0$ such that for $V = \{\textbf{z}\in \mathcal{Z}: \|\textbf{z}\| < k_0\}$ we have
$$\sup_{\textbf{z}\in V} \Expect\tau^{\textbf{z}}(\varepsilon, V) < \infty.$$
\section{Proof of \lemt{minorant_measure}}\label{sect:lemma_mixing}
We prove existence of a lower bound for $\inf_{\textbf{z} \in V} \PB\{Z^{\textbf{z}}(U) \in B\}$ where\\
$V= \{\textbf{z}\in \mathcal{Z}: \|\textbf{z}\| < k_0\}$ (see the end of previous section). By Theorem 19.2 from Sato (1999), the L\'evy process $X(t)$ can be represented as a sum $X^1(t) + X^2(t)$ of two independent processes, a jump process $X^1(t)$ and a Gaussian process $X^2(t)$ with drift. We consider cases where at least one coordinate is close enough to zero. If all the coordinates of $\textbf{z}$ are bounded away from zero then the proof follows similar lines.
Since random variables $\xi^{(1)}_{ij}$, $i, j\in [1, N]$, are strictly positive, there are constants $k^+_1, k^-_1 >0$ such that
$$p_1 = \PB\{A_1\} \equiv \PB\left\{(\xi^{(1)}_{ij})_{i, j=1}^N \in [k^-_1, k^+_1]^{N^2}\right\} >0.$$
Without loss of generality, we assume that $z_1 < k^-_1/6$.
First, we bound the jump process $X^1(t)$ in the time interval $[0, 2]$, which includes the time interval $[0, U]$, and take time instant $t_0 \leq 1/2$ such that $\PB\{X^1_1(t_0) < k^-_1/6\} >0$. Denote
$$A_2 = \left\{\max_{1\leq i \leq N} (X^1_i(2)) < k_2\right\} \cap \left\{X^1_1(t_0) < k^-_1/6\right\},$$
and take a constant $k_2 >0$ such that $p_2 = \PB\{A_2\} >0.$ Next, we use the condition that the Gaussian process $X^2(t)$ is non-degenerate and none of its coordinates is a deterministic line. Thus, we denote
\begin{equation*}
\begin{split}
A_3 & = \left\{\max_{1\leq i \leq N} \sup_{0\leq t \leq \frac{1}{2}} (X^2_i(t)) \leq k_3-k_2 - k^-_1\right\} \cap \left\{\min_{1\leq i \leq N} \inf_{0\leq t \leq \frac{1}{2}} (X^2_i(t)) \geq -\frac{k^-_1}{2}\right\}\\
& \cap \left\{\inf_{0\leq t \leq t_0} (X^2_1(t)) \leq -\frac{k^-_1}{3}\right\}
\end{split}
\end{equation*}
and take a constant $k_3 > k_2 - k^-_1$ such that $p_3 = \PB\{A_3\} >0$. One can show that, given $A_2 \cap A_3$, the first spike occurs up to time $t_0$ and the second one can occur only after time $1/2$.
Denote the new set $D = A_1 \cap A_2 \cap A_3$. From independence of $X^1$, $X^2$ and $(\xi^{(1)}_{ij})_{i, j =1}^N$ we have $\PB\{D\} = p_1p_2p_3 >0$. We have
\begin{equation}\label{eq:away_from_zero_after_spike}
D \subseteq \left\{\frac{k^-_1}{2} \leq Z_i^{\textbf{z}}\left( \frac{1}{2}\right) + X_i^1(U) - X_i^1\left( \frac{1}{2}\right) \leq k^+_1 + k_2 + k_3, \ i = 1, \ldots, N\right\}.
\end{equation}
We restrict ourselves to events without the second spike up to time $U$. Denote
$$G_2(t_1, t_2, k) = \left\{\min_{1\leq i \leq N} \inf_{t_1\leq s \leq t_2} (X^2_i(s)) > -k\right\}.$$
Using \eq{away_from_zero_after_spike}, we get that, given $G_2\left( 1/2, U, k^-_1/2\right) \cap D$, the second spike occurs after time $U$.
Let $K = k^+_1 + k_2 + k_3$. We prove that, for any point $\textbf{y} \in (K, 2K)^N$ and a measurable set $\Delta \subset [0, K]^N$, there is a number $p>0$ such that
$$
\PB\left\{\left\{Z^{\textbf{z}}(U) \in \textbf{y} + \Delta\right\} \cap G_2\left( \frac{1}{2}, U, \frac{k^-_1}{2}\right) \cap D \right\} \\
\geq p \lambda(\Delta),
$$
where $\lambda$ is the Lebesgue measure. Denote, $\widehat{y}_i = y_i - Z_i^{\textbf{z}}\left( 1/2\right) - \left( X_i^1(U) - X_i^1\left( 1/2\right)\right) $, for $i\in[1, N]$. Then we have
$$\left\{Z^{\textbf{z}}(U) \in \textbf{y} + \Delta\right\} = \left\{ X^2(U) - X^2\left( \frac{1}{2} \right) \in\widehat{\textbf{y}} + \Delta\right\}.$$
Since $X^2$ is a Markov process, the events $ \left\{ X^2(U) - X^2\left( 1/2 \right) \in\widehat{\textbf{y}} + \Delta\right\} \cap G_2\left( 1/2, U, k^-_1/2\right)$ and $D$ are independent, conditioned on a value of $\widehat{\textbf{y}}$. Thus, we have
\begin{equation*}
\begin{split}
&\PB\left\{\left\{ X^2(U) - X^2\left( \frac{1}{2} \right) \in\widehat{\textbf{y}} + \Delta\right\} \cap G_2\left( \frac{1}{2}, U, \frac{k^-_1}{2}\right) \cap D \right\}\\
& = \Expect\left( \PB\left\{\left\{ X^2(U) - X^2\left( \frac{1}{2} \right) \in\widehat{\textbf{y}} + \Delta\right\} \cap G_2\left( \frac{1}{2}, U, \frac{k^-_1}{2}\right) \mid \ \widehat{\textbf{y}}\right\} \PB\left\{D \mid \ \widehat{\textbf{y}}\right\}\right)
\end{split}
\end{equation*}
Next, we need a technical lemma regarding a monotonicity property of the Brownian bridge.
\begin{Lemma}\label{lem:brownian_bridge}
For any $t, k>0$ and $\Delta \subset [0, \infty)^N$ we have
$$\PB\{G_2(0, t, k) \mid \ X^2(t) \in \Delta\} \geq \PB\{G_2(0, t, k) \mid \ X^2(t) = \textbf{0}\} >0.$$
\end{Lemma}
Given $D$, we have $\widehat{y}_i \geq 0$, for $i \in [1, N]$, and we can use \lemt{brownian_bridge} to obtain
\begin{multline*}
\PB\left\{\left\{ X^2(U) - X^2\left( \frac{1}{2} \right) \in\widehat{\textbf{y}} + \Delta\right\} \cap G_2\left( \frac{1}{2}, U, \frac{k^-_1}{2}\right) \mid \ \widehat{\textbf{y}}\right\}\\
\overset{a.s.}{\ge} \PB\left\{ X^2(U) - X^2\left( \frac{1}{2} \right) \in\widehat{\textbf{y}} + \Delta\mid \ \widehat{\textbf{y}}\right\} \PB\left\lbrace G_2\left( \frac{1}{2}, U, \frac{k^-_1}{2}\right) \mid \ X^2\left( U\right) -X^2\left( \frac{1}{2}\right) = \textbf{0} \right\rbrace.
\end{multline*}
The density of $X^2(t)$ is bounded away from zero on any compact set, and $X^2$ and $U$ are independent. Therefore, there exists $p_4 >0$ such that, given $D$, for a measurable set $\Delta \subseteq [0, K]^N$ we have
$$\PB\left\{ X^2(U) - X^2\left( \frac{1}{2} \right) \in\widehat{\textbf{y}} + \Delta\mid \ \widehat{\textbf{y}}\right\} \overset{a.s.}{\geq} p_4\lambda(\Delta).$$
Denote $p'_4 = p_4\PB\left\lbrace G_2\left( \frac{1}{2}, U, \frac{k^-_1}{2}\right) \mid \ X^2\left( U\right) -X^2\left( \frac{1}{2}\right) = \textbf{0} \right\rbrace >0$. Combining altogether, we get that if $z_1 < k^-_1/6$ then
$$\PB\{Z^{\textbf{z}}(U) \in \textbf{y} + \Delta\} \geq p_1p_2p_3p'_4\lambda(\Delta),$$
for $\textbf{y} \in (K, 2K]$ and $\Delta \subseteq [0, K]^N$.
|
\section{Introduction}\label{sec:intro}
\subsection{Results}\label{sec:results}
Let $T: X \to X$ be a continuous map of a compact metric space $(X,d)$. In the topological dynamical system $(X,T)$, the \emph{set of return times} of a point $x \in X$ to a non-empty, open set $U \subseteq X$ is
\[R(x,U) \vcentcolon= \{n \in \mathbb{N} \ | \ T^n x \in U \}.\]
Much is known about the relationship between dynamical properties of the system $(X,T)$ and the additive combinatorial properties of the sets $R(x,U)$. For example, if $(X,T)$ is \emph{minimal} (that is, for all $x \in X$, the set $\{T^n x \ | \ n \in \mathbb{N}\}$ is dense in $X$),
then every set of return times $R(x,U)$ is \emph{syndetic}, meaning that there exists $N \in \mathbb{N}$ such that $R(x,U)$ has non-empty intersection with every interval of $N$ consecutive positive integers. This connection between dynamics and additive combinatorics has had a strong influence in Ramsey Theory; we discuss some of the history behind this connection and put our main results into context in \cref{sec:historyandcontext}.
In this paper, we consider the relationship between the dynamical properties of the system $(X,T)$ and the \emph{multiplicative} combinatorial properties of the sets $R(x,U)$. Our first main result concerns geometric progressions, configurations of the form $\{nm,nm^2,\ldots,nm^\ell\}$, in sets of return times in minimal systems.
\begin{theorem}\label{thm:mainthm}
Let $(X,T)$ be a minimal dynamical system. There exists a residual set $X' \subseteq X$ such that for all $x \in X'$ and all non-empty, open $U \subseteq X$, the set $R(x,U)$ contains arbitrarily long geometric progressions.
\end{theorem}
We are able to strengthen the conclusion of \cref{thm:mainthm} in special classes of dynamical systems. The system $(X,T)$ is \emph{totally minimal} if for all $n \in \mathbb{N}$, the system $(X,T^n)$ is minimal.
Our main result for totally minimal systems makes use of the \emph{multiplicative upper Banach density}, defined for $A \subseteq \mathbb{N}$ by
\begin{align}\label{eqn:densityinseminplusintromult}
d_\times^*(A) \vcentcolon= \limsup_{n \to \infty} \ \max_{m \in \mathbb{N}} \ \frac{|A \cap \{m p_1^{e_1} \cdots p_n^{e_n} \ | \ e_1, \ldots, e_{n} \in \{1, \ldots, n\} \}|}{n^n},
\end{align}
where $(p_n)_{n \in \mathbb{N}}$ is an enumeration of the primes. This density, introduced and studied by Bergelson \cite{bmultlarge}, is the multiplicative analogue of the additive upper Banach density in $\mathbb{N}$. It is independent of the chosen enumeration of the primes; see \cref{def:multdensity} for an equivalent definition of $d_\times^*$ and the remark following it.
\begin{theorem}\label{thm:mainthmwithtotallymin}
Let $(X,T)$ be a totally minimal dynamical system. There exists a residual set $X' \subseteq X$ such that for all non-empty, open $U \subseteq X$, there exists $\eta > 0$ such that for all $x \in X'$, the set $R(x,U)$ satisfies $d_\times^*(R(x,U)) \geq \eta$.
\end{theorem}
Szemer\'edi's theorem \cite{szemeredi} on arithmetic progressions can be used to prove that any set of positive multiplicative upper Banach density contains arbitrarily long geometric progressions; see \cref{thm:multszem}. Therefore, \cref{thm:mainthmwithtotallymin} strengthens \cref{thm:mainthm} when the system $(X,T)$ is totally minimal. In fact, sets of positive multiplicative upper Banach density contain \emph{geo-arithmetic configurations}, combinatorial configurations of the form $\big\{ c (a+id)^j \ \big| \ 1 \leq i, j \leq \ell \big\}$ that are much richer than simply geometric progressions; see \cref{thm:geoarithmeticpatterns}.
Without the assumption of total minimality, \emph{local obstructions} appear that prevent return time sets from having positive multiplicative density.
For instance, the set $4\mathbb{N}-2$ is a set of return times in a four point rotation (which is minimal but not totally minimal), but it has zero multiplicative upper Banach density. The set $4\mathbb{N}-2$ is, however, multiplicatively large in a different sense: it is a coset of the multiplicative subsemigroup $2\mathbb{N}-1$.
We resolve local obstructions by measuring multiplicative density not along $\mathbb{N}$, but along cosets of multiplicative subsemigroups of $\mathbb{N}$. A \emph{multiplicative subsemigroup} of $\mathbb{N}$ is a subset $S \subseteq \mathbb{N}$ that is closed under multiplication, and a \emph{coset} of $S$ is a set of the form $nS$ for $n \in \mathbb{N}$. The multiplicative upper Banach density $d^*_{nS}$ for subsets of $nS$ can be defined analogously to $d_\times^*$ in (\ref{eqn:densityinseminplusintromult}) (using dilates of so-called \emph{F{\o}lner sequences} in $S$) or as in \cref{def:multdensity}. For convenience, when $A \subseteq \mathbb{N}$, we write $d^*_{nS}(A)$ to mean $d^*_{nS}(A \cap nS)$.
For the special class of distal systems, we prove an analogue of \cref{thm:mainthmwithtotallymin} without the assumption of total minimality. A system $(X,T)$ is called \emph{distal} if for all $x, y \in X$ with $x \neq y$, $\inf_{n \in \mathbb{N}} d(T^n x, T^n y) > 0$. Distal systems encompass limits of group extensions of group rotations and form important building blocks in the various structure theories of minimal dynamical systems.
The following theorem shows that the local obstructions described above are the only types of obstructions to positive multiplicative density in distal systems.
\begin{theorem}\label{thm:mainthmwithdistal}
Let $(X,T)$ be a minimal distal system. There exists a residual set $X' \subseteq X$ such that for all non-empty, open $U \subseteq X$, there exists $\eta > 0$ such that for all $x \in X'$, there exists a multiplicative subsemigroup $S$ of $\mathbb{N}$ and $n \in \mathbb{N}$ such that the set $R(x,U)$ satisfies $d_{nS}^* \big( R(x,U) \big) \geq \eta$.
\end{theorem}
Sets with positive multiplicative density along a coset of a multiplicative subsemigroup contain an abundance of multiplicative configurations, including arbitrarily long geometric progressions and geo-arithmetic configurations; see \cref{thm:multszem,thm:geoarithmeticpatterns}.
Our final main result is purely combinatorial but indirectly concerns nilsystems, a subclass of distal systems that encompasses algebraic group extensions of group rotations.\footnote{A \emph{nilsystem} is a topological dynamical system $(X,T)$ where $X$ is a compact homogeneous space of a nilpotent Lie group $G$ and $T$ is a translation of $X$ by an element of $G$.} A subset of $\mathbb{N}$ is called \emph{$\text{IP}_r$}, $r \in \mathbb{N}$, if it contains a set of the form
\begin{align}\left\{ \sum_{i \in I} x_i \ \middle| \ \emptyset \neq I \subseteq \{1, \ldots, r\} \right\}, \quad x_1, \ldots, x_r \in \mathbb{N}.\end{align}
A subset of $\mathbb{N}$ is called \emph{$\text{IP}_r^*$} if it has non-empty intersection with every $\text{IP}_r$ set in $\mathbb{N}$. Such sets arose first in the work of Furstenberg and Katznelson \cite{furstenbergkatznelsonipszem} on the multidimensional IP Szemer\'{e}di theorem and were recently used by Bergelson and Leibman \cite{BLiprstarcharacterization} to characterize nilsystems: roughly speaking, a system $(X,T)$ is a nilsystem if and only if for all non-empty, open $U \subseteq X$, there exists $r \in \mathbb{N}$ such that for every $x \in U$, the set $R(x,U)$ is $\text{IP}_r^*$.
The following theorem addresses the multiplicative properties of additive translates of $\text{IP}_r^*$ sets.
\begin{theorem}\label{thm:translatesofiprstararelarge}
Let $A \subseteq \mathbb{N}$ be an $\text{IP}_r^*$ set. For all $t \in \mathbb{Z}$, there exists a multiplicative subsemigroup $S$ of $\mathbb{N}$ and $n \in \mathbb{N}$ such that $d_{nS}^* \big(A + t \big) > 0$.
\end{theorem}
Whereas the previous results rely on tools and techniques from dynamics, the statement and proof of \cref{thm:translatesofiprstararelarge} are entirely combinatorial. This imparts two advantages: we avoid the machinery necessary to work with nilsystems, and the result concerns a wider class of sets.\footnote{While every set of the form $R(x,U)$ in a minimal nilsystem is $\text{IP}_r^*$ for some $r \in \mathbb{N}$, not every $\text{IP}_r^*$ set in $\mathbb{N}$ contains a set of return times from a minimal nilsystem; see \cref{{ex:propertyofnilsystem}}.} In particular, our result implies that return time sets in minimal nilsystems contain arbitrarily large geo-arithmetic configurations, hence arbitrarily long geometric progressions. When applied to sets of natural numbers that arise in polynomial Diophantine approximation, \cref{thm:translatesofiprstararelarge} yields the following corollary. Denote by $\{x\}$ the fractional part of $x \in \mathbb{R}$.
\begin{corollary}\label{cor:polyapprox}
Let $p_1, \ldots, p_k \in \mathbb{R}[x]$ be non-constant polynomials that are linearly independent in the following sense: for all $h_1, \ldots, h_k \in \mathbb{Z}$, not all zero, at least one of the non-constant coefficients of $\sum_{i=1}^k h_i p_i$ is irrational. Let $I_1, \ldots, I_k \subseteq [0,1)$ be sets that are open when $[0,1)$ is identified with the 1-torus. The set
\[A \vcentcolon= \big\{ n \in \mathbb{Z} \ \big| \ \text{for all } i \in \{1,\ldots,k\}, \ \{p_i(n)\} \in I_i \big\}\]
has positive multiplicative upper Banach density in a coset of a multiplicative subsemigroup of $\mathbb{N}$. As a consequence, for all $n \in \mathbb{N}$, there exist $a,c,d \in \mathbb{N}$ such that $\big\{ c (a+id)^j \ \big| \ 1 \leq i, j \leq n \big\} \subseteq A$.
\end{corollary}
\begin{comment}
The following corollary is an application of \cref{thm:translatesofiprstararelarge} to sets of natural numbers that arise in polynomial Diophantine approximation. Denote by $\{ x \}$ the fractional part of $x \in \mathbb{R}$.
\begin{corollary}\label{cor:polyapprox}
Let $p_1, \ldots, p_k \in \mathbb{R}[x]$ be non-constant polynomials that are linear independent in the following sense: for all $h_1, \ldots, h_k \in \mathbb{Z}$, not all zero, at least one of the non-constant coefficients of $\sum_{i=1}^k h_i p_i$ is irrational. Let $I_1, \ldots, I_k \subseteq [0,1)$ be sets that are open when $[0,1)$ is identified with the 1-torus. The set
\[A \vcentcolon= \big\{ n \in \mathbb{Z} \ \big| \ \text{for all } i \in \{1,\ldots,k\}, \ \{p_i(n)\} \in I_i \big\}\]
has positive multiplicative upper Banach density in a coset of a multiplicative subsemigroup of $\mathbb{N}$. Moreover, for all $n \in \mathbb{N}$, there exist $a,c,d \in \mathbb{N}$ such that $\big\{ c (a+id)^j \ \big| \ 1 \leq i, j \leq n \big\} \subseteq A$.
\end{corollary}
\end{comment}
\subsection{Motivation and historical context} \label{sec:historyandcontext}
Van der Waerden's theorem on arithmetic progressions \cite{vdw} is one of the most celebrated results in Ramsey Theory. An equivalent formulation due to Kakeya and Morimoto \cite[Theorem I]{kakeyamorimoto} states that every syndetic subset of $\mathbb{N}$ contains arbitrarily long arithmetic progressions. The multiplicative analogue of this result asserts that every \emph{multiplicatively syndetic} subset of $\mathbb{N}$ (that is, a set $A \subseteq \mathbb{N}$ for which there exists $N \in \mathbb{N}$ such that $A \cup A/2 \cup \cdots \cup A/N = \mathbb{N}$, where $A/n \vcentcolon= \{m \in \mathbb{N} \ | \ mn \in A\}$) contains arbitrarily long \emph{geometric progressions}.
The following long-standing open problem in Ramsey Theory features both of these additive and multiplicative notions and is the primary motivation for our work.
\begin{question}[{\cite{bbhsaddimpliesmult}}]\label{question:mainquestion}
Does every additively syndetic subset of $\mathbb{N}$ contain arbitrarily long geometric progressions?
\end{question}
Just as in other problems in Ramsey theory involving both addition and multiplication -- most notably the $\{x+y,xy\}$ problem that was recently resolved in \cite{moreiraxyproblem} -- analysis is complicated by the combination of addition and multiplication. Until now, very little progress has been made on \cref{question:mainquestion}; in fact, it is still unknown whether or not syndetic subsets of $\mathbb{N}$ contain a square integer ratio. Recent work in \cite{patilgeomsyndetic} addresses the set of integer ratios of elements of syndetic sets.
Being unable to make progress on the problem in its full generality, it is natural to restrict the class of syndetic subsets under consideration. Each of our main results concerns such a restriction: \cref{thm:mainthm} lends some evidence toward a positive answer to \cref{question:mainquestion} by showing that many syndetic sets of dynamical origin contain arbitrarily long geometric progressions; \cref{thm:mainthmwithtotallymin,thm:mainthmwithdistal} show that much more is true with further restrictions on the dynamics: syndetic sets arising from these systems have positive multiplicative density; and \cref{thm:translatesofiprstararelarge} shows that members of a combinatorially defined subclass of syndetic sets also have positive multiplicative density. \\
The idea to approach problems in Ramsey Theory and combinatorial number theory with tools from dynamics goes back to the work of Furstenberg \cite{foriginal} in the measure-theoretic setting and Furstenberg and Weiss \cite{furstenbergweiss} in the topological setting. The basic idea is that the existence of combinatorial configurations in subsets of $\mathbb{N}$ can be reformulated in the language of dynamics to be about the recurrence of points and sets. Consider, for example, that the set $A \subseteq \mathbb{N}$ contains an arithmetic progression of length $k+1$ and step size $n$ if and only if
\begin{equation}
\label{eqn:discretemultiplerecurrence}
A \cap \big( A - n \big) \cap \big( A - 2n \big) \cap \cdots \cap \big( A - kn \big) \neq \emptyset.
\end{equation}
Results concerning the recurrence of sets in topological dynamical systems can be made to apply to sets in the positive integers via \emph{correspondence principles} which, roughly speaking, turn the would-be dynamical system $(\mathbb{N},n \mapsto n+1)$ into a genuine one and convert the expression \eqref{eqn:discretemultiplerecurrence} into one similar to \eqref{eqn:topologicalmultiplerecurrence} below regarding the recurrence of open sets.
Exemplifying this approach, the following topological dynamical result implies (and, in fact, can be shown to be equivalent to) van der Waerden's theorem.
\begin{theorem}[{\cite[Theorem 1.5]{furstenbergweiss}}]\label{thm:dvdw}
Let $(X,T)$ be a minimal dynamical system. For all non-empty, open $U \subseteq X$ and for all $k \in \mathbb{N}$, there exists $n \in \mathbb{N}$ such that
\begin{equation}
\label{eqn:topologicalmultiplerecurrence}
U \cap T^{-n} U \cap \cdots \cap T^{-kn} U \neq \emptyset.
\end{equation}
\end{theorem}
Interestingly, \cref{thm:dvdw} can be reformulated in terms of return time sets and multiplicative density. A set $A \subseteq \mathbb{N}$ has multiplicative upper Banach density equal to 1, $d_\times^*(A) = 1$, if and only if for all finite $F \subseteq \mathbb{N}$, there exists $n \in \mathbb{N}$ such that $nF \subseteq A$. (This equivalence is not apparent from (\ref{eqn:densityinseminplusintromult}) but is immediate from \cref{def:multdensity}. Such sets are called \emph{multiplicatively thick.}) We demonstrate the equivalence between \cref{thm:dvdw} and the following theorem at the beginning of \cref{sec:proofofmainthm}.
\begin{theorem}\label{thm:dvdwreform}
Let $(X,T)$ be a minimal dynamical system. There exists a residual set $X' \subseteq X$ such that for all $x \in X'$ and all non-empty, open $U \subseteq X$ containing $x$, the set $R(x,U)$ satisfies $d_\times^*(R(x,U)) = 1$, i.e., it is multiplicatively thick.
\end{theorem}
Though \cref{thm:dvdwreform} strongly resembles \cref{thm:mainthm} -- the set $U$ is not required to be a neighborhood of the point $x$ in \cref{thm:mainthm} -- one cannot hope to easily derive the latter from the former by translating the return time sets. Evidence for this is given by the fact that there are examples of sets which are multiplicatively large but whose additive translates are all multiplicatively very small; one can construct, for example, a multiplicatively thick set $A \subseteq \mathbb{N}$ with the property that for all $t \in \mathbb{Z} \setminus \{0\}$, the set $A+t$ has zero multiplicative density in all cosets of all non-trivial multiplicative subsemigroups of $\mathbb{N}$.
The following theorem, a result of Glasner's reformulated in a similar vein, is an improvement to \cref{thm:dvdw} in the case that $(X,T)$ is weakly mixing. A system $(X,T)$ is \emph{weakly mixing} if the system $(X^2, T \times T)$ contains a point with a dense forward orbit. A minimal, weakly mixing system is totally minimal, so our \cref{thm:mainthmwithtotallymin} can be understood to make less of an assumption on the dynamics and arrive at a similar, but weaker, conclusion.
\begin{theorem}[{\cite[Corollary 2.5]{glasnertopergodicdecomp}}]\label{thm:glasnerthickreturns}
Let $(X,T)$ be a minimal, weakly mixing dynamical system. There exists a residual set $X' \subseteq X$ such that for all $x \in X'$ and all non-empty, open sets $U \subseteq X$, the set $R(x,U)$ satisfies $d_\times^*(R(x,U)) = 1$.
\end{theorem}
These theorems exemplify the historical precedent that motivated our approach to \cref{question:mainquestion} by considering syndetic sets arising in dynamics. As \cref{question:mainquestion} is purely combinatorial, our ultimate goal is to understand the multiplicative configurations contained in arbitrary syndetic subsets of the natural numbers. If one is looking for additive configurations in syndetic sets then this is achieved historically via dynamics by making use of translation invariance: since arithmetic progressions and additive density are translation invariant, one can transfer the problem of finding such configurations (as in van der Waerden's theorem) to an analogous dynamical problem on symbolic shift space. The same approach does not work as easily for multiplicative configurations. The fact that geometric progressions and multiplicative density are not translation invariant presents the most serious obstacle faced in this work.
\subsection{Outline of the paper}
In \cref{sec:defs}, we gather definitions and prove some initial lemmas on additive and multiplicative density, topological and measurable dynamics, and set-valued maps. We define the rational topological Kronecker factor of a system in \cref{sec:ratkronanddistallemma} and prove a key lemma about distal systems. In \cref{sec:dynamicsondiag}, we establish some preliminary results concerning dynamics on the orbit closure of the diagonal. This is followed by proofs of the main results, \cref{thm:mainthm}, \cref{thm:mainthmwithtotallymin,thm:mainthmwithdistal}, and \cref{thm:translatesofiprstararelarge}, in Sections~\ref{sec:proofofmainthm}, \ref{sec:proofofdistalresults}, and \ref{sec:combresults}, respectively. We conclude the work with \cref{sec:concludingremarks,sec:remarksandquestions} by exhibiting some syndetic sets which do not arise from dynamics and collecting some questions for further consideration.
\subsection{Acknowledgements}
Thanks goes to Joel Moreira for permission to include his previously unpublished \cref{lem:joels-lemma0,lem:joels-lemma} in this paper. Gratitude is also extended to the referees for a number of helpful comments and corrections.
\section{Definitions and preliminary results}\label{sec:defs}
In this section we gather definitions and preliminary results that will be necessary later on. Denote by $\mathbb{N}$ the set of positive integers and by $\mathbb{N}_0$ the set $\mathbb{N} \cup \{0\}$.
\subsection{Set algebra, additive and multiplicative density}
For $A \subseteq \mathbb{N}$ and $n \in \mathbb{N}$, define
\[A-n \vcentcolon= \{m \in \mathbb{N} \ | \ m+n \in A\} \text{ and } A / n \vcentcolon= \{m \in \mathbb{N} \ | \ mn \in A\}.\]
The set $A$ is \emph{syndetic} if there exists a finite set $F \subseteq \mathbb{N}$ for which
\[A-F \vcentcolon= \bigcup_{n \in F} (A-n) = \mathbb{N}.\]
This is equivalent to the set $A$ having bounded gaps: if $A = \{a_1 < a_2 < \cdots\}$, then $A$ is syndetic if and only if $\sup_{i \in \mathbb{N}} (a_{i+1}-a_i)$ is finite.
A \emph{mean on} $\mathbb{N}$ is a positive linear functional of norm 1 on $B(\mathbb{N})$, the Banach space of bounded, real-valued functions on $\mathbb{N}$ with the supremum norm. A mean $\lambda$ is \emph{(additively) translation invariant} if for all $f \in B(\mathbb{N})$ and all $m \in \mathbb{N}$, $\lambda \big( n \mapsto f(n + m) \big) = \lambda ( f )$. Abusing notation, for $A \subseteq \mathbb{N}$, we write $\lambda(A)$ to mean $\lambda(\mathbbm{1}_A)$, where $\mathbbm{1}_A \in B(\mathbb{N})$ is the indicator function of $A$; if $\lambda$ is translation invariant, then for all $m \in \mathbb{N}$, $\lambda(A-m) = \lambda(A+m) = \lambda(A)$.
The following is an easy consequence of the pigeonhole principle that will be used repeatedly throughout this work.
\begin{lemma}\label{lem:meanpigeon}
Let $\lambda$ be a mean on $\mathbb{N}$ and $A_1, \ldots, A_k \subseteq \mathbb{N}$. If $\eta > 0$ is such that for all $i \in \{1,\ldots,k\}$, $\lambda(A_i) > \eta$, then there exists $n \in \mathbb{N}$ for which $\big|\{ 1 \leq i \leq k \ | \ n \in A_i \} \big| > \eta k$.
\end{lemma}
\begin{proof}
Define $g(n) = \sum_{i=1}^k \mathbbm{1}_{A_i}(n) / k$. Because $\lambda$ is positive and $\lambda( g ) > \eta$, there exists $n \in \mathbb{N}$ for which $g(n) > \eta$, as was to be shown.
\end{proof}
A \emph{multiplicative subsemigroup} $S$ of $\mathbb{N}$ is a subset of $\mathbb{N}$ that is closed under multiplication, and a \emph{coset} of $S$ is a set of the form $nS$ for $n \in \mathbb{N}$. The multiplicative subsemigroups that will appear most frequently in this paper are
\[S_N := \big\{n \in \mathbb{N} \ \big| \ (n,N) = 1 \big\},\]
the natural numbers coprime to a given positive integer $N \in \mathbb{N}$.
\begin{definition}\label{def:multdensity}
Let $S$ be a multiplicative subsemigroup of $\mathbb{N}$, $n \in \mathbb{N}$, and $A \subseteq nS$. The \emph{multiplicative upper Banach density} of $A$ in $nS$ is
\[d^*_{nS}(A) = \sup \big\{ \alpha \geq 0 \ \big| \ \forall F \subseteq nS \text{ finite}, \ \exists s \in S, \ |sF \cap A| \geq \alpha|F| \big\}.\]
When $S = \mathbb{N}$, we write $d^*_\times$ instead of $d^*_\mathbb{N}$. When $A \subseteq \mathbb{N}$ is not a subset of $nS$, the symbol $d^*_{nS}(A)$ is understood to mean $d^*_{nS}(A \cap nS)$. Note that for all $A \subseteq \mathbb{N}$, $d^*_{nS}(A) = d^*_S(A/n)$.
\end{definition}
\begin{remark}\label{rmk:densityequivalent}
The upper Banach density $d^*_S$ that was just defined for multiplicative subsemigroups $S$ is equivalent to the upper Banach density defined via multiplicative F{\o}lner sequences for $(S,\cdot)$ (as in (\ref{eqn:densityinseminplusintromult}) in the case of $(\mathbb{N},\cdot)$) or multiplicatively invariant means on $S$. For a proof, see \cite[Theorem 3.5]{BGpaperonearxiv}.
\end{remark}
A set $A \subseteq \mathbb{N}$ is \emph{GP-rich} if it contains arbitrarily long geometric progressions, subsets of the form $\{nm,nm^2,\ldots,nm^\ell\}$. The following theorem is a simple consequence of \cref{def:multdensity} and Szemer\'edi's theorem \cite{szemeredi} on arithmetic progressions.
\begin{theorem}\label{thm:multszem}
Let $S$ be a multiplicative subsemigroup of $\mathbb{N}$, $n \in \mathbb{N}$, and $A \subseteq nS$. If $d^*_{nS}(A) > 0$, then $A$ is GP-rich.
\end{theorem}
\begin{proof}
Let $0 < \varepsilon < d^*_{nS}(A)$ and $\ell \in \mathbb{N}$. By Szemer\'edi's theorem, there exists $L \in \mathbb{N}$ so that all subsets of $\{1, \ldots, L\}$ with relative density at least $\varepsilon$ contain an arithmetic progression of length $\ell$. It follows that for all $m, s \in \mathbb{N}$, all subsets of $s \{m,m^2,\ldots,m^L\}$ of relative density at least $\varepsilon$ contain a geometric progression of length $\ell$.
Fix $m \in S$, and put $F=\{nm,nm^2,\ldots,nm^L\} \subseteq nS$. By the definition of multiplicative upper Banach density, because $d^*_{nS}(A) > \varepsilon$, there exists $s \in S$ such that the set $A$ has relative density at least $\varepsilon$ in $\{snm,snm^2,\ldots,snm^L\}$; in particular, $A$ contains a geometric progression of length $\ell$.
\end{proof}
In fact, sets with positive multiplicative density contain much richer combinatorial configurations than simply geometric progressions. Bergelson \cite{bmultlarge} showed that a set $A \subseteq \mathbb{N}$ with positive multiplicative density contains \emph{geo-arithmetic configurations} such as the ones appearing in the following theorem. We will use the following extension of his result for applications in this paper.
\begin{theorem}\label{thm:geoarithmeticpatterns}
Let $n, N \in \mathbb{N}$, and $A \subseteq nS_N$. If $d^*_{nS_N}(A) > 0$, then for all $\ell \in \mathbb{N}$, there exist $a,c,d \in \mathbb{N}$ such that $\big\{ c (a+id)^j \ \big| \ 1 \leq i, j \leq \ell \big\} \subseteq A$.
\end{theorem}
\begin{proof}
It is quick to verify that the multiplicative subsemigroup $S_N$ satisfies the conditions in \cite[Theorem 8.8]{BGpaperonearxiv}. Let $\ell \in \mathbb{N}$, and apply that theorem to $A/n$ (using that $d^*_{S_N}(A / n) > 0$) with the endomorphisms $\varphi_i: S_N \to S_N$ defined by $\varphi_i(m) = m^i$ and with the finite set $F$ equal to an arithmetic progression long enough to guarantee that the subset $F'$ (in the notation of \cite[Theorem 8.8]{BGpaperonearxiv}) contains an arithmetic progression of length $\ell$. This yields the desired configuration in the set $A/n$; multiplying by $n$ yields the configuration in the set $A$.
\end{proof}
\subsection{Topological and measurable dynamics}
Let $X$ and $Y$ be topological spaces and $A \subseteq X$. The set $A$ is \emph{residual} if it contains a dense $G_\delta$ set, and it is \emph{meager} if its complement is residual. A map $f: X \to Y$ is \emph{semiopen} if $f(A)$ has non-empty interior when $A$ has non-empty interior.
\begin{lemma}\label{lem:imageofresidual}
Let $X$ be a complete metric space and $Y$ be a Hausdorff topological space. If $A \subseteq X$ is residual and $f: X \to Y$ is a continuous, semiopen surjection, then $f(A) \subseteq Y$ is residual.
\end{lemma}
\begin{proof}
This is proved in \cite[Lemma 4.25]{phelpsbook} under the assumption that $f$ is an open map. The same proof, with the obvious adjustments (i.e., in the Banach-Mazur game, the winning strategy of $B$ comes by choosing the open set $V_i$ to be the set of interior points of $f(B_i)$, which is non-empty because $f$ is semiopen) gives the result in the case that $f$ is semiopen.
\end{proof}
In a metric space $(X,d)$, the open ball of radius $r$ centered at $x$ will be denoted $B(x,r)$. The set $A$ is \emph{$\varepsilon$-dense} if for all $x \in X$, there exists $a \in A$ such that $d(x,a)<\varepsilon$.
A \emph{topological dynamical system} $(X,T)$ is a compact metric space $X$ paired with a continuous map $T: X \to X$; we will usually refer to $(X,T)$ as simply a \emph{system}. The system $(X,T)$ is \emph{invertible} if $T$ is a homeomorphism. The set $A$ is \emph{$T$ invariant} if $TA \subseteq A$.
Given $x \in X$ and $U \subseteq X$, the \emph{set of return times of $x$ to $U$} is
\[R_T(x,U) = \{n \in \mathbb{N} \ | \ T^n x \in U \}. \]
The letter $U$ will usually be used for a non-empty, open subset of $X$, and we will usually write $R(x,U)$ instead of $R_T(x,U)$ when the map $T$ is understood.
Note the standard manipulations
\[R_{T^n}(x,U) = R_T(x,U) / n \quad \text{ and } \quad R_T(T^nx,U) = R_T(x,U) - n.\]
The \emph{(forward) orbit of a point $x \in X$ under $T$} is $o_T(x) \vcentcolon= \{ T^{n}x \ | \ n \in \mathbb{N}_0\} $ while that of a subset $Y \subseteq X$ is $o_{T}(Y) \vcentcolon= \bigcup_{n \in \mathbb{N}_0} T^{n}Y$. We denote the corresponding closures with $\overline{o}_T(x)$ and $\overline{o}_T(Y)$ respectively. A system $(X,T)$ is \emph{minimal} if all points have a dense forward orbit and \emph{totally minimal} if for all $n \in \mathbb{N}$, the system $(X,T^n)$ is minimal.
Given continuous maps $T,S: X \to X$ that commute, the \emph{(forward) orbit of a point $x \in X$ under $T$ and $S$} is $o_{T,S}(x) = \{ T^{n} S^m x \ | \ n, m \in \mathbb{N}_0\} ,$ and we denote its closure with $\overline{o}_{T,S}(x).$ The system $(X,T,S)$ is \emph{minimal} if all points have a dense forward orbit. In the case that $T$ and $S$ are invertible, minimality is equivalent by the following lemma to all points having a dense ``two-sided'' orbit. We will make use of this lemma in Section~\ref{sec:dynamicsondiag}.
\begin{lemma}\label{lem:semigroupandgroupminimality}
Let $X$ be a compact metric space, and let $S$ be a commutative subsemigroup of the group of homeomorphisms of $X$. Denote by $G$ the group generated by $S$. The following are equivalent:
\begin{enumerate}
\item \label{eqn:minimalone} For all $x \in X$, $\overline{Sx} = X$.
\item \label{eqn:minimaltwo} For all $x \in X$, $\overline{Gx} = X$.
\end{enumerate}
\end{lemma}
\begin{proof}
Clearly (\ref{eqn:minimalone}) implies (\ref{eqn:minimaltwo}) since $S \subseteq G$. Suppose (\ref{eqn:minimaltwo}) holds, and let $x \in X$. By adjoining the identity map to $S$ if necessary, we may assume without loss of generality that $S$ contains the identity. Let
\[A = \bigcap_{s \in S} \overline{S sx}.\]
We will show that $A = X$; this will conclude the proof since $A \subseteq \overline{Sx}$. Since the collection $\{\overline{S sx} \ | \ s \in S\}$ has the finite intersection property and $X$ is compact, $A$ is compact and non-empty. We claim that for all $g \in G$, $gA \subseteq A$. Let $g \in G$; since $S$ generates $G$ and is commutative, there exist $s_1, s_2 \in S$ such that $g=s_1 s_2^{-1}$. Since $g: X \to X$ is a homeomorphism,
\[gA = \bigcap_{s \in S} g\overline{S sx} = \bigcap_{s \in S} \overline{S sgx}.\]
Let $y \in gA$; we will show that $y \in A$. Let $s \in S$. Since $y \in gA$ and $s_2 s \in S$, $y \in \overline{S s_2 s s_1 s_2^{-1}x}$. Since $Ss_1 \subseteq S$, we see that $y \in \overline{S s x}$, and since $s \in S$ was arbitrary, this shows $y \in A$. Since $y \in gA$ was arbitrary, $gA \subseteq A$. Now by (\ref{eqn:minimaltwo}), for any $a \in A$, $X = \overline{Ga} \subseteq A$, meaning $A = X$, as was to be shown.
\end{proof}
The system $(X,T)$ is \emph{distal} if for all $x, y \in X$ with $x \neq y$, $\inf_{n \in \mathbb{N}} d(T^n x, T^n y) > 0$. Distal systems appear in the statement of \cref{thm:mainthmwithdistal}. Distality is a fundamental concept in understanding the structure of topological and measurable dynamical systems; see \cite{glasnertopstructuretheory} and the references therein. In this work, the definition of distality and the fact that distal systems are invertible\footnote{This follows immediately from the fact that the Ellis enveloping semigroup is a group; see \cite[Theorem 3.1]{furstenbergdistalstructuretheory}.} will suffice.
\begin{definition}[cf. {\cite[Definition 6.8.10]{downarowiczbook}}]\label{def:naturalext}
Let $(X,T)$ be a system with $T$ surjective. The \emph{(topological) natural extension} of $(X,T)$ is the system $(W,T)$, where
\[W \vcentcolon= \big\{ (w_i)_{i \in \mathbb{Z}} \in X^\mathbb{Z} \ \big| \ \forall \ i \in \mathbb{Z}, \ w_{i+1} = Tw_i \big\}\]
inherits the topology from the product topology on $X^\mathbb{Z}$, $T = \restr{\sigma}{W}$ is the restriction of the left shift on $X^\mathbb{Z}$ to $W$, and $\pi: (W,T) \to (X,T)$ is defined by $\pi((w_i)_{i \in \mathbb{Z}}) = w_0$.
\end{definition}
\begin{lemma}\label{lem:naturalextensionisminimal}
Let $(X,T)$ be a system with $T$ surjective, and let $(W,T)$ be its natural extension.
\begin{enumerate}
\item \label{eqn:naturalone} The system $(W,T)$ is invertible.
\item \label{eqn:naturaltwo} For all $n \in \mathbb{N}$, the system $(X,T^n)$ is minimal if and only if $(W,T^n)$ is minimal.
\item \label{eqn:naturalthree} If $(X,T)$ is minimal, then the factor map $(W,T) \to (X,T)$ is semiopen.
\end{enumerate}
\end{lemma}
\begin{proof}
Statement (\ref{eqn:naturalone}) follows from the definition of $(W,T)$.
Statement (\ref{eqn:naturaltwo}) follows from two facts: factors of minimal systems are minimal, and, when $(X,T^n)$ is minimal, $W$ is the only closed, $T^n$-invariant subset of $W$ that surjects onto $X$.
Statement (\ref{eqn:naturalthree}) follows from (\ref{eqn:naturaltwo}) and the more general fact that a factor map $\pi: (W,T) \to (X,T)$ of minimal systems is semiopen. Let $U \subseteq W$ be open, and let $V \subseteq U$ be closed with non-empty interior. Since the system $(W,T)$ is minimal, there exists $h \in \mathbb{N}$ so that $\bigcup_{n=1}^h T^{-n} V = W$. Applying the factor map, $\bigcup_{n=1}^h T^{-n} \pi (V) = X$. Each $T^{-n} \pi (V)$ is closed, so by the Baire Category Theorem, there exists $n \in \{1,\ldots,h\}$ so that $T^{-n} \pi (V)$ has non-empty interior. By \cite[Theorem 2.4]{kstpaper}, the map $T$ is semiopen, hence $T^n$ is semiopen, too. It follows that $T^nT^{-n} \pi (V) \subseteq \pi (V) \subseteq \pi (U)$ has non-empty interior, as was to be shown.
\end{proof}
We will frequently make use of probability measures on compact metric spaces. Unless otherwise stated, all measures appearing in this work are Borel probability measures. If $\mu$ is a measure on $X$, we write $T \mu$ for the push-forward measure defined for $A \subseteq X$ by $T\mu(A) = \mu(T^{-1}A)$. The measure $\mu$ is \emph{$T$-invariant} if $T\mu = \mu$.
We thank Joel Moreira for permission to include \cref{lem:joels-lemma0,lem:joels-lemma}, previously unpublished, in this paper.
\begin{lemma}\label{lem:joels-lemma0}
Let $(X,T)$ be a system and $\mu$ be a $T$-invariant probability measure on $X$. Suppose that $x\in X$ is such that $\overline{o}(x) = X$. For all $f\in C(X)$, $\varepsilon>0$, and $N_0\in\mathbb{N}$, there exist $m,N\in\mathbb{N}$ with $N\geq N_0$ such that
\[ \left| \int_X f \; d\mu -\frac{1}{N}\sum_{i=m}^{m+N-1} f(T^i x) \right|<\varepsilon.\]
\end{lemma}
\begin{proof}
Let $f \in C(X)$, $\varepsilon > 0$, and $N_0 \in \mathbb{N}$. Let $M = \max_{x \in X} |f(x)|$, $I=\int_X f\;d\mu$, and for $m, N\in \mathbb{N}$, put
\[A(m,N) \vcentcolon= \frac{1}{N}\sum_{i=m}^{m+N-1}f(T^i x). \]
Note that for all $N>2M/\varepsilon$,
\begin{align}\label{E:1jl} \max \big\{|A(m,N)-A(m,N+1)|, |A(m,N)-A(m+1,N)| \big\}\leq \frac{2M}{N}<\varepsilon.\end{align}
By the ergodic decomposition, there exist ergodic, $T$-invariant, Borel probability measures $\mu_1$ and $\mu_2$ such that
\[ I_1:=\int_X f\;d\mu_1\leq I\leq I_2:= \int_X f\;d\mu_2.\]
By \cite[Proposition 3.9]{furstenberg-book}, for each $i = 1,2$, there exist $N_i>\max\{N_0,2M/\varepsilon$\} and $m_i\in \mathbb{N}$ such that $|I_i-A(m_i,N_i)|<\varepsilon.$ It follows that
\[ I-\varepsilon\leq I_2-\varepsilon< A(m_2,N_2) \quad \text{and} \quad A(m_1,N_1)< I_1+\varepsilon\leq I+\varepsilon.\]
By (\ref{E:1jl}), there exist $m, N \in \mathbb{N}$ between $m_1, m_2$ and $N_1, N_2$, respectively, for which $I-\varepsilon< A(m,N)< I+\varepsilon,$ as was to be shown.
\end{proof}
\begin{lemma}\label{lem:joels-lemma}
Let $(X,T)$ be a system and $\mu$ be a $T$-invariant probability measure on $X$. Suppose that $x\in X$ is such that $\overline{o}(x) = X$. There exists an additively invariant mean $\lambda$ on $\mathbb{N}$ such that for all $f\in C(X)$,
\[\lambda \big(n \mapsto f(T^n x) \big) = \int_X f \ d\mu. \]
\end{lemma}
\begin{proof}
Let
\begin{align*}
\Omega &\vcentcolon= \big\{ \omega_f: n \mapsto f(T^n x) \ \big| \ f \in C(X) \big\} \subseteq B(\mathbb{N}),\\
\chi &\vcentcolon= \big\{ n \mapsto \zeta(n+1)-\zeta(n) \ \big| \ \zeta \in B(\mathbb{N}) \big\} \subseteq B(\mathbb{N}).
\end{align*}
Note that because $\overline{o}(x) = X$, the map $f \mapsto \omega_f$ is a bijection from $C(X)$ to $\Omega$. In what follows, when we write $\omega_f \in \Omega$, we are implicitly specifying both $\omega_f \in \Omega$ and the corresponding $f \in C(X)$.
We claim that for all $\omega_f \in \Omega \cap \chi$, $\int_X f \ d\mu = 0$. Indeed, there exists $\zeta \in B(\mathbb{N})$ such that for all $n \in \mathbb{N}$,
\[ w_f(n)= f(T^n x) = \zeta(n+1)-\zeta(n).\]
Because $\zeta$ is bounded, for all $\varepsilon>0$, all sufficiently large $N \in \mathbb{N}$, and all $m \in \mathbb{N}$,
\[ \left| \frac{1}{N}\sum_{n=m}^{m+N-1}f(T^n x)\right| = \frac{|\zeta(m+N)-\zeta(m)|}{N}<\varepsilon.\]
It follows now by \cref{lem:joels-lemma0} that $\int_X f \ d\mu =0$.
Define a linear function $\lambda: \Omega + \chi \to\mathbb{R}$ by $\lambda(\omega_f+c) = \int_X f \ d\mu$. The previous paragraph shows that $\lambda$ is well defined. The plan is to extend $\lambda$ using the Hahn-Banach theorem to a positive linear functional; any such extension of $\lambda$ will satisfy the conclusions of the lemma.
First, we claim that $\|\lambda\|\leq 1.$ To see this, let $\tau=\omega_f + c\in \Omega + \chi,$ and let $\zeta \in B(\mathbb{N})$ be such that $c(n) = \zeta(n+1)-\zeta(n)$. By \cref{lem:joels-lemma0}, for all $\varepsilon>0,$ there exist $N > \|\zeta\|/\varepsilon$ and $m\in\mathbb{N}$ such that
\[\left|\int_X f \ d\mu-\frac{1}{N}\sum_{n=m}^{m+N-1}\omega_f(n)\right|<\varepsilon.\]
By the same reasoning as before,
\[ \left|\frac{1}{N}\sum_{n=m}^{m+N-1} \big(\tau(n)-\omega_f(n) \big)\right|=\left| \frac{1}{N}\sum_{n=m}^{m+N-1} c(n)\right|=\frac{|\zeta(m+N)-\zeta(m)|}{N}<2\varepsilon.\]
It follows that
\[\left| \lambda(\tau)-\frac{1}{N}\sum_{n=m}^{m+N-1}\tau(n)\right|=\left|\int_X f \ d\mu-\frac{1}{N}\sum_{n=m}^{m+N-1}\tau(n)\right|< 3\varepsilon.\]
This shows that there exists $n \in \{m, \ldots, m+N-1\}$ for which $|\tau(n)|\geq |\lambda(\tau)|-3\varepsilon,$ meaning $\|\tau\|\geq |\lambda(\tau)|-3\varepsilon.$ Since $\varepsilon > 0$ was arbitrary, $|\lambda(\tau)|\leq \|\tau\|.$
By the Hahn-Banach theorem, $\lambda$ extends to a linear functional on $B(\mathbb{N})$ (which we still call $\lambda$) with norm $\| \lambda \| \leq 1$. We have only to show that $\lambda$ is positive and translation invariant. To show positivity, let $\mathbbm{1} \in B(\mathbb{N})$ denote the constant one function, and note that $\tau(\mathbbm{1}) = 1$. Suppose $\tau \in B(\mathbb{N})$ is positive. Since $0\leq \tau/\|\tau\|\leq 1,$ we have $\big|\lambda(\mathbbm{1}-\tau/\|\tau\|) \big|\leq 1,$ and it follows that
\[ \frac{\lambda(\tau)}{\|\tau\|}=\lambda\left(\frac{\tau}{\|\tau\|}\right)=1-\lambda\left(\mathbbm{1}-\frac{\tau}{\|\tau\|}\right)\geq 0.\]
To show invariance, let $\tau\in B(\mathbb{N})$, and define $c \in B(\mathbb{N})$ by $c(n) = \tau(n+1)- \tau(n).$ Since $c \in \chi$, $\lambda(c)=0,$ meaning $\lambda\big(n \mapsto \tau(n+1) \big)=\lambda ( \tau ).$
\end{proof}
\subsection{Set-valued maps}
Let $(X,d)$ be a compact metric space. For $A \subseteq X$ and $\delta > 0$, let
\begin{align}\label{eqn:defofdeltaneighborhood}[A]_\delta \vcentcolon= \{x \in X \ | \ \exists a \in A, \ d(x,a) \leq \delta \}.\end{align}
The set of all non-empty, closed subsets of $X$ is denoted by $\mathcal{F}(X)$. The Hausdorff metric, defined between $F, H \in \mathcal{F}(X)$ by
\[d_H(F,H) \vcentcolon= \inf\{ \delta > 0 \ | \ F \subseteq [H]_\delta \text{ and } H \subseteq [F]_\delta\},\]
makes $(\mathcal{F}(X),d_H)$ a compact metric space.
\begin{definition}
\label{def:lsc_hausdorffmetric}
Let $X$ and $Y$ be compact metric spaces. A map $\varphi:X \to \mathcal{F}(Y)$ is \emph{lower semicontinuous} ({\sc lsc}{}) \emph{at $x \in X$} if for all $\varepsilon > 0$, there exists $\delta > 0$ such that for all $x' \in X$ with $d(x,x') < \delta$, $\varphi(x) \subseteq [\varphi(x')]_\varepsilon$.
\end{definition}
\begin{lemma}\label{lem:orbislsc}
Let $(X,T)$ be a system. The map $\overline{o}_T:X \to \mathcal{F}(X)$ is {\sc lsc}{}. In particular, it is Borel measurable: for all Borel subsets $B \subseteq \mathcal{F}(X)$, the set $\overline{o}_T^{-1}(B) \subseteq X$ is Borel.
\end{lemma}
\begin{proof}
For convenience, we will write $\overline{o}$ in place of $\overline{o}_T$. Let $x \in X$ and $\varepsilon>0$. Since $\overline{o}(x)$ is compact, there exist $m_1,\ldots, m_k\in \mathbb{N}_0$ such that $\{T^{m_1}x,\ldots,T^{m_k}x\}$ is an $\varepsilon/2$-dense subset of $\overline{o}(x)$. Because each $T^{m_i}$ is continuous, there exists $\delta > 0$ such that for all $x' \in X$ with $d(x,x') < \delta$, for all $1 \leq i \leq k$, $d(T^{m_i}x,T^{m_i}x') < \varepsilon / 2$.
We claim now that for all $x' \in B(x,\delta)$, $\overline{o}(x) \subseteq [\overline{o}(x')]_\varepsilon$. Let $x' \in B(x,\delta)$ and $y \in \overline{o}(x)$. There exists $1 \leq i \leq k$ for which $d(T^{m_i}x,y) < \varepsilon/2$, and by the triangle inequality, $d(T^{m_i}x',y) < \varepsilon$. This means $y \in [\overline{o}(x')]_\varepsilon$, as was to be shown.
The second statement follows from the fact that when $X$ and $Y$ are compact metric spaces, all {\sc lsc}{} functions $\varphi:X \to \mathcal{F}(Y)$ are Borel measurable; see Lemma~17.5, Theorem~17.15, and Theorem~18.10 in \cite{AliprantisBorderBook}.
\end{proof}
\begin{lemma}\label{lem:fortvariant}
Let $(X,T)$ be an invertible system, and denote by $\Xi$ the set of points of discontinuity of the map $\overline{o}_T: X \to \mathcal{F}(X)$. There exists a countable family $\{B_i\}_{i \in \mathbb{N}}$ of closed, $T$-invariant, empty-interior subsets of $X$ for which $\Xi \subseteq \bigcup_i B_i$.
\end{lemma}
\begin{proof}
For convenience, we will write $\overline{o}$ in place of $\overline{o}_T$. For $A \in \mathcal{F}(X)$ and $\varepsilon > 0$, let $M(A,\varepsilon)$ be the largest positive integer $n$ for which there exist $a_1, \ldots, a_n \in A$ satisfying, for all $i \neq j$, $d(a_i,a_j) > \varepsilon$, and let $U(A,\varepsilon) = \{x \in X \ | \ \exists a \in A, \ d(x,a) < \delta \}$. For $n \in \mathbb{N}$ and $\varepsilon > 0$, let
\[B_{n, \varepsilon} \vcentcolon= \left\{ x \in X \ \middle| \ \begin{gathered}M(\overline{o}(x),\varepsilon) \leq n, \text{ and } \\ \forall \varepsilon' \in (0,3\varepsilon), \ \forall U \ni x \text{ open}, \ \exists y \in U, \ \overline{o}(y) \not\subseteq U(\overline{o}(x),\varepsilon') \end{gathered} \right\}.\]
It is proved in \cite[Theorem 1]{fort} that each $B_{n,\varepsilon}$ is closed with empty interior and that $\Xi \subseteq \bigcup\{B_{n,\varepsilon} \ | \ n \in \mathbb{N}, \ \varepsilon \in \mathbb{Q}_+ \}$.
It remains to be shown that each $B_{n,\varepsilon}$ is $T$-invariant, i.e., $TB_{n,\varepsilon} \subseteq B_{n,\varepsilon}$. Let $x \in B_{n,\varepsilon}$. Since $\overline{o}(Tx) \subseteq \overline{o}(x)$, $M(\overline{o}(Tx),\varepsilon) \leq M(\overline{o}(x),\varepsilon) \leq n$. Let $0<\varepsilon'<3\varepsilon$ and $W$ be an open neighborhood of $Tx$. Since $x \in T^{-1}W$ and $x \in B_{n,\varepsilon}$, there exists $y \in T^{-1}W$ with $d(x,y) < \varepsilon'$ such that $\overline{o}(y) \not\subseteq U(\overline{o}(x),\varepsilon')$. Consider $Ty \in W$; it will complete the proof to show that $\overline{o}(Ty) \not\subseteq U(\overline{o}(Tx),\varepsilon')$. Since $d(x,y) < \varepsilon'$, $y \in U(\overline{o}(x),\varepsilon')$. Since $\overline{o}(y) \not\subseteq U(\overline{o}(x),\varepsilon')$, it follows that $\overline{o}(Ty) \not\subseteq U(\overline{o}(x),\varepsilon')$. Because $U(\overline{o}(Tx),\varepsilon') \subseteq U(\overline{o}(x),\varepsilon')$, this implies that $\overline{o}(Ty) \not\subseteq U(\overline{o}(Tx),\varepsilon')$.
\end{proof}
\begin{lemma}\label{lem:hausconvergence}
Let $X$ be a compact metric space. Suppose $(\mu_n)_{n \in \mathbb{N}}$ is a sequence of Borel probability measures converging in the weak-$\ast$ topology to a probability measure $\mu$. If $(H_n)_{n \in \mathbb{N}}$ is a sequence of closed subsets of $X$ such that $\mathop{\mathrm{supp}} \mu_n \subseteq H_n$ and $H$ is a closed subset of $X$ such that $H_n \to H$ in the Hausdorff metric, then $\mathop{\mathrm{supp}} \mu \subseteq H$.
\end{lemma}
\begin{proof}
We must prove that $\mu(H) = 1$. Since $H = \bigcap_{n \in \mathbb{N}} [H]_{1/n}$, in order to prove that $\mu(H) = 1$, it suffices to prove that for all $\delta > 0$, $\mu([H]_\delta) = 1$.
Fix $\delta > 0$. Convergence in the Hausdorff metric implies that $H_n \subseteq [H]_\delta$ for all sufficiently large $n \in \mathbb{N}$. By the properties of weak convergence of measures,
\[\mu([H]_\delta) \geq \limsup_{n \to \infty} \mu_n([H]_\delta) \geq \limsup_{n \to \infty} \mu_n(H_n) = 1,\]
meaning $\mu([H]_\delta) = 1$, as was to be shown.
\end{proof}
\section{The rational topological kronecker factor}\label{sec:ratkronanddistallemma}
Let $(X,T)$ be a minimal system. According to \cite[Theorem 3.1]{ye1992}, for all $n \in \mathbb{N}$, the set $X$ decomposes into a disjoint union of $d_n = d_n(X,T) \in \mathbb{N}$ clopen sets
\[X = X_{n,0} \cup \cdots \cup X_{n,d_n-1},\]
where $d_n$ divides $n$, for all $k \in \mathbb{Z}$, $T^k X_{n,j} = X_{n,j+k \pmod {d_n}}$, and the systems $(X_{n,j},T^n)$ are minimal. To save on notation, the second component of the index on $X_{n,j+k}$ and on related expressions will be implicitly understood to be taken modulo $d_n$. The notation $(X_{N,j})_{n,i}$ will mean the $i^{\text{th}}$ of the $d_n(X_{N,j},T^N)$ many $T^{Nn}$-minimal components of the system $(X_{N,j},T^N)$. With this definition, it is quick to check that $(X_{N,j})_{n,i} = X_{nN,iN + j}$.
For $U \subseteq X$ and a probability measure $\mu$ on $X$, we write
\begin{align}\label{eqn:defofuni}U_{n,i} \vcentcolon= U \cap X_{n,i}, \text{ and } \mu_{n,i} \vcentcolon= d_{n} \restr{\mu}{X_{n,i}}.\end{align}
Note that if $\mu$ is $T$-invariant, then $T\mu_{n,i} = \mu_{n,i+1}$ and $\mu_{n,i}$ is $T^n$-invariant. Though $\mu_{n,i}$ is technically a measure on $X$, we will sometimes regard $\mu_{n,i}$ as a measure on $X_{n,i}$ so that $(X_{n,i},T^n,\mu_{n,i})$ is a measure preserving system. This allows us to define the symbol $(\mu_{N,j})_{n,i}$ as in (\ref{eqn:defofuni}); regarded as measures on $X$, it is quick to check that $(\mu_{N,j})_{n,i} = \mu_{nN,iN+j}$.
\begin{lemma}\label{lem:totalminimalatresolutionU}
Let $(X,T)$ be a minimal system and $U \subseteq X$ be a non-empty, open set. There exists $N \in \mathbb{N}$ such that for all $n \in \mathbb{N}$ with $(n,N)=1$ and all $x \in X$, the $T^n$-orbit closure of $x$ has non-empty intersection with $U$.
\end{lemma}
\begin{proof}
It is equivalent to show that there exists $N\in\mathbb{N}$ such that for all $n\in \mathbb{N}$ with $(n,N)=1$, $\bigcup_{j=1}^\infty (T^n)^{-j} U=X$.
Because $T$ is minimal, there exists $h \in \mathbb{N}$ for which $X=\bigcup_{j=1}^h T^{-j}U$. The conclusion of the lemma will follow if we show that the set
\[ B\vcentcolon= \left\{n\in\mathbb{N} \ \middle| \ \bigcup_{j=1}^\infty (T^n)^{-j}U\neq X \right\} \]
does not contain $h$ pairwise coprime elements.
Let $n\in B$. There exists $0\leq i_n\leq d_n-1$ such that $U_{n,i_n}=\emptyset$. Indeed, if this was not the case, then for all $0 \leq i \leq d_n-1$, the set $U_{n,i}$ would be a non-empty, open subset of $X_{n,i}$. It would follow by the minimality of $(X_{n,i},T^n)$ that
\[X=\bigcup_{i=0}^{d_n-1} X_{n,i}=\bigcup_{i=0}^{d_n-1} \bigcup_{j=1}^\infty (T^n)^{-j}U_{n,i}=\bigcup_{j=1}^\infty (T^n)^{-j}U,\]
contradicting the fact that $n \in B$.
Suppose for a contradiction that $n_1, \ldots, n_h\in B$ are pairwise coprime. Since $d_{n_i}$ divides $n_i$, the numbers $d_{n_1}, \ldots, d_{n_h}$ are also pairwise coprime. For each $1\leq j\leq h$, let $0\leq i_{n_j}\leq d_{n_j}-1$ be the index for which $U_{n_j,i_{n_j}}=\emptyset$. By the Chinese Remainder Theorem, there exists $\ell \in \mathbb{N}$ such that for all $1\leq j\leq h,$ $\ell+j\equiv i_{n_j} \pmod{d_{n_j}}$. Using the fact that $U_{n_j,i_{n_j}}=\emptyset$,
\[ (T^{-(\ell+j)}U) \cap X_{n_j,0}\subseteq T^{-(\ell+j)}(U\cap T^{\ell+j}X_{n_j,0})=T^{-(\ell+j)}U_{n_j,i_{n_j}}=\emptyset.\]
Thus, for all $1 \leq j \leq h$, $(T^{-(\ell+j)}U)\cap X_{n_j,0} = \emptyset$.
On the other hand, since $d_{n_1}, \ldots, d_{n_h}$ are pairwise coprime, again by the Chinese Remainder Theorem,
\[\bigcup_{n = 1}^\infty T^{-n} \bigcap_{j=1}^h X_{n_j,0} = \bigcup_{n = 1}^\infty \bigcap_{j=1}^h X_{n_j,-n} = X,\]
which implies that $\bigcap_{j=1}^h X_{n_j,0}\neq \emptyset$. Since $\bigcup_{j=1}^h T^{-j}U=X$, $\bigcup_{j=1}^h T^{-(\ell+j)}U=X$. Putting these facts together, we see
\[\bigcap_{j=1}^h X_{n_j,0} = \left(\bigcup_{j=1}^h T^{-(\ell+j)}U \right) \cap \left(\bigcap_{j=1}^h X_{n_j,0}\right) \subseteq \bigcup_{j=1}^h \left((T^{-(\ell+j)}U)\cap X_{n_j,0} \right),\]
a contradiction since the leftmost set was shown to be non-empty while the rightmost set was shown to be empty.
\end{proof}
\begin{proposition}\label{thm:epsdenseorbits}
Let $(X,T)$ be a minimal system. For all $\varepsilon > 0$, there exists $N \in \mathbb{N}$ such that for all $n \in \mathbb{N}$ with $(n,N)=1$ and all $x \in X$, the $T^n$-orbit of $x$ is $\varepsilon$-dense in $X$.
\end{proposition}
\begin{proof}
Let $\varepsilon>0.$ Since $X$ is compact, there exist $x_1, \ldots, x_h \in X$ for which $X=\bigcup_{i=1}^h B(x_i,\varepsilon/2)$. For each $1\leq i\leq h,$ \cref{lem:totalminimalatresolutionU} gives the existence of $N_i\in\mathbb{N}$ such that for all $n\in\mathbb{N}$ with $(n,N_i)=1$ and all $x\in X$, $\overline{o}_{T^n}(x)\cap B(x_i,\varepsilon/2)\neq \emptyset.$ We claim that $N \vcentcolon= \prod_{i=1}^h N_i$ has the required property.
Let $n\in\mathbb{N}$ with $(n,N)=1$, $x\in X$, and $1 \leq i \leq h$. Since $(n,N_i)=1,$ \cref{lem:totalminimalatresolutionU} gives that $\overline{o}_{T^n}(x)\cap B(x_i,\varepsilon/2)\neq \emptyset.$ Since $1 \leq i \leq h$ was arbitrary and $X=\bigcup_{i=1}^h B(x_i,\varepsilon/2)$, the orbit $\overline{o}_{T^n}(x)$ is $\varepsilon$-dense.
\end{proof}
The family $\{\mathbb{Z} / d_n\mathbb{Z} \ | \ n \in \mathbb{N}\}$, directed via the maps $\mathbb{Z} / d_{nm}\mathbb{Z} \to \mathbb{Z} / d_n\mathbb{Z}$, gives rise to the \emph{rational topological Kronecker factor} of $(X,T)$: the system
\[\mathbf{K}_{\text{rat}}(X,T) \vcentcolon= \big(Z \vcentcolon= \varprojlim_{n \in \mathbb{N}} \mathbb{Z} / d_n \mathbb{Z},T\big),\]
where $T: (a_n)_{n \in \mathbb{N}} \mapsto (a_n + 1)_{n \in \mathbb{N}}$ is a minimal rotation of the compact abelian group $Z$. Defining $Z_{n,i}$ just as it was defined for $X$ at the beginning of this section, we see that $d_n(Z,T) = d_n(X,T)$ and, by the topology on $Z$, that the factor map $\pi: (X,T) \to (Z,T)$ is defined uniquely by the property $\pi(X_{n,i}) = Z_{n,i}$. Also, note that for any non-empty, open set $V \subseteq Z$, there exists $n,i \in \mathbb{N}$ such that $Z_{n,i} \subseteq V$.
The goal for the remainder of this section is to prove \cref{lem:distalhavegoodmeasure}, a result related to \cref{lem:totalminimalatresolutionU} on the measure of the sets $U \cap X_{N,j}$ in distal systems. This will be accomplished with the help of the topological Kronecker factor of $(X,T)$.
\begin{definition}\label{def:totallyvisible}
Let $(X,T)$ be a system, $U \subseteq X$ be open, non-empty, and $\mu$ be a $T$-invariant probability measure on $X$. The set $U$ is \emph{totally visible by $\mu$} if
\[\inf_{n,i \in \mathbb{N}} \mu_{n,i}(U) > 0.\]
The set $U$ is \emph{totally visible} if it is totally visible by some $T$-invariant probability measure $\mu$ on $X$.
\end{definition}
\begin{remark}\label{rmk:totallyminimalsystems}
If $(X,T)$ is totally minimal, then for all $n, i \in \mathbb{N}$, $X_{n,i} = X$ and $\mathbf{K}_{\text{rat}}(X,T)$ is trivial. For any $T$-invariant measure $\mu$ and all $n, i \in \mathbb{N}$, $\mu_{n,i} = \mu$. It follows that in totally minimal systems, all open sets are totally visible by any invariant probability measure.
\end{remark}
\begin{lemma}\label{lem:distalhavegoodmeasure}
Let $(X,T)$ be a minimal, distal system, and let $U \subseteq X$ be a non-empty, open set. There exists $N, j \in \mathbb{N}$ so that $U_{N,j}$ is totally visible in the system $(X_{N,j},T^N)$.
\end{lemma}
\begin{proof}
Let $W \subseteq V \subseteq U$ and $\varepsilon > 0$ be such that $W$ and $V$ are non-empty and open, $\overline{V} \subseteq U$, and for all $x \in W$, $B(x,2\varepsilon) \subseteq \overline{V}$. Put $(Z,T)= \mathbf{K}_{\text{rat}}(X,T)$. By \cite[Theorem 8.1]{furstenbergdistalstructuretheory}, the factor map $\pi: X \to Z$ is open, so there exists $N, j \in \mathbb{N}$ such that $Z_{N,j} \subseteq \pi W$.
Let $z_0 \in Z$. By \cite[Lemma 8.1]{furstenbergdistalstructuretheory}, there exists a finite set $F \subseteq \pi^{-1}(\{z_0\})$ such that for all $n \in \mathbb{Z}$, the set $T^n F$ is $\varepsilon$-dense in the fiber $\pi^{-1}(\{T^{n}z_0\})$.
Put $\eta = (2|F|)^{-1}$, and let
\[\nu = \frac{1}{|F|} \sum_{f \in F} \delta_f.\]
We claim that for all $n, i \in \mathbb{N}$ and all $k \in \mathbb{Z}$,
\begin{align}\label{eqn:fiberinequality}T^k \nu \big(\overline{V} \cap X_{nN,Ni+j} \big) \geq \eta T^k \delta_{z_0} \big( Z_{nN,Ni+j} \big).\end{align}
To see why, note that the right hand side is zero unless $T^k z_0 \in Z_{nN,Ni+j}$. Suppose that $T^k z_0 \in Z_{nN,Ni+j}$. Because $Z_{nN,Ni+j} \subseteq Z_{N,j} \subseteq \pi W$, there exists a point $x \in \pi^{-1}(\{T^k z_0\}) \cap W$. Since $x \in W$, $B(x,2\varepsilon) \subseteq \overline{V}$. Because $T^kF$ is $\varepsilon$-dense in $\pi^{-1}(\{T^{k}z_0\})$, at least one point of $T^k F$ is in $\overline{V}$. This combined with the fact that $T^k \nu$ is supported on $\pi^{-1}(\{T^{k}z_0\}) \subseteq X_{nN,Ni+j}$ implies that $T^k \nu \big(\overline{V} \cap X_{nN,Ni+j} \big) \geq \eta$, showing (\ref{eqn:fiberinequality}).
Let $\mu$ be a weak-$\ast$ limit point of the set $\big\{ N^{-1} \sum_{k=0}^{N-1} T^k \nu \ \big| \ N \in \mathbb{N} \big\}$. We claim that the set $U_{N,j}$ is visible by the measure $\mu_{N,j}$ in the system $(X_{N,j}, T^N)$. We must show that for all $n, i \in \mathbb{N}$,
\begin{align}\label{eqn:visibilityinequality}(\mu_{N,j})_{n,i} (U_{N,j}) \geq \eta.\end{align}
Let $n, i \in \mathbb{N}$, and recall that $d_n = d_n(X,T) = d_n(Z,T)$. Using (\ref{eqn:fiberinequality}), we see
\begin{align*}(\mu_{N,j})_{n,i} (U_{N,j}) &= \mu_{nN,Ni+j}(U)\\
&= d_{nN} \mu(U \cap X_{nN,Ni+j})\\
&\geq d_{nN}\mu(\overline{V} \cap X_{nN,Ni+j})\\
&\geq d_{nN}\liminf_{N \to \infty} \frac 1{N} \sum_{k=0}^{N - 1} T^k \nu (\overline{V} \cap X_{nN,Ni+j})\\
& \geq d_{nN}\eta \liminf_{N \to \infty} \frac 1{N} \sum_{k=0}^{N - 1} T^k \delta_{z_0} \big( Z_{nN,Ni+j} \big) = \eta,
\end{align*}
where the last equality follows from the ergodic theorem because $z_0$ is generic for the Haar measure on $Z$ and the Haar measure of $Z_{nN,Ni+j}$ is $1/d_{nN}$. This establishes (\ref{eqn:visibilityinequality}), concluding the proof.
\end{proof}
\section{Dynamics on the orbit closure of the diagonal}\label{sec:dynamicsondiag}
Suppose $(X,T)$ is minimal and invertible, and fix $\ell \in \mathbb{N}$ and $\vec m \in \mathbb{N}^\ell$. Put $M = \text{{\sc lcm}}(\vec m)$, and let $d_M = d_M(X,T)$ be as described in Section~\ref{sec:ratkronanddistallemma}. We will now prove some preliminary results concerning dynamics of points along the diagonal of $X^\ell$ and points of continuity of the orbit closure map $\overline{o}_{T^{m_1} \times \cdots \times T^{m_\ell}}$.
Let $\Delta: X \to X^\ell$, $x \mapsto (x,\ldots,x)$, be the diagonal injection. Let
\begin{align*}
\Delta(T) &\vcentcolon= T \times \cdots \times T,\\
T^{\vec m} &\vcentcolon= T^{m_1} \times \cdots \times T^{m_\ell},\\
X^{\Delta} & \vcentcolon= \overline{o}_{T^{\vec m}}\big(\Delta(X) \big) \vcentcolon= \overline{\bigcup_{n \in \mathbb{N}_0} (T^{\vec m})^{n} \Delta(X)} \subseteq X^\ell,\\
X^{\Delta}_{M,j} &\vcentcolon= \overline{o}_{T^{\vec m}}\big(\Delta(X_{M,j}) \big) \subseteq X^{\Delta}, \ j \in \{0,\ldots,d_M-1\}.
\end{align*}
Note that because $T$ is a homeomorphism, $\Delta(T)$ and $T^{\vec m}$ are commuting homeomorphisms of $X^\ell$.
\begin{theorem}\label{thm:glasnersystemisminimal}
The maps $\Delta(T)$ and $T^{\vec m}$ are homeomorphisms of $X^{\Delta}$, and the system $\big(X^{\Delta},\Delta(T), T^{\vec m} \big)$ is minimal.
\end{theorem}
\begin{proof}
Let
\begin{align}\label{eqn:definitionofxsuperdelta}\mathbb{X}^\Delta \vcentcolon= \overline{\bigcup_{n \in \mathbb{Z}} (T^{\vec m})^{n} \Delta(X)} \subseteq X^\ell.\end{align}
It is immediate that $\Delta(T)$ and $T^{\vec m}$ are homeomorphisms of $\mathbb{X}^\Delta$. It is proved in \cite[Theorem 5.1]{glasnertopergodicdecomp} that the system $(\mathbb{X}^\Delta, \Delta(T), T^{\vec m})$ is minimal in the case that $\vec m = (1, 2, \ldots, \ell)$. Since factors of minimal systems are minimal, and since $(\mathbb{X}^\Delta,\Delta(T), T^{\vec m})$ is a factor of a system to which Glasner's theorem applies (for example, the one corresponding to the vector $(1,2,\ldots, \max_i \vec m_i)$), it is minimal.
Let $\vec x \in X^{\Delta} \subseteq \mathbb{X}^\Delta$. Since $\mathbb{X}^\Delta$ is minimal, by \cref{lem:semigroupandgroupminimality},
\[\mathbb{X}^\Delta = \overline{\big\{\Delta(T)^n (T^{\vec m})^k \vec x \ \big| \ n, k \in \mathbb{N} \big\}} \subseteq X^{\Delta}.\]
This shows that $X^{\Delta} = \mathbb{X}^\Delta$. Therefore, $\Delta(T)$ and $T^{\vec m}$ are homeomorphisms of $X^{\Delta}$ and $\big(X^{\Delta},\Delta(T), T^{\vec m} \big)$ is minimal.
\end{proof}
\begin{theorem}\label{thm:subsystemsareminimal}
The $X^{\Delta}_{M,j}$'s are mutually disjoint, clopen, and
\[X^{\Delta} = X^{\Delta}_{M,0} \cup \cdots \cup X^{\Delta}_{M,d_M-1}.\]
The maps $\Delta(T)^M$ and $T^{\vec m}$ are homeomorphisms of $X^{\Delta}_{M,j}$, and the system $\big(X^{\Delta}_{M,j},\allowbreak \Delta(T)^M, T^{\vec m} \big)$ is minimal.
\end{theorem}
\begin{proof}
Since $X = \bigcup_{j=0}^{d_M-1} X_{M,j}$, it follows immediately from the definition of $X^{\Delta}$ that
\[X^{\Delta} = X^{\Delta}_{M,0} \cup \cdots \cup X^{\Delta}_{M,d_M-1}.\]
We will show next that the $X^{\Delta}_{M,j}$'s are mutually disjoint. Since they are closed, disjointness will imply that the $X^{\Delta}_{M,j}$'s are open, hence clopen.
Suppose $j, j' \in \{0,\ldots,d_M-1\}$ are such that $X^{\Delta}_{M,j} \cap X^{\Delta}_{M,j'} \neq \emptyset$; we will show that $j = j'$. Let $ \vec x = (x_1,\ldots,x_\ell) \in X^{\Delta}_{M,j} \cap X^{\Delta}_{M,j'}$. By the definition of $X^{\Delta}_{M,j}$, there exist sequences $(n_k)_{k \in \mathbb{N}}, (n_k')_{k \in \mathbb{N}} \subseteq \mathbb{N}$, $(y_k)_{k \in \mathbb{N}} \subseteq X_{M,j}$, and $(y_k')_{k \in \mathbb{N}} \subseteq X_{M,j'}$ so that
\[ \lim_{k \to \infty}(T^{\vec m})^{n_k} \Delta(y_k) = \lim_{k \to \infty} (T^{\vec m})^{n_k'} \Delta(y_k') =\vec x.\]
It follows that for each $i \in \{1,\ldots,\ell\}$,
\[ \lim_{k \to \infty} T^{m_i n_k} y_k = \lim_{k \to \infty} T^{m_i n_k'} y_k' = x_i.\]
For each $i \in \{1, \ldots, \ell\}$, let $j_i \in \{0,\ldots,d_M-1\}$ be such that $x_i \in X_{M,j_i}$. Since $T^{m_i n_k} y_k \in X_{M,j+ m_i n_k}$ and $x_i \in X_{M,j_i}$, it must be that $j+m_i n_k \equiv j_i \pmod {d_M}$ for all sufficiently large $k$. Similarly, we can conclude that $j'+m_i n_k' \equiv j_i \pmod {d_M}$, meaning $m_i (n_k' - n_k) \equiv j - j' \pmod {d_M}$ for all sufficiently large $k$. This implies that $j-j'$ is a multiple of $(m_i,d_M)$ for all $i \in \{1,\ldots,\ell\}$, whereby $(M,d_M)$ divides $j-j'$. Since $(M,d_M) = d_M$, it follows that $j \equiv j' \pmod {d_M}$, implying $j=j'$.
For $j \in \{0,\ldots, d_M-1\}$, let
\[\mathbb{X}^\Delta_{M,j} \vcentcolon= \overline{\bigcup_{n \in \mathbb{Z}} (T^{\vec m})^{n} \Delta(X_{M,j})} \subseteq X^\ell.\]
Since $T X_{M,j} = X_{M,j+1}$, we have that $\Delta(T)^M$ and $T^{\vec m}$ are homeomorphisms of $\mathbb{X}^\Delta_{M,j}$. It also follows that $\Delta(T)\mathbb{X}^\Delta_{M,j} = \mathbb{X}^\Delta_{M,j+1}$ and that
\[\mathbb{X}^\Delta = \mathbb{X}^\Delta_{M,0} \cup \cdots \cup \mathbb{X}^\Delta_{M,d_M-1},\]
where $\mathbb{X}^\Delta$ is as defined in (\ref{eqn:definitionofxsuperdelta}). It was shown in that proof that $X^{\Delta} = \mathbb{X}^\Delta$; combining this with the facts that $X^{\Delta}_{M,j} \subseteq \mathbb{X}^\Delta_{M,j}$ and that the $X^{\Delta}_{M,j}$'s are mutually disjoint, we see $\mathbb{X}^\Delta_{M,j} = X^{\Delta}_{M,j}$. This shows that $\Delta(T)^M$ and $T^{\vec m}$ are homeomorphisms of $X^{\Delta}_{M,j}$.
To show that $(X^{\Delta}_{M,j},\Delta(T)^M,T^{\vec m})$ is minimal, we will show that every point has a dense orbit, starting with points on the diagonal. Let $x \in X_{M,j}$ and consider
\[Y \vcentcolon= \overline{\big\{\Delta(T)^{Mn} (T^{\vec m})^k \Delta(x) \ \big| \ n, k \in \mathbb{N}_0 \big\}} \subseteq X^{\Delta}_{M,j}.\]
Since $(X_{M,j},T^M)$ is minimal, $\Delta(X_{M,j}) \subseteq Y$, and, moreover, for all $k \in \mathbb{N}$, $(T^{\vec m})^k \allowbreak \Delta(X_{M,j}) \subseteq Y$. Since $Y$ is closed, it follows that $X^{\Delta}_{M,j} \subseteq Y$, which implies that $Y = X^{\Delta}_{M,j}$. Thus, points on the diagonal have a dense orbit.
Let $\vec x \in X^{\Delta}_{M,j}$, and let $Y$ be the $(\Delta(T)^M,T^{\vec m})$-orbit closure of $\vec x$. Let $w \in X_{M,j}$, and note that $\Delta(w) \in X^{\Delta}_{M,j}$. By \cref{thm:glasnersystemisminimal}, the system $\big(X^{\Delta},\Delta(T), T^{\vec m} \big)$ is minimal, so there exists a sequence $\big((a_n,b_n) \big)_{n \in \mathbb{N}} \subseteq \mathbb{N}^2$ for which $(T^{\vec m})^{a_n} \Delta(T)^{b_n}\vec x \to \Delta(w)$ as $n \to \infty$. By passing to a subsequence, we may assume that there exists $0 \leq b \leq M -1$ such that for all $n \in \mathbb{N}$, $b_n \equiv b \pmod M$. Since $((T^{\vec m})^{a_n} \Delta(T)^{b_n}\vec x)_{n \in \mathbb{N}} \subseteq X^{\Delta}_{M,j+b}$, $\Delta(w) \in X^{\Delta}_{M,j+b}$. Since the $X^{\Delta}_{M,j}$'s are disjoint and $\Delta(w) \in X^{\Delta}_{M,j} \cap X^{\Delta}_{M,j+b}$, $b= 0$, whereby $\Delta(w) \in Y$. It follows now from the previous paragraph that $Y = X^{\Delta}_{M,j}$.
\end{proof}
\begin{lemma}\label{lem:nonemptyinteriortwo}
For all open, non-empty $U \subseteq X_{M,j}$, the set
\[X_U^\Delta \vcentcolon= \overline{o}_{T^{\vec m}}\big(\Delta(U) \big)\]
has non-empty interior in $X^{\Delta}_{M,j}$.
\end{lemma}
\begin{proof}
Let $U \subseteq X_{M,j}$ be open, non-empty. Since $(X_{M,j},T^M)$ is minimal, there exists $h \in \mathbb{N}$ such that
\[X_{M,j}^\Delta = \bigcup_{i=1}^h X_{T^{-Mi}U}^\Delta = \bigcup_{i=1}^h \Delta(T)^{-Mi} X_{U}^\Delta.\]
Since $X_{M,j}^\Delta$ is a Baire space (it is a compact metric space), some $\Delta(T)^{-Mi} X_{U}^\Delta$ has non-empty interior. Since $\Delta(T)^M$ is a homeomorphism, it is open, implying that $X_{U}^\Delta$ has non-empty interior.
\end{proof}
\begin{proposition}\label{prop:pointsofcontinuityalongthediagonal}
Let $\Omega$ be the set of points of continuity of the map $\overline{o}_{T^{\vec m}}: X^{\Delta} \to \mathcal{F}(X^{\Delta})$. The set $\Omega \cap \Delta(X)$ is a residual subset of $\Delta(X)$. For $j \in \{0,\ldots, d_M-1\}$, the set $\Omega_{M,j} \vcentcolon= \Omega \cap X^{\Delta}_{M,j}$ is the set of points of continuity of the map $\overline{o}_{T^{\vec m}}: X^{\Delta}_{M,j} \to \mathcal{F}(X^{\Delta}_{M,j})$ and $\Omega_{M,j} \cap \Delta(X_{M,j})$ is a residual subset of $\Delta(X_{M,j})$.
\end{proposition}
\begin{proof}
Let $\Xi = X^{\Delta} \setminus \Omega$ be the set of points of discontinuity of the map $\overline{o}_{T^{\vec m}}: X^{\Delta} \to \mathcal{F}(X^{\Delta})$. By \cref{lem:fortvariant}, there exists a countable family $\{B_i\}_{i \in \mathbb{N}}$ of closed, $T^{\vec m}$-invariant, empty-interior subsets of $X^{\Delta}$ for which $\Xi \subseteq \bigcup_i B_i$. We claim that each $B_i \cap \Delta(X)$ is a closed set with empty interior in $\Delta(X)$. It is closed because $\Delta(X)$ is closed. Suppose for a contradiction that $U \subseteq X$ is open and is such that $\Delta(U) \subseteq B_i$. Since $B_i$ is $T^{\vec m}$-invariant and closed, $X_U^\Delta \subseteq B_i$. It follows by \cref{lem:nonemptyinteriortwo} that $B_i$ has non-empty interior, a contradiction.
Note that $\Delta(X) \cap \Xi \subseteq \bigcup_i (B_i \cap \Delta(X))$ is a cover of $\Delta(X) \cap \Xi$ with closed sets with empty interior, meaning $\Delta(X) \cap \Xi$ is a meager subset of $\Delta(X)$. Since $\Delta(X)= \big(\Delta(X) \cap \Omega \big) \cup \big(\Delta(X) \cap \Xi \big)$, the set $\Delta(X) \cap \Omega$ is a residual subset of $\Delta(X)$.
Let $j \in \{0,\ldots, d_M-1\}$. Since $\overline{o}_{T^{\vec m}}: X^{\Delta}_{M,j} \to \mathcal{F}(X^{\Delta}_{M,j})$ is the restriction of the map $\overline{o}_{T^{\vec m}}: X^{\Delta} \to \mathcal{F}(X^{\Delta})$ to $X^{\Delta}_{M,j}$, the set of its points of continuity is $\Omega_{M,j} = \Omega \cap X^{\Delta}_{M,j}$. Since $\Delta(X_{M,j})$ is an open subset of $\Delta(X)$ and $\Omega$ is residual in $\Delta(X)$, the set $\Omega \cap \Delta(X_{M,j}) = \Omega_{M,j} \cap \Delta(X_{M,j})$ is a residual subset of $\Delta(X_{M,j})$.
\end{proof}
\section{Results on minimal systems} \label{sec:proofofmainthm}
We begin this section by demonstrating the equivalence between \cref{thm:dvdw,thm:dvdwreform}, dynamical formulations of van der Waerden's theorem from the Introduction.
\begin{proof}[Proof of equivalence between \cref{thm:dvdw,thm:dvdwreform}]
A point $x\in X$ is called \emph{$k$-recurrent} if for all $\varepsilon>0$ there exists $n\in\mathbb{N}$ such that $\max_{i=1,\ldots,k} d(x, T^{in}x)<\varepsilon$, and it is called \emph{multiply recurrent} (cf. \cite[page 9]{furstenberg-book}) if it is $k$-recurrent for all positive integers $k$. Let $E_k$ and $E$ denote the set of $k$-recurrent and multiply recurrent points in $X$, respectively, and note that $E = \bigcap_{k \in \mathbb{N}} E_k$. We will show that \cref{thm:dvdw,thm:dvdwreform} are both equivalent to the fact that $E$ is a residual subset of $X$.
First we will show that \cref{thm:dvdw} implies that $E$ is a residual subset of $X$. Let $k \in \mathbb{N}$, and for an open set $U \subseteq X$, let
\[U' \vcentcolon= \bigcup_{n \in \mathbb{N}} \big( U \cap T^{-n}U \cap \cdots \cap T^{-kn} U \big).\]
It follows from \cref{thm:dvdw} that $U'$ is an open, dense subset of $U$. (The set $U'$ is open by definition and non-empty by \cref{thm:dvdw}. To see that it is a dense subset of $U$, let $V \subseteq U$ be non-empty and open, and note that $U' \cap V \supseteq U' \cap V' = V'$, which is non-empty by \cref{thm:dvdw}.) Since $X$ is compact, for all $\ell \in \mathbb{N}$, there exists a finite open cover $X = U_{\ell, 1} \cup U_{\ell, 2} \cup \cdots \cup U_{\ell, m_\ell}$ where each $U_{\ell,i}$ has diameter less than $1/\ell$. Since each $U_{\ell, i}'$ is open and dense in $U_{\ell,i}$, the set $U_\ell' \vcentcolon= \bigcup_{i=1}^{m_\ell} U_{\ell,i}'$ is an open, dense subset of $X$. We claim that $E_k = \bigcap_{\ell \in \mathbb{N}} U_\ell'$. It will follow that $E_k$ is a residual subset of $X$, and since $E = \bigcap_{k \in \mathbb{N}} E_k$, it will finally follow that $E$ is a residual subset of $X$.
To see that $E_k = \bigcap_{\ell \in \mathbb{N}} U_\ell'$, let $x \in E_k$ and $\ell \in \mathbb{N}$, and let $i \in \{1, \ldots, m_\ell\}$ be such that $x \in U_{\ell,i}$. Let $\varepsilon > 0$ be sufficiently small so that $B(x,\varepsilon) \subseteq U_{\ell,i}$. Since $x \in E_k$, there exists $n \in \mathbb{N}$ for which $x$, $T^nx$, \dots, $T^{kn}x \in B(x, \varepsilon) \subseteq U_{\ell,i}$, which implies that $x \in U_{\ell,i}' \subseteq U_\ell'$. To see the reverse inclusion, suppose $x \in U_\ell'$ for all $\ell \in \mathbb{N}$. To see that $x \in E_k$, let $\varepsilon > 0$, and let $\ell > 1/ \varepsilon$. There exists $i \in \{1, \ldots, m_\ell\}$ such that $x \in U_{\ell,i}'$, meaning that there exists $n \in \mathbb{N}$ such that $x$, $T^n x$, \dots, $T^{kn}x \in U_{\ell,i}$. Since the diameter of $U_{\ell,i}$ is less than $1/\ell$, it follows that $\max_{i=1,\ldots,k} d(x, T^{in}x)<1/ \ell < \varepsilon$. Since $\varepsilon$ was arbitrary, this shows that $x \in E_k$.
To see that $E$ being a residual subset of $X$ implies \cref{thm:dvdw}, let $U \subseteq X$ be non-empty, open and let $k \in \mathbb{N}$. Since $E$ is a dense subset of $X$, there exists $x \in E \cap U$. Let $\varepsilon > 0$ such that $B(x,\varepsilon) \subseteq U$. Since $x$ is $k$-recurrent, there exists $n \in \mathbb{N}$ such that $x$, $T^nx$, \dots, $T^{nk}x \in B(x,\varepsilon) \subseteq U$, meaning that $U \cap T^{-n} U \cap \cdots \cap T^{-kn} U \neq \emptyset$.
Next we will demonstrate the equivalence of \cref{thm:dvdwreform} and $E$ being a residual subset of $X$. First, let us prove \cref{thm:dvdwreform} assuming that $E$ is residual. We will show that every $x \in E$ is such that for any open set $U\subseteq X$ containing $x$, the set $R(x,U)$ satisfies $d_\times^*(R(x,U)) = 1$. Let $U \subseteq X$ be an open set containing $x$. To show that $d_\times^*(R(x,U)) = 1$, we will show that for all finite $F \subseteq \mathbb{N}$, there exists $n \in \mathbb{N}$ such that $nF \subseteq R(x,U)$. Let $k\in \mathbb{N}$ be such that $F\subseteq \{1,\ldots, k\}$, and let $\varepsilon>0$ be such that $B(x,\varepsilon)\subseteq U$. Since $x$ is multiply recurrent, there exists $n\in \mathbb{N}$ such that for all $i \in \{1,\ldots,k\}$, $d(x, T^{in}x)<\varepsilon$. This implies that $\{n,2n,\ldots,kn\}\subseteq R(x,B(x,\varepsilon))\subseteq R(x,U)$, whereby $nF\subseteq R(x,U)$, as desired.
Finally, let us prove that \cref{thm:dvdwreform} implies that $E$ is residual. Let $X'\subseteq X$ be the residual set guaranteed by \cref{thm:dvdwreform}. We will show that $X'\subseteq E$, which will prove that $E$ is residual. Let $x\in X'$. Let $\varepsilon>0$, and set $U\coloneqq B(x,\varepsilon)$. Since $d_\times^*(R(x,U)) = 1$, for all finite $F \subseteq \mathbb{N}$, there exists $n \in \mathbb{N}$ such that $nF \subseteq R(x,U)$. In particular, for any $k\in \mathbb{N}$, there exists $n \in \mathbb{N}$ such that $\{n,2n,\ldots,nk\} \subseteq R(x,U)$, which implies that $\max_{i=1,\ldots,k}d(x,T^{in}x)<\varepsilon$. Since $\varepsilon > 0$ and $k \in \mathbb{N}$ were arbitrary, this shows that $x$ is a multiply recurrent point.
\end{proof}
Let $(X,T)$ be a topological dynamical system, $x \in X$, and $U \subseteq X$ be non-empty, open. Finding a configuration of the form $\{n,nm,nm^2\}$ in $R(x,U)$ is equivalent to showing that the $T \times T^{m} \times T^{m^2}$-orbit closure of $(x,x,x)$ in $X^3$ has non-empty intersection with $U\times U \times U$. This observation motivates the approach we use in the proof of \cref{thm:mainthm}.
The first step in the proof is to show that for any $\ell \in \mathbb{N}$ and $\vec m \in \mathbb{N}^\ell$, the $T^{\vec m}$-orbit closure of many points $\Delta(x) = (x,\ldots,x)$ along the diagonal in $X^{\Delta}$ supports a measure $\nu_x^{\Delta}$ whose marginals on $X$ give mass $\eta > 0$ to $U$. The second step is to use \cref{lem:joels-lemma} to find a mean with respect to which the point $\Delta(x)$ is $\nu_x^{\Delta}$-generic. Because each coordinate of $\Delta(x)$ spends an $\eta$-proportion of time in $U$ under $T^{\vec m}$, there must be many times for which an $\eta$-proportion of the coordinates are simultaneously in $U$. Szemer\'edi's theorem then allows us to finish the argument by taking $\vec m$ to be a sufficiently long geometric progression.
\begin{proposition}\label{thm:measureonorbitstwo}
Let $(X,T)$ be an invertible, minimal dynamical system, and let $U \subseteq X$ be open, non-empty. There exists $\eta > 0$ such that for all $\varepsilon > 0$, there exists $N \in \mathbb{N}$ such that for all finite sets $F \subseteq S_N$, there exists an $\varepsilon$-dense subset $X_\varepsilon \subseteq X$ such that for all $x \in X_\varepsilon$, there exists $F' \subseteq F$ with $|F'| > \eta |F|$ and $n \in \mathbb{N}$ such that $nF' \subseteq R(x,U)$.
\end{proposition}
\begin{proof}
Let $V \subseteq X$ open, non-empty such that $\overline{V} \subseteq U$, and let $\mu$ be any $T$-invariant probability measure on $X$. Since $T$ is minimal, $\eta \vcentcolon= (\mu(V)/2)^2 > 0$. Let $\varepsilon > 0$, and let $N \in \mathbb{N}$ be as given in \cref{thm:epsdenseorbits}.
Let $F = \{m_1, \ldots, m_\ell\} \subseteq S_N$, and put $\vec m = (m_1,\ldots,m_\ell) \in S_N^\ell$. Let
\[X^{\Delta} \vcentcolon= \overline{o}_{T^{\vec m}}\big(\Delta(X) \big).\]
Let $\mu^{\Delta}$ be any weak-$\ast$ limit point of the set $\big\{ N^{-1} \sum_{n=0}^{N-1} (T^{\vec m})^n \Delta(\mu) \ \big| \ N \in \mathbb{N} \big\}$, where $\Delta(\mu)$ denotes the push-forward of $\mu$ under the map $\Delta$. It follows that $\mu^{\Delta}$ is a $T^{\vec m}$-invariant probability measure on $X^{\Delta}$ with marginals $\pi_i \mu^{\Delta} = \mu$, where $\pi_i: X^{\Delta} \to X$ is the projection onto the $i^{\text{th}}$ coordinate.
By \cref{lem:orbislsc}, the map $\overline{o}_{T^{\vec m}}: X^{\Delta} \to \mathcal{F}(X^{\Delta})$ is lower semicontinuous, hence Borel. Denote by $\mathcal{B}_{X^{\Delta}}$ and $\mathcal{B}_{\mathcal{F}(X^{\Delta})}$ the Borel $\sigma$-algebras of $X^{\Delta}$ and $\mathcal{F}(X^{\Delta})$, respectively. Let $\mathcal{A}$ be the pull-back of $\mathcal{B}_{\mathcal{F}(X^{\Delta})}$ through $\overline{o}_{T^{\vec m}}$. Because $\overline{o}_{T^{\vec m}}$ is Borel, $\mathcal{A}$ is a sub-$\sigma$-algebra of $\mathcal{B}_{X^{\Delta}}$, and because $\mathcal{B}_{\mathcal{F}(X^{\Delta})}$ is countably generated, so too is $\mathcal{A}$. Note that $\mathfrak{a}(\vec x) \vcentcolon= \overline{o}_{T^{\vec m}}^{-1}\big(\{\overline{o}_{T^{\vec m}}(\vec x)\} \big) \in \mathcal{A}$ is the atom of $\mathcal{A}$ containing $\vec x \in X^{\Delta}$.
Disintegrating $\mu^{\Delta}$ with respect to $\mathcal{A}$ (see, e.g., \cite[Theorem 5.14]{EW11}), there exists a $\mu^{\Delta}$-co-null set $X^{\Delta}_0\subseteqX^{\Delta}$ and, for each $\vec x\inX^{\Delta}_0$, a Borel probability measure $\mu^{\Delta}_{\vec x}$ supported on $\mathfrak{a}(\vec x) \subseteq \overline{o}_{T^{\vec m}}(\vec x)$ such that $\mu^{\Delta}_{\vec x}=\mu^{\Delta}_{\vec y}$ whenever $\mathfrak{a}(\vec x)=\mathfrak{a}(\vec y)$ and such that
\[\mu^{\Delta} = \int_{X^{\Delta}_0} \mu^{\Delta}_{\vec x} \ d \mu^{\Delta} (\vec x).\]
By the essential uniqueness of this disintegration, the $T^{\vec m}$-invariance of $\mu^{\Delta}$, and the fact that $T^{\vec m} \mathfrak{a}(\vec x) = \mathfrak{a}(T^{\vec m} \vec x)$, it follows that $T^{\vec m}\mu^{\Delta}_{\vec x} = \mu^{\Delta}_{T^{\vec m}\vec x}$. Note that if $\vec x$ is a $T^{\vec m}$-recurrent point, then $\overline{o}_{T^{\vec m}}(\vec x) = \overline{o}_{T^{\vec m}}(T^{\vec m}\vec x)$. For such points, $\mathfrak{a}(\vec x) = \mathfrak{a}(T^{\vec m} \vec x)$, whereby $T^{\vec m} \mu^{\Delta}_{\vec x} = \mu^{\Delta}_{\vec x}$, meaning $\mu^{\Delta}_{\vec x}$ is $T^{\vec m}$-invariant. By \cite[Proposition 4.2.2]{brinstuckbook}, $\mu^{\Delta}$-almost every point $\vec x \in X^{\Delta}$ is $T^{\vec m}$-recurrent, so by passing to a $\mu^{\Delta}$-co-null subset of $X^{\Delta}_0$, we may assume that for all $\vec x \in X^{\Delta}_0$, the measure $\mu^{\Delta}_{\vec x}$ is $T^{\vec m}$-invariant.
For $i \in \{1, \ldots, \ell\}$, let
\[X^{\Delta}_i = \big\{ \vec x \in X^{\Delta}_0 \ \big| \ \pi_i \mu^{\Delta}_{\vec x}(\overline{V}) > \sqrt{\eta} \big\}.\]
Since
\[2 \sqrt{\eta} < \mu(\overline{V}) = \pi_i \mu^{\Delta}(\overline{V}) = \int_{X^{\Delta}_0} \pi_i \mu^{\Delta}_{\vec x}(\overline{V}) \ d \mu^{\Delta}(\vec x),\]
we have by Chebyshev's inequality that $\mu^{\Delta} (X^{\Delta}_i) > \sqrt{\eta}$. By the pigeonhole principle (with reasoning similar to that in the proof of \cref{lem:meanpigeon}), there exists $I \subseteq \{1, \ldots, \ell\}$ with $|I| > \sqrt{\eta} \ell$ for which $X^{\Delta}_0 \cap \bigcap_{i \in I} X^{\Delta}_i \neq \emptyset$. Let $\vec w$ be an element of this set.
Put $M = \text{{\sc lcm}}(\vec m)$. By \cref{prop:pointsofcontinuityalongthediagonal}, there exists a point $x \in X$ for which $\vec x \vcentcolon= \Delta(x)$ is a point of continuity of the map $\overline{o}_{T^{\vec m}}$. By \cref{thm:glasnersystemisminimal}, there exists a sequence $\big((a_n,b_n) \big)_{n \in \mathbb{N}} \subseteq \mathbb{N}^2$ for which $(T^{\vec m})^{a_n} \Delta(T)^{b_n}\vec w \to \vec x$ as $n \to \infty$. By passing to a subsequence if necessary, there exists $0 \leq b \leq M-1$ such that for all $n \in \mathbb{N}$, $b_n \equiv b \pmod M$. By \cref{prop:pointsofcontinuityalongthediagonal}, the points of continuity of the map $\overline{o}_{T^{\vec m}}$ are $\Delta(T)$-invariant. Therefore, by replacing $x$ with $T^{M-b}x$ and $b_n$ with $b_n + M - b$, we may assume that $b=0$, that is, that $(b_n)_{n \in \mathbb{N}} \subseteq M\mathbb{N}$.
Let $\nu^{\Delta}_x$ be a weak-$\ast$ limit point of the set $\big\{\Delta(T)^{b_n} \mu^{\Delta}_{\vec w} \ \big| \ n \in \mathbb{N} \big\}$; by passing to a subsequence, we may assume without loss of generality that $\Delta(T)^{b_n} \mu^{\Delta}_{\vec w} \to \nu^{\Delta}_x$ as $n \to \infty$. We will show now that $\nu^{\Delta}_x$ is a $T^{\vec m}$-invariant probability measure supported on $x^\Delta \vcentcolon= \overline{o}_{T^{\vec m}}\big(\Delta(x) \big)$ such that for all $i \in I$, $\pi_i \nu^{\Delta}_x (\overline{V}) > \sqrt{\eta}$.
That $\nu^{\Delta}_x$ is a probability measure is immediate from its definition. The measure $\nu^{\Delta}_x$ is $T^{\vec m}$-invariant because $\Delta(T)$ and $T^{\vec m}$ commute, are continuous, and $\mu^{\Delta}_{\vec w}$ is $T^{\vec m}$-invariant. Since $(T^{\vec m})^{a_n}\Delta(T)^{b_n} \vec w \to \vec x$ and $\vec x$ is a point of continuity of $\overline{o}_{T^{\vec m}}$, $\overline{o}_{T^{\vec m}}((T^{\vec m})^{a_n}\Delta(T)^{b_n}\vec w) \to x^\Delta$ as $n \to \infty$. Combined with the fact that $\mathop{\mathrm{supp}} ((T^{\vec m})^{a_n}\Delta(T)^{b_n}\mu^{\Delta}_{\vec w}) \subseteq \overline{o}_{T^{\vec m}}((T^{\vec m})^{a_n}\Delta(T)^{b_n}\vec w)$, \cref{lem:hausconvergence} gives that the measure $\nu^{\Delta}_x$ is supported on $x^\Delta$.
Let $i \in I$. Because $\mu^{\Delta}_{\vec w}$ is $T^{\vec m}$-invariant, the measure $\pi_i \mu^{\Delta}_{\vec w}$ is $T^{m_i}$-invariant, and hence, for all $n \in \mathbb{N}$, $T^{b_n}$-invariant. By properties of weak-$\ast$ convergence of measures and this invariance,
\begin{align}\label{eqn:measureboundedfrombelow}
\begin{aligned}
\pi_i \nu^{\Delta}_x (\overline{V}) &\geq \pi_i \limsup_{n \to \infty} (T^{\vec m})^{a_n}\Delta(T)^{b_n} \mu^{\Delta}_{\vec w} (\overline V)\\
&= \limsup_{n \to \infty} T^{b_n} \pi_i \mu^{\Delta}_{\vec w}(\overline V)\\
&= \pi_i \mu^{\Delta}_{\vec w}(\overline V) > \sqrt{\eta}.\end{aligned}
\end{align}
Applied to the system $(x^{\Delta},T^{\vec m},\nu^{\Delta}_x)$, \cref{lem:joels-lemma} gives the existence of an additively invariant mean $\lambda$ on $\mathbb{N}$ such that for all $g \in C(x^{\Delta})$,
\[ \lambda \big(n \mapsto g(T^{m_1 n}x, T^{m_2 n}x,\ldots,T^{m_\ell n}x) \big) = \int_{x^{\Delta}} g \ d\nu^{\Delta}_x.\]
By Urysohn's lemma, there exists a continuous function $f:X\to[0,1]$ with $f(y)=1$ for all $y\in \overline{V}$ and $f(y)=0$ for all $y\notin U$. By (\ref{eqn:measureboundedfrombelow}), for all $i \in I$,
\begin{align*}
\lambda \big(R(x,U)/m_i\big) &\geq \lambda \big( n \mapsto f(T^{m_i n}x)\big)\\
&= \lambda \big( n \mapsto ( f \circ \pi_i)(T^{m_1 n}x, T^{m_2 n}x,\ldots,T^{m_\ell n}x)\big)\\
&= \int_{x^{\Delta}} f \circ \pi_i \ d\nu^{\Delta}_x\\
&\geq \pi_i\nu^{\Delta}_x(\overline{V}) > \sqrt{\eta}.
\end{align*}
Put $X_\varepsilon = \{T^{Mk} x \ | \ k \in \mathbb{N}\}$. Since $M \in S_N$, by \cref{thm:epsdenseorbits}, the set $X_\varepsilon$ is $\varepsilon$-dense in $X$. Therefore, to conclude the proof, it suffices to show that for every $k \in \mathbb{N}$, there exists $F' \subseteq F$ with $|F'| > \eta |F|$ and $n \in \mathbb{N}$ such that $nF' \subseteq R(T^{Mk}x,U)$.
Let $k \in \mathbb{N}$. By the translation invariance of $\lambda$, for all $i \in I$,
\[\lambda \left(\frac{R(T^{Mk}x,U)}{m_i} \right) = \lambda \left( \frac{R(x,U) - Mk}{m_i} \right) = \lambda \left(\frac{R(x,U)}{m_i} - \frac{M}{m_i}k\right) > \sqrt{\eta}.\]
It follows by \cref{lem:meanpigeon} that there exists $I' \subseteq I$ with $|I'| > \sqrt{\eta} |I|$ and $n \in \mathbb{N}$ such that for all $i \in I'$, $m_i n \in R(T^{Mk}x,U)$. Setting $F' = \{m_i \ | \ i \in I'\}$, we see $|F'| > \eta |F|$ and $nF' \subseteq R(T^{Mk}x,U)$, as was to be shown.
\end{proof}
Now we can prove \cref{thm:mainthm}, restated here.
\begin{mainthm*}
Let $(X,T)$ be a minimal dynamical system. There exists a residual set $X' \subseteq X$ such that for all $x \in X'$ and all non-empty, open $U \subseteq X$, the set $R(x,U)$ contains arbitrarily long geometric progressions.
\end{mainthm*}
\begin{proof}
It suffices to prove the statement for $(X,T)$ invertible. To see why, let $\pi: (W,T) \to (X,T)$ be the natural extension (\cref{def:naturalext}). Suppose that \cref{thm:mainthm} holds for $(W,T)$: there exists a residual set $W' \subseteq W$ such that for all $w \in W'$ and all non-empty, open $V \subseteq X$, the set $R(w,V)$ contains arbitrarily long geometric progressions. By \cref{lem:imageofresidual,lem:naturalextensionisminimal}, the set $X' \vcentcolon= \pi W'$ is residual. Let $x \in X'$ and $U \subseteq X$ open, non-empty. Choose $w \in \pi^{-1}(\{x\}) \cap W'$, and note that $R(w,\pi^{-1} U) = R(x,U)$. It follows that $R(x,U)$ contains arbitrarily long geometric progressions.
By taking a countable basis of open sets and a countable intersection of residual sets, the residual set $X'$ is allowed to depend on the set $U$. Let $U \subseteq X$ be open, non-empty. Let $\eta > 0$ be as guaranteed by \cref{thm:measureonorbitstwo} for the set $U$.
For $\ell \in \mathbb{N}$, put
\[G_\ell \vcentcolon= \bigcup_{\substack{m,n \in \mathbb{N} \\ m \geq 2}} \big(T^{-nm} U \cap \cdots \cap T^{-nm^\ell} U \big).\]
This is precisely the set of points $x \in X$ for which $R(x,U)$ contains a geometric progression of length $\ell$. The set of those points $x \in X$ for which $R(x,U)$ contains arbitrarily long geometric progressions is thus $X'\vcentcolon= \bigcap_{\ell \in \mathbb{N}} G_\ell$. We will show that $X'$ is residual by showing that each $G_\ell$ is open and dense in $X$. Since $G_\ell$ is open by definition, we have only to show that $G_\ell$ is $\varepsilon$-dense in $X$ for any $\varepsilon > 0$.
Fix $\ell \in \mathbb{N}$ and $\varepsilon > 0$. Let $N \in \mathbb{N}$ be as guaranteed by \cref{thm:measureonorbitstwo}. By Szemer\'edi's theorem \cite{szemeredi} and the argument in the proof of \cref{thm:multszem}, there exists $L \in \mathbb{N}$ such that any subset of any geometric progression of length $L$ of relative density at least $\eta$ contains a geometric progression of length $\ell$. Let $F \subseteq S_N$ be a geometric progression of length $L$. Let $X_\varepsilon \subseteq X$ be as guaranteed by \cref{thm:measureonorbitstwo}. To show that $G_\ell$ is $\varepsilon$-dense, it suffices now to show that $X_\varepsilon \subseteq G_\ell$.
Let $x \in X_\varepsilon$. By \cref{thm:measureonorbitstwo}, there exists $F' \subseteq F$ with $|F'| > \eta |F|$ and $n \in \mathbb{N}$ such that $nF' \subseteq R(x,U)$. By Szemer\'edi's theorem, the set $F'$ contains a geometric progression of length $\ell$, hence so does $nF'$. Because $R(x,U)$ contains a geometric progression of length $\ell$, the point $x$ belongs to $G_\ell$. It follows that $X_\varepsilon \subseteq G_\ell$, as was to be shown.
\end{proof}
\section{Results on totally minimal and distal systems}\label{sec:proofofdistalresults}
To prove \cref{thm:mainthmwithtotallymin,thm:mainthmwithdistal}, we need the following strengthening of \cref{thm:measureonorbitstwo} using total visibility; recall \cref{def:totallyvisible}.
\begin{proposition}\label{prop:existenceofmean}
Let $(X,T)$ be an invertible, minimal system and $V \subseteq U \subseteq X$ be open, non-empty sets with $\overline{V} \subseteq U$. If $V$ is totally visible, then there exists $\eta > 0$ and a residual set $X' \subseteq X$ such that for all $x \in X'$ and all finite sets $F \subseteq \mathbb{N}$, there exists $F' \subseteq F$ with $|F'| > \eta |F|$ and an additively invariant mean $\lambda$ on $\mathbb{N}$ such that for all $f \in F'$, $\lambda \big(R(x,U)/f \big) > \eta$.
\end{proposition}
\begin{proof}
Let $\mu$ be a $T$-invariant probability measure on $X$ for which $V$ is totally visible. Let $\eta > 0$ be a third of the infimum from \cref{def:totallyvisible}. By taking a countable intersection of residual sets, it suffices to show: \emph{for all finite sets $F \subseteq \mathbb{N}$, there exists a residual set $X' \subseteq X$ such that for all $x \in X'$, there exists $F' \subseteq F$ with $|F'| > \eta |F|$ and an additively invariant mean $\lambda$ on $\mathbb{N}$ such that for all $f \in F'$, $\lambda \big(R(x,U)/f \big) > \eta$.}
Let $F = \{m_1, \ldots, m_\ell\} \subseteq \mathbb{N}$, and put $\vec m = (m_1,\ldots,m_\ell) \in \mathbb{N}^\ell$ and $M = \text{{\sc lcm}}(\vec m)$. Let $X' \subseteq X$ be the set of points $x \in X$ such that $\Delta(x)$ is a point of continuity of the map $\overline{o}_{T^{\vec m}}: X^{\Delta} \to \mathcal{F}(X^{\Delta})$. By \cref{prop:pointsofcontinuityalongthediagonal}, $X'$ is a residual subset of $X$. By the same proposition, for $j \in \{0, \ldots, d_{M}-1\}$, the set $X_{M,j}' \vcentcolon= X' \cap X_{M,j}$ is the set of points $x \in X_{M,j}$ for which $\Delta(x)$ is a point of continuity of the map $\overline{o}_{T^{\vec m}}: X^{\Delta}_{M,j} \to \mathcal{F}(X^{\Delta}_{M,j})$, and $X_{M,j}'$ is residual in $X_{M,j}$.
Fix $j \in \{0, \ldots, d_{M}-1\}$. The measure $\mu_{M,j}$ is supported on $X_{M,j}$ and, because $V$ is totally visible by $\mu$, satisfies: for all $i \in \{1,\ldots,\ell\}$,
\begin{align}\label{eqn:liminfiscorrect}\liminf_{N \to \infty} \frac 1N \sum_{k=0}^{N-1}T^{m_ik}\mu_{M,j}(\overline{V}) = \liminf_{N \to \infty} \frac 1N \sum_{k=0}^{N-1}\mu_{M,j + m_ik}(\overline{V}) > 2 \eta.\end{align}
Let $\mu^{\Delta}_{M,j}$ be any weak-$\ast$ limit point of the set $\big\{ N^{-1} \sum_{n=0}^{N-1} (T^{\vec m})^n \Delta(\mu_{M,j}) \ \big| \ N \in \mathbb{N} \big\}$. It follows that $\mu^{\Delta}_{M,j}$ is a $T^{\vec m}$-invariant probability measure on $X^{\Delta}_{M,j}$, and (\ref{eqn:liminfiscorrect}) implies that for all $i \in \{1,\ldots, \ell\}$, $\pi_i \mu^{\Delta}_{M,j} (\overline{V}) > 2 \eta$, where $\pi_i: X^{\Delta} \to X$ is the projection onto the $i^{\text{th}}$ coordinate.
At this point, we repeat verbatim Paragraphs~3 through 5 of the proof of \cref{thm:measureonorbitstwo} with $X^{\Delta}_{M,j}$ in place of $X^{\Delta}$, $\mu^{\Delta}_{M,j}$ in place of $\mu^{\Delta}$, and with $\eta$ in place of $\sqrt{\eta}$. We get $I \subseteq \{1, \ldots, \ell\}$ with $|I| > \eta \ell$, $\vec w \in X^{\Delta}_{M,j}$, and a $T^{\vec m}$-invariant probability measure\footnote{Perhaps a more fitting notation for this measure would be $\mu^{\Delta}_{M,j,\vec w}$, which we avoid for notational simplicity.} $\mu^{\Delta}_{\vec w}$ supported on $\overline{o}_{T^{\vec m}}(\vec w)$ with the property that for all $i \in I$, $\pi_i \mu^{\Delta}_{\vec w}(\overline{V}) > \eta$.
Let $x \in X_{M,j}'$, and write $\vec x \vcentcolon= \Delta(x)$. By \cref{thm:subsystemsareminimal}, the system $(X^{\Delta}_{M,j},T^{\vec m},\allowbreak \Delta(T)^M)$ is minimal, so there exists a sequence $\big((a_n,b_n) \big)_{n \in \mathbb{N}} \subseteq \mathbb{N}^2$ for which $(T^{\vec m})^{a_n} \Delta(T)^{b_n M}\vec w \to \vec x$ as $n \to \infty$.
Let $\nu^{\Delta}_x$ be a weak-$\ast$ limit point of the set $\big\{\Delta(T)^{b_n M} \mu^{\Delta}_{\vec w} \ \big| \ n \in \mathbb{N} \big\}$; by passing to a subsequence, we may assume without loss of generality that $\Delta(T)^{b_n M} \mu^{\Delta}_{\vec w} \to \nu^{\Delta}_x$ as $n \to \infty$. By repeating verbatim Paragraphs~8 and 9 in the proof of \cref{thm:measureonorbitstwo}, we show that $\nu^{\Delta}_x$ is a $T^{\vec m}$-invariant probability measure supported on $\overline{o}_{T^{\vec m}}\big(\Delta(x) \big)$ such that for all $i \in I$, $\pi_i \nu^{\Delta}_x (\overline{V}) > \eta$. An application of \cref{lem:joels-lemma} just as in Paragraph~10 in the proof of \cref{thm:measureonorbitstwo} gives the existence of an additively invariant mean $\lambda$ on $\mathbb{N}$ such that for all $f \in F' \vcentcolon= \{m_i \ | \ i \in I\}$, $\lambda \big(R(x,U)/f \big) > \eta$, as was to be shown.
\end{proof}
It was explained in \cref{rmk:totallyminimalsystems} that all non-empty, open sets in a totally minimal system are totally visible by every $T$-invariant probability measure. We use this fact to prove \cref{thm:mainthmwithtotallymin}.
\begin{proof}[Proof of \cref{thm:mainthmwithtotallymin}]
By the same initial argument in the proof of \cref{thm:mainthm}, it suffices to prove this theorem in the case that $(X,T)$ is invertible.
By taking a countable basis of open sets and a countable intersection of residual sets, the residual set $X'$ is allowed to depend on the set $U$. Let $U \subseteq X$ be open, non-empty, and let $V \subseteq U$ be open, non-empty such that $\overline{V} \subseteq U$. Since $(X,T)$ is totally minimal, the set $V$ is totally visible. Let $\eta > 0$ and $X' \subseteq X$ be as guaranteed by \cref{prop:existenceofmean}. We will show that for all $x \in X'$, the set $R(x,U)$ satisfies $d_\times^*(R(x,U)) \geq \eta^2$
Let $x \in X'$ and $F \subseteq \mathbb{N}$ be finite. Let $F' \subseteq F$ and $\lambda$ be as guaranteed by \cref{prop:existenceofmean}: $|F'| > \eta |F|$, and for all $f \in F'$, $\lambda \big(R(x,U)/f \big) > \eta$. By \cref{lem:meanpigeon}, there exists $F'' \subseteq F'$ with $|F''| > \eta |F'| > \eta^2 |F|$ and $n \in \mathbb{N}$ such that $nF'' \subseteq R(x,U)$. Since $F \subseteq \mathbb{N}$ was arbitrary, this shows $d_\times^*(R(x,U)) \geq \eta^2$.
\end{proof}
The following lemma gives a sufficient condition on a set $A \subseteq \mathbb{N}$ for all of its translates to have positive multiplicative density in a coset of a multiplicative semigroup, and it will allow us to prove \cref{thm:mainthmwithdistal,thm:translatesofiprstararelarge}.
\begin{lemma}\label{lem:gprichtranslates}
Let $A \subseteq \mathbb{N}$, and suppose that there exists $\eta > 0$ and $N \in \mathbb{N}$ for which the following holds:
\begin{align}\label{eqn:propforrichtranslations} \begin{gathered} \text{for all $F \subseteq \mathbb{N}$ and $a \in \mathbb{Z}$, there exists $F' \subseteq F$ with $|F'| > \eta |F|$}\\ \text{and a translation invariant mean $\lambda$ on $\mathbb{N}$ such that} \\ \text{for all $f \in F'$, $\lambda \left( \frac{A-aN}{Nf} \right) > \eta$.}\end{gathered}\end{align}
Then, for all $t \in \mathbb{Z}$,
\[d^*_{(N,t) S_{N/(N,t)}}\big(A+t \big) \geq \eta^2 (N,t) / N.\]
\end{lemma}
\begin{proof}
Let $t \in \mathbb{Z}$, and put $K=N/(N,t)$. We will show that $d^*_{S_K}\big((A+t)/(N,t)\big) \geq \eta^2 / K$.
Let $F \subseteq S_K$ be finite, and let $F' \subseteq F$ with $|F'| \geq |F| / K$ be such that all elements of $F'$ are congruent modulo $K$ to some $f_0 \in S_K$. Denote by $\Pi F'$ the product of the elements of $F'$. Since $(\Pi F',K) = 1$, there exists $a \in \mathbb{Z}$ and $b \in \mathbb{N}$ with $(b,K) = (t / (N,t),K) = 1$ so that
\[b \Pi F' - aK = \frac t{(N,t)}.\]
Put $c = bf_0^{|F'| - 1}$ and note that $(c,K) = 1$. Let $f \in F'$. Since $b \Pi F' / f \equiv c \pmod {K}$, $b\Pi F' \equiv c f \pmod {fK}$. Therefore,
\begin{align}
\notag c f - aK &\equiv \frac t{(N,t)} \pmod {fK}, \text{ whereby}\\
\label{eqn:mainequivalence} c (N,t) f - aN &\equiv t \pmod {fN}.
\end{align}
Summarizing, we have found $a \in \mathbb{Z}$ and $c \in \mathbb{N}$ with $(c,K) = 1$ so that for all $f \in F'$, the congruence in (\ref{eqn:mainequivalence}) holds.
By the assumptions in (\ref{eqn:propforrichtranslations}), there exists $F'' \subseteq F'$ with $|F''| > \eta|F'|$ and a translation invariant mean $\lambda$ on $\mathbb{N}$ such that for all $f \in F''$, $\lambda \left( (A-aN)/(Nf) \right) > \eta$. For all $f \in F''$, $-aN \equiv t-c (N,t)f \pmod{Nf}$, so by the translation invariance of $\lambda$,
\[\lambda \left( \frac{A+t-c(N,t)f}{Nf} \right) = \lambda \left( \frac{A-aN}{Nf} \right) > \eta.\]
By \cref{lem:meanpigeon}, there exists $F''' \subseteq F''$ with $|F'''| > \eta |F''|$ such that $\bigcap_{f \in F'''} (A+t-c(N,t)f)/(Nf)$ is non-empty; let $n$ be an element of this set. Now $Nn + c(N,t) \in \mathbb{N}$ is such that $(Nn+c(N,t))F''' \subseteq A+t$, meaning that
\[(Kn+c)F''' \subseteq \frac{A+t}{(N,t)}.\]
Since $(Kn+c,K) = (c,K) = 1$, we have shown that $Kn+c \in S_K$ satisfies $|(Kn+c)F \cap (A+t)/(N,t)| \geq |F'''| > \eta^2 |F| / K$. Since $F \subseteq S_K$ was arbitrary, this shows $d^*_{S_K}\big((A+t)/(N,t)\big) \geq \eta^2 / K$.
\end{proof}
We are now able to prove the following theorem, a strengthening of \cref{thm:mainthmwithdistal}.
\begin{theorem}\label{thm:mainthmwithdistalbetter}
Let $(X,T)$ be a minimal distal system. There exists a residual set $X' \subseteq X$ such that for all non-empty, open $U \subseteq X$, there exists $N \in \mathbb{N}$ and $\eta > 0$ such that for all $x \in X'$, there exists $i \in \mathbb{N}$ such that for all $t \in \mathbb{Z}$,
\[d_{(N,t)S_{N/(N,t)}}^* \big( R(x,U)+i+t \big) \geq \eta (N,t) / N.\]
In particular, putting $t=-i$, we see that the set $R(x,U)$ has positive multiplicative density in a coset of a multiplicative subsemigroup of $\mathbb{N}$.
\end{theorem}
\begin{proof}
Because $(X,T)$ is distal, it is invertible. By taking a countable basis of open sets and a countable intersection of residual sets, the residual set $X'$ is allowed to depend on the set $U$. Let $U \subseteq X$ be open, non-empty, and let $V \subseteq U$ be open, non-empty with $\overline{V} \subseteq U$. By \cref{lem:distalhavegoodmeasure}, there exists $N \in \mathbb{N}$ and $j \in \{0,\ldots,d_N-1\}$ such that the set $V_{N,j}$ is totally visible in the system $(X_{N,j},T^N)$.
Now $(X_{N,j},T^N)$ is an invertible, minimal system, $\overline{V_{N,j}} \subseteq U_{N,j}$, and $V_{N,j}$ is totally visible. Let $\sigma > 0$ be the ``$\eta$'' as guaranteed by \cref{prop:existenceofmean}, and let $X_{N,j}' \subseteq X_{N,j}$ be as guaranteed by the same proposition. Put $\eta = \sigma^2$. Since $X_{N,j}'$ is residual, so is $\bigcap_{n \in \mathbb{Z}} T^{Nn} X_{N,j}'$; thus, replacing the former set with the latter, we may assume that $X_{N,j}'$ is $T^N$-invariant.
We will verify next that for every $x \in X_{N,j}'$, the set $R(x,U)$ satisfies the conditions in (\ref{eqn:propforrichtranslations}) in \cref{lem:gprichtranslates} with $\sigma$ as ``$\eta$'' and $N$ as it is. Let $x \in X_{N,j}'$ and put $A = R(x,U)$. Let $F \subseteq \mathbb{N}$ and $a \in \mathbb{Z}$. Because $X_{N,j}'$ is $T^N$-invariant, $T^{aN}x \in X_{N,j}'$. By \cref{prop:existenceofmean}, there exists $F' \subseteq F$ with $|F'| > \sigma |F|$ and an additively invariant mean $\lambda$ on $\mathbb{N}$ such that for all $f \in F'$,
\[\lambda\left( \frac{A-aN}{Nf} \right) = \lambda \left( \frac{R_{T^N}(T^{aN}x,U)}{f} \right) > \sigma.\]
This shows that the conditions in \cref{lem:gprichtranslates} are satisfied.
Put $X' = \bigcup_{i=0}^{d_N-1} T^i X_{N,j}'$. Since $T$ is a homeomorphism, the set $X'$ is residual. We will show that every $x \in X'$ satisfies the conclusions of the theorem.
Let $x \in X'$. By the definition of $X'$, there exists $i \in \mathbb{N}$ so that $T^{-i} x \in X_{N,j}'$. Since $R(T^{-i} x,U) = R(x,U) + i$ satisfies the conditions in \cref{lem:gprichtranslates}, for all $t \in \mathbb{Z}$,
\[d^*_{(N,t)S_{N/(N,t)}}\big( R(x,U)+i+t \big) \geq \sigma^2 (N,t) / N = \eta (N,t) / N,\]
as was to be shown.
\end{proof}
\section{Results on \texorpdfstring{IP$_r^*$}{IPr*} sets and nilsystems}\label{sec:combresults}
Not every syndetic subset of $\mathbb{N}$ arises as the set of returns $R(x,U)$ in a dynamical system (see \cref{ex:setwithoutipnaughtstarshifts,ex:propertyofnilsystem} below), so \cref{thm:mainthm} lends only some evidence in favor of the conjecture that all syndetic subsets of $\mathbb{N}$ contain arbitrarily long geometric progressions. In this section, we show that a subclass of syndetic sets, translates of $\text{IP}_r^*$ sets, do have positive multiplicative density in cosets of multiplicative subsemigroups of $\mathbb{N}$ and, hence, are GP-rich.
\begin{definition}
A subset of $\mathbb{N}$ is called \emph{$\text{IP}_r$}, $r \in \mathbb{N}$, if it contains a \emph{finite sums set with $r$ generators}, a set of the form
\begin{align}\text{FS}(x_1, \ldots, x_r) \vcentcolon= \left\{ \sum_{i \in I} x_i \ \middle| \ \emptyset \neq I \subseteq \{1, \ldots, r\} \right\}, \quad x_1, \ldots, x_r \in \mathbb{N}.\end{align}
A set which is $\text{IP}_r$ for all $r \in \mathbb{N}$ is called \emph{$\text{IP}_0$}. A subset of $\mathbb{N}$ is called \emph{$\text{IP}_r^*$} if it has non-empty intersection with every $\text{IP}_r$ set in $\mathbb{N}$, and it is called \emph{$\text{IP}_0^*$} if it has non-empty intersection with every $\ip_0$ set in $\mathbb{N}$ (equivalently, if it is $\text{IP}_r^*$ for some $r \in \mathbb{N}$). The \emph{rank} of an $\text{IP}_0^*$ set is the minimal $r \in \mathbb{N}$ for which it is an $\text{IP}_r^*$ set.
\end{definition}
As a warm-up to the proof of \cref{thm:translatesofiprstararelarge}, we will show that every $\text{IP}_r^*$ set $A$ is syndetic. Let $F \subseteq \mathbb{N} \setminus A$ be a maximal finite sums set, and put $F_0 = F \cup \{0\}$. For every $m \in \mathbb{N}$, $F \cup (F_0+m)$ is a finite sums set that, by the maximality of $F$, has non-empty intersection with $A$. Since $A \cap F = \emptyset$, it must be that $A \cap (F_0+m) \neq \emptyset$, meaning $m \in A-F_0$. Since $m$ was arbitrary, $A-F_0 = \mathbb{N}$, meaning $A$ is syndetic: the set $F_0 + 1 \subseteq \mathbb{N}$ is such that $\bigcup_{f \in F_0+1} (A-f) = \mathbb{N}$.
In fact, we get a quantitative measure on the syndeticity of the set $A$. Since $A$ is $\text{IP}_r^*$, $F$ is $\text{IP}_s$ for some $s \leq r-1$, meaning $|F_0+1| \leq 2^s$. It follows by additivity that for any additively invariant mean $\lambda$ on $\mathbb{N}$, $\lambda(A) \geq 2^{-s}$. This is the basis for applying \cref{lem:gprichtranslates} in the proof of the following theorem, from which \cref{thm:translatesofiprstararelarge} immediately follows.
\begin{theorem}\label{thm:translatesofiprstararelargebetter}
Let $A \subseteq \mathbb{N}$ be an $\text{IP}_r^*$ set. There exists an $N \in \mathbb{N}$ such that for all $t \in \mathbb{Z}$,
\begin{align}\label{eqn:inequalityfortranslatediprstarset}d^*_{(N,t)S_{N/(N,t)}} \big( A+t \big) \geq \frac{(N,t)}{2^{2r+2}N}.\end{align}
\end{theorem}
\begin{proof}
It is quick to check that for all $n \in \mathbb{N}$, $\mathop{\mathrm{rank}}(A/n) \leq \mathop{\mathrm{rank}}(A) \leq r$. It follows that there exists $N \in \mathbb{N}$ for which $\mathop{\mathrm{rank}}(A/N) = \min_{n \in \mathbb{N}} \mathop{\mathrm{rank}}(A/n)$. Let $s = \mathop{\mathrm{rank}}(A/N) \leq r$ be this minimal rank.
We will show that $A$ satisfies the conditions in \cref{lem:gprichtranslates} with $N$ as it is and $\eta = 2^{-(s+1)}$. It suffices to show the following: \emph{for all $f \in \mathbb{N}$ and $a \in \mathbb{Z}$, at most $2^{s}$-many translates of the set $(A - aN)/(Nf) = (A/N - a)/f$ are sufficient to cover all but finitely many elements of $\mathbb{N}$.} Indeed, this ensures that for every translation invariant mean $\lambda$ on $\mathbb{N}$, $\lambda\big( (A - aN) / (Nf) \big) \geq 2^{-s} > \eta$.
Let $f \in \mathbb{N}$ and $a \in \mathbb{Z}$. Since $\mathop{\mathrm{rank}}(A/fN) = s$, there exists an $\text{IP}_{s-1}$ set $G' \subseteq \mathbb{N} \setminus (A/fN)$. (If $s = 1$, then take $G' = \emptyset$.) Set $G \vcentcolon= fG' \cup\{0\}$, and note that $|G| \leq 2^s$. We claim that $A/N - G = \mathbb{N}$. Let $m \in \mathbb{N}$. Since $(fG') \cup \big(G + m\big)$ is an $\text{IP}_s$ set and $(fG') \cap (A/N) = \emptyset$ and $\mathop{\mathrm{rank}}(A/N) = s$, we see that $(A/N) \cap \big(G + m\big) \neq \emptyset$, implying that $m \in A/N - G$.
Since $A/N - G = \mathbb{N}$, the set $\mathbb{N} \setminus \big(A/N - a - G\big)$ is finite. Dividing by $f$, we see that $\mathbb{N} \setminus \big((A/N - a)/f - (G' \cup\{0\})\big)$ is co-finite, as was to be shown.
\end{proof}
As a corollary to \cref{thm:translatesofiprstararelargebetter} and \cref{thm:geoarithmeticpatterns}, we see that translates of $\text{IP}_r^*$ sets contain arbitrarily long geo-arithmetic configurations and so, in particular, are GP-rich.
We now derive three consequences of \cref{thm:translatesofiprstararelargebetter} based on the connection between nilsystems and $\ip_0^*$ sets discussed in \cref{sec:results}.
A \emph{nilsystem} is a topological dynamical system $(X,T)$ where $X$ is a compact homogeneous space of a nilpotent Lie group $G$ and $T$ is a translation of $X$ by an element of $G$. The key fact in each of these consequences follows from \cite[Theorem 0.2]{BLiprstarcharacterization}: \emph{in a nilsystem $(X,T)$, for all non-empty, open $U \subseteq X$ and all $x \in U$, the set $R(x,U)$ is $\ip_0^*$.}\footnote{Strictly speaking, \cite[Theorem 0.2]{BLiprstarcharacterization} concerns invertible systems and $\ip_0^*$ subsets of $\mathbb{Z}$. This theorem applies to our setting by noting that nilsystems are invertible and that if a set $A$ is $\text{IP}_r^*$ in $\mathbb{Z}$, then $A \cap \mathbb{N}$ is $\text{IP}_r^*$ in $\mathbb{N}$.}
First, we show that sets of returns in minimal nilsystems have positive multiplicative density.
\begin{corollary}
Let $(X,T)$ be a minimal nilsystem. For all $x \in X$ and all non-empty, open $U \subseteq X$, the set $R(x,U)$ has positive multiplicative density in a coset of a multiplicative subsemigroup of $\mathbb{N}$.
\end{corollary}
\begin{proof}
Because $(X,T)$ is minimal, there exists $n \in \mathbb{N}$ such that $T^n x \in U$. By \cite[Theorem 0.2]{BLiprstarcharacterization}, the set $R(T^n x, U)$ is $\ip_0^*$. It follows from \cref{thm:translatesofiprstararelargebetter} that translates of $R(T^n x,U)$ have positive multiplicative density in a coset of a multiplicative subsemigroup of $\mathbb{N}$. To finish, note that $R(x,U) \supseteq R(T^nx,U) + n$, a translate of $R(T^n x,U)$.
\end{proof}
A natural question is whether or not our main results can be enhanced by improving ``positive upper Banach density'' to ``multiplicatively piecewise syndetic.'' A set $A \subseteq \mathbb{N}$ is \emph{multiplicatively syndetic} if there exists a finite set $F \subseteq \mathbb{N}$ such that $\bigcup_{f \in F} A/f = \mathbb{N}$. A set $C \subseteq \mathbb{N}$ is \emph{multiplicatively piecewise syndetic} if there exists a multiplicatively syndetic set $A \subseteq \mathbb{N}$ and a set $B \subseteq \mathbb{N}$ with $d_\times^*(B) = 1$ such that $C = A \cap B$. Multiplicatively piecewise syndetic sets have positive multiplicative upper Banach density, but, by \cite[Theorem 6.4]{bcrzpaper}, there exist subsets of $\mathbb{N}$ of multiplicative density arbitrarily close to $1$ that are not multiplicatively piecewise syndetic.
We will argue now that the conclusion of \cref{thm:mainthmwithtotallymin} cannot be improved to show that the set $R(x,U)$ is, in general, multiplicatively piecewise syndetic. Suppose $(X,T)$ is a totally minimal nilsystem and $U$ and $V$ are non-empty, disjoint open sets. Let $x \in U$ and put $A = R(x,U)$ and $B = R(x,V)$. Since $U \cap V = \emptyset$, $A \cap B = \emptyset$. By \cite[Theorem 0.2]{BLiprstarcharacterization}, the set $A$ is an $\ip_0^*$ set, and \cite[Corollary 7.3]{BGpaperonearxiv} gives that $A$ has non-empty intersection with all multiplicatively piecewise syndetic subsets of $\mathbb{N}$. Since $A \cap B = \emptyset$, it follows that the set $B$ is not multiplicatively piecewise syndetic.
Third, we prove \cref{cor:polyapprox} from the introduction, an application of \cref{thm:translatesofiprstararelargebetter} to finding geo-arithmetic configurations in sets arising from polynomial Diophantine approximation.
\begin{proof}[Proof of \cref{cor:polyapprox}]
By writing the set $A$ as a set of return times of a point to an open set in a minimal nilsystem, it is shown in \cite[Theorem 6.14]{BGpaperonearxiv} that $A$ is a translated $\ip_0^*$ set. It follows by \cref{thm:translatesofiprstararelargebetter} that there exists $K \in \mathbb{N}$ such that $A$ has positive multiplicative upper Banach density in a coset of the multiplicative subsemigroup $S_K$. The stated geo-arithmetic configurations can be found in $A$ by using \cref{thm:geoarithmeticpatterns}.
\end{proof}
\section{Syndetic sets not arising from dynamics}\label{sec:concludingremarks}
As mentioned in \cref{sec:historyandcontext}, the fact that geometric progressions and multiplicative density are not translation invariant prevents us from being able to deduce results on arbitrary syndetic subsets of $\mathbb{N}$ from our dynamical ones. Still, one might hope that an arbitrary syndetic set or $\text{IP}_r^*$ set takes the form $R(x,U)$, or at least contains a set of the form $R(x,U)$, where $x$ and $U$ are a point and a non-empty, open set in a minimal system or nilsystem. We show in this section that this is not the case.
\begin{lemma}\label{lem:propertyofdynamicsyndetic}
Let $(X,T)$ be a minimal system, $x \in X$, and $U \subseteq X$ open, non-empty. There exists $N \in \mathbb{N}$ such that for all $t \in \mathbb{N}$ and all $n \in S_N$, the set $(R(x,U) - t) / n$ is syndetic.
\end{lemma}
\begin{proof}
Let $N \in \mathbb{N}$ be from \cref{lem:totalminimalatresolutionU}, and let $t \in \mathbb{N}$ and $n \in S_N$. Since $\bigcup_{\ell=1}^\infty (T^n)^{-\ell} U=X$ and $X$ is compact, for all $y \in X$, the set $R_{T^n}(y,U)$ is syndetic; in particular, the set $R_{T^n}(T^tx,U) = (R(x,U) - t) / n$ is syndetic, as was to be shown.
\end{proof}
\begin{example}\label{ex:setwithoutipnaughtstarshifts}
There exists a syndetic set $A \subseteq \mathbb{N}$ such that for all $t,n \in \mathbb{N}$ with $n \geq 2$, the set $(A-t)/n$ is not syndetic. By \cref{lem:propertyofdynamicsyndetic}, it follows that $A$ does not contain a set of the form $R(x,U)$ where $x$ and $U$ are a point and a non-empty, open subset of a minimal dynamical system.
To construct such a set, let $\{T_{t,n}\}_{t,n \in \mathbb{N}, n \geq 2}$ be a family of thick subsets of $\mathbb{N}$ (i.e., subsets containing arbitrarily long intervals) with the property that if $m \in T_{t,n}$ and $m + 1 \in T_{t',n'}$, then $t = t'$ and $n=n'$. Put
\[A = \left(\bigcup_{\substack{t,n \in \mathbb{N} \\ n \geq 2}} \big(T_{t,n} \setminus (n\mathbb{N}+t) \big)\right) \cup \left(\mathbb{N} \setminus \bigcup_{\substack{t,n \in \mathbb{N} \\ n \geq 2}} T_{t,n} \right).\]
We claim that $A$ is syndetic; in fact, we will show that $A \cup (A-1) = \mathbb{N}$. Indeed, let $m \in \mathbb{N}$. If $m \not\in \bigcup T_{t,n}$, then $m \in A$. Otherwise, there exists $t, n \in \mathbb{N}$ with $n \geq 2$ such that $m \in T_{t,n}$. If $m+1 \not\in \bigcup T_{t,n}$, then $m \in A-1$. Otherwise, $m+1 \in \bigcup T_{t,n}$, which implies that $m+1 \in T_{t,n}$. Since $n \geq 2$, at least one of $m$ and $m+1$ is in the set $T_{t,n} \setminus (n\mathbb{N}+t)$, meaning $m \in A \cup (A-1)$. In any case, we have shown that $m \in A \cup (A-1)$, which, since $m \in \mathbb{N}$ was arbitrary, implies that $A \cup (A-1) = \mathbb{N}$. By construction, however, for all $t, n \in \mathbb{N}$ with $n \geq 2$, the set $(A-t)/n$ has empty intersection with the thick set $(T_{t,n} - t)/n$, meaning $(A-t)/n$ is not syndetic.
\end{example}
\begin{lemma}\label{lem:propertyofnilsystem}
Let $(X,T)$ be a minimal nilsystem, $x \in X$, and $U \subseteq X$ open, non-empty. The set $R(x,U)$ is non-empty, and for all $t \in R(x,U)$, the set $R(x,U) - t$ is $\ip_0^*$.
\end{lemma}
\begin{proof}
Because $(X,T)$ is minimal, the set $R(x,U)$ is non-empty. Let $t \in R(x,U)$. Because $T^t x \in U$, it follows from \cite[Theorem 0.2]{BLiprstarcharacterization} that the set $R(x,U) - t = R(T^t x, U)$ is $\text{IP}_0^*$.
\end{proof}
\begin{example}\label{ex:propertyofnilsystem}
There exists an $\text{IP}_2^*$ set $A \subseteq \mathbb{N}$ such that for all $t \in \mathbb{Z} \setminus \{0\}$, the set $A-t$ is not $\text{IP}_0^*$. By \cref{lem:propertyofnilsystem}, it follows that $A$ does not contain a set of the form $R(x,U)$ where $x$ and $U$ are a point and a non-empty, open subset of a minimal nilsystem.
To construct such a set, let $(m_n)_{n \in \mathbb{N}} \subseteq \mathbb{Z} \setminus \{0\}$ be a sequence with the property that for all $t \in \mathbb{Z} \setminus \{0\}$, there are infinitely many $n \in \mathbb{N}$ such that $m_n=t$. Choose a sequence $(r_n)_{n \in \mathbb{N}} \subseteq \mathbb{N}$ that is increasing sufficiently rapidly so that the set $B \vcentcolon= \bigcup_{n \in \mathbb{N}} (r_n \{1, \ldots, n\} + m_n) \subseteq \mathbb{N}$ is not $\text{IP}_2$, that is, does not contain a configuration of the form $\{x, y, x+y\}$. For $t \in \mathbb{Z}\setminus \{0\}$, define
\[B_t \vcentcolon= \bigcup_{\substack{n \in \mathbb{N} \\ m_n = t}} \big(r_n \{1, \ldots, n\} + t\big) \subseteq B,\]
and note that $B_t - t$ is an $\text{IP}_0$ set. Set $A = \mathbb{N} \setminus B$. Since $B$ is not $\text{IP}_2$, the set $A$ is $\text{IP}_2^*$, and for $t \in \mathbb{Z} \setminus \{0\}$, $(A-t) \cap (B_t-t) = \emptyset$, implying that $A-t$ is not $\text{IP}_0^*$.
\end{example}
\section{Concluding remarks and questions}\label{sec:remarksandquestions}
We collect here a number of further questions and open problems, beginning with ones of a dynamical nature.
There are two primary avenues for improvement in the main dynamical theorems, Theorems~\ref{thm:mainthm}, \ref{thm:mainthmwithtotallymin}, and \ref{thm:mainthmwithdistal}: upgrading the conclusions by saying more about the multiplicative combinatorial structure of return time sets $R(x,U)$, and enlarging the set of points $X'$ about which we can address the sets $R(x,U)$. In the first direction, it is natural to speculate how much the conclusion of \cref{thm:mainthm} can be upgraded.
\begin{question}\label{quest:posdenreturns}
Let $(X,T)$ be a minimal dynamical system. Does there exist a residual set of points $X' \subseteq X$ such that for all $x \in X'$ and all non-empty, open $U \subseteq X$, the set $R(x,U)$ has positive multiplicative density in a coset of a multiplicative subsemigroup of $\mathbb{N}$?
\end{question}
The conclusion in this question could be further upgraded to, \emph{``the set $R(x,U)$ has multiplicative density $1$ in a coset of a multiplicative subsemigroup of $\mathbb{N}$?''} If true, such a result would lend further evidence toward the stronger conjectures about the multiplicative combinatorial structure of additively syndetic sets outlined below.
In the second direction, it is natural to ask about the nature of return time sets $R(x,U)$ for points $x$ outside of $X'$, the residual subset of $X$ that appears in each of the main dynamical theorems. A positive answer to the following question would improve \cref{thm:mainthm}.
\begin{question}\label{quest:gpsforallpoints}
Let $(X,T)$ be a minimal dynamical system. Is it true that for all $x \in X$ and all non-empty, open $U \subseteq X$, the set $R(x,U)$ contains arbitrarily long geometric progressions?
\end{question}
There is a positive answer to Questions~\ref{quest:posdenreturns} and \ref{quest:gpsforallpoints} in the case that $(X,T)$ is an irrational rotation of the $1$-torus. It can be shown in that case that for all $x \in \mathbb{T}$ and all non-empty, open $U \subseteq \mathbb{T}$, there exist $n, N \in \mathbb{N}$ such that $d_{n S_N}^*\big( R(x,U) \big) = 1$.
While we are not able to answer these questions in more generality, we do know that systems in which the return time sets $R(x,U)$ are multiplicatively large for \emph{all} points $x \in X$ enjoy some rather strong dynamical properties. The following lemma outlines some of the (equivalent) dynamical consequences of assuming that every return times set $R(x,U)$ is \emph{multiplicatively thick} in $\mathbb{N}$: for all finite $F \subseteq \mathbb{N}$, there exists $m \in \mathbb{N}$ so that $mF \subseteq R(x,U)$.
\begin{lemma}\label{lem:equivformsofallmultthick}
Let $(X,T)$ be a dynamical system, and for $n \in \mathbb{N}$, let $\Delta_n = \Delta_n(X) \subseteq X^n$ be the diagonal $\{(x,\ldots,x) \in X^n \ | \ x \in X \}$. The following are equivalent:
\begin{enumerate}
\item \label{item:equivone} for all $x \in X$ and all non-empty, open $U \subseteq X$, the set $R(x,U)$ is multiplicatively thick in $\mathbb{N}$;
\item \label{item:equivtwo} for all $x \in X$ and all $n \in \mathbb{N}$, the $T \times T^2 \times \cdots \times T^n$-orbit closure of $(x,\ldots,x)$ contains the diagonal $\Delta_n$;
\item \label{item:equivthree} for all non-empty, open $U \subseteq X$ and all $n \in \mathbb{N}$,
\[\bigcup_{m=1}^\infty (T \times T^2 \times \cdots \times T^n)^{-m} \big( U \times \cdots \times U\big) \supseteq \Delta_n;\]
\item \label{item:equivfour} for all non-empty, open $U \subseteq X$ and all $n \in \mathbb{N}$,
\[\bigcup_{m=1}^\infty \bigcap_{i=1}^n T^{-mi} U = X;\]
\item \label{item:equivfive} \label{item:syndeticandthickcondition}for all $x \in X$ and all non-empty, open $U \subseteq X$, the set $R(x,U)$ satisfies: for all $n \in \mathbb{N}$, there exists a finite $F \subseteq \mathbb{N}$ such that for all $\ell \in \mathbb{N}_0$, there exists $m \in F$ such that $\ell + m\{1, \ldots, n\} \subseteq R(x,U)$.
\end{enumerate}
\end{lemma}
\begin{proof}
We will show that each condition implies the one following it; that condition (\ref{item:equivfive}) implies condition (\ref{item:equivone}) is immediate by taking $\ell = 0$.
(\ref{item:equivone}) implies (\ref{item:equivtwo}): Let $x \in X$ and $n \in \mathbb{N}$. Any non-empty, open subset $V$ of $\Delta_n$ contains a set of the form $( U \times \cdots \times U ) \cap \Delta_n$, where $U \subseteq X$ is non-empty, open. By (\ref{item:equivone}), there exists $m \in \mathbb{N}$ such that $m\{1,\ldots,n\} \subseteq R(x,U)$. This means that $(T \times T^2 \times \cdots \times T^n)^m(x,\ldots,x) \in U \times \cdots \times U$. Since $V$ was arbitrary, this shows that the $T \times T^2 \times \cdots \times T^n$-orbit closure of $(x,\ldots,x)$ contains the diagonal $\Delta_n$.
(\ref{item:equivtwo}) implies (\ref{item:equivthree}): Let $U \subseteq X$ be non-empty, open and $n \in \mathbb{N}$. Let $(x,\ldots,x) \in \Delta_n$. By (\ref{item:equivtwo}), there exists $m \in \mathbb{N}$ such that $(T \times T^2 \times \cdots \times T^n)^m(x,\ldots,x) \in U \times \cdots \times U$. This implies that $(x,\ldots,x) \in (T \times T^2 \times \cdots \times T^n)^{-m} \big(U \times \cdots \times U \big)$.
(\ref{item:equivthree}) implies (\ref{item:equivfour}): Let $U \subseteq X$ be non-empty, open and $n \in \mathbb{N}$. Let $x \in X$. By (\ref{item:equivthree}), there exists $m \in \mathbb{N}$ such that $(x,\ldots,x) \in (T \times T^2 \times \cdots \times T^n)^{-m} \big(U \times \cdots \times U \big)$, meaning that $x \in \bigcap_{i=1}^n T^{-mi} U$.
(\ref{item:equivfour}) implies (\ref{item:equivfive}): Let $x \in X$ and $U \subseteq X$ be non-empty, open. Let $n \in \mathbb{N}$. By (\ref{item:equivfour}) and the compactness of $X$, there exists a finite $F \subseteq \mathbb{N}$ such that
\[\bigcup_{m \in F} \bigcap_{i=1}^n T^{-mi} U = X.\]
Let $\ell \in \mathbb{N}_0$. There exists $m \in F$ such that $T^\ell x \in \bigcap_{i=1}^n T^{-mi} U$, meaning that $\ell + m\{1,\ldots,n\} \subseteq R(x,U)$.
\end{proof}
The $n=1$ case of condition (\ref{item:equivfour}) in \cref{lem:equivformsofallmultthick} is equivalent to the minimality of $(X,T)$. Condition (\ref{item:syndeticandthickcondition}) is easily seen to imply that $R(x,U)$ is both additively syndetic (the gap size is bounded by $\max F$) and multiplicatively thick. This is to be expected: as soon as the set $R(x,U)$ is non-empty for all $x \in X$ and all non-empty, open $U \subseteq X$, the system $(X,T)$ must be minimal and hence the sets $R(x,U)$ must be additively syndetic.
We proceed now with some open questions of a combinatorial nature related to the main motivating question, \cref{question:mainquestion}. The most basic open combinatorial question is whether or not syndetic sets contain a square ratio.
\begin{question}\label{quest:squareratios}
Do all additively syndetic subsets of $\mathbb{N}$ contain a configuration of the form $\{x, xy^2\}$ for $x, y \in \mathbb{N}$?
\end{question}
Going beyond square ratios and geometric progressions, the results in \cref{thm:mainthmwithtotallymin,thm:mainthmwithdistal} suggest that syndetic subsets of $\mathbb{N}$ may have positive multiplicative density in a coset of some multiplicative subsemigroup. In fact, the improvement of \cref{thm:mainthmwithdistal} in \cref{thm:mainthmwithdistalbetter} suggests the possibility that finitely many subsemigroups suffice to capture the multiplicative density of a syndetic set and all of its translates.
\begin{question}\label{quest:posdensityofsyndincoset} Let $A \subseteq \mathbb{N}$ be additively syndetic.
\begin{enumerate}
\item Do there exist $n, N \in \mathbb{N}$ such that $d^*_{nS_N}(A) > 0$?
\item Do there exist $i, N \in \mathbb{N}$ such that for all $t \in \mathbb{Z}$, $d^*_{(t,N) S_{N / (N,t)}}(A + i + t) > 0$?
\end{enumerate}
\end{question}
We have not even been able to rule out the possibility that syndetic sets have full multiplicative density in a coset of some non-trivial multiplicative subsemigroup. A positive answer to the following question would yield a positive answer not only to \cref{question:mainquestion}, but to Questions~\ref{quest:posdenreturns}, \ref{quest:gpsforallpoints}, \ref{quest:squareratios}, and \ref{quest:posdensityofsyndincoset} (1). Here $S_{N,1}$ denotes the multiplicative subsemigroup of positive integers congruent to 1 modulo $N$.
\begin{question}\label{quest:thickinsubsemi}
Let $A \subseteq \mathbb{N}$ be additively syndetic. Do there exist $n, N \in \mathbb{N}$ such that $d^*_{nS_{N,1}}(A) = 1$?\footnote{Since publication, this question has been answered in the negative: there exists a set $A \subseteq \mathbb{N}$ for which $A \cup (A-1) = \mathbb{N}$ but for which no such $n$ and $N$ exist.}
\end{question}
Being unable to answer \cref{quest:thickinsubsemi} for arbitrary syndetic sets, it makes sense to narrow the scope by asking the same question for combinatorially defined subclasses of syndetic sets.
\begin{question}\label{quest:thickinsubsemiforiprstar}
Let $A \subseteq \mathbb{N}$ be additively $\text{IP}_0^*$. Is it true that for all $t \in \mathbb{Z}$, there exist $n, N \in \mathbb{N}$ such that $d^*_{nS_{N,1}}(A+t) = 1$?
\end{question}
It is a consequence of \cite[Corollary 7.3]{BGpaperonearxiv} that the answer to \cref{quest:thickinsubsemiforiprstar} is ``yes'' when $t=0$ with $n=N=1$. Still, it is entirely possible that some or all of the questions posed here have a negative answer in general.
\bibliographystyle{alphanum}
|
\section{Introduction}
The maxima of Gaussian processes play an essential role in many aspects of probability and statistics, and the literature describing them is highly developed~\cite{Leadbetter:1983,Adler:1990,Lifshits:1995,Ledoux:Talagrand:2013, Talagrand:2014}. Within this area, a variety of questions are related to showing that the maximum of a process is unlikely to deviate far above, or below, its mean.
However, in comparison to the set of tools for handling the upper tail of a maximum, there are relatively few approaches for the lower tail. (Additional commentary related to this distinction may be found in~\cite[p.viii]{Talagrand:2014}~\cite[Sec.4.2]{Li:Shao:2001}~\cite[Sec.18]{Lifshits:1995}.)
In this paper, we consider the setting of a finite index set, where $X=(X_1,\dots,X_n)$ is a Gaussian vector with $\mathbb{E}[X_i]=0$ and $\operatorname{var}(X_i)=1$ for all $1\leq i\leq n$. The correlation matrix of $X$ is denoted by $R\in\mathbb{R}^{n\times n}$, and throughout the paper, we make the basic assumption that there is some fixed constant $\rho_0\in (0,1)$ such that
\begin{equation}\label{eqn:assumcor}
\max_{i\neq j}R_{ij}\leq \rho_0.
\end{equation}
In particular, the matrix $R$ is not required to be invertible.
Letting
\begin{equation}\label{eqn:maxdef}
M_n(X) = \max_{1\leq i\leq n} X_i,
\end{equation}
our aim is to derive a~non-asymptotic upper bound on
$\P(M_n(X) \leq t)$, where $t$ is a suitable point in the lower tail.
\subsection{Background.}\label{sec:background}
To provide background on lower-tail bounds for $M_n(X)$, we now briefly review some leading results.
Under the preceding conditions, the well-known concentration inequality for Lipschitz functions of Gaussian vectors implies that for any $s>0$,
\begin{equation}\label{eqn:BTS}
\P\Big( M_n(X) \ \leq \ \text{med}[M_n(X)]-s\Big) \ \leq \ e^{-s^2/2},
\end{equation}
where $\text{med}[\cdot]$ is any median~\cite{Sudakov:1974,Borell:1975,Ledoux:Talagrand:2013}.
Although this bound is broadly applicable, it can fail to describe lower-tail probabilities smaller than $\mathcal{O}(n^{-1})$. To see this, consider using~\eqref{eqn:BTS} to bound the probability $\P(M_n(X)\leq \delta_0\text{med}[M_n(X)])$ for some $\delta_0\in(0,1)$. If the entries of $X$ are independent, then $\text{med}[M_n(X)]=\sqrt{2\log(n)}(1+o(1))$, and
it follows that~\eqref{eqn:BTS} cannot give a bound better than $\mathcal{O}(n^{-1})$ in this case.
Furthermore, such a bound is far too large. In fact, when the entries of $X$ are independent, the actual size of this probability is smaller than $\mathcal{O}(n^{-\kappa_0})$ for \emph{any} $\kappa_0> 1$, no matter which $\delta_0$ is chosen. This limitation is also of more general interest, because it is a prototypical instance of the ``superconcentration phenomenon'', which has been a topic of much ongoing research activity~\cite{Chatterjee:2014}.
Very recently, a breakthrough result of Paouris and Valettas~\cite[eqn.~1.5]{Paouris:2018} improved upon~\eqref{eqn:BTS} by showing that the inequality
\begin{equation}\label{eqn:paouris}
\P\Big(M_n(X)\leq \text{med}[M_n(X)]-s\Big) \ \leq \ \textstyle\frac{1}{2}\exp\Big\{-\textstyle\frac{\pi}{1024}\frac{s^2}{\operatorname{var}[M_n(X)]}\Big\},
\end{equation}
holds for any $s>0$. The improvement over~\eqref{eqn:BTS} is most readily apparent when the entries of $X$ are independent, because the bound \mbox{$\operatorname{var}(M_n(X))=\mathcal{O}(1/\log(n))$} is known to hold in that case~\cite[Prop. 9.5]{Chatterjee:2014}.
Notably, the result underlying~\eqref{eqn:paouris} is also quite general, since it allows $M_n(X)$ to be replaced with any convex function of $X$.
Another interesting feature of~\eqref{eqn:paouris} is that it specifically describes the lower tail, and does not have an analogue for the upper tail (whereas the bound~\eqref{eqn:BTS} does).
Despite the progress achieved by the bound~\eqref{eqn:paouris}, it is natural to suspect that the constant $\frac{\pi}{1024}$ can be sharpened, and our main result (Theorem~\ref{thm:main}) leads to such an improvement in the following way. First note that if $\delta_0\in(0,1)$ is fixed, then~\eqref{eqn:paouris} becomes
\begin{equation}\label{eqn:mypaouris}
\P\Big(M_n(X) \ \leq \ \delta_0\,\text{med}[M_n(X)]\Big) \ \leq \ \textstyle\frac{1}{2}\exp\Big\{-\textstyle\frac{\pi (1-\delta_0)^2}{1024}\frac{\text{med}[M_n(X)]^2}{\operatorname{var}[M_n(X)]}\Big\}.
\end{equation}
In the particular case that $R_{ij}=\rho_0$ for all $i\neq j$, it is possible to show that
\begin{equation}
\lim_{n\to\infty} \ \textstyle\frac{\text{med}[M_n(X)]^2}{\operatorname{var}[M_n(X)]}\frac{1}{2\log(n)} \ = \ \frac{1-\rho_0}{\rho_0},
\end{equation}
which is handled later in Proposition~\ref{prop:extra}. Hence, in a worst-case sense with respect to the dependence condition~\eqref{eqn:assumcor},
the exponent in~\eqref{eqn:mypaouris} cannot be asymptotically better than
\begin{equation}\label{eqn:newexp}
-\textstyle\frac{2\pi}{1024}\frac{1-\rho_0}{\rho_0}(1-\delta_0)^2\, \log(n).
\end{equation}
As an improvement, our main result provides an inequality similar to~\eqref{eqn:mypaouris}, which holds uniformly with respect to~\eqref{eqn:assumcor}, and has an exponent equal to
\begin{equation}
\ \ \ \ \ \ \ \ \ \, \ \ \ \ -\textstyle\frac{1-\rho_0}{\rho_0}(1-\delta_0)^2\, \log(n)+\mathcal{O}(\log\log(n)).
\end{equation}
Thus, up to a $\log\log(n)$ term, the $\frac{2\pi}{1024}$ in~\eqref{eqn:newexp} can be replaced with 1. (From a purely heuristic standpoint, this substitution makes sense, because it corresponds to replacing $\frac{\pi}{1024}$ with $\frac{1}{2}$ in~\eqref{eqn:paouris}, which gives the usual form of a Gaussian tail bound.)
One more result to mention is a lower-tail bound due to Hartigan~\cite[Theorem 3.4]{Hartigan:2014}. This bound is of a different form than those mentioned previously, and is based upon the ``residual variables'' defined by $V_i=X_i-\mathbb{E}[X_i|X_1,\dots,X_{i-1}]$. If we put $\kappa(n,\varepsilon)=2\log(n/\sqrt{2\pi})-2\log\log(1/\varepsilon)$ for some $\varepsilon\in (0,1)$, and define $\sigma_n^2=\min_{1\leq i\leq n} \operatorname{var}(V_i)$, then as long as $\kappa(n,\varepsilon)\geq 6$, the bound may be written as
\begin{equation}\label{eqn:hartigan}
\P\bigg(M_n(X) \ \leq \ \sigma_n \sqrt{ \kappa(n,\varepsilon) -\log(\kappa(n,\varepsilon))}-\sqrt{1-\sigma_n^2}|\Phi^{-1}(\varepsilon)|\bigg) \ \leq \ 2\varepsilon,
\end{equation}
where $\Phi$ is the standard normal distribution function.\footnote{When referencing the main result in the paper~\cite{Hartigan:2014}, note that a square root is omitted, and that the quantity $\sigma\tau\Phi^{-1}(\alpha)$ should be replaced with $\tau\Phi^{-1}(\alpha)$. In addition, the $\lambda_n\sqrt{2\log(n)}$ in the abstract should be replaced with $\sqrt{2\lambda_n\log(n)}.$} For a suitable choice of $\varepsilon$, this bound has some qualitative similarities with the Paouris-Valettas bound~\eqref{eqn:paouris} and our main result in Theorem~\ref{thm:main}. However, due to the intricate dependence on the parameters $(\varepsilon, n,\sigma_n,\rho_0)$, it is not obvious if the bound~\eqref{eqn:hartigan} is sharp under~\eqref{eqn:assumcor}. By contrast, the simple form of the bound given in Theorem~\ref{thm:main} makes it possible to see immediately that the bound is sharp. Also, it is notable that all three arguments underlying~\eqref{eqn:paouris},~\eqref{eqn:hartigan}, and Theorem~\ref{thm:main} are entirely different.
\subsection{Remarks on proof and outline}\label{sec:proofremarks} At a high level, our proof is based on the classical idea of comparing $X$ with another Gaussian vector whose entries are equicorrelated. This idea seems to have its roots in the early work of Berman~\cite{Berman:1962,Berman:1964} on the maxima of stationary Gaussian processes, and hinges on Slepian's lemma~\cite{Slepian:1962}: If $X'\in\mathbb{R}^n$ is a centered Gaussian vector with standardized entries, and if its correlation matrix $R'$ satisfies
$$R_{ij} \ \leq \ R_{ij}'$$
for all $(i,j)$, then the inequality
\begin{equation}\label{eqn:slepian}
\P\big(M_n(X) \leq t\big) \ \leq \ \P\big(M_n(X') \leq t\big)
\end{equation}
holds for all $t\in\mathbb{R}$. For the purpose of deriving an upper bound on $\P(M_n(X)\leq t)$, the equicorrelated case where $R_{ij}'=\rho_0$ for all $i\neq j$ is of special importance, because it is ``maximal'' with respect to~\eqref{eqn:slepian}. More precisely, for any fixed $\rho_0$, the inequality~\eqref{eqn:slepian} holds over the entire set of matrices $R$ with $\max_{i\neq j}R_{ij}\leq \rho_0$.
Although this general strategy has been used in a variety of works to analyze $M_n(X)$, the associated results have been predominantly asymptotic, and focused on deviations of order $o(1)$. (We refer to the papers~\cite{Berman:1964,Pickands:1967,Pickands:1969,Mittal:1975}, as well as Chapter 6 of the book~\cite{Leadbetter:1983} for an overview.) Alternatively, we study deviations that are proportional to $\sqrt{\log(n)}$, which extend much deeper into the lower tail.
The remainder of the paper is organized as follows. After presenting our main result in Section~\ref{sec:main}, we discuss some consequences in Section~\ref{sec:conseq}, showing how the result can be improved when $R$ has additional structure, and how it can be applied to small-ball probabilities. The proof of the main result is given in Section 4. With regard to notation, the symbols $c$ and $C$ always denote constants that do not depend on $n$, but their dependence on other parameters may change from line to line. The abbreviation $\ell_2(n):=\log\log(\max\{n,3\})$ is also used.
\section{Main result}\label{sec:main}
To clarify the statement of our main result, we first state a basic proposition describing the sizes of $\mathbb{E}[M_n(X)]$ and $\text{med}[M_n(X)]$ under~\eqref{eqn:assumcor}. More specifically, the proposition shows that the value $\sqrt{2(1-\rho_0)\log(n)}$ is a natural reference level for a lower-tail bound. Although this fact could be considered well known to specialists, it is not easily referenced in the form given below, and so we provide a short proof at the end of Section~\ref{sec:proofs}.
\begin{proposition}\label{prop:expec}
Let $\mu_n$ stand for either $\mathbb{E}[M_n(X)]$ or $\textup{med}[M_n(X)]$. Fix a constant $\rho_0\in(0,1)$, and suppose the condition~\eqref{eqn:assumcor} holds. Then, there is a universal constant $c_0>0$, such that
\begin{equation}\label{eqn:expeclower}
\begin{split}
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \mu_n & \ \geq \ \sqrt{2(1-\rho_0)\log(n)} - c_0\sqrt{\ell_2(n)}.
\end{split}
\end{equation}
Furthermore, if $R_{ij}=\rho_0$ for all $i\neq j$, then
\begin{equation}\label{eqn:expecupper}
\mu_n \ \leq \ \sqrt{2(1-\rho_0)\log(n)}.
\end{equation}
\end{proposition}
\noindent The following theorem is the main result of the paper.
\begin{theorem}\label{thm:main}
Fix two constants $\delta_0,\rho_0\in(0,1)$, and suppose the condition~\eqref{eqn:assumcor} holds. In addition, define the constants
\begin{equation}\label{eqn:alphadef}
\alpha_0=\textstyle\frac{(1-\rho_0)(1-\delta_0)^2}{\rho_0} \ \ \ \ \text{ and } \ \ \ \ \beta_0=\textstyle\frac{(1-\rho_0)(1-\delta_0)}{\rho_0}.
\end{equation}
Then, there is a constant $C>0$ depending only on $(\delta_0,\rho_0)$, such that
\begin{equation}\label{eqn:mainbound}
\ \P\bigg(M_n(X)\, \leq \, \delta_0\sqrt{2(1-\rho_0)\log(n)}\bigg) \ \leq \ C\, n^{-\alpha_0}\log(n)^{\frac{\beta_0-1}{2}}.
\end{equation}
Furthermore, the bound~\eqref{eqn:mainbound} is sharp in the sense that if $R_{ij}=\rho_0$ for all $i\neq j$, then there is a constant $c>0$ depending only on $(\delta_0,\rho_0)$, such that
\begin{equation}\label{eqn:mainlower}
\P\bigg(M_n(X) \, \leq \, \delta_0\sqrt{2(1-\rho_0)\log(n)}\bigg) \ \geq \ c\, n^{-\alpha_0}\log(n)^{\frac{\beta_0-1}{2}}.
\end{equation}
\end{theorem}
To comment on some basic features of the theorem, first note that the dominant exponent satisfies $\alpha_0\to\infty$ as $\rho_0\to 0$. Hence, the bound respects the fact that the lower-tail probability decays faster than any power of $n^{-1}$ when the entries of $X$ are independent (as discussed in Section~\ref{sec:background}). Second, the theorem conforms with the reference level motivated by Proposition~\ref{prop:expec}, since~\eqref{eqn:mainbound} implies that $\text{med}[M_n(X)]$ cannot be much less than $\sqrt{2(1-\rho_0)\log(n)}$.
\section{Some consequences}\label{sec:conseq}
This brief section presents two direct corollaries of Theorem~\ref{thm:main}.
\subsection{Taking advantage of correlation structure}
If the matrix $R$ satisfies $\max_{i\neq j}R_{ij}= \rho_0$, but most off-diagonal entries are substantially less than $\rho_0$, then the upper bound in Theorem~\ref{thm:main} can be conservative. Fortunately, the following simple corollary shows this is not necessarily a problem. Roughly speaking, if there is a number $\tilde\rho_0\in (0,\rho_0)$ such that most off-diagonal entries satisfy $R_{ij}\leq \tilde \rho_0$, then Theorem~\ref{thm:main} can be essentially re-used with $\rho_0$ being replaced by $\tilde \rho_0$. Furthermore, this can be done even when $\rho_0=1$.
\begin{corollary}\label{cor:simple}
Fix two constants $\delta_0,\tilde{\rho}_0\in(0,1)$. Suppose there is an index set $J\subset\{1,\dots,n\}$ with cardinality $\tilde n\geq 2$, such that $R_{ij}\leq \tilde \rho_0$ for all distinct $i,j\in J$.
Lastly, define the constants
$$ \tilde\alpha_0 \ = \ \textstyle\frac{(1-\tilde\rho_0)(1-\delta_0)^2}{\tilde \rho_0} \text{ \ \ \ \ and \ \ \ \ } \tilde \beta_0 \ = \ \textstyle\frac{(1-\tilde\rho_0)(1-\delta_0)}{\tilde\rho_0}.$$
Then, there is a constant $ C>0$ depending only on $(\delta_0,\tilde\rho_0)$ such that
\begin{equation}\label{eqn:mainrefined}
\P\bigg(M_n(X)\, \leq \, \delta_0\sqrt{2(1-\tilde\rho_0)\log(\tilde n)}\bigg) \ \leq \ C\, \tilde n^{-\tilde{\alpha}_0}\,\log(\tilde n)^{\frac{\tilde{\beta}_0-1}{2}}.
\end{equation}
\end{corollary}
If detailed knowledge of $R$ is available, the corollary can be further refined by carefully choosing $\tilde \rho_0$ and $J$. Indeed, there is a tradeoff involved, because a smaller value of $\tilde\rho_0$ improves the exponent $\tilde \alpha_0$, but reduces $\tilde n$.
\subsection{Upper bounds for small-ball probabilities} Small-ball probabilities are known to be an important ingredient in the study of Gaussian processes~\cite{Li:Shao:2001}, and they have the form $\P(\|X\|_{\infty}\leq t)$, where $\|X\|_{\infty}=\max_{1\leq i\leq n}|X_i|$, and $t$ is suitably small. Such probabilities are also a special case of the previous lower-tail probabilities, since they can be written as $\P(M_{2n}(Y)\leq t)$ with $Y=(-X,X)\in\mathbb{R}^{2n}$.
One of the well-known challenges involved in analyzing small-ball probabilities is that relatively few general results are available. In this regard, two noteworthy bounds are due to Lata\l a-Oleszkiewicz~\cite[Theorem 4]{Latala:2005}, and Paouris-Valettas~\cite[Theorem 3.3]{Paouris:2018}. When $X\in\mathbb{R}^n$ is a centered Gaussian vector with standardized entries, these bounds respectively state that if $\delta_0\in (0,1/2)$, then
\begin{equation}\label{eqn:smallLatala}
\ \ \ \P\Big(\|X\|_{\infty} \leq \delta_0 m_n\Big)
\ \leq \ \textstyle\frac{1}{2}\exp\Big\{-\frac{1}{4}m_n^2\log(\frac{1}{2\delta_0})\Big\},
\end{equation}
and
\begin{equation}\label{eqn:smallPaouris}
\P\Big(\|X\|_{\infty} \leq \delta_0 m_n\Big)
\ \leq \ \textstyle\frac{1}{2}\exp\Big\{- c\,\textstyle\frac{m_n^2}{\varsigma_n^2}\log(\frac{1}{\delta_0})\Big\},
\end{equation}
where $m_n=\text{med}[\|X\|_{\infty}]$, $\varsigma_n^2=\operatorname{var}[\|X\|_{\infty}]$, and $c>0$ is a universal constant. Due to the fact that the constant $c>0$ is not easily specified, the first bound~\eqref{eqn:smallLatala} has an advantage insofar as its exponent is explicit. On the other hand, the second bound~\eqref{eqn:smallPaouris} may be preferred if $\varsigma_n^2\to 0$ as $n\to\infty$ (as in the case of independent variables, cf.~\cite[Prop. 9.5]{Chatterjee:2014}).
In cases where $\delta_0$ is not too small, the following corollary of Theorem~\ref{thm:main} can enhance some aspects of the bounds~\eqref{eqn:smallLatala} and~\eqref{eqn:smallPaouris}.
\begin{corollary}\label{cor:small}
Fix two numbers $\delta_0,\rho_0\in(0,1)$, and let $(\alpha_0,\beta_0)$ be as in~\eqref{eqn:alphadef}. Also, let $C>0$ be as in~\eqref{eqn:mainbound}, and suppose $\max_{i\neq j}|R_{ij}|\leq \rho_0$. Then,
\begin{equation}\label{eqn:firstsmall}
\ \P\Big(\|X\|_{\infty}\leq \delta_0\sqrt{2(1-\rho_0)\log(2n)}\Big) \ \leq \ C\,(2n)^{-\alpha_0}\log(2n)^{\frac{\beta_0-1}{2}}.
\end{equation}
In addition, if $R_{ij}\in[0,\rho_0]$ for all $i\neq j$, then
\begin{equation}\label{eqn:secondsmall}
\P\Big(\|X\|_{\infty}\leq \delta_0\sqrt{2(1-\rho_0)\log(n)}\Big) \ \leq \ C^2\, n^{-2\alpha_0}\log(n)^{\beta_0-1}.
\end{equation}
\end{corollary}
Although it is difficult to make an exact comparison with the bounds~\eqref{eqn:smallLatala} and~\eqref{eqn:smallPaouris}, it is possible to point out some basic differences.
To do this, recall that $\alpha_0$ can be made arbitrarily large when $\rho_0$ is small enough, and that $m_n\leq \sqrt{2\log(2n)}$~(cf.~\cite{Kwapien:1994}).
Based on these observations, the right side of~\eqref{eqn:firstsmall} will be of smaller order than the right side of~\eqref{eqn:smallLatala} when $\rho_0$ is sufficiently small (for a given $\delta_0$). With regard to the bound~\eqref{eqn:smallPaouris}, it is known from~\cite[Prop. 9.5]{Chatterjee:2014} that there is a universal constant $c_1>0$ for which
$$\frac{1}{\varsigma_n^2}\ \geq \ \frac{c_1}{\rho_0+\textstyle\frac{1}{\log(n)}}.$$
Hence, when $\rho_0$ is small, the bounds~\eqref{eqn:firstsmall} and~\eqref{eqn:smallPaouris} are roughly similar, except that the exponent in~\eqref{eqn:firstsmall} is explicit. On this point, it is also worth mentioning that even if the $c$ in~\eqref{eqn:smallPaouris} can be specified, it is still difficult to obtain good constants when bounding~$\varsigma_n^2$.\\
\noindent \emph{Proof of Corollary~\ref{cor:small}.} The first bound~\eqref{eqn:firstsmall} follows immediately from Theorem~\ref{thm:main} by considering the vector $Y=(X,-X)\in\mathbb{R}^{2n}$. To prove the second bound~\eqref{eqn:secondsmall},
let $R'\in \mathbb{R}^{n\times n}$ denote the correlation matrix whose off-diagonal entries are equal to $\rho_0$, and let $X'\sim N(0,R')$. In addition, let $Y'\in\mathbb{R}^{2n}$ be the random vector obtained by concatenating two independent copies of $X'$. Due to the assumption that $R_{ij}\in[0,\rho_0]$ for all $i\neq j$, the correlation matrices of $Y$ and $Y'$ are related by the following entrywise inequality,
\begin{equation}
\text{cor}(Y) \ = \ \Big(\begin{array}{cc}
R & \!\!-R\\
-R & R
\end{array}\Big) \ \leq \ \Big(\begin{array}{cc}
R' & \!\!0\\
0 & R'
\end{array}\Big) \ = \ \text{cor}(Y').
\end{equation}
Therefore, the vectors $Y$ and $Y'$ satisfy the conditions of Slepian's lemma (which allows for singular correlation matrices). In turn, if we define the number
\begin{equation}\label{eqn:tndef}
t_n= \delta_0\sqrt{2(1-\rho_0)\log(n)},
\end{equation}
then we obtain
\begin{equation}
\begin{split}
\P\big(\|X\|_{\infty}\leq t_n\big) & \ = \ \P\big(M_{2n}(Y) \leq t_n\big)\\[0.3cm]
%
& \ \leq \ \P\big(M_{2n}(Y')\leq t_n\big)\\[0.3cm]
%
& \ = \ \Big\{\P\big(M_n(X')\leq t_n\big)\Big\}^2\\[0.3cm]
%
& \ \leq \ C^2\, n^{-2\alpha_0}\log(n)^{\beta_0-1},
\end{split}
\end{equation}
as needed.\qed
\section{Proofs for Section~\ref{sec:main}}\label{sec:proofs}
\noindent\textbf{Remark.} Throughout the proofs, we always assume that $n$ is sufficiently large for any given expression to make sense --- since this can be accommodated by an adjustment of constants in the results.\\
\noindent \textbf{Proof of the upper bound~\eqref{eqn:mainbound}.} Let $X'\sim N(0,R')$ be an equicorrelated Gaussian vector, as in the proof of Corollary~\ref{cor:small}.
It follows that if $Z_0,Z_1,\dots,Z_n$ denote independent $N(0,1)$ random variables, then the coordinates of $X'$ may be jointly represented as
\begin{equation}\label{eqn:correp}
X_i' = \sqrt{\rho_0}Z_0+\sqrt{1-\rho_0}Z_i.
\end{equation}
This yields a valuable representation of the maximum,
\begin{equation}\label{eqn:maxrep}
M_n(X')= \sqrt{\rho_0}Z_0+\sqrt{1-\rho_0}\max_{1\leq i\leq n}Z_i.
\end{equation}
In other words, $M_n(X')$ is the convolution of the variables on the right side. Recalling the definition of $t_n$ in~\eqref{eqn:tndef}, a
simple calculation then gives an exact formula,
\begin{equation}\label{eqn:mainint}
\P\Big(M_n(X') \leq t_n\Big) = \int_{-\infty}^{\infty}\psi_n(s)ds,
\end{equation}
where the integrand is defined by
\begin{equation}
\psi_n(s) \ = \ \phi(s)\Phi^n \Big(\textstyle\frac{t_n-\sqrt{\rho_0}s}{\sqrt{1-\rho_0}}\Big),
\end{equation}
with $\phi$ being the standard normal density.\\
The remainder of the proof consists in bounding integral $\int_{-\infty}^{\infty}\psi_n(s)ds$ over the intervals
$(-\infty,-c_n]$, $[-c_n,0]$, and $[0,\infty)$, where we put
\begin{equation}\label{eqn:cnformula}
c_n=\sqrt{\textstyle\frac{2(1-\rho_0)\log(n)}{\rho_0}}\Big\{1-\delta_0-\textstyle\frac{\frac 14 \ell_2(n)}{\log(n)}\Big\}.
\end{equation}
The choice of this value as a cut-off point for decomposing the integral~\eqref{eqn:mainint} is an important element of the proof. In particular, the $1/4$ coefficient on the $\ell_2(n)/\log(n)$ term is essential.
Handling the integral over $[-c_n,0]$ is the most involved portion of the proof, and is postponed to Lemma~\ref{lem:hard}.
Once the proof of that lemma is complete, it will be straightforward to control the integral over $[0,\infty)$, which is done in Lemma~\ref{lem:easy}. For the moment, we only handle the interval $(-\infty,-c_n]$, since it requires no further preparation. Indeed, we have
\begin{equation}\label{eqn:firstmills}
\begin{split}
\int_{-\infty}^{-c_n} \psi_n(s) & \ \leq \ \Phi(-c_n)\\[0.2cm]
& \ \leq \ \frac{e^{-c_n^2/2}}{\sqrt{2\pi}c_n} \ \ \ \ \text{(Mill's inequality)}\\[0.4cm]
& \ \leq \ \textstyle\frac{C}{\sqrt{\log(n)}}\, n^{-\alpha_0} \, \exp\big\{\frac{\beta_0}{2} \ell_2(n)\big\}\\[0.3cm]
& \ = \ C n^{-\alpha_0} \log(n)^{\frac{\beta_0-1}{2}}
\end{split}
\end{equation}
where $C$ only depends on $\delta_0$ and $\rho_0$. This completes the proof of the upper bound~\eqref{eqn:mainbound}.\\
\noindent \textbf{Proof of the lower bound~\eqref{eqn:mainlower}.} It suffices to derive a lower bound on the integral $\int_{-\infty}^{-c_n}\psi_n(s)ds$. Using the fact that the function $\Phi^n \Big(\textstyle\frac{t_n-\sqrt{\rho_0}s}{\sqrt{1-\rho_0}}\Big)$ is decreasing in $s$, we have
\begin{equation}\label{eqn:firstlower}
\int_{-\infty}^{-c_n}\psi_n(s)ds \ \geq \ \Phi^n \Big(\textstyle\frac{t_n+\sqrt{\rho_0}c_n}{\sqrt{1-\rho_0}}\Big) \Phi(-c_n).
\end{equation}
If we define the number
\begin{equation}
\begin{split}
r_n& \ = \ \textstyle\frac{t_n+\sqrt{\rho_0}c_n}{\sqrt{1-\rho_0}}\\[0.2cm]
& \ = \ \sqrt{2\log(n)}\Big\{1-\textstyle\frac{\frac 14\ell_2(n)}{\log(n)}\Big\},
\end{split}
\end{equation}
then Mill's inequality gives
\begin{equation}
\begin{split}
\Phi(r_n) & \ \geq \ 1-\textstyle\frac{1}{\sqrt{2\pi}r_n} e^{-r_n^2/2}\\[0.2cm]
&\ \geq \ 1- \textstyle\frac{C}{\sqrt{\log(n)}}e^{-\log(n)}e^{\frac12 \ell_2(n)}\\[0.2cm]
& \ = \ 1 -\textstyle\frac{C}{n}.
\end{split}
\end{equation}
Hence, the limit $(1-C/n)^n\to \exp(-C)$ as $n\to\infty$ shows that $\Phi^n(r_n)$ is lower bounded by a positive constant for all large $n$. Finally, regarding the factor $\Phi(-c_n)$ in~\eqref{eqn:firstlower}, the lower-bound version Mill's inequality~\cite[Theorem 1.4]{Durrett:2005} gives
\begin{equation}
\begin{split}
\Phi(-c_n) & \ \geq \ \Big\{\textstyle\frac{1}{c_n}-\textstyle\frac{1}{c_n^3}\Big\}\frac{1}{\sqrt{2\pi}}e^{-\frac{1}{2}c_n^2},\\[0.3cm]
& \ \geq \ c n^{-\alpha_0}\log(n)^{\frac 12(\beta_0-1)},
\end{split}
\end{equation}
where the second step follows from a calculation very similar to~\eqref{eqn:firstmills}. This completes the proof of the lower bound~\eqref{eqn:mainlower} and the theorem. \qed
\begin{lemma}\label{lem:hard}
There is a constant $C>0$ depending only on $(\rho_0,\delta_0)$ such that
\begin{equation}
\int_{-c_n}^{0}\psi_n(s)ds \ \leq \ C\, n^{-\alpha_0} \log(n)^{\frac 12(\beta_0-1)} .
\end{equation}
\end{lemma}
\proof
To simplify the analysis of $\psi_n(s)$, we will parameterize $s$ as function of a number $\delta$ lying in an interval $\mathcal{I}_n$ defined by
$$\mathcal{I}_n=\Big[0\, , \, 1-\delta_0-\textstyle\frac{\frac 14 \ell_2(n)}{\log(n)}\Big].$$
Specifically, we write
$$s=s_n(\delta) \ = -a_n\delta,$$
where
$$a_n=\sqrt{\textstyle\frac{2(1-\rho_0)\log(n)}{\rho_0}}.$$
In particular, note that $s_n(\cdot)$ maps $\mathcal{I}_n$ to the interval $[-c_n,0]$. As a means of simplifying the factor $\Phi^n\big(\textstyle\ts\frac{t_n-\sqrt{\rho_0}s}{\sqrt{1-\rho_0}}\big)$ in the definition of $\psi_n(s)$, let
\begin{equation}\label{eqn:undef}
\begin{split}
u_n(\delta) &=\textstyle\frac{t_n-\sqrt{\rho_0}s_n(\delta)}{\sqrt{1-\rho_0}}\\[0.3cm]
&= \sqrt{2\log(n)}(\delta_0+\delta).
\end{split}
\end{equation}
It follows that for all $\delta\in\mathcal{I}_n$,
\begin{equation}\label{eqn:logpsiformula}
\log \psi_n(s_n(\delta)) = -\textstyle\frac{(1-\rho_0)\cdot\log(n)\cdot \delta^2}{\rho_0}+n\log \Phi(u_n(\delta))-\log(\sqrt{2\pi}).
\end{equation}
Due to the lower-bound form of Mill's inequality, we have
\begin{equation}\label{eqn:cubicmills}
\begin{split}
\Phi(u_n(\delta)) &\ \leq \ 1-\textstyle\frac{1}{\sqrt{2\pi}}\Big\{\textstyle\frac{1}{u_n(\delta)}-\textstyle\frac{1}{u_n^3(\delta)}\Big\}\displaystyle\exp({-\textstyle\frac{1}{2}u_n(\delta)^2}),\\[0.3cm]
& \ = \ 1-b_n(\delta)
\end{split}
\end{equation}
where $b_n(\delta)$ is defined by the last line. When $n$ is sufficiently large, it is simple to check that the condition $0<b_n(\delta)<1$ holds for all $\delta\in\mathcal{I}_n$, which gives
$$\log(1-b_n(\delta)) \ \leq \ -b_n(\delta).$$
Combining the last few steps, the following bound holds for all $\delta\in\mathcal{I}_n$,
\begin{equation}\label{eqn:logupperbound}
\begin{split}
n\log \Phi(u_n(\delta)) \ & \ \leq -n\, b_n(\delta)\\[0.2cm]
& \ \leq \ \textstyle\frac{-cn}{\sqrt{\log(n)}}\exp\big\{-\log(n)(\delta_0+\delta)^2\big\},
\end{split}
\end{equation}
where $c>0$ depends only on $\delta_0$ and $\rho_0$.\\
The work up to this point provides us with a useful majorant for $\psi_n$. By looking at the equation~\eqref{eqn:logpsiformula} and the bound~\eqref{eqn:logupperbound}, it is clear that if we define the function
\begin{equation}\label{eqn:upperfunc}
\begin{split}
f_n(\delta)&=\textstyle\frac{(1-\rho_0)\log(n)\delta^2}{\rho_0}+\textstyle\frac{cn}{\sqrt{\log(n)}}\displaystyle\exp\big\{{-\log(n)(\delta_0+\delta)^2}\big\},
\end{split}
\end{equation}
then the bound
\begin{equation}\label{eqn:upperfuncbound}
\psi_n(s_n(\delta)) \ \leq \ \exp\{-f_n(\delta)\}
\end{equation}
holds for all $\delta\in\mathcal{I}_n$. Integrating this bound gives
\begin{equation}\label{eqn:firstcov}
\int_{-c_n}^0 \psi_n(s)ds \ \leq \ a_n \!\int_{\mathcal{I}_n} \exp\{-f_n(\delta)\} d\delta.
\end{equation}
We now parameterize $\delta$ as a function of a positive number $\eta$ using
$$\delta=\delta_n(\eta) = 1-\delta_0-\textstyle\frac{\eta\ell_2(n)}{\log(n)}.$$
Specifically, if we define the interval
$$\mathcal{J}_n =\Big[\textstyle\frac{1}{4}\, , \, \frac{(1-\delta_0)\log(n)}{\ell_2(n)}\Big],$$
then $\delta_n(\cdot)$ maps $\mathcal{J}_n$ to $\mathcal{I}_n$, and the integral bound~\eqref{eqn:firstcov} becomes
\begin{equation}\label{eqn:secondcov}
\begin{split}
\int_{-c_n}^0 \psi_n(s)ds
& \ \leq \textstyle\frac{a_n \ell_2(n)}{\log(n)} \displaystyle \int_{\mathcal{J}_n} \exp\{-f_n(\delta_n(\eta))\}d\eta.
\end{split}
\end{equation}
The remainder of the proof will be divided into two parts, in which the integral over $\mathcal{J}_n$ is decomposed with the subintervals
$$\mathcal{J}_n' = \Big[\textstyle\frac{1}{4}\, , \, \textstyle\frac{\log(n)^{1/4}}{\ell_2(n)}\Big] \ \ \ \ \ \text{ and } \ \ \ \ \ \mathcal{J}_n'' = \Big[\textstyle\frac{\log(n)^{1/4}}{\ell_2(n)},\textstyle\frac{(1-\delta_0)\log(n)}{\ell_2(n)}\Big].$$
In handling these subintervals, it will be convenient to label the summands of $f_n$ in line~\eqref{eqn:upperfunc} according to
$$f_n(\delta) = g_n(\delta) + h_n(\delta).$$
\noindent \emph{The integral over $\mathcal{J}_n'$.}
By expanding out the square $\delta_n(\eta)^2$, and dropping the smallest positive term, the following bound holds for any $\eta$,
$$\exp\big\{\!-\!g_n(\delta_n(\eta))\big\} \ \leq \ n^{-\alpha_0}\exp\Big\{2\beta_0 \ell_2(n)\eta\Big\}.$$
In addition, if we expand the square $(\delta_0+\delta_n(\eta))^2$, and use the fact that every $\eta\in\mathcal{J}_n'$ is bounded above by $\log(n)^{1/4}/\ell_2(n)$, then a short calculation gives
\begin{equation*}
\begin{split}
-h_n(\delta_n(\eta))
& \ \leq \ -c e^{\ell_2(n)(2\eta-\frac{1}{2})},
\end{split}
\end{equation*}
for small enough $c>0$. Directly combining the last two steps gives
\begin{equation}\label{eqn:quadstep}
\exp\big\{\!-\!f_n(\delta_n(\eta))\big\} \ \leq \ n^{-\alpha_0}\exp\Big\{ 2\beta_0\eta \ell_2(n) - ce^{\ell_2(n)(2\eta-\frac{1}{2})} \Big\}.
\end{equation}
To simplify the previous bound, define
$$x(\eta)=2\eta-\textstyle\frac{1}{2}.$$
Since $x(\eta)$ is non-negative for all $\eta\in\mathcal{J}_n'$, we may approximate $\exp\{\ell_2(n)x(\eta)\}$ from below using a second-order Taylor expansion
$$e^{\ell_2(n) x(\eta)} \ \geq \ 1+\ell_2(n) x(\eta)+\textstyle\frac{1}{2}\ell_2(n)^2 x(\eta)^2.$$
After some arithmetic, the bound~\eqref{eqn:quadstep} becomes
\begin{equation}\label{eqn:expbound}
\begin{split}
\exp\big\{\!-\!f_n(\delta_n(\eta))\big\} & \ \leq \ e^{\frac{\beta_0}{2}\ell_2(n)}\cdot e^{-c} \cdot n^{-\alpha_0}\cdot \varphi_n(x(\eta)),
\end{split}
\end{equation}
where we define the function
$$\varphi_n(x)=\exp\Big\{ -\big[\textstyle\frac{c}{2}\ell_2(n)^2\big] x^2 + \big[(\beta_0-c)\ell_2(n)\big] x \Big\}.$$
Integrating the bound~\eqref{eqn:expbound} over $\mathcal{J}_n'$, we obtain
\begin{equation}\label{eqn:nearly}
\small
\begin{split}
\textstyle\frac{a_n \ell_2(n)}{\log(n)} \displaystyle \int_{\mathcal{J}_n'}\exp\{-f_n(\delta_n(\eta))\}d\eta
& \ \ \leq \ \ C\cdot \ell_2(n) \cdot \log(n)^{\frac{1}{2}(\beta_0-1)} \cdot n^{-\alpha_0} \displaystyle \int_{0}^{\infty}\textstyle \varphi_n(x)dx.
\end{split}
\end{equation}
To handle the last integral, note that the function $\varphi_n(x)$ can be written in the form $\varphi_n(x)=\exp(-a x^2+bx)$,
and that the following elementary bound holds for any $a>0$ and $b\in\mathbb{R}$,
\begin{equation*}
\int_0^{\infty} e^{-a x^2+b x}dx \ \leq \ \textstyle\frac{\sqrt{\pi}}{\sqrt a} \exp\Big(\textstyle\frac{b^2}{4a}\Big).
\end{equation*}
Therefore,
$$\displaystyle \int_{0}^{\infty}\varphi_n(x)dx \ \leq \ \textstyle\frac{C}{\ell_2(n)}. $$
Combining this with the bound~\eqref{eqn:nearly} completes the work on $\mathcal{J}_n'$.\\
\noindent \emph{The integral over $\mathcal{J}_n''$.} First note that $\exp\{-f_n(\delta)\} \leq \exp\{-h_n(\delta)\}$. Also, the function $\exp\{-h_n(\delta_n(\eta))\}$ is decreasing in $\eta$, and so if we denote the left endpoint of $\mathcal{J}_n''$ as $\eta_n:=\log(n)^{1/4}/\ell_2(n)$, then we have the following height-width integral bound
\begin{equation*}
\begin{split}
\textstyle\frac{a_n \ell_2(n)}{\log(n)} \displaystyle \int_{\mathcal{J}_n''}\exp\big\{\!-\!f_n(\delta_n(\eta))\big\}d\eta & \ \leq \ \textstyle\frac{a_n \ell_2(n)}{\log(n)} \cdot |\mathcal{J}_n''| \cdot \exp\{-h_n(\delta_n(\eta_n))\}\\[0.3cm]
& \ \leq \ C\cdot \sqrt{\log(n)} \cdot \exp\Big\{\textstyle\frac{-c}{\sqrt{\log(n)}}e^{2\log(n)^{1/4}} \Big\}.
\end{split}
\end{equation*}
It is simple to check that this bound is of smaller order than $n^{-\alpha_0}\log(n)^{\frac 12(\beta_0-1)}$. This completes the proof of Lemma~\ref{lem:hard}\qed
~\\
\begin{lemma}\label{lem:easy}
There is a constant $C>0$ depending only on $(\delta_0,\rho_0)$ such that
\begin{equation}
\int_{0}^{\infty}\psi_n(s)ds \ \leq \ C\, n^{-\alpha_0} \log(n)^{\frac 12(\beta_0-1)} .
\end{equation}
\end{lemma}
\proof The function $\Phi^n \Big(\textstyle\frac{t_n-\sqrt{\rho_0}s}{\sqrt{1-\rho_0}}\Big)$ is decreasing in $s$, and so
$$\int_0^{\infty}\psi_n(s)ds \leq \ \Phi^n \Big(\textstyle\frac{t_n}{\sqrt{1-\rho_0}}\Big).$$
Then, the inequality~\eqref{eqn:logupperbound} gives
\begin{equation}
\begin{split}
\Phi^n \Big(\textstyle\frac{t_n}{\sqrt{1-\rho_0}}\Big) & \ \leq \ \exp(-h_n(0))\\[0.3cm]
& \ = \ \exp\Big\{-\textstyle\frac{cn}{\sqrt{\log(n)}}e^{-\log(n)\delta_0^2}\Big\}\\[0.3cm]
%
& \ = \ \exp\Big\{\textstyle\frac{-c n^{1-\delta_0^2}}{\sqrt{\log(n)}}\Big\},
\end{split}
\end{equation}
which is clearly of smaller order than the stated bound. \qed
\subsection*{Proof of Proposition~\ref{prop:expec}}
We first show~\eqref{eqn:expeclower} in the case where $\mu_n=\mathbb{E}[M_n(X)]$. Let $X'\in\mathbb{R}^n$ be as in the proof of Corollary ~\ref{cor:small}. By integrating the inequality~\eqref{eqn:slepian}, Slepian's lemma implies
\begin{equation}
\mathbb{E}\big[M_n(X)\big] \ \geq \ \mathbb{E}\big[M_n(X')\big].
\end{equation}
In turn, the representation~\eqref{eqn:maxrep} gives
\begin{equation}\label{eqn:twostep}
\begin{split}
\mathbb{E}\big[M_n(X')\big] & \ = \ \sqrt{1-\rho_0}\, \mathbb{E}\Big[\max_{1\leq i\leq n}Z_i\Big]\\[0.3cm]
& \ \geq \ \sqrt{2(1-\rho_0)\log(n)} - c_0\sqrt{\ell_2(n)}
\end{split}
\end{equation}
for some universal constant $c_0>0$, where the second step can be obtained from~\cite[p.66]{Massart:2007}.
To handle $\mu_n=\text{med}[M_n(X)]$, note that
$$\Big|\mathbb{E}[M_n(X)]-\text{med}[M_n(X)]\Big| \ \leq \ \sqrt{\operatorname{var}[M_n(X)]} \ \leq 1,$$
where the second step follows from the Gaussian Poincar\'e inequality~\cite[Theorem 3.20]{Massart:2013}.
To show the upper bound~\eqref{eqn:expecupper}, note that we may take $X=X'$. When $\mu_n=\mathbb{E}[M_n(X')]$, the result follows from the first line of~\eqref{eqn:twostep} and the standard inequality $\mathbb{E}[\max_{1\leq i\leq n}Z_i] \leq \sqrt{2\log(n)}$. When $\mu_n=\text{med}[M_n(X')]$, the result follows from the fact that $\text{med}[M_n(X')]\leq \mathbb{E}[M_n(X')]$, since $M_n(\cdot)$ is a continuous convex function~\cite{Kwapien:1994}. \qed
~\\
\begin{proposition}\label{prop:extra}
If $R_{ij}=\rho_0$ for all $i\neq j$, then
\begin{equation}
\lim_{n\to\infty} \ \textstyle\frac{\textup{med}[M_n(X)]^2}{\operatorname{var}[M_n(X)]}\frac{1}{2\log(n)} \ = \ \frac{1-\rho_0}{\rho_0}.
\end{equation}
\end{proposition}
\proof Due to Proposition~\ref{prop:expec},
$$\lim_{n\to\infty}\textstyle\frac{\text{med}[M_n(X)]^2}{2\log(n)}=1-\rho_0.$$
To show that $\operatorname{var}[M_n(X)]\to \rho_0$, the representation~\eqref{eqn:maxrep} gives
\begin{equation}
\begin{split}
\operatorname{var}[M_n(X)] &= \rho_0+(1-\rho_0)\operatorname{var}\Big[\max_{1\leq i\leq n} Z_i\Big]\\[0.3cm]
& =\rho_0+\mathcal{O}\big(\textstyle\frac{1}{\log(n)}\big),
\end{split}
\end{equation}
where the second line follows from~\cite[Prop. 9.5]{Chatterjee:2014}.\qed
\subsection*{Acknowledgement} The author thanks Junwen Yao for carefully reading an initial draft of this work.
\bibliographystyle{alpha}
|
\section{Introduction} \label{Sec:intro}
\IEEEPARstart{S}{ensors} like Passive Infrared (PIR)~\cite{fang2006path}, camera~\cite{wang2003silhouette}, fingerprint scanner~\cite{jain2000biometric}, accelerometer~\cite{Begg_svm_gait}, interferometric reflectance imaging sensor (IRIS)~\cite{iris_lie}, microphone~\cite{Wang}, mm wave scanner~\cite{body_shape_mm_wave_person}, pressure sensor fabric~\cite{carpet_Zhou}etc. have been extensively used in biometric systems. However, most of these popular biometric sensors need active cooperation of the individual. Some require them to stand in front of a camera while others need them to place their finger on a scanner or keep an eye on the IRIS. However, in footfall based biometric system, the individuals are only required to walk through the active region of the sensor. The main advantages of this type of biometric system are: seismic sensors can be easily camouflaged; evading detection is impossible because footstep patterns are inimitable; it does not breach individual's privacy; less sensitive to environmental parameters and beyond the capacity of an individual to decode and manufacture the raw signal. Human identification system has significant applications in various areas such as office premises, classrooms, shopping malls, military areas, buildings, hospitals etc. It can be used for registering attendance of students in a classroom, employees in an office or workers in a workshop just by placing a seismic sensor or an array of sensors at the entrance gate. It has the capability to detect intruders in high security areas by identifying footsteps of unregistered users. Applications can also be developed to control home appliances in smart buildings using footfall~\cite{jun2015controlling}.
This paper implements a human identification system using the signatures of the seismic waves generated from footfalls. The seismic vibration created by rhythmic movement of heel and toe during walking is captured using a geophone. Processing a seismic signal in the ``pay-as-you-go" cloud in real time requires high computational power and bandwidth. So, in our implementation, we shifted the computational functions to the edge of the network using Fog computing paradigm~\cite{Preden_fog_mist,fog_computing_Tang}.
The main contributions of the paper are as follows: i) It proposes and implements a Fog based architecture for person identification using seismic waves generated by footfalls. ii) A footstep database containing a total of 46000 footfall events from 8 individuals were created.
iii) It proposes a basis pursuit based data compression technique \textit{DS8BP} for reducing bandwidth and energy during wireless transmission of the raw seismic signal. iv) Experiments have been conducted to obtain the best suited classifier (both in terms of accuracy and computational complexity), minimum sampling frequency of the signal and number of footsteps required per sample to achieve a respectable accuracy. v) Studies have also been carried out to find the minimum number of footsteps required (for each individual) to train a classifier for achieving a certain classification accuracy. vi) Experimental validation of the proposed model \textit{DS8BP} have been done and compared with other techniques.
The remainder of the paper is organized as follows: In Section~\ref{Sec:rel_work} we discuss the related works on footfall based human identification system. In Section~\ref{Sec:sys_arch} we introduce the Fog based paradigm for implementing an IoT based person identification system. Section~\ref{Sec:result_diss} presents the performance analysis of the proposed techniques. Finally, conclusion and future works are discussed in Section~\ref{Sec:future_con}.
\begin{table*}[]
\centering
\caption{Comparison among existing footstep based biometric system.}
\label{Tb:survey}
\resizebox{\textwidth}{!}{\begin{tabular}{|l|l|l|l|l|l|l|l|}
\hline
\multirow{2}{*}{Author(s)} & \multirow{2}{*}{Year} & \multirow{2}{*}{Organ} & \multirow{2}{*}{Sensor} & \multicolumn{2}{c|}{Database} & \multirow{2}{*}{ $\text{Algorithm}^\text{\textdagger}$} & \multirow{2}{*}{$\text{Accuracy}^\ast$} \\ \cline{5-6}
& & & & \# Person(s) & \# Samples & & \\ \hline
%
Addlesee \textit{et al.} \cite{Addlesee} & 1997 & Weight & Load cell & 15 & 300 & HMM & IA: $<$50\% \\ \hline
%
Middleton \textit{et al.} \cite{middleton_floor} & 2005 & Foot pressure & Pressure sensor & 15 & 180 & Decision trees & IA: 80\% \\ \hline
%
Suutala \textit{et al.} \cite{SUUTALA200821} & 2007 & Foot pressure & Electromechanical film & 11 & 440 & SVM & IA: 93.96\% \\ \hline
%
Qian \textit{et al.} \cite{Qian} & 2010 & Footfall & Pressure sensing floor & 11 & 5690 & LDA & IA: 92\% \\ \hline
%
Miyoshi \textit{et al.} \cite{Miyoshi_microphone} & 2011 & Footfall & Microphone & 12 & 720 & GMM & IA: 92.8\% \\ \hline
%
Yun \textit{et al.} \cite{Yun} & 2011 & Footfall & Photo sensor & 10 & 5000 & MLP & IA: 96.64\% \\ \hline
%
V.Rodriguez \textit{et al.} \cite{Rodriguez_footstep_mat} & 2013 & Foot pressure & Piezoelectric sensor & 120 & 20,000 & BTime, BSpace & EER: (5-15)\% \\ \hline
%
\multirow{3}{*}{Wang \textit{et al.} \cite{Wang}} & \multirow{3}{*}{2015} & \multirow{3}{*}{Walking interval} & Mic (carpet platform) & 15 & 450 & \multirow{3}{*}{Similarity based} & Pre: 91.78\% \\ \cline{4-6} \cline{8-8}
& & & Mic (carpet platform) & 15 & 600 & & Pre: 87.50\% \\ \cline{4-6} \cline{8-8}
& & & Mic (wooden platform) & 15 & 450 & & Pre: 86.68\% \\ \hline
%
%
Pan \textit{et al} \cite{Pan_Indoor_Person_Identification} & 2015 & Footfall & Seismic & 5 & - & SVM & IA: 93\% \\ \hline
%
Serra \textit{et al.} \cite{serra} & 2016 & Foot pressure & Polymer floor sensor & 13 & 529 & SVM & IA: 76.9\% \\ \hline
%
Pan \textit{et al}\cite{Pan_2017} & 2017 & Footfall & Seismic & 10 & & SVM & IA:96\% \\ \hline
%
%
Zohu \textit{et al.} \cite{carpet_Zhou} & 2017 & Foot pressure & Fabric based pressure sensor & 13 & 529 & SVM & IA: 76.9\% \\ \hline
%
Proposed & 2017 & Footfall & Seismic & 8 & 46,489 & SVM & IA: 90\%, 95\% $^{\ast \ast}$ \\ \hline
\multicolumn{8}{|l|}{\begin{tabular}[c]{@{}l@{}}
\textdagger Hidden Markov Model (HMM); Decision trees (DT); Linear Discrement analysis (LDA); Gaussian Mixture Model (GMM); Multiple Layer Perceptron (MLP); \\ Hamming Distance (HD); Support Vector Machine (SVM); Artificial Neural Network (ANN); Linear Regression (LR)
\\ $\ast$ Identification Accuracy (IA); (EER); Precision (Pre); False Acceptance Rate (FAR); False Rejection Rate (FRR)
\\ $\ast \ast$ 90\% (5 footstep/sample); 95\% (10 footstep/sample)
\end{tabular}} \\ \hline
\end{tabular}}
\end{table*}
\section{Related Works} \label{Sec:rel_work}
Broadly biometric properties can be divided into two categories, behavioural (gait, keystroke, signature, voice, footsteps) and physiological (face, fingerprint, palmprint, signature, DNA etc)~\cite{Wang,jain_biometric}. A biometric system~\cite{fog_face} collects biometric data from a person, extracts useful information from it, compares the extracted features with a standard database and predicts the identity of the person using pattern recognition techniques. Details of the existing biometric systems are presented chronologically in Table~\ref{Tb:survey}.
Yun \textit{et al.} \cite{Yun} proposed a system for human identification called UbiFloorII, which consisted of uniformly distributed photo interrupters that converted reflected light to voltage. They developed software modules to extract gait patterns of users from footfall samples. Their system consisted of 24 square tiles of length 30 cm and each tile consisted of 64 photo sensors (sensor density = 700 per $m^2$). V.Rodriguez \textit{et al.} \cite{Rodriguez_footstep_mat} used spatiotemporal information of the footstep signals collected using piezoelectric sensor for person recognition. They observed that the identification accuracies remain almost same (5\% to 15\% EER) in case of temporal and spatial approaches. However, the accuracy increases if the temporal and spatial features are fused together. They used two $45 \times 30 \ cm^2$ mats each containing 88 piezoelectric sensors (sensor density - 650 per $m^2$). Zohu \textit{et al.} \cite{carpet_Zhou} used a flexible fabric based pressure sensor for person identification and achieved an accuracy of 76.9\% with 13 participants. The fabric consisted of parallel electrodes on top and bottom layer separated by a non-conducting polymer substrate. The mat produced 120-by-54 pressure points and had a sensing area of 1.8\,m by 0.8\,m. Thus, person identification using footfall pressure based techniques require high number of sensors and have a small sensing region. On the contrary, in case of seismic signal based techniques, only a single geophone is required and they have a circular sensing range of radius 2.5 m(approx) in concrete. So, it is easier to scale up a system using seismic signal based approach.
DeLoney \textit{et al.} \, \cite{deloney2008person} proposed a person identification system by classifying sounds generated by footsteps. They used spectro-temporal modulations for classification. They achieved an accuracy more than 60\% for identifying 9 people wearing three different types of shoes on two types of floors. Miyoshi \textit{et al.} \cite{Miyoshi_microphone} used a microphone based technique for person identification. They recorded footstep data, detected footfalls, extracted features like Mel-Frequency Cepstral Coefficients (MFCCs), $\Delta$MFCCs, and $\Delta$Logarithm Powers, and used k-Nearest Neighbour (k-NN) and GMM to predict the identity of the person. Wang \textit{et al.} \cite{Wang} also proposed a microphone based person identification system. They collected datasets from 15 examinees in three different types of scenarios: examinees walking on concrete passage wearing their favorite shoe, examinees walking on a concrete passage wearing 4 types of shoes, and examinees walking on a wooden passage. Guo and Wang \cite{guo_mic_iet} used a Bayesian decision classifier to predict identities of individuals using features from footstep data recorded with the help of a microphone. The main drawback of using a microphone is that it gets affected by environmental noise and audible sounds.
Pan \textit{et al.} \,\cite{Pan_Indoor_Person_Identification} were the first research group to use structural vibration for human being identification. They extracted features related to gait patterns from different individuals. They calculated prediction accuracy using features from a single footstep and also from consecutive footsteps. They considered five footsteps with high SNR (Signal to Noise ratio) for prediction and achieved an average accuracy of 83\%. They have also performed confidence level thresholding which increased their accuracy to 96.5\%. Confidence level thresholding eliminated potential incorrect classification cases (it discarded almost 50\% of them) and tagged them as non-classifiable.
\begin{figure*}[]
\centering
\includegraphics[width=\linewidth]{Figure_seismic/Fog_Sys_Arch.pdf}
\caption{Three layer Fog computing architecture for person identification using seismic waves generate from footfalls. }
\label{fig:sysarch}
\end{figure*}
\section{System Architecture for footstep based Person Identification} \label{Sec:sys_arch}
In this paper, we present a Fog computing based architecture for implementing human identification system using seismic sensors (refer Fig.~\ref{fig:sysarch}). The proposed system is developed to monitor people moving around different regions in real time. Depending on applications, the entire monitoring area (housing complexes, colleges, factories, hospitals etc.) is divided into zones like buildings (for housing complex), departments(for college), factory floors (for factory), hospital blocks (for hospitals) etc. and sub-zones like room (of buildings), classrooms (of departments), workshop sheds (of factory floors), and hospital wards (of hospital blocks) etc. Each sub-zone consists of a single or multiple sensors responsible for identifying humans. A huge amount of bandwidth and energy will be required to transmit the entire seismic data to the cloud (for processing), as these sensors generate enormous amount of data. To overcome this problem we shifted the computational parts from the IoT cloud~\cite{iot_survey} to the edge of the network. Commercially available clouds (ThinkSpeak, Google Cloud, Microsoft Azure etc.) are costly, bear high latency and require huge communication bandwidth. So, it is not possible to implement a cloud based system which is connected to a large number of geographically distributed bulk data generating sensors. Hence for implementing the system we propose a three layer architecture as shown in Fig.~\ref{fig:sysarch}: i) \textit{Things} (Sensor + Low end Processor: $EP$ + Transceiver) ii) \textit{Fog} ( Embedded Processor: $EP^{++}$ + Transceiver) iii) \textit{Cloud} (Server). The proposed hierarchical structure diminishes the communication bandwidth and reduces the burden of the cloud by parallelizing the computational processes.
\textit{Things}: In each of the \textit{Thing}, a lower end embedded processor($EP$ -- Raspberry Pi Zero/ Orange Pi Zero) is connected to seismic sensor(s) and a low power long range transceiver module (XBee 868LP). The data transmission rate of the transceiver nodes are usually in the range of 80-100~Kbps (data rate of XBee 868LP is 80 kbps) which makes it impractical to transmit each sample of the seismic signal recorded by $EP$ to the \textit{Fog}. So, $EP$ uses an event extraction technique to find the portion of signal that represents a footfall and then compresses the event before transmitting. Each sub-zone has a single \textit{Thing} and all the \textit{Things} within a zone are connected to a single $EP^{++}$ present in \textit{Fog}.
\textit{Fog}: The embedded processor ($EP^{++}$ -- Raspberry Pi 3 model B) present in the Fog is responsible for collecting and processing data from all the \textit{Things} present within its zone. It receives the footfall signal over ZIGBEE network, decompresses it, extracts important features from it and finally classifies the signal for human identification. The final results are stored in a local database present within the Fog.
\textit{Cloud}: It communicates only with the $EP^{++}$ over Ethernet/Wi-Fi and hosts a single database which contains information related to human movement in different zones and sub-zones. To reduce the communication bandwidth, the Cloud periodically connects with the Fog and updates its database tables. These tables present in the Cloud and Fog can be used to design different applications for attendance, surveillance, activity monitoring~\cite{bodhi_intruder}, fall detection~\cite{fall_detection_Wang}, appliance control in smart homes~\cite{jun2015controlling} etc.
The activities performed by the processors present in the \textit{Things} and \textit{Fog} are explained below.
\begin{figure}[!b]
\centering
\includegraphics[width=\linewidth]{Figure_seismic/SeismicEvents.pdf}
\caption{Seismic Signal corresponding to ten consecutive footsteps. Each black dot ($\bullet$) represents a single footstep. }
\label{fig:seismic signal}
\end{figure}
\setcounter{MaxMatrixCols}{25}
\begin{figure*}[]
\begin{align*}
\mathbf{\boldsymbol{D}} & =\begin{pmatrix}
\vdots & \vdots & \vdots & \vdots & \vdots & \vdots &
\vdots & \vdots & \vdots & \vdots & \vdots & \vdots & & &
\vdots & \vdots & \vdots & \vdots & \vdots & \vdots
\\
\rotatebox{90}{\hspace{15pt} $f(t;\tau=0)$ } & \rotatebox{90}{$f_c(t;\tau=0,\omega=5)$} & \rotatebox{90}{$f_s(t;\tau=0,\omega=5)$} & \hdots & \rotatebox{90}{$f_c(t;\tau=0,\omega=250)$} &
\rotatebox{90}{$f_s(t;\tau=0,\omega=250)$} &
\rotatebox{90}{\hspace{15pt} $f(t;\tau=\frac{1}{L-1})$} & \rotatebox{90}{$f_c(t;\tau=\frac{1}{L-1},\omega=5)$} & \rotatebox{90}{$f_s(t;\tau=\frac{1}{L-1},\omega=5)$} & \hdots & \rotatebox{90}{$f_c(t;\tau=\frac{1}{L-1},\omega=250)$} &
\rotatebox{90}{$f_s(t;\tau=\frac{1}{L-1},\omega=250)$}
& \hdots & \hdots &
\rotatebox{90}{\hspace{15pt} $f(t;\tau=1)$ } & \rotatebox{90}{$f_c(t;\tau=1,\omega=5)$} & \rotatebox{90}{$f_s(t;\tau=1,\omega=5)$} & \hdots & \rotatebox{90}{$f_c(t;\tau=1,\omega=250)$} &
\rotatebox{90}{$f_s(t;\tau=1,\omega=250)$}
\\
\vdots & \vdots & \vdots & \vdots & \vdots & \vdots &
\vdots & \vdots & \vdots & \vdots & \vdots & \vdots & & &
\vdots & \vdots & \vdots & \vdots & \vdots & \vdots
\end{pmatrix}
\end{align*}
\caption{The Dictionary ($\boldsymbol{D} \in \Re^{L \times (1 + 2|w|)L}$) used for compressing the footfall events by \textit{DS8BP}.}
\label{Fig:Dictionary}
\end{figure*}
\subsection{Signal Acquisition} \label{SubSec:sig_rec}
$EP$s convert the analog seismic waves into digital by using an ADC of resolution 16 bits and sampling frequency($f_s$) of 8 kHz. Fig.~\ref{fig:seismic signal} shows a seismic signal containing human footsteps. The geophone used in the paper has an indoor range of 2.5\,m and an outdoor range of 25\,m.
\subsection{Event Extraction}
\label{SubSec:event_extraction}
An Event extraction technique is used to detect and fetch portions of seismic signal that represent footfalls. A fixed threshold based event detection technique fails to capture the events with time varying amplitudes (refer Fig.~\ref{fig:seismic signal}). So, we used an adaptive threshold based event extraction technique for footfall signal~\cite{bodhi_intruder, Anch1701:Predicting}. It was observed that if two consecutive footfalls occur within a short period of time, the event extraction technique detects them as a single event. This happens when an individual is walking very close ($<0.25m$) to the seismic sensor and a new footfall starts before the previous one damps out. The event detection technique also picks up floor/ground vibrations of high magnitude (movements of tables and chairs) other than footfall. However, these events have a short lifespan and die out quickly. Therefore, if the width of the detected event is too large then it is actually a combination of two footfalls and if it is too small then it is actually a noise. To get rid of these types of false events, an upper bound of 0.437\,sec and a lower bound of 0.144\,sec were set on the width of the events. These bounds were set by studying different types of footfall signals.
\begin{algorithm}[!b]
\caption{Compression Algorithm: \textit{DS8BP}}
\label{Algo:compression}
\begin{algorithmic}[1]
\State $\it{Sig}$ \Comment{Input: Seismic signal (corresponding a footstep) after event extraction }
\State $L_{GC}$ \Comment{Input: Lower threshold on the number of atoms}
\State $H_{GC}$ \Comment{Input: Higher threshold on the number of atoms}
\State $\it{Sig_{comp}}=[\ ]$ \Comment{Output: Compressed Signal }
\State $\it{I}=[ \ ]$ \Comment{Output: Atom Index }
\State $\it{L}=[ \ ]$ \Comment{Output: Length of the original signal}
\State $L = length(Sig)$
\State $\textbf{D}= \textit{generate\_dictionary}(L)$
\State $Sig_{ds}= \textit{Downsample}(Sig,8)$
\State $\hat x = \underset{x}{min} ||\textbf{D}x - Sig_{ds}||_2 + ||x||_1 $
\State $ [ \hat x_{sort}, \ \hat x_{index} ]= sort(\hat x,decreasing)$
\State $\hat x_{{energy}} = \sum_{i}(\hat x_{sort_i})^2$
\State $\hat x_{sortL }= \hat x_{sort }(1:L-1) $
\State $\hat x_{sortE} = \frac{Lower\_Tri(L-1,L-1)*\hat x_{sortL }}{\hat x_{{energy}}} $
\State $ \hat i= \max_i \ s.t. \ \sum_{i} \hat x_{sortE} < 1$
\State $I = x_{index}(1: \hat i)$
\If {$ L_{GC} \leq length(I) \leq H_{GC}$}
\State $\widehat{\textbf{D}} = \textbf{D}(:,I)$
\State $Sig_{comp}=\widehat{\textbf{D}}^{\dagger}*Sig_{ds}$
\Else
\State Discard $Sig$
\EndIf
\end{algorithmic}
\end{algorithm}
\subsection{Signal Compression and Decompression} \label{SubSec:sig_compress_decompress}
$EP$'s sample the seismic signals at 8kHz i.e each of them generates 8000 sample points per second. Transmitting data at such a high rate will require a high bandwidth and consume a lot of power. So, we proposed a basis pursuit~\cite{comp_sen_WSN} based compression technique (\textit{DS8BP}) to reduce the length of the signal. Down-sampling the raw signal is also an alternative to decrease signal length, however the proposed technique has capability to compress the signal much below the down-sampled signal in spite of maintaining an acceptable recovery error (refer Section~\ref{SubSec:pref_DS8BP}).
\textit{DS8BP} projects a down sampled version ($Sig_{ds}$) of a finite length seismic event ($Sig$) on an over complete dictionary $\boldsymbol{D}$. $\boldsymbol{D}$ is selected in such a manner that footfall events can be expressed by linear combinations of few columns (atoms) of $\boldsymbol{D}$ i.e coefficients of the columns of $D$ that represent $Sig_{ds}$ are always sparse in nature. It was observed that footfall signals can be represented by linear combinations of Gabor functions. Fig.~\ref{Fig:Dictionary} shows the dictionary $\boldsymbol{D} (\in \Re^{L \times (1 + 2|w|)L})$, where $f(t,\tau) = e^{ \frac{(t-\tau)^2}{\sigma^2}}$, $f_c(t,\tau,\omega) = e^{ \frac{(t-\tau)^2}{\sigma^2}}cos(\omega t)$, $f_s(t,\tau,\omega) = e^{ \frac{(t-\tau)^2}{\sigma^2}}sin(\omega t)$, $\omega = \{ 5,10,15, \hdots, 250 \}$, $\sigma = 0.5 \text{ (set heuristacally)}$, $t=\tau = \{ 0,\frac{1}{L-1},\hdots, 1 \}$, $|t|=|\tau| = L$ and $L$= length of the down-sampled signal $Sig_{ds}$ ($|.|$ represents the cardinality of the set). To generate the dictionary we only need to specify the value of $L$. Maximum value of $\omega$ is 250 Hz, as it is sufficient for representation of a footfall event.
Algorithm\ref{Algo:compression} shows the pseudocode for implementing \textit{DS8BP}. $\hat x$ represents coefficients of the columns in D required for decompressing the signal($Sig_{ds}$). $\hat x$ is obtained using LASSO regularization (Line 10 in Algorithm~\ref{Algo:compression}), so it is always sparse in nature. A new matrix $\widehat D$ is formed by eliminating the columns of $D$ whose corresponding coefficients in $\hat x$ are negligible. An energy based technique (Line 9 \-- 14 in Algorithm~\ref{Algo:compression}) is used to find the columns (Atom index($I$) ) of D. Seismic events ($Sig$) are discarded if the number of atoms required to represent the signal is lower than $L_{GC}$ or higher than $H_{GC}$ (refer Section.~\ref{SubSec:pref_DS8BP}). Finally, the compressed signal ($Sig_{comp}$) is obtained by projecting the down-sampled signal ($Sig_{ds}$) on the matrix $\widehat D$. The original signal is down-sampled before performing the compression technique to reduce the size of the dictionary. LASSO regularisation and pseudoinverse ($\dagger$) cannot be performed within an $EP$ if the size of $D$ (proportional to $L$ -- refer Fig.~\ref{Fig:Dictionary}) is very large. The compressed footfall event $Sig_{comp}$ is transmitted to $EP^{++}$ using XBee.
$EP^{++}$ recovers the original signal using Algorithm~\ref{Algo:de_compression}. For decompression we require following three quantities \--- a) compressed signal ($Sig_{comp}$), b) position of the significant atoms ($I$) and, c) length of the down-sampled signal ($L$) (to generate dictionary $\textbf{D}$). Using this information the original signal can be recovered by weighted summation of the columns of $\textbf{D}$ indexed by $I$. The elements in vector $Sig_{comp}$ represent the weight by which the selected columns of $D$ are multiplied.
\begin{algorithm}[!h]
\caption{De-Compression Algorithm}
\label{Algo:de_compression}
\begin{algorithmic}[1]
\State $\it{Sig_{comp}}$ \Comment{Input: Compressed Signal }
\State $\it{I}$ \Comment{Input: Atom Index }
\State $\it{L}$ \Comment{Input: Original Signal Length }
\State $\it{Sig_{rec}=[ \ ]}$ \Comment{Output: Recovered Signal}
\State $\textbf{D}= \textit{generate\_dictionary}(L)$
\State $\widehat {\textbf{D}} = \textbf{D}(:,I)$
\State $\it{Sig_{rec}}= \widehat {\textbf{D}} * Sig_{comp}$
\end{algorithmic}
\end{algorithm}
If $M$ is the length of the compressed signal ($Sig_{comp}$) then $10M +2 $ bytes ($M \times 8 + M \times 2 + 2$) of data are required to transmit a single footfall event. Elements of $Sig_{comp}$ are represented by 4 bytes (double precision floating point format) whereas $I$ and $L$ are represented by 2 bytes.
\subsection{Feature Engineering} \label{SubSec:feature_engineering}
Features are the most vital factor in any biometric system. We collected both time and frequency domain related features from the decompressed footfall signals. The length of a footfall and the gap between two consecutive footsteps (cadence) were also included in the feature vector. Features related to the energy of the signal in time domain were not considered, in order to make the feature set distance (between the person and the sensor) independent. From Fig.~\ref{fig:seismic signal}, it can be seen that the amplitudes of consecutive footfalls do not remain the same. Therefore, a person walking close or far from the geophone will make no difference until they are within the sensing range of the sensor. The extracted features are shown in Table~\ref{tab:features_table}.
\begin{table}[!t]
\centering
\scriptsize
\caption{Features extracted from a single seismic event.}
\begin{tabular}{|c|c|c|}
\hline
\textbf{Features} & \textbf{Parameter} & \textbf{No. of features} \\ \hline
\multirow{3}{*}{\textbf{Time Domain}} & Standard Deviation & 1 \\ \cline{2-3}
& Skewness & 1 \\ \cline{2-3}
& Kurtosis & 1 \\ \hline
\multirow{4}{*}{\textbf{Hilbert Transform}} & Mean & 1 \\ \cline{2-3}
& Standard Deviation & 1 \\ \cline{2-3}
& Skewness & 1 \\ \cline{2-3}
& Kurtosis & 1 \\ \hline
\multirow{4}{*}{\textbf{Frequency}} & Energy in 0-2 Hz. & 1 \\ \cline{2-3}
& Energy in 2-4 Hz. & 1 \\ \cline{2-3}
& $\vdots$ & $\vdots$ \\ \cline{2-3}
& Energy in 248-250 Hz. & 1 \\ \hline
\textbf{Gait Property 1} & Cadence (rhythm of footsteps) & 1 \\ \hline
\textbf{Gait Property 2} & Duration of a footstep(in samples) & 1 \\ \hline
\multicolumn{2}{|c|}{\textbf{Total Features}} & \textbf{135} \\ \hline
\end{tabular}
\label{tab:features_table}
\end{table}
\subsection{Classification and Data storage} \label{SubSec:classification}
Trained models are preloaded in $EP^{++}$ to predict the identity of individuals using the feature vectors. The final results are stored in a local database within $EP^{++}$. $EP^{++}$ uploads the results to the central database in the cloud\cite{think_speak} periodically.
\begin{table}[!b]
\caption{Details of the Data Set.}
\centering
\begin{tabular}{|c|c||c|c|}
\hline
\begin{tabular}[c]{@{}c@{}}Class / \\ Person\end{tabular} & No. of Footstep & \begin{tabular}[c]{@{}c@{}}Class /\\ Person\end{tabular} & No. of Footstep \\ \hline
1 & 5994 & 5 & 5431 \\ \hline
2 & 6251 & 6 & 5554\\ \hline
3 & 5299 & 7 & 6091\\ \hline
4 & 5809 & 8 & 6060\\ \hline
\end{tabular}
\label{Tb:dataset}
\end{table}
\section{Results and Discussion} \label{Sec:result_diss}
Footstep data of eight individuals (four males and four females) were collected using a geophone of 2.88\,V/mm/sec sensitivity and 100 gain. Average age of the individuals was between 20 to 25 years. The sensor was placed on the floor of the lab and each individual was made to walk (barefooted) in circles of radius 1\,m to 2.5\,m around it. The dataset consisted more than one hour of recorded seismic signals of each individual (collected over a period of one month). Footfall events were extracted from the signal using the adaptive threshold based event extraction technique (Section~\ref{SubSec:event_extraction}). The final dataset was created by manually annotating the extracted footfall events. The details of the dataset are shown in Table\,\ref{Tb:dataset}. To the best of our knowledge, this is the largest footstep database collected to date containing 46,489 footsteps. Certain criteria required for hardware implementation were studied: best suited classifier, optimal sampling frequency, and number of training data (footfalls) required per individual. Also it is difficult to achieve high prediction accuracy by using features only from a single footstep, as the feature space of different individuals remains overlapped. So, we have considered the mean of the features extracted from $F$ consecutive footsteps as a single sample. A study has also been conducted to obtain the optimal value of $F$. Performances\footnote{All the computations were performed using a 64 bit operating system running on an Intel(R) Core(TM) i7-4790 (CPU @ 3.60GHz x 8 ) processor with 16.00 GB RAM} of different multiclass classifiers~\cite{Pattern_recognition_bishop} (Logistic Regression (LR), SVM linear (SVM-Lin), SVM Gaussian (SVM-RBF), single hidden layer Artificial Neural Network (ANN), Linear Discriminate Analysis (LDA)) were tested using the dataset (Table\,\ref{Tb:dataset}).
\subsection{Selection of Optimal Number of footsteps/sample } \label{SubSec:optimal_no_footstep}
Table~\ref{Tb:accu_foot_per_sample} shows the performance of different classifiers on datasets of varying footsteps per sample. It can be observed that the performance of the classifiers escalates as the number of footstep/sample increases. SVM-RBF outperformed the rest of the classifiers when single footstep/sample was considered by achieving an accuracy of 71.2\%. Its performance was $11\%$ better than LR and $7\%$ better than SVM-Lin. It proved that the classes were not linearly separable in the feature space of 1 footstep/sample. There was a drastic increment in performance ( 9\% for SVM RBF and 11\%-12\% for the rest of the classifiers) as we moved from 1 to 2 footstep per sample, as in case of 1 footstep/sample the feature cadence was absent. However, there was only 5\%-6\% improvement in accuracy in 3 footsteps/sample over 2 footsteps/sample. We achieved an accuracy of 90\,\% with SVM-RBF by considering 5 footsteps/sample compared to 83\,\% achieved in~\cite{Pan_Indoor_Person_Identification}. They also considered five footsteps of highest SNR as a sample. As more number of footsteps were considered in a single sample, classes in the feature space become more distinct. All the classifiers had similar accuracies (about $97\%$) when 25 footsteps/sample were considered. In that case, the feature space of the classes became linearly separable and all machine learning techniques (even those not used in the paper) could easily classify the data with high accuracy.
The selection of an optimal number of footstep depends on the type of application for which the system is being used. However, 7 footsteps/sample is a good choice in terms of accuracy and data acquisition, as it is possible to collect 7 consecutive footsteps using a single geophone (sensing diameter of a geophone in indoor is 5~m approximately). 15-20 footsteps per sample can also be be used in applications which require very high accuracy. For collecting 15-20 consecutive footsteps in indoor condition an array of seismic sensors can to be used.
\begin{table}[]
\centering
\caption{Accuracy of different classifiers obtained by varying the number of footsteps/sample. ($f_s = 8 kHz$)}
\label{Tb:accu_foot_per_sample}
\resizebox{.49\textwidth}{!}{\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Classifiers\\ Accuracy(\%)\end{tabular}}} & \multicolumn{8}{c|}{\textbf{Number of Footsteps/sample}} \\ \cline{2-9}
& \textbf{1} & \textbf{2} &\textbf{3} & \textbf{5} & \textbf{7} & \textbf{10} & \textbf{15} & \textbf{25} \\ \hline
\textbf{SVM Lin.} & 64.5 & 76.4 & 82.0 & 88.2 & 91.1 & 94.4 & 96.6 & 98.0 \\ \hline
\textbf{SVM RBF} & 71.2 & 80.0 & 85.3 & 90.4 & 92.2 & 95.3 & 96.1 & 98.1 \\ \hline
\textbf{LR} & 60.9 & 73.9 & 79.4 & 85.7 & 90.5 & 93.5 & 96.0 & 97.3 \\ \hline
\textbf{LDA} & 54.5 & 65.8 & 71.8 & 80.1 & 84.8 & 88.4 & 93.8 & 96.2\\ \hline
\textbf{ANN} & 61.9 & 73.8 & 80.3 & 85.9 & 89.6 & 92.3 & 95.1 & 97.4 \\ \hline
\end{tabular}}
\end{table}
\subsection{Selection of best suitable Classifiers for person identification } \label{Sec:best_classifier}
The performances of the classifiers were obtained using a dataset where each sample represented features from 7 consecutive footsteps ($F=7$) sampled at 8~kHz. 10 fold cross validation technique was used to avoid over and underfitting of the classifiers. The test set was normalized by the mean and standard deviation obtained by normalizing the training set to imitate real time scenarios. The performance parameters used for classifier selection were accuracy, precision, recall, F1 score, and computational complexity.
It was observed that SVM-RBF outperformed rest of the techniques by achieving a prediction accuracy of 92.29\%. SVM-RBF was followed by SVM-Lin and LR, by achieving an accuracy of 91.90\% and 90.13\%. Out of all the classifier, LDA's performance was the worst (accuracy=84.76\%). It is due to the reason that LDA assumes the features are drawn from a Gaussian distribution and all the classes share a common covariance matrix. Out of all the classifiers, LR was computationally least expensive as it separated the classes in the dataset with straight lines. ANN performed better than LDA (accuracy of 88.97\%), however it required a lot of computational power for model training.
\subsection{Selection of Optimal Sampling Frequency} \label{Sec:samp_freq_opt}
The seismic waveform generated by footfalls are low frequency signals and can be sampled at a much lower sampling rate($f_s$) than 8 kHz. Reduction in the sampling rate decreases event detection time (EDT: total time to detect and extract all the events of a finite length signal divided by the total number of events) and feature extraction time per sample (FET). However, reducing the sampling frequency below a certain point deteriorates classifiers accuracy as downsampling affects the integrity of the signal. So, we have tried to obtain a sampling frequency below 8 kHz where the accuracy of the classifier remains unaltered. In this study, to obtain the optimal sampling frequency, we used SVM-RBF on a dataset containing 7 consecutive footsteps per sample. It was observed (refer Fig.~\ref{fig:accu_fs}) that accuracy remained unaltered (decreased by less than 2\%) when the sampling frequency was reduced from 8 kHz to 500 HZ whereas the EDT and FET reduced by 78\% and 99\%.
\begin{figure}[]
\centering
\includegraphics[width=\linewidth]{Figure_seismic/F1_EDT_FET.pdf}
\caption{Variation of F1 score, Event Detection time (EDT) and Feature extraction time (FET) of the signal with change in $f_s$. (Classifier used SVM-RBF \& each sample contains 7 footsteps)}
\label{fig:accu_fs}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{Figure_seismic/Learning_curve.pdf}
\caption{Learning Curve of all the classifiers on seismic data set. (Each sample has 7 footsteps, $f_s$=8 kHz)}
\label{fig:learning_curve}
\end{figure}
\subsection{Selection of Optimal Number of Training samples per class} \label{Sec:train_sample_opt}
From an implementation point of view, it is important to know the required number of training samples per class to obtain the desired classification accuracy. The number of training samples can be obtained from the learning curve of the classifiers (refer Fig.~\ref{fig:learning_curve}). It can be found that to achieve an accuracy greater than 85\% we need 875 footsteps per class i.e 8 minutes of walking (assuming a normal human has approximately 100 footsteps per minute). But to get an accuracy of 90\%, 17 minutes of seismic data are required. SVM-RBF has the highest learning rate compared to the other classifiers followed by LR, whereas ANN has the poorest learning rate. Using the learning curve (Fig.~\ref{fig:learning_curve}) a designer can easily pick an appropriate value for the number of footstep required to train his/her classifier.
In spite of the fact that LR has the least prediction time and has almost similar performance parameters to SVM-RBF, the latter is the best suited classifier for footstep data. SVM-RBF has better performance parameters (even when samples have lower number of footsteps: 71.2\% for 1 footstep/sample) and a better learning curve than the others. In case of single footstep/sample, the accuracy of LR is only 60\%. Table~\ref{Tb:conf_svm-rbf} shows the confusion matrix of a test instance of SVM-RBF classifier with 7 footsteps/sample and 500Hz as the sampling frequency.
\begin{table}[!h]
\centering
\caption{Confusion Matrix of the SVM-RBF for 7 footsteps/sample and 500 Hz sampling frequency.}
\noindent\begin{tabular}{cc*{8}{|E}|}
\cline{3-10}
\multirow{8}{*}{\rotatebox[origin=c]{90}{\textbf{Predicted Class}}}
&\textbf{P1} & 0.94 & 0 & 0 & 0 & 0 & 0.01 & 0 & 0 \\ \cline{3-10}
&\textbf{P2} & 0 & 0.99 & 0 & 0 & 0 & 0 & 0 & 0 \\ \cline{3-10}
&\textbf{P3} & 0.01 & 0 & 0.98 & 0.06 & 0 & 0.01 & 0 & 0.05 \\ \cline{3-10}
&\textbf{P4} & 0.04 & 0 & 0 & 0.98 & 0 & 0 & 0 & 0 \\ \cline{3-10}
&\textbf{P5} & 0 & 0 & 0 & 0 & 0.94 & 0.03 & 0.02 & 0 \\ \cline{3-10}
&\textbf{P6} & 0 & 0 & 0 & 0.02 & 0.05 & 0.93 & 0 & 0.03 \\ \cline{3-10}
&\textbf{P7} & 0 & 0 & 0 & 0 & 0 & 0 & 0.95 & 0 \\ \cline{3-10}
&\textbf{P8} & 0 & 0.01 & 0.02 & 0 & 0.02 & 0.02 & 0.03 & 0.92 \\ \cline{3-10}
& \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{P1}} & \multicolumn{1}{c}{\textbf{P2}}
& \multicolumn{1}{c}{\textbf{P3}} & \multicolumn{1}{c}{\textbf{P4}} & \multicolumn{1}{c}{\textbf{P5}}
& \multicolumn{1}{c}{\textbf{P6}} & \multicolumn{1}{c}{\textbf{P7}} & \multicolumn{1}{c}{\textbf{P8}}\\
& \multicolumn{1}{c}{} & \multicolumn{8}{c}{\textbf{Actual Class}}
\end{tabular}
\label{Tb:conf_svm-rbf}
\end{table}
\subsection{Performance of the proposed Compression and De-compression Technique} \label{SubSec:pref_DS8BP}
The proposed technique \textit{DS8BP} (discussed in Section.~\ref{SubSec:sig_compress_decompress}) used for signal compression first down samples the 8 kHz ($Sig$) signal to 1 kHz ($Sig_{ds}$)) and then uses basis pursuit technique to decompress it. The signals which have low compression factors (length $Sig_{comp}$/length $Sig_{ds}$) need large number of basis functions (i.e columns of $D$) for representation. Footsteps are low frequency signals and can be represented by a handful of basis functions. A seismic event with very low compression factor is not a footfall, they might be high amplitude noise wrongly picked up by the event extraction technique. Noise constitutes different frequencies from low to high and requires a large number of basis function for representation. Hence, signals whose compression factors were less than 5 ($L_{GC}$) were eliminated. Also, those signals that had very high compression factor above 40 ($H_{GC}$) were not recoverable as they had very few basis functions. \textit{DS8BP} acts as a filter for eliminating signals falsely picked up by the event extraction technique. The values of $L_{GC}$ and $H_{GC}$ were set experimentally.
The compression technique \textit{DS8BP} was compared with another technique named \textit{DS16}. \textit{DS16} decimated the seismic signal by a factor of 16. The decimation factor was kept at 16 (i.e. sampling frequency 500 Hz) as it was the minimum sampling frequency upto which the integrity of the signal remained intact (refer Sec.\ref{Sec:samp_freq_opt}). Average length of a footfall event was around 250.62 millisecond (i.e 2005 samples/event when sampled at 8 kHz). By using the footfall events in the dataset (Table ~\ref{Tb:dataset}), it was observed that in case of \textit{DS8BP}, the average compression factor (length $Sig_{comp}$/length $ Sig_{ds}$) of a single footstep was 13.54 and the average length of a compressed footfall event ($Sig_{comp}$) was 18.51 ($M$). It was also noticed that \textit{DS8BP} compresses the original 8 kHz signal by a factor of 108.32 ($8 \times 13.54$) whereas \textit{DS16} compresses only by a factor of 16.
A wireless transmitter (XBee 868LR) with a data rate of 80 kbps will requires 18.71 millisecond \ $(\frac{(10M + 2) \times 8}{80 \times 1000})$ and 100 millisecond $(\frac{(125 \times 8) \times 8}{80 \times 1000})$ to transmit a single footfall event when compressed by \textit{DS8BP} and \textit{DS16}.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{Figure_seismic/Orig_Recov.pdf}
\caption{ The original ($Sig$) and the decompressed signal ($Sig_{rec}$) of a footfall event. The compression and decompression were performed using Algorithm~\ref{Algo:compression} and Algorithm~\ref{Algo:de_compression}.}
\label{fig:orig_recov}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{Figure_seismic/csvsnon_cs.pdf}
\caption{Performance analysis of classifiers on recovered signals after being compressed by $DS16$ and $DS8BP$.}
\label{fig:cs_noncs_acc}
\end{figure}
Fig.~\ref{fig:orig_recov} shows the original ($Sig_{ds}$), and recovered footfall event($Sig_{rec}$) corresponding to a single footstep obtained using \textit{DS8BP}. It can be seen that the recovered signal completely matches with the original signal and does not contain high frequency components. The recovered signal is formed by a linear combination of a handful of Gabor functions (20 in this case of Fig.~\ref{fig:orig_recov}), with a maximum frequency of 250 Hz. So there is no basis that represents the high frequency components of the original signal. Fig.~\ref{fig:cs_noncs_acc} shows the effect of the compression techniques on classifiers' performance. The solid lines and the dotted lines represent the accuracies of classifiers (SVM-RBF, SVM-Lin, LR) on decompressed signals. The signals are compressed by using \textit{DS16} and \textit{DS8BP} respectively. Fig.~\ref{fig:cs_noncs_acc} shows the accuracies of the classifiers w.r.t the number of footsteps per sample. In all the three classifiers, signals compressed by \textit{DS8BP} technique achieved higher classification accuracy as compared to signals compressed by \textit{DS16}. This is due to the fact that \textit{DS8BP} de-noises the signals as it works as a smoothing function. The accuracy of SVM-RBF, SVM-Lin, and LR increases by (2\%-4.3\%) in case of 1 footstep/sample and (.8\%-2.4\%) with 3 footsteps/sample when \textit{DS8BP} is used instead of \textit{DS16}.
\section{Conclusion and Future Work } \label{Sec:future_con}
In this paper, we have implemented a Fog computing architecture based person identification system using seismic events generated from their footfalls. The following observations are made \--- clubbing consecutive footsteps per sample increases the prediction accuracy, downsampling the 8 kHz signal to 500 Hz does not affect classifiers' accuracy, although it reduces FET(feature extraction time) and EDT(event detection time). The proposed signal compression technique \textit{DS8BP} reduces the 8 kHz seismic signal by 108 times and increases the classifiers' accuracy by (2-4)\% in 1 footstep/sample scenario. \textit{DS8BP} also increases the security of the system as its impossible to decode the compressed signal without knowing the dictionary ($D$).
The proposed set up for human identification assumes that a single person was present within the active region of the sensor. This assumption is valid in certain indoor scenario, however in an outdoor environment there are chances that more than one person can step into the active arena of the geophone. So, techniques that can separate combined footstep signals of two (or more) individuals might be explored in future. As this work has achieved very high accuracy for person identification, it can be extended to detecting intruder i.e. non registered persons (anomaly detection). Its applications can be further extended to surveillance, border monitoring etc. Deep Neural algorithms: Convolution Neural Network, Recurrent Neural Network (RNN) can be used for feature extraction and classification.
\bibliographystyle{IEEEtran}
\section{Selection of Optimal Number of footsteps/sample}
Fig.~S\ref{fig:acc_footstep_each_clfr} shows the performance of different classifiers as footsteps/sample are increased from 1 to 25. It can be clearly observed that as the number of footsteps increases, the performance of linear classifiers (SVM-Lin, LR, LDA) become almost equal to that of non-linear classifiers (SVM-RBF, ANN). The poor performance of the linear classifiers suggests that the classes in the feature space of single footstep/sample are overlapped and are not linearly separable. The overlapping among the classes decreases with the increase in the number of footsteps, as it reduces the effect of noise in the feature space.
\begin{figure}[!h]
\centering
\includegraphics[width=\linewidth]{Figure_seismic/acc_footstps.pdf}
\caption{Accuracy of different classifier obtained with increasing number of footfalls/sample. (~$f_s$ is 8kHz~)}
\label{fig:acc_footstep_each_clfr}
\end{figure}
\section{Selection of best suitable Classifiers for person identification}
Table~S\ref{Tb:perf_diff_classifiers} shows the performances (accuracy, precision, recall and F1 Score) of each of the classifiers on the footstep database. It also displays the F1 score of all the individual classes (P1,P2, $\cdots$, P8) and prediction time of a single sample. All the computations were performed using a 64 bit operating system running on an Intel(R) Core(TM) i7-4790 (CPU @ 3.60GHz x 8 ) processor with 16.00 GB RAM. Data set used for performance analysis contained features extracted from 7 consecutive footsteps recorded with a sampling frequency ($f_s$) of 8 kHz.
\begin{table*}[!t]
\centering
\caption{Performance of different classifiers on the footstep seismic data. (Each sample consisted of 7 footsteps, $f_s$ = 8 kHz)}
\label{Tb:perf_diff_classifiers}
\resizebox{\textwidth}{!}{\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{\textbf{Classifier}} & \multirow{2}{*}{\textbf{Value}} & \multirow{2}{*}{\textbf{Accuracy}} & \multirow{2}{*}{\textbf{Precision}} & \multirow{2}{*}{\textbf{Recall}} & \multirow{2}{*}{\textbf{F1 Score}} & \multicolumn{8}{c|}{\textbf{Individual F1 Scores}} & \textbf{Time(ms)} \\ \cline{7-15}
& & & & & & \textbf{P1} & \textbf{P2} & \textbf{P3} & \textbf{P4} & \textbf{P5} & \textbf{P6} & \textbf{P7} & \textbf{P8} & \\ \hline
\textbf{LR} & Mean & 90.13 & 90.01 & 90.04 & 90.09 & 93.70 & 96.15 & 91.12 & 88.86 & 87.51 & 88.80 & 90.51 & 83.46 & 0.1 \\ \hline
\textbf{} & std. & 1.32 & 1.30 & 1.30 & 1.29 & 2.24 & 1.71 & 2.18 & 2.68 & 3.29 & 2.72 & 1.50 & 2.43 & \\ \hline
\textbf{SVM-Lin} & Mean & 91.90 & 91.82 & 91.81 & 91.89 & 94.94 & 97.59 & 92.58 & 92.22 & 89.46 & 90.67 & 91.91 & 85.23 & 22.6 \\ \hline
\textbf{} & std. & 0.98 & 1& 1& 1.02 & 2.22 & 0.59 & 2.11 & 2.07 & 1.75 & 2.44 & 1.55 & 3.28 & \\ \hline
\textbf{SVM-RBF} & Mean & 92.29 & 92.19 & 92.20 & 92.25 & 94.72 & 97.82 & 92.36 & 91.57 & 90.20 & 92.52 & 92.68 & 85.66 & 32.8 \\ \hline
\textbf{} & std. & 1.23 & 1.25 & 1.22 & 1.27 & 1.52 & 1.41 & 3.01 & 1.55 & 2.85 & 1.92 & 2.73 & 1.76 & \\ \hline
\textbf{ANN} & Mean & 88.97 & 88.85 & 88.87 & 88.97 & 91.38 & 95.43 & 90.21 & 87.53 & 87.31 & 88.21 & 90.65 & 80.12 & 1.7 \\ \hline
\textbf{} & std. & 1.30 & 1.32 & 1.31 & 1.28 & 1.49 & 0.57 & 3.36 & 2.14 & 3.97 & 3.30 & 2.96 & 4.00 & \\ \hline
\textbf{LDA} & Mean & 84.79 & 84.81 & 84.77 & 85.23 & 86.34 & 95.19 & 86.91 & 78.38 & 84.33 & 85.64 & 87.32 & 74.37 & 2.5 \\ \hline
& std. & 1.17 & 1.18 & 1.20 & 1.15 & 1.46 & 0.52 & 3.05 & 2.26 & 2.06 & 2.64 & 1.93 & 3.01 & \\ \hline
\end{tabular}}
\end{table*}
LR~\cite{Pattern_recognition_bishop} is a binary classifier with categorical dependent variable i.e. class 1 and class 2 are represented by 0 and 1. The function used by LR for prediction is:
\begin{equation}
\label{Eq:LR}
f_{LR}(\boldsymbol{x};\boldsymbol{\theta})=\frac{1}{1+ e^{-\boldsymbol{\theta^T} \boldsymbol{x}} }
\end{equation}
where $\boldsymbol{\theta} (\in \mathfrak{R}^{n+1})$
is the model parameter or weights, $\boldsymbol{x}\in\mathfrak{R}^{n+1} \ ([x_0=1, \ x_1, \ \cdots \ x_n ]^T) $ is the feature vector corresponding to a single sample, and $n$ is the number of features. A test sample $\boldsymbol{x_{test}}$ belongs to class 1 if $f_{LR}(\boldsymbol{x_{test}};\boldsymbol{\theta}) > 0.5 $ and to class 0 if $f_{LR}(\boldsymbol{x_{test}};\boldsymbol{\theta}) < 0.5 $. Multi-class classification by LR was implemented by using one-vs.-rest technique. The simplicity of the LR model makes it easily implementable in a lower end embedded processors. However, its biggest drawback is its performance with data having lower number of footsteps per sample (60\,\% in case of single footstep/sample, refer Fig.~S\ref{fig:acc_footstep_each_clfr}).
SVM-Lin~\cite{svm} showed almost similar prediction accuracy with respect to LR, as both of them linearly separates the dataset. However here the classes are labelled as +1 (class 1) and -1 (class 2). It is modelled using the function
\begin{equation}
\label{Eq:SVM-lin}
f_{SVM-Lin}(\boldsymbol{x};\{ \boldsymbol{w},b\})= sign( \boldsymbol{w}^T\boldsymbol{x} + b )
\end{equation}
where $\boldsymbol{w}(\in \mathfrak{R}^{n})$ is the weight vector, $\boldsymbol{x}$ is the feature vector and $b \in \mathfrak{R}$ is a scalar.
A grid-search was performed to find the optimal value of the hyper parameter $C (\in \mathfrak{R})$ and was set to 1. From Fig.~S\ref{fig:acc_footstep_each_clfr}, it can be seen that SVM-Lin outperforms LR in terms of accuracy by 2-3\,\% when number of footsteps/sample is below 7. It is due to the soft margin of SVM-Lin, as it allows some miss-classification in the training dataset during model training.
SVM is well known for its Kernel trick which works very efficiently when the classes in the feature space are not linearly separable. SVM-RBF is modeled by the function
\begin{equation}
\label{Eq:SVM-RBF}
f_{SVM-RBF}(\boldsymbol{x};\{ \alpha_i,K,\boldsymbol{x_{sv}},b\})= sign (\sum_{i = 1}^{m} \alpha_i K(\boldsymbol{x_{sv}^{(i)}},\boldsymbol{x})+ b )
\end{equation}
where $K$ is the Kernel function, $\boldsymbol{x_{sv}^{(i)}}$ is the $i^{th}$ support vector, $\alpha_i$ is a positive constant, $\boldsymbol{x}$ is the feature vector of the test sample, $b \in \mathfrak{R}$ is a scalar and $m$ is the number of support vectors. The Kernel function for RBF is
\begin{equation}
\label{Eq:kernel}
K (\boldsymbol{x_{sv}^{(i)}},\boldsymbol{x}) = exp(-\gamma ||\boldsymbol{x_{sv}^{(i)}} - \boldsymbol{x}||^2)
\end{equation}
Soft-margin SVM-RBF has two hyper parameters $C$ and $\gamma$. The values of C and $\gamma$ were obtained using a grid search and were found to be 100 and 0.001. SVM-RBF performed consistently well among the other classifiers because, its Kernel($K$) transformed the non linearly separable low dimension features to a linearly separable high dimensional feature. Table~S\ref{Tb:conf_svm-rbf_25footstep} shows the confusion matrix of SVM-RBF on a dataset obtained using features from 25 consecutive footfalls.
ANN~\cite{ANN} (single hidden layer) was the most computationally expensive learning technique used in this paper. It took maximum amount of training time w.r.t to the other classifiers. However, its prediction time was better than SVM-Lin and SVM-RBF. The number of neural net used in the single layer ANN was 40 (obtained by grid search). Performance-wise it does not produce better results than other classifiers. The performance of LDA~\cite{lda} was worst among all the classifiers. It is because LDA assumes the features are drawn from a Gaussian distribution and they share a common covariance matrix.
\begin{table}[!h]
\caption{Confusion Matrix of SVM-RBF for 25 footsteps/sample and 500 Hz sampling frequency}
\label{Tb:conf_svm-rbf_25footstep}
\noindent\begin{tabular}{cc*{8}{|E}|}
\cline{3-10}
\multirow{8}{*}{\rotatebox[origin=c]{90}{\textbf{Predicted Class}}}
&\textbf{P1} & 1 & 0 & 0 & 0.01 & 0 & 0 & 0 & 0 \\ \cline{3-10}
&\textbf{P2} & 0 & 1& 0 & 0 & 0 & 0 & 0 & 0 \\ \cline{3-10}
&\textbf{P3} & 0 & 0 & 0.96 & 0 & 0 & 0 & 0 & 0 \\ \cline{3-10}
&\textbf{P4} & 0 & 0 & 0.02 & 0.98 & 0 & 0 & 0 & 0 \\ \cline{3-10}
&\textbf{P5} & 0 & 0 & 0 & 0 & 1 & 0 & 0.02 & 0.02 \\ \cline{3-10}
&\textbf{P6} & 0 & 0 & 0 & 0.01 & 0 & 1& 0 & 0 \\ \cline{3-10}
&\textbf{P7} & 0 & 0 & 0 & 0 & 0 & 0 & 0.96 & 0.02 \\ \cline{3-10}
&\textbf{P8} & 0 & 0 & 0.02 & 0 & 0 & 0 & 0.02 & 0.96 \\ \cline{3-10}
& \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{P1}} & \multicolumn{1}{c}{\textbf{P2}}
& \multicolumn{1}{c}{\textbf{P3}} & \multicolumn{1}{c}{\textbf{P4}} & \multicolumn{1}{c}{\textbf{P5}}
& \multicolumn{1}{c}{\textbf{P6}} & \multicolumn{1}{c}{\textbf{P7}} & \multicolumn{1}{c}{\textbf{P8}}\\
& \multicolumn{1}{c}{} & \multicolumn{8}{c}{\textbf{Actual Class}}
\end{tabular}
\end{table}
\begin{table*}[]
\centering
\caption{Prediction accuracy of classifiers with and without compression technique on the footfall event. (NC: Not compressed ($f_s$=8 kHz))}
\label{Tb:comp_ds8bp_ds16}
\begin{tabular}{|c|c|c|l|c|c|c|}
\hline
\multirow{2}{*}{\textbf{Classifiers}} & \multirow{2}{*}{\textbf{Compression}} & \multicolumn{5}{c|}{\textbf{Number of Footsteps/sample}} \\ \cline{3-7}
& & \textbf{1} & \multicolumn{1}{c|}{\textbf{2}} & \textbf{3} & \textbf{5} & \textbf{7} \\ \hline
\multirow{3}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}SVM\\ RBF\end{tabular}}} & \textbf{NC} & 71.20 & 80.00 & 85.3 & 90.40 & 92.20 \\ \cline{2-7}
& \textbf{DS16} & 70.49 & 80.15 & 84.69 & 89.69 & 91.33 \\ \cline{2-7}
& \textbf{DS8BP} & 73.37 & 81.22 & 85.51 & 90.16 & 93.10 \\ \hline
\multirow{3}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}SVM\\ Linear\end{tabular}}} & \textbf{NC} & 64.50 & 76.40 & 82.00 & 88.20 & 91.10 \\ \cline{2-7}
& \textbf{DS16} & 64.03 & 76.11 & 81.46 & 87.17 & 90.36 \\ \cline{2-7}
& \textbf{DS8BP} & 67.51 & 78.15 & 83.85 & 89.08 & 92.63 \\ \hline
\multirow{3}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Logistic\\ Regression\end{tabular}}} & \textbf{NC} & 60.90 & 73.90 & 79.40 & 85.70 & 90.50 \\ \cline{2-7}
& \textbf{DS16} & 60.62 & 73.10 & 78.59 & 85.77 & 89.34 \\ \cline{2-7}
& \textbf{DS8BP} & 64.92 & 75.50 & 80.16 & 85.77 & 91.40 \\ \hline
\end{tabular}
\end{table*}
\section{Performance of the proposed Compression and De-compression Technique}
Fig.S\ref{fig:compression_hist} shows the histogram of compression factors i.e. length(compressed signal-$Sig_{comp}$)/length(downsampled signal-$ Sig_{ds}$) achieved by \textit{DS8BP}. The histogram includes all the footfall signals (Table III of the main manuscript) whose compression factor is between 5($L_{GC}$) and 40($H_{GC}$). The average compression factor achieved by \textit{DS8BP} is 13.54 with a standard deviation of 4.68.
The datagram used for transmitting footfall signal from $EP$ to $EP^{++}$ over a Zigbee network is shown in Fig.~S\ref{fig:datagram}. Fig.~S\ref{fig:compression_sys} draws a graphical comparison between \textit{DS16} and \textit{DS8BP}.
\begin{figure}[]
\centering
\includegraphics[width=\linewidth]{Figure_seismic/compress_hist.pdf}
\caption{Histogram of Compression Factor [length($Sig_{comp}$)/length($ Sig_{ds}$] achieved by \textit{DS8BP}. }
\label{fig:compression_hist}
\end{figure}
\begin{figure}[]
\centering
\includegraphics[width=\linewidth]{Figure_seismic/datagram}
\caption{Datagram representation of a single footstep. $10 \times M+2$~bytes are required to represent a single footstep.}
\label{fig:datagram}
\end{figure}
\begin{figure}[]
\centering
\includegraphics[width=.9\linewidth]{Figure_seismic/CompressionSys}
\caption{Comparison between \textit{DS8BP} and \textit{DS16} regarding comprehensibility capability. }
\label{fig:compression_sys}
\end{figure}
Table.~S\ref{Tb:comp_ds8bp_ds16} shows the prediction accuracy of SVM-RBF, SVM-Lin, and LR when compression techniques \textit{DS8BP} and \textit{DS16} are used on the original footfall signal. In case of \textit{DS8BP}, the features are extracted from the de-compressed signal ($Sig_{rec}$, Algorithm 2 in the main manuscript). And in \textit{DS16} features are extracted from the downsampled (decimation factor = 16) footfall signals. In NC (No Compression) features are extracted from the original 8 kHz signal. Table.~S\ref{Tb:comp_ds8bp_ds16} clearly shows that the performance of NC and \textit{DS16} is almost similar. However, \textit{DS8BP} increases classifiers accuracy (2\%-4\%), (1\%-2\%) in case of single footstep/sample and 7 footsteps/sample when compared to other techniques.
\begin{figure}[]
\centering
\begin{subfigure}[b]{.5\textwidth}
\includegraphics[width=.9\linewidth]{Figure_seismic/PrototypeFog}
\caption{Floor plan for implementing human identification system using seismic sensor. }
\label{fig:PrototypeFog}
\end{subfigure}%
\begin{subfigure}[b]{.45\textwidth}
\includegraphics[width=\linewidth]{Figure_seismic/thing.jpg}
\caption{A \textit{Thing} constituted the following components ; 1)~Geophone; 2)~Xbee 868 LP; 3)~Antennae; 4)~Raspberry Pi Zero ($EP$); 5)~Sound Card (ADC); 6)~Battery. }
\label{fig:EP}
\end{subfigure}%
\begin{subfigure}[b]{0.35\textwidth}
\centering
\includegraphics[width=\linewidth,]{Figure_seismic/fog.jpg}
\caption{A \textit{Fog} unit consisted of the following components: Raspberry Pi 3 model B ($EP^{++}$); Xbee 868 LP; Battery. }
\label{fig:EP++}
\end{subfigure}%
\caption{Floor planning and hardware components for implementing person identification system.}\label{fig:hardare_comp}
\end{figure}
\section{Hardware Implementation}
We have implemented \textit{Fog} computing based human identification system in our lab using geophones. Fig.~S\ref{fig:PrototypeFog} shows the floor plan of our implementation. We have divided the lab (IIA-307) into two zones: Z-1 and Z-2. Each of the zones consisted only a single Sub-Zone (SZ(a)) and each Sub-Zone had only one \textit{Thing}. Fig.~S\ref{fig:EP} and Fig.~S\ref{fig:EP++} show components used for implementing a \textit{Things} and a \textit{Fog Unit}. Fig.~S\ref{fig:flowchart} shows the flowcharts of algorithms running on the processors present in \textit{Things}, \textit{Fog}, and \textit{Cloud}. $EP$ records a $t$ sec signal and extracts footfall events from it. It then compresses the footfall events individually using \textit{DS8BP} and transmit them to $EP^{++}$ using XBee~\cite{faludi2010building}. The value of $t$ depends on the application for which the system is being used. Here, we have set the value of $t$ to 10 sec. The processor in the \textit{Fog} ($EP^{++}$) extracts features from the footfall events and then classifies the signal. It finally stores the predicted results in its local database. The \textit{Cloud} fetches information from these local databases (present in the $EP^{++}$s) periodically and updates its own database. $T_j$ ($j=1,...,L$) is an application specific variable, it determines the frequency by which the \textit{Cloud} updates its tables. Here $F_i$ represents $i^{th}$ \textit{Fog Unit} (in our experiment i={1,2}). Six students had volunteered (three in each zone) for our experiment on human identification. \footnote{To prevent identity disclosure the students were denoted as Person\textit{i} (where \textit{i}=1,2,3).}
Fig.S\ref{fig:table_fog1} and Fig.S\ref{fig:table_fog2} displays the local databases present with in the two \textit{Fog Unit}s. Fig.S\ref{fig:table_cloud} shows the global database present in the \textit{Cloud} which periodically copies the database of the \textit{Fog Units}.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{Figure_seismic/SupplimentryFlowChart}
\caption{Flow charts of Algorithms running in \textit{$EP$}, \textit{$EP^{++}$} and \textit{Cloud}.}
\label{fig:flowchart}
\end{figure}
\begin{figure}[]
\centering
\begin{subfigure}[b]{.5\textwidth}
\includegraphics[width=\linewidth]{Figure_seismic/Fog1_table}
\caption{Local database (PersonFog1) hosted in $EP^{++}$ of \textit{Fog Unit 1}. }
\label{fig:table_fog1}
\end{subfigure}%
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=\linewidth,]{Figure_seismic/Fog2_table}
\caption{Local database (PersonsFog2) hosted in $EP^{++}$ of \textit{Fog Unit 2}. }
\label{fig:table_fog2}
\end{subfigure}%
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=\linewidth,]{Figure_seismic/ThingSpeak2.jpg}
\caption{Global dtabase in the \textit{Cloud}. }
\label{fig:table_cloud}
\end{subfigure}%
\caption{Local databases in the \textit{Fog} and global database in the \textit{Cloud}.}\label{fig:tables_fog_cloud}
\end{figure}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:int}
The gravitational wave detections by the LIGO and Virgo collaborations opened the door to test the strong field regime of gravity~\cite{Abbott:2016blz,Abbott:2016nmj,Abbott:2017oio,Abbott:2017vtc,TheLIGOScientific:2017qsa,Abbott:2017gyy}. In this regime, especially in the merger of compact objects, general relativity (GR) shows all its nuances and one has to study it through the use of full numerical relativity machinery. Additionally, with the possibility of GR not being the true theory of gravity, these scenarios are ideal to observe signatures of alternative theories of gravity~\cite{TheLIGOScientific:2016src} (see also~\cite{Konoplya:2016pmh,Yunes:2016jcc,Berti:2015itd,Cardoso:2016ryw,Berti:2018cxi,Berti:2018vdi,Barack:2018yly}).
Since the discovery of GR, a plethora of alternative theories of gravity have been proposed, many of them motivated by problems related to the nonrenormalizabity of GR. An interesting class of alternative theories of gravity is the dynamical Chern-Simons (dCS) theory (see Ref.~\cite{Alexander:2009tp} for a review). The theory was introduced by Jackiw and Pi~\cite{Jackiw:1990mb}, but it can also arise from heterotic superstring scenarios~\cite{Campbell:1990fu,Campbell:1992hc}. We can write the dCS action as (we will use natural units $G=c=\hbar=1$)\footnote{We could also add a constant in the scalar field action, but this can be absorbed in a redefinition of the scalar field~\cite{Pani:2011xj}.}
\begin{align}
S&=\frac{1}{16\pi}\int d^4x\sqrt{-g}\Big\{R
+\frac{\alpha}{4}\varphi\,{^*RR}\nonumber\\
&-\frac{1}{2}[g^{ab}\nabla_a\varphi\nabla_b\varphi+V_s(\varphi)]\Big\}+S_{\rm mat},
\end{align}
where $g$ is the metric determinant, $R$ is the Ricci scalar, $\varphi$ is the scalar field,
\begin{equation}
{^*RR}=R_{abcd}^{*}R^{bacd}=\frac{1}{2}R_{abcd}\epsilon^{baef}{R^{cd}}_{ef},
\end{equation}
the Pontryagin density, and $V_s(\varphi)$ is the scalar field potential. The dCS theory can also be thought of as a subclass of a more general class of theories of gravity called quadratic gravity~\cite{Alexander:2009tp,Yunes:2011we,Pani:2011gy}. Usually, the scalar field potential is set to zero, but here we will consider it to describe a massive scalar, i.e.,
\begin{equation}
V_s=\mu^2|\varphi|^2,
\end{equation}
where $\mu$ is the scalar field mass. The consideration of a massive scalar field can be thought of as a dominant term in an expansion of the potential $V_s$ in powers of $\varphi$. As such, we will consider it to be relatively small, scaled with the black hole (BH) mass as $M\mu\leq 0.5$.
Moreover, small mass ranges are particularly interesting when rotation is considered, as the superradiant instability is highly suppressed for large $\mu M$~\cite{Zouros:1979iw,Dolan:2007mj}.
We will investigate dynamical features of massive dCS by using perturbation theory. Because the Pontryagin density vanishes for spherically symmetric spacetimes~\cite{Grumiller:2007rv}, spherical BHs are still described by the Scwarzschild spacetime, i.e.,
\begin{equation}
ds^2=-f(r)dt^2+f(r)^{-1}dr^2+r^2d\Omega^2,
\end{equation}
where $f(r)=1-2M/r$.
Metric and scalar field perturbations of BHs in dCS gravity can be studied by using the standard formalism introduced by Regge-Wheller~\cite{Regge:1957td} and Zerilli~\cite{Zerilli:1970se}. Perturbations of spherically symmetric BHs in dCS were studied in the past~\cite{Alexander:2009tp,Molina:2010fb,Pani:2011xj}, and we will closely follow the definitions of Ref.~\cite{Pani:2011gy}. One can find that the even sector of the perturbation is precisely the same as that in GR, as scalar field perturbations couple only with the axial sector. Therefore, we will only study the axial sector in this paper. Within this picture, one can compute the quasinormal modes (QNMs) of the system, which are natural frequencies of the BH, behaving as purely ingoing waves at the horizon and outgoing waves at infinity~\cite{Berti:2009kk}.
The QNMs in massless dCS was investigated in detail in Ref.~\cite{Molina:2010fb}, where the behavior of the mode frequencies were analyzed for a wide range of the parameter space. Additionally, the authors briefly discussed the influence of the mass term into the QNMs of static BHs in dCS.
Essentially, with the influence of the scalar mass, the behavior of the scalar modes as a function of the coupling constant is analogous to the massless case. Moreover, the gravitational mode dependence on the scalar mass is very mild. However, a study of one important part of the spectrum that arises when the mass term is present is still lacking: the quasibound (QB) states~\cite{Deruelle:1974zy,Damour:1976kh}. QB states are localized long-lived solutions,
slowly leaking to the BH\footnote{We will see that in the dCS the
QB mode will also leak to infinity mainly through the gravitational channel [see Eq.~\eqref{eq:binf2}].}.
This part of the spectrum is important because it is responsible
for the superradiant instabilities rotating spacetime solutions~\cite{Dolan:2007mj}.
In this work, we explore the QB states of massive dCS gravity, considering as the background the Schwarzschild spacetime. We compute the monopolar and dipolar modes, considering different values for the coupling constant. We show that the coupling impacts mostly the imaginary part of the modes, influencing, therefore, the decay time of initial perturbations that excites the modes. We obtain in this frequency window that the spacetime is stable, corroborating previous studies in the literature~\cite{Molina:2010fb,Kimura:2018nxk}. We also study how the mass term affects the time evolution of initial signals. We perform time evolutions of Gaussian wave packets, showing that the gravitational sector of the perturbations presents, at late times, a behavior similar to that of massive scalars.
The remainder of this paper is organized as follows. In Sec.~\ref{sec:pert}, we present the equations describing axial perturbations in massive dCS gravity. We present the direct integration method we will use to find the QB frequencies and the Breit-Wigner method to verify the results. We also present the setup for the time evolution of initial Gaussian profiles. In Sec.~\ref{sec:results}, we present the numerical results for the QB modes and the results for the time evolution of gravitational and scalar perturbations. Finally, in Sec.~\ref{sec:conclusion}, we present our discussion and conclusions.
\section{Perturbations of BHs in massive dCS gravity}
\label{sec:pert}
The equations describing axial perturbations of Schwarzschild black holes in massive dCS can be written as\footnote{See Ref.~\cite{Pani:2011gy} for a derivation.}
\begin{equation}
\left(\frac{\partial^2}{\partial r_\star^2}-\frac{\partial^2}{\partial t^2}-{\bf V}\right){\bf \Psi}(t,r_\star)=0,
\label{eq:perturbation}
\end{equation}
where $r_\star$ is the tortoise coordinates, defined by $dr_\star=f(r)^{-1}dr$, and ${\bf \Psi}=\{\psi_g,\Theta\}$, in which $\psi_g$ represents the gravitational part of the perturbations and $\Theta$ is the scalar one. The potential $\bf V$ is given by
\begin{equation}
{\bf V}=\left[
\begin{array}{cc}
V_{11} & V_{12}\\
V_{21} & V_{22}
\end{array}\right],
\label{eq:system}
\end{equation}
with
\begin{align}
V_{11}&=f\left(\frac{\ell( \ell+1)}{r^2}-\frac{6M}{r^3}\right),\\
V_{12}&=f\frac{96 \pi \alpha M}{r^5},\qquad V_{21}=f\frac{6 \alpha M (\ell+2)!}{r^5(\ell-2)!},\label{eq:crosspot}\\
V_{22}&=f\left[\frac{\ell(\ell+1)}{r^2}\left(1+\frac{576\pi M^2\alpha^2}{r^6}\right)+\frac{2M}{r^3}+\mu^2\right].\label{eq:v22}
\end{align}
To compute the modes of the spacetime, we separate the time dependence by ${\bf \Psi}(t,r)={\bf \Psi}(r) e^{-i\omega t}$ and supplement the system of equations with proper boundary conditions. These boundary conditions usually represent purely outgoing waves at infinity and ingoing waves at the horizon, which are referred to as QNMs~\cite{Berti:2009kk}. However, because of the mass term, one can also have QB states in which the scalar field is suppressed at large distances. Note that, due to the form of $V_{11}$, the gravitational field is not suppressed, behaving as an outgoing wave at infinity. Here, we will explore QB states for the scalar fields in dCS.
Near the horizon, we require purely ingoing boundary condition. We have
\begin{equation}
{\bf \Psi}(r_\star \to-\infty)\approx e^{-i\omega r_\star}\sum_{i=0}^{N}(r-2M)^i\{B_{g,i},B_{\Theta,i}\} ,
\label{eq:expanhori}
\end{equation}
where the coefficients $B_{g,i}$ and $B_{\Theta,i}$ are obtained by expanding the differential equations in powers of $(r-2M)$ near the horizon, solving for coefficients iteratively up to the order of ${\cal O}[(r-2M)^N]$.
The boundary condition at the horizon, given by Eq.~\eqref{eq:expanhori}, depends only on two constants $(B_{g,0},B_{\Theta,0})$.
We can, therefore, obtain numerically two independent ingoing solutions at the horizon, namely, $\bf \Psi_{1}^{(-)}$ and $\bf \Psi_{2}^{(-)}$, by choosing $(B_{g,0},B_{\Theta,0})=(1,0)$ and $(0,1)$ and integrating Eq.~\eqref{eq:perturbation} outwards. A generic solution of Eq.~\eqref{eq:perturbation} that is ingoing at the event horizon can be obtained by
\begin{equation}
{\bf \Psi}_{-}=\beta_{1}^{(-)}{\bf \Psi_{1}^{(-)}}+\beta_{2}^{(-)}{\bf \Psi_{2}^{(-)}}.
\label{eq:qnm1}
\end{equation}
At large distances, the boundary condition is more involved. We have to deal with two different asymptotic forms in order to properly describe the boundary conditions at infinity (see Appendix B of Ref.~\cite{Cardoso:2009pk} for a different approach). The first boundary condition can be obtained by making
\begin{equation}
{\bf \Psi}(r_\star \to\infty)\sim r^{\mp\nu}e^{\pm k r_{\star}}\sum_{i=0}^N \left\{\frac{A_{g,i}}{r^{i+5}},\frac{A_{\Theta,i}}{r^i} \right\},
\label{eq:binf1}
\end{equation}
where $k=\sqrt{\mu^2-\omega^2}$, and $\nu=M\mu^2/k$. The upper signal indicates the QNM condition (${\rm Re}(k)>0$), and the bottom indicates the QB condition (${\rm Im}(k)<0$)~\cite{Rosa:2011my,Dolan:2007mj}. The second boundary condition is given by
\begin{equation}
{\bf \Psi}(r_\star \to\infty)\sim e^{i\omega r_{\star}}\sum_{i=0}^N \left\{\frac{A_{g,i}}{r^{i}},\frac{A_{\Theta,i}}{r^{i+5}} \right\}.
\label{eq:binf2}
\end{equation}
It is worth noticing that the appearance of a damped term in the gravitational part of Eq.~\eqref{eq:binf1}
and an outgoing wave term in the scalar part of Eq.~\eqref{eq:binf2} is due to the coupling between the two fields.
In fact, this is the reason why the first contribution of these is $\propto r^{-5}$ [see the potentials $V_{12}$ and $V_{21}$
in Eq.~\eqref{eq:crosspot}]. Inserting Eqs.~\eqref{eq:binf1} and~\eqref{eq:binf2} into Eq.~\eqref{eq:perturbation} and expanding in powers of $r^{-1}$ leads to recurrence relations for the coefficients. In the end, these recurrence relations can be solved for the coefficients as functions of $A_{\Theta,0}$ [in the case of Eq.~\eqref{eq:binf1}] and of $A_{g,0}$ [in the case of Eq.~\eqref{eq:binf2}]. Therefore, as it happens for the boundary conditions at the horizon, we can construct two independent solutions integrating the differential Eq.~\eqref{eq:perturbation} from the numerical infinity, using $(A_{g,0},A_{\Theta,0})=(1,0)$ and $(0,1)$, namely, $\bf \Psi_1^{(+)}$ and $\bf \Psi_2^{(+)}$, respectively. Hence, the general solution satisfying the proper boundary conditions at infinity is
\begin{equation}
{\bf \Psi}_+=\beta_1^{(+)}{\bf \Psi_1^{(+)}}+\beta_2^{(+)}{\bf \Psi_2^{(+)}}.
\label{eq:qnm2}
\end{equation}
QB states can be found by requiring both boundary conditions at infinity and at the horizon to be satisfied. This means that, from Eqs.~\eqref{eq:qnm1} and~\eqref{eq:qnm2},
\begin{align}
{{\bf \Psi}_-}&={{\bf \Psi}_+},\label{eq:cond1}\\
\frac{d}{dr_\star}{{\bf \Psi}_-}&=\frac{d}{dr_\star}{{\bf \Psi}_+},\label{eq:cond2}
\end{align}
holds when $\omega$ is the frequency of the QB state. In practice, Eqs.~\eqref{eq:cond1} and \eqref{eq:cond2} are computed at some intermediate point, say $r=r_m$. As the system is linear, we can set one of the $\beta$ coefficients to unity and use three of the components of Eqs.~\eqref{eq:cond1} and \eqref{eq:cond2} to solve for the remaining $\beta$. The last component can be solved to find the mode frequencies $\omega$. Equivalently, we can impose that the determinant
\begin{equation}
\det({\bf \Psi_{1}^{(-)}},{\bf \Psi_{2}^{(-)}},{\bf \Psi_{1}^{(+)}},{\bf \Psi_{2}^{(+)}})
\label{eq:det}
\end{equation}
vanishes when the frequency is the eigenvalue of the system, generating complex values for the frequency, namely, $\omega=\omega_R+i\,\omega_I$, with $(\omega_R,\omega_I)$ being real values.
The method described above is usually referred to as direct integration method (DI)~\cite{Chandrasekhar:1975zza,Pani:2013pma,Macedo:2016wgh}, and it has been applied in the past to other classes of quadratic gravity~\cite{Blazquez-Salcedo:2016enn,Blazquez-Salcedo:2016yka}.
Another method to compute the modes and frequencies that is more efficient is the continued fraction (CF) method~\cite{Leaver:1985ax,Nollert:1993zz}. The CF method relies on finding an appropriate expression to represent the fields such that it satisfies the proper boundary conditions of the problem (see, e.g., Refs.~\cite{Pani:2013pma,Macedo:2016wgh}).
For coupled systems, such as the one presented in this paper, the situations are more complex because we have to consider a matrix-valued recurrence relation (see Refs.~\cite{Cardoso:2009pk,Pani:2013pma} for instance). For our specific problem, we also have to find an expression that handles both with the behavior of outgoing waves and damped for the perturbations as presented in Eqs.~\eqref{eq:binf1} and~\eqref{eq:binf2}. Therefore, a more robust method---such as the CF---to find the modes is beyond the scope of this paper. Notwithstanding, the DI method usually handles QB states computations well, as in the case of the parameter space explored in this paper.
\subsection{Monopolar and dipolar modes}
Differently from GR, axial perturbations in dCS gravity pre\-sent monopolar ($\ell=0$) and dipolar ($\ell=1$) radiative modes, linked with the scalar degree of freedom of the theory. One can show that the these modes can be described by the same equation, given by
\begin{equation}
\left(\frac{d}{dr_\star}-V_{22}\right)\Theta=0,
\label{eq:dec}
\end{equation}
where $V_{22}$ is given by Eq.~\eqref{eq:v22}. Because the system reduces to a single second-order differential equation,
it is much simpler to analyze, compared to the $\ell>1$ modes. Note that for $\ell=0$, the modes
have a behavior that is precisely the same as that of a massive scalar field in the Schwarzschild background because the term proportional
to $\alpha$ vanishes. On the other hand, the $\ell=1$ mode does have an additional term proportional to the coupling parameter.
The $\ell=1$ mode is especially important because it is this mode that presents the most prominent superradiant instabilitiy
in rotating BH spacetimes~\cite{Dolan:2012yt}. Therefore, one can analyze if the coupling makes the mode less susceptible
to this instability by looking into the influence of the coupling into the imaginary part.
One may be tempted to adopt Eq.~\eqref{eq:dec} as an approximation scheme for higher multipoles, which would be the case for a Dudley-Finley-like scheme~\cite{Dudley:1977zz}. However, we verified that, although it reduces to the GR case in the limit $\alpha\to 0$, Eq.~~\eqref{eq:dec} \textit{does not} reproduce the behavior of the modes computed using the full set of equations~\eqref{eq:perturbation}. This reflects the importance of the gravitational feedback into the scalar QB modes.
\subsection{Breit-Wigner method for the bound states}
Another method that we can exploit is the Breit-Wigner resonant method~\cite{1991RSPSA.434..449C,Berti:2009wx,Pani:2013pma,Macedo:2018yoi}. This method works well for modes with high-quality factors, i.e.,
$|\omega_R/\omega_I|\gg 1.$
Essentially, we have that the determinant given by Eq.~\eqref{eq:det} near the QB frequencies can be expanded as~\cite{Pani:2013pma}
\begin{equation}
|\det({\bf S})|^2\propto (\omega-\omega_R)^2+\omega_I^2.
\label{eq:bw}
\end{equation}
The modes can be found by sweeping $\omega$ through real values, finding the above determinant, and fitting the modes to Eq.~\eqref{eq:bw}. As QB states are usually high-quality factor modes, we can use the Breit-Wigner method as a check for the modes computed using the full direction integration method.
\subsection{Time evolution of initial data}
In order to understand the influence of the QB states into gravitational wave signals, we also perform a time evolution of the system, similarly to the one presented in Refs.~\cite{Gundlach:1993tp,Molina:2010fb}. We write the system~\eqref{eq:perturbation} using the light-cone coordinates, $u=r_\star-t$ and $v=r_\star+t$, obtaining
\begin{equation}
4\frac{\partial^2}{\partial u\,\partial v}{\bf \Psi}=-{\bf V}{\bf \Psi}.
\label{eq:systemp}
\end{equation}
We solve the system~\eqref{eq:systemp} with Gaussian initial conditions, i.e.,
\begin{align}
{\bf \Psi}(0,v)&=\left[
\begin{array}{l}
A_1 \exp\left(-\frac{(v-v_{c1})^2}{2 \sigma_1}\right)\\
A_2 \exp\left(-\frac{(v-v_{c2})^2}{2 \sigma_2}\right)
\end{array}\right],~{\rm and}~\label{eq:init1}\\
{\bf \Psi}(u,0)&=\left[
\begin{array}{l}
0\\
0
\end{array}\right].
\label{eq:init2}
\end{align}
In general, it is expected that the results, after a transient regime, do not depend on the choice of initial data~\cite{Nollert:1993zz}. As such, for our purpose, it is sufficient to consider only the initial Gaussian data of the type given by Eqs.~\eqref{eq:init1} and~\eqref{eq:init2}.
For massive fields, time evolutions shed light in many directions. One can study the frequency of the QB states by analyzing the late-time signal through a fre\-quen\-cy-fil\-ter technique~\cite{Dolan:2012yt}. One can also look into the influence of the scalar field on the
gravitational signals, which can be studied, for instance, through only gravitational initial data [i.e., setting $A_2=0$ in Eq.~\eqref{eq:init1}]. Because of the coupling, after an initial time, the scalar field will have an influence in the gravitational wave signal (and \textit{vice versa}). Therefore, the mass of the scalar field will have an influence on the power-law tail. For the massless case, the power-law tail behavior is given by $\propto t^{-(2\ell+3)}$. For the massive case, it is given by~\cite{Koyama:2001ee,Koyama:2001qw,Burko:2004jn,Witek:2012tr}
\begin{equation}
{\bf \Psi}\sim t^{p} \sin(\omega_c t),
\label{eq:latetime}
\end{equation}
with $p=-(\ell+3/2)$ at intermediate times and $p=-5/6$ at very late times, and $\omega_c$ spans over values close to the scalar field's mass, typically related to the QB state frequencies. This gives us possible distinctive signatures of alternative theories with extra massive degrees of freedom.
\section{Numerical results}\label{sec:results}
Here, we present our numerical results, parametrizing them by $\zeta=16\pi\alpha^2/M^4$. In order to validate our numerical code, we computed the QNMs in the massless case and performed time evolutions in the GR limit. For the massless case, we compare our the results with the ones presented in Ref.~\cite{Molina:2010fb}, and the time evolution and massive modes with~\cite{Burko:2004jn,Dolan:2007mj}. Our results are in excellent agreement with the ones presented in the literature.
\subsection{Quasibound states}
\begin{figure}%
\includegraphics[width=0.95\columnwidth]{mode1.pdf}%
\caption{Imaginary part of the fundamental QB frequencies, normalized by the GR value, as a function of the coupling constant $\zeta$. \textit{Top panel}: For the $\ell=1$ mode we see that, in the mass range studied here, the mode decreases in absolute value for high values of the coupling $\zeta$, meaning that it is becoming less stable. \textit{Bottom panel:} The $\ell=2$ mode behaves differently from the $\ell=1$ mode, increasing in absolute value, becoming more stable. For both $\ell=1$ and $\ell=2$, these effects are attenuated for higher values of the mass.}%
\label{fig:mode1}%
\end{figure}
We apply the DI procedure described in Sec.~\ref{sec:pert}. We focus on the dipolar and quadrupolar modes, but we also verify that the results for $\ell=3$ are qualitatively similar to the $\ell=2$ ones. We recall that the monopolar mode ($\ell=0$) is precisely the same as that in GR. The dipolar case is important in order study possible implications of the coupling constant into the superradiant instability, once rotation is considered. The quadrupolar mode can show possible signatures in gravitational wave observations.
\begin{figure*}%
\includegraphics[width=\columnwidth]{time1.pdf}\includegraphics[width=\columnwidth]{time1_2.pdf}%
\caption{Time evolution of an initial quadrupolar ($\ell=2$) gravitational Gaussian profile for different scalar masses, $\zeta=1$ (left panels) and $\zeta=10$ (right panels). \textit{Top panel:} Gravitational perturbation is initially insensitive to the scalar influence, but later on, it becomes contaminated by it, presenting an oscillatory behavior. \textit{Bottom panel:} Scalar perturbation is initially smaller than the gravitational perturbation, becoming larger after some time. }%
\label{fig:time1}%
\end{figure*}
\begin{figure*}%
\includegraphics[width=\columnwidth]{time2.pdf}\includegraphics[width=\columnwidth]{time2_2.pdf}%
\caption{Time evolution of gravitational wave packets considering different values for the coupling $\zeta$, $M\mu=0.1$ (left panels) and $M\mu=0.2$ (right panels). We see that the characteristic oscillation appears earlier for higher values of the coupling $\zeta$, but the frequency of these oscillations is basically the same (see the inset).}%
\label{fig:time2}%
\end{figure*}
In Fig.~\ref{fig:mode1} we show the imaginary part of the $\ell=1$ and $\ell=2$ fundamental QB frequency,
normalized by the GR value ($\omega_{I}^{(GR)}$), for different values of the mass $M\mu$. The behavior of the dipolar modes is very different from that of the quadrupolar one.
In the top panel of Fig.~\ref{fig:mode1}, we see that the dipolar mode becomes less stable as the coupling $\zeta$ grows, meaning that the imaginary part decreases in absolute value. This implies that these modes live even longer than they do in GR, making them more susceptible to the superradiant instability once rotation is considered. The opposite happens for the imaginary part of the $\ell=2$ mode (bottom panel of Fig.~\ref{fig:mode1}).
We see that the imaginary part increases in absolute value, becoming more stable and, therefore, more damped. We confirm these results by estimating the mode frequencies looking into the Breit-Wigner expression presented in Sec.~\ref{sec:pert}, comparing with the DI result.
We note that the behavior described above switches when considering $M\mu\sim0.5$ for small couplings. As we can see in Fig.~\ref{fig:mode1}, there is a maximum in the $\ell=1$ case and a minimum for the $\ell=2$ case. Nonetheless, for high values of the coupling, the behavior of the modes is the same as that described above, i.e., $\ell=1$ modes gets less stable and $\ell=2$ more stable.
For the real part of the modes, we find that in all cases it is basically insensitive to changes in the coupling $\zeta$, the differences being $<0.1\%$ from the GR modes in the range considered here. Hence, in the massive dCS gravity, the oscillation of the perturbations remains basically the same, but the damping time depends on the coupling.
In all cases described here, we performed a numerical search for unstable QB modes, finding none. Our numerical results further corroborate the conclusion that the Schwarzschild spacetime in massive dCS gravity is stable. Note that this conclusion is supported from studies of another independent method, the $S$-deformation method~\cite{Kimura:2018nxk}.
\subsection{Time evolution of gravitational pulses}
Here, focus on numerical evolution of an initially pure gravitational Gaussian packet in order to understand the influence of the scalar sector on gravitational waves, considering $A_2=0$ in Eq.~\eqref{eq:init1}. The Gaussian packet is centered at $v_{c1}=10M$ and has a width $\sigma=M$, and the wave functions are extracted at $r_\star=50M$.
In Fig.~\ref{fig:time1} we show the result of a time evolution considering different values for the scalar field mass, considering $\zeta=1$ (left panels) and $\zeta=10$ (right panels). We can see that the initial gravitational response for all cases is the same. After some time, the gravitational field becomes contaminated with the scalar field, which impacts the behavior of the late-time tails not only in the power-law profile, but also by oscillations. The frequency of these oscillations is predominantly given by the QB frequencies, similarly to what happens in the GR case~\cite{Burko:2004jn} (see also Ref.~\cite{Degollado:2014vsa}). The scalar field perturbation is initially smaller than the gravitational one, as expected because the pulse is initially only gravitational. However, the scalar perturbation grows in time, becoming larger at some intermediate times depending on $M\mu$ and $\zeta$.
\begin{figure}%
\includegraphics[width=\columnwidth]{time_late}%
\caption{Time evolution of gravitational wave packets for $M\mu=0.3$, considering $\zeta=1$ and $\zeta=10$. In this plot, we can see that the intermediate and very late time behavior of the gravitational perturbation is similar to the massive case in pure GR, given by Eq.~\eqref{eq:latetime}.}%
\label{fig:tail}%
\end{figure}
To see the role of the coupling constant on the gravitational signal, in Fig.~\ref{fig:time2} we plot the results for the time evolution with $M\mu=0.1$ and $M\mu=0.2$, considering $\zeta=1$ and $\zeta=10$.
For $\zeta=10$ the scalar field has an impact on the gravitational perturbations at earlier times than in the $\zeta=1$ case.
Additionally, in the late-time regime, the amplitude of the gravitational wave is larger in the $\zeta=10$ than it is for $\zeta=1$, demonstrating a clear influence of the cross terms in the late-time behavior. Also, we can see that the oscillation of the late field is the same for the two cases, which corroborates our result that the real part of the modes is almost insensitive to changes in $\zeta$.
In Fig.~\ref{fig:tail} we plot the results for the time evolution with $M\mu=0.3$, considering $\zeta=1$ and $\zeta=10$. We can readily see that the power law for the gravitational perturbation is not the same as that in GR. Instead, we see that the behavior fits the ones given by Eq.~\eqref{eq:latetime}, which shows further imprints of the scalar field's mass onto the gravitational sector of the perturbations.
\section{Discussion and conclusion}\label{sec:conclusion}
Perturbation of BH spacetimes is a timely topic, helping us to understand the dynamics of these objects with a linear system of equations. In this work, we studied axial perturbations of static, spherically symmetric BHs in dCS gravity with a massive scalar field. Our work extends previous works in the literature, analyzing the behavior of QB states, which are long-living modes (partially) damped by the mass term. We found that the real part of the modes is insensitive to the coupling constant, and that the dipolar mode becomes less stable and the quadrupole more stable, as the coupling constant increases. We also analyzed the influence of the mass term into gravitational signals propagating in the BH spacetime, through the study of time evolutions of gravitational Gaussian packets. We showed that the scalar field mass has a direct influence on the late-time behavior of the signals.
Considering the QB states, our results show that BHs in massive dCS gravity can be, in principle,
more susceptible to the superradiant instabilities of rotating black holes when compared to the GR counterpart. This susceptibility
will depend on how the perturbations are coupled when rotation is considered. The results presented here give an indication of QB mode dependence on the coupling constant. A further study considering rotating black holes is needed to present the whole picture in this scenario and, therefore, is a logical extension of the present paper.
For the time evolution of gravitational perturbations, our results show that the theory presents clear signatures at late times, which could be potentially observed once we can access the gravitational tail. Additionally, this late-time behavior could be a signature of theories that are coupled with massive fields, not being exclusive to massive dCS gravity. For instance, within GR, if one considers the gravitational radiation triggered by the interaction of massive scalar field packets with black holes, the power-law tail will oscillate with a characteristic frequency that depends essentially on the QB frequency of the scalar field~\cite{Degollado:2014vsa}, similarly to what we described here. Clearly, because the scalar field in dCS is directly coupled with the gravitational perturbations, the oscillation is more pronounced, as it could possibly be the case of other similar theories of gravity.
\section*{Acknowledgments}
It is a pleasure to thank Hector O. Silva, Leandro A. Oliveira, Luís C. B. Crispino, Paolo Pani, and Vitor Cardoso for useful comments and discussions. I would like to thank Conselho Nacional de Desenvolvimento Científico e Tecnológico
(CNPq), Coordenação de Aperfeiçoamento de Pessoal
de Nível Superior (CA\-PES), and Fundação Amazônia de Amparo a Estudos e Pesquisas (FAPESPA), from Brazil, for partial financial
support.
|
\section{Introduction}
In this paper we propose the Single-equation Penalized Error Correction Selector (SPECS) as a tool to perform automated modelling of a potentially large number of time series of unknown order of integration. In many economic applications, datasets will contain possibly (co)integrated time series, which has to be taken into account in the statistical analysis. Traditional approaches include modelling the full system of time series as a vector error correction model (VECM), estimated by methods such as maximum likelihood estimation \citep{Johansen1995}, or transforming all variables to stationarity before performing further analysis. However, both methods have considerable drawbacks when the dimension of the dataset increases.
While the VECM approach allows for flexible modelling of potentially cointegrated series, these estimators suffer from the curse of dimensionality due to the large number of parameters to estimate. In practice they therefore quickly become difficult to interpret and computationally intractable on even moderately sized datasets. As such, to reliably apply such full-system estimators requires non-trivial a priori choices on the relevance of specific variables to keep the dimension manageable. Moreover, often one only has a single variable of interest, and estimating the parameter-heavy full system is not necessary.
On the other hand, the alternative strategy of prior transformations to stationarity is more easily compatible with single variables of interest and larger dimensions, but requires either a priori knowledge of the order of integration of individual variables, or pre-testing for unit roots, which is prone to errors in particular if the number of variables is large \citep[cf.][]{Smeekes2020}. Additionally, this approach ignores the presence of cointegration among the variables, which may have detrimental effects on the subsequent analysis.
SPECS is a form of penalized regression designed to sparsely estimate a conditional error correction model (CECM). We demonstrate that SPECS possesses the oracle property as defined in \citet{Fan2001}; in particular, SPECS simultaneously allows for consistent estimation of the non-zero coefficients and the correct recovery of sparsity patterns in the single-equation model. It therefore provides a fully data-driven way of selecting the relevant variables from a potentially large dataset of (co)integrated time series. Moreover, due to the flexible specification of the single-equation model, SPECS is able to take into account cointegration in the dataset without requiring any form of pre-testing for unit roots or testing for the cointegrating rank, and can thus be applied ``as is'' to any dataset containing an (unknown) mix of stationary and integrated time series. As a companion to this paper, an \textit{R} package is made available that implements a fast and easy-to-interpret algorithm for SPECS estimation, and provides immediate access to the dataset used in the empirical application.\footnote{\href{https://github.com/wijler/specs}{\textcolor{blue}{https://github.com/wijler/specs}}}.
Single-equation error correction models are frequently employed in tests for cointegration \citep[e.g.][]{Engle1987,Phillips1990b,Boswijk1994,Banerjee1998} as well as in forecasting applications \citep[e.g.][]{EngleY1987,Chou1996}, but require a weak exogeneity assumption for asymptotically efficient inference \citep{Johansen1992}. Weak exogeneity entails the existence of a single cointegrating vector that only appears in the marginal equation for the variable of interest. If this assumption holds, our procedure can be interpreted as an alternative to cointegration testing in the ECM framework \citep{Boswijk1994,Palm2010}. However, weak exogeneity may not be realistic in large datasets and we provide detailed illustrations of the implications of failure of this assumption and demonstrate that absent of weak exogeneity our procedure consistently estimates a linear combination of the true cointegrating vectors. While this impedes inference on the cointegrating relations, when the main aim of the model is nowcasting or forecasting, our procedure remains theoretically justifiable and provides empirical researchers with a simple and powerful tool for automated analysis of high-dimensional non-stationary datasets. In addition, for modeling a single variable of interest using a large set of potential regressors, SPECS provides a variable selection mechanism, allowing the researcher to discard variables that are irrelevant for this particular analysis. Our simulation results demonstrate strong selective capabilities in both low and high dimensions. Furthermore, a simulated nowcasting application highlights the importance of incorporating cointegration in the data as our proposed estimators obtain higher nowcast accuracies in comparison to a penalized autoregressive distributed lag (ADL) model. This finding is confirmed in an empirical application, where SPECS is employed to nowcast Dutch unemployment rates with the use of a dataset containing Google Trends series.
The use of penalized regression in time series analysis has gained in popularity, with a wide range of variants showing promising performance in applications \citep[see][for a recent overview]{Smeekes2018b}. Recent literature has also seen the development of methods for analyzing high-dimensional (co)integrated time series.
\citet{Kock2016} proposes the adaptive lasso to estimate an augmented Dickey-Fuller regression. While this univariate model is inherently different from ours, it provides an insightful demonstration of how the lasso may be used as an alternative to testing for non-stationarity, paralleling our suggestion to consider SPECS as an alternative for cointegration testing under the assumption of weak exogeneity.
For VECM systems, \citet{Wilms2016} propose a penalized maximum likelihood approach, with shrinkage performed on the cointegrating vectors, the coefficients regulating the short-run dynamics and the covariance matrix. While their method is shown to obtain forecast gains relative to the traditional Johansen method, no theoretical results are provided. \citet{Liao2015} provide an automated method of joint rank selection and parameter estimation with the use of an adaptive penalty and derive oracle properties in a fixed-dimensional framework. Next to this theoretical limitation on its applicability to large datasets, practical implementation is further complicated due to reliance on the eigenvalue decomposition of an asymmetric matrix, which introduces complex values into the corresponding objective function. As noted by \citet[][p. 424]{Liang2019}, this results in a non-standard harmonic function optimization problem. \citet{Liang2019} propose joint
parameter estimation and rank determination by employing a penalty that makes use of the $QR$-decomposition of the long-run coefficient matrix. This method possesses oracle-like properties under a high-dimensional asymptotic regime, but it requires the availability of an initial OLS estimator, thereby preventing applications on datasets in which the number of variables exceeds, or is close to, the number of available time series observations. Additionally, estimation of the long-run and short-run dynamics is performed sequentially rather than simultaneously, necessitating a two-step procedure.
In a single-equation setting, \citet{Lee2018} derive fixed-dimensional oracle properties for the adaptive lasso applied to predictive regressions where the regressors are allowed to be of mixed orders of integration. However, as a consequence of their model formulation in which all variables enter in levels, their estimator appears to be susceptible to spurious regression when the regressors are not cointegrated.
Finally, outside the penalized regression framework, \citet{Zhang2018b} propose an eigenvalue decomposition to estimate the cointegrating space in the presence of any integer and fractional order of integration of the variables. However, the estimation procedure proposed by \citeauthor{Zhang2018b} does not perform variable selection, nor does it provide explicit estimates of the transient dynamics in a VECM. \citet{Onatski2019} develop a novel inference procedure for the cointegrating rank in high dimensions. Similar to the Johansen procedure, their test is based on the squared canonical correlations, for which they derive the limit spectral distribution under joint asymptotics with the use of arguments from random matrix theory.
Our proposed method provides several contributions to this existing literature. First, our theoretical results are derived in a high-dimensional framework where the number of parameters is allowed to grow with the sample size. This requires non-standard theoretical results on bounds of the smallest eigenvalue of a matrix of (co)integrated regressors, similar to those in \citet{Zhang2018b}, which are further developed in this paper. Second, unlike many of the penalized regression methods surveyed above, the practical implementation of SPECS is straightforward for large datasets, including cases where the number of parameters is larger than the time dimension. Third, our method completely removes the need for pre-testing for the order of integration or cointegrating rank, and is not sensitive to spurious regression. Fourth, to the best of our knowledge, our paper is the first to explicitly allow for the presence of deterministic components in the theory, a crucial feature for many applications.
The paper is structured as follows. In Section \ref{sec:model} we discuss the data generating process. Section \ref{Sec:SPECS} describes the SPECS estimator. The main theoretical results of the paper are presented in Section \ref{Sec:Theory}. Section \ref{sec: simulations} contains several simulation studies, followed by an empirical application in Section \ref{Sec:application}. We conclude in Section \ref{Sec:Conclusion}. The main proofs and preliminary lemmas needed for them are contained in Appendix \ref{App:main}, while Appendix \ref{App:eigenvalues} contains results on minimum eigenvalue bounds. Finally, Appendix \ref{App:supp} contains supplementary material on proofs of preliminary lemmas and additional theorems, as well as further details on the empirical application.
A word on notation. For any an $N$-dimensional vector $\bm x$, $\norm{\bm x}_p = \left(\sum_{i=1}^N x_i^p\right)^{1/p}$ denotes the $\ell_p$-norm, while for any matrix $\bm D$ with $N$ columns, $\norm{\bm D}_p = \underset{\bm x \in \mathbb{R}^N}{\text{max}} \frac{\norm{\bm D x}_p}{\norm{\bm x}_p}$ is the corresponding induced norm and $\norm{\bm D}_F$ denotes the Frobenius norm. For an index set $S \subset \{1, \ldots, N\}$, let $\bm x_{S}$ be the vector containing the elements of $\bm x$ corresponding to $S$. Similarly, for a matrix $\bm D$ with $N$ rows, $\bm D_{S}$ is the sub-matrix containing the rows of $\bm D$ indexed by $S$. The orthogonal complement of $\bm D$ is denoted by $\bm D_\perp$, such that $\bm D_\perp^\prime \bm D = \bm{0}$. When $\bm D$ is a square matrix, we denote its $N$ ordered eigenvalues by $\lambda_1(\bm D) \geq \ldots \geq \lambda_N(\bm D)$ and we use $\bm D \succ 0$ to denote that the matrix is positive definite. We use $\bm \iota_N$ to denote a vector of ones of length $N$ and $\bm I_N$ to denote the $N$-dimensional identity matrix. We use $\overset{p}{\to}$ ($\overset{d}{\to}$) to denote convergence in probability (distribution) and $\overset{d}{=}$ denotes equivalence in distribution. Finally, we frequently make use of an arbitrary positive and finite constant $K$ whose value may change throughout the paper, but is always independent of the time and cross-sectional dimensions.
\section{The High-Dimensional Error Correction Model}\label{sec:model}
In this section we first discuss the data generating process for the vector time series along with the assumptions made. Next we transform the multivariate model to a single equation describing our variable of interest.
\subsection{Data Generating Process}
Assume one is interested in modelling a single variable of interest, say $y_t$, based on an $N$-dimensional time series $\bm z_t = (y_t,\bm x_t^\prime)$ observed at $t=1,\ldots,T$. Let $\bm z_t$ be described by
\begin{equation}\label{eq:z1}
\bm z_t = \bm \mu + \bm \tau t + \bm \zeta_t,
\end{equation}
with the stochastic component given by
\begin{equation}\label{eq:z2}
\begin{split}
\Delta\bm \zeta_t &= \bm A\bm B^\prime\bm \zeta_{t-1} + \sum_{j=1}^p \bm \varPhi_j \Delta \bm \zeta_{t-j} + \bm \epsilon_t,
\end{split}
\end{equation}
where $\bm A$ and $\bm B$ are $(N \times r)$-dimensional matrices containing the adjustment rates and cointegrating vectors, respectively. The innovations $\bm \epsilon_t = (\epsilon_{1,t},\bm \epsilon_{2,t}^\prime)^\prime$ satisfy the following assumptions:
\begin{assumption}\label{Ass:moments}
The sequence of innovations $\lbrace\bm \epsilon_t\rbrace_{t\geq 1}$ is an $N$-dimensional martingale difference sequence (m.d.s.) with $\E{(\bm \epsilon_t \bm \epsilon_t^\prime)}=\bm \varSigma_\epsilon$. Furthermore, we assume that
\begin{enumerate}[(1)]
\item There exists an $m > 2$, such that $\max_{1 \leq i \leq N, 1 \leq t \leq T} \E\abs{\epsilon_{i,t}}^{2m} \leq K_m$, and
\item There exist constants $\phi_\min,\phi_\max > 0$, such that $\phi_\min \leq \lambda_\min\left(\bm \varSigma_\epsilon\right) < \lambda_\max\left(\bm \varSigma_\epsilon\right) \leq \phi_\max$.
\end{enumerate}
\end{assumption}
This assumption implies that $\bm \epsilon_t$ is an martingale difference sequence with at least (a bit more than) four moments existing. The eigenvalue bounds in the second part place some restrictions on the dependence among the elements of $\bm \epsilon_t$, ruling out for instance a strong common factor affecting all errors. However, a wide range of contemporaneous dependence structures, such as spatial dependence, is still allowed.
The model can be rewritten into a VECM form by substituting \eqref{eq:z1} into \eqref{eq:z2} to obtain
\begin{equation}\label{eq:VECM}
\begin{split}
\Delta \bm z_t &= \bm A\bm B^\prime \left(\bm z_{t-1}-\bm \mu-\bm \tau (t-1)\right) + \tau^* + \sum_{j=1}^p \bm \varPhi_j\Delta \bm z_{t-j} + \bm \epsilon_t,\\
\end{split}
\end{equation}
where $\tau^* = (I-\sum_{j=1}^p\bm \varPhi_j)\bm \tau$. From this representation, it can directly be observed that the presence of a constant in \eqref{eq:z1} results in a constant within the cointegrating relationship if $\bm B^\prime \bm \mu \neq \bm{0}$. Furthermore, the linear trend in \eqref{eq:z1} appears as a constant in the differenced series and may additionally appear as a trend within the cointegrating vector if $\bm B^\prime \bm \tau \neq \bm{0}$, the latter implying that the equilibrium error $\bm B^\prime \bm z_t$ is a trend stationary process.
The following assumption asserts that the process is (at most) I(1), and the Granger Representation Theorem \citep[e.g.][p. 49]{Johansen1995} can be applied.
\begin{assumption}\label{Ass: GRT}
Define $\bm A(z):= (1-z)\bm I_N-\bm A\bm B^\prime z - \sum_{j=1}^p \bm \varPhi_j (1-z) z^j$.
\begin{enumerate}[(1)]
\item The determinantal equation $\abs{\bm A(z)}$ has all roots on or outside the unit circle.
\item $\bm A$ and $\bm B$ are $N \times r$ matrices with $1 \leq r \leq N$ and $\rank(\bm A) = \rank(\bm B) = r$.
\item The $\left((N-r) \times (N-r)\right)$ matrix $\bm A_\perp^\prime \left(I_N - \sum_{j=1}^p \bm \varPhi_j\right)\bm B_\perp$ is invertible.
\end{enumerate}
\end{assumption}
Assumption \ref{Ass: GRT} enables \eqref{eq:VECM} to be written as a vector moving average (VMA) process
\begin{equation}\label{eq:GRT}
\bm z_t = \bm C\bm s_t + \bm \mu + \bm \tau t + \bm C(L)\bm \epsilon_t + \bm C\bm z_0,
\end{equation}
where $\bm C=\bm B_\perp\left(\bm A_\perp^\prime \left(\bm I_N - \sum_{j=1}^p\bm \varPhi_j\right) \bm B_\perp\right)^{-1}\bm A_\perp^\prime$, $\bm s_t = \sum_{s=1}^t \bm \epsilon_s$, $\bm C(L)\bm \epsilon_t$ is a stationary linear process and $\bm z_0$ are initial values. Without loss of generality, we assume henceforth that $\bm z_0 = \bm{0}$.
We need a further restriction on the dependence in the VMA representation in the form of the following assumption, which ensures norm-summability of the coefficients in the Beveridge-Nelson decomposition.
\begin{assumption}\label{Ass:Dependence}
There exists a $K<\infty$ such that $\bm C$ in \eqref{eq:GRT} satisfies $\norm{\bm C}_\infty \leq K$. In addition, the matrix lag polynomial $\bm C(L)$ is given by $\bm C(z) = \sum_{l=0}^\infty \bm C_lz^l$ and satisfies $\sum_{l=0}^\infty l\norm{\bm C_l}_\infty \leq K$.
\end{assumption}
\subsection{Single-Equation Representation}\label{sec:ser}
The number of parameters to estimate in \eqref{eq:VECM} is at least $2Nr + N^2p$, such that the system quickly grows too large to accurately estimate based on traditional methods. As we assume a single variable $y_t$ is of interest, we therefore instead consider the lighter parameterized single-equation model for $y_t$. To ensure that the variables modelling the variation in $y_t$ remain exogenous, we orthogonalize the errors driving the single-equation model, say $\epsilon_{y,t}$, from the errors driving the marginal equations of the endogenous variables $\bm x_t$. This is achieved by decomposing $\epsilon_{1,t}$ into its best linear prediction based on $\bm \epsilon_{2,t}$ and the corresponding orthogonal prediction error. To this end, partition the covariance matrix of $\bm \epsilon_t$ as
\begin{equation}\label{eq:covariance partition}
\bm \varSigma_\epsilon = \begin{bmatrix}
\E(\epsilon_{1,t})^2 & \E(\epsilon_{1,t}\bm \epsilon_{2,t}^\prime) \\
\E(\epsilon_{1,t}\bm \epsilon_{2,t}) & \E(\bm \epsilon_{2,t}\bm \epsilon_{2,t}^\prime) \\
\end{bmatrix} = \begin{bmatrix}
\sigma_{11} & \bm \sigma_{21}^\prime \\
\bm \sigma_{21} & \bm \varSigma_{22} \\
\end{bmatrix},
\end{equation}
such that we obtain
\begin{equation}\label{eq:error decomposition}
\begin{split}
\epsilon_{1,t} &= (0,\bm \sigma_{21}^\prime\bm \varSigma_{22}^{-1}) \bm \epsilon_t + \left(1, - \bm \sigma_{21}^\prime\bm \varSigma_{22}^{-1}\right)\bm \epsilon_t = \hat{\epsilon}_{1,t} + \epsilon_{y,t}.
\end{split}
\end{equation}
Define $\pi_0 = \bm \varSigma_{22}^{-1}\bm \sigma_{21}$. Then, writing out \eqref{eq:error decomposition} in terms of the observable time series results in the single-equation model
\begin{equation}\label{eq:CECM}
\begin{split}
\Delta \bm y_t &= \left(1,-\bm \pi_0^\prime \right)\left(\bm A\bm B^\prime (\bm z_{t-1} - \bm \mu - \bm \tau(t-1)) + \bm \tau^* + \sum_{j=1}^p \bm \varPhi_j^\prime \Delta \bm z_{t-j} \right) + \bm \pi_0^\prime \Delta \bm x_t + \epsilon_{y,t}\\
&= \bm \delta^\prime \bm z_{t-1} + \bm \pi^\prime \bm w_t + \mu_0 + \tau_0(t-1) + \epsilon_{y,t},
\end{split}
\end{equation}
where $\bm \delta^\prime = \left(1,-\bm \pi_0^\prime \right)\bm A\bm B^\prime$, $\bm \pi = (\bm \pi_0^\prime,\ldots,\bm \pi_p^\prime)^\prime$ with $\bm \pi_j^\prime = (1,-\bm \pi_0^\prime)\bm \varPhi_j$ for $j=1,\ldots,p$, $\mu_0 = (1,-\bm \pi_0^\prime)\left(\bm A\bm B^\prime\bm \mu + \bm \tau^*\right)$ and $\tau_0 = (1,-\bm \pi_0^\prime)\bm \tau^*$. Note that $\bm \delta$ is a vector of length $N$, whereas $\bm \pi$ is a vector of length $M = N(p+1)-1$. Additionally, $\bm w_t=(\Delta \bm x_t^\prime,\Delta \bm z_{t-1}^\prime,\ldots,\Delta \bm z_{t-p}^\prime)^\prime$ and $\epsilon_{y,t} = (1 - \bm \pi_0^\prime)\bm \epsilon_t $. Finally, we write the single-equation model in matrix notation as
\begin{equation}\label{eq:CECM_matrix}
\begin{split}
\Delta \bm y &= \bm Z_{-1}\delta + \bm W\pi + \bm \iota_T \mu_0 + \bm t\tau_0 + \bm \epsilon_y = \bm V\bm \gamma + \bm D\bm \theta + \bm \epsilon_y,
\end{split}
\end{equation}
where $\bm Z_{-1} = (\bm z_0,\ldots,\bm z_{T-1})^\prime$, $\bm W = (\bm w_t,\ldots,\bm w_T)^\prime$, $\bm t = (0,\ldots,T-1)^\prime$, $\bm V=(\bm Z_{-1},\bm W)$, $\bm D = (\bm \iota_T,\bm t)$, $\bm \gamma = (\bm \delta^\prime,\bm \pi^\prime)^\prime$ and $\bm \theta = (\mu_0,\tau_0)^\prime$.
\begin{remark}
The single-equation model may similarly be derived under the assumption of normal errors. In this framework, $\epsilon_{y,t}$ has the conditional normal distribution from which \eqref{eq:CECM} can be obtained \citep[cf.][]{Boswijk1994}. A benefit of assuming normality is that, under the additional assumption of weak exogeneity, the OLS estimates based on \eqref{eq:CECM} are optimal in the mean-squared sense. However, the assumption of normality is unnecessarily restrictive when the, perhaps overly, ambitious goal of complete and correct specification is abandoned.
\end{remark}
\begin{remark}
An additional benefit of the conditional error-correction model, as opposed to the predictive regressions specified in levels considered in \citet{Lee2018}, is that the former avoids spurious regression. In the case where all variables in $\bm z_t$ are integrated of order one and independent of one another, the left-hand side of \eqref{eq:CECM_matrix} would remain stationary. Intuitively, any ``best fitting'' linear combination between the stationary component $\Delta y_t$ and $(\bm z_{t-1}^\prime,\bm w_t^\prime)^\prime$ would seek to minimize the contribution of the variables in $\bm z_t$, as their stochastically trending nature substantially inflates the fitting error. This behaviour is well-documented for the fixed-dimensional OLS estimator -- cf. \citet[][A.9]{Boswijk1994} in which $\hat{\bm \delta}_{OLS}$ turns out to be superconsistent -- and carries over to SPECS in high-dimensions.
\end{remark}
In general, the implied cointegrating vector $\bm \delta$ in the single-equation model for $y_t$ contains a linear combination of the cointegrating vectors in $\bm B$ with their weights being given by $\left(1,-\bm \pi_0^\prime \right)\bm A$. Since the marginal equations of $\bm x_t$ contain information about the cointegrating relationship, efficient estimation within the single-equation model is only attained under an assumption of weak exogeneity. \citet{Johansen1992} shows that sufficient conditions for weak exogeneity to hold are (i) normality of $\bm \epsilon_t$, (ii) $\rank(\bm A \bm B^\prime) = 1$, i.e. there is a single cointegrating $N$-dimensional cointegrating vector $\bm \beta$, and (iii) the vector of adjustment rates takes on the form $\bm \alpha = (\alpha_1,\bm{0}^\prime)^\prime$. However, these conditions are rather restrictive when considering high-dimensional economic datasets that are likely to possess multiple cointegrating relationships and complex covariance structures across the errors. Therefore, we opt to derive our results without assuming weak exogeneity, while acknowledging that direct interpretation of the estimated cointegrating vector will only be valid in the presence of weak exogeneity. Furthermore, we believe that whether the potential loss of asymptotic efficiency in our more parsimonious single-equation model translates to inferior performance in finite samples ultimately remains an empirical question.
As we consider sparse estimation of this single-equation model, let us briefly touch upon the required sparsity. For measuring the sparsity, we work directly in the single-equation representation.\footnote{In absence of weak exogeneity, it may not be directly obvious how we obtain a sparse single-equation model from the VECM. We therefore provide a more detailed discussion of the interpretation of sparsity absent of weak exogeneity in Section \ref{Sec:Sparsity}. In this section we just take the single-equation model directly as starting point.} Let $S_\delta = \lbrace i \vert \delta_i \neq 0\rbrace$ denote the index set of the non-zero elements in $\bm \delta$, with its cardinality denoted by $ \abs{S_\delta}$, and let $S_{\pi}$ be defined accordingly for $\bm \pi$. In addition, let $r^*$ denote the dimension of the cointegration space of $\bm z_{S_\delta,t}$, i.e. the number of independent linear stationary combinations of $\bm z_{S_\delta,t}$ (cf. Remark \ref{Rem:r}), and define $s_\delta = \abs{S_\delta} - r^*$ and $s_\pi = \abs{S_\pi}+r^*$ as the number of ``effective'' relevant non-stationary and stationary variables, respectively. Our estimation goal will then be to obtain estimates of $S_\delta$ and $S_{\pi}$, as well as estimate $\bm \delta_{S_\delta}$ and $\bm \pi_{S_\pi}$. To obtain consistency, we need the following assumptions on the amount of sparsity.
\begin{assumption}\label{ass:sparsity}
Assume that
(1) $s_\delta = o(T^{1/4})$; (2) $s_\pi = o(\sqrt{T})$ and (3) $\max\{s_\delta, \sqrt{s_\pi}\} = o(\gamma_\min \sqrt{T})$, where $\gamma_{\min} = \min\{\abs{\gamma_i}: \gamma_i \neq 0\}$.
\end{assumption}
Parts (1) and (2) put restrictions on how fast the number of relevant parameters is allowed to grow. The ``effective'' number of relevant stationary variables ($s_\pi$) is allowed to grow faster than the ``effective'' number of integrated variables ($s_\delta$), as a result of the collinearity induced by the stochastic trends (cf. Remark \ref{Rem:min_eig_zero}). Part (3) puts an additional restriction on the number of relevant coefficients as a function of the smallest non-zero coefficient. Clearly, if all coefficients are assumed to be fixed, (3) is not binding. In fact, one can allow $\gamma_\min$ to shrink at a rate up to $T^{-1/4}$ before it becomes binding. This assumption may therefore be interpreted as determining the fastest rate at which the population coefficients are allowed to decrease, as a function of $T$, $s_\delta$ and $s_\pi$, to still ensure it can be consistently picked up by our estimation method.
\subsection{Rotations and Bounds on Eigenvalues} \label{sec:rot}
Bounds on eigenvalues play a crucial role in establishing consistency properties of lasso-type penalized regression methods. However, due the mixed integrated nature of our data, where parts of the regressors are stationary, and other parts are only stationary after rotation, the object of our assumptions is not the sample covariance matrix directly, but instead a carefully transformed version. Under Assumptions \ref{Ass:moments}-\ref{Ass:Dependence}, it is then possible to ensure eigenvalue conditions on the sample covariance matrices. Before we can state the assumption, we must therefore establish some further notation and rotations to be used later.
Let $\bm \gamma = (\bm \delta^\prime, \bm \pi^\prime)^\prime$ and $S_{\gamma}$ its active set. Without loss of generality, we partition the data matrix as $\bm V=(\bm V_{S_\gamma},\bm V_{S_\gamma^c})$, with $\bm V_{S_\gamma} = (\bm Z_{-1,S_\delta},\bm W_{S_\pi})$ representing the time series carrying non-zero coefficients in the population single-equation model, henceforth referred to as the set of relevant variables. In the presence of cointegration, it follows from \eqref{eq:GRT} that the relevant lagged levels can be written as
\begin{equation}\label{eq:GRT_relevant}
\begin{split}
\bm z_{S_\delta,t} &= \bm C_{S_\delta}\bm s_t + \bm \mu_{S_\delta} + \bm \tau_{S_\delta}t + \bm u_{S_\delta,t},\quad
\bm C_{S_\delta} = \bm B_{\perp,S_\delta}\left(\bm A_\perp^\prime \left(\bm I_N - \sum_{j=1}^p\bm \varPhi_j\right) \bm B_\perp\right)^{-1}\bm A_\perp^\prime
\end{split}
\end{equation}
where $\bm B_{\perp,S_\delta}$ is an $(\abs{S_\delta} \times (N-r))$-dimensional matrix containing the rows of $\bm B_\perp$ indexed by $S_\delta$ and $\bm u_{S_\delta,t} = \bm C_{S_\delta}(L)\bm \epsilon_t$. The left null space of $\bm B_{\perp,S_\delta}$, defined as $\bm B^* = \left\lbrace \bm x \in \mathbb{R}^{\abs{S_\delta}} \vert \ \bm B_{\perp,S_\delta}^\prime \bm x = \bm{0}\right\rbrace$, contains the linear combinations that convert $\bm z_{S_\delta,t}$ to a stationary process. Accordingly, we also refer to this null space as the cointegrating space of $\bm z_{S_\delta,t}$. By construction, $\bm \delta_{S_\delta} \in \bm B^*$, such that this cointegrating space is non-empty whenever $\bm \delta \neq \bm{0}$. In this case, we define $\bm B_{S_\delta}$ as a $(\abs{S_\delta} \times r^*)$-dimensional basis matrix of $\bm B^*$, with $r^* \leq \abs{S_\delta}$ representing the dimension of the cointegrating space.\footnote{The matrix $\bm B_{S_\delta}$ is not uniquely defined. However, in most instances, including those contained in the current work, identification of the span of $\bm B_{S_\delta}$ is sufficient.}
Similarly, we define $\bm B_{S_\delta,\perp}$ as a basis matrix of the left null-space of $\bm B_{S_\delta}$, i.e. a $\left(\abs{S_\delta}\times(\abs{S_\delta}-r^*)\right)$-dimensional matrix of full column rank with the property that $\bm B_{S_\delta,\perp}^\prime \bm B_{S_\delta} = \bm{0}$. Then, we are able to define a $\bm Q$-transformation that decomposes the reduced system into a stationary and non-stationary contribution as
\begin{equation}\label{eq:Q}
\begin{split}
\bm Q &= \begin{bmatrix}
\bm B_{S_\delta}^\prime & \bm{0}\\
\bm{0} & \bm I_{\abs{S_\pi}}\\
\bm B_{S_\delta,\perp}^\prime & \bm{0}
\end{bmatrix}, \quad \bm Q^{-1} = \begin{bmatrix}
\bm B_{S_\delta}\left(\bm B_{S_\delta}^\prime\bm B_{S_\delta}\right)^{-1} & \bm{0} & \bm B_{S_\delta,\perp}\left(\bm B_{S_\delta,\perp}^\prime \bm B_{S_\delta,\perp}\right)^{-1}\\
\bm{0} & \bm I_{\abs{S_\pi}} & 0
\end{bmatrix}.
\end{split}
\end{equation}
For the case $\bm \delta = \bm{0}$, we define $\bm Q = \bm I_{\abs{S_\pi}}$. Post-multiplication of the data matrix by $\bm Q^\prime$ gives
\begin{equation}\label{eq:Q-transformed}
\bm V_{S_\gamma}\bm Q^\prime = \begin{bmatrix}
\bm Z_{-1,S_\delta}\bm B_{S_\delta} & \bm W_{S_\pi} & \bm Z_{-1,S_\delta}\bm B_{S_\delta,\perp}
\end{bmatrix}
\end{equation}
which we refer to as the $\bm Q$-transformed version of $\bm V_{S_\gamma}$. The first $s_\pi = \abs{S_\pi} + r^*$ columns of \eqref{eq:Q-transformed}, corresponding to $(\bm Z_{-1,S_\delta}\bm B_{S_\delta},\bm W_{S_\pi})$, contain independent stationary linear combinations of the variables that are relevant to $\Delta y_t$ in the single-equation model. The remaining $s_\delta = \abs{S_\delta}-r^*$ columns, given by $\bm Z_{-1,S_\delta}\bm B_{S_\delta,\perp}$, contain all linearly independent combinations that are integrated of order one.
\begin{remark}\label{Rem:r}
We may interpret $r^*$ as the ``effective'' cointegration rank, where ``effective'' relates to variable of interest $y_t$. Essentially, we remove all variables not relevant to $y_t$ in the long-run ($S_\delta^c$) and then reconstruct a VECM from the remaining variables, which now has rank $r^*$.\end{remark}
Finally, we construct a transformed version of the sample covariance matrix based on $\bm V_{S_\gamma}$, which plays a crucial role in the development of our theory. First, to regress out the deterministic components of the observed time series in \eqref{eq:VECM}, we define the matrix $\bm M = \bm I_T - \bm D\left(\bm D^\prime\bm D\right)^{-1}\bm D^\prime$.\footnote{Note that $\bm D$ may vary depending on the deterministic specification of the model; setting $\bm D = (\bm \iota_T,\bm t)$ allows for both a non-zero constant and linear trend, while simply setting $\bm M = \bm I_T$ may be desired (although not required) when it is believed that $\bm \mu=\bm \tau=\bm{0}$.} Then, after rotating by $\bm Q$ and regressing out the deterministic components by $\bm M$, the stationary and non-stationary components are scaled via the matrix $\bm S_T = \diag(\sqrt{T}\bm I_{s_\pi},\frac{T}{\sqrt{s_\delta}}\bm I_{s_\delta})$. Hence, our transformed sample covariance matrix is defined as
\begin{align}
\label{eq:Sigma_hat}
&\hat{\bm \varSigma} = \bm S_T^{-1}\bm Q\bm V_{S_\gamma}^\prime\bm M\bm V_{S_\gamma}\bm Q^\prime\bm S_T^{-1} = \begin{bmatrix}
\hat{\bm \varSigma}_{11} & \hat{\bm \varSigma}_{12}\\
\hat{\bm \varSigma}_{21} & \hat{\bm \varSigma}_{22}
\end{bmatrix},\\
&\text{with} \quad
\label{eq:Sigma_11}
\hat{\bm \varSigma}_{11} = \frac{1}{T}\begin{bmatrix}
\bm B_{S_\delta}^\prime\bm Z_{-1,S_\delta}^\prime \bm M \bm Z_{-1,S_\delta}\bm B_{S_\delta} & \bm B_{S_\delta}^\prime\bm Z_{-1,S_\delta}^\prime \bm M \bm W_{S_\pi}\\
\bm W_{S_\pi}^\prime \bm M \bm Z_{-1,S_\delta}\bm B_{S_\delta} & \bm W_{S_\pi}^\prime \bm M \bm W_{S_\pi}
\end{bmatrix}
\end{align}
and $\hat{\bm \varSigma}_{22} = \frac{s_\delta}{T^2}\bm B_{S_\delta,\perp}^\prime\bm Z_{-1,S_\delta}^\prime\bm M\bm Z_{-1,S_\delta}\bm B_{S_\delta,\perp}$.
We can now state the eigenvalue assumptions.
\begin{assumption}\label{Ass:eigenvalues}
Assume that, on a set with probability converging to 1 as $T,N,p \to \infty$, there exists a constant $\phi>0$, such that
$\underset{\bm x \in \mathbb{R}^{s_\pi}}{\text{inf}} \frac{\bm x^\prime \hat{\bm \varSigma}_{11}\bm x}{\bm x^\prime \bm x} \geq \phi$
an
$\underset{\bm x \in \mathbb{R}^{s_\delta}}{\text{inf}} \frac{\bm x^\prime \hat{\bm \varSigma}_{22} \bm x}{\bm x^\prime \bm x} \geq \phi$.
\end{assumption}
The first part of Assumption \ref{Ass:eigenvalues} applies to stationary data and is known to hold when the minimum eigenvalue of the corresponding population covariance matrix is bounded away from zero \citep[e.g.][Section B.2]{Medeiros2016}. The second part, however, applies to integrated variables and requires arguments that are unique to the non-stationary setting. In particular, we note the necessity of applying a scaling by $\frac{s_\delta}{T^2}$, rather than the usual $\frac{1}{T^2}$ one may expect from the fixed-dimensional literature, cf. Remark \ref{Rem:min_eig_zero}. In Appendix \ref{App:eigenvalues}, we show several cases under which Assumption \ref{Ass:eigenvalues} is satisfied.
\begin{remark}\label{Rem:min_eig_zero}
As an illustration of the problems with adopting the usual scaling by $T^{-2}$, consider the simple example of an $s$-dimensional white noise sequence $\bm u_t \overset{i.i.d.}{\sim} \mathcal{N}(\bm{0},\bm I_s)$ and define $\bm h_t = \sum_{j=1}^t \bm u_j$. Then, in Lemma \ref{Lemma:BM_int} in Appendix \ref{App:eigenvalues} we show that $\Prob\left(\lambda_\min\left(\frac{1}{T^2}\sum_{t=1}^T\bm h_t\bm h_t^\prime\right) > \phi \right) \to 0$, as $s,T \to \infty$, regardless of their relative rates. Hence, even in this simple case we cannot assume that the minimum eigenvalue is bounded away from zero if we stick to the $T^{-2}$ scaling.
\end{remark}
\begin{remark}
There are several noteworthy instances in which $\lambda_\min\left(\hat{\bm \varSigma}_{22}\right)$ is bounded away from zero with arbitrarily high probability without the need for Assumption \ref{Ass:eigenvalues}. Assume that the dimension of the orthogonal complement of the cointegrating space in the subset of relevant non-stationary variables converges to a finite constant, i.e. $s_\delta \to K$. Then, based on a standard functional central limit theorem,
\begin{equation*}
\hat{\bm \varSigma}_{22} \overset{d}{\to} K\bm B_{S_\delta,\perp}^\prime\bm C_{S_\delta}\left(\int_0^1 \tilde{\bm B}(r)\tilde{\bm B}^\prime(r)dr\right)\bm C_{S_\delta}^\prime\bm B_{S_\delta,\perp} \overset{d}{=} \int_0^1 \bm B^*(r)\bm B^{*\prime}(r)dr,
\end{equation*}
where $\tilde{\bm B}(r)$ is an $s_\delta$-dimensional Gaussian process, described in the proof of Lemma A.2 in \citet{Phillips1990a}, and $\bm B^*(r)$ is simply a linearly transformed version. By the same lemma, it follows that $\int_0^1 \bm B^*(r)\bm B^{*\prime}(r)dr$ is positive-definite almost surely. Then, for any $\epsilon>0$, we may choose $\phi(\epsilon) > 0$ such that
\begin{equation*}
\Prob\left(\lambda_{\min}\left(\hat{\bm \varSigma}_{22}\right) \leq \phi(\epsilon)\right) \to \Prob\left(\lambda_\min\left(\int_0^1 \bm B^*(r)\bm B^{*\prime}(r)dr\right) \leq \phi(\epsilon)\right) \leq \epsilon.
\end{equation*}
A straightforward case in which $s_\delta$ remains finite is to simply assume that the number of relevant integrated variables stays finite, i.e. $\abs{S_\delta} \leq K$. However, a more general example occurs when the dimension of the cointegrating space of $\bm z_{S_\delta,t}$ diverges at the rate $\abs{S_\delta}$. This occurs in the case of a non-stationary factor model with stationary idiosyncratic components, as proposed by \citet{Banerjee2014}. Further illustrations are provided in Remark \ref{Rem:factor}.
\end{remark}
\section{The Single-Equation Penalized Error Correction Selector}\label{Sec:SPECS}
Despite the dimension reduction obtained from moving towards a single-equation representation, regularization remains a necessity in high dimensions. The single-equation model \eqref{eq:CECM} contains a total of $N(p+2) + 1$ parameters, compared to the $2N(r+1) + N^2p$ parameters in the full-system VECM in \eqref{eq:VECM}, resulting in a substantial reduction in dimensionality. However, the dimension may still grow large when either: (1) the number of potentially relevant variables is large or (ii) when the number of lagged differences required to appropriately model the short-run dynamics is large. Therefore, we consider the use of $\ell_1$-regularization to enable estimation in high dimensions.
The resulting estimator, henceforth referred to as the Single-equation Penalized Error Correction Selector (SPECS), is defined as the minimizer of the following objective function:
\begin{equation}\label{eq:SPECS}
G_T\left(\bm \gamma,\bm \theta\right) = \norm{\Delta \bm y - \bm V\bm \gamma - \bm D\bm \theta}_2^2 + \lambda_I\sum_{i=1}^{N+M} \omega_i\abs{\gamma_i} + \lambda_G\norm{\bm \delta}_2,
\end{equation}
where $M = (N+1)p-1$ refers to the number of transformed variables in $\bm w_t$, i.e. the length of $\bm \pi$. We denote the minimizers of \eqref{eq:SPECS} by $\hat{\bm \gamma}$. The group penalty, regulated by $\lambda_G$, serves to promote exclusion of the lagged levels as a group when there is no cointegration present in the data. In this case, the model is effectively estimated in differences and corresponds to a conditional model derived from a vector autoregressive model specified in differences. The individual $\ell_1$-penalties, regulated by $\lambda_I$, serve to enforce sparsity in the coefficient vectors $\bm \delta$ and $\bm \pi$ respectively.
The penalty of each coefficient $\gamma_i$ is weighted by $\omega_i$ to enable simultaneous estimation and selection consistency of the coefficients. Therefore, SPECS resembles a sparse group lasso \citep[e.g.][]{Simon2013} with adaptive weighting, applied to the conditional error correction model. The weights $\omega_i$ in \eqref{eq:SPECS} are typically derived from an initial estimation procedure such as OLS (if the number of variables is small enough), ridge, or lasso. In particular, let $\hat{\bm \gamma}_I$ denote initial estimates obtained for $\bm \gamma$ using one of the aforementioned methods. The weights can then be constructed as $\omega_i = \abs{\hat\gamma_{I,i}}^{-k}$ for some $k>0$. As the coefficients of the irrelevant variables tend to zero, this will ``blow up'' the weights for these coefficients, making them unlikely to be selected in the final estimation. On the other hand, the weights of the relevant coefficients converge to a positive constant leaving them unaffected. This wedge between the weights of relevant and irrelevant coefficients is exactly needed to achieve selection consistency. As demonstrated by \citet{Zou2006a}, under such assumptions on the weights, the adaptive lasso attains simultaneous selection and estimation consistency, without the necessity for the rather stringent irrepresentable condition in \citet{Zhao2006}.\footnote{In fact, as the adaptive lasso can be written as a regular lasso on a transformed design matrix, the irrepresentable condition, while still needed, operates on this transformed design matrix and becomes a weighted irrepresentable condition. This condition is then in turn implied by appropriate assumptions on the weights. In this paper we directly take this route rather than going via an irrepresentable condition. Section 7.5 of \citet{Buhlmann2011} provides details on the links between these assumptions.} To maintain generality we work with general weights without specifying how they are obtained, and therefore define appropriate assumptions directly on the weights. In Section \ref{Sec:Init_est} we then return to weight construction and propose a feasible way to construct weights that are theoretically shown to satisfy our assumptions.
\begin{assumption}\label{Ass:Regularization} Assume that the weights and regularization penalties satisfy:
\begin{enumerate}
\item\label{ass:reg_om_max} $\omega_{S_\gamma,\max} = o_p(T^\xi$) for some $\xi > 0$, where $\omega_{S, \max} = \max\{\omega_i: i \in S\}$.
\item\label{ass:reg_lambda} $\lambda_I = o\left(\frac{\left(s_\delta + \sqrt{s_\pi}\right)T^{1/2-\xi}} {\sqrt{s_\delta + s_\pi}} \right)$ and $\lambda_G = o(\sqrt{T})$.
\item\label{ass:reg_om_min} Let $\omega_{S, \min} = \min\{\omega_i: i \in S\}$. Then
\begin{align*}
\omega_{S_\delta^c,\min}^{-1} &= o_p \left(\min\left\{(s_\delta + s_\pi)^{-1/2} T^{-1/2 - \xi} N^{-1/2}, \lambda_I (s_\delta + \sqrt{s_\pi})^{-1} T^{-1} N^{-1/2} \right\} \right) ,\\
\omega_{S_\pi^c,\min}^{-1} &= o_p \left(\min\left\{(s_\delta + s_\pi)^{-1/2} T^{-\xi} (Np)^{-1/2}, \lambda_I (s_\delta + \sqrt{s_\pi})^{-1} (TNp)^{-1/2} \right\} \right).
\end{align*}
\end{enumerate}
\end{assumption}
Part (\ref{ass:reg_om_max}) puts an upper bound on the rate at which the weights corresponding to the relevant variables diverge. Part (\ref{ass:reg_lambda}) restricts the maximum admissible growth rate of the penalty. Exceeding this rate would in an excess of shrinkage bias that impedes estimation consistency. Finally, part (\ref{ass:reg_om_min}) states that the weights of the irrelevant variables -- interacting with the penalty parameter $\lambda_I$ -- grow sufficiently fast in order to guarantee that irrelevant variables are removed from the model with probability converging to one. The required minimum growth rate of the penalty parameter is inversely related to the growth rate of the weights of the irrelevant variables; faster diverging weights require less penalization to identify irrelevant variables.
\begin{remark} \label{rem:group}
The only restriction that Assumption \ref{Ass:Regularization} imposes on the growth rate of the group penalty is that $\frac{\lambda_{G} }{\sqrt{T}} \to 0$, which is necessary for preventing shrinkage bias induced by the group penalty from impeding estimation consistency. Since $\lambda_{G} = 0$ is an admissible value, it follows that the theoretical results presented in the following section apply to the minimizer of
$G^*_T(\bm \gamma,\bm \theta) = \norm{\Delta \bm y - \bm V\bm \gamma - \bm D\bm \theta}_2^2 + \lambda_I\sum_{i=1}^{N+M}\omega_i\abs{\gamma_i}$
as well, as long as the remaining conditions are satisfied.
\end{remark}
\begin{remark} \label{rem:dets}
Note that the deterministic components $\bm \theta$ are left unpenalized in \eqref{eq:SPECS}, as their inclusion in the model is desirable to enable identification of the limiting distribution of the estimators. Similar to the classical Frisch-Wraugh-Lovell Theorem, \citet{Yamada2017} show that the inclusion of unpenalized components is equivalent to performing the estimation after regressing out those components. In other words, we may define $\bm M = \bm I_T - \bm D\left(\bm D^\prime\bm D\right)^{-1}\bm D^\prime$ and note that
\begin{equation*}
\hat{\bm \gamma} = \argmin_{\bm \gamma} \norm{\bm M\left(\Delta \bm y - \bm V\bm \gamma\right)}_2^2 + \lambda_I\sum_{i=1}^{N+M} \omega_i\abs{\gamma_i} + \lambda_G\norm{\bm \delta}_2.
\end{equation*}
If one believes that the trend or constant are zero, one may reflect this knowledge in the construction of $\bm M$, with the convention that $\bm M=\bm I_T$ when $\bm \mu=\bm \tau=\bm{0}$.
\end{remark}
Two common data-driven ways to select the tuning parameters $\lambda_I$ and $\lambda_G$ are using cross-validation and information criteria. As standard $K$-fold cross-validation does not respect the time order of the data, we instead consider a time series cross-validation (TSCV) scheme as proposed by e.g.~\citet{Hyndman2018} and \citet{Wilms2017}, where for different values of $\bm \lambda = (\lambda_I, \lambda_G)^\prime$ the model is estimated on the first part of the sample, and its prediction for the next observation is recorded. The sample is then recursively moved forward towards the end, and the $\bm \lambda$ with the lowest mean squared prediction error is selected. We refer to \citet{Smeekes2018a} for details on the implementation and a comparison with traditional $K$-fold cross-validation.
While cross-validation works well for prediction \citep{chetverikov2016}, it tends to generally select fairly low penalty levels and therefore includes many variables. An alternative way to select $\bm \lambda$ is using information criteria, where we find the value of $\bm \lambda$ as
\begin{equation*}
\hat{\bm \lambda}_{IC} = \argmin_{\bm \lambda} \ln \left(\frac{1}{T} \norm{\Delta \bm y - \bm V \hat{\bm \gamma}(\bm \lambda) - \bm D \hat\theta}_2^2 \right) + \frac{C_T \widehat{df}(\bm \lambda)}{T},
\end{equation*}
where $\hat{\bm \gamma}(\bm \lambda)$ and $\hat{\bm \theta}$ denote the minimizers of $G_T\left(\bm \gamma,\bm \theta\right)$ in \eqref{eq:SPECS} for a particular value of $\bm \lambda$.\footnote{As explained in Remark \ref{rem:dets}, $\hat{\bm \theta}$ does not depend on $\bm \lambda$.} In addition, $\widehat{df}(\bm \lambda)$ is an estimate of the degrees of freedom and $C_T$ is the criterion-specific penalty; for the latter we use the Bayesian Information Criterion \citep[BIC]{Schwarz1978} with $C_T= \ln(T)$.
\citet{Zou2007} show that for the (adaptive) lasso the number of non-zero coefficients is an appropriate estimate for the degrees of freedom for model selection using information criteria. For group lasso penalties, estimating the degrees of freedom is more complicated. \citet{Yuan2006} propose a heuristic rule, but this requires the least squares estimator which is not available for large $N$. Alternative rules are provided by \citet{Breheny2009} and \citet{vaiter2012} among others, but none are theoretically valid in our setting. For this reason we propose a simple, heuristic rule where we set $\widehat{df}(\bm \lambda)$ equal to the number of non-zero coefficients. Essentially this means we ignore the strength of the group penalty on the complexity of the model as long as the group is selected, thereby overestimating $df(\bm \lambda)$. As a consequence, we will only choose non-zero values of $\lambda_G$ if they either improve the fit directly or result in setting the whole group to zero without affecting the fit too much. This is an intentional choice, consistent with our theoretical treatment of the group penalty. As discussed in Remark \ref{rem:group}, the group penalty is not necessary and consistency can be achieved even with $\lambda_G = 0$, and can therefore be seen as an optional add-on penalty.
Finally, we note that in practice both methods require the respective objective function to be minimized for a two-dimensional grid of values for $\bm \lambda$. By choosing the lower and upper bounds of the grid carefully, one can ensure that the selected tuning parameters satisfy the assumptions listed in the next subsection. Of course, even though this ensures the theoretical validity of the selection method, its practical performance can still vary considerably. Therefore we investigate the practical performance of BIC and TSCV in the simulations and empirical application respectively.
\section{Theoretical Results}\label{Sec:Theory}
In this section we derive the asymptotic properties of SPECS, describe the construction of the weights and discuss implications for particular model specifications.
\subsection{Asymptotic Properties}\label{Sec:Asymptotics}
The first result that we pursue is that of selection consistency, i.e. the ability of an estimation procedure to select the correct set of relevant variables with probability converging to one. In fact, \citet{Zhao2006} define a stronger property referred to as sign consistency, which additionally requires the procedure to identify the correct signs of the non-zero coefficients with probability converging to one. In the following theorem, we derive sign consistency of SPECS.
\begin{theorem}\label{Thm:Selection_Consistency}
Under Assumptions \ref{Ass:moments}-\ref{Ass:Regularization}, as $T,N,p, \to \infty$ it holds that
$\Prob\left(\emph{sign}\left(\hat{\bm \gamma}\right) = \emph{sign}\left(\bm \gamma\right)\right) \to 1$.
\end{theorem}
Theorem \ref{Thm:Selection_Consistency} provides an asymptotic justification for implementing SPECS as a high-dimensional variable selection device. Furthermore, selection consistency is a crucial property when one aims to obtain interpretable solutions or even utilize the estimator as an alternative to classical tests for cointegration. An example of a traditional test for cointegration is the ECM-test by \citet{Banerjee1998} which looks at the $t$-ratio of the ordinary least squares coefficient of the lagged dependent variable. Alternatively, \citet{Boswijk1994} proposes to test for the joint significance of the least squares coefficients of all lagged variables with a Wald-type test. In our case, one could interpret exclusion of the lagged levels of the dependent variable, or the lagged levels of all variables, as evidence against the presence of cointegration. However, as discussed, an assumption of weak exogeneity is necessary when the aim is a direct interpretation of the estimated cointegration vector. Notwithstanding this caveat, selection consistency offers valuable insights when viewed as a screening mechanism that excludes irrelevant variables even in the absence of weak exogeneity. Moreover, since the set of variables included is strictly smaller than the time series dimension, it is possible to apply a traditional consistent estimator to the selected set of variables \citep[e.g.][]{Belloni2013}. However, ideally SPECS would contain desirable properties that omit the need of a second estimation procedure. For this reason, we establish the simultaneous consistency of the estimated coefficients in the following theorem.
\begin{theorem}\label{Thm:Estimation_Consistency}
Let $\bm S_T = \emph{diag}\left(\sqrt{T}\bm I_{s_\pi},\frac{T}{\sqrt{s_\delta}}\bm I_{s_\delta}\right)$ and $\bm Q$ as defined in \eqref{eq:Q}. Under the same assumptions as in Theorem \ref{Thm:Selection_Consistency}, it holds that
$\norm{\bm S_T\bm Q^{\prime -1}\left(\hat{\bm \gamma}_{S_{\gamma}} - \bm \gamma_{S_\gamma}\right)}_2 = O_p\left(s_\delta + \sqrt{s_\pi}\right)$.
\end{theorem}
The estimation consistency derived in Theorem \ref{Thm:Estimation_Consistency} does not place any restrictions on the relative growth rates of $T,N,p$, because it relies solely on high-level assumptions stated in the preceding section. However, when we derive sufficient conditions for the eigenvalue assumptions in Assumption \ref{Ass:eigenvalues} in Appendix \ref{App:eigenvalues} and provide a feasible method to construct weights that satisfy Assumption \ref{Ass:Regularization} in Section \ref{Sec:Init_est}, these restrictions do appear. We refer to Section \ref{sec:rates} for an explicit discussion.
\begin{remark}\label{Rem:derotate}
As an immediate consequence of Theorem \ref{Thm:Estimation_Consistency}, we have
$\norm{\hat{\bm \gamma}_{S_{\gamma}} - \bm \gamma_{S_\gamma}}_2 = O_p\left(\frac{s_\delta + \sqrt{s_\pi}}{\sqrt{T}}\right)$,
such that SPECS attains $\sqrt{T}$-consistency when $s_\delta$ and $s_\pi$ remain finite. To see this, note that by the assumption on $s_\delta$, it holds that $\frac{T}{\sqrt{s_\delta}} \geq \sqrt{T}$ for sufficiently large $T$. Then,
\begin{equation*}
\norm{\bm S_T\bm Q^{\prime -1}\left(\hat{\bm \gamma}_{S_{\gamma}} - \bm \gamma_{S_\gamma}\right)}_2 \geq \sqrt{T}\norm{\bm Q^{\prime -1}\left(\hat{\bm \gamma}_{S_{\gamma}} - \bm \gamma_{S_\gamma}\right)}_2.
\end{equation*}
Moreover, since the basis matrices $\bm B_{S_\delta}$ and $\bm B_{S_\delta,\perp}$ are not uniquely defined, we may impose a normalization such that $\norm{\bm Q}_2 \leq 1$. Then,
\begin{equation*}
\begin{split}
&\norm{\hat{\bm \gamma}_{S_{\gamma}} - \bm \gamma_{S_\gamma}}_2 = \norm{\bm Q^\prime\bm Q^{\prime -1}\left(\hat{\bm \gamma}_{S_{\gamma}} - \bm \gamma_{S_\gamma}\right)}_2
\leq \norm{\bm Q}_2\norm{\bm Q^{\prime -1}\left(\hat{\bm \gamma}_{S_{\gamma}} - \bm \gamma_{S_\gamma}\right)}_2 \leq \norm{\bm Q^{\prime -1}\left(\hat{\bm \gamma}_{S_{\gamma}} - \bm \gamma_{S_\gamma}\right)}_2,
\end{split}
\end{equation*}
such that $\norm{\bm S_T\bm Q^{\prime -1}\left(\hat{\bm \gamma}_{S_{\gamma}} - \bm \gamma_{S_\gamma}\right)}_2 \geq \sqrt{T}\norm{\hat{\bm \gamma}_{S_{\gamma}} - \bm \gamma_{S_\gamma}}_2$.
\end{remark}
As a corollary to Theorem \ref{Thm:Estimation_Consistency}, it is possible to establish a relationship between the limit distribution of SPECS and the OLS estimator based on the subset of relevant variables.
\begin{corollary}\label{Cor:OLS_oracle}
Define the OLS oracle estimator as $\hat{\bm \gamma}_{OLS,S_\gamma} = \argmin_{\bm \gamma}\norm{\bm M(\Delta \bm y - \bm V_{S_\gamma}\bm \gamma)}_2^2$. Then, with $\xi>0$ as in Assumption \ref{Ass:Regularization}, under the same assumptions as Theorem \ref{Thm:Selection_Consistency} it holds that
\begin{equation}
\norm{\bm S_T\bm Q^{\prime -1}\left(\hat{\bm \gamma}_{S_\gamma} - \hat{\bm \gamma}_{OLS,S_\gamma}\right)}_2 = o_p\left(\frac{\lambda_I(\sqrt{s_\delta} + \sqrt{s_\pi})}{T^{1/2 - \xi}}\right).
\end{equation}
\end{corollary}
The oracle results in Corollary \ref{Cor:OLS_oracle}, combined with the sign consistency from Theorem \ref{Thm:Selection_Consistency}, are suggestive of a post-selection inferential procedure. In particular, one may implement a two-step estimation procedure in which SPECS is used to perform variable selection in the first step and a regular OLS regression is performed on the selected variables in the second step. Then, after strengthening part \ref{ass:reg_lambda} of Assumption \ref{Ass:Regularization} to
$\lambda_I = o\left(\frac{T^{1/2-\xi}}{\sqrt{s_\delta} + \sqrt{s_\pi}}\right)$,
Corollary \ref{Cor:OLS_oracle} seems to validate the use of the regular OLS distribution for this two-step estimator, essentially ignoring the variable selection from the first stage. For example, in the case where $\abs{S_\gamma}$ remains finite, one could use the standard fixed-dimensional results \citep[e.g.][]{Boswijk1994} to perform inference. However, such a post-selection inferential procedure should be treated with caution, as it is well known that the selection step impacts the sampling properties of the estimator \citep[see][]{Leeb2005}. The convergence results of many selection procedures, SPECS included, hold pointwise only, i.e. the finite-sample distributions do not converge uniformly over the parameter space to their asymptotic distribution. The practical implication is that for certain values in the parameter space, relying on the oracle properties for post-selection test statistics may provide strongly misleading results. While developing a valid post-selection inference procedure to, for example, test for cointegration is certainly of interest, the field of valid post-selection inference is, despite its rapid development, still in its infancy. None of the currently existing methods, such as those considered in \citet{Berk2013}, \citet{Vandegeer2014}, \citet{Lee2016} or \citet{Chernozhukov2018}, can easily be adapted to - let alone validated in - our setting. Developing such a method therefore requires a full new theory which is outside the scope of the current paper.
\subsection{Initial Estimates}\label{Sec:Init_est}
In this section, we provide the reader with a directly implementable method to construct weights that satisfy Assumption \ref{Ass:Regularization}. As discussed in Section 2.2, we construct the weights as $\omega_i = \abs{\hat{\gamma}_{I,i}}^{-k}$. For our initial estimator we focus here on the ridge estimator, from which we can derive results for OLS as a special case, and comment on the lasso later on in the section.
Note that the power $k$ gives one the flexibility to adjust how big the wedge between relevant and irrelevant variables is. To illustrate, assume that $\hat{\gamma}_{I,i} = \gamma_i + O_p\left(T^{-a}\right)$ for all $i$. Then, it is clear that $\omega_i = O_p(1)$ when $\gamma_i \neq 0$ and $\omega_i = O_p\left(T^{ka}\right)$ when $\gamma_i = 0$. Therefore, larger values of $k$ will increase the rate at which the weights corresponding to the irrelevant variables diverge. Based on this principle, the availability of a consistent initial estimator allows us to construct weights that satisfy the conditions in Assumption \ref{Ass:Regularization}. However, while the idea of adjusting divergence rates through imposing varying values of $k$ seems theoretically attractive, large values of $k$ result in substantial amplification of finite-sample estimation error. As a result, the finite-sample performance of the lasso becomes unstable for large $k$, such that in practice one may want to set the value for $k$ as low as theoretically admissible.
Regardless of the choice of $k$, the basic ingredient for good adaptive weights is a consistent initial estimator. Therefore, we derive the consistency of the ridge estimator. Recall that the ridge estimator is defined as the minimizer of the following objective function:
\begin{equation}\label{eq:ridge}
G_R(\bm \gamma,\bm \theta) := \norm{\Delta \bm y - \bm V\bm \gamma - \bm D\bm \theta}_2^2 + \lambda_R\norm{\bm \gamma}_2^2.
\end{equation}
The properties of the ridge estimator are well-studied in the stationary setting \citep[e.g][Section 3.4.1]{Hastie2008}. However, to the best of our knowledge, no explicit results are available in the high-dimensional non-stationary case considered here.
In order to derive consistency of the ridge estimator, we redefine the transformed sample covariance matrix from Section \ref{sec:rot} and the corresponding bound on its minimum eigenvalue. Let $N_\delta = N-r$, $M_\pi = M+r$ and define the new scaling and rotation matrices as $\bm S_R = \diag\left(\sqrt{T}\bm I_{M_\pi},\frac{T}{\sqrt{N_\delta}}\bm I_{N_\delta}\right)$ and
\begin{equation*}
\bm Q_R = \begin{bmatrix}
\left(\bm B^\prime\bm B\right)^{-1/2}\bm B^\prime & 0\\
0 & \bm I_M\\
\left(\bm B_\perp^\prime\bm B_\perp\right)^{-1/2}\bm B_\perp^\prime & 0
\end{bmatrix},
\end{equation*}
respectively. The new transformed covariance matrix, based on the full dataset, is given by
\begin{equation}\label{eq:Sigma_R}
\hat{\bm \varSigma}_R = \bm S_R^{-1}\bm Q_R\bm V^\prime\bm M\bm V\bm Q_R\bm S_R^{-1} = \begin{bmatrix}
\hat{\bm \varSigma}_{R,11} & \hat{\bm \varSigma}_{R,12}\\
\hat{\bm \varSigma}_{R,21} & \hat{\bm \varSigma}_{R,22}
\end{bmatrix},\\
\end{equation}
with
$\hat{\bm \varSigma}_{R,11} = \frac{1}{T}\begin{bmatrix}
\bm B^\prime\bm Z_{-1}^\prime \bm M \bm Z_{-1}\bm B & \bm B^\prime\bm Z_{-1}^\prime \bm M \bm W\\
\bm W^\prime \bm M \bm Z_{-1}\bm B & \bm W^\prime \bm M \bm W
\end{bmatrix},$
and $\hat{\bm \varSigma}_{R,22} = \frac{N_\delta}{T^2}\bm B_\perp^\prime\bm Z_{-1}^\prime\bm M\bm Z_{-1}\bm B_\perp$.
Then, we extend the minimum eigenvalue bound in Assumption \ref{Ass:eigenvalues} to \eqref{eq:Sigma_R} as follows.
\begin{assumption}\label{Ass:eig_ridge}
Assume that, on a set with probability converging to 1 as $T,N,p \to \infty$, there exists a constant $\phi_R>0$, such that
$\underset{\bm x \in \mathbb{R}^{M_\pi}}{\text{inf}} \frac{\bm x^\prime \hat{\bm \varSigma}_{R,11}\bm x}{\bm x^\prime \bm x} \geq \phi_R$
an
$\underset{\bm x \in \mathbb{R}^{N_\delta}}{\text{inf}} \frac{\bm x^\prime \hat{\bm \varSigma}_{R,22} \bm x}{\bm x^\prime \bm x} \geq \phi_R.$
\end{assumption}
We now derive the convergence rate of the ridge estimator under a further restriction on the growth rates of $N,M$. The consistency of the ridge estimator is given in the following theorem.
\begin{theorem}\label{Thm:ridge}
Assume that $\frac{N_\delta}{T^{1/4}} \to 0$, $\frac{M_\pi}{\sqrt{T}} \to 0$, and $\lambda_R = O\left(\frac{\left(N_\delta + \sqrt{M_\pi}\right)\sqrt{T}}{\sqrt{\abs{S_\delta} + \abs{S_\pi}}}\right)$. Then, under Assumptions \ref{Ass:moments}-\ref{Ass:Dependence} and \ref{Ass:eig_ridge}, it holds that
$\norm{\bm S_R\bm Q_R^{\prime -1}\left(\hat{\bm \gamma}_R - \bm \gamma\right)}_2 = O_p\left(N_\delta + \sqrt{M_\pi}\right)$.
\end{theorem}
Similar to Remark \ref{Rem:derotate}, it follows from Theorem \ref{Thm:ridge} that
$\norm{\hat{\bm \gamma}_R - \bm \gamma}_2 = O_p\left(\frac{N_\delta + \sqrt{M_\pi}}{\sqrt{T}}\right)$.
Based on the assumption that $\frac{N_\delta}{T^{1/4}} \to 0$ and $\frac{M_\pi}{\sqrt{T}} \to 0$ in Theorem \ref{Thm:ridge}, it follows directly that $\norm{\hat{\bm \gamma}_R - \bm \gamma}_2 = o_p(1)$, and therefore ridge can be used to construct weights that satisfy our Assumption \ref{Ass:Regularization}. The exact values of $k$ that are needed theoretically vary depending on the number of (total and relevant) variables in the dataset; we return to this issue in Section \ref{sec:rates}.
The attentive reader may note that the admissible growth rates of $N_\delta,M_\pi$ in Theorem \ref{Thm:ridge} are the same as those initially assumed on the subsets of relevant variables, i.e. $s_\delta,s_\pi$, in Theorem \ref{Thm:Selection_Consistency}. The restriction imposed on the number of stochastic trends, $\frac{N_\delta}{T^{1/4}} \to 0$, corresponds closely to that of Corollary 2.1 in \citet{Liang2019}, who consider (co)integrated processes as well and roughly require that $\frac{N}{T^{1/4 - \nu}} \to 0$ for some $\nu > 0$. The growth rate of the total number of (implied) stationary variables is restricted to $\frac{M_\pi}{\sqrt{T}} \to 0$. While this may seem limited in comparison to the admissible (near) exponential growth in the stationary setting with i.i.d. Gaussian errors
\citep[e.g.][Thm 3]{KockCallot2015}, we stress that our time series framework is more general, allowing not only for integrated processes, but also substantial dependence in the stationary component. Regarding the latter, our assumptions closely match those in the second row of Table 6 of \citet{Medeiros2016} with $\zeta = 1$, where our allowed growth rates are only slightly slower.
Ideally, we would like to allow for faster rates of divergence for the set of the irrelevant variables. A prospective strategy to attain this, would be to implement the lasso as an initial estimator, the consistency of which may be derived with the use of a compatibility condition \citep[see for example][Ch. 6]{Buhlmann2011}. While desirable, deriving the validity of an appropriate compatibility condition is a considerable task. In addition to the difficulty of showing the theoretical validity of a compatibility condition in the non-stationary setting considered here, the use of a compatibility condition is further complicated by the fact that the stochastic trends asymptotically dominate the variation. More specifically, in order to attain a non-singular limit matrix, a rotation similar to $\bm Q$ is required that separates the stationary and non-stationary components in the full dataset. The standard compatibility condition would have to be adjusted in a non-trivial manner to account for such a rotation. Consequently, we leave the development of a suitable compatibility condition to future research, and instead focus on the ridge estimator under the more stringent growth rates on the number of variables. In the simulations we explore settings beyond these restrictive assumptions, and our adaptive weights continue to function in this case as well. We therefore conjecture that the suitability of the ridge estimator can be extended to a more general setting.
\begin{remark}
Theorem \ref{Thm:ridge} imposes no minimum growth rate of the penalty term $\lambda_R$ in \eqref{eq:ridge}. Therefore, in the case where $M+N < T$, the choice $\lambda_R = 0$ is both theoretically admissible and computationally feasible, such that consistency of the OLS estimator follows as a by-product of our result. Similarly, under the conditions imposed in Theorem \ref{Thm:ridge}, the lasso can also be shown to be a consistent initial estimator. In particular, Assumption \ref{Ass:eig_ridge} allows for the derivation of
a minimum eigenvalue bound for the sample covariance matrix of the full data set, which enables application of standard proofs of consistency that are familiar from the fixed-dimensional setting. Due to space consideration, we refrain from providing a full proof on this conjecture, but refer the interested reader to Theorem 3.1 in \citet{Liao2015}, the proof of which may be adjusted to fit the current setting.
\end{remark}
\subsection{Implications for Particular Model Specifications}
To fully appreciate the theoretical results in the preceding section, a detailed understanding of the generality provided by the set of imposed assumptions is helpful. For example, as the results are derived without requiring weak exogeneity, our set of assumptions allows for the presence of stationary variables in the data. However, in the absence of weak exogeneity, model interpretation becomes non-standard and the notion of sparsity carries non-trivial annotations. Therefore, in this section we elaborate on several relevant model specifications to demonstrate the flexibility of the single-equation model and highlight the practical implications of variable selection in such a general framework.
\subsubsection{Sparsity and Weak Exogeneity}\label{Sec:Sparsity}
The benefit of $\ell_1$-regularized estimation stems from its ability to identify sparse parameter structures. However, the concept of sparsity in the conditional models here considered merits additional clarification, as the potential absence of weak exogeneity obscures standard interpretability. Accordingly, in this section we comment on the interplay between weak exogeneity and sparsity and provide several illustrative examples of sparse DGPs. For simplicity of illustration, we assume in this and the following section that $\bm \mu=\bm \tau=\bm{0}$.
In Section \ref{sec:model} we argue that the coefficients regulating the long-run dynamics in the conditional model are generally derived from linear combinations of the cointegrating vectors in the VECM representation \eqref{eq:VECM}. By decomposing the matrix with adjustment rates as $\bm A = (
\bm \alpha_1, \bm A_2^\prime)^\prime$, we obtain the explicit construction
$\bm \delta = \bm B(\bm \alpha_1 - \bm A_2^\prime\bm \varSigma_{\epsilon,22}^{-1}\bm \sigma_{\epsilon,21})$.
Hence, it follows that $\delta_i=0$ if the sparsity condition
$\bm \beta_i^\prime \left(\bm \alpha_1 - \bm A_2^\prime\bm \varSigma_{\epsilon,22}^{-1}\bm \sigma_{\epsilon,21}\right) = 0$
is satisfied, where $\bm \beta_i$ is the $i$-th row of $\bm B$. While this condition may hold in a variety of non-trivial ways, specific cases of interest that lead to sparsity in $\bm \delta$ can be derived. For example, an integrated variable $x_{i,t}$ that does not cointegrate with any of the variables in the system ($\bm \beta_i = \bm{0}$), will carry a zero coefficient in the derived single-equation long-run equilibrium.
As a more general example, assume that the researcher observes the $N$-dimensional time series $\bm z_t = (\bm z_{1,t}^\prime,\bm z_{2,t}^\prime)^\prime = (y_t,\bm x^\prime_t)^\prime$, from time $t=1,\ldots,T$, where $\bm z_{1,t} = (y_t,\bm x_{1,t}^\prime)^\prime$ is an $N_1$-dimensional time series and $\bm z_{2,t}$ is an $N_2$-dimensional time series. Moreover,
\begin{equation}\label{eq:DGP_SVECM}
\begin{split}
\begin{bmatrix}
\Delta \bm z_{1,t}\\
\Delta \bm z_{2,t}
\end{bmatrix} &= \begin{bmatrix}
\bm \varPi_{11} & \bm \varPi_{12}\\
\bm \varPi_{21} & \bm \varPi_{22}
\end{bmatrix} \begin{bmatrix}
\bm z_{1,t-1}\\
\bm z_{2,t-1}
\end{bmatrix} + \sum_{j=1}^p \begin{bmatrix}
\bm \varPhi_{j,11} & \bm \varPhi_{j,12}\\
\bm \varPhi_{j,21} & \bm \varPhi_{j,22}
\end{bmatrix} \begin{bmatrix}
\Delta \bm z_{1,t-j}\\
\Delta \bm z_{2,t-j}
\end{bmatrix} + \begin{bmatrix}
\bm \epsilon_{1,t}\\
\bm \epsilon_{2,t}
\end{bmatrix}\\
&= \bm \varPi\bm z_{t-1} + \sum_{j=1}^p\bm \varPhi_j \Delta \bm z_{t-j} + \bm \epsilon_t.
\end{split}
\end{equation}
In addition, assume that $\bm \varSigma_\epsilon = \E\left(\bm \epsilon_t\bm \epsilon_t^\prime\right)$ satisfies Assumption \ref{Ass:moments} and can be decomposed as
\begin{equation}\label{eq:Sigma_eps_res}
\begin{split}
&\bm \varSigma_\epsilon = \begin{bmatrix}
\bm \varSigma_{\epsilon,11} & \bm{0}\\
\bm{0} & \bm \varSigma_{\epsilon,22}
\end{bmatrix},\text{ with } \bm \varSigma_{\epsilon,11} = \begin{bmatrix}
\sigma_{1,11} & \bm \sigma_{1,21}^\prime\\
\bm \sigma_{1,21} & \bm \varSigma_{1,22}
\end{bmatrix}.
\end{split}
\end{equation}
Then, the quantities appearing in the single-equation model in \eqref{eq:CECM} take on the form
\begin{equation}\label{eq:CECM_coefs}
\begin{split}
\bm \pi_0 &= \begin{bmatrix}
\bm \varSigma_{1,22}^{-1} & \bm{0}\\
\bm{0} & \bm \varSigma_{\epsilon,22}^{-1}
\end{bmatrix}\begin{bmatrix}
\bm \sigma_{1,21}\\
\bm{0}
\end{bmatrix} = \begin{bmatrix}
\bm \pi_{0,1}\\
\bm{0}
\end{bmatrix},\\
\bm \delta &= \begin{bmatrix}
\bm \varPi_{11}^\prime & \bm \varPi_{21}^\prime\\
\bm \varPi_{12}^\prime & \bm \varPi_{22}^\prime
\end{bmatrix}\begin{bmatrix}
1\\
-\bm \pi_0\\
\end{bmatrix} = \begin{bmatrix}
\bm \varPi_{11}^\prime\\
\bm \varPi_{12}^\prime
\end{bmatrix}\begin{bmatrix}
1\\
-\bm \pi_{0,1}
\end{bmatrix} = \begin{bmatrix}
\bm \delta_1\\
\bm \delta_2
\end{bmatrix},\\
\bm \pi_j &= \begin{bmatrix}
\bm \varPhi_{j,11}^\prime & \bm \varPhi_{j,21}^\prime\\
\bm \varPhi_{j,12}^\prime & \bm \varPhi_{j,22}^\prime
\end{bmatrix}\begin{bmatrix}
1\\
-\bm \pi_0
\end{bmatrix} = \begin{bmatrix}
\bm \varPhi_{j,11}^\prime\\
\bm \varPhi_{j,12}^\prime
\end{bmatrix}\begin{bmatrix}
1\\
-\bm \pi_{0,1}
\end{bmatrix} = \begin{bmatrix}
\bm \pi_{j,1}\\
\bm \pi_{j,2}
\end{bmatrix}.
\end{split}
\end{equation}
The definitions in \eqref{eq:CECM_coefs} demonstrate that, under the restriction that the errors driving $\bm z_{1,t}$ and $\bm z_{2,t}$ are uncorrelated, sparsity in the single-equation model arises when (a subset of) $\bm z_{2,t}$ does not Granger-Cause $\bm z_{1,t}$. For example, in the extreme case where $\bm \varPi_{12} = \bm{0}$ and $\bm \varPhi_{12} = \bm{0}$, we have $\bm \delta_2 = \bm{0}$ and $\bm \pi_{j,2} = 0$, respectively.
Consequently, then the single-equation model reads as
\begin{equation}\label{eq:CECM_sparse}
\begin{split}
\Delta y_t &= \bm \delta^\prime \bm z_{t-1} + \bm \pi_0^\prime \Delta \bm x_t + \sum_{j=1}^p \bm \pi_j^\prime\Delta \bm z_{t-j} + \epsilon_{y,t}\\
&= \bm \delta_1^\prime \bm z_{1,t-1} + \bm \pi_{0,1}^{\prime}\Delta \bm x_{1,t} + \sum_{j=1}^p \bm \pi_{1,j}^\prime \Delta \bm z_{1,t-j} + \epsilon_{y,t}.
\end{split}
\end{equation}
As an interesting special case, consider the decomposition in \eqref{eq:DGP_SVECM} in which $z_{2,t}=\epsilon_{2,t}$ is scalar-valued with $\E(\epsilon_{2,t}\bm \epsilon_{1,t}) = \bm{0}$. Then, it is straightforward to see that $\bm \pi_{12}=\bm \pi_{21}=\bm{0}$, $\pi_{22} = -1$ and, consequently, $\delta_N = 0$. This finding highlights that stationary variables result in sparsity in $\bm \delta$ only when they are fully exogenous, as said variables may enter the implied cointegrating vector through their correlation structure with the other variables in the system. This further demonstrates the difficulty of direct interpretation of $\bm \delta$ without imposing additional restrictions on the DGP. From a prediction perspective, however, the model's ability to include stationary variables through their correlation structure is clearly a desirable feature.
Finally, we consider a DGP in which $\bm \varSigma_\epsilon$ follows a Toeplitz structure with $\sigma_{\epsilon,ij} = \rho^{\abs{i-j}}$. After partitioning $\bm \varSigma_\epsilon$ as in \eqref{eq:covariance partition}, we can rewrite
\begin{equation}\label{eq:Toeplitz proof}
\begin{split}
\bm \sigma_{\epsilon,21} = \begin{bmatrix}
\rho^1\\
\vdots \\
\rho^{N-1}
\end{bmatrix} &= \begin{bmatrix}
\rho^0 & \ldots & \rho^{N-2}\\
\vdots & \ddots & \vdots\\
\rho^{N-2} & \ldots & \rho^0
\end{bmatrix}\begin{bmatrix}
\rho^1\\
0\\
\vdots\\
0
\end{bmatrix} =\bm \varSigma_{\epsilon,22}\bm \pi_0,
\end{split}
\end{equation}
thus showing that $\bm \pi_0 = \bm \varSigma_{\epsilon,22}^{-1}\bm \sigma_{\epsilon,21} = (\rho, 0, \ldots, 0)^\prime$.\footnote{It is straightforward to show that this property carries over to covariance matrices with a block-diagonal Toeplitz structure, with each block $\bm \varSigma_\epsilon^{(k)}$ having the form $\sigma^{(k)}_{i,j}=\rho_{(k)}^{\abs{i-j}}$. The number of non-zero elements in the resulting vector $\bm \pi_0$ will equal the number of blocks in the covariance matrix.} As $\bm \delta^\prime = (1,-\bm \pi_0^\prime)\bm A\bm B^\prime$, this implies that only the long-run equilibria that occur in the equations for $\Delta y_t$ or its cross-sectionally neighbouring variable will be part of the linear combination in the derived the single-equation model. Consequently, any variables in the dataset that are not contained in the equilibria occurring in these equations will induce sparsity in $\bm \delta$.
\subsubsection{Mixed Orders of Integration}
One of the most prominent benefits of SPECS is the ability to model potentially non-stationary and cointegrated data without the need to adopt a pre-testing procedure with the aim of checking, and potentially correcting, for the order of integration or to decide on the appropriate cointegrating rank of the system. The assumptions under which our theory is developed are compatible with a wide variety of DGPs, including settings where the dataset contains an arbitrary mix of $I(1)$ and $I(0)$ variables. The researcher simply transforms the dataset according to \eqref{eq:CECM} and SPECS provides consistent estimation of the parameters and identification of the correct implied sparsity pattern. The purpose of this section is to demonstrate this feature by means of some illustrative examples.
The central idea underlying the above feature is that a single-equation model can be derived from any system admitting a finite order VECM representation. In a VECM system containing variables with mixed orders of integration, however, each stationary variable adds an additional trivial cointegrating vector. Such a vector corresponds to a unit vector that equals 1 on the index of the stationary variable. For illustrative purposes, we consider the following general example. Define $\bm z_t = (\bm z_{1,t}^\prime,\bm z_{2,t}^\prime)^\prime$, where $\bm z_{1,t} \sim I(0)$ and $\bm z_{2,t} \sim I(1)$ and possibly cointegrated. Let the dimensions of $\bm z_{1,t}$ and $\bm z_{2,t}$ be $N_1$ and $N_2$ respectively. Then, $\bm z_t$ admits the representation
\begin{equation}\label{eq:VECM_stat}
\begin{split}
\begin{bmatrix}
\Delta \bm z_{1,t}\\
\Delta \bm z_{2,t}
\end{bmatrix} &= \begin{bmatrix}
-\bm I_{N_1} & \bm{0}\\
\bm{0} & \bm \varPi_{22}
\end{bmatrix}\begin{bmatrix}
\bm z_{1,t-1}\\
\bm z_{2,t-2}
\end{bmatrix} + \bm \varPhi(L)\Delta \bm z_{t-1} + \bm \epsilon_t= \bm A\bm B^\prime\bm z_{t-1} + \bm \varPhi(L)\Delta \bm z_{t-1} + \bm \epsilon_t,
\end{split}
\end{equation}
where $\bm \varPhi(L)$ corresponds to a $p$-dimensional matrix lag polynomial by Assumption \ref{Ass: GRT} and $\bm \epsilon_t$ satisfies the conditions in Assumption \ref{Ass:moments}. As long as the design of \eqref{eq:VECM_stat} conforms to Assumption \ref{Ass: GRT} and \ref{Ass:Dependence}, our main results apply to this setting and both selection and estimation consistency is maintained. For the extreme case in which all variables are integrated of order one, but none are cointegrate, we define $\bm A=\bm B = \bm{0}$. Clearly, it follows that $\bm \delta=\bm{0}$, such that the single-equation model can be seen as a conditional model obtained from a VAR specified in differences. In the other extreme case, when the levels of all variables in the VECM are weakly stationary, decomposition \eqref{eq:VECM_stat} would simply lead to a VECM in which $-\bm A=\bm B=\bm I_N$, thereby enabling the results in Section \ref{Sec:Asymptotics} to carry through.\footnote{When all variables are stationary, SPECS can also be shown to consistently estimate the parameters based on the well-documented properties of the adaptive lasso in stationary time series settings, such as those considered in \citet{Medeiros2016} and \citet{Masini2019}.}
\subsubsection{Rates of Convergence} \label{sec:rates}
We conclude our theoretical analysis with a detailed illustration of the attainable rates of convergence in different asymptotic frameworks. The rates of convergence of $\hat{\bm \gamma}_R$ and $\hat{\bm \gamma}$, as well as the specific construction of the initial weights, are dependent on the growth rates of $N,p,r,\abs{S_\delta}$ and $\abs{S_\pi}$. Because of the trade-off between the admissible dimension and the rate of convergence, the choice of the desired asymptotic framework is likely dependent on the specific application. For example, typical macro-economic applications are characterized by short panel datasets which would require a framework in which the cross-sectional dimension grows as fast as theoretically admissible. On the other hand, in applications with a large number of time series observations, such as forecasting based on high-frequency data, the assumption that the number of (potentially) relevant variables grows slow relative to the available time periods seems reasonable. Therefore, to aid interpretation of our results, we provide an overview with different asymptotic frameworks and the corresponding penalty parameters, weight constructions and convergence rates of the initial estimator in Table \ref{Tab4:settings}. The weights for $\delta_i$ and $\pi_j$ are constructed as $\omega_i = \abs{\hat{\delta}_{R,i}}^{-k_\delta}$ and $\omega_{N+j} = \abs{\hat{\pi}_{R,j}}^{-k_\pi}$.
\begin{table}
\caption{Dimensions, Penalties, Weights and Convergence Rates}
\begin{tabularx}{\textwidth}{XXXXXllXX}
\hline
$N$ & $p$ & $r$ & $\abs{S_\delta}$ & $\abs{S_\pi}$ & $k_\delta$ & $k_\pi$ & $\lambda_R,\lambda_I$ & $\norm{\hat{\bm \gamma} - \bm \gamma}_2$ \tabularnewline
\hline
\hline
fixed & fixed & fixed & fixed & fixed & 2 & 1 & $KT^{2/5}$ & $O_p\left(T^{-1/2}\right)$ \tabularnewline
$T^{1/4}$ & fixed & fixed & fixed & fixed & 3 & 1 & $KT^{2/5}$ & $O_p\left(T^{-1/2}\right)$ \tabularnewline
$T^{1/4}$ & $T^{1/4}$ & fixed & fixed & fixed & 3 & 2 & $KT^{2/5}$ & $O_p\left(T^{-1/2}\right)$ \tabularnewline
$T^{1/4}$ & $T^{1/4}$ & $T^{1/4}$ & fixed & fixed & 3 & 2 & $KT^{2/5}$ & $O_p\left(T^{-1/2}\right)$ \tabularnewline
$T^{1/4}$ & $T^{1/4}$ & $T^{1/4}$ & fixed & $T^{1/4}$ & 4 & 2 & $KT^{2/5}$ & $O_p\left(T^{-3/8}\right)$ \tabularnewline
$T^{1/4}$ & $T^{1/4}$ & fixed & $T^{1/4}$ & $T^{1/4}$ & 4 & 2 & $KT^{2/5}$ & $O_p\left(T^{-1/4}\right)$ \tabularnewline
$T^{1/4}$ & $T^{1/4}$ & $T^{1/4}$ & $T^{1/4}$ & $T^{1/4}$ & 4 & 2 & $KT^{2/5}$ & $O_p\left(T^{-3/8}\right)$ \tabularnewline
\hline
\end{tabularx}
\label{Tab4:settings}
\caption*{This table displays possible settings for the weights ($k_\delta,k_\pi$) and penalty parameters ($\lambda_I,\lambda_R$) that satisfy Assumption \ref{Ass:Regularization} under a variety of asymptotic frameworks ($N,r,p,\abs{S_\delta},\abs{S_\pi}$). The convergence rate of SPECS is displayed in the last column.}
\end{table}
The first row of Table \ref{Tab4:settings} corresponds to the classic fixed-dimensional case. It is reassuring that, similar to the OLS estimator, SPECS obtains $\sqrt{T}$-convergence, with the additional benefit of allowing for consistent recovery of the sparsity pattern. In fact the next three rows highlight that when $N$, $p$ or $r$ diverge, while the number of relevant variables remains fixed, SPECS maintains its $\sqrt{T}$-convergence as long as the penalty weights $k_\delta$ and $k_\pi$ are adjusted appropriately. In the fifth row, we allow the number of relevant stationary variables, i.e. $\abs{S_\pi}$ to diverge as well. This setting may be preferred when the integrated time series remain persistent after being transformed to stationarity by differencing. We observe that consistency is maintained, although even sharper weights are required and the rate of convergence has reduced to $T^{3/8}$. In the sixth row we additionally allow the number of relevant non-stationary, i.e. $\abs{S_\delta}$, to increase, whereas the number of cointegrating vectors remains fixed. The increased number of non-zero coefficients corresponding to non-stationary variables reduces the rate of convergence to $T^{1/4}$. Interestingly, in the last row we let the dimension of the cointegrating subspace $r$ grow at the same rate. As illustrated in Remark \ref{Rem:factor}, this setting naturally occurs when the data is modelled by a non-stationary factor model with idiosyncratic components. In this framework, the number of stochastic trends driving the subset of relevant variables, i.e. $s_\delta$, remains fixed, which positively affects the convergence rate of SPECS.
\bigskip
We consider the theoretical results presented in this section to be of a double nature. On the one hand, it is reassuring that consistent estimation remains feasible in growing dimensions and that suitable weights are available. On the other hand, we acknowledge that the required restrictions on the growth rate of the number of variables seem to caution against application of penalized regression in very high-dimensional settings. However, it is worth noting that the restrictions on $N$ and $p$ largely result from the use of ridge regression as an initial estimator. Indeed, the availability of a novel compatibility condition could justify the use of the lasso as an initial estimator and will allow for generalization of our theoretical results to even higher dimensional asymptotic frameworks. We consider this an interesting avenue for future research.
\begin{remark}\label{Rem:factor}
The VECM \eqref{eq:DGP_SVECM} can be rewritten into a non-stationary factor model with stationary idiosyncratic components, similarly to \citet{Banerjee2014}. Based on the VMA representation of $\bm z_t$ in \eqref{eq:GRT}, with $\bm C$ a matrix of reduced rank, we can rewrite the process as
\begin{equation}\label{eq:ns_factors}
\bm z_t = \bm C\bm s_t + \bm \mu + \bm \tau t + \bm u_t = \bm \varLambda\bm f_t+ \bm \mu + \bm \tau t + \bm u_t,
\end{equation}
where $\bm \varLambda = \bm B_\perp\left(\bm A_\perp^\prime\left(\bm I-\sum_{j=1}^p\bm \varPhi_j\right)\bm B_\perp\right)^{-1}$, $\bm f_t = \bm A_\perp^\prime\bm s_t$ and $\bm u_t = \bm C(L)\bm \epsilon_t + \bm z_0$. This representation is particularly relevant in relation to the growth rate of $N_\delta = N-r$. Typically, the theory for consistent estimation of \eqref{eq:ns_factors} is derived under the assumption that the $N_\delta$ factors remain fixed, while letting both $N$ and $T$ go to infinity. Hence, in this framework, noting that $s_\delta \leq N_\delta$, the assumptions that $\frac{s_\delta}{T^{1/4}} \to 0$ and $\frac{N_\delta}{T^{1/4}} \to 0$ in Theorems \ref{Thm:Selection_Consistency}-\ref{Thm:ridge} are automatically satisfied. Consequently, the convergence rates of the initial and final estimators are given by $\norm{\hat{\bm \gamma}_R - \bm \gamma}_2 = O_p\left(\sqrt{\frac{M_\pi}{T}}\right)$ and $\norm{\hat{\bm \gamma}-\bm \gamma}_2 = O_p\left(\sqrt{\frac{s_\pi}{T}}\right)$.
\end{remark}
\section{Simulations}\label{sec: simulations}
In this section we analyze the selective capabilities and predictive performance of SPECS by means of simulations. We estimate the single-equation model according to the objective function \eqref{eq:SPECS} with the following settings for the penalty rates:
\begin{enumerate}
\item Ordinary Least Squares (OLS: $\lambda_G=0$, $\lambda_I=0$),
\item Autoregressive Distributed Lag (ADL: $\lambda_G = 0$, $\lambda_I > 0$, $\omega_i = \infty$ for $i=1,\ldots,N$),
\item SPECS - no group penalty (SPECS$_1$: $\lambda_G = 0$, $\lambda_I > 0$),
\item SPECS - group penalty (SPECS$_2$: $\lambda_G > 0$, $\lambda_I > 0$)\footnote{As a helpful reminder, the reader may relate the subscript to the number of penalty categories included in the estimation; SPECS$_1$ only contains an individual penalty whereas SPECS$_2$ contains both a group penalty and and individual penalty.}.
\end{enumerate}
The OLS estimator is only included when feasible according to the dimension of the model to estimate and we additionally include a penalized autoregressive distributed lag model (ADL) with all variables entering in first differences. The latter model can be interpreted as the conditional model one would obtain when ignoring cointegration in the data and specifying a VAR in differences as a model for the full system. The resulting conditional model is the same as the CECM that we consider, but with the built-in restriction $\bm \delta=\bm{0}$.
We estimate the solutions for a grid of penalty values and construct the weights from an initial ridge estimator as proposed in Section \ref{Sec:Init_est}. For ADL and SPECS$_1$, we consider 100 possible values for $\lambda_I$ and choose the final model based on the BIC criterion. Alternatively, for SPECS$_2$, the model selection takes place over a two-dimensional grid consisting of 100 values for $\lambda_I$ and 10 possible values for $\lambda_G$, with model selection again being based on the BIC criterion. The weights are defined by $\omega_i = \abs{\hat{\gamma}_{R,i}}^{-k}$, where $k=2$ for $i \in \lbrace 1,\ldots,N\rbrace$ and $k=1$ for $i \in \lbrace N+1,\ldots,N+M \rbrace$.
We consider three different settings under which we analyze the performance of our estimators; the first setting aims to analyze the effects of dimensionality and weak exogeneity, the second setting explores the effect of the variables' orders of integration and the third setting considers the performance in non-sparse settings. Each setting is described in detail below.
\subsection{Dimensionality and Weak Exogeneity}\label{Sec:dim_we}
In the first part of our simulation study we focus on the effects of dimensionality and weak exogeneity on a (co)integrated dataset. Our simulation DGP takes the form
\begin{equation}\label{eq:general DGP sim}
\Delta \bm z_t = \bm A\bm B^\prime \bm z_{t-1} + \bm \varPhi_1 \Delta \bm z_{t-1} + \bm \epsilon_t,
\end{equation}
with $t=1,\ldots, T=100$, $\bm \epsilon_t \sim \mathcal{N}(0,\bm \varSigma)$ and $\sigma_{ij} = 0.8^{|i-j|}$. Furthermore, $\bm \varPhi_1$, the coefficient matrix regulating the short-run dynamics is generated as $0.4 \cdot \bm I_N$, where $N$ varies depending on the specific DGP considered. Based on this DGP, the single-equation model takes on the form
\begin{equation*}
\Delta y_t = \bm \delta^\prime \bm z_{t-1} + \bm \pi_0^\prime \Delta \bm x_t + \bm \pi_1^\prime \Delta \bm z_{t-1} + \epsilon_{y,t},
\end{equation*}
with $\bm \pi_0$ and $\bm \pi_1$ as defined in \eqref{eq:CECM}. We consider a total of four different settings, corresponding to different combinations of (i) dimensionality (low/high) and (ii) weak exogeneity (present/absent). The corresponding parameter settings and implied cointegrating vector $\bm \delta$ are given in Table \ref{table:DGP WE and Dim}.
\begin{table}[t]
\caption{Simulation Design for the First Study (Dimensionality and Weak Exogeneity)}\label{table:DGP WE and Dim}
\begin{tabularx}{\textwidth}{cccc}
\hline
Low Dimension & $\bm A$ & $\bm B$ & $\bm \delta$ \tabularnewline
\hline
WE & $\alpha_{1}\cdot\begin{bmatrix}1\\
\bm{0}_{9\times 1}
\end{bmatrix}$ & $\begin{bmatrix}\tilde{\bm \iota}\\
\bm{0}_{5\times 1}
\end{bmatrix}$ & $\alpha_{1}\cdot\bm B$ \tabularnewline
No WE & $\alpha_{1}\cdot
\bm B$
&
$\begin{bmatrix}\tilde{\bm \iota} & \bm{0}_{5\times 1}\\
\bm{0}_{5\times 1} & \tilde{\bm \iota}
\end{bmatrix}$
& $(1+\rho)\alpha_{1}\cdot
\begin{bmatrix}\tilde{\bm \iota}\\
\bm{0}_{5\times 1}
\end{bmatrix}$ \tabularnewline
\hline
High Dimension & $\bm A$ & $\bm B$ & $\bm \delta$\tabularnewline
\hline
WE & $\alpha_{1}\cdot\begin{bmatrix}1\\
\bm{0}_{49 \times 1}
\end{bmatrix}$ & $\begin{bmatrix}\tilde{\bm \iota}\\
\bm{0}_{45\times 1}
\end{bmatrix}$ & $\alpha_{1}\cdot\bm B$ \tabularnewline
No WE & $\alpha_{1}\cdot\bm B$ & $\begin{bmatrix}\tilde{\bm \iota} &\bm{0}_{5\times 1} & \bm{0}_{5\times 1}\\
\bm{0}_{5\times 1} & \tilde{\bm \iota} & \bm{0}_{5\times 1}\\
\bm{0}_{5\times 1} & \bm{0}_{5\times 1} & \tilde{\bm \iota}\\
\bm{0}_{35\times 1} & \bm{0}_{35\times 1} & \bm{0}_{35\times 1}
\end{bmatrix}$ & $(1+\rho)\alpha_{1}\cdot\begin{bmatrix}\tilde{\bm \iota}\\
\bm{0}_{45\times 1}
\end{bmatrix}$ \tabularnewline
\hline
\end{tabularx}
\caption*{Notes: The low-dimensional (high-dimensional) design corresponds to a system with $N=10$ ($N=50$) unique time series and $N^\prime = 31$ ($N^\prime = 151$) parameters to estimate. Furthermore, $\tilde{\bm \iota} = (1, -\bm \iota_4^\prime)^\prime$ and $\alpha_1 = -0.5,-0.45,\ldots,0$ regulates the adjustment rate towards the equilibrium.}
\end{table}
We measure the selective capabilities based on three metrics. The pseudo-power of the models measures the ability to appropriately pick up the presence of cointegration in the underlying DGP. For the OLS procedure we perform the Wald test proposed by \citet{Boswijk1994}. When the OLS fitting procedure is unfeasible due to the high-dimensionality, we perform the Wald test on the subset of variables included after fitting SPECS$_1$ and refer to this approach as Wald-PS (where PS stands for post-selection). Despite the caveats of oracle-based post-selection inference discussed after Corollary \ref{Cor:OLS_oracle}, the inclusion of Wald-PS still offers valuable insights regarding the performance one may expect of such a procedure in light of the aforementioned limitation. SPECS is used as an alternative to this cointegration test by simply checking whether at least one of the lagged levels is included in the model. The percentage of trials in which cointegration is found is then reported as the pseudo-power.
Second, for each trial the Proportion of Correct Selection (PCS) measures the proportion of correctly selected variables, while the Proportion of Incorrect Selection (PICS) describes, as the name may suggest, the proportion of incorrectly selected variables. They are given by
\begin{equation*}
PCS = \frac{\abs{\left\lbrace \hat{\gamma}_j \neq 0\right\rbrace \cap \left\lbrace\gamma_j \neq 0\right\rbrace}}{\abs{\left\lbrace \gamma_j \neq 0 \right\rbrace}}; \qquad
PICS = \frac{\abs{\left\lbrace \hat{\gamma}_j \neq 0\right\rbrace \cap \left\lbrace\gamma_j = 0\right\rbrace}}{\abs{\left\lbrace \gamma_j = 0 \right\rbrace}}.
\end{equation*}
The PCS and PICS are calculated for SPECS$_1$ and SPECS$_2$ and averaged over all trials.
Finally, we consider the predictive performance in a simulated nowcasting application, where we implicitly assume that the information on the latest realization of $\bm x_T$ arrives before the realization of $y_T$. These situations frequently occur in practice, see \citet{Giannone2008} and the references therein for an overview as well as the empirical application considered in Section \ref{Sec:application}. Due to the construction of the single-equation model, in which contemporaneous values of the conditioning variables contribute to the contemporaneous variation in the dependent variable, our proposed method is particularly well-suited to this application. For any of the considered fitting procedures, the nowcast is given by $\hat{y}_T = \hat{\bm \delta}^\prime \bm z_{T-1} + \hat{\bm \pi}^\prime \Delta \bm x_T + \hat{\bm \phi}^\prime \Delta \bm z_{T-1}$, where by construction $\hat{\bm \delta}=\bm{0}$ in the ADL model. For each method we record the root mean squared nowcast error (RMSNE) relative to the OLS oracle procedure fitted on the relevant variables.
\begin{sidewaysfigure}
\includegraphics[width=\textwidth]{SimOneOrder}
\caption{Pseudo-Power, Proportion of Correct Selection (PCS), Proportion of Incorrect Selection (PICS) and Root Mean Squared Nowcast Error (RMSNE) for Low- and High-Dimensional specifications. The adjustment rate multiplier $a$ is on the horizontal axis.}\label{Fig: metrics1}
\end{sidewaysfigure}
Figure \ref{Fig: metrics1} visually displays the evolution of our performance metrics over a range of values for $a$, representing increasingly faster rates of adjustment towards the long-run equilibrium. The first row of plots shows near-perfect performance of SPECS over all metrics. The pseudo-size is slightly lower than the size of the Wald test when the latter is controlled at 5\%, whereas the pseudo-power quickly approaches one. Following expectations, the pseudo-size for SPECS$_2$ is slightly lower as a result of the additional group penalty. Focussing on the selection of variables, we find that for faster adjustment rates, SPECS is able to exactly identify the sparsity pattern with very high frequency, as demonstrated by the PCS approaching 100\% and the PICS staying near 0\%. Furthermore, the MSNE obtained by our methods is close to the OLS oracle method and is substantially lower than the MSNE obtained by the ADL model for faster adjustment rates, while being almost identical absent of cointegration. The picture remains qualitatively similar when moving away from weak exogeneity while staying in a low-dimensional framework, although the gain in predictive performance over the ADL has decreased somewhat. We postulate that the ADL may benefit from a bias-variance tradeoff, given that the correctly specified single-equation model is sub-optimal in terms of efficiency absent of weak exogeneity compared to a full system estimator. Nonetheless, SPECS is clearly preferred.
The performance in the high-dimensional setting is displayed in rows 3 and 4 of Figure \ref{Fig: metrics1}. When the conditioning variables are weakly exogenous with respect to the parameters of interest, the selective capabilities remain strong. The pseudo-power demonstrates the attractive prospect of using our method as an alternative to cointegration testing, especially when taking into consideration that the traditional Wald test is infeasible in the current setting. In addition, the nowcasting performance remains far superior to that of the misspecified ADL. The last row depicts the performance absent of weak exogeneity. In this setting, exact identification of the implied cointegrating vector occurs less frequently, which seems to negatively impact the nowcasting performance. However, the misspecified ADL is still outperformed, despite the deterioration in the selective capabilities of our method.
\subsection{Mixed Orders of Integration}\label{sec:sim_mixed}
We next analyze the performance of SPECS on datasets containing variables with mixed orders of integration. The aim of this section is to gain an understanding of the relative performance of SPECS when not all time series are (co)integrated and to compare the performance of SPECS to traditional approaches that rely on pre-testing. The latter goal is attained by adding an additional penalized ADL model to the comparison, namely one in which the data is first corrected for non-stationarity based on a pre-testing procedure in which an Augmented Dickey-Fuller (ADF) test is performed on the individual series. We refer to this procedure as the ADL-ADF model. Based on the general DGP \eqref{eq:general DGP sim}, we distinguish four different cases, corresponding to (i) different orders of the dependent variable ($I(0)$/$I(1)$) and (ii) different degrees of persistence in the stationary variables (low/high). The choice to include varying degrees of persistence is motivated by the conjecture that the performance of the pre-testing procedure incorporated in the ADL-ADF model may deteriorate when the degree of persistence increases, which in turn translates to a decrease in the overall performance of the procedure.
\begin{table}[t]
\caption{Simulation Design for the Second Study (Mixed Orders of Integration)}\label{table:DGP mixed order}
\begin{tabularx}{\textwidth}{cccc}
\hline
Mixed Order & $\bm A$ & $\bm B$ & $\bm \delta$ \tabularnewline
\hline
$y \sim I(0)$ & $\begin{bmatrix}
1 & 0 & \bm{0}_{1\times24}\\
\bm{0}_{15\times 1} & a\bm B^* & \bm{0}_{15\times 24}\\
\bm{0}_{10\times 1} & \bm{0}_{10\times 3} & \bm{0}_{10 \times 24}\\
\bm{0}_{24 \times 1} & \bm{0}_{24 \times 3} & \bm I_{24}
\end{bmatrix}$ & $\begin{bmatrix}
-b & 0 & \bm{0}_{1\times24}\\
\bm{0}_{15\times 1} & \bm B^* & \bm{0}_{15\times 24}\\
\bm{0}_{10\times 1} & \bm{0}_{10\times 3} & \bm{0}_{10 \times 24}\\
\bm{0}_{24 \times 1} & \bm{0}_{24 \times 3} & -\bm{B}_{24\times 24}
\end{bmatrix}$ & $\begin{bmatrix}
-1\\
-\rho a\tilde{\bm \iota}\\
\bm{0}_{44\times 1}
\end{bmatrix}$ \tabularnewline
$y \sim I(1)$ & $\begin{bmatrix}
a\bm B^* & \bm{0}_{15\times 25}\\
\bm{0}_{10 \times 3} & \bm{0}_{10 \times 25}\\
\bm{0}_{25 \times 3} & \bm I_{25}
\end{bmatrix}$ & $\begin{bmatrix}
\bm B^* & \bm{0}_{15\times 25}\\
\bm{0}_{10 \times 3} & \bm{0}_{10 \times 25}\\
\bm{0}_{25 \times 3} & -\tilde{\bm B}_{25 \times 25}
\end{bmatrix}$ & $(1+\rho)a \cdot \begin{bmatrix}
\tilde{\bm \iota}\\
\bm{0}_{45\times 1}
\end{bmatrix}$, \tabularnewline
\hline
\end{tabularx}
\caption*{Notes: see notes in Table \ref{table:DGP WE and Dim}. Additionally, we define $b=1$ ($b \sim U(0,0.2)$) and $\tilde{\bm B}$ as a diagonal matrix with $b_{ii}=1$ ($b_{ii} \sim U(0,0.2)$) in the absence (presence) of persistence, and $\bm B^* = (\bm{1}_{3\times 3} \otimes \tilde{\bm \iota})$.}
\end{table}
The parameter settings for the varying DGPs, displayed in Table \ref{table:DGP mixed order}, are chosen such that they allow for a subset of stationary variables in the system. In particular, we first consider a scenario in which the dependent variable itself admits a stationary autoregressive representation in levels. In addition, based on their cross-sectional ordering, the first 15 variables after $y$ are cointegrated based on three cointegrating vectors, the next 10 variables are non-cointegrated random walks, and the last 24 variables all admit a stationary autoregressive structure in levels. The degree of persistence in the stationary variables is regulated by the diagonal matrix $\tilde{\bm B}$ in $\bm B$, with elements $b_{ii} = 1$ in the low persistence case and $b_{ii} \sim U(0,0.2)$ in the high persistence case. It can be seen from the last column in Table \ref{table:DGP mixed order}, that in line with the stationarity of the dependent variable, the first element in $\bm \delta$ will always be equal to $-1$, whereas an additional five-dimensional cointegrating vector enters the single-equation model for positive values of $a$. For the scenario in which the dependent variable is integrated of order one, the first 15 variables (including $y$) are all cointegrated based on three cointegrating vectors, the next 10 variables are non-cointegrated random walks, whereas the last 15 variables all admit a stationary autoregressive representation. The persistence in the stationary variables is regulated similar to the previous case. Now, however, it is clear from the last column in Table \ref{table:DGP mixed order} that $\bm \delta \neq \bm{0}$ only if $a > 0$, such that lagged levels only enter the single-equation when $y$ is cointegrated with its neighbouring variables. We display the performance of the models in Figure \ref{Fig: metrics2}.
\begin{sidewaysfigure}
\includegraphics[width=\textwidth]{SimMixedOrder}
\caption{Pseudo-Power, Proportion of Correct Selection (PCS), Proportion of Incorrect Selection (PICS) and Root Mean Squared Nowcast Error (RMSNE) for four Mixed Order specifications. The adjustment rate multiplier $a$ is on the horizontal axis.}\label{Fig: metrics2}
\end{sidewaysfigure}
In the first row of Figure \ref{Fig: metrics2}, corresponding to $y \sim I(0)$ and low persistence, SPECS correctly selects the lagged dependent variable in all simulation trials, such that the pseudo-power is always 1. Interestingly, PCS also seems constant around 35\%. Upon closer inspection, we find that SPECS chooses an alternative representation of the single-equation model in which the contribution of the non-trivial cointegrating vector seems to be absorbed in the lagged level of the dependent variable. While the resulting model differs from the implied oracle model, which is indeed accurately estimated by the OLS oracle procedure, the model choice seems motivated by a favourable bias-variance trade-off. In line with this conjecture, the nowcast performance of SPECS occasionally exceeds the OLS oracle procedure's where a larger number of parameters is estimated. The standard ADL nowcasts are again inferior, whereas the ADL-ADF model seems to benefit from correct identification of the stationarity of the dependent variable, which is particularly relevant given that the dependent variable itself is a main component in the optimal forecast. However, the nowcast accuracy of SPECS is almost identical to that of the ADL-ADF model, a finding that we interpret as reassuring and confirmatory of our claim that SPECS may be used without any pre-testing procedure. Moreover, the absence of strong persistence in the stationary variables idealizes the results of the ADL-ADF procedure.
In typical macroeconomic applications many time series that are considered as I(0) display much slower mean reversion and, consequently, are more difficult to correctly identify as being stationary.\footnote{For example, the ten time series in the popular Fred-MD dataset which \citet{McCracken2016} propose to be I(0), i.e. the series corresponding to a tcode of one, all display strong persistence or near unit root behaviour, with the smallest estimated AR(1) coefficient exceeding 0.86.} Accordingly, in row 2 we display the result for a DGP where the stationary variables display more persistent behaviour. The performance of SPECS remains largely unaffected, whereas the nowcasting performance of the ADL-ADF model deteriorates drastically. We stress the relevance of this result, given that the estimation of ADL models after pre-testing for non-stationarity is fairly common practice. Somewhat surprisingly, the ADL model in differences nowcasts almost as well as SPECS here. Overall, however, the nowcast accuracy of SPECS remains the highest and, equally important, most stable across all specifications.
Continuing the analysis of mixed order datasets, rows 3 and 4 of Figure \ref{Fig: metrics2} display the results for DGPs where the dependent variable is generated as being integrated of order one. The pseudo-power plot clearly reflects that $\bm \delta \neq \bm{0}$ only when $a>0$. Furthermore, while SPECS performs well at removing the irrelevant variables, the relevant variables are not all selected correctly, resulting in somewhat lower values for the PCS metric. Nevertheless, the nowcast performance remains superior to that of the ADL model, especially in the presence of cointegration with fast adjustment rates.
\subsection{Non-sparse Data Generating Processes}\label{Sec:adverse}
To avoid idealizing the results through a choice of DGPs that suits our estimator, this section considers the performance of the penalized regression estimators in two different non-sparse settings. First, we consider an explicitly constructed VECM that contains many small, but non-zero coefficients. Second, we consider a DGP that contains a non-stationary factor structure on which the single-equation model is likely misspecified.
The non-sparse VECM is generated according to \eqref{eq:general DGP sim} with $\bm B = \bm I_3 \otimes \tilde{\bm \iota}$, where $\tilde{\bm \iota} = (1,-\bm \iota_4^\prime)^\prime$, and $\bm A = a\bm B$ for $a=0,-0.05,\ldots,-0.5$. Hence, $N=15$ and the total number of parameters to estimate (including a constant and linear trend) is $N(p+2)+1=46$. A major difference with Section \ref{Sec:dim_we} is that we do not generate the covariance matrix of the errors as a Toeplitz-matrix, the latter being a crucial driver of sparsity in the preceding sections. Instead, we implement the procedure detailed in \citet[][p. 277-278]{Chang2004}, in which we
generate a $(N\times N)$ matrix $\bm U$ with $u_{ij} \sim U(0,1)$ to construct the orthonormal matrix $\bm H = \bm U\left(\bm U^\prime\bm U\right)^{-1/2}$,
and generate a set of $N$ eigenvalues, $\lambda_1,\ldots,\lambda_N$, where $\lambda_1 = 0.01$, $\lambda_N=1$ and $\lambda_2,\ldots,\lambda_{N-1} \sim U(0.1,1)$ to construct $\bm \varLambda = diag(\lambda_1,\ldots,\lambda_N)$. We then
construct the covariance matrix as $\bm \varSigma = \bm H\bm \varLambda\bm H^\prime$.
At each simulation trial, we generate a new $\bm \varSigma$ such that the results cannot be attributed to a specific random draw of the covariance matrix. Based on this construction, $\bm \pi_0$, as defined below \eqref{eq:CECM}, and $\bm \delta$ are non-sparse vectors with small elements; even in the setting with the strongest cointegration, i.e. $a=-0.5$, the median magnitude of the coefficients in $\bm \delta$ across all trial is only 0.12. As before, we set $T=100$ and perform 1,000 simulation trials.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{Sim_nonsparse}
\caption{Pseudo-Power, Proportion of Correct Selection (PCS), Proportion of Incorrect Selection (PICS) and Root Mean Squared Nowcast Error (RMSNE). The adjustment rate multiplier $a$ is on the horizontal axis.}
\label{Fig:non-sparse}
\end{figure}
The results are displayed in Figure \ref{Fig:non-sparse}, which contain a number of interesting results. Unsurprisingly, all estimators obtain a substantially lower (pseudo-)power in the current framework. The $\ell_1$-regularized estimators seem more sensitive to this than the traditional Wald estimator considered in \citet{Boswijk1994}. In line with the weak power, we observe that the PCS for both SPECS$_1$ and SPECS$_2$ is low, with on average only 0.75 out of 15 variables being included in levels.\footnote{The PICS is zero for all $a>0$, simply because the DGP is non-sparse, and is omitted accordingly.} Appropriate inference in the current setting is a difficult task and direct application of SPECS without alteration does not seem to be a feasible strategy. The development of a uniformly valid post-selection inference procedure, such as the desparsified lasso of \citet{Vandegeer2014}, may alleviate some of these issues. While we consider this an interesting avenue of research, it is outside the scope of the current paper.
While these results may seem discouraging, the results on the nowcast accuracy display a different story. The mean-squared nowcast errors, relative to the OLS oracle procedure, are almost always below one and are similar for the SPECS and penalized ADL estimators. This highlights that the signal of the long-run component is so weak, that the estimation of a misspecified model which ignores cointegration benefits from a favourable bias-variance tradeoff. Therefore, the conclusion remains that SPECS obtains superior predictive performance relative to methods that ignore cointegration when the long-run component provides a strong signal, without sacrificing performance absent of cointegration or in the presence of very weak cointegration.
The second, and final, DGP that we consider contains a non-stationary factor structure and corresponds to setting III in \citet[][p. 92]{Palm2011}. We allow for contemporaneous correlation and dynamic structures in both the error processes driving the ``observable'' data and the idiosyncratic component in the factor structure. The DGP is given by $\bm z_t = \bm \lambda f_t + \bm \omega_t$, where $\bm z_t$ is a $(50 \times 1)$ time series process, $f_t$ is a single scalar factor and
\begin{equation*}
f_t = \phi f_{t-1} + \zeta_t, \qquad \omega_{i,t} = \theta_i\omega_{i,t-1} + v_{i,t}.
\end{equation*}
Furthermore,
\begin{align*}
\bm v_t &= \bm A_1\bm v_{t-1} + \bm \epsilon_{1,t} + \bm B_1 \bm \epsilon_{1,t-1},\qquad
\zeta_t = \alpha_2\zeta_{t-1} + \epsilon_{2,t} + \beta_2\epsilon_{2,t-1},
\end{align*}
where $\bm \epsilon_{1,t} \sim \mathcal{N}(\bm{0},\bm \varSigma)$, with $\bm \varSigma$ again generated as in \citet{Chang2004}, and $\epsilon_{2,t} \sim \mathcal{N}(0,1)$.
The comparison focuses exclusively on the nowcasting performance for a setting without dynamics ($\bm A_1=\bm B_1=\bm{0}$ and $\alpha_2=\beta_2=0$) and a setting with dynamics ($\alpha_2=\beta_2=0.4$). The construction of $\bm A_1$ and $\bm B_1$ is analogous to \citet[][p. 93]{Palm2011}. We report the RMSNEs of SPECS relative to the ADL in Table \ref{table:factor dgp}. Given that the single-equation model is misspecified in this setup, it is unreasonable to expect SPECS to outperform. Indeed, we observe that the RMSNEs are all very close to one and, while in most cases the ADL model performs slightly better, the difference seems negligible. Hence, the risk of using SPECS to estimate a misspecified model in the sense considered here, does not seem to be higher than the use of the alternative ADL model, whereas the relative merits of SPECS when applied to a wide range of correctly specified models are evident from the first part of the simulations.
\begin{table}
\caption{Nowcasting performance on a DGP with a non-stationary factor.}
\label{table:factor dgp}
\begin{tabularx}{\textwidth}{lXXX}
\hline
\multicolumn{4}{c}{Root Mean Squared Nowcast Error}\tabularnewline
& SPECS$_1$ & SPECS$_2$ & SPECS$_1$ - OLS\tabularnewline
\hline
No Dynamics & 1.07 & 1.11 & 0.99\tabularnewline
Dynamics & 1.02 & 1.02 & 1.01\tabularnewline
\hline
\end{tabularx}
\end{table}
\section{Empirical Application}\label{Sec:application}
Inspired by \citet{Choi2012}, we consider nowcasting Dutch unemployment with SPECS based on Google Trends data. Google Trends are time series consisting of normalized indices depicting the volume of search queries entered in Google, originating from a certain geographical area. The Dutch unemployment rates are made available by Statistics Netherlands, an autonomous administrative body focussing on the collection and publication of statistical information. These rates are published on a monthly basis with new releases being made available on the 15th of each new month. This misalignment of publication dates clearly illustrate a practically relevant scenario where improvements upon forward looking predictions of Dutch unemployment rates may be obtained by utilizing contemporaneous Google Trends series.
We collect a novel dataset containing seasonally unadjusted Dutch unemployment rates from the website of Statistics Netherlands\footnote{\href{http://statline.cbs.nl/StatWeb/publication/?VW=T&DM=SLEN&PA=80479eng&LA=EN}{\textcolor{blue}{http://statline.cbs.nl/StatWeb/publication/?VW=T\&DM=SLEN\&PA=80479eng\&LA=EN}}} and a set of manually selected Google Trends time series containing unemployment related search queries, such as ``Vacancy", ``Resume" and ``Unemployment Benefits". The dataset comprises of monthly observations ranging from January 2004 to December 2017. While the full dataset contains 100 unique search queries, a number of these contain zeroes for large sub-periods, indicating insufficient search volumes for those particular series. Consequently, we remove all series that are perfectly correlated over any sub-period consisting of 20\% of the total sample.\footnote{The dataset and corresponding \textit{R} package are available at \href{https://github.com/wijler/specs}{\textcolor{blue}{https://github.com/wijler/specs}}.}
The benchmark model we consider is an ADL model fitted to the differenced data. In detail, let $y_t$ and $\bm x_t$ be the scalar unemployment rate and the vector of Google Trends series observed at time $t$, respectively, and define $\bm z_t = (y_t,\bm x_t^\prime)^\prime$. The benchmark ADL estimator fits
\begin{equation*}
\Delta y_t = \bm \pi_0^\prime \Delta \bm x_t + \sum_{j=1}^p \bm \pi_j^\prime \Delta \bm z_{t-j} + \mu_0 + \tau_0(t-1) + \epsilon_t.
\end{equation*}
However, this estimator ignores the order of integration of individual time series by differencing the whole dataset, while it is common practice to transform individual series to stationarity based on a preliminary test for unit roots. Hence, similar to Sections \ref{sec:sim_mixed} and \ref{Sec:adverse}, we include an additional ADL model where the decision to difference is based on a preliminary ADF test and refer to this method as ADL-ADF.\footnote{We note that none of the time series were found to be integrated of order 2. The outcome of the ADF test is reported for each time series in the online Appendix \ref{App: Variables}.} Finally, SPECS estimates
\begin{equation*}
\Delta y_t = \bm \delta^\prime \bm z_{t-1} + \bm \pi_0^\prime \Delta \bm x_t + \sum_{j=1}^p \bm \pi_j^\prime \Delta \bm z_{t-i} + \mu_0 + \tau_0(t-1) + \epsilon_t.
\end{equation*}
All tuning parameters are obtained by time series cross-validation and we use $k=1.1$ based on a preliminary analysis.\footnote{Comparing the nowcast accuracy for varying $k \in [0,4]$, we found the highest accuracy for $k = 1.1$.} The first nowcast is made by fitting the models on a window containing the first two-thirds of the complete sample, i.e. $t=1,\ldots,T_{c}$ with $T_c = \lceil \frac{2}{3}T \rceil$, based on which the nowcast for $\Delta y_{T_c+1}$ is produced. This procedure is repeated by rolling the window forward by one observation until the end of the sample is reached, producing a total of 54 pseudo out-of-sample nowcasts. Table \ref{Table:RMSNE Empirical} reports the MSNE relative to the ADL model for $p = 1,3,6$.
\begin{table}
\begin{tabularx}{\textwidth}{XXXXX}
\hline
$p$ &
\# of parameters & ADL-ADF & SPECS$_1$ & SPECS$_2$\tabularnewline
\hline
\hline
1 & 262 & 1.27 & 0.99 & 1.07\tabularnewline
3 & 436 & 1.06 & 0.82{*} & 0.88\tabularnewline
6 & 697 & 0.90 & 0.90 & 0.84{*}\tabularnewline
\hline
\end{tabularx}
\caption{Mean-Squared Nowcast Errora relative to the ADL model for varying number of lagged differences $p$. * denotes rejection by the Diebold-Mariano test at a 10\% significance level.}
\label{Table:RMSNE Empirical}
\end{table}
The ADL-ADF estimator does not perform better than the regular ADL model for $p=1,3$, indicating that the potential for errors in pre-testing might lead to unfavourable results. SPECS performs well and is able to obtain smaller mean-squared nowcast errors than the ADL benchmark across almost all specifications, with the combination SPECS$_2$ and $p=1$ being the exception. Moreover, for SPECS$_1$ ($p=3$) and SPECS$_2$ ($p=6$), we find the differences in MSNE to be significant at the 10\% level according to the Diebold-Mariano test. The overall (unreported) MSNE is lowest for the SPECS$_1$ estimator based on $p=3$ lagged differences. Given that the addition of lagged levels to the models improves the nowcast performance, the premise of cointegrating relationships between Dutch unemployment rates and Google Trends series seems likely. To further explore the presence of cointegration among our time series we group our variables in five categories; (1) Application Training, (2) General, (3) Job Search, (4) Recruitment Agencies (RA) and (5) Social Security. We narrow down our focus to the nowcasts of models with three lagged difference included, $p=3$, estimated by SPECS$_1$. In Figure \ref{Fig:empirical} we visually display the share of nowcasts in which the lagged levels of each variable are included in the estimated model. In addition, it depicts the selection stability of those variables, where a green colour indicates that a given variables is included in a given nowcast, and red vice versa. The figure also displays the actual unemployment rates compared to the nowcasted values.
\begin{figure}[t]
\includegraphics[width=\textwidth]{nowcastplots}
\caption{\textit{Top-left}: Selection frequency, measured as the percentage of all nowcasts the variable was selected. \textit{Bottom-left}: Selection stability with green indicating a variable was included in the nowcast model and red indicating exclusion. \textit{Right}: Actual versus predicted unemployed labour force (ULF) in levels and differences.}
\label{Fig:empirical}
\end{figure}
Figure \ref{Fig:empirical} highlights that only few variables are consistently selected for all nowcasts, although in each category we can distinguish some variables that are included at higher frequencies. The variable whose lagged levels are always selected is ``Vakantiebaan", which is a search query for a temporary job during the summer holiday. We postulate that this variable is selected by SPECS to account for seasonality in the Dutch unemployment rates. In an unreported exercise we estimate the model with the addition of a set of eleven unpenalized dummies representing different months of the year. While the variable ``Vakantiebaan" is never selected, the mean squared nowcast error increases substantially. Hence, we opt to adhere to our standard model under the caveat that for at least one of the lagged levels included, seasonality effects rather than cointegration seem a more appropriate explanation for its inclusion. Other frequently included variables are queries for vacancies (``uwv.vacatures", 78\%), unemployment (``werkloos", 76\%) and social benefits (``ww uitkering", 72\%), where the stated percentages indicate the proportion of nowcast models in which the respective variables are selected. Furthermore, the last bar represents the frequency in which the lagged level of the Dutch unemployment rate is selected, which occurs for 43 out of 54 nowcasts (80\%). The frequent selection of the lagged level of unemployment rates in conjunction with the other lagged levels is indicative of the presence of cointegration among unemployment and Google Trends series. However, we do not attach any structural meaning to the found equilibria based on the difficulty of interpretation when one does not assume the presence of weak exogeneity.
In an attempt to gain insights into the temporal stability of our estimator, we visually display the selection stability in the bottom-left part of Figure \ref{Fig:empirical}. Generally, we see that for the early and later period of the sample very few time series enter the model in levels, whereas for the middle part of the sample the majority of variables are selected. The exact reason for these patterns to occur is unknown and raises questions on the stability of Google trends as informative predictors of Dutch unemployment rates. Standard feasible explanations concern structural instability in the DGP, seasonality effects or data idiosyncrasies. However, there are additional peculiarities specific to the use of Google trends such as normalization, data hubris and search algorithm dynamics, all of which might result in unstable performance \citep[cf.][]{Lazer2014}. Since the focus of this application is on the relative performance between our estimator and a common benchmark model, rather than on a structural analysis of the relation between Google Trends and unemployment rates, we leave this issue aside as it is outside the scope of the paper. Instead, we focus on the relative empirical performance of our methods, which, notwithstanding the aforementioned caveats, we deem convincingly favourable for SPECS. Finally, on the right of Figure \ref{Fig:empirical} we display the realized and predicted unemployment rates in levels and differences. Both the penalized ADL model and SPECS seem to follow the actual unemployment rates with reasonable accuracy, with the largest nowcast errors occurring in the first half of 2014. Prior to this period the unemployment rates had been steadily rising in the aftermath of the economic recession, whereas 2014 marks the start of a recovery period. Given that the models are fit on historical data, it is natural that the estimators overestimate the unemployment rate shortly after the start of the economic recovery. Perhaps not entirely coincidental, the start of the period over which the majority of lagged levels are included by SPECS coincides with this recovery period as well, thereby hinting towards structural instability in the DGP as a plausible cause for the observed selection instability.
\section{Conclusion}\label{Sec:Conclusion}
In this paper, we propose the use of SPECS as an automated approach for sparse single-equation error correction modelling in high-dimensional settings. SPECS is an intuitive estimator that applies penalized regression to a conditional error-correction model. We show that SPECS possesses the oracle property and is able to consistently select the long-run and short-run dynamics in the underlying DGP. These results are derived with the aid of a novel bound on the minimum eigenvalue of the sample covariance matrix containing integrated process, which may be of independent interest. Additionally, in pursuit of suitable weights that aid in the identification of the subset of relevant variables, we derive the consistency of the ridge estimator applied to the same model and demonstrate how ridge regression may be used to construct these weights.
We document favourable finite sample performance of SPECS by means of simulations and an empirical application. The simulation exercise confirms strong selective and predictive capabilities in both low and high dimensions with convincing gains over a benchmark penalized ADL model that ignores cointegration in the dataset. Furthermore, the simulation results demonstrate that the selective capabilities of SPECS remain adequate absent of weak exogeneity and the nowcasting performance remains superior to the benchmark. Finally, we consider an empirical application in which we nowcast the Dutch unemployment rate with the use of Google Trends series. Across all three different dynamic specifications considered, SPECS attains higher nowcast accuracy, thus confirming the findings from our simulation study. As a result, we believe that our proposed estimator, which is easily implemented with readily available tools at a low computational cost, offers a valuable tool for practitioners by enabling automated model estimation on relatively large and potentially non-stationary datasets and, most importantly, allowing to take into account potential (co)integration without requiring pre-testing procedures.
Finally, we highlight several important sources through which the assumptions and asymptotic framework may be generalized further. Sharper and more direct eigenvalue bounds can be utilized to cast SPECS into an even higher-dimensional setting. Similarly, a suitable compatibility condition can be used to validate the lasso as an initial estimator, resulting in improved weights and, again, a less restrictive asymptotic framework. These topics remain subject to our continuing investigation.
\label{Bibliography}
\bibliographystyle{apalike}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.